fix: link subagent sessions to parent and hide from session list

Subagent sessions spawned by delegate_task were created with
parent_session_id=NULL and source=cli, making them indistinguishable
from user sessions in hermes sessions list and /resume.

Changes:
- delegate_tool.py: pass parent_agent.session_id to child agent
- run_agent.py: accept parent_session_id param, pass to create_session
- hermes_state.py list_sessions_rich: filter parent_session_id IS NULL
  by default (opt-in include_children=True for callers that need them)
- hermes_state.py delete_session: delete child sessions first (FK)
- hermes_state.py prune_sessions: delete children before parents (FK)

session_search already handles parent_session_id correctly — child
sessions are filtered from recent list and resolved to parent root
in full-text search results.

Fixes #5122

.dockerignore

@@ -0,0 +1,15 @@
+# Git
+.git
+.gitignore
+.gitmodules
+
+# Dependencies
+node_modules
+
+# CI/CD
+.github
+
+# Environment files
+.env
+
+*.md

.env.example

@@ -7,18 +7,19 @@
 # OpenRouter provides access to many models through one API
 # All LLM calls go through OpenRouter - no direct provider keys needed
 # Get your key at: https://openrouter.ai/keys
-OPENROUTER_API_KEY=
+# OPENROUTER_API_KEY=
 
-# Default model to use (OpenRouter format: provider/model)
-# Examples: anthropic/claude-opus-4.6, openai/gpt-4o, google/gemini-3-flash-preview, zhipuai/glm-4-plus
-LLM_MODEL=anthropic/claude-opus-4.6
+# Default model is configured in ~/.hermes/config.yaml (model.default).
+# Use 'hermes model' or 'hermes setup' to change it.
+# LLM_MODEL is no longer read from .env — this line is kept for reference only.
+# LLM_MODEL=anthropic/claude-opus-4.6
 
 # =============================================================================
 # LLM PROVIDER (z.ai / GLM)
 # =============================================================================
 # z.ai provides access to ZhipuAI GLM models (GLM-4-Plus, etc.)
 # Get your key at: https://z.ai or https://open.bigmodel.cn
-GLM_API_KEY=
+# GLM_API_KEY=
 # GLM_BASE_URL=https://api.z.ai/api/paas/v4  # Override default base URL
 
 # =============================================================================
@@ -28,7 +29,7 @@ GLM_API_KEY=
 # Get your key at: https://platform.kimi.ai (Kimi Code console)
 # Keys prefixed sk-kimi- use the Kimi Code API (api.kimi.com) by default.
 # Legacy keys from platform.moonshot.ai need KIMI_BASE_URL override below.
-KIMI_API_KEY=
+# KIMI_API_KEY=
 # KIMI_BASE_URL=https://api.kimi.com/coding/v1  # Default for sk-kimi- keys
 # KIMI_BASE_URL=https://api.moonshot.ai/v1      # For legacy Moonshot keys
 # KIMI_BASE_URL=https://api.moonshot.cn/v1       # For Moonshot China keys
@@ -38,33 +39,67 @@ KIMI_API_KEY=
 # =============================================================================
 # MiniMax provides access to MiniMax models (global endpoint)
 # Get your key at: https://www.minimax.io
-MINIMAX_API_KEY=
+# MINIMAX_API_KEY=
 # MINIMAX_BASE_URL=https://api.minimax.io/v1  # Override default base URL
 
 # MiniMax China endpoint (for users in mainland China)
-MINIMAX_CN_API_KEY=
+# MINIMAX_CN_API_KEY=
 # MINIMAX_CN_BASE_URL=https://api.minimaxi.com/v1  # Override default base URL
 
+# =============================================================================
+# LLM PROVIDER (OpenCode Zen)
+# =============================================================================
+# OpenCode Zen provides curated, tested models (GPT, Claude, Gemini, MiniMax, GLM, Kimi)
+# Pay-as-you-go pricing. Get your key at: https://opencode.ai/auth
+# OPENCODE_ZEN_API_KEY=
+# OPENCODE_ZEN_BASE_URL=https://opencode.ai/zen/v1  # Override default base URL
+
+# =============================================================================
+# LLM PROVIDER (OpenCode Go)
+# =============================================================================
+# OpenCode Go provides access to open models (GLM-5, Kimi K2.5, MiniMax M2.5)
+# $10/month subscription. Get your key at: https://opencode.ai/auth
+# OPENCODE_GO_API_KEY=
+
+# =============================================================================
+# LLM PROVIDER (Hugging Face Inference Providers)
+# =============================================================================
+# Hugging Face routes to 20+ open models via unified OpenAI-compatible endpoint.
+# Free tier included ($0.10/month), no markup on provider rates.
+# Get your token at: https://huggingface.co/settings/tokens
+# Required permission: "Make calls to Inference Providers"
+# HF_TOKEN=
+# OPENCODE_GO_BASE_URL=https://opencode.ai/zen/go/v1  # Override default base URL
+
 # =============================================================================
 # TOOL API KEYS
 # =============================================================================
 
+# Exa API Key - AI-native web search and contents
+# Get at: https://exa.ai
+# EXA_API_KEY=
+
+# Parallel API Key - AI-native web search and extract
+# Get at: https://parallel.ai
+# PARALLEL_API_KEY=
+
 # Firecrawl API Key - Web search, extract, and crawl
 # Get at: https://firecrawl.dev/
-FIRECRAWL_API_KEY=
+# FIRECRAWL_API_KEY=
+
 
 # FAL.ai API Key - Image generation
 # Get at: https://fal.ai/
-FAL_KEY=
+# FAL_KEY=
 
 # Honcho - Cross-session AI-native user modeling (optional)
 # Builds a persistent understanding of the user across sessions and tools.
 # Get at: https://app.honcho.dev
 # Also requires ~/.honcho/config.json with enabled=true (see README).
-HONCHO_API_KEY=
+# HONCHO_API_KEY=
 
 # =============================================================================
-# TERMINAL TOOL CONFIGURATION (mini-swe-agent backend)
+# TERMINAL TOOL CONFIGURATION
 # =============================================================================
 # Backend type: "local", "singularity", "docker", "modal", or "ssh"
 # Terminal backend is configured in ~/.hermes/config.yaml (terminal.backend).
@@ -147,10 +182,10 @@ TERMINAL_LIFETIME_SECONDS=300
 
 # Browserbase API Key - Cloud browser execution
 # Get at: https://browserbase.com/
-BROWSERBASE_API_KEY=
+# BROWSERBASE_API_KEY=
 
 # Browserbase Project ID - From your Browserbase dashboard
-BROWSERBASE_PROJECT_ID=
+# BROWSERBASE_PROJECT_ID=
 
 # Enable residential proxies for better CAPTCHA solving (default: true)
 # Routes traffic through residential IPs, significantly improves success rate
@@ -182,7 +217,7 @@ BROWSER_INACTIVITY_TIMEOUT=120
 # Uses OpenAI's API directly (not via OpenRouter).
 # Named VOICE_TOOLS_OPENAI_KEY to avoid interference with OpenRouter.
 # Get at: https://platform.openai.com/api-keys
-VOICE_TOOLS_OPENAI_KEY=
+# VOICE_TOOLS_OPENAI_KEY=
 
 # =============================================================================
 # SLACK INTEGRATION
@@ -197,6 +232,21 @@ VOICE_TOOLS_OPENAI_KEY=
 # Slack allowed users (comma-separated Slack user IDs)
 # SLACK_ALLOWED_USERS=
 
+# =============================================================================
+# TELEGRAM INTEGRATION
+# =============================================================================
+# Telegram Bot Token - From @BotFather (https://t.me/BotFather)
+# TELEGRAM_BOT_TOKEN=
+# TELEGRAM_ALLOWED_USERS=                  # Comma-separated user IDs
+# TELEGRAM_HOME_CHANNEL=                   # Default chat for cron delivery
+# TELEGRAM_HOME_CHANNEL_NAME=              # Display name for home channel
+
+# Webhook mode (optional — for cloud deployments like Fly.io/Railway)
+# Default is long polling. Setting TELEGRAM_WEBHOOK_URL switches to webhook mode.
+# TELEGRAM_WEBHOOK_URL=https://my-app.fly.dev/telegram
+# TELEGRAM_WEBHOOK_PORT=8443
+# TELEGRAM_WEBHOOK_SECRET=                 # Recommended for production
+
 # WhatsApp (built-in Baileys bridge — run `hermes whatsapp` to pair)
 # WHATSAPP_ENABLED=false
 # WHATSAPP_ALLOWED_USERS=15551234567
@@ -253,11 +303,11 @@ IMAGE_TOOLS_DEBUG=false
 
 # Tinker API Key - RL training service
 # Get at: https://tinker-console.thinkingmachines.ai/keys
-TINKER_API_KEY=
+# TINKER_API_KEY=
 
 # Weights & Biases API Key - Experiment tracking and metrics
 # Get at: https://wandb.ai/authorize
-WANDB_API_KEY=
+# WANDB_API_KEY=
 
 # RL API Server URL (default: http://localhost:8080)
 # Change if running the rl-server on a different host/port

.envrc

@@ -0,0 +1 @@
+use flake

.github/workflows/deploy-site.yml

@@ -6,6 +6,8 @@ on:
     paths:
       - 'website/**'
       - 'landingpage/**'
+      - 'skills/**'
+      - 'optional-skills/**'
       - '.github/workflows/deploy-site.yml'
   workflow_dispatch:
 
@@ -19,6 +21,8 @@ concurrency:
 
 jobs:
   build-and-deploy:
+    # Only run on the upstream repository, not on forks
+    if: github.repository == 'NousResearch/hermes-agent'
     runs-on: ubuntu-latest
     environment:
       name: github-pages
@@ -32,6 +36,16 @@ jobs:
           cache: npm
           cache-dependency-path: website/package-lock.json
 
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Install PyYAML for skill extraction
+        run: pip install pyyaml
+
+      - name: Extract skill metadata for dashboard
+        run: python3 website/scripts/extract-skills.py
+
       - name: Install dependencies
         run: npm ci
         working-directory: website

.github/workflows/docker-publish.yml

@@ -0,0 +1,79 @@
+name: Docker Build and Publish
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+  release:
+    types: [published]
+
+concurrency:
+  group: docker-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  build-and-push:
+    # Only run on the upstream repository, not on forks
+    if: github.repository == 'NousResearch/hermes-agent'
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          submodules: recursive
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Build image
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: Dockerfile
+          load: true
+          tags: nousresearch/hermes-agent:test
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
+
+      - name: Test image starts
+        run: |
+          docker run --rm \
+            -v /tmp/hermes-test:/opt/data \
+            --entrypoint /opt/hermes/docker/entrypoint.sh \
+            nousresearch/hermes-agent:test --help
+
+      - name: Log in to Docker Hub
+        if: github.event_name == 'push' && github.ref == 'refs/heads/main' || github.event_name == 'release'
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+
+      - name: Push image (main branch)
+        if: github.event_name == 'push' && github.ref == 'refs/heads/main'
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: Dockerfile
+          push: true
+          tags: |
+            nousresearch/hermes-agent:latest
+            nousresearch/hermes-agent:${{ github.sha }}
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
+
+      - name: Push image (release)
+        if: github.event_name == 'release'
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: Dockerfile
+          push: true
+          tags: |
+            nousresearch/hermes-agent:latest
+            nousresearch/hermes-agent:${{ github.event.release.tag_name }}
+            nousresearch/hermes-agent:${{ github.sha }}
+          cache-from: type=gha
+          cache-to: type=gha,mode=max

.github/workflows/docs-site-checks.yml

@@ -0,0 +1,42 @@
+name: Docs Site Checks
+
+on:
+  pull_request:
+    paths:
+      - 'website/**'
+      - '.github/workflows/docs-site-checks.yml'
+  workflow_dispatch:
+
+jobs:
+  docs-site-checks:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version: 20
+          cache: npm
+          cache-dependency-path: website/package-lock.json
+
+      - name: Install website dependencies
+        run: npm ci
+        working-directory: website
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Install Python dependencies
+        run: python -m pip install ascii-guard pyyaml
+
+      - name: Extract skill metadata for dashboard
+        run: python3 website/scripts/extract-skills.py
+
+      - name: Lint docs diagrams
+        run: npm run lint:diagrams
+        working-directory: website
+
+      - name: Build Docusaurus
+        run: npm run build
+        working-directory: website

.github/workflows/nix.yml

@@ -0,0 +1,40 @@
+name: Nix
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    paths:
+      - 'flake.nix'
+      - 'flake.lock'
+      - 'nix/**'
+      - 'pyproject.toml'
+      - 'uv.lock'
+      - 'hermes_cli/**'
+      - 'run_agent.py'
+      - 'acp_adapter/**'
+
+concurrency:
+  group: nix-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  nix:
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest]
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 30
+    steps:
+      - uses: actions/checkout@v4
+      - uses: DeterminateSystems/nix-installer-action@main
+      - uses: DeterminateSystems/magic-nix-cache-action@main
+      - name: Check flake
+        if: runner.os == 'Linux'
+        run: nix flake check --print-build-logs
+      - name: Build package
+        if: runner.os == 'Linux'
+        run: nix build --print-build-logs
+      - name: Evaluate flake (macOS)
+        if: runner.os == 'macOS'
+        run: nix flake show --json > /dev/null

.github/workflows/supply-chain-audit.yml

@@ -0,0 +1,192 @@
+name: Supply Chain Audit
+
+on:
+  pull_request:
+    types: [opened, synchronize, reopened]
+
+permissions:
+  pull-requests: write
+  contents: read
+
+jobs:
+  scan:
+    name: Scan PR for supply chain risks
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Scan diff for suspicious patterns
+        id: scan
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -euo pipefail
+
+          BASE="${{ github.event.pull_request.base.sha }}"
+          HEAD="${{ github.event.pull_request.head.sha }}"
+
+          # Get the full diff (added lines only)
+          DIFF=$(git diff "$BASE".."$HEAD" -- . ':!uv.lock' ':!*.lock' ':!package-lock.json' ':!yarn.lock' || true)
+
+          FINDINGS=""
+          CRITICAL=false
+
+          # --- .pth files (auto-execute on Python startup) ---
+          PTH_FILES=$(git diff --name-only "$BASE".."$HEAD" | grep '\.pth$' || true)
+          if [ -n "$PTH_FILES" ]; then
+            CRITICAL=true
+            FINDINGS="${FINDINGS}
+          ### 🚨 CRITICAL: .pth file added or modified
+          Python \`.pth\` files in \`site-packages/\` execute automatically when the interpreter starts — no import required. This is the exact mechanism used in the [litellm supply chain attack](https://github.com/BerriAI/litellm/issues/24512).
+
+          **Files:**
+          \`\`\`
+          ${PTH_FILES}
+          \`\`\`
+          "
+          fi
+
+          # --- base64 + exec/eval combo (the litellm attack pattern) ---
+          B64_EXEC_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -iE 'base64\.(b64decode|decodebytes|urlsafe_b64decode)' | grep -iE 'exec\(|eval\(' | head -10 || true)
+          if [ -n "$B64_EXEC_HITS" ]; then
+            CRITICAL=true
+            FINDINGS="${FINDINGS}
+          ### 🚨 CRITICAL: base64 decode + exec/eval combo
+          This is the exact pattern used in the [litellm supply chain attack](https://github.com/BerriAI/litellm/issues/24512) — base64-decoded strings passed to exec/eval to hide credential-stealing payloads.
+
+          **Matches:**
+          \`\`\`
+          ${B64_EXEC_HITS}
+          \`\`\`
+          "
+          fi
+
+          # --- base64 decode/encode (alone — legitimate uses exist) ---
+          B64_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -iE 'base64\.(b64decode|b64encode|decodebytes|encodebytes|urlsafe_b64decode)|atob\(|btoa\(|Buffer\.from\(.*base64' | head -20 || true)
+          if [ -n "$B64_HITS" ]; then
+            FINDINGS="${FINDINGS}
+          ### ⚠️ WARNING: base64 encoding/decoding detected
+          Base64 has legitimate uses (images, JWT, etc.) but is also commonly used to obfuscate malicious payloads. Verify the usage is appropriate.
+
+          **Matches (first 20):**
+          \`\`\`
+          ${B64_HITS}
+          \`\`\`
+          "
+          fi
+
+          # --- exec/eval with string arguments ---
+          EXEC_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -E '(exec|eval)\s*\(' | grep -v '^\+\s*#' | grep -v 'test_\|mock\|assert\|# ' | head -20 || true)
+          if [ -n "$EXEC_HITS" ]; then
+            FINDINGS="${FINDINGS}
+          ### ⚠️ WARNING: exec() or eval() usage
+          Dynamic code execution can hide malicious behavior, especially when combined with base64 or network fetches.
+
+          **Matches (first 20):**
+          \`\`\`
+          ${EXEC_HITS}
+          \`\`\`
+          "
+          fi
+
+          # --- subprocess with encoded/obfuscated commands ---
+          PROC_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -E 'subprocess\.(Popen|call|run)\s*\(' | grep -iE 'base64|decode|encode|\\x|chr\(' | head -10 || true)
+          if [ -n "$PROC_HITS" ]; then
+            CRITICAL=true
+            FINDINGS="${FINDINGS}
+          ### 🚨 CRITICAL: subprocess with encoded/obfuscated command
+          Subprocess calls with encoded arguments are a strong indicator of payload execution.
+
+          **Matches:**
+          \`\`\`
+          ${PROC_HITS}
+          \`\`\`
+          "
+          fi
+
+          # --- Network calls to non-standard domains ---
+          EXFIL_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -iE 'requests\.(post|put)\(|httpx\.(post|put)\(|urllib\.request\.urlopen' | grep -v '^\+\s*#' | grep -v 'test_\|mock\|assert' | head -10 || true)
+          if [ -n "$EXFIL_HITS" ]; then
+            FINDINGS="${FINDINGS}
+          ### ⚠️ WARNING: Outbound network calls (POST/PUT)
+          Outbound POST/PUT requests in new code could be data exfiltration. Verify the destination URLs are legitimate.
+
+          **Matches (first 10):**
+          \`\`\`
+          ${EXFIL_HITS}
+          \`\`\`
+          "
+          fi
+
+          # --- setup.py / setup.cfg install hooks ---
+          SETUP_HITS=$(git diff --name-only "$BASE".."$HEAD" | grep -E '(setup\.py|setup\.cfg|__init__\.pth|sitecustomize\.py|usercustomize\.py)$' || true)
+          if [ -n "$SETUP_HITS" ]; then
+            FINDINGS="${FINDINGS}
+          ### ⚠️ WARNING: Install hook files modified
+          These files can execute code during package installation or interpreter startup.
+
+          **Files:**
+          \`\`\`
+          ${SETUP_HITS}
+          \`\`\`
+          "
+          fi
+
+          # --- Compile/marshal/pickle (code object injection) ---
+          MARSHAL_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -iE 'marshal\.loads|pickle\.loads|compile\(' | grep -v '^\+\s*#' | grep -v 'test_\|re\.compile\|ast\.compile' | head -10 || true)
+          if [ -n "$MARSHAL_HITS" ]; then
+            FINDINGS="${FINDINGS}
+          ### ⚠️ WARNING: marshal/pickle/compile usage
+          These can deserialize or construct executable code objects.
+
+          **Matches:**
+          \`\`\`
+          ${MARSHAL_HITS}
+          \`\`\`
+          "
+          fi
+
+          # --- Output results ---
+          if [ -n "$FINDINGS" ]; then
+            echo "found=true" >> "$GITHUB_OUTPUT"
+            if [ "$CRITICAL" = true ]; then
+              echo "critical=true" >> "$GITHUB_OUTPUT"
+            else
+              echo "critical=false" >> "$GITHUB_OUTPUT"
+            fi
+            # Write findings to a file (multiline env vars are fragile)
+            echo "$FINDINGS" > /tmp/findings.md
+          else
+            echo "found=false" >> "$GITHUB_OUTPUT"
+            echo "critical=false" >> "$GITHUB_OUTPUT"
+          fi
+
+      - name: Post warning comment
+        if: steps.scan.outputs.found == 'true'
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          SEVERITY="⚠️ Supply Chain Risk Detected"
+          if [ "${{ steps.scan.outputs.critical }}" = "true" ]; then
+            SEVERITY="🚨 CRITICAL Supply Chain Risk Detected"
+          fi
+
+          BODY="## ${SEVERITY}
+
+          This PR contains patterns commonly associated with supply chain attacks. This does **not** mean the PR is malicious — but these patterns require careful human review before merging.
+
+          $(cat /tmp/findings.md)
+
+          ---
+          *Automated scan triggered by [supply-chain-audit](/.github/workflows/supply-chain-audit.yml). If this is a false positive, a maintainer can approve after manual review.*"
+
+          gh pr comment "${{ github.event.pull_request.number }}" --body "$BODY"
+
+      - name: Fail on critical findings
+        if: steps.scan.outputs.critical == 'true'
+        run: |
+          echo "::error::CRITICAL supply chain risk patterns detected in this PR. See the PR comment for details."
+          exit 1

.github/workflows/tests.yml

@@ -34,9 +34,37 @@ jobs:
       - name: Run tests
         run: |
           source .venv/bin/activate
-          python -m pytest tests/ -q --ignore=tests/integration --tb=short -n auto
+          python -m pytest tests/ -q --ignore=tests/integration --ignore=tests/e2e --tb=short -n auto
         env:
           # Ensure tests don't accidentally call real APIs
           OPENROUTER_API_KEY: ""
           OPENAI_API_KEY: ""
           NOUS_API_KEY: ""
+
+  e2e:
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v5
+
+      - name: Set up Python 3.11
+        run: uv python install 3.11
+
+      - name: Install dependencies
+        run: |
+          uv venv .venv --python 3.11
+          source .venv/bin/activate
+          uv pip install -e ".[all,dev]"
+
+      - name: Run e2e tests
+        run: |
+          source .venv/bin/activate
+          python -m pytest tests/e2e/ -v --tb=short
+        env:
+          OPENROUTER_API_KEY: ""
+          OPENAI_API_KEY: ""
+          NOUS_API_KEY: ""

.gitignore

@@ -53,3 +53,8 @@ environments/benchmarks/evals/
 
 # Release script temp files
 .release_notes.md
+mini-swe-agent/
+
+# Nix
+.direnv/
+result

.gitmodules

@@ -1,6 +1,3 @@
-[submodule "mini-swe-agent"]
-	path = mini-swe-agent
-	url = https://github.com/SWE-agent/mini-swe-agent
 [submodule "tinker-atropos"]
 	path = tinker-atropos
 	url = https://github.com/nousresearch/tinker-atropos

AGENTS.md

@@ -5,7 +5,7 @@ Instructions for AI coding assistants and developers working on the hermes-agent
 ## Development Environment
 
 ```bash
-source .venv/bin/activate  # ALWAYS activate before running Python
+source venv/bin/activate  # ALWAYS activate before running Python
 ```
 
 ## Project Structure
@@ -23,6 +23,7 @@ hermes-agent/
 │   ├── prompt_caching.py     # Anthropic prompt caching
 │   ├── auxiliary_client.py   # Auxiliary LLM client (vision, summarization)
 │   ├── model_metadata.py     # Model context lengths, token estimation
+│   ├── models_dev.py         # models.dev registry integration (provider-aware context)
 │   ├── display.py            # KawaiiSpinner, tool preview formatting
 │   ├── skill_commands.py     # Skill slash commands (shared CLI/gateway)
 │   └── trajectory.py         # Trajectory saving helpers
@@ -37,6 +38,7 @@ hermes-agent/
 │   ├── tools_config.py   # `hermes tools` — enable/disable tools per platform
 │   ├── skills_hub.py     # `/skills` slash command (search, browse, install)
 │   ├── models.py         # Model catalog, provider model lists
+│   ├── model_switch.py   # Shared /model switch pipeline (CLI + gateway)
 │   └── auth.py           # Provider credential resolution
 ├── tools/                # Tool implementations (one file per tool)
 │   ├── registry.py       # Central tool registry (schemas, handlers, dispatch)
@@ -44,7 +46,7 @@ hermes-agent/
 │   ├── terminal_tool.py  # Terminal orchestration
 │   ├── process_registry.py # Background process management
 │   ├── file_tools.py     # File read/write/search/patch
-│   ├── web_tools.py      # Firecrawl search/extract
+│   ├── web_tools.py      # Web search/extract (Parallel + Firecrawl)
 │   ├── browser_tool.py   # Browserbase browser automation
 │   ├── code_execution_tool.py # execute_code sandbox
 │   ├── delegate_tool.py  # Subagent delegation
@@ -129,14 +131,51 @@ Messages follow OpenAI format: `{"role": "system/user/assistant/tool", ...}`. Re
 - **KawaiiSpinner** (`agent/display.py`) — animated faces during API calls, `┊` activity feed for tool results
 - `load_cli_config()` in cli.py merges hardcoded defaults + user config YAML
 - **Skin engine** (`hermes_cli/skin_engine.py`) — data-driven CLI theming; initialized from `display.skin` config key at startup; skins customize banner colors, spinner faces/verbs/wings, tool prefix, response box, branding text
-- `process_command()` is a method on `HermesCLI` (not in commands.py)
+- `process_command()` is a method on `HermesCLI` — dispatches on canonical command name resolved via `resolve_command()` from the central registry
 - Skill slash commands: `agent/skill_commands.py` scans `~/.hermes/skills/`, injects as **user message** (not system prompt) to preserve prompt caching
 
-### Adding CLI Commands
+### Slash Command Registry (`hermes_cli/commands.py`)
 
-1. Add to `COMMANDS` dict in `hermes_cli/commands.py`
-2. Add handler in `HermesCLI.process_command()` in `cli.py`
-3. For persistent settings, use `save_config_value()` in `cli.py`
+All slash commands are defined in a central `COMMAND_REGISTRY` list of `CommandDef` objects. Every downstream consumer derives from this registry automatically:
+
+- **CLI** — `process_command()` resolves aliases via `resolve_command()`, dispatches on canonical name
+- **Gateway** — `GATEWAY_KNOWN_COMMANDS` frozenset for hook emission, `resolve_command()` for dispatch
+- **Gateway help** — `gateway_help_lines()` generates `/help` output
+- **Telegram** — `telegram_bot_commands()` generates the BotCommand menu
+- **Slack** — `slack_subcommand_map()` generates `/hermes` subcommand routing
+- **Autocomplete** — `COMMANDS` flat dict feeds `SlashCommandCompleter`
+- **CLI help** — `COMMANDS_BY_CATEGORY` dict feeds `show_help()`
+
+### Adding a Slash Command
+
+1. Add a `CommandDef` entry to `COMMAND_REGISTRY` in `hermes_cli/commands.py`:
+```python
+CommandDef("mycommand", "Description of what it does", "Session",
+           aliases=("mc",), args_hint="[arg]"),
+```
+2. Add handler in `HermesCLI.process_command()` in `cli.py`:
+```python
+elif canonical == "mycommand":
+    self._handle_mycommand(cmd_original)
+```
+3. If the command is available in the gateway, add a handler in `gateway/run.py`:
+```python
+if canonical == "mycommand":
+    return await self._handle_mycommand(event)
+```
+4. For persistent settings, use `save_config_value()` in `cli.py`
+
+**CommandDef fields:**
+- `name` — canonical name without slash (e.g. `"background"`)
+- `description` — human-readable description
+- `category` — one of `"Session"`, `"Configuration"`, `"Tools & Skills"`, `"Info"`, `"Exit"`
+- `aliases` — tuple of alternative names (e.g. `("bg",)`)
+- `args_hint` — argument placeholder shown in help (e.g. `"<prompt>"`, `"[name]"`)
+- `cli_only` — only available in the interactive CLI
+- `gateway_only` — only available in messaging platforms
+- `gateway_config_gate` — config dotpath (e.g. `"display.tool_progress_command"`); when set on a `cli_only` command, the command becomes available in the gateway if the config value is truthy. `GATEWAY_KNOWN_COMMANDS` always includes config-gated commands so the gateway can dispatch them; help/menus only show them when the gate is open.
+
+**Adding an alias** requires only adding it to the `aliases` tuple on the existing `CommandDef`. No other file changes needed — dispatch, help text, Telegram menu, Slack mapping, and autocomplete all update automatically.
 
 ---
 
@@ -171,6 +210,10 @@ registry.register(
 
 The registry handles schema collection, dispatch, availability checking, and error wrapping. All handlers MUST return a JSON string.
 
+**Path references in tool schemas**: If the schema description mentions file paths (e.g. default output directories), use `display_hermes_home()` to make them profile-aware. The schema is generated at import time, which is after `_apply_profile_override()` sets `HERMES_HOME`.
+
+**State files**: If a tool stores persistent state (caches, logs, checkpoints), use `get_hermes_home()` for the base directory — never `Path.home() / ".hermes"`. This ensures each profile gets its own state.
+
 **Agent-level tools** (todo, memory): intercepted by `run_agent.py` before `handle_function_call()`. See `todo_tool.py` for the pattern.
 
 ---
@@ -235,6 +278,7 @@ hermes_cli/skin_engine.py    # SkinConfig dataclass, built-in skins, YAML loader
 | Spinner verbs | `spinner.thinking_verbs` | `display.py` |
 | Spinner wings (optional) | `spinner.wings` | `display.py` |
 | Tool output prefix | `tool_prefix` | `display.py` |
+| Per-tool emojis | `tool_emojis` | `display.py` → `get_tool_emoji()` |
 | Agent name | `branding.agent_name` | `banner.py`, `cli.py` |
 | Welcome message | `branding.welcome` | `cli.py` |
 | Response box label | `branding.response_label` | `cli.py` |
@@ -318,8 +362,69 @@ in config.yaml (or `HERMES_BACKGROUND_NOTIFICATIONS` env var):
 
 ---
 
+## Profiles: Multi-Instance Support
+
+Hermes supports **profiles** — multiple fully isolated instances, each with its own
+`HERMES_HOME` directory (config, API keys, memory, sessions, skills, gateway, etc.).
+
+The core mechanism: `_apply_profile_override()` in `hermes_cli/main.py` sets
+`HERMES_HOME` before any module imports. All 119+ references to `get_hermes_home()`
+automatically scope to the active profile.
+
+### Rules for profile-safe code
+
+1. **Use `get_hermes_home()` for all HERMES_HOME paths.** Import from `hermes_constants`.
+   NEVER hardcode `~/.hermes` or `Path.home() / ".hermes"` in code that reads/writes state.
+   ```python
+   # GOOD
+   from hermes_constants import get_hermes_home
+   config_path = get_hermes_home() / "config.yaml"
+
+   # BAD — breaks profiles
+   config_path = Path.home() / ".hermes" / "config.yaml"
+   ```
+
+2. **Use `display_hermes_home()` for user-facing messages.** Import from `hermes_constants`.
+   This returns `~/.hermes` for default or `~/.hermes/profiles/<name>` for profiles.
+   ```python
+   # GOOD
+   from hermes_constants import display_hermes_home
+   print(f"Config saved to {display_hermes_home()}/config.yaml")
+
+   # BAD — shows wrong path for profiles
+   print("Config saved to ~/.hermes/config.yaml")
+   ```
+
+3. **Module-level constants are fine** — they cache `get_hermes_home()` at import time,
+   which is AFTER `_apply_profile_override()` sets the env var. Just use `get_hermes_home()`,
+   not `Path.home() / ".hermes"`.
+
+4. **Tests that mock `Path.home()` must also set `HERMES_HOME`** — since code now uses
+   `get_hermes_home()` (reads env var), not `Path.home() / ".hermes"`:
+   ```python
+   with patch.object(Path, "home", return_value=tmp_path), \
+        patch.dict(os.environ, {"HERMES_HOME": str(tmp_path / ".hermes")}):
+       ...
+   ```
+
+5. **Gateway platform adapters should use token locks** — if the adapter connects with
+   a unique credential (bot token, API key), call `acquire_scoped_lock()` from
+   `gateway.status` in the `connect()`/`start()` method and `release_scoped_lock()` in
+   `disconnect()`/`stop()`. This prevents two profiles from using the same credential.
+   See `gateway/platforms/telegram.py` for the canonical pattern.
+
+6. **Profile operations are HOME-anchored, not HERMES_HOME-anchored** — `_get_profiles_root()`
+   returns `Path.home() / ".hermes" / "profiles"`, NOT `get_hermes_home() / "profiles"`.
+   This is intentional — it lets `hermes -p coder profile list` see all profiles regardless
+   of which one is active.
+
 ## Known Pitfalls
 
+### DO NOT hardcode `~/.hermes` paths
+Use `get_hermes_home()` from `hermes_constants` for code paths. Use `display_hermes_home()`
+for user-facing print/log messages. Hardcoding `~/.hermes` breaks profiles — each profile
+has its own `HERMES_HOME` directory. This was the source of 5 bugs fixed in PR #3575.
+
 ### DO NOT use `simple_term_menu` for interactive menus
 Rendering bugs in tmux/iTerm2 — ghosting on scroll. Use `curses` (stdlib) instead. See `hermes_cli/tools_config.py` for the pattern.
 
@@ -327,17 +432,33 @@ Rendering bugs in tmux/iTerm2 — ghosting on scroll. Use `curses` (stdlib) inst
 Leaks as literal `?[K` text under `prompt_toolkit`'s `patch_stdout`. Use space-padding: `f"\r{line}{' ' * pad}"`.
 
 ### `_last_resolved_tool_names` is a process-global in `model_tools.py`
-When subagents overwrite this global, `execute_code` calls after delegation may fail with missing tool imports. Known bug.
+`_run_single_child()` in `delegate_tool.py` saves and restores this global around subagent execution. If you add new code that reads this global, be aware it may be temporarily stale during child agent runs.
+
+### DO NOT hardcode cross-tool references in schema descriptions
+Tool schema descriptions must not mention tools from other toolsets by name (e.g., `browser_navigate` saying "prefer web_search"). Those tools may be unavailable (missing API keys, disabled toolset), causing the model to hallucinate calls to non-existent tools. If a cross-reference is needed, add it dynamically in `get_tool_definitions()` in `model_tools.py` — see the `browser_navigate` / `execute_code` post-processing blocks for the pattern.
 
 ### Tests must not write to `~/.hermes/`
 The `_isolate_hermes_home` autouse fixture in `tests/conftest.py` redirects `HERMES_HOME` to a temp dir. Never hardcode `~/.hermes/` paths in tests.
 
+**Profile tests**: When testing profile features, also mock `Path.home()` so that
+`_get_profiles_root()` and `_get_default_hermes_home()` resolve within the temp dir.
+Use the pattern from `tests/hermes_cli/test_profiles.py`:
+```python
+@pytest.fixture
+def profile_env(tmp_path, monkeypatch):
+    home = tmp_path / ".hermes"
+    home.mkdir()
+    monkeypatch.setattr(Path, "home", lambda: tmp_path)
+    monkeypatch.setenv("HERMES_HOME", str(home))
+    return home
+```
+
 ---
 
 ## Testing
 
 ```bash
-source .venv/bin/activate
+source venv/bin/activate
 python -m pytest tests/ -q          # Full suite (~3000 tests, ~3 min)
 python -m pytest tests/test_model_tools.py -q   # Toolset resolution
 python -m pytest tests/test_cli_init.py -q       # CLI config loading

CONTRIBUTING.md

@@ -72,8 +72,9 @@ export VIRTUAL_ENV="$(pwd)/venv"
 
 # Install with all extras (messaging, cron, CLI menus, dev tools)
 uv pip install -e ".[all,dev]"
-uv pip install -e "./mini-swe-agent"
-uv pip install -e "./tinker-atropos"
+
+# Optional: RL training submodule
+# git submodule update --init tinker-atropos && uv pip install -e "./tinker-atropos"
 
 # Optional: browser tools
 npm install
@@ -136,7 +137,7 @@ hermes-agent/
 │   ├── auth.py                   # Provider resolution, OAuth, Nous Portal
 │   ├── models.py                 # OpenRouter model selection lists
 │   ├── banner.py                 # Welcome banner, ASCII art
-│   ├── commands.py               # Slash command definitions + autocomplete
+│   ├── commands.py               # Central slash command registry (CommandDef), autocomplete, gateway helpers
 │   ├── callbacks.py              # Interactive callbacks (clarify, sudo, approval)
 │   ├── doctor.py                 # Diagnostics
 │   ├── skills_hub.py             # Skills Hub CLI + /skills slash command
@@ -147,7 +148,7 @@ hermes-agent/
 │   ├── approval.py               # Dangerous command detection + per-session approval
 │   ├── terminal_tool.py          # Terminal orchestration (sudo, env lifecycle, backends)
 │   ├── file_operations.py        # read_file, write_file, search, patch, etc.
-│   ├── web_tools.py              # web_search, web_extract (Firecrawl + Gemini summarization)
+│   ├── web_tools.py              # web_search, web_extract (Parallel/Firecrawl + Gemini summarization)
 │   ├── vision_tools.py           # Image analysis via multimodal models
 │   ├── delegate_tool.py          # Subagent spawning and parallel task execution
 │   ├── code_execution_tool.py    # Sandboxed Python with RPC tool access

Dockerfile

@@ -0,0 +1,25 @@
+FROM debian:13.4
+
+# Install system dependencies in one layer, clear APT cache
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        build-essential nodejs npm python3 python3-pip ripgrep ffmpeg gcc python3-dev libffi-dev && \
+    rm -rf /var/lib/apt/lists/*
+
+COPY . /opt/hermes
+WORKDIR /opt/hermes
+
+# Install Python and Node dependencies in one layer, no cache
+RUN pip install --no-cache-dir -e ".[all]" --break-system-packages && \
+    npm install --prefer-offline --no-audit && \
+    npx playwright install --with-deps chromium --only-shell && \
+    cd /opt/hermes/scripts/whatsapp-bridge && \
+    npm install --prefer-offline --no-audit && \
+    npm cache clean --force
+
+WORKDIR /opt/hermes
+RUN chmod +x /opt/hermes/docker/entrypoint.sh
+
+ENV HERMES_HOME=/opt/data
+VOLUME [ "/opt/data" ]
+ENTRYPOINT [ "/opt/hermes/docker/entrypoint.sh" ]

MANIFEST.in

@@ -0,0 +1,4 @@
+graft skills
+graft optional-skills
+global-exclude __pycache__
+global-exclude *.py[cod]

README.md

@@ -2,7 +2,7 @@
   <img src="assets/banner.png" alt="Hermes Agent" width="100%">
 </p>
 
-# Hermes Agent ⚕
+# Hermes Agent ☤
 
 <p align="center">
   <a href="https://hermes-agent.nousresearch.com/docs/"><img src="https://img.shields.io/badge/Docs-hermes--agent.nousresearch.com-FFD700?style=for-the-badge" alt="Documentation"></a>
@@ -62,6 +62,24 @@ hermes doctor       # Diagnose any issues
 
 📖 **[Full documentation →](https://hermes-agent.nousresearch.com/docs/)**
 
+## CLI vs Messaging Quick Reference
+
+Hermes has two entry points: start the terminal UI with `hermes`, or run the gateway and talk to it from Telegram, Discord, Slack, WhatsApp, Signal, or Email. Once you're in a conversation, many slash commands are shared across both interfaces.
+
+| Action | CLI | Messaging platforms |
+|---------|-----|---------------------|
+| Start chatting | `hermes` | Run `hermes gateway setup` + `hermes gateway start`, then send the bot a message |
+| Start fresh conversation | `/new` or `/reset` | `/new` or `/reset` |
+| Change model | `/model [provider:model]` | `/model [provider:model]` |
+| Set a personality | `/personality [name]` | `/personality [name]` |
+| Retry or undo the last turn | `/retry`, `/undo` | `/retry`, `/undo` |
+| Compress context / check usage | `/compress`, `/usage`, `/insights [--days N]` | `/compress`, `/usage`, `/insights [days]` |
+| Browse skills | `/skills` or `/<skill-name>` | `/skills` or `/<skill-name>` |
+| Interrupt current work | `Ctrl+C` or send a new message | `/stop` or send a new message |
+| Platform-specific status | `/platforms` | `/status`, `/sethome` |
+
+For the full command lists, see the [CLI guide](https://hermes-agent.nousresearch.com/docs/user-guide/cli) and the [Messaging Gateway guide](https://hermes-agent.nousresearch.com/docs/user-guide/messaging).
+
 ---
 
 ## Documentation
@@ -126,16 +144,14 @@ Quick start for contributors:
 ```bash
 git clone https://github.com/NousResearch/hermes-agent.git
 cd hermes-agent
-git submodule update --init mini-swe-agent   # required terminal backend
 curl -LsSf https://astral.sh/uv/install.sh | sh
-uv venv .venv --python 3.11
-source .venv/bin/activate
+uv venv venv --python 3.11
+source venv/bin/activate
 uv pip install -e ".[all,dev]"
-uv pip install -e "./mini-swe-agent"
 python -m pytest tests/ -q
 ```
 
-> **RL Training (optional):** To work on the RL/Tinker-Atropos integration, also run:
+> **RL Training (optional):** To work on the RL/Tinker-Atropos integration:
 > ```bash
 > git submodule update --init tinker-atropos
 > uv pip install -e "./tinker-atropos"

RELEASE_v0.3.0.md

@@ -0,0 +1,377 @@
+# Hermes Agent v0.3.0 (v2026.3.17)
+
+**Release Date:** March 17, 2026
+
+> The streaming, plugins, and provider release — unified real-time token delivery, first-class plugin architecture, rebuilt provider system with Vercel AI Gateway, native Anthropic provider, smart approvals, live Chrome CDP browser connect, ACP IDE integration, Honcho memory, voice mode, persistent shell, and 50+ bug fixes across every platform.
+
+---
+
+## ✨ Highlights
+
+- **Unified Streaming Infrastructure** — Real-time token-by-token delivery in CLI and all gateway platforms. Responses stream as they're generated instead of arriving as a block. ([#1538](https://github.com/NousResearch/hermes-agent/pull/1538))
+
+- **First-Class Plugin Architecture** — Drop Python files into `~/.hermes/plugins/` to extend Hermes with custom tools, commands, and hooks. No forking required. ([#1544](https://github.com/NousResearch/hermes-agent/pull/1544), [#1555](https://github.com/NousResearch/hermes-agent/pull/1555))
+
+- **Native Anthropic Provider** — Direct Anthropic API calls with Claude Code credential auto-discovery, OAuth PKCE flows, and native prompt caching. No OpenRouter middleman needed. ([#1097](https://github.com/NousResearch/hermes-agent/pull/1097))
+
+- **Smart Approvals + /stop Command** — Codex-inspired approval system that learns which commands are safe and remembers your preferences. `/stop` kills the current agent run immediately. ([#1543](https://github.com/NousResearch/hermes-agent/pull/1543))
+
+- **Honcho Memory Integration** — Async memory writes, configurable recall modes, session title integration, and multi-user isolation in gateway mode. By @erosika. ([#736](https://github.com/NousResearch/hermes-agent/pull/736))
+
+- **Voice Mode** — Push-to-talk in CLI, voice notes in Telegram/Discord, Discord voice channel support, and local Whisper transcription via faster-whisper. ([#1299](https://github.com/NousResearch/hermes-agent/pull/1299), [#1185](https://github.com/NousResearch/hermes-agent/pull/1185), [#1429](https://github.com/NousResearch/hermes-agent/pull/1429))
+
+- **Concurrent Tool Execution** — Multiple independent tool calls now run in parallel via ThreadPoolExecutor, significantly reducing latency for multi-tool turns. ([#1152](https://github.com/NousResearch/hermes-agent/pull/1152))
+
+- **PII Redaction** — When `privacy.redact_pii` is enabled, personally identifiable information is automatically scrubbed before sending context to LLM providers. ([#1542](https://github.com/NousResearch/hermes-agent/pull/1542))
+
+- **`/browser connect` via CDP** — Attach browser tools to a live Chrome instance through Chrome DevTools Protocol. Debug, inspect, and interact with pages you already have open. ([#1549](https://github.com/NousResearch/hermes-agent/pull/1549))
+
+- **Vercel AI Gateway Provider** — Route Hermes through Vercel's AI Gateway for access to their model catalog and infrastructure. ([#1628](https://github.com/NousResearch/hermes-agent/pull/1628))
+
+- **Centralized Provider Router** — Rebuilt provider system with `call_llm` API, unified `/model` command, auto-detect provider on model switch, and direct endpoint overrides for auxiliary/delegation clients. ([#1003](https://github.com/NousResearch/hermes-agent/pull/1003), [#1506](https://github.com/NousResearch/hermes-agent/pull/1506), [#1375](https://github.com/NousResearch/hermes-agent/pull/1375))
+
+- **ACP Server (IDE Integration)** — VS Code, Zed, and JetBrains can now connect to Hermes as an agent backend, with full slash command support. ([#1254](https://github.com/NousResearch/hermes-agent/pull/1254), [#1532](https://github.com/NousResearch/hermes-agent/pull/1532))
+
+- **Persistent Shell Mode** — Local and SSH terminal backends can maintain shell state across tool calls — cd, env vars, and aliases persist. By @alt-glitch. ([#1067](https://github.com/NousResearch/hermes-agent/pull/1067), [#1483](https://github.com/NousResearch/hermes-agent/pull/1483))
+
+- **Agentic On-Policy Distillation (OPD)** — New RL training environment for distilling agent policies, expanding the Atropos training ecosystem. ([#1149](https://github.com/NousResearch/hermes-agent/pull/1149))
+
+---
+
+## 🏗️ Core Agent & Architecture
+
+### Provider & Model Support
+- **Centralized provider router** with `call_llm` API and unified `/model` command — switch models and providers seamlessly ([#1003](https://github.com/NousResearch/hermes-agent/pull/1003))
+- **Vercel AI Gateway** provider support ([#1628](https://github.com/NousResearch/hermes-agent/pull/1628))
+- **Auto-detect provider** when switching models via `/model` ([#1506](https://github.com/NousResearch/hermes-agent/pull/1506))
+- **Direct endpoint overrides** for auxiliary and delegation clients — point vision/subagent calls at specific endpoints ([#1375](https://github.com/NousResearch/hermes-agent/pull/1375))
+- **Native Anthropic auxiliary vision** — use Claude's native vision API instead of routing through OpenAI-compatible endpoints ([#1377](https://github.com/NousResearch/hermes-agent/pull/1377))
+- Anthropic OAuth flow improvements — auto-run `claude setup-token`, reauthentication, PKCE state persistence, identity fingerprinting ([#1132](https://github.com/NousResearch/hermes-agent/pull/1132), [#1360](https://github.com/NousResearch/hermes-agent/pull/1360), [#1396](https://github.com/NousResearch/hermes-agent/pull/1396), [#1597](https://github.com/NousResearch/hermes-agent/pull/1597))
+- Fix adaptive thinking without `budget_tokens` for Claude 4.6 models — by @ASRagab ([#1128](https://github.com/NousResearch/hermes-agent/pull/1128))
+- Fix Anthropic cache markers through adapter — by @brandtcormorant ([#1216](https://github.com/NousResearch/hermes-agent/pull/1216))
+- Retry Anthropic 429/529 errors and surface details to users — by @0xbyt4 ([#1585](https://github.com/NousResearch/hermes-agent/pull/1585))
+- Fix Anthropic adapter max_tokens, fallback crash, proxy base_url — by @0xbyt4 ([#1121](https://github.com/NousResearch/hermes-agent/pull/1121))
+- Fix DeepSeek V3 parser dropping multiple parallel tool calls — by @mr-emmett-one ([#1365](https://github.com/NousResearch/hermes-agent/pull/1365), [#1300](https://github.com/NousResearch/hermes-agent/pull/1300))
+- Accept unlisted models with warning instead of rejecting ([#1047](https://github.com/NousResearch/hermes-agent/pull/1047), [#1102](https://github.com/NousResearch/hermes-agent/pull/1102))
+- Skip reasoning params for unsupported OpenRouter models ([#1485](https://github.com/NousResearch/hermes-agent/pull/1485))
+- MiniMax Anthropic API compatibility fix ([#1623](https://github.com/NousResearch/hermes-agent/pull/1623))
+- Custom endpoint `/models` verification and `/v1` base URL suggestion ([#1480](https://github.com/NousResearch/hermes-agent/pull/1480))
+- Resolve delegation providers from `custom_providers` config ([#1328](https://github.com/NousResearch/hermes-agent/pull/1328))
+- Kimi model additions and User-Agent fix ([#1039](https://github.com/NousResearch/hermes-agent/pull/1039))
+- Strip `call_id`/`response_item_id` for Mistral compatibility ([#1058](https://github.com/NousResearch/hermes-agent/pull/1058))
+
+### Agent Loop & Conversation
+- **Anthropic Context Editing API** support ([#1147](https://github.com/NousResearch/hermes-agent/pull/1147))
+- Improved context compaction handoff summaries — compressor now preserves more actionable state ([#1273](https://github.com/NousResearch/hermes-agent/pull/1273))
+- Sync session_id after mid-run context compression ([#1160](https://github.com/NousResearch/hermes-agent/pull/1160))
+- Session hygiene threshold tuned to 50% for more proactive compression ([#1096](https://github.com/NousResearch/hermes-agent/pull/1096), [#1161](https://github.com/NousResearch/hermes-agent/pull/1161))
+- Include session ID in system prompt via `--pass-session-id` flag ([#1040](https://github.com/NousResearch/hermes-agent/pull/1040))
+- Prevent closed OpenAI client reuse across retries ([#1391](https://github.com/NousResearch/hermes-agent/pull/1391))
+- Sanitize chat payloads and provider precedence ([#1253](https://github.com/NousResearch/hermes-agent/pull/1253))
+- Handle dict tool call arguments from Codex and local backends ([#1393](https://github.com/NousResearch/hermes-agent/pull/1393), [#1440](https://github.com/NousResearch/hermes-agent/pull/1440))
+
+### Memory & Sessions
+- **Improve memory prioritization** — user preferences and corrections weighted above procedural knowledge ([#1548](https://github.com/NousResearch/hermes-agent/pull/1548))
+- Tighter memory and session recall guidance in system prompts ([#1329](https://github.com/NousResearch/hermes-agent/pull/1329))
+- Persist CLI token counts to session DB for `/insights` ([#1498](https://github.com/NousResearch/hermes-agent/pull/1498))
+- Keep Honcho recall out of the cached system prefix ([#1201](https://github.com/NousResearch/hermes-agent/pull/1201))
+- Correct `seed_ai_identity` to use `session.add_messages()` ([#1475](https://github.com/NousResearch/hermes-agent/pull/1475))
+- Isolate Honcho session routing for multi-user gateway ([#1500](https://github.com/NousResearch/hermes-agent/pull/1500))
+
+---
+
+## 📱 Messaging Platforms (Gateway)
+
+### Gateway Core
+- **System gateway service mode** — run as a system-level systemd service, not just user-level ([#1371](https://github.com/NousResearch/hermes-agent/pull/1371))
+- **Gateway install scope prompts** — choose user vs system scope during setup ([#1374](https://github.com/NousResearch/hermes-agent/pull/1374))
+- **Reasoning hot reload** — change reasoning settings without restarting the gateway ([#1275](https://github.com/NousResearch/hermes-agent/pull/1275))
+- Default group sessions to per-user isolation — no more shared state across users in group chats ([#1495](https://github.com/NousResearch/hermes-agent/pull/1495), [#1417](https://github.com/NousResearch/hermes-agent/pull/1417))
+- Harden gateway restart recovery ([#1310](https://github.com/NousResearch/hermes-agent/pull/1310))
+- Cancel active runs during shutdown ([#1427](https://github.com/NousResearch/hermes-agent/pull/1427))
+- SSL certificate auto-detection for NixOS and non-standard systems ([#1494](https://github.com/NousResearch/hermes-agent/pull/1494))
+- Auto-detect D-Bus session bus for `systemctl --user` on headless servers ([#1601](https://github.com/NousResearch/hermes-agent/pull/1601))
+- Auto-enable systemd linger during gateway install on headless servers ([#1334](https://github.com/NousResearch/hermes-agent/pull/1334))
+- Fall back to module entrypoint when `hermes` is not on PATH ([#1355](https://github.com/NousResearch/hermes-agent/pull/1355))
+- Fix dual gateways on macOS launchd after `hermes update` ([#1567](https://github.com/NousResearch/hermes-agent/pull/1567))
+- Remove recursive ExecStop from systemd units ([#1530](https://github.com/NousResearch/hermes-agent/pull/1530))
+- Prevent logging handler accumulation in gateway mode ([#1251](https://github.com/NousResearch/hermes-agent/pull/1251))
+- Restart on retryable startup failures — by @jplew ([#1517](https://github.com/NousResearch/hermes-agent/pull/1517))
+- Backfill model on gateway sessions after agent runs ([#1306](https://github.com/NousResearch/hermes-agent/pull/1306))
+- PID-based gateway kill and deferred config write ([#1499](https://github.com/NousResearch/hermes-agent/pull/1499))
+
+### Telegram
+- Buffer media groups to prevent self-interruption from photo bursts ([#1341](https://github.com/NousResearch/hermes-agent/pull/1341), [#1422](https://github.com/NousResearch/hermes-agent/pull/1422))
+- Retry on transient TLS failures during connect and send ([#1535](https://github.com/NousResearch/hermes-agent/pull/1535))
+- Harden polling conflict handling ([#1339](https://github.com/NousResearch/hermes-agent/pull/1339))
+- Escape chunk indicators and inline code in MarkdownV2 ([#1478](https://github.com/NousResearch/hermes-agent/pull/1478), [#1626](https://github.com/NousResearch/hermes-agent/pull/1626))
+- Check updater/app state before disconnect ([#1389](https://github.com/NousResearch/hermes-agent/pull/1389))
+
+### Discord
+- `/thread` command with `auto_thread` config and media metadata fixes ([#1178](https://github.com/NousResearch/hermes-agent/pull/1178))
+- Auto-thread on @mention, skip mention text in bot threads ([#1438](https://github.com/NousResearch/hermes-agent/pull/1438))
+- Retry without reply reference for system messages ([#1385](https://github.com/NousResearch/hermes-agent/pull/1385))
+- Preserve native document and video attachment support ([#1392](https://github.com/NousResearch/hermes-agent/pull/1392))
+- Defer discord adapter annotations to avoid optional import crashes ([#1314](https://github.com/NousResearch/hermes-agent/pull/1314))
+
+### Slack
+- Thread handling overhaul — progress messages, responses, and session isolation all respect threads ([#1103](https://github.com/NousResearch/hermes-agent/pull/1103))
+- Formatting, reactions, user resolution, and command improvements ([#1106](https://github.com/NousResearch/hermes-agent/pull/1106))
+- Fix MAX_MESSAGE_LENGTH 3900 → 39000 ([#1117](https://github.com/NousResearch/hermes-agent/pull/1117))
+- File upload fallback preserves thread context — by @0xbyt4 ([#1122](https://github.com/NousResearch/hermes-agent/pull/1122))
+- Improve setup guidance ([#1387](https://github.com/NousResearch/hermes-agent/pull/1387))
+
+### Email
+- Fix IMAP UID tracking and SMTP TLS verification ([#1305](https://github.com/NousResearch/hermes-agent/pull/1305))
+- Add `skip_attachments` option via config.yaml ([#1536](https://github.com/NousResearch/hermes-agent/pull/1536))
+
+### Home Assistant
+- Event filtering closed by default ([#1169](https://github.com/NousResearch/hermes-agent/pull/1169))
+
+---
+
+## 🖥️ CLI & User Experience
+
+### Interactive CLI
+- **Persistent CLI status bar** — always-visible model, provider, and token counts ([#1522](https://github.com/NousResearch/hermes-agent/pull/1522))
+- **File path autocomplete** in the input prompt ([#1545](https://github.com/NousResearch/hermes-agent/pull/1545))
+- **`/plan` command** — generate implementation plans from specs ([#1372](https://github.com/NousResearch/hermes-agent/pull/1372), [#1381](https://github.com/NousResearch/hermes-agent/pull/1381))
+- **Major `/rollback` improvements** — richer checkpoint history, clearer UX ([#1505](https://github.com/NousResearch/hermes-agent/pull/1505))
+- **Preload CLI skills on launch** — skills are ready before the first prompt ([#1359](https://github.com/NousResearch/hermes-agent/pull/1359))
+- **Centralized slash command registry** — all commands defined once, consumed everywhere ([#1603](https://github.com/NousResearch/hermes-agent/pull/1603))
+- `/bg` alias for `/background` ([#1590](https://github.com/NousResearch/hermes-agent/pull/1590))
+- Prefix matching for slash commands — `/mod` resolves to `/model` ([#1320](https://github.com/NousResearch/hermes-agent/pull/1320))
+- `/new`, `/reset`, `/clear` now start genuinely fresh sessions ([#1237](https://github.com/NousResearch/hermes-agent/pull/1237))
+- Accept session ID prefixes for session actions ([#1425](https://github.com/NousResearch/hermes-agent/pull/1425))
+- TUI prompt and accent output now respect active skin ([#1282](https://github.com/NousResearch/hermes-agent/pull/1282))
+- Centralize tool emoji metadata in registry + skin integration ([#1484](https://github.com/NousResearch/hermes-agent/pull/1484))
+- "View full command" option added to dangerous command approval — by @teknium1 based on design by community ([#887](https://github.com/NousResearch/hermes-agent/pull/887))
+- Non-blocking startup update check and banner deduplication ([#1386](https://github.com/NousResearch/hermes-agent/pull/1386))
+- `/reasoning` command output ordering and inline think extraction fixes ([#1031](https://github.com/NousResearch/hermes-agent/pull/1031))
+- Verbose mode shows full untruncated output ([#1472](https://github.com/NousResearch/hermes-agent/pull/1472))
+- Fix `/status` to report live state and tokens ([#1476](https://github.com/NousResearch/hermes-agent/pull/1476))
+- Seed a default global SOUL.md ([#1311](https://github.com/NousResearch/hermes-agent/pull/1311))
+
+### Setup & Configuration
+- **OpenClaw migration** during first-time setup — by @kshitijk4poor ([#981](https://github.com/NousResearch/hermes-agent/pull/981))
+- `hermes claw migrate` command + migration docs ([#1059](https://github.com/NousResearch/hermes-agent/pull/1059))
+- Smart vision setup that respects the user's chosen provider ([#1323](https://github.com/NousResearch/hermes-agent/pull/1323))
+- Handle headless setup flows end-to-end ([#1274](https://github.com/NousResearch/hermes-agent/pull/1274))
+- Prefer curses over `simple_term_menu` in setup.py ([#1487](https://github.com/NousResearch/hermes-agent/pull/1487))
+- Show effective model and provider in `/status` ([#1284](https://github.com/NousResearch/hermes-agent/pull/1284))
+- Config set examples use placeholder syntax ([#1322](https://github.com/NousResearch/hermes-agent/pull/1322))
+- Reload .env over stale shell overrides ([#1434](https://github.com/NousResearch/hermes-agent/pull/1434))
+- Fix is_coding_plan NameError crash — by @0xbyt4 ([#1123](https://github.com/NousResearch/hermes-agent/pull/1123))
+- Add missing packages to setuptools config — by @alt-glitch ([#912](https://github.com/NousResearch/hermes-agent/pull/912))
+- Installer: clarify why sudo is needed at every prompt ([#1602](https://github.com/NousResearch/hermes-agent/pull/1602))
+
+---
+
+## 🔧 Tool System
+
+### Terminal & Execution
+- **Persistent shell mode** for local and SSH backends — maintain shell state across tool calls — by @alt-glitch ([#1067](https://github.com/NousResearch/hermes-agent/pull/1067), [#1483](https://github.com/NousResearch/hermes-agent/pull/1483))
+- **Tirith pre-exec command scanning** — security layer that analyzes commands before execution ([#1256](https://github.com/NousResearch/hermes-agent/pull/1256))
+- Strip Hermes provider env vars from all subprocess environments ([#1157](https://github.com/NousResearch/hermes-agent/pull/1157), [#1172](https://github.com/NousResearch/hermes-agent/pull/1172), [#1399](https://github.com/NousResearch/hermes-agent/pull/1399), [#1419](https://github.com/NousResearch/hermes-agent/pull/1419)) — initial fix by @eren-karakus0
+- SSH preflight check ([#1486](https://github.com/NousResearch/hermes-agent/pull/1486))
+- Docker backend: make cwd workspace mount explicit opt-in ([#1534](https://github.com/NousResearch/hermes-agent/pull/1534))
+- Add project root to PYTHONPATH in execute_code sandbox ([#1383](https://github.com/NousResearch/hermes-agent/pull/1383))
+- Eliminate execute_code progress spam on gateway platforms ([#1098](https://github.com/NousResearch/hermes-agent/pull/1098))
+- Clearer docker backend preflight errors ([#1276](https://github.com/NousResearch/hermes-agent/pull/1276))
+
+### Browser
+- **`/browser connect`** — attach browser tools to a live Chrome instance via CDP ([#1549](https://github.com/NousResearch/hermes-agent/pull/1549))
+- Improve browser cleanup, local browser PATH setup, and screenshot recovery ([#1333](https://github.com/NousResearch/hermes-agent/pull/1333))
+
+### MCP
+- **Selective tool loading** with utility policies — filter which MCP tools are available ([#1302](https://github.com/NousResearch/hermes-agent/pull/1302))
+- Auto-reload MCP tools when `mcp_servers` config changes without restart ([#1474](https://github.com/NousResearch/hermes-agent/pull/1474))
+- Resolve npx stdio connection failures ([#1291](https://github.com/NousResearch/hermes-agent/pull/1291))
+- Preserve MCP toolsets when saving platform tool config ([#1421](https://github.com/NousResearch/hermes-agent/pull/1421))
+
+### Vision
+- Unify vision backend gating ([#1367](https://github.com/NousResearch/hermes-agent/pull/1367))
+- Surface actual error reason instead of generic message ([#1338](https://github.com/NousResearch/hermes-agent/pull/1338))
+- Make Claude image handling work end-to-end ([#1408](https://github.com/NousResearch/hermes-agent/pull/1408))
+
+### Cron
+- **Compress cron management into one tool** — single `cronjob` tool replaces multiple commands ([#1343](https://github.com/NousResearch/hermes-agent/pull/1343))
+- Suppress duplicate cron sends to auto-delivery targets ([#1357](https://github.com/NousResearch/hermes-agent/pull/1357))
+- Persist cron sessions to SQLite ([#1255](https://github.com/NousResearch/hermes-agent/pull/1255))
+- Per-job runtime overrides (provider, model, base_url) ([#1398](https://github.com/NousResearch/hermes-agent/pull/1398))
+- Atomic write in `save_job_output` to prevent data loss on crash ([#1173](https://github.com/NousResearch/hermes-agent/pull/1173))
+- Preserve thread context for `deliver=origin` ([#1437](https://github.com/NousResearch/hermes-agent/pull/1437))
+
+### Patch Tool
+- Avoid corrupting pipe chars in V4A patch apply ([#1286](https://github.com/NousResearch/hermes-agent/pull/1286))
+- Permissive `block_anchor` thresholds and unicode normalization ([#1539](https://github.com/NousResearch/hermes-agent/pull/1539))
+
+### Delegation
+- Add observability metadata to subagent results (model, tokens, duration, tool trace) ([#1175](https://github.com/NousResearch/hermes-agent/pull/1175))
+
+---
+
+## 🧩 Skills Ecosystem
+
+### Skills System
+- **Integrate skills.sh** as a hub source alongside ClawHub ([#1303](https://github.com/NousResearch/hermes-agent/pull/1303))
+- Secure skill env setup on load ([#1153](https://github.com/NousResearch/hermes-agent/pull/1153))
+- Honor policy table for dangerous verdicts ([#1330](https://github.com/NousResearch/hermes-agent/pull/1330))
+- Harden ClawHub skill search exact matches ([#1400](https://github.com/NousResearch/hermes-agent/pull/1400))
+- Fix ClawHub skill install — use `/download` ZIP endpoint ([#1060](https://github.com/NousResearch/hermes-agent/pull/1060))
+- Avoid mislabeling local skills as builtin — by @arceus77-7 ([#862](https://github.com/NousResearch/hermes-agent/pull/862))
+
+### New Skills
+- **Linear** project management ([#1230](https://github.com/NousResearch/hermes-agent/pull/1230))
+- **X/Twitter** via x-cli ([#1285](https://github.com/NousResearch/hermes-agent/pull/1285))
+- **Telephony** — Twilio, SMS, and AI calls ([#1289](https://github.com/NousResearch/hermes-agent/pull/1289))
+- **1Password** — by @arceus77-7 ([#883](https://github.com/NousResearch/hermes-agent/pull/883), [#1179](https://github.com/NousResearch/hermes-agent/pull/1179))
+- **NeuroSkill BCI** integration ([#1135](https://github.com/NousResearch/hermes-agent/pull/1135))
+- **Blender MCP** for 3D modeling ([#1531](https://github.com/NousResearch/hermes-agent/pull/1531))
+- **OSS Security Forensics** ([#1482](https://github.com/NousResearch/hermes-agent/pull/1482))
+- **Parallel CLI** research skill ([#1301](https://github.com/NousResearch/hermes-agent/pull/1301))
+- **OpenCode** CLI skill ([#1174](https://github.com/NousResearch/hermes-agent/pull/1174))
+- **ASCII Video** skill refactored — by @SHL0MS ([#1213](https://github.com/NousResearch/hermes-agent/pull/1213), [#1598](https://github.com/NousResearch/hermes-agent/pull/1598))
+
+---
+
+## 🎙️ Voice Mode
+
+- Voice mode foundation — push-to-talk CLI, Telegram/Discord voice notes ([#1299](https://github.com/NousResearch/hermes-agent/pull/1299))
+- Free local Whisper transcription via faster-whisper ([#1185](https://github.com/NousResearch/hermes-agent/pull/1185))
+- Discord voice channel reliability fixes ([#1429](https://github.com/NousResearch/hermes-agent/pull/1429))
+- Restore local STT fallback for gateway voice notes ([#1490](https://github.com/NousResearch/hermes-agent/pull/1490))
+- Honor `stt.enabled: false` across gateway transcription ([#1394](https://github.com/NousResearch/hermes-agent/pull/1394))
+- Fix bogus incapability message on Telegram voice notes (Issue [#1033](https://github.com/NousResearch/hermes-agent/issues/1033))
+
+---
+
+## 🔌 ACP (IDE Integration)
+
+- Restore ACP server implementation ([#1254](https://github.com/NousResearch/hermes-agent/pull/1254))
+- Support slash commands in ACP adapter ([#1532](https://github.com/NousResearch/hermes-agent/pull/1532))
+
+---
+
+## 🧪 RL Training
+
+- **Agentic On-Policy Distillation (OPD)** environment — new RL training environment for agent policy distillation ([#1149](https://github.com/NousResearch/hermes-agent/pull/1149))
+- Make tinker-atropos RL training fully optional ([#1062](https://github.com/NousResearch/hermes-agent/pull/1062))
+
+---
+
+## 🔒 Security & Reliability
+
+### Security Hardening
+- **Tirith pre-exec command scanning** — static analysis of terminal commands before execution ([#1256](https://github.com/NousResearch/hermes-agent/pull/1256))
+- **PII redaction** when `privacy.redact_pii` is enabled ([#1542](https://github.com/NousResearch/hermes-agent/pull/1542))
+- Strip Hermes provider/gateway/tool env vars from all subprocess environments ([#1157](https://github.com/NousResearch/hermes-agent/pull/1157), [#1172](https://github.com/NousResearch/hermes-agent/pull/1172), [#1399](https://github.com/NousResearch/hermes-agent/pull/1399), [#1419](https://github.com/NousResearch/hermes-agent/pull/1419))
+- Docker cwd workspace mount now explicit opt-in — never auto-mount host directories ([#1534](https://github.com/NousResearch/hermes-agent/pull/1534))
+- Escape parens and braces in fork bomb regex pattern ([#1397](https://github.com/NousResearch/hermes-agent/pull/1397))
+- Harden `.worktreeinclude` path containment ([#1388](https://github.com/NousResearch/hermes-agent/pull/1388))
+- Use description as `pattern_key` to prevent approval collisions ([#1395](https://github.com/NousResearch/hermes-agent/pull/1395))
+
+### Reliability
+- Guard init-time stdio writes ([#1271](https://github.com/NousResearch/hermes-agent/pull/1271))
+- Session log writes reuse shared atomic JSON helper ([#1280](https://github.com/NousResearch/hermes-agent/pull/1280))
+- Atomic temp cleanup protected on interrupts ([#1401](https://github.com/NousResearch/hermes-agent/pull/1401))
+
+---
+
+## 🐛 Notable Bug Fixes
+
+- **`/status` always showing 0 tokens** — now reports live state (Issue [#1465](https://github.com/NousResearch/hermes-agent/issues/1465), [#1476](https://github.com/NousResearch/hermes-agent/pull/1476))
+- **Custom model endpoints not working** — restored config-saved endpoint resolution (Issue [#1460](https://github.com/NousResearch/hermes-agent/issues/1460), [#1373](https://github.com/NousResearch/hermes-agent/pull/1373))
+- **MCP tools not visible until restart** — auto-reload on config change (Issue [#1036](https://github.com/NousResearch/hermes-agent/issues/1036), [#1474](https://github.com/NousResearch/hermes-agent/pull/1474))
+- **`hermes tools` removing MCP tools** — preserve MCP toolsets when saving (Issue [#1247](https://github.com/NousResearch/hermes-agent/issues/1247), [#1421](https://github.com/NousResearch/hermes-agent/pull/1421))
+- **Terminal subprocesses inheriting `OPENAI_BASE_URL`** breaking external tools (Issue [#1002](https://github.com/NousResearch/hermes-agent/issues/1002), [#1399](https://github.com/NousResearch/hermes-agent/pull/1399))
+- **Background process lost on gateway restart** — improved recovery (Issue [#1144](https://github.com/NousResearch/hermes-agent/issues/1144))
+- **Cron jobs not persisting state** — now stored in SQLite (Issue [#1416](https://github.com/NousResearch/hermes-agent/issues/1416), [#1255](https://github.com/NousResearch/hermes-agent/pull/1255))
+- **Cronjob `deliver: origin` not preserving thread context** (Issue [#1219](https://github.com/NousResearch/hermes-agent/issues/1219), [#1437](https://github.com/NousResearch/hermes-agent/pull/1437))
+- **Gateway systemd service failing to auto-restart** when browser processes orphaned (Issue [#1617](https://github.com/NousResearch/hermes-agent/issues/1617))
+- **`/background` completion report cut off in Telegram** (Issue [#1443](https://github.com/NousResearch/hermes-agent/issues/1443))
+- **Model switching not taking effect** (Issue [#1244](https://github.com/NousResearch/hermes-agent/issues/1244), [#1183](https://github.com/NousResearch/hermes-agent/pull/1183))
+- **`hermes doctor` reporting cronjob as unavailable** (Issue [#878](https://github.com/NousResearch/hermes-agent/issues/878), [#1180](https://github.com/NousResearch/hermes-agent/pull/1180))
+- **WhatsApp bridge messages not received** from mobile (Issue [#1142](https://github.com/NousResearch/hermes-agent/issues/1142))
+- **Setup wizard hanging on headless SSH** (Issue [#905](https://github.com/NousResearch/hermes-agent/issues/905), [#1274](https://github.com/NousResearch/hermes-agent/pull/1274))
+- **Log handler accumulation** degrading gateway performance (Issue [#990](https://github.com/NousResearch/hermes-agent/issues/990), [#1251](https://github.com/NousResearch/hermes-agent/pull/1251))
+- **Gateway NULL model in DB** (Issue [#987](https://github.com/NousResearch/hermes-agent/issues/987), [#1306](https://github.com/NousResearch/hermes-agent/pull/1306))
+- **Strict endpoints rejecting replayed tool_calls** (Issue [#893](https://github.com/NousResearch/hermes-agent/issues/893))
+- **Remaining hardcoded `~/.hermes` paths** — all now respect `HERMES_HOME` (Issue [#892](https://github.com/NousResearch/hermes-agent/issues/892), [#1233](https://github.com/NousResearch/hermes-agent/pull/1233))
+- **Delegate tool not working with custom inference providers** (Issue [#1011](https://github.com/NousResearch/hermes-agent/issues/1011), [#1328](https://github.com/NousResearch/hermes-agent/pull/1328))
+- **Skills Guard blocking official skills** (Issue [#1006](https://github.com/NousResearch/hermes-agent/issues/1006), [#1330](https://github.com/NousResearch/hermes-agent/pull/1330))
+- **Setup writing provider before model selection** (Issue [#1182](https://github.com/NousResearch/hermes-agent/issues/1182))
+- **`GatewayConfig.get()` AttributeError** crashing all message handling (Issue [#1158](https://github.com/NousResearch/hermes-agent/issues/1158), [#1287](https://github.com/NousResearch/hermes-agent/pull/1287))
+- **`/update` hard-failing with "command not found"** (Issue [#1049](https://github.com/NousResearch/hermes-agent/issues/1049))
+- **Image analysis failing silently** (Issue [#1034](https://github.com/NousResearch/hermes-agent/issues/1034), [#1338](https://github.com/NousResearch/hermes-agent/pull/1338))
+- **API `BadRequestError` from `'dict'` object has no attribute `'strip'`** (Issue [#1071](https://github.com/NousResearch/hermes-agent/issues/1071))
+- **Slash commands requiring exact full name** — now uses prefix matching (Issue [#928](https://github.com/NousResearch/hermes-agent/issues/928), [#1320](https://github.com/NousResearch/hermes-agent/pull/1320))
+- **Gateway stops responding when terminal is closed on headless** (Issue [#1005](https://github.com/NousResearch/hermes-agent/issues/1005))
+
+---
+
+## 🧪 Testing
+
+- Cover empty cached Anthropic tool-call turns ([#1222](https://github.com/NousResearch/hermes-agent/pull/1222))
+- Fix stale CI assumptions in parser and quick-command coverage ([#1236](https://github.com/NousResearch/hermes-agent/pull/1236))
+- Fix gateway async tests without implicit event loop ([#1278](https://github.com/NousResearch/hermes-agent/pull/1278))
+- Make gateway async tests xdist-safe ([#1281](https://github.com/NousResearch/hermes-agent/pull/1281))
+- Cross-timezone naive timestamp regression for cron ([#1319](https://github.com/NousResearch/hermes-agent/pull/1319))
+- Isolate codex provider tests from local env ([#1335](https://github.com/NousResearch/hermes-agent/pull/1335))
+- Lock retry replacement semantics ([#1379](https://github.com/NousResearch/hermes-agent/pull/1379))
+- Improve error logging in session search tool — by @aydnOktay ([#1533](https://github.com/NousResearch/hermes-agent/pull/1533))
+
+---
+
+## 📚 Documentation
+
+- Comprehensive SOUL.md guide ([#1315](https://github.com/NousResearch/hermes-agent/pull/1315))
+- Voice mode documentation ([#1316](https://github.com/NousResearch/hermes-agent/pull/1316), [#1362](https://github.com/NousResearch/hermes-agent/pull/1362))
+- Provider contribution guide ([#1361](https://github.com/NousResearch/hermes-agent/pull/1361))
+- ACP and internal systems implementation guides ([#1259](https://github.com/NousResearch/hermes-agent/pull/1259))
+- Expand Docusaurus coverage across CLI, tools, skills, and skins ([#1232](https://github.com/NousResearch/hermes-agent/pull/1232))
+- Terminal backend and Windows troubleshooting ([#1297](https://github.com/NousResearch/hermes-agent/pull/1297))
+- Skills hub reference section ([#1317](https://github.com/NousResearch/hermes-agent/pull/1317))
+- Checkpoint, /rollback, and git worktrees guide ([#1493](https://github.com/NousResearch/hermes-agent/pull/1493), [#1524](https://github.com/NousResearch/hermes-agent/pull/1524))
+- CLI status bar and /usage reference ([#1523](https://github.com/NousResearch/hermes-agent/pull/1523))
+- Fallback providers + /background command docs ([#1430](https://github.com/NousResearch/hermes-agent/pull/1430))
+- Gateway service scopes docs ([#1378](https://github.com/NousResearch/hermes-agent/pull/1378))
+- Slack thread reply behavior docs ([#1407](https://github.com/NousResearch/hermes-agent/pull/1407))
+- Redesigned landing page with Nous blue palette — by @austinpickett ([#974](https://github.com/NousResearch/hermes-agent/pull/974))
+- Fix several documentation typos — by @JackTheGit ([#953](https://github.com/NousResearch/hermes-agent/pull/953))
+- Stabilize website diagrams ([#1405](https://github.com/NousResearch/hermes-agent/pull/1405))
+- CLI vs messaging quick reference in README ([#1491](https://github.com/NousResearch/hermes-agent/pull/1491))
+- Add search to Docusaurus ([#1053](https://github.com/NousResearch/hermes-agent/pull/1053))
+- Home Assistant integration docs ([#1170](https://github.com/NousResearch/hermes-agent/pull/1170))
+
+---
+
+## 👥 Contributors
+
+### Core
+- **@teknium1** — 220+ PRs spanning every area of the codebase
+
+### Top Community Contributors
+
+- **@0xbyt4** (4 PRs) — Anthropic adapter fixes (max_tokens, fallback crash, 429/529 retry), Slack file upload thread context, setup NameError fix
+- **@erosika** (1 PR) — Honcho memory integration: async writes, memory modes, session title integration
+- **@SHL0MS** (2 PRs) — ASCII video skill design patterns and refactoring
+- **@alt-glitch** (2 PRs) — Persistent shell mode for local/SSH backends, setuptools packaging fix
+- **@arceus77-7** (2 PRs) — 1Password skill, fix skills list mislabeling
+- **@kshitijk4poor** (1 PR) — OpenClaw migration during setup wizard
+- **@ASRagab** (1 PR) — Fix adaptive thinking for Claude 4.6 models
+- **@eren-karakus0** (1 PR) — Strip Hermes provider env vars from subprocess environment
+- **@mr-emmett-one** (1 PR) — Fix DeepSeek V3 parser multi-tool call support
+- **@jplew** (1 PR) — Gateway restart on retryable startup failures
+- **@brandtcormorant** (1 PR) — Fix Anthropic cache control for empty text blocks
+- **@aydnOktay** (1 PR) — Improve error logging in session search tool
+- **@austinpickett** (1 PR) — Landing page redesign with Nous blue palette
+- **@JackTheGit** (1 PR) — Documentation typo fixes
+
+### All Contributors
+
+@0xbyt4, @alt-glitch, @arceus77-7, @ASRagab, @austinpickett, @aydnOktay, @brandtcormorant, @eren-karakus0, @erosika, @JackTheGit, @jplew, @kshitijk4poor, @mr-emmett-one, @SHL0MS, @teknium1
+
+---
+
+**Full Changelog**: [v2026.3.12...v2026.3.17](https://github.com/NousResearch/hermes-agent/compare/v2026.3.12...v2026.3.17)

RELEASE_v0.4.0.md

@@ -0,0 +1,400 @@
+# Hermes Agent v0.4.0 (v2026.3.23)
+
+**Release Date:** March 23, 2026
+
+> The platform expansion release — OpenAI-compatible API server, 6 new messaging adapters, 4 new inference providers, MCP server management with OAuth 2.1, @ context references, gateway prompt caching, streaming enabled by default, and a sweeping reliability pass with 200+ bug fixes.
+
+---
+
+## ✨ Highlights
+
+- **OpenAI-compatible API server** — Expose Hermes as an `/v1/chat/completions` endpoint with a new `/api/jobs` REST API for cron job management, hardened with input limits, field whitelists, SQLite-backed response persistence, and CORS origin protection ([#1756](https://github.com/NousResearch/hermes-agent/pull/1756), [#2450](https://github.com/NousResearch/hermes-agent/pull/2450), [#2456](https://github.com/NousResearch/hermes-agent/pull/2456), [#2451](https://github.com/NousResearch/hermes-agent/pull/2451), [#2472](https://github.com/NousResearch/hermes-agent/pull/2472))
+
+- **6 new messaging platform adapters** — Signal, DingTalk, SMS (Twilio), Mattermost, Matrix, and Webhook adapters join Telegram, Discord, and WhatsApp. Gateway auto-reconnects failed platforms with exponential backoff ([#2206](https://github.com/NousResearch/hermes-agent/pull/2206), [#1685](https://github.com/NousResearch/hermes-agent/pull/1685), [#1688](https://github.com/NousResearch/hermes-agent/pull/1688), [#1683](https://github.com/NousResearch/hermes-agent/pull/1683), [#2166](https://github.com/NousResearch/hermes-agent/pull/2166), [#2584](https://github.com/NousResearch/hermes-agent/pull/2584))
+
+- **@ context references** — Claude Code-style `@file` and `@url` context injection with tab completions in the CLI ([#2343](https://github.com/NousResearch/hermes-agent/pull/2343), [#2482](https://github.com/NousResearch/hermes-agent/pull/2482))
+
+- **4 new inference providers** — GitHub Copilot (OAuth + token validation), Alibaba Cloud / DashScope, Kilo Code, and OpenCode Zen/Go ([#1924](https://github.com/NousResearch/hermes-agent/pull/1924), [#1879](https://github.com/NousResearch/hermes-agent/pull/1879) by @mchzimm, [#1673](https://github.com/NousResearch/hermes-agent/pull/1673), [#1666](https://github.com/NousResearch/hermes-agent/pull/1666), [#1650](https://github.com/NousResearch/hermes-agent/pull/1650))
+
+- **MCP server management CLI** — `hermes mcp` commands for installing, configuring, and authenticating MCP servers with full OAuth 2.1 PKCE flow ([#2465](https://github.com/NousResearch/hermes-agent/pull/2465))
+
+- **Gateway prompt caching** — Cache AIAgent instances per session, preserving Anthropic prompt cache across turns for dramatic cost reduction on long conversations ([#2282](https://github.com/NousResearch/hermes-agent/pull/2282), [#2284](https://github.com/NousResearch/hermes-agent/pull/2284), [#2361](https://github.com/NousResearch/hermes-agent/pull/2361))
+
+- **Context compression overhaul** — Structured summaries with iterative updates, token-budget tail protection, configurable summary endpoint, and fallback model support ([#2323](https://github.com/NousResearch/hermes-agent/pull/2323), [#1727](https://github.com/NousResearch/hermes-agent/pull/1727), [#2224](https://github.com/NousResearch/hermes-agent/pull/2224))
+
+- **Streaming enabled by default** — CLI streaming on by default with proper spinner/tool progress display during streaming mode, plus extensive linebreak and concatenation fixes ([#2340](https://github.com/NousResearch/hermes-agent/pull/2340), [#2161](https://github.com/NousResearch/hermes-agent/pull/2161), [#2258](https://github.com/NousResearch/hermes-agent/pull/2258))
+
+---
+
+## 🖥️ CLI & User Experience
+
+### New Commands & Interactions
+- **@ context completions** — Tab-completable `@file`/`@url` references that inject file content or web pages into the conversation ([#2482](https://github.com/NousResearch/hermes-agent/pull/2482), [#2343](https://github.com/NousResearch/hermes-agent/pull/2343))
+- **`/statusbar`** — Toggle a persistent config bar showing model + provider info in the prompt ([#2240](https://github.com/NousResearch/hermes-agent/pull/2240), [#1917](https://github.com/NousResearch/hermes-agent/pull/1917))
+- **`/queue`** — Queue prompts for the agent without interrupting the current run ([#2191](https://github.com/NousResearch/hermes-agent/pull/2191), [#2469](https://github.com/NousResearch/hermes-agent/pull/2469))
+- **`/permission`** — Switch approval mode dynamically during a session ([#2207](https://github.com/NousResearch/hermes-agent/pull/2207))
+- **`/browser`** — Interactive browser sessions from the CLI ([#2273](https://github.com/NousResearch/hermes-agent/pull/2273), [#1814](https://github.com/NousResearch/hermes-agent/pull/1814))
+- **`/cost`** — Live pricing and usage tracking in gateway mode ([#2180](https://github.com/NousResearch/hermes-agent/pull/2180))
+- **`/approve` and `/deny`** — Replaced bare text approval in gateway with explicit commands ([#2002](https://github.com/NousResearch/hermes-agent/pull/2002))
+
+### Streaming & Display
+- Streaming enabled by default in CLI ([#2340](https://github.com/NousResearch/hermes-agent/pull/2340))
+- Show spinners and tool progress during streaming mode ([#2161](https://github.com/NousResearch/hermes-agent/pull/2161))
+- Show reasoning/thinking blocks when `show_reasoning` enabled ([#2118](https://github.com/NousResearch/hermes-agent/pull/2118))
+- Context pressure warnings for CLI and gateway ([#2159](https://github.com/NousResearch/hermes-agent/pull/2159))
+- Fix: streaming chunks concatenated without whitespace ([#2258](https://github.com/NousResearch/hermes-agent/pull/2258))
+- Fix: iteration boundary linebreak prevents stream concatenation ([#2413](https://github.com/NousResearch/hermes-agent/pull/2413))
+- Fix: defer streaming linebreak to prevent blank line stacking ([#2473](https://github.com/NousResearch/hermes-agent/pull/2473))
+- Fix: suppress spinner animation in non-TTY environments ([#2216](https://github.com/NousResearch/hermes-agent/pull/2216))
+- Fix: display provider and endpoint in API error messages ([#2266](https://github.com/NousResearch/hermes-agent/pull/2266))
+- Fix: resolve garbled ANSI escape codes in status printouts ([#2448](https://github.com/NousResearch/hermes-agent/pull/2448))
+- Fix: update gold ANSI color to true-color format ([#2246](https://github.com/NousResearch/hermes-agent/pull/2246))
+- Fix: normalize toolset labels and use skin colors in banner ([#1912](https://github.com/NousResearch/hermes-agent/pull/1912))
+
+### CLI Polish
+- Fix: prevent 'Press ENTER to continue...' on exit ([#2555](https://github.com/NousResearch/hermes-agent/pull/2555))
+- Fix: flush stdout during agent loop to prevent macOS display freeze ([#1654](https://github.com/NousResearch/hermes-agent/pull/1654))
+- Fix: show human-readable error when `hermes setup` hits permissions error ([#2196](https://github.com/NousResearch/hermes-agent/pull/2196))
+- Fix: `/stop` command crash + UnboundLocalError in streaming media delivery ([#2463](https://github.com/NousResearch/hermes-agent/pull/2463))
+- Fix: allow custom/local endpoints without API key ([#2556](https://github.com/NousResearch/hermes-agent/pull/2556))
+- Fix: Kitty keyboard protocol Shift+Enter for Ghostty/WezTerm (attempted + reverted due to prompt_toolkit crash) ([#2345](https://github.com/NousResearch/hermes-agent/pull/2345), [#2349](https://github.com/NousResearch/hermes-agent/pull/2349))
+
+### Configuration
+- **`${ENV_VAR}` substitution** in config.yaml ([#2684](https://github.com/NousResearch/hermes-agent/pull/2684))
+- **Real-time config reload** — config.yaml changes apply without restart ([#2210](https://github.com/NousResearch/hermes-agent/pull/2210))
+- **`custom_models.yaml`** for user-managed model additions ([#2214](https://github.com/NousResearch/hermes-agent/pull/2214))
+- **Priority-based context file selection** + CLAUDE.md support ([#2301](https://github.com/NousResearch/hermes-agent/pull/2301))
+- **Merge nested YAML sections** instead of replacing on config update ([#2213](https://github.com/NousResearch/hermes-agent/pull/2213))
+- Fix: config.yaml provider key overrides env var silently ([#2272](https://github.com/NousResearch/hermes-agent/pull/2272))
+- Fix: log warning instead of silently swallowing config.yaml errors ([#2683](https://github.com/NousResearch/hermes-agent/pull/2683))
+- Fix: disabled toolsets re-enable themselves after `hermes tools` ([#2268](https://github.com/NousResearch/hermes-agent/pull/2268))
+- Fix: platform default toolsets silently override tool deselection ([#2624](https://github.com/NousResearch/hermes-agent/pull/2624))
+- Fix: honor bare YAML `approvals.mode: off` ([#2620](https://github.com/NousResearch/hermes-agent/pull/2620))
+- Fix: `hermes update` use `.[all]` extras with fallback ([#1728](https://github.com/NousResearch/hermes-agent/pull/1728))
+- Fix: `hermes update` prompt before resetting working tree on stash conflicts ([#2390](https://github.com/NousResearch/hermes-agent/pull/2390))
+- Fix: use git pull --rebase in update/install to avoid divergent branch error ([#2274](https://github.com/NousResearch/hermes-agent/pull/2274))
+- Fix: add zprofile fallback and create zshrc on fresh macOS installs ([#2320](https://github.com/NousResearch/hermes-agent/pull/2320))
+- Fix: remove `ANTHROPIC_BASE_URL` env var to avoid collisions ([#1675](https://github.com/NousResearch/hermes-agent/pull/1675))
+- Fix: don't ask IMAP password if already in keyring or env ([#2212](https://github.com/NousResearch/hermes-agent/pull/2212))
+- Fix: OpenCode Zen/Go show OpenRouter models instead of their own ([#2277](https://github.com/NousResearch/hermes-agent/pull/2277))
+
+---
+
+## 🏗️ Core Agent & Architecture
+
+### New Providers
+- **GitHub Copilot** — Full OAuth auth, API routing, token validation, and 400k context. ([#1924](https://github.com/NousResearch/hermes-agent/pull/1924), [#1896](https://github.com/NousResearch/hermes-agent/pull/1896), [#1879](https://github.com/NousResearch/hermes-agent/pull/1879) by @mchzimm, [#2507](https://github.com/NousResearch/hermes-agent/pull/2507))
+- **Alibaba Cloud / DashScope** — Full integration with DashScope v1 runtime, model dot preservation, and 401 auth fixes ([#1673](https://github.com/NousResearch/hermes-agent/pull/1673), [#2332](https://github.com/NousResearch/hermes-agent/pull/2332), [#2459](https://github.com/NousResearch/hermes-agent/pull/2459))
+- **Kilo Code** — First-class inference provider ([#1666](https://github.com/NousResearch/hermes-agent/pull/1666))
+- **OpenCode Zen and OpenCode Go** — New provider backends ([#1650](https://github.com/NousResearch/hermes-agent/pull/1650), [#2393](https://github.com/NousResearch/hermes-agent/pull/2393) by @0xbyt4)
+- **NeuTTS** — Local TTS provider backend with built-in setup flow, replacing the old optional skill ([#1657](https://github.com/NousResearch/hermes-agent/pull/1657), [#1664](https://github.com/NousResearch/hermes-agent/pull/1664))
+
+### Provider Improvements
+- **Eager fallback** to backup model on rate-limit errors ([#1730](https://github.com/NousResearch/hermes-agent/pull/1730))
+- **Endpoint metadata** for custom model context and pricing; query local servers for actual context window size ([#1906](https://github.com/NousResearch/hermes-agent/pull/1906), [#2091](https://github.com/NousResearch/hermes-agent/pull/2091) by @dusterbloom)
+- **Context length detection overhaul** — models.dev integration, provider-aware resolution, fuzzy matching for custom endpoints, `/v1/props` for llama.cpp ([#2158](https://github.com/NousResearch/hermes-agent/pull/2158), [#2051](https://github.com/NousResearch/hermes-agent/pull/2051), [#2403](https://github.com/NousResearch/hermes-agent/pull/2403))
+- **Model catalog updates** — gpt-5.4-mini, gpt-5.4-nano, healer-alpha, haiku-4.5, minimax-m2.7, claude 4.6 at 1M context ([#1913](https://github.com/NousResearch/hermes-agent/pull/1913), [#1915](https://github.com/NousResearch/hermes-agent/pull/1915), [#1900](https://github.com/NousResearch/hermes-agent/pull/1900), [#2155](https://github.com/NousResearch/hermes-agent/pull/2155), [#2474](https://github.com/NousResearch/hermes-agent/pull/2474))
+- **Custom endpoint improvements** — `model.base_url` in config.yaml, `api_mode` override for responses API, allow endpoints without API key, fail fast on missing keys ([#2330](https://github.com/NousResearch/hermes-agent/pull/2330), [#1651](https://github.com/NousResearch/hermes-agent/pull/1651), [#2556](https://github.com/NousResearch/hermes-agent/pull/2556), [#2445](https://github.com/NousResearch/hermes-agent/pull/2445), [#1994](https://github.com/NousResearch/hermes-agent/pull/1994), [#1998](https://github.com/NousResearch/hermes-agent/pull/1998))
+- Inject model and provider into system prompt ([#1929](https://github.com/NousResearch/hermes-agent/pull/1929))
+- Tie `api_mode` to provider config instead of env var ([#1656](https://github.com/NousResearch/hermes-agent/pull/1656))
+- Fix: prevent Anthropic token leaking to third-party `anthropic_messages` providers ([#2389](https://github.com/NousResearch/hermes-agent/pull/2389))
+- Fix: prevent Anthropic fallback from inheriting non-Anthropic `base_url` ([#2388](https://github.com/NousResearch/hermes-agent/pull/2388))
+- Fix: `auxiliary_is_nous` flag never resets — leaked Nous tags to other providers ([#1713](https://github.com/NousResearch/hermes-agent/pull/1713))
+- Fix: Anthropic `tool_choice 'none'` still allowed tool calls ([#1714](https://github.com/NousResearch/hermes-agent/pull/1714))
+- Fix: Mistral parser nested JSON fallback extraction ([#2335](https://github.com/NousResearch/hermes-agent/pull/2335))
+- Fix: MiniMax 401 auth resolved by defaulting to `anthropic_messages` ([#2103](https://github.com/NousResearch/hermes-agent/pull/2103))
+- Fix: case-insensitive model family matching ([#2350](https://github.com/NousResearch/hermes-agent/pull/2350))
+- Fix: ignore placeholder provider keys in activation checks ([#2358](https://github.com/NousResearch/hermes-agent/pull/2358))
+- Fix: Preserve Ollama model:tag colons in context length detection ([#2149](https://github.com/NousResearch/hermes-agent/pull/2149))
+- Fix: recognize Claude Code OAuth credentials in startup gate ([#1663](https://github.com/NousResearch/hermes-agent/pull/1663))
+- Fix: detect Claude Code version dynamically for OAuth user-agent ([#1670](https://github.com/NousResearch/hermes-agent/pull/1670))
+- Fix: OAuth flag stale after refresh/fallback ([#1890](https://github.com/NousResearch/hermes-agent/pull/1890))
+- Fix: auxiliary client skips expired Codex JWT ([#2397](https://github.com/NousResearch/hermes-agent/pull/2397))
+
+### Agent Loop
+- **Gateway prompt caching** — Cache AIAgent per session, keep assistant turns, fix session restore ([#2282](https://github.com/NousResearch/hermes-agent/pull/2282), [#2284](https://github.com/NousResearch/hermes-agent/pull/2284), [#2361](https://github.com/NousResearch/hermes-agent/pull/2361))
+- **Context compression overhaul** — Structured summaries, iterative updates, token-budget tail protection, configurable `summary_base_url` ([#2323](https://github.com/NousResearch/hermes-agent/pull/2323), [#1727](https://github.com/NousResearch/hermes-agent/pull/1727), [#2224](https://github.com/NousResearch/hermes-agent/pull/2224))
+- **Pre-call sanitization and post-call tool guardrails** ([#1732](https://github.com/NousResearch/hermes-agent/pull/1732))
+- **Auto-recover** from provider-rejected `tool_choice` by retrying without ([#2174](https://github.com/NousResearch/hermes-agent/pull/2174))
+- **Background memory/skill review** replaces inline nudges ([#2235](https://github.com/NousResearch/hermes-agent/pull/2235))
+- **SOUL.md as primary agent identity** instead of hardcoded default ([#1922](https://github.com/NousResearch/hermes-agent/pull/1922))
+- Fix: prevent silent tool result loss during context compression ([#1993](https://github.com/NousResearch/hermes-agent/pull/1993))
+- Fix: handle empty/null function arguments in tool call recovery ([#2163](https://github.com/NousResearch/hermes-agent/pull/2163))
+- Fix: handle API refusal responses gracefully instead of crashing ([#2156](https://github.com/NousResearch/hermes-agent/pull/2156))
+- Fix: prevent stuck agent loop on malformed tool calls ([#2114](https://github.com/NousResearch/hermes-agent/pull/2114))
+- Fix: return JSON parse error to model instead of dispatching with empty args ([#2342](https://github.com/NousResearch/hermes-agent/pull/2342))
+- Fix: consecutive assistant message merge drops content on mixed types ([#1703](https://github.com/NousResearch/hermes-agent/pull/1703))
+- Fix: message role alternation violations in JSON recovery and error handler ([#1722](https://github.com/NousResearch/hermes-agent/pull/1722))
+- Fix: `compression_attempts` resets each iteration — allowed unlimited compressions ([#1723](https://github.com/NousResearch/hermes-agent/pull/1723))
+- Fix: `length_continue_retries` never resets — later truncations got fewer retries ([#1717](https://github.com/NousResearch/hermes-agent/pull/1717))
+- Fix: compressor summary role violated consecutive-role constraint ([#1720](https://github.com/NousResearch/hermes-agent/pull/1720), [#1743](https://github.com/NousResearch/hermes-agent/pull/1743))
+- Fix: remove hardcoded `gemini-3-flash-preview` as default summary model ([#2464](https://github.com/NousResearch/hermes-agent/pull/2464))
+- Fix: correctly handle empty tool results ([#2201](https://github.com/NousResearch/hermes-agent/pull/2201))
+- Fix: crash on None entry in `tool_calls` list ([#2209](https://github.com/NousResearch/hermes-agent/pull/2209) by @0xbyt4, [#2316](https://github.com/NousResearch/hermes-agent/pull/2316))
+- Fix: per-thread persistent event loops in worker threads ([#2214](https://github.com/NousResearch/hermes-agent/pull/2214) by @jquesnelle)
+- Fix: prevent 'event loop already running' when async tools run in parallel ([#2207](https://github.com/NousResearch/hermes-agent/pull/2207))
+- Fix: strip ANSI at the source — clean terminal output before it reaches the model ([#2115](https://github.com/NousResearch/hermes-agent/pull/2115))
+- Fix: skip top-level `cache_control` on role:tool for OpenRouter ([#2391](https://github.com/NousResearch/hermes-agent/pull/2391))
+- Fix: delegate tool — save parent tool names before child construction mutates global ([#2083](https://github.com/NousResearch/hermes-agent/pull/2083) by @ygd58, [#1894](https://github.com/NousResearch/hermes-agent/pull/1894))
+- Fix: only strip last assistant message if empty string ([#2326](https://github.com/NousResearch/hermes-agent/pull/2326))
+
+### Session & Memory
+- **Session search** and management slash commands ([#2198](https://github.com/NousResearch/hermes-agent/pull/2198))
+- **Auto session titles** and `.hermes.md` project config ([#1712](https://github.com/NousResearch/hermes-agent/pull/1712))
+- Fix: concurrent memory writes silently drop entries — added file locking ([#1726](https://github.com/NousResearch/hermes-agent/pull/1726))
+- Fix: search all sources by default in `session_search` ([#1892](https://github.com/NousResearch/hermes-agent/pull/1892))
+- Fix: handle hyphenated FTS5 queries and preserve quoted literals ([#1776](https://github.com/NousResearch/hermes-agent/pull/1776))
+- Fix: skip corrupt lines in `load_transcript` instead of crashing ([#1744](https://github.com/NousResearch/hermes-agent/pull/1744))
+- Fix: normalize session keys to prevent case-sensitive duplicates ([#2157](https://github.com/NousResearch/hermes-agent/pull/2157))
+- Fix: prevent `session_search` crash when no sessions exist ([#2194](https://github.com/NousResearch/hermes-agent/pull/2194))
+- Fix: reset token counters on new session for accurate usage display ([#2101](https://github.com/NousResearch/hermes-agent/pull/2101) by @InB4DevOps)
+- Fix: prevent stale memory overwrites by flush agent ([#2687](https://github.com/NousResearch/hermes-agent/pull/2687))
+- Fix: remove synthetic error message injection, fix session resume after repeated failures ([#2303](https://github.com/NousResearch/hermes-agent/pull/2303))
+- Fix: quiet mode with `--resume` now passes conversation_history ([#2357](https://github.com/NousResearch/hermes-agent/pull/2357))
+- Fix: unify resume logic in batch mode ([#2331](https://github.com/NousResearch/hermes-agent/pull/2331))
+
+### Honcho Memory
+- Honcho config fixes and @ context reference integration ([#2343](https://github.com/NousResearch/hermes-agent/pull/2343))
+- Self-hosted / Docker configuration documentation ([#2475](https://github.com/NousResearch/hermes-agent/pull/2475))
+
+---
+
+## 📱 Messaging Platforms (Gateway)
+
+### New Platform Adapters
+- **Signal Messenger** — Full adapter with attachment handling, group message filtering, and Note to Self echo-back protection ([#2206](https://github.com/NousResearch/hermes-agent/pull/2206), [#2400](https://github.com/NousResearch/hermes-agent/pull/2400), [#2297](https://github.com/NousResearch/hermes-agent/pull/2297), [#2156](https://github.com/NousResearch/hermes-agent/pull/2156))
+- **DingTalk** — Adapter with gateway wiring and setup docs ([#1685](https://github.com/NousResearch/hermes-agent/pull/1685), [#1690](https://github.com/NousResearch/hermes-agent/pull/1690), [#1692](https://github.com/NousResearch/hermes-agent/pull/1692))
+- **SMS (Twilio)** ([#1688](https://github.com/NousResearch/hermes-agent/pull/1688))
+- **Mattermost** — With @-mention-only channel filter ([#1683](https://github.com/NousResearch/hermes-agent/pull/1683), [#2443](https://github.com/NousResearch/hermes-agent/pull/2443))
+- **Matrix** — With vision support and image caching ([#1683](https://github.com/NousResearch/hermes-agent/pull/1683), [#2520](https://github.com/NousResearch/hermes-agent/pull/2520))
+- **Webhook** — Platform adapter for external event triggers ([#2166](https://github.com/NousResearch/hermes-agent/pull/2166))
+- **OpenAI-compatible API server** — `/v1/chat/completions` endpoint with `/api/jobs` cron management ([#1756](https://github.com/NousResearch/hermes-agent/pull/1756), [#2450](https://github.com/NousResearch/hermes-agent/pull/2450), [#2456](https://github.com/NousResearch/hermes-agent/pull/2456))
+
+### Telegram Improvements
+- MarkdownV2 support — strikethrough, spoiler, blockquotes, escape parentheses/braces/backslashes/backticks ([#2199](https://github.com/NousResearch/hermes-agent/pull/2199), [#2200](https://github.com/NousResearch/hermes-agent/pull/2200) by @llbn, [#2386](https://github.com/NousResearch/hermes-agent/pull/2386))
+- Auto-detect HTML tags and use `parse_mode=HTML` ([#1709](https://github.com/NousResearch/hermes-agent/pull/1709))
+- Telegram group vision support + thread-based sessions ([#2153](https://github.com/NousResearch/hermes-agent/pull/2153))
+- Auto-reconnect polling after network interruption ([#2517](https://github.com/NousResearch/hermes-agent/pull/2517))
+- Aggregate split text messages before dispatching ([#1674](https://github.com/NousResearch/hermes-agent/pull/1674))
+- Fix: streaming config bridge, not-modified, flood control ([#1782](https://github.com/NousResearch/hermes-agent/pull/1782), [#1783](https://github.com/NousResearch/hermes-agent/pull/1783))
+- Fix: edited_message event crashes ([#2074](https://github.com/NousResearch/hermes-agent/pull/2074))
+- Fix: retry 409 polling conflicts before giving up ([#2312](https://github.com/NousResearch/hermes-agent/pull/2312))
+- Fix: topic delivery via `platform:chat_id:thread_id` format ([#2455](https://github.com/NousResearch/hermes-agent/pull/2455))
+
+### Discord Improvements
+- Document caching and text-file injection ([#2503](https://github.com/NousResearch/hermes-agent/pull/2503))
+- Persistent typing indicator for DMs ([#2468](https://github.com/NousResearch/hermes-agent/pull/2468))
+- Discord DM vision — inline images + attachment analysis ([#2186](https://github.com/NousResearch/hermes-agent/pull/2186))
+- Persist thread participation across gateway restarts ([#1661](https://github.com/NousResearch/hermes-agent/pull/1661))
+- Fix: gateway crash on non-ASCII guild names ([#2302](https://github.com/NousResearch/hermes-agent/pull/2302))
+- Fix: thread permission errors ([#2073](https://github.com/NousResearch/hermes-agent/pull/2073))
+- Fix: slash event routing in threads ([#2460](https://github.com/NousResearch/hermes-agent/pull/2460))
+- Fix: remove bugged followup messages + `/ask` command ([#1836](https://github.com/NousResearch/hermes-agent/pull/1836))
+- Fix: graceful WebSocket reconnection ([#2127](https://github.com/NousResearch/hermes-agent/pull/2127))
+- Fix: voice channel TTS when streaming enabled ([#2322](https://github.com/NousResearch/hermes-agent/pull/2322))
+
+### WhatsApp & Other Adapters
+- WhatsApp: outbound `send_message` routing ([#1769](https://github.com/NousResearch/hermes-agent/pull/1769) by @sai-samarth), LID format self-chat ([#1667](https://github.com/NousResearch/hermes-agent/pull/1667)), `reply_prefix` config fix ([#1923](https://github.com/NousResearch/hermes-agent/pull/1923)), restart on bridge child exit ([#2334](https://github.com/NousResearch/hermes-agent/pull/2334)), image/bridge improvements ([#2181](https://github.com/NousResearch/hermes-agent/pull/2181))
+- Matrix: correct `reply_to_message_id` parameter ([#1895](https://github.com/NousResearch/hermes-agent/pull/1895)), bare media types fix ([#1736](https://github.com/NousResearch/hermes-agent/pull/1736))
+- Mattermost: MIME types for media attachments ([#2329](https://github.com/NousResearch/hermes-agent/pull/2329))
+
+### Gateway Core
+- **Auto-reconnect** failed platforms with exponential backoff ([#2584](https://github.com/NousResearch/hermes-agent/pull/2584))
+- **Notify users when session auto-resets** ([#2519](https://github.com/NousResearch/hermes-agent/pull/2519))
+- **Reply-to message context** for out-of-session replies ([#1662](https://github.com/NousResearch/hermes-agent/pull/1662))
+- **Ignore unauthorized DMs** config option ([#1919](https://github.com/NousResearch/hermes-agent/pull/1919))
+- Fix: `/reset` in thread-mode resets global session instead of thread ([#2254](https://github.com/NousResearch/hermes-agent/pull/2254))
+- Fix: deliver MEDIA: files after streaming responses ([#2382](https://github.com/NousResearch/hermes-agent/pull/2382))
+- Fix: cap interrupt recursion depth to prevent resource exhaustion ([#1659](https://github.com/NousResearch/hermes-agent/pull/1659))
+- Fix: detect stopped processes and release stale locks on `--replace` ([#2406](https://github.com/NousResearch/hermes-agent/pull/2406), [#1908](https://github.com/NousResearch/hermes-agent/pull/1908))
+- Fix: PID-based wait with force-kill for gateway restart ([#1902](https://github.com/NousResearch/hermes-agent/pull/1902))
+- Fix: prevent `--replace` mode from killing the caller process ([#2185](https://github.com/NousResearch/hermes-agent/pull/2185))
+- Fix: `/model` shows active fallback model instead of config default ([#1660](https://github.com/NousResearch/hermes-agent/pull/1660))
+- Fix: `/title` command fails when session doesn't exist in SQLite yet ([#2379](https://github.com/NousResearch/hermes-agent/pull/2379) by @ten-jampa)
+- Fix: process `/queue`'d messages after agent completion ([#2469](https://github.com/NousResearch/hermes-agent/pull/2469))
+- Fix: strip orphaned `tool_results` + let `/reset` bypass running agent ([#2180](https://github.com/NousResearch/hermes-agent/pull/2180))
+- Fix: prevent agents from starting gateway outside systemd management ([#2617](https://github.com/NousResearch/hermes-agent/pull/2617))
+- Fix: prevent systemd restart storm on gateway connection failure ([#2327](https://github.com/NousResearch/hermes-agent/pull/2327))
+- Fix: include resolved node path in systemd unit ([#1767](https://github.com/NousResearch/hermes-agent/pull/1767) by @sai-samarth)
+- Fix: send error details to user in gateway outer exception handler ([#1966](https://github.com/NousResearch/hermes-agent/pull/1966))
+- Fix: improve error handling for 429 usage limits and 500 context overflow ([#1839](https://github.com/NousResearch/hermes-agent/pull/1839))
+- Fix: add all missing platform allowlist env vars to startup warning check ([#2628](https://github.com/NousResearch/hermes-agent/pull/2628))
+- Fix: media delivery fails for file paths containing spaces ([#2621](https://github.com/NousResearch/hermes-agent/pull/2621))
+- Fix: duplicate session-key collision in multi-platform gateway ([#2171](https://github.com/NousResearch/hermes-agent/pull/2171))
+- Fix: Matrix and Mattermost never report as connected ([#1711](https://github.com/NousResearch/hermes-agent/pull/1711))
+- Fix: PII redaction config never read — missing yaml import ([#1701](https://github.com/NousResearch/hermes-agent/pull/1701))
+- Fix: NameError on skill slash commands ([#1697](https://github.com/NousResearch/hermes-agent/pull/1697))
+- Fix: persist watcher metadata in checkpoint for crash recovery ([#1706](https://github.com/NousResearch/hermes-agent/pull/1706))
+- Fix: pass `message_thread_id` in send_image_file, send_document, send_video ([#2339](https://github.com/NousResearch/hermes-agent/pull/2339))
+- Fix: media-group aggregation on rapid successive photo messages ([#2160](https://github.com/NousResearch/hermes-agent/pull/2160))
+
+---
+
+## 🔧 Tool System
+
+### MCP Enhancements
+- **MCP server management CLI** + OAuth 2.1 PKCE auth ([#2465](https://github.com/NousResearch/hermes-agent/pull/2465))
+- **Expose MCP servers as standalone toolsets** ([#1907](https://github.com/NousResearch/hermes-agent/pull/1907))
+- **Interactive MCP tool configuration** in `hermes tools` ([#1694](https://github.com/NousResearch/hermes-agent/pull/1694))
+- Fix: MCP-OAuth port mismatch, path traversal, and shared handler state ([#2552](https://github.com/NousResearch/hermes-agent/pull/2552))
+- Fix: preserve MCP tool registrations across session resets ([#2124](https://github.com/NousResearch/hermes-agent/pull/2124))
+- Fix: concurrent file access crash + duplicate MCP registration ([#2154](https://github.com/NousResearch/hermes-agent/pull/2154))
+- Fix: normalise MCP schemas + expand session list columns ([#2102](https://github.com/NousResearch/hermes-agent/pull/2102))
+- Fix: `tool_choice` `mcp_` prefix handling ([#1775](https://github.com/NousResearch/hermes-agent/pull/1775))
+
+### Web Tool Backends
+- **Tavily** as web search/extract/crawl backend ([#1731](https://github.com/NousResearch/hermes-agent/pull/1731))
+- **Parallel** as alternative web search/extract backend ([#1696](https://github.com/NousResearch/hermes-agent/pull/1696))
+- **Configurable web backend** — Firecrawl/BeautifulSoup/Playwright selection ([#2256](https://github.com/NousResearch/hermes-agent/pull/2256))
+- Fix: whitespace-only env vars bypass web backend detection ([#2341](https://github.com/NousResearch/hermes-agent/pull/2341))
+
+### New Tools
+- **IMAP email** reading and sending ([#2173](https://github.com/NousResearch/hermes-agent/pull/2173))
+- **STT (speech-to-text)** tool using Whisper API ([#2072](https://github.com/NousResearch/hermes-agent/pull/2072))
+- **Route-aware pricing estimates** ([#1695](https://github.com/NousResearch/hermes-agent/pull/1695))
+
+### Tool Improvements
+- TTS: `base_url` support for OpenAI TTS provider ([#2064](https://github.com/NousResearch/hermes-agent/pull/2064) by @hanai)
+- Vision: configurable timeout, tilde expansion in file paths, DM vision with multi-image and base64 fallback ([#2480](https://github.com/NousResearch/hermes-agent/pull/2480), [#2585](https://github.com/NousResearch/hermes-agent/pull/2585), [#2211](https://github.com/NousResearch/hermes-agent/pull/2211))
+- Browser: race condition fix in session creation ([#1721](https://github.com/NousResearch/hermes-agent/pull/1721)), TypeError on unexpected LLM params ([#1735](https://github.com/NousResearch/hermes-agent/pull/1735))
+- File tools: strip ANSI escape codes from write_file and patch content ([#2532](https://github.com/NousResearch/hermes-agent/pull/2532)), include pagination args in repeated search key ([#1824](https://github.com/NousResearch/hermes-agent/pull/1824) by @cutepawss), improve fuzzy matching accuracy + position calculation refactor ([#2096](https://github.com/NousResearch/hermes-agent/pull/2096), [#1681](https://github.com/NousResearch/hermes-agent/pull/1681))
+- Code execution: resource leak and double socket close fix ([#2381](https://github.com/NousResearch/hermes-agent/pull/2381))
+- Delegate: thread safety for concurrent subagent delegation ([#1672](https://github.com/NousResearch/hermes-agent/pull/1672)), preserve parent agent's tool list after delegation ([#1778](https://github.com/NousResearch/hermes-agent/pull/1778))
+- Fix: make concurrent tool batching path-aware for file mutations ([#1914](https://github.com/NousResearch/hermes-agent/pull/1914))
+- Fix: chunk long messages in `send_message_tool` before platform dispatch ([#1646](https://github.com/NousResearch/hermes-agent/pull/1646))
+- Fix: add missing 'messaging' toolset ([#1718](https://github.com/NousResearch/hermes-agent/pull/1718))
+- Fix: prevent unavailable tool names from leaking into model schemas ([#2072](https://github.com/NousResearch/hermes-agent/pull/2072))
+- Fix: pass visited set by reference to prevent diamond dependency duplication ([#2311](https://github.com/NousResearch/hermes-agent/pull/2311))
+- Fix: Daytona sandbox lookup migrated from `find_one` to `get/list` ([#2063](https://github.com/NousResearch/hermes-agent/pull/2063) by @rovle)
+
+---
+
+## 🧩 Skills Ecosystem
+
+### Skills System Improvements
+- **Agent-created skills** — Caution-level findings allowed, dangerous skills ask instead of block ([#1840](https://github.com/NousResearch/hermes-agent/pull/1840), [#2446](https://github.com/NousResearch/hermes-agent/pull/2446))
+- **`--yes` flag** to bypass confirmation in `/skills install` and uninstall ([#1647](https://github.com/NousResearch/hermes-agent/pull/1647))
+- **Disabled skills respected** across banner, system prompt, and slash commands ([#1897](https://github.com/NousResearch/hermes-agent/pull/1897))
+- Fix: skills custom_tools import crash + sandbox file_tools integration ([#2239](https://github.com/NousResearch/hermes-agent/pull/2239))
+- Fix: agent-created skills with pip requirements crash on install ([#2145](https://github.com/NousResearch/hermes-agent/pull/2145))
+- Fix: race condition in `Skills.__init__` when `hub.yaml` missing ([#2242](https://github.com/NousResearch/hermes-agent/pull/2242))
+- Fix: validate skill metadata before install and block duplicates ([#2241](https://github.com/NousResearch/hermes-agent/pull/2241))
+- Fix: skills hub inspect/resolve — 4 bugs in inspect, redirects, discovery, tap list ([#2447](https://github.com/NousResearch/hermes-agent/pull/2447))
+- Fix: agent-created skills keep working after session reset ([#2121](https://github.com/NousResearch/hermes-agent/pull/2121))
+
+### New Skills
+- **OCR-and-documents** — PDF/DOCX/XLS/PPTX/image OCR with optional GPU ([#2236](https://github.com/NousResearch/hermes-agent/pull/2236), [#2461](https://github.com/NousResearch/hermes-agent/pull/2461))
+- **Huggingface-hub** bundled skill ([#1921](https://github.com/NousResearch/hermes-agent/pull/1921))
+- **Sherlock OSINT** username search ([#1671](https://github.com/NousResearch/hermes-agent/pull/1671))
+- **Meme-generation** — Image generator with Pillow ([#2344](https://github.com/NousResearch/hermes-agent/pull/2344))
+- **Bioinformatics** gateway skill — index to 400+ bio skills ([#2387](https://github.com/NousResearch/hermes-agent/pull/2387))
+- **Inference.sh** skill (terminal-based) ([#1686](https://github.com/NousResearch/hermes-agent/pull/1686))
+- **Base blockchain** optional skill ([#1643](https://github.com/NousResearch/hermes-agent/pull/1643))
+- **3D-model-viewer** optional skill ([#2226](https://github.com/NousResearch/hermes-agent/pull/2226))
+- **FastMCP** optional skill ([#2113](https://github.com/NousResearch/hermes-agent/pull/2113))
+- **Hermes-agent-setup** skill ([#1905](https://github.com/NousResearch/hermes-agent/pull/1905))
+
+---
+
+## 🔌 Plugin System Enhancements
+
+- **TUI extension hooks** — Build custom CLIs on top of Hermes ([#2333](https://github.com/NousResearch/hermes-agent/pull/2333))
+- **`hermes plugins install/remove/list`** commands ([#2337](https://github.com/NousResearch/hermes-agent/pull/2337))
+- **Slash command registration** for plugins ([#2359](https://github.com/NousResearch/hermes-agent/pull/2359))
+- **`session:end` lifecycle event** hook ([#1725](https://github.com/NousResearch/hermes-agent/pull/1725))
+- Fix: require opt-in for project plugin discovery ([#2215](https://github.com/NousResearch/hermes-agent/pull/2215))
+
+---
+
+## 🔒 Security & Reliability
+
+### Security
+- **SSRF protection** for vision_tools and web_tools ([#2679](https://github.com/NousResearch/hermes-agent/pull/2679))
+- **Shell injection prevention** in `_expand_path` via `~user` path suffix ([#2685](https://github.com/NousResearch/hermes-agent/pull/2685))
+- **Block untrusted browser-origin** API server access ([#2451](https://github.com/NousResearch/hermes-agent/pull/2451))
+- **Block sandbox backend creds** from subprocess env ([#1658](https://github.com/NousResearch/hermes-agent/pull/1658))
+- **Block @ references** from reading secrets outside workspace ([#2601](https://github.com/NousResearch/hermes-agent/pull/2601) by @Gutslabs)
+- **Malicious code pattern pre-exec scanner** for terminal_tool ([#2245](https://github.com/NousResearch/hermes-agent/pull/2245))
+- **Harden terminal safety** and sandbox file writes ([#1653](https://github.com/NousResearch/hermes-agent/pull/1653))
+- **PKCE verifier leak** fix + OAuth refresh Content-Type ([#1775](https://github.com/NousResearch/hermes-agent/pull/1775))
+- **Eliminate SQL string formatting** in `execute()` calls ([#2061](https://github.com/NousResearch/hermes-agent/pull/2061) by @dusterbloom)
+- **Harden jobs API** — input limits, field whitelist, startup check ([#2456](https://github.com/NousResearch/hermes-agent/pull/2456))
+
+### Reliability
+- Thread locks on 4 SessionDB methods ([#1704](https://github.com/NousResearch/hermes-agent/pull/1704))
+- File locking for concurrent memory writes ([#1726](https://github.com/NousResearch/hermes-agent/pull/1726))
+- Handle OpenRouter errors gracefully ([#2112](https://github.com/NousResearch/hermes-agent/pull/2112))
+- Guard print() calls against OSError ([#1668](https://github.com/NousResearch/hermes-agent/pull/1668))
+- Safely handle non-string inputs in redacting formatter ([#2392](https://github.com/NousResearch/hermes-agent/pull/2392), [#1700](https://github.com/NousResearch/hermes-agent/pull/1700))
+- ACP: preserve session provider on model switch, persist sessions to disk ([#2380](https://github.com/NousResearch/hermes-agent/pull/2380), [#2071](https://github.com/NousResearch/hermes-agent/pull/2071))
+- API server: persist ResponseStore to SQLite across restarts ([#2472](https://github.com/NousResearch/hermes-agent/pull/2472))
+- Fix: `fetch_nous_models` always TypeError from positional args ([#1699](https://github.com/NousResearch/hermes-agent/pull/1699))
+- Fix: resolve merge conflict markers in cli.py breaking startup ([#2347](https://github.com/NousResearch/hermes-agent/pull/2347))
+- Fix: `minisweagent_path.py` missing from wheel ([#2098](https://github.com/NousResearch/hermes-agent/pull/2098) by @JiwaniZakir)
+
+### Cron System
+- **`[SILENT]` response** — cron agents can suppress delivery ([#1833](https://github.com/NousResearch/hermes-agent/pull/1833))
+- **Scale missed-job grace window** with schedule frequency ([#2449](https://github.com/NousResearch/hermes-agent/pull/2449))
+- **Recover recent one-shot jobs** ([#1918](https://github.com/NousResearch/hermes-agent/pull/1918))
+- Fix: normalize `repeat<=0` to None — jobs deleted after first run when LLM passes -1 ([#2612](https://github.com/NousResearch/hermes-agent/pull/2612) by @Mibayy)
+- Fix: Matrix added to scheduler delivery platform_map ([#2167](https://github.com/NousResearch/hermes-agent/pull/2167) by @buntingszn)
+- Fix: naive ISO timestamps without timezone — jobs fire at wrong time ([#1729](https://github.com/NousResearch/hermes-agent/pull/1729))
+- Fix: `get_due_jobs` reads `jobs.json` twice — race condition ([#1716](https://github.com/NousResearch/hermes-agent/pull/1716))
+- Fix: silent jobs return empty response for delivery skip ([#2442](https://github.com/NousResearch/hermes-agent/pull/2442))
+- Fix: stop injecting cron outputs into gateway session history ([#2313](https://github.com/NousResearch/hermes-agent/pull/2313))
+- Fix: close abandoned coroutine when `asyncio.run()` raises RuntimeError ([#2317](https://github.com/NousResearch/hermes-agent/pull/2317))
+
+---
+
+## 🧪 Testing
+
+- Resolve all consistently failing tests ([#2488](https://github.com/NousResearch/hermes-agent/pull/2488))
+- Replace `FakePath` with `monkeypatch` for Python 3.12 compat ([#2444](https://github.com/NousResearch/hermes-agent/pull/2444))
+- Align Hermes setup and full-suite expectations ([#1710](https://github.com/NousResearch/hermes-agent/pull/1710))
+
+---
+
+## 📚 Documentation
+
+- Comprehensive docs update for recent features ([#1693](https://github.com/NousResearch/hermes-agent/pull/1693), [#2183](https://github.com/NousResearch/hermes-agent/pull/2183))
+- Alibaba Cloud and DingTalk setup guides ([#1687](https://github.com/NousResearch/hermes-agent/pull/1687), [#1692](https://github.com/NousResearch/hermes-agent/pull/1692))
+- Detailed skills documentation ([#2244](https://github.com/NousResearch/hermes-agent/pull/2244))
+- Honcho self-hosted / Docker configuration ([#2475](https://github.com/NousResearch/hermes-agent/pull/2475))
+- Context length detection FAQ and quickstart references ([#2179](https://github.com/NousResearch/hermes-agent/pull/2179))
+- Fix docs inconsistencies across reference and user guides ([#1995](https://github.com/NousResearch/hermes-agent/pull/1995))
+- Fix MCP install commands — use uv, not bare pip ([#1909](https://github.com/NousResearch/hermes-agent/pull/1909))
+- Replace ASCII diagrams with Mermaid/lists ([#2402](https://github.com/NousResearch/hermes-agent/pull/2402))
+- Gemini OAuth provider implementation plan ([#2467](https://github.com/NousResearch/hermes-agent/pull/2467))
+- Discord Server Members Intent marked as required ([#2330](https://github.com/NousResearch/hermes-agent/pull/2330))
+- Fix MDX build error in api-server.md ([#1787](https://github.com/NousResearch/hermes-agent/pull/1787))
+- Align venv path to match installer ([#2114](https://github.com/NousResearch/hermes-agent/pull/2114))
+- New skills added to hub index ([#2281](https://github.com/NousResearch/hermes-agent/pull/2281))
+
+---
+
+## 👥 Contributors
+
+### Core
+- **@teknium1** (Teknium) — 280 PRs
+
+### Community Contributors
+- **@mchzimm** (to_the_max) — GitHub Copilot provider integration ([#1879](https://github.com/NousResearch/hermes-agent/pull/1879))
+- **@jquesnelle** (Jeffrey Quesnelle) — Per-thread persistent event loops fix ([#2214](https://github.com/NousResearch/hermes-agent/pull/2214))
+- **@llbn** (lbn) — Telegram MarkdownV2 strikethrough, spoiler, blockquotes, and escape fixes ([#2199](https://github.com/NousResearch/hermes-agent/pull/2199), [#2200](https://github.com/NousResearch/hermes-agent/pull/2200))
+- **@dusterbloom** — SQL injection prevention + local server context window querying ([#2061](https://github.com/NousResearch/hermes-agent/pull/2061), [#2091](https://github.com/NousResearch/hermes-agent/pull/2091))
+- **@0xbyt4** — Anthropic tool_calls None guard + OpenCode-Go provider config fix ([#2209](https://github.com/NousResearch/hermes-agent/pull/2209), [#2393](https://github.com/NousResearch/hermes-agent/pull/2393))
+- **@sai-samarth** (Saisamarth) — WhatsApp send_message routing + systemd node path ([#1769](https://github.com/NousResearch/hermes-agent/pull/1769), [#1767](https://github.com/NousResearch/hermes-agent/pull/1767))
+- **@Gutslabs** (Guts) — Block @ references from reading secrets ([#2601](https://github.com/NousResearch/hermes-agent/pull/2601))
+- **@Mibayy** (Mibay) — Cron job repeat normalization ([#2612](https://github.com/NousResearch/hermes-agent/pull/2612))
+- **@ten-jampa** (Tenzin Jampa) — Gateway /title command fix ([#2379](https://github.com/NousResearch/hermes-agent/pull/2379))
+- **@cutepawss** (lila) — File tools search pagination fix ([#1824](https://github.com/NousResearch/hermes-agent/pull/1824))
+- **@hanai** (Hanai) — OpenAI TTS base_url support ([#2064](https://github.com/NousResearch/hermes-agent/pull/2064))
+- **@rovle** (Lovre Pešut) — Daytona sandbox API migration ([#2063](https://github.com/NousResearch/hermes-agent/pull/2063))
+- **@buntingszn** (bunting szn) — Matrix cron delivery support ([#2167](https://github.com/NousResearch/hermes-agent/pull/2167))
+- **@InB4DevOps** — Token counter reset on new session ([#2101](https://github.com/NousResearch/hermes-agent/pull/2101))
+- **@JiwaniZakir** (Zakir Jiwani) — Missing file in wheel fix ([#2098](https://github.com/NousResearch/hermes-agent/pull/2098))
+- **@ygd58** (buray) — Delegate tool parent tool names fix ([#2083](https://github.com/NousResearch/hermes-agent/pull/2083))
+
+---
+
+**Full Changelog**: [v2026.3.17...v2026.3.23](https://github.com/NousResearch/hermes-agent/compare/v2026.3.17...v2026.3.23)

RELEASE_v0.5.0.md

@@ -0,0 +1,348 @@
+# Hermes Agent v0.5.0 (v2026.3.28)
+
+**Release Date:** March 28, 2026
+
+> The hardening release — Hugging Face provider, /model command overhaul, Telegram Private Chat Topics, native Modal SDK, plugin lifecycle hooks, tool-use enforcement for GPT models, Nix flake, 50+ security and reliability fixes, and a comprehensive supply chain audit.
+
+---
+
+## ✨ Highlights
+
+- **Nous Portal now supports 400+ models** — The Nous Research inference portal has expanded dramatically, giving Hermes Agent users access to over 400 models through a single provider endpoint
+
+- **Hugging Face as a first-class inference provider** — Full integration with HF Inference API including curated agentic model picker that maps to OpenRouter analogues, live `/models` endpoint probe, and setup wizard flow ([#3419](https://github.com/NousResearch/hermes-agent/pull/3419), [#3440](https://github.com/NousResearch/hermes-agent/pull/3440))
+
+- **Telegram Private Chat Topics** — Project-based conversations with functional skill binding per topic, enabling isolated workflows within a single Telegram chat ([#3163](https://github.com/NousResearch/hermes-agent/pull/3163))
+
+- **Native Modal SDK backend** — Replaced swe-rex dependency with native Modal SDK (`Sandbox.create.aio` + `exec.aio`), eliminating tunnels and simplifying the Modal terminal backend ([#3538](https://github.com/NousResearch/hermes-agent/pull/3538))
+
+- **Plugin lifecycle hooks activated** — `pre_llm_call`, `post_llm_call`, `on_session_start`, and `on_session_end` hooks now fire in the agent loop and CLI/gateway, completing the plugin hook system ([#3542](https://github.com/NousResearch/hermes-agent/pull/3542))
+
+- **Improved OpenAI Model Reliability** — Added `GPT_TOOL_USE_GUIDANCE` to prevent GPT models from describing intended actions instead of making tool calls, plus automatic stripping of stale budget warnings from conversation history that caused models to avoid tools across turns ([#3528](https://github.com/NousResearch/hermes-agent/pull/3528))
+
+- **Nix flake** — Full uv2nix build, NixOS module with persistent container mode, auto-generated config keys from Python source, and suffix PATHs for agent-friendliness ([#20](https://github.com/NousResearch/hermes-agent/pull/20), [#3274](https://github.com/NousResearch/hermes-agent/pull/3274), [#3061](https://github.com/NousResearch/hermes-agent/pull/3061)) by @alt-glitch
+
+- **Supply chain hardening** — Removed compromised `litellm` dependency, pinned all dependency version ranges, regenerated `uv.lock` with hashes, added CI workflow scanning PRs for supply chain attack patterns, and bumped deps to fix CVEs ([#2796](https://github.com/NousResearch/hermes-agent/pull/2796), [#2810](https://github.com/NousResearch/hermes-agent/pull/2810), [#2812](https://github.com/NousResearch/hermes-agent/pull/2812), [#2816](https://github.com/NousResearch/hermes-agent/pull/2816), [#3073](https://github.com/NousResearch/hermes-agent/pull/3073))
+
+- **Anthropic output limits fix** — Replaced hardcoded 16K `max_tokens` with per-model native output limits (128K for Opus 4.6, 64K for Sonnet 4.6), fixing "Response truncated" and thinking-budget exhaustion on direct Anthropic API ([#3426](https://github.com/NousResearch/hermes-agent/pull/3426), [#3444](https://github.com/NousResearch/hermes-agent/pull/3444))
+
+---
+
+## 🏗️ Core Agent & Architecture
+
+### New Provider: Hugging Face
+- First-class Hugging Face Inference API integration with auth, setup wizard, and model picker ([#3419](https://github.com/NousResearch/hermes-agent/pull/3419))
+- Curated model list mapping OpenRouter agentic defaults to HF equivalents — providers with 8+ curated models skip live `/models` probe for speed ([#3440](https://github.com/NousResearch/hermes-agent/pull/3440))
+- Added glm-5-turbo to Z.AI provider model list ([#3095](https://github.com/NousResearch/hermes-agent/pull/3095))
+
+### Provider & Model Improvements
+- `/model` command overhaul — extracted shared `switch_model()` pipeline for CLI and gateway, custom endpoint support, provider-aware routing ([#2795](https://github.com/NousResearch/hermes-agent/pull/2795), [#2799](https://github.com/NousResearch/hermes-agent/pull/2799))
+- Removed `/model` slash command from CLI and gateway in favor of `hermes model` subcommand ([#3080](https://github.com/NousResearch/hermes-agent/pull/3080))
+- Preserve `custom` provider instead of silently remapping to `openrouter` ([#2792](https://github.com/NousResearch/hermes-agent/pull/2792))
+- Read root-level `provider` and `base_url` from config.yaml into model config ([#3112](https://github.com/NousResearch/hermes-agent/pull/3112))
+- Align Nous Portal model slugs with OpenRouter naming ([#3253](https://github.com/NousResearch/hermes-agent/pull/3253))
+- Fix Alibaba provider default endpoint and model list ([#3484](https://github.com/NousResearch/hermes-agent/pull/3484))
+- Allow MiniMax users to override `/v1` → `/anthropic` auto-correction ([#3553](https://github.com/NousResearch/hermes-agent/pull/3553))
+- Migrate OAuth token refresh to `platform.claude.com` with fallback ([#3246](https://github.com/NousResearch/hermes-agent/pull/3246))
+
+### Agent Loop & Conversation
+- **Improved OpenAI model reliability** — `GPT_TOOL_USE_GUIDANCE` prevents GPT models from describing actions instead of calling tools + automatic budget warning stripping from history ([#3528](https://github.com/NousResearch/hermes-agent/pull/3528))
+- **Surface lifecycle events** — All retry, fallback, and compression events now surface to the user as formatted messages ([#3153](https://github.com/NousResearch/hermes-agent/pull/3153))
+- **Anthropic output limits** — Per-model native output limits instead of hardcoded 16K `max_tokens` ([#3426](https://github.com/NousResearch/hermes-agent/pull/3426))
+- **Thinking-budget exhaustion detection** — Skip useless continuation retries when model uses all output tokens on reasoning ([#3444](https://github.com/NousResearch/hermes-agent/pull/3444))
+- Always prefer streaming for API calls to prevent hung subagents ([#3120](https://github.com/NousResearch/hermes-agent/pull/3120))
+- Restore safe non-streaming fallback after stream failures ([#3020](https://github.com/NousResearch/hermes-agent/pull/3020))
+- Give subagents independent iteration budgets ([#3004](https://github.com/NousResearch/hermes-agent/pull/3004))
+- Update `api_key` in `_try_activate_fallback` for subagent auth ([#3103](https://github.com/NousResearch/hermes-agent/pull/3103))
+- Graceful return on max retries instead of crashing thread ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Count compression restarts toward retry limit ([#3070](https://github.com/NousResearch/hermes-agent/pull/3070))
+- Include tool tokens in preflight estimate, guard context probe persistence ([#3164](https://github.com/NousResearch/hermes-agent/pull/3164))
+- Update context compressor limits after fallback activation ([#3305](https://github.com/NousResearch/hermes-agent/pull/3305))
+- Validate empty user messages to prevent Anthropic API 400 errors ([#3322](https://github.com/NousResearch/hermes-agent/pull/3322))
+- GLM reasoning-only and max-length handling ([#3010](https://github.com/NousResearch/hermes-agent/pull/3010))
+- Increase API timeout default from 900s to 1800s for slow-thinking models ([#3431](https://github.com/NousResearch/hermes-agent/pull/3431))
+- Send `max_tokens` for Claude/OpenRouter + retry SSE connection errors ([#3497](https://github.com/NousResearch/hermes-agent/pull/3497))
+- Prevent AsyncOpenAI/httpx cross-loop deadlock in gateway mode ([#2701](https://github.com/NousResearch/hermes-agent/pull/2701)) by @ctlst
+
+### Streaming & Reasoning
+- **Persist reasoning across gateway session turns** with new schema v6 columns (`reasoning`, `reasoning_details`, `codex_reasoning_items`) ([#2974](https://github.com/NousResearch/hermes-agent/pull/2974))
+- Detect and kill stale SSE connections ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Fix stale stream detector race causing spurious `RemoteProtocolError` ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Skip duplicate callback for `<think>`-extracted reasoning during streaming ([#3116](https://github.com/NousResearch/hermes-agent/pull/3116))
+- Preserve reasoning fields in `rewrite_transcript` ([#3311](https://github.com/NousResearch/hermes-agent/pull/3311))
+- Preserve Gemini thought signatures in streamed tool calls ([#2997](https://github.com/NousResearch/hermes-agent/pull/2997))
+- Ensure first delta is fired during reasoning updates ([untagged commit](https://github.com/NousResearch/hermes-agent))
+
+### Session & Memory
+- **Session search recent sessions mode** — Omit query to browse recent sessions with titles, previews, and timestamps ([#2533](https://github.com/NousResearch/hermes-agent/pull/2533))
+- **Session config surfacing** on `/new`, `/reset`, and auto-reset ([#3321](https://github.com/NousResearch/hermes-agent/pull/3321))
+- **Third-party session isolation** — `--source` flag for isolating sessions by origin ([#3255](https://github.com/NousResearch/hermes-agent/pull/3255))
+- Add `/resume` CLI handler, session log truncation guard, `reopen_session` API ([#3315](https://github.com/NousResearch/hermes-agent/pull/3315))
+- Clear compressor summary and turn counter on `/clear` and `/new` ([#3102](https://github.com/NousResearch/hermes-agent/pull/3102))
+- Surface silent SessionDB failures that cause session data loss ([#2999](https://github.com/NousResearch/hermes-agent/pull/2999))
+- Session search fallback preview on summarization failure ([#3478](https://github.com/NousResearch/hermes-agent/pull/3478))
+- Prevent stale memory overwrites by flush agent ([#2687](https://github.com/NousResearch/hermes-agent/pull/2687))
+
+### Context Compression
+- Replace dead `summary_target_tokens` with ratio-based scaling ([#2554](https://github.com/NousResearch/hermes-agent/pull/2554))
+- Expose `compression.target_ratio`, `protect_last_n`, and `threshold` in `DEFAULT_CONFIG` ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Restore sane defaults and cap summary at 12K tokens ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Preserve transcript on `/compress` and hygiene compression ([#3556](https://github.com/NousResearch/hermes-agent/pull/3556))
+- Update context pressure warnings and token estimates after compaction ([untagged commit](https://github.com/NousResearch/hermes-agent))
+
+### Architecture & Dependencies
+- **Remove mini-swe-agent dependency** — Inline Docker and Modal backends directly ([#2804](https://github.com/NousResearch/hermes-agent/pull/2804))
+- **Replace swe-rex with native Modal SDK** for Modal backend ([#3538](https://github.com/NousResearch/hermes-agent/pull/3538))
+- **Plugin lifecycle hooks** — `pre_llm_call`, `post_llm_call`, `on_session_start`, `on_session_end` now fire in the agent loop ([#3542](https://github.com/NousResearch/hermes-agent/pull/3542))
+- Fix plugin toolsets invisible in `hermes tools` and standalone processes ([#3457](https://github.com/NousResearch/hermes-agent/pull/3457))
+- Consolidate `get_hermes_home()` and `parse_reasoning_effort()` ([#3062](https://github.com/NousResearch/hermes-agent/pull/3062))
+- Remove unused Hermes-native PKCE OAuth flow ([#3107](https://github.com/NousResearch/hermes-agent/pull/3107))
+- Remove ~100 unused imports across 55 files ([#3016](https://github.com/NousResearch/hermes-agent/pull/3016))
+- Fix 154 f-strings, simplify getattr/URL patterns, remove dead code ([#3119](https://github.com/NousResearch/hermes-agent/pull/3119))
+
+---
+
+## 📱 Messaging Platforms (Gateway)
+
+### Telegram
+- **Private Chat Topics** — Project-based conversations with functional skill binding per topic, enabling isolated workflows within a single Telegram chat ([#3163](https://github.com/NousResearch/hermes-agent/pull/3163))
+- **Auto-discover fallback IPs via DNS-over-HTTPS** when `api.telegram.org` is unreachable ([#3376](https://github.com/NousResearch/hermes-agent/pull/3376))
+- **Configurable reply threading mode** ([#2907](https://github.com/NousResearch/hermes-agent/pull/2907))
+- Fall back to no `thread_id` on "Message thread not found" BadRequest ([#3390](https://github.com/NousResearch/hermes-agent/pull/3390))
+- Self-reschedule reconnect when `start_polling` fails after 502 ([#3268](https://github.com/NousResearch/hermes-agent/pull/3268))
+
+### Discord
+- Stop phantom typing indicator after agent turn completes ([#3003](https://github.com/NousResearch/hermes-agent/pull/3003))
+
+### Slack
+- Send tool call progress messages to correct Slack thread ([#3063](https://github.com/NousResearch/hermes-agent/pull/3063))
+- Scope progress thread fallback to Slack only ([#3488](https://github.com/NousResearch/hermes-agent/pull/3488))
+
+### WhatsApp
+- Download documents, audio, and video media from messages ([#2978](https://github.com/NousResearch/hermes-agent/pull/2978))
+
+### Matrix
+- Add missing Matrix entry in `PLATFORMS` dict ([#3473](https://github.com/NousResearch/hermes-agent/pull/3473))
+- Harden e2ee access-token handling ([#3562](https://github.com/NousResearch/hermes-agent/pull/3562))
+- Add backoff for `SyncError` in sync loop ([#3280](https://github.com/NousResearch/hermes-agent/pull/3280))
+
+### Signal
+- Track SSE keepalive comments as connection activity ([#3316](https://github.com/NousResearch/hermes-agent/pull/3316))
+
+### Email
+- Prevent unbounded growth of `_seen_uids` in EmailAdapter ([#3490](https://github.com/NousResearch/hermes-agent/pull/3490))
+
+### Gateway Core
+- **Config-gated `/verbose` command** for messaging platforms — toggle tool output verbosity from chat ([#3262](https://github.com/NousResearch/hermes-agent/pull/3262))
+- **Background review notifications** delivered to user chat ([#3293](https://github.com/NousResearch/hermes-agent/pull/3293))
+- **Retry transient send failures** and notify user on exhaustion ([#3288](https://github.com/NousResearch/hermes-agent/pull/3288))
+- Recover from hung agents — `/stop` hard-kills session lock ([#3104](https://github.com/NousResearch/hermes-agent/pull/3104))
+- Thread-safe `SessionStore` — protect `_entries` with `threading.Lock` ([#3052](https://github.com/NousResearch/hermes-agent/pull/3052))
+- Fix gateway token double-counting with cached agents — use absolute set instead of increment ([#3306](https://github.com/NousResearch/hermes-agent/pull/3306), [#3317](https://github.com/NousResearch/hermes-agent/pull/3317))
+- Fingerprint full auth token in agent cache signature ([#3247](https://github.com/NousResearch/hermes-agent/pull/3247))
+- Silence background agent terminal output ([#3297](https://github.com/NousResearch/hermes-agent/pull/3297))
+- Include per-platform `ALLOW_ALL` and `SIGNAL_GROUP` in startup allowlist check ([#3313](https://github.com/NousResearch/hermes-agent/pull/3313))
+- Include user-local bin paths in systemd unit PATH ([#3527](https://github.com/NousResearch/hermes-agent/pull/3527))
+- Track background task references in `GatewayRunner` ([#3254](https://github.com/NousResearch/hermes-agent/pull/3254))
+- Add request timeouts to HA, Email, Mattermost, SMS adapters ([#3258](https://github.com/NousResearch/hermes-agent/pull/3258))
+- Add media download retry to Mattermost, Slack, and base cache ([#3323](https://github.com/NousResearch/hermes-agent/pull/3323))
+- Detect virtualenv path instead of hardcoding `venv/` ([#2797](https://github.com/NousResearch/hermes-agent/pull/2797))
+- Use `TERMINAL_CWD` for context file discovery, not process cwd ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Stop loading hermes repo AGENTS.md into gateway sessions (~10k wasted tokens) ([#2891](https://github.com/NousResearch/hermes-agent/pull/2891))
+
+---
+
+## 🖥️ CLI & User Experience
+
+### Interactive CLI
+- **Configurable busy input mode** + fix `/queue` always working ([#3298](https://github.com/NousResearch/hermes-agent/pull/3298))
+- **Preserve user input on multiline paste** ([#3065](https://github.com/NousResearch/hermes-agent/pull/3065))
+- **Tool generation callback** — streaming "preparing terminal…" updates during tool argument generation ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Show tool progress for substantive tools, not just "preparing" ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Buffer reasoning preview chunks and fix duplicate display ([#3013](https://github.com/NousResearch/hermes-agent/pull/3013))
+- Prevent reasoning box from rendering 3x during tool-calling loops ([#3405](https://github.com/NousResearch/hermes-agent/pull/3405))
+- Eliminate "Event loop is closed" / "Press ENTER to continue" during idle — three-layer fix with `neuter_async_httpx_del()`, custom exception handler, and stale client cleanup ([#3398](https://github.com/NousResearch/hermes-agent/pull/3398))
+- Fix status bar shows 26K instead of 260K for token counts with trailing zeros ([#3024](https://github.com/NousResearch/hermes-agent/pull/3024))
+- Fix status bar duplicates and degrades during long sessions ([#3291](https://github.com/NousResearch/hermes-agent/pull/3291))
+- Refresh TUI before background task output to prevent status bar overlap ([#3048](https://github.com/NousResearch/hermes-agent/pull/3048))
+- Suppress KawaiiSpinner animation under `patch_stdout` ([#2994](https://github.com/NousResearch/hermes-agent/pull/2994))
+- Skip KawaiiSpinner when TUI handles tool progress ([#2973](https://github.com/NousResearch/hermes-agent/pull/2973))
+- Guard `isatty()` against closed streams via `_is_tty` property ([#3056](https://github.com/NousResearch/hermes-agent/pull/3056))
+- Ensure single closure of streaming boxes during tool generation ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Cap context pressure percentage at 100% in display ([#3480](https://github.com/NousResearch/hermes-agent/pull/3480))
+- Clean up HTML error messages in CLI display ([#3069](https://github.com/NousResearch/hermes-agent/pull/3069))
+- Show HTTP status code and 400 body in API error output ([#3096](https://github.com/NousResearch/hermes-agent/pull/3096))
+- Extract useful info from HTML error pages, dump debug on max retries ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Prevent TypeError on startup when `base_url` is None ([#3068](https://github.com/NousResearch/hermes-agent/pull/3068))
+- Prevent update crash in non-TTY environments ([#3094](https://github.com/NousResearch/hermes-agent/pull/3094))
+- Handle EOFError in sessions delete/prune confirmation prompts ([#3101](https://github.com/NousResearch/hermes-agent/pull/3101))
+- Catch KeyboardInterrupt during `flush_memories` on exit and in exit cleanup handlers ([#3025](https://github.com/NousResearch/hermes-agent/pull/3025), [#3257](https://github.com/NousResearch/hermes-agent/pull/3257))
+- Guard `.strip()` against None values from YAML config ([#3552](https://github.com/NousResearch/hermes-agent/pull/3552))
+- Guard `config.get()` against YAML null values to prevent AttributeError ([#3377](https://github.com/NousResearch/hermes-agent/pull/3377))
+- Store asyncio task references to prevent GC mid-execution ([#3267](https://github.com/NousResearch/hermes-agent/pull/3267))
+
+### Setup & Configuration
+- Use explicit key mapping for returning-user menu dispatch instead of positional index ([#3083](https://github.com/NousResearch/hermes-agent/pull/3083))
+- Use `sys.executable` for pip in update commands to fix PEP 668 ([#3099](https://github.com/NousResearch/hermes-agent/pull/3099))
+- Harden `hermes update` against diverged history, non-main branches, and gateway edge cases ([#3492](https://github.com/NousResearch/hermes-agent/pull/3492))
+- OpenClaw migration overwrites defaults and setup wizard skips imported sections — fixed ([#3282](https://github.com/NousResearch/hermes-agent/pull/3282))
+- Stop recursive AGENTS.md walk, load top-level only ([#3110](https://github.com/NousResearch/hermes-agent/pull/3110))
+- Add macOS Homebrew paths to browser and terminal PATH resolution ([#2713](https://github.com/NousResearch/hermes-agent/pull/2713))
+- YAML boolean handling for `tool_progress` config ([#3300](https://github.com/NousResearch/hermes-agent/pull/3300))
+- Reset default SOUL.md to baseline identity text ([#3159](https://github.com/NousResearch/hermes-agent/pull/3159))
+- Reject relative cwd paths for container terminal backends ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Add explicit `hermes-api-server` toolset for API server platform ([#3304](https://github.com/NousResearch/hermes-agent/pull/3304))
+- Reorder setup wizard providers — OpenRouter first ([untagged commit](https://github.com/NousResearch/hermes-agent))
+
+---
+
+## 🔧 Tool System
+
+### API Server
+- **Idempotency-Key support**, body size limit, and OpenAI error envelope ([#2903](https://github.com/NousResearch/hermes-agent/pull/2903))
+- Allow Idempotency-Key in CORS headers ([#3530](https://github.com/NousResearch/hermes-agent/pull/3530))
+- Cancel orphaned agent + true interrupt on SSE disconnect ([#3427](https://github.com/NousResearch/hermes-agent/pull/3427))
+- Fix streaming breaks when agent makes tool calls ([#2985](https://github.com/NousResearch/hermes-agent/pull/2985))
+
+### Terminal & File Operations
+- Handle addition-only hunks in V4A patch parser ([#3325](https://github.com/NousResearch/hermes-agent/pull/3325))
+- Exponential backoff for persistent shell polling ([#2996](https://github.com/NousResearch/hermes-agent/pull/2996))
+- Add timeout to subprocess calls in `context_references` ([#3469](https://github.com/NousResearch/hermes-agent/pull/3469))
+
+### Browser & Vision
+- Handle 402 insufficient credits error in vision tool ([#2802](https://github.com/NousResearch/hermes-agent/pull/2802))
+- Fix `browser_vision` ignores `auxiliary.vision.timeout` config ([#2901](https://github.com/NousResearch/hermes-agent/pull/2901))
+- Make browser command timeout configurable via config.yaml ([#2801](https://github.com/NousResearch/hermes-agent/pull/2801))
+
+### MCP
+- MCP toolset resolution for runtime and config ([#3252](https://github.com/NousResearch/hermes-agent/pull/3252))
+- Add MCP tool name collision protection ([#3077](https://github.com/NousResearch/hermes-agent/pull/3077))
+
+### Auxiliary LLM
+- Guard aux LLM calls against None content + reasoning fallback + retry ([#3449](https://github.com/NousResearch/hermes-agent/pull/3449))
+- Catch ImportError from `build_anthropic_client` in vision auto-detection ([#3312](https://github.com/NousResearch/hermes-agent/pull/3312))
+
+### Other Tools
+- Add request timeouts to `send_message_tool` HTTP calls ([#3162](https://github.com/NousResearch/hermes-agent/pull/3162)) by @memosr
+- Auto-repair `jobs.json` with invalid control characters ([#3537](https://github.com/NousResearch/hermes-agent/pull/3537))
+- Enable fine-grained tool streaming for Claude/OpenRouter ([#3497](https://github.com/NousResearch/hermes-agent/pull/3497))
+
+---
+
+## 🧩 Skills Ecosystem
+
+### Skills System
+- **Env var passthrough** for skills and user config — skills can declare environment variables to pass through ([#2807](https://github.com/NousResearch/hermes-agent/pull/2807))
+- Cache skills prompt with shared `skill_utils` module for faster TTFT ([#3421](https://github.com/NousResearch/hermes-agent/pull/3421))
+- Avoid redundant file re-read for skill conditions ([#2992](https://github.com/NousResearch/hermes-agent/pull/2992))
+- Use Git Trees API to prevent silent subdirectory loss during install ([#2995](https://github.com/NousResearch/hermes-agent/pull/2995))
+- Fix skills-sh install for deeply nested repo structures ([#2980](https://github.com/NousResearch/hermes-agent/pull/2980))
+- Handle null metadata in skill frontmatter ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Preserve trust for skills-sh identifiers + reduce resolution churn ([#3251](https://github.com/NousResearch/hermes-agent/pull/3251))
+- Agent-created skills were incorrectly treated as untrusted community content — fixed ([untagged commit](https://github.com/NousResearch/hermes-agent))
+
+### New Skills
+- **G0DM0D3 godmode jailbreaking skill** + docs ([#3157](https://github.com/NousResearch/hermes-agent/pull/3157))
+- **Docker management skill** added to optional-skills ([#3060](https://github.com/NousResearch/hermes-agent/pull/3060))
+- **OpenClaw migration v2** — 17 new modules, terminal recap for migrating from OpenClaw to Hermes ([#2906](https://github.com/NousResearch/hermes-agent/pull/2906))
+
+---
+
+## 🔒 Security & Reliability
+
+### Security Hardening
+- **SSRF protection** added to `browser_navigate` ([#3058](https://github.com/NousResearch/hermes-agent/pull/3058))
+- **SSRF protection** added to `vision_tools` and `web_tools` (hardened) ([#2679](https://github.com/NousResearch/hermes-agent/pull/2679))
+- **Restrict subagent toolsets** to parent's enabled set ([#3269](https://github.com/NousResearch/hermes-agent/pull/3269))
+- **Prevent zip-slip path traversal** in self-update ([#3250](https://github.com/NousResearch/hermes-agent/pull/3250))
+- **Prevent shell injection** in `_expand_path` via `~user` path suffix ([#2685](https://github.com/NousResearch/hermes-agent/pull/2685))
+- **Normalize input** before dangerous command detection ([#3260](https://github.com/NousResearch/hermes-agent/pull/3260))
+- Make tirith block verdicts approvable instead of hard-blocking ([#3428](https://github.com/NousResearch/hermes-agent/pull/3428))
+- Remove compromised `litellm`/`typer`/`platformdirs` from deps ([#2796](https://github.com/NousResearch/hermes-agent/pull/2796))
+- Pin all dependency version ranges ([#2810](https://github.com/NousResearch/hermes-agent/pull/2810))
+- Regenerate `uv.lock` with hashes, use lockfile in setup ([#2812](https://github.com/NousResearch/hermes-agent/pull/2812))
+- Bump dependencies to fix CVEs + regenerate `uv.lock` ([#3073](https://github.com/NousResearch/hermes-agent/pull/3073))
+- Supply chain audit CI workflow for PR scanning ([#2816](https://github.com/NousResearch/hermes-agent/pull/2816))
+
+### Reliability
+- **SQLite WAL write-lock contention** causing 15-20s TUI freeze — fixed ([#3385](https://github.com/NousResearch/hermes-agent/pull/3385))
+- **SQLite concurrency hardening** + session transcript integrity ([#3249](https://github.com/NousResearch/hermes-agent/pull/3249))
+- Prevent recurring cron job re-fire on gateway crash/restart loop ([#3396](https://github.com/NousResearch/hermes-agent/pull/3396))
+- Mark cron session as ended after job completes ([#2998](https://github.com/NousResearch/hermes-agent/pull/2998))
+
+---
+
+## ⚡ Performance
+
+- **TTFT startup optimizations** — salvaged easy-win startup improvements ([#3395](https://github.com/NousResearch/hermes-agent/pull/3395))
+- Cache skills prompt with shared `skill_utils` module ([#3421](https://github.com/NousResearch/hermes-agent/pull/3421))
+- Avoid redundant file re-read for skill conditions in prompt builder ([#2992](https://github.com/NousResearch/hermes-agent/pull/2992))
+
+---
+
+## 🐛 Notable Bug Fixes
+
+- Fix gateway token double-counting with cached agents ([#3306](https://github.com/NousResearch/hermes-agent/pull/3306), [#3317](https://github.com/NousResearch/hermes-agent/pull/3317))
+- Fix "Event loop is closed" / "Press ENTER to continue" during idle sessions ([#3398](https://github.com/NousResearch/hermes-agent/pull/3398))
+- Fix reasoning box rendering 3x during tool-calling loops ([#3405](https://github.com/NousResearch/hermes-agent/pull/3405))
+- Fix status bar shows 26K instead of 260K for token counts ([#3024](https://github.com/NousResearch/hermes-agent/pull/3024))
+- Fix `/queue` always working regardless of config ([#3298](https://github.com/NousResearch/hermes-agent/pull/3298))
+- Fix phantom Discord typing indicator after agent turn ([#3003](https://github.com/NousResearch/hermes-agent/pull/3003))
+- Fix Slack progress messages appearing in wrong thread ([#3063](https://github.com/NousResearch/hermes-agent/pull/3063))
+- Fix WhatsApp media downloads (documents, audio, video) ([#2978](https://github.com/NousResearch/hermes-agent/pull/2978))
+- Fix Telegram "Message thread not found" killing progress messages ([#3390](https://github.com/NousResearch/hermes-agent/pull/3390))
+- Fix OpenClaw migration overwriting defaults ([#3282](https://github.com/NousResearch/hermes-agent/pull/3282))
+- Fix returning-user setup menu dispatching wrong section ([#3083](https://github.com/NousResearch/hermes-agent/pull/3083))
+- Fix `hermes update` PEP 668 "externally-managed-environment" error ([#3099](https://github.com/NousResearch/hermes-agent/pull/3099))
+- Fix subagents hitting `max_iterations` prematurely via shared budget ([#3004](https://github.com/NousResearch/hermes-agent/pull/3004))
+- Fix YAML boolean handling for `tool_progress` config ([#3300](https://github.com/NousResearch/hermes-agent/pull/3300))
+- Fix `config.get()` crashes on YAML null values ([#3377](https://github.com/NousResearch/hermes-agent/pull/3377))
+- Fix `.strip()` crash on None values from YAML config ([#3552](https://github.com/NousResearch/hermes-agent/pull/3552))
+- Fix hung agents on gateway — `/stop` now hard-kills session lock ([#3104](https://github.com/NousResearch/hermes-agent/pull/3104))
+- Fix `_custom` provider silently remapped to `openrouter` ([#2792](https://github.com/NousResearch/hermes-agent/pull/2792))
+- Fix Matrix missing from `PLATFORMS` dict ([#3473](https://github.com/NousResearch/hermes-agent/pull/3473))
+- Fix Email adapter unbounded `_seen_uids` growth ([#3490](https://github.com/NousResearch/hermes-agent/pull/3490))
+
+---
+
+## 🧪 Testing
+
+- Pin `agent-client-protocol` < 0.9 to handle breaking upstream release ([#3320](https://github.com/NousResearch/hermes-agent/pull/3320))
+- Catch anthropic ImportError in vision auto-detection tests ([#3312](https://github.com/NousResearch/hermes-agent/pull/3312))
+- Update retry-exhaust test for new graceful return behavior ([#3320](https://github.com/NousResearch/hermes-agent/pull/3320))
+- Add regression tests for null metadata frontmatter ([untagged commit](https://github.com/NousResearch/hermes-agent))
+
+---
+
+## 📚 Documentation
+
+- Update all docs for `/model` command overhaul and custom provider support ([#2800](https://github.com/NousResearch/hermes-agent/pull/2800))
+- Fix stale and incorrect documentation across 18 files ([#2805](https://github.com/NousResearch/hermes-agent/pull/2805))
+- Document 9 previously undocumented features ([#2814](https://github.com/NousResearch/hermes-agent/pull/2814))
+- Add missing skills, CLI commands, and messaging env vars to docs ([#2809](https://github.com/NousResearch/hermes-agent/pull/2809))
+- Fix api-server response storage documentation — SQLite, not in-memory ([#2819](https://github.com/NousResearch/hermes-agent/pull/2819))
+- Quote pip install extras to fix zsh glob errors ([#2815](https://github.com/NousResearch/hermes-agent/pull/2815))
+- Unify hooks documentation — add plugin hooks to hooks page, add `session:end` event ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Clarify two-mode behavior in `session_search` schema description ([untagged commit](https://github.com/NousResearch/hermes-agent))
+- Fix Discord Public Bot setting for Discord-provided invite link ([#3519](https://github.com/NousResearch/hermes-agent/pull/3519)) by @mehmoodosman
+- Revise v0.4.0 changelog — fix feature attribution, reorder sections ([untagged commit](https://github.com/NousResearch/hermes-agent))
+
+---
+
+## 👥 Contributors
+
+### Core
+- **@teknium1** — 157 PRs covering the full scope of this release
+
+### Community Contributors
+- **@alt-glitch** (Siddharth Balyan) — 2 PRs: Nix flake with uv2nix build, NixOS module, and persistent container mode ([#20](https://github.com/NousResearch/hermes-agent/pull/20)); auto-generated config keys and suffix PATHs for Nix builds ([#3061](https://github.com/NousResearch/hermes-agent/pull/3061), [#3274](https://github.com/NousResearch/hermes-agent/pull/3274))
+- **@ctlst** — 1 PR: Prevent AsyncOpenAI/httpx cross-loop deadlock in gateway mode ([#2701](https://github.com/NousResearch/hermes-agent/pull/2701))
+- **@memosr** (memosr.eth) — 1 PR: Add request timeouts to `send_message_tool` HTTP calls ([#3162](https://github.com/NousResearch/hermes-agent/pull/3162))
+- **@mehmoodosman** (Osman Mehmood) — 1 PR: Fix Discord docs for Public Bot setting ([#3519](https://github.com/NousResearch/hermes-agent/pull/3519))
+
+### All Contributors
+@alt-glitch, @ctlst, @mehmoodosman, @memosr, @teknium1
+
+---
+
+**Full Changelog**: [v2026.3.23...v2026.3.28](https://github.com/NousResearch/hermes-agent/compare/v2026.3.23...v2026.3.28)

RELEASE_v0.6.0.md

@@ -0,0 +1,249 @@
+# Hermes Agent v0.6.0 (v2026.3.30)
+
+**Release Date:** March 30, 2026
+
+> The multi-instance release — Profiles for running isolated agent instances, MCP server mode, Docker container, fallback provider chains, two new messaging platforms (Feishu/Lark and WeCom), Telegram webhook mode, Slack multi-workspace OAuth, 95 PRs and 16 resolved issues in 2 days.
+
+---
+
+## ✨ Highlights
+
+- **Profiles — Multi-Instance Hermes** — Run multiple isolated Hermes instances from the same installation. Each profile gets its own config, memory, sessions, skills, and gateway service. Create with `hermes profile create`, switch with `hermes -p <name>`, export/import for sharing. Full token-lock isolation prevents two profiles from using the same bot credential. ([#3681](https://github.com/NousResearch/hermes-agent/pull/3681))
+
+- **MCP Server Mode** — Expose Hermes conversations and sessions to any MCP-compatible client (Claude Desktop, Cursor, VS Code, etc.) via `hermes mcp serve`. Browse conversations, read messages, search across sessions, and manage attachments — all through the Model Context Protocol. Supports both stdio and Streamable HTTP transports. ([#3795](https://github.com/NousResearch/hermes-agent/pull/3795))
+
+- **Docker Container** — Official Dockerfile for running Hermes Agent in a container. Supports both CLI and gateway modes with volume-mounted config. ([#3668](https://github.com/NousResearch/hermes-agent/pull/3668), closes [#850](https://github.com/NousResearch/hermes-agent/issues/850))
+
+- **Ordered Fallback Provider Chain** — Configure multiple inference providers with automatic failover. When your primary provider returns errors or is unreachable, Hermes automatically tries the next provider in the chain. Configure via `fallback_providers` in config.yaml. ([#3813](https://github.com/NousResearch/hermes-agent/pull/3813), closes [#1734](https://github.com/NousResearch/hermes-agent/issues/1734))
+
+- **Feishu/Lark Platform Support** — Full gateway adapter for Feishu (飞书) and Lark with event subscriptions, message cards, group chat, image/file attachments, and interactive card callbacks. ([#3799](https://github.com/NousResearch/hermes-agent/pull/3799), [#3817](https://github.com/NousResearch/hermes-agent/pull/3817), closes [#1788](https://github.com/NousResearch/hermes-agent/issues/1788))
+
+- **WeCom (Enterprise WeChat) Platform Support** — New gateway adapter for WeCom (企业微信) with text/image/voice messages, group chats, and callback verification. ([#3847](https://github.com/NousResearch/hermes-agent/pull/3847))
+
+- **Slack Multi-Workspace OAuth** — Connect a single Hermes gateway to multiple Slack workspaces via OAuth token file. Each workspace gets its own bot token, resolved dynamically per incoming event. ([#3903](https://github.com/NousResearch/hermes-agent/pull/3903))
+
+- **Telegram Webhook Mode & Group Controls** — Run the Telegram adapter in webhook mode as an alternative to polling — faster response times and better for production deployments behind a reverse proxy. New group mention gating controls when the bot responds: always, only when @mentioned, or via regex triggers. ([#3880](https://github.com/NousResearch/hermes-agent/pull/3880), [#3870](https://github.com/NousResearch/hermes-agent/pull/3870))
+
+- **Exa Search Backend** — Add Exa as an alternative web search and content extraction backend alongside Firecrawl and DuckDuckGo. Set `EXA_API_KEY` and configure as preferred backend. ([#3648](https://github.com/NousResearch/hermes-agent/pull/3648))
+
+- **Skills & Credentials on Remote Backends** — Mount skill directories and credential files into Modal and Docker containers, so remote terminal sessions have access to the same skills and secrets as local execution. ([#3890](https://github.com/NousResearch/hermes-agent/pull/3890), [#3671](https://github.com/NousResearch/hermes-agent/pull/3671), closes [#3665](https://github.com/NousResearch/hermes-agent/issues/3665), [#3433](https://github.com/NousResearch/hermes-agent/issues/3433))
+
+---
+
+## 🏗️ Core Agent & Architecture
+
+### Provider & Model Support
+- **Ordered fallback provider chain** — automatic failover across multiple configured providers ([#3813](https://github.com/NousResearch/hermes-agent/pull/3813))
+- **Fix api_mode on provider switch** — switching providers via `hermes model` now correctly clears stale `api_mode` instead of hardcoding `chat_completions`, fixing 404s for providers with Anthropic-compatible endpoints ([#3726](https://github.com/NousResearch/hermes-agent/pull/3726), [#3857](https://github.com/NousResearch/hermes-agent/pull/3857), closes [#3685](https://github.com/NousResearch/hermes-agent/issues/3685))
+- **Stop silent OpenRouter fallback** — when no provider is configured, Hermes now raises a clear error instead of silently routing to OpenRouter ([#3807](https://github.com/NousResearch/hermes-agent/pull/3807), [#3862](https://github.com/NousResearch/hermes-agent/pull/3862))
+- **Gemini 3.1 preview models** — added to OpenRouter and Nous Portal catalogs ([#3803](https://github.com/NousResearch/hermes-agent/pull/3803), closes [#3753](https://github.com/NousResearch/hermes-agent/issues/3753))
+- **Gemini direct API context length** — full context length resolution for direct Google AI endpoints ([#3876](https://github.com/NousResearch/hermes-agent/pull/3876))
+- **gpt-5.4-mini** added to Codex fallback catalog ([#3855](https://github.com/NousResearch/hermes-agent/pull/3855))
+- **Curated model lists preferred** over live API probe when the probe returns fewer models ([#3856](https://github.com/NousResearch/hermes-agent/pull/3856), [#3867](https://github.com/NousResearch/hermes-agent/pull/3867))
+- **User-friendly 429 rate limit messages** with Retry-After countdown ([#3809](https://github.com/NousResearch/hermes-agent/pull/3809))
+- **Auxiliary client placeholder key** for local servers without auth requirements ([#3842](https://github.com/NousResearch/hermes-agent/pull/3842))
+- **INFO-level logging** for auxiliary provider resolution ([#3866](https://github.com/NousResearch/hermes-agent/pull/3866))
+
+### Agent Loop & Conversation
+- **Subagent status reporting** — reports `completed` status when summary exists instead of generic failure ([#3829](https://github.com/NousResearch/hermes-agent/pull/3829))
+- **Session log file updated during compression** — prevents stale file references after context compression ([#3835](https://github.com/NousResearch/hermes-agent/pull/3835))
+- **Omit empty tools param** — sends no `tools` parameter when empty instead of `None`, fixing compatibility with strict providers ([#3820](https://github.com/NousResearch/hermes-agent/pull/3820))
+
+### Profiles & Multi-Instance
+- **Profiles system** — `hermes profile create/list/switch/delete/export/import/rename`. Each profile gets isolated HERMES_HOME, gateway service, CLI wrapper. Token locks prevent credential collisions. Tab completion for profile names. ([#3681](https://github.com/NousResearch/hermes-agent/pull/3681))
+- **Profile-aware display paths** — all user-facing `~/.hermes` paths replaced with `display_hermes_home()` to show the correct profile directory ([#3623](https://github.com/NousResearch/hermes-agent/pull/3623))
+- **Lazy display_hermes_home imports** — prevents `ImportError` during `hermes update` when modules cache stale bytecode ([#3776](https://github.com/NousResearch/hermes-agent/pull/3776))
+- **HERMES_HOME for protected paths** — `.env` write-deny path now respects HERMES_HOME instead of hardcoded `~/.hermes` ([#3840](https://github.com/NousResearch/hermes-agent/pull/3840))
+
+---
+
+## 📱 Messaging Platforms (Gateway)
+
+### New Platforms
+- **Feishu/Lark** — Full adapter with event subscriptions, message cards, group chat, image/file attachments, interactive card callbacks ([#3799](https://github.com/NousResearch/hermes-agent/pull/3799), [#3817](https://github.com/NousResearch/hermes-agent/pull/3817))
+- **WeCom (Enterprise WeChat)** — Text/image/voice messages, group chats, callback verification ([#3847](https://github.com/NousResearch/hermes-agent/pull/3847))
+
+### Telegram
+- **Webhook mode** — run as webhook endpoint instead of polling for production deployments ([#3880](https://github.com/NousResearch/hermes-agent/pull/3880))
+- **Group mention gating & regex triggers** — configurable bot response behavior in groups: always, @mention-only, or regex-matched ([#3870](https://github.com/NousResearch/hermes-agent/pull/3870))
+- **Gracefully handle deleted reply targets** — no more crashes when the message being replied to was deleted ([#3858](https://github.com/NousResearch/hermes-agent/pull/3858), closes [#3229](https://github.com/NousResearch/hermes-agent/issues/3229))
+
+### Discord
+- **Message processing reactions** — adds a reaction emoji while processing and removes it when done, giving visual feedback in channels ([#3871](https://github.com/NousResearch/hermes-agent/pull/3871))
+- **DISCORD_IGNORE_NO_MENTION** — skip messages that @mention other users/bots but not Hermes ([#3640](https://github.com/NousResearch/hermes-agent/pull/3640))
+- **Clean up deferred "thinking..."** — properly removes the "thinking..." indicator after slash commands complete ([#3674](https://github.com/NousResearch/hermes-agent/pull/3674), closes [#3595](https://github.com/NousResearch/hermes-agent/issues/3595))
+
+### Slack
+- **Multi-workspace OAuth** — connect to multiple Slack workspaces from a single gateway via OAuth token file ([#3903](https://github.com/NousResearch/hermes-agent/pull/3903))
+
+### WhatsApp
+- **Persistent aiohttp session** — reuse HTTP sessions across requests instead of creating new ones per message ([#3818](https://github.com/NousResearch/hermes-agent/pull/3818))
+- **LID↔phone alias resolution** — correctly match Linked ID and phone number formats in allowlists ([#3830](https://github.com/NousResearch/hermes-agent/pull/3830))
+- **Skip reply prefix in bot mode** — cleaner message formatting when running as a WhatsApp bot ([#3931](https://github.com/NousResearch/hermes-agent/pull/3931))
+
+### Matrix
+- **Native voice messages via MSC3245** — send voice messages as proper Matrix voice events instead of file attachments ([#3877](https://github.com/NousResearch/hermes-agent/pull/3877))
+
+### Mattermost
+- **Configurable mention behavior** — respond to messages without requiring @mention ([#3664](https://github.com/NousResearch/hermes-agent/pull/3664))
+
+### Signal
+- **URL-encode phone numbers** and correct attachment RPC parameter — fixes delivery failures with certain phone number formats ([#3670](https://github.com/NousResearch/hermes-agent/pull/3670)) — @kshitijk4poor
+
+### Email
+- **Close SMTP/IMAP connections on failure** — prevents connection leaks during error scenarios ([#3804](https://github.com/NousResearch/hermes-agent/pull/3804))
+
+### Gateway Core
+- **Atomic config writes** — use atomic file writes for config.yaml to prevent data loss during crashes ([#3800](https://github.com/NousResearch/hermes-agent/pull/3800))
+- **Home channel env overrides** — apply environment variable overrides for home channels consistently ([#3796](https://github.com/NousResearch/hermes-agent/pull/3796), [#3808](https://github.com/NousResearch/hermes-agent/pull/3808))
+- **Replace print() with logger** — BasePlatformAdapter now uses proper logging instead of print statements ([#3669](https://github.com/NousResearch/hermes-agent/pull/3669))
+- **Cron delivery labels** — resolve human-friendly delivery labels via channel directory ([#3860](https://github.com/NousResearch/hermes-agent/pull/3860), closes [#1945](https://github.com/NousResearch/hermes-agent/issues/1945))
+- **Cron [SILENT] tightening** — prevent agents from prefixing reports with [SILENT] to suppress delivery ([#3901](https://github.com/NousResearch/hermes-agent/pull/3901))
+- **Background task media delivery** and vision download timeout fixes ([#3919](https://github.com/NousResearch/hermes-agent/pull/3919))
+- **Boot-md hook** — example built-in hook to run a BOOT.md file on gateway startup ([#3733](https://github.com/NousResearch/hermes-agent/pull/3733))
+
+---
+
+## 🖥️ CLI & User Experience
+
+### Interactive CLI
+- **Configurable tool preview length** — show full file paths by default instead of truncating at 40 chars ([#3841](https://github.com/NousResearch/hermes-agent/pull/3841))
+- **Tool token context display** — `hermes tools` checklist now shows estimated token cost per toolset ([#3805](https://github.com/NousResearch/hermes-agent/pull/3805))
+- **/bg spinner TUI fix** — route background task spinner through the TUI widget to prevent status bar collision ([#3643](https://github.com/NousResearch/hermes-agent/pull/3643))
+- **Prevent status bar wrapping** into duplicate rows ([#3883](https://github.com/NousResearch/hermes-agent/pull/3883)) — @kshitijk4poor
+- **Handle closed stdout ValueError** in safe print paths — fixes crashes when stdout is closed during gateway thread shutdown ([#3843](https://github.com/NousResearch/hermes-agent/pull/3843), closes [#3534](https://github.com/NousResearch/hermes-agent/issues/3534))
+- **Remove input() from /tools disable** — eliminates freeze in terminal when disabling tools ([#3918](https://github.com/NousResearch/hermes-agent/pull/3918))
+- **TTY guard for interactive CLI commands** — prevent CPU spin when launched without a terminal ([#3933](https://github.com/NousResearch/hermes-agent/pull/3933))
+- **Argparse entrypoint** — use argparse in the top-level launcher for cleaner error handling ([#3874](https://github.com/NousResearch/hermes-agent/pull/3874))
+- **Lazy-initialized tools show yellow** in banner instead of red, reducing false alarm about "missing" tools ([#3822](https://github.com/NousResearch/hermes-agent/pull/3822))
+- **Honcho tools shown in banner** when configured ([#3810](https://github.com/NousResearch/hermes-agent/pull/3810))
+
+### Setup & Configuration
+- **Auto-install matrix-nio** during `hermes setup` when Matrix is selected ([#3802](https://github.com/NousResearch/hermes-agent/pull/3802), [#3873](https://github.com/NousResearch/hermes-agent/pull/3873))
+- **Session export stdout support** — export sessions to stdout with `-` for piping ([#3641](https://github.com/NousResearch/hermes-agent/pull/3641), closes [#3609](https://github.com/NousResearch/hermes-agent/issues/3609))
+- **Configurable approval timeouts** — set how long dangerous command approval prompts wait before auto-denying ([#3886](https://github.com/NousResearch/hermes-agent/pull/3886), closes [#3765](https://github.com/NousResearch/hermes-agent/issues/3765))
+- **Clear __pycache__ during update** — prevents stale bytecode ImportError after `hermes update` ([#3819](https://github.com/NousResearch/hermes-agent/pull/3819))
+
+---
+
+## 🔧 Tool System
+
+### MCP
+- **MCP Server Mode** — `hermes mcp serve` exposes conversations, sessions, and attachments to MCP clients via stdio or Streamable HTTP ([#3795](https://github.com/NousResearch/hermes-agent/pull/3795))
+- **Dynamic tool discovery** — respond to `notifications/tools/list_changed` events to pick up new tools from MCP servers without reconnecting ([#3812](https://github.com/NousResearch/hermes-agent/pull/3812))
+- **Non-deprecated HTTP transport** — switched from `sse_client` to `streamable_http_client` ([#3646](https://github.com/NousResearch/hermes-agent/pull/3646))
+
+### Web Tools
+- **Exa search backend** — alternative to Firecrawl and DuckDuckGo for web search and extraction ([#3648](https://github.com/NousResearch/hermes-agent/pull/3648))
+
+### Browser
+- **Guard against None LLM responses** in browser snapshot and vision tools ([#3642](https://github.com/NousResearch/hermes-agent/pull/3642))
+
+### Terminal & Remote Backends
+- **Mount skill directories** into Modal and Docker containers ([#3890](https://github.com/NousResearch/hermes-agent/pull/3890))
+- **Mount credential files** into remote backends with mtime+size caching ([#3671](https://github.com/NousResearch/hermes-agent/pull/3671))
+- **Preserve partial output** when commands time out instead of losing everything ([#3868](https://github.com/NousResearch/hermes-agent/pull/3868))
+- **Stop marking persisted env vars as missing** on remote backends ([#3650](https://github.com/NousResearch/hermes-agent/pull/3650))
+
+### Audio
+- **.aac format support** in transcription tool ([#3865](https://github.com/NousResearch/hermes-agent/pull/3865), closes [#1963](https://github.com/NousResearch/hermes-agent/issues/1963))
+- **Audio download retry** — retry logic for `cache_audio_from_url` matching the existing image download pattern ([#3401](https://github.com/NousResearch/hermes-agent/pull/3401)) — @binhnt92
+
+### Vision
+- **Reject non-image files** and enforce website-only policy for vision analysis ([#3845](https://github.com/NousResearch/hermes-agent/pull/3845))
+
+### Tool Schema
+- **Ensure name field** always present in tool definitions, fixing `KeyError: 'name'` crashes ([#3811](https://github.com/NousResearch/hermes-agent/pull/3811), closes [#3729](https://github.com/NousResearch/hermes-agent/issues/3729))
+
+### ACP (Editor Integration)
+- **Complete session management surface** for VS Code/Zed/JetBrains clients — proper task lifecycle, cancel support, session persistence ([#3675](https://github.com/NousResearch/hermes-agent/pull/3675))
+
+---
+
+## 🧩 Skills & Plugins
+
+### Skills System
+- **External skill directories** — configure additional skill directories via `skills.external_dirs` in config.yaml ([#3678](https://github.com/NousResearch/hermes-agent/pull/3678))
+- **Category path traversal blocked** — prevents `../` attacks in skill category names ([#3844](https://github.com/NousResearch/hermes-agent/pull/3844))
+- **parallel-cli moved to optional-skills** — reduces default skill footprint ([#3673](https://github.com/NousResearch/hermes-agent/pull/3673)) — @kshitijk4poor
+
+### New Skills
+- **memento-flashcards** — spaced repetition flashcard system ([#3827](https://github.com/NousResearch/hermes-agent/pull/3827))
+- **songwriting-and-ai-music** — songwriting craft and AI music generation prompts ([#3834](https://github.com/NousResearch/hermes-agent/pull/3834))
+- **SiYuan Note** — integration with SiYuan note-taking app ([#3742](https://github.com/NousResearch/hermes-agent/pull/3742))
+- **Scrapling** — web scraping skill using Scrapling library ([#3742](https://github.com/NousResearch/hermes-agent/pull/3742))
+- **one-three-one-rule** — communication framework skill ([#3797](https://github.com/NousResearch/hermes-agent/pull/3797))
+
+### Plugin System
+- **Plugin enable/disable commands** — `hermes plugins enable/disable <name>` for managing plugin state without removing them ([#3747](https://github.com/NousResearch/hermes-agent/pull/3747))
+- **Plugin message injection** — plugins can now inject messages into the conversation stream on behalf of the user via `ctx.inject_message()` ([#3778](https://github.com/NousResearch/hermes-agent/pull/3778)) — @winglian
+- **Honcho self-hosted support** — allow local Honcho instances without requiring an API key ([#3644](https://github.com/NousResearch/hermes-agent/pull/3644))
+
+---
+
+## 🔒 Security & Reliability
+
+### Security Hardening
+- **Hardened dangerous command detection** — expanded pattern matching for risky shell commands and added file tool path guards for sensitive locations (`/etc/`, `/boot/`, docker.sock) ([#3872](https://github.com/NousResearch/hermes-agent/pull/3872))
+- **Sensitive path write checks** in approval system — catch writes to system config files through file tools, not just terminal ([#3859](https://github.com/NousResearch/hermes-agent/pull/3859))
+- **Secret redaction expansion** — now covers ElevenLabs, Tavily, and Exa API keys ([#3920](https://github.com/NousResearch/hermes-agent/pull/3920))
+- **Vision file rejection** — reject non-image files passed to vision analysis to prevent information disclosure ([#3845](https://github.com/NousResearch/hermes-agent/pull/3845))
+- **Category path traversal blocking** — prevent directory traversal in skill category names ([#3844](https://github.com/NousResearch/hermes-agent/pull/3844))
+
+### Reliability
+- **Atomic config.yaml writes** — prevent data loss during gateway crashes ([#3800](https://github.com/NousResearch/hermes-agent/pull/3800))
+- **Clear __pycache__ on update** — prevent stale bytecode from causing ImportError after updates ([#3819](https://github.com/NousResearch/hermes-agent/pull/3819))
+- **Lazy imports for update safety** — prevent ImportError chains during `hermes update` when modules reference new functions ([#3776](https://github.com/NousResearch/hermes-agent/pull/3776))
+- **Restore terminalbench2 from patch corruption** — recovered file damaged by patch tool's secret redaction ([#3801](https://github.com/NousResearch/hermes-agent/pull/3801))
+- **Terminal timeout preserves partial output** — no more lost command output on timeout ([#3868](https://github.com/NousResearch/hermes-agent/pull/3868))
+
+---
+
+## 🐛 Notable Bug Fixes
+
+- **OpenClaw migration model config overwrite** — migration no longer overwrites model config dict with a string ([#3924](https://github.com/NousResearch/hermes-agent/pull/3924)) — @0xbyt4
+- **OpenClaw migration expanded** — covers full data footprint including sessions, cron, memory ([#3869](https://github.com/NousResearch/hermes-agent/pull/3869))
+- **Telegram deleted reply targets** — gracefully handle replies to deleted messages instead of crashing ([#3858](https://github.com/NousResearch/hermes-agent/pull/3858))
+- **Discord "thinking..." persistence** — properly cleans up deferred response indicators ([#3674](https://github.com/NousResearch/hermes-agent/pull/3674))
+- **WhatsApp LID↔phone aliases** — fixes allowlist matching failures with Linked ID format ([#3830](https://github.com/NousResearch/hermes-agent/pull/3830))
+- **Signal URL-encoded phone numbers** — fixes delivery failures with certain formats ([#3670](https://github.com/NousResearch/hermes-agent/pull/3670))
+- **Email connection leaks** — properly close SMTP/IMAP connections on error ([#3804](https://github.com/NousResearch/hermes-agent/pull/3804))
+- **_safe_print ValueError** — no more gateway thread crashes on closed stdout ([#3843](https://github.com/NousResearch/hermes-agent/pull/3843))
+- **Tool schema KeyError 'name'** — ensure name field always present in tool definitions ([#3811](https://github.com/NousResearch/hermes-agent/pull/3811))
+- **api_mode stale on provider switch** — correctly clear when switching providers via `hermes model` ([#3857](https://github.com/NousResearch/hermes-agent/pull/3857))
+
+---
+
+## 🧪 Testing
+
+- Resolved 10+ CI failures across hooks, tiktoken, plugins, and skill tests ([#3848](https://github.com/NousResearch/hermes-agent/pull/3848), [#3721](https://github.com/NousResearch/hermes-agent/pull/3721), [#3936](https://github.com/NousResearch/hermes-agent/pull/3936))
+
+---
+
+## 📚 Documentation
+
+- **Comprehensive OpenClaw migration guide** — step-by-step guide for migrating from OpenClaw/Claw3D to Hermes Agent ([#3864](https://github.com/NousResearch/hermes-agent/pull/3864), [#3900](https://github.com/NousResearch/hermes-agent/pull/3900))
+- **Credential file passthrough docs** — document how to forward credential files and env vars to remote backends ([#3677](https://github.com/NousResearch/hermes-agent/pull/3677))
+- **DuckDuckGo requirements clarified** — note runtime dependency on duckduckgo-search package ([#3680](https://github.com/NousResearch/hermes-agent/pull/3680))
+- **Skills catalog updated** — added red-teaming category and optional skills listing ([#3745](https://github.com/NousResearch/hermes-agent/pull/3745))
+- **Feishu docs MDX fix** — escape angle-bracket URLs that break Docusaurus build ([#3902](https://github.com/NousResearch/hermes-agent/pull/3902))
+
+---
+
+## 👥 Contributors
+
+### Core
+- **@teknium1** — 90 PRs across all subsystems
+
+### Community Contributors
+- **@kshitijk4poor** — 3 PRs: Signal phone number fix ([#3670](https://github.com/NousResearch/hermes-agent/pull/3670)), parallel-cli to optional-skills ([#3673](https://github.com/NousResearch/hermes-agent/pull/3673)), status bar wrapping fix ([#3883](https://github.com/NousResearch/hermes-agent/pull/3883))
+- **@winglian** — 1 PR: Plugin message injection interface ([#3778](https://github.com/NousResearch/hermes-agent/pull/3778))
+- **@binhnt92** — 1 PR: Audio download retry logic ([#3401](https://github.com/NousResearch/hermes-agent/pull/3401))
+- **@0xbyt4** — 1 PR: OpenClaw migration model config fix ([#3924](https://github.com/NousResearch/hermes-agent/pull/3924))
+
+### Issues Resolved from Community
+@Material-Scientist ([#850](https://github.com/NousResearch/hermes-agent/issues/850)), @hanxu98121 ([#1734](https://github.com/NousResearch/hermes-agent/issues/1734)), @penwyp ([#1788](https://github.com/NousResearch/hermes-agent/issues/1788)), @dan-and ([#1945](https://github.com/NousResearch/hermes-agent/issues/1945)), @AdrianScott ([#1963](https://github.com/NousResearch/hermes-agent/issues/1963)), @clawdbot47 ([#3229](https://github.com/NousResearch/hermes-agent/issues/3229)), @alanfwilliams ([#3404](https://github.com/NousResearch/hermes-agent/issues/3404)), @kentimsit ([#3433](https://github.com/NousResearch/hermes-agent/issues/3433)), @hayka-pacha ([#3534](https://github.com/NousResearch/hermes-agent/issues/3534)), @primmer ([#3595](https://github.com/NousResearch/hermes-agent/issues/3595)), @dagelf ([#3609](https://github.com/NousResearch/hermes-agent/issues/3609)), @HenkDz ([#3685](https://github.com/NousResearch/hermes-agent/issues/3685)), @tmdgusya ([#3729](https://github.com/NousResearch/hermes-agent/issues/3729)), @TypQxQ ([#3753](https://github.com/NousResearch/hermes-agent/issues/3753)), @acsezen ([#3765](https://github.com/NousResearch/hermes-agent/issues/3765))
+
+---
+
+**Full Changelog**: [v2026.3.28...v2026.3.30](https://github.com/NousResearch/hermes-agent/compare/v2026.3.28...v2026.3.30)

RELEASE_v0.7.0.md

@@ -0,0 +1,290 @@
+# Hermes Agent v0.7.0 (v2026.4.3)
+
+**Release Date:** April 3, 2026
+
+> The resilience release — pluggable memory providers, credential pool rotation, Camofox anti-detection browser, inline diff previews, gateway hardening across race conditions and approval routing, and deep security fixes across 168 PRs and 46 resolved issues.
+
+---
+
+## ✨ Highlights
+
+- **Pluggable Memory Provider Interface** — Memory is now an extensible plugin system. Third-party memory backends (Honcho, vector stores, custom DBs) implement a simple provider ABC and register via the plugin system. Built-in memory is the default provider. Honcho integration restored to full parity as the reference plugin with profile-scoped host/peer resolution. ([#4623](https://github.com/NousResearch/hermes-agent/pull/4623), [#4616](https://github.com/NousResearch/hermes-agent/pull/4616), [#4355](https://github.com/NousResearch/hermes-agent/pull/4355))
+
+- **Same-Provider Credential Pools** — Configure multiple API keys for the same provider with automatic rotation. Thread-safe `least_used` strategy distributes load across keys, and 401 failures trigger automatic rotation to the next credential. Set up via the setup wizard or `credential_pool` config. ([#4188](https://github.com/NousResearch/hermes-agent/pull/4188), [#4300](https://github.com/NousResearch/hermes-agent/pull/4300), [#4361](https://github.com/NousResearch/hermes-agent/pull/4361))
+
+- **Camofox Anti-Detection Browser Backend** — New local browser backend using Camoufox for stealth browsing. Persistent sessions with VNC URL discovery for visual debugging, configurable SSRF bypass for local backends, auto-install via `hermes tools`. ([#4008](https://github.com/NousResearch/hermes-agent/pull/4008), [#4419](https://github.com/NousResearch/hermes-agent/pull/4419), [#4292](https://github.com/NousResearch/hermes-agent/pull/4292))
+
+- **Inline Diff Previews** — File write and patch operations now show inline diffs in the tool activity feed, giving you visual confirmation of what changed before the agent moves on. ([#4411](https://github.com/NousResearch/hermes-agent/pull/4411), [#4423](https://github.com/NousResearch/hermes-agent/pull/4423))
+
+- **API Server Session Continuity & Tool Streaming** — The API server (Open WebUI integration) now streams tool progress events in real-time and supports `X-Hermes-Session-Id` headers for persistent sessions across requests. Sessions persist to the shared SessionDB. ([#4092](https://github.com/NousResearch/hermes-agent/pull/4092), [#4478](https://github.com/NousResearch/hermes-agent/pull/4478), [#4802](https://github.com/NousResearch/hermes-agent/pull/4802))
+
+- **ACP: Client-Provided MCP Servers** — Editor integrations (VS Code, Zed, JetBrains) can now register their own MCP servers, which Hermes picks up as additional agent tools. Your editor's MCP ecosystem flows directly into the agent. ([#4705](https://github.com/NousResearch/hermes-agent/pull/4705))
+
+- **Gateway Hardening** — Major stability pass across race conditions, photo media delivery, flood control, stuck sessions, approval routing, and compression death spirals. The gateway is substantially more reliable in production. ([#4727](https://github.com/NousResearch/hermes-agent/pull/4727), [#4750](https://github.com/NousResearch/hermes-agent/pull/4750), [#4798](https://github.com/NousResearch/hermes-agent/pull/4798), [#4557](https://github.com/NousResearch/hermes-agent/pull/4557))
+
+- **Security: Secret Exfiltration Blocking** — Browser URLs and LLM responses are now scanned for secret patterns, blocking exfiltration attempts via URL encoding, base64, or prompt injection. Credential directory protections expanded to `.docker`, `.azure`, `.config/gh`. Execute_code sandbox output is redacted. ([#4483](https://github.com/NousResearch/hermes-agent/pull/4483), [#4360](https://github.com/NousResearch/hermes-agent/pull/4360), [#4305](https://github.com/NousResearch/hermes-agent/pull/4305), [#4327](https://github.com/NousResearch/hermes-agent/pull/4327))
+
+---
+
+## 🏗️ Core Agent & Architecture
+
+### Provider & Model Support
+- **Same-provider credential pools** — configure multiple API keys with automatic `least_used` rotation and 401 failover ([#4188](https://github.com/NousResearch/hermes-agent/pull/4188), [#4300](https://github.com/NousResearch/hermes-agent/pull/4300))
+- **Credential pool preserved through smart routing** — pool state survives fallback provider switches and defers eager fallback on 429 ([#4361](https://github.com/NousResearch/hermes-agent/pull/4361))
+- **Per-turn primary runtime restoration** — after fallback provider use, the agent automatically restores the primary provider on the next turn with transport recovery ([#4624](https://github.com/NousResearch/hermes-agent/pull/4624))
+- **`developer` role for GPT-5 and Codex models** — uses OpenAI's recommended system message role for newer models ([#4498](https://github.com/NousResearch/hermes-agent/pull/4498))
+- **Google model operational guidance** — Gemini and Gemma models get provider-specific prompting guidance ([#4641](https://github.com/NousResearch/hermes-agent/pull/4641))
+- **Anthropic long-context tier 429 handling** — automatically reduces context to 200k when hitting tier limits ([#4747](https://github.com/NousResearch/hermes-agent/pull/4747))
+- **URL-based auth for third-party Anthropic endpoints** + CI test fixes ([#4148](https://github.com/NousResearch/hermes-agent/pull/4148))
+- **Bearer auth for MiniMax Anthropic endpoints** ([#4028](https://github.com/NousResearch/hermes-agent/pull/4028))
+- **Fireworks context length detection** ([#4158](https://github.com/NousResearch/hermes-agent/pull/4158))
+- **Standard DashScope international endpoint** for Alibaba provider ([#4133](https://github.com/NousResearch/hermes-agent/pull/4133), closes [#3912](https://github.com/NousResearch/hermes-agent/issues/3912))
+- **Custom providers context_length** honored in hygiene compression ([#4085](https://github.com/NousResearch/hermes-agent/pull/4085))
+- **Non-sk-ant keys** treated as regular API keys, not OAuth tokens ([#4093](https://github.com/NousResearch/hermes-agent/pull/4093))
+- **Claude-sonnet-4.6** added to OpenRouter and Nous model lists ([#4157](https://github.com/NousResearch/hermes-agent/pull/4157))
+- **Qwen 3.6 Plus Preview** added to model lists ([#4376](https://github.com/NousResearch/hermes-agent/pull/4376))
+- **MiniMax M2.7** added to hermes model picker and OpenCode ([#4208](https://github.com/NousResearch/hermes-agent/pull/4208))
+- **Auto-detect models from server probe** in custom endpoint setup ([#4218](https://github.com/NousResearch/hermes-agent/pull/4218))
+- **Config.yaml single source of truth** for endpoint URLs — no more env var vs config.yaml conflicts ([#4165](https://github.com/NousResearch/hermes-agent/pull/4165))
+- **Setup wizard no longer overwrites** custom endpoint config ([#4180](https://github.com/NousResearch/hermes-agent/pull/4180), closes [#4172](https://github.com/NousResearch/hermes-agent/issues/4172))
+- **Unified setup wizard provider selection** with `hermes model` — single code path for both flows ([#4200](https://github.com/NousResearch/hermes-agent/pull/4200))
+- **Root-level provider config** no longer overrides `model.provider` ([#4329](https://github.com/NousResearch/hermes-agent/pull/4329))
+- **Rate-limit pairing rejection messages** to prevent spam ([#4081](https://github.com/NousResearch/hermes-agent/pull/4081))
+
+### Agent Loop & Conversation
+- **Preserve Anthropic thinking block signatures** across tool-use turns ([#4626](https://github.com/NousResearch/hermes-agent/pull/4626))
+- **Classify think-only empty responses** before retrying — prevents infinite retry loops on models that produce thinking blocks without content ([#4645](https://github.com/NousResearch/hermes-agent/pull/4645))
+- **Prevent compression death spiral** from API disconnects — stops the loop where compression triggers, fails, compresses again ([#4750](https://github.com/NousResearch/hermes-agent/pull/4750), closes [#2153](https://github.com/NousResearch/hermes-agent/issues/2153))
+- **Persist compressed context** to gateway session after mid-run compression ([#4095](https://github.com/NousResearch/hermes-agent/pull/4095))
+- **Context-exceeded error messages** now include actionable guidance ([#4155](https://github.com/NousResearch/hermes-agent/pull/4155), closes [#4061](https://github.com/NousResearch/hermes-agent/issues/4061))
+- **Strip orphaned think/reasoning tags** from user-facing responses ([#4311](https://github.com/NousResearch/hermes-agent/pull/4311), closes [#4285](https://github.com/NousResearch/hermes-agent/issues/4285))
+- **Harden Codex responses preflight** and stream error handling ([#4313](https://github.com/NousResearch/hermes-agent/pull/4313))
+- **Deterministic call_id fallbacks** instead of random UUIDs for prompt cache consistency ([#3991](https://github.com/NousResearch/hermes-agent/pull/3991))
+- **Context pressure warning spam** prevented after compression ([#4012](https://github.com/NousResearch/hermes-agent/pull/4012))
+- **AsyncOpenAI created lazily** in trajectory compressor to avoid closed event loop errors ([#4013](https://github.com/NousResearch/hermes-agent/pull/4013))
+
+### Memory & Sessions
+- **Pluggable memory provider interface** — ABC-based plugin system for custom memory backends with profile isolation ([#4623](https://github.com/NousResearch/hermes-agent/pull/4623))
+- **Honcho full integration parity** restored as reference memory provider plugin ([#4355](https://github.com/NousResearch/hermes-agent/pull/4355)) — @erosika
+- **Honcho profile-scoped** host and peer resolution ([#4616](https://github.com/NousResearch/hermes-agent/pull/4616))
+- **Memory flush state persisted** to prevent redundant re-flushes on gateway restart ([#4481](https://github.com/NousResearch/hermes-agent/pull/4481))
+- **Memory provider tools** routed through sequential execution path ([#4803](https://github.com/NousResearch/hermes-agent/pull/4803))
+- **Honcho config** written to instance-local path for profile isolation ([#4037](https://github.com/NousResearch/hermes-agent/pull/4037))
+- **API server sessions** persist to shared SessionDB ([#4802](https://github.com/NousResearch/hermes-agent/pull/4802))
+- **Token usage persisted** for non-CLI sessions ([#4627](https://github.com/NousResearch/hermes-agent/pull/4627))
+- **Quote dotted terms in FTS5 queries** — fixes session search for terms containing dots ([#4549](https://github.com/NousResearch/hermes-agent/pull/4549))
+
+---
+
+## 📱 Messaging Platforms (Gateway)
+
+### Gateway Core
+- **Race condition fixes** — photo media loss, flood control, stuck sessions, and STT config issues resolved in one hardening pass ([#4727](https://github.com/NousResearch/hermes-agent/pull/4727))
+- **Approval routing through running-agent guard** — `/approve` and `/deny` now route correctly when the agent is blocked waiting for approval instead of being swallowed as interrupts ([#4798](https://github.com/NousResearch/hermes-agent/pull/4798), [#4557](https://github.com/NousResearch/hermes-agent/pull/4557), closes [#4542](https://github.com/NousResearch/hermes-agent/issues/4542))
+- **Resume agent after /approve** — tool result is no longer lost when executing blocked commands ([#4418](https://github.com/NousResearch/hermes-agent/pull/4418))
+- **DM thread sessions seeded** with parent transcript to preserve context ([#4559](https://github.com/NousResearch/hermes-agent/pull/4559))
+- **Skill-aware slash commands** — gateway dynamically registers installed skills as slash commands with paginated `/commands` list and Telegram 100-command cap ([#3934](https://github.com/NousResearch/hermes-agent/pull/3934), [#4005](https://github.com/NousResearch/hermes-agent/pull/4005), [#4006](https://github.com/NousResearch/hermes-agent/pull/4006), [#4010](https://github.com/NousResearch/hermes-agent/pull/4010), [#4023](https://github.com/NousResearch/hermes-agent/pull/4023))
+- **Per-platform disabled skills** respected in Telegram menu and gateway dispatch ([#4799](https://github.com/NousResearch/hermes-agent/pull/4799))
+- **Remove user-facing compression warnings** — cleaner message flow ([#4139](https://github.com/NousResearch/hermes-agent/pull/4139))
+- **`-v/-q` flags wired to stderr logging** for gateway service ([#4474](https://github.com/NousResearch/hermes-agent/pull/4474))
+- **HERMES_HOME remapped** to target user in system service unit ([#4456](https://github.com/NousResearch/hermes-agent/pull/4456))
+- **Honor default for invalid bool-like config values** ([#4029](https://github.com/NousResearch/hermes-agent/pull/4029))
+- **setsid instead of systemd-run** for `/update` command to avoid systemd permission issues ([#4104](https://github.com/NousResearch/hermes-agent/pull/4104), closes [#4017](https://github.com/NousResearch/hermes-agent/issues/4017))
+- **'Initializing agent...'** shown on first message for better UX ([#4086](https://github.com/NousResearch/hermes-agent/pull/4086))
+- **Allow running gateway service as root** for LXC/container environments ([#4732](https://github.com/NousResearch/hermes-agent/pull/4732))
+
+### Telegram
+- **32-char limit on command names** with collision avoidance ([#4211](https://github.com/NousResearch/hermes-agent/pull/4211))
+- **Priority order enforced** in menu — core > plugins > skills ([#4023](https://github.com/NousResearch/hermes-agent/pull/4023))
+- **Capped at 50 commands** — API rejects above ~60 ([#4006](https://github.com/NousResearch/hermes-agent/pull/4006))
+- **Skip empty/whitespace text** to prevent 400 errors ([#4388](https://github.com/NousResearch/hermes-agent/pull/4388))
+- **E2E gateway tests** added ([#4497](https://github.com/NousResearch/hermes-agent/pull/4497)) — @pefontana
+
+### Discord
+- **Button-based approval UI** — register `/approve` and `/deny` slash commands with interactive button prompts ([#4800](https://github.com/NousResearch/hermes-agent/pull/4800))
+- **Configurable reactions** — `discord.reactions` config option to disable message processing reactions ([#4199](https://github.com/NousResearch/hermes-agent/pull/4199))
+- **Skip reactions and auto-threading** for unauthorized users ([#4387](https://github.com/NousResearch/hermes-agent/pull/4387))
+
+### Slack
+- **Reply in thread** — `slack.reply_in_thread` config option for threaded responses ([#4643](https://github.com/NousResearch/hermes-agent/pull/4643), closes [#2662](https://github.com/NousResearch/hermes-agent/issues/2662))
+
+### WhatsApp
+- **Enforce require_mention in group chats** ([#4730](https://github.com/NousResearch/hermes-agent/pull/4730))
+
+### Webhook
+- **Platform support fixes** — skip home channel prompt, disable tool progress for webhook adapters ([#4660](https://github.com/NousResearch/hermes-agent/pull/4660))
+
+### Matrix
+- **E2EE decryption hardening** — request missing keys, auto-trust devices, retry buffered events ([#4083](https://github.com/NousResearch/hermes-agent/pull/4083))
+
+---
+
+## 🖥️ CLI & User Experience
+
+### New Slash Commands
+- **`/yolo`** — toggle dangerous command approvals on/off for the session ([#3990](https://github.com/NousResearch/hermes-agent/pull/3990))
+- **`/btw`** — ephemeral side questions that don't affect the main conversation context ([#4161](https://github.com/NousResearch/hermes-agent/pull/4161))
+- **`/profile`** — show active profile info without leaving the chat session ([#4027](https://github.com/NousResearch/hermes-agent/pull/4027))
+
+### Interactive CLI
+- **Inline diff previews** for write and patch operations in the tool activity feed ([#4411](https://github.com/NousResearch/hermes-agent/pull/4411), [#4423](https://github.com/NousResearch/hermes-agent/pull/4423))
+- **TUI pinned to bottom** on startup — no more large blank spaces between response and input ([#4412](https://github.com/NousResearch/hermes-agent/pull/4412), [#4359](https://github.com/NousResearch/hermes-agent/pull/4359), closes [#4398](https://github.com/NousResearch/hermes-agent/issues/4398), [#4421](https://github.com/NousResearch/hermes-agent/issues/4421))
+- **`/history` and `/resume`** now surface recent sessions directly instead of requiring search ([#4728](https://github.com/NousResearch/hermes-agent/pull/4728))
+- **Cache tokens shown** in `/insights` overview so total adds up ([#4428](https://github.com/NousResearch/hermes-agent/pull/4428))
+- **`--max-turns` CLI flag** for `hermes chat` to limit agent iterations ([#4314](https://github.com/NousResearch/hermes-agent/pull/4314))
+- **Detect dragged file paths** instead of treating them as slash commands ([#4533](https://github.com/NousResearch/hermes-agent/pull/4533)) — @rolme
+- **Allow empty strings and falsy values** in `config set` ([#4310](https://github.com/NousResearch/hermes-agent/pull/4310), closes [#4277](https://github.com/NousResearch/hermes-agent/issues/4277))
+- **Voice mode in WSL** when PulseAudio bridge is configured ([#4317](https://github.com/NousResearch/hermes-agent/pull/4317))
+- **Respect `NO_COLOR` env var** and `TERM=dumb` for accessibility ([#4079](https://github.com/NousResearch/hermes-agent/pull/4079), closes [#4066](https://github.com/NousResearch/hermes-agent/issues/4066)) — @SHL0MS
+- **Correct shell reload instruction** for macOS/zsh users ([#4025](https://github.com/NousResearch/hermes-agent/pull/4025))
+- **Zero exit code** on successful quiet mode queries ([#4613](https://github.com/NousResearch/hermes-agent/pull/4613), closes [#4601](https://github.com/NousResearch/hermes-agent/issues/4601)) — @devorun
+- **on_session_end hook fires** on interrupted exits ([#4159](https://github.com/NousResearch/hermes-agent/pull/4159))
+- **Profile list display** reads `model.default` key correctly ([#4160](https://github.com/NousResearch/hermes-agent/pull/4160))
+- **Browser and TTS** shown in reconfigure menu ([#4041](https://github.com/NousResearch/hermes-agent/pull/4041))
+- **Web backend priority** detection simplified ([#4036](https://github.com/NousResearch/hermes-agent/pull/4036))
+
+### Setup & Configuration
+- **Allowed_users preserved** during setup and quiet unconfigured provider warnings ([#4551](https://github.com/NousResearch/hermes-agent/pull/4551)) — @kshitijk4poor
+- **Save API key to model config** for custom endpoints ([#4202](https://github.com/NousResearch/hermes-agent/pull/4202), closes [#4182](https://github.com/NousResearch/hermes-agent/issues/4182))
+- **Claude Code credentials gated** behind explicit Hermes config in wizard trigger ([#4210](https://github.com/NousResearch/hermes-agent/pull/4210))
+- **Atomic writes in save_config_value** to prevent config loss on interrupt ([#4298](https://github.com/NousResearch/hermes-agent/pull/4298), [#4320](https://github.com/NousResearch/hermes-agent/pull/4320))
+- **Scopes field written** to Claude Code credentials on token refresh ([#4126](https://github.com/NousResearch/hermes-agent/pull/4126))
+
+### Update System
+- **Fork detection and upstream sync** in `hermes update` ([#4744](https://github.com/NousResearch/hermes-agent/pull/4744))
+- **Preserve working optional extras** when one extra fails during update ([#4550](https://github.com/NousResearch/hermes-agent/pull/4550))
+- **Handle conflicted git index** during hermes update ([#4735](https://github.com/NousResearch/hermes-agent/pull/4735))
+- **Avoid launchd restart race** on macOS ([#4736](https://github.com/NousResearch/hermes-agent/pull/4736))
+- **Missing subprocess.run() timeouts** added to doctor and status commands ([#4009](https://github.com/NousResearch/hermes-agent/pull/4009))
+
+---
+
+## 🔧 Tool System
+
+### Browser
+- **Camofox anti-detection browser backend** — local stealth browsing with auto-install via `hermes tools` ([#4008](https://github.com/NousResearch/hermes-agent/pull/4008))
+- **Persistent Camofox sessions** with VNC URL discovery for visual debugging ([#4419](https://github.com/NousResearch/hermes-agent/pull/4419))
+- **Skip SSRF check for local backends** (Camofox, headless Chromium) ([#4292](https://github.com/NousResearch/hermes-agent/pull/4292))
+- **Configurable SSRF check** via `browser.allow_private_urls` ([#4198](https://github.com/NousResearch/hermes-agent/pull/4198)) — @nils010485
+- **CAMOFOX_PORT=9377** added to Docker commands ([#4340](https://github.com/NousResearch/hermes-agent/pull/4340))
+
+### File Operations
+- **Inline diff previews** on write and patch actions ([#4411](https://github.com/NousResearch/hermes-agent/pull/4411), [#4423](https://github.com/NousResearch/hermes-agent/pull/4423))
+- **Stale file detection** on write and patch — warns when file was modified externally since last read ([#4345](https://github.com/NousResearch/hermes-agent/pull/4345))
+- **Staleness timestamp refreshed** after writes ([#4390](https://github.com/NousResearch/hermes-agent/pull/4390))
+- **Size guard, dedup, and device blocking** on read_file ([#4315](https://github.com/NousResearch/hermes-agent/pull/4315))
+
+### MCP
+- **Stability fix pack** — reload timeout, shutdown cleanup, event loop handler, OAuth non-blocking ([#4757](https://github.com/NousResearch/hermes-agent/pull/4757), closes [#4462](https://github.com/NousResearch/hermes-agent/issues/4462), [#2537](https://github.com/NousResearch/hermes-agent/issues/2537))
+
+### ACP (Editor Integration)
+- **Client-provided MCP servers** registered as agent tools — editors pass their MCP servers to Hermes ([#4705](https://github.com/NousResearch/hermes-agent/pull/4705))
+
+### Skills System
+- **Size limits for agent writes** and **fuzzy matching for skill patch** — prevents oversized skill writes and improves edit reliability ([#4414](https://github.com/NousResearch/hermes-agent/pull/4414))
+- **Validate hub bundle paths** before install — blocks path traversal in skill bundles ([#3986](https://github.com/NousResearch/hermes-agent/pull/3986))
+- **Unified hermes-agent and hermes-agent-setup** into single skill ([#4332](https://github.com/NousResearch/hermes-agent/pull/4332))
+- **Skill metadata type check** in extract_skill_conditions ([#4479](https://github.com/NousResearch/hermes-agent/pull/4479))
+
+### New/Updated Skills
+- **research-paper-writing** — full end-to-end research pipeline (replaced ml-paper-writing) ([#4654](https://github.com/NousResearch/hermes-agent/pull/4654)) — @SHL0MS
+- **ascii-video** — text readability techniques and external layout oracle ([#4054](https://github.com/NousResearch/hermes-agent/pull/4054)) — @SHL0MS
+- **youtube-transcript** updated for youtube-transcript-api v1.x ([#4455](https://github.com/NousResearch/hermes-agent/pull/4455)) — @el-analista
+- **Skills browse and search page** added to documentation site ([#4500](https://github.com/NousResearch/hermes-agent/pull/4500)) — @IAvecilla
+
+---
+
+## 🔒 Security & Reliability
+
+### Security Hardening
+- **Block secret exfiltration** via browser URLs and LLM responses — scans for secret patterns in URL encoding, base64, and prompt injection vectors ([#4483](https://github.com/NousResearch/hermes-agent/pull/4483))
+- **Redact secrets from execute_code sandbox output** ([#4360](https://github.com/NousResearch/hermes-agent/pull/4360))
+- **Protect `.docker`, `.azure`, `.config/gh` credential directories** from read/write via file tools and terminal ([#4305](https://github.com/NousResearch/hermes-agent/pull/4305), [#4327](https://github.com/NousResearch/hermes-agent/pull/4327)) — @memosr
+- **GitHub OAuth token patterns** added to redaction + snapshot redact flag ([#4295](https://github.com/NousResearch/hermes-agent/pull/4295))
+- **Reject private and loopback IPs** in Telegram DoH fallback ([#4129](https://github.com/NousResearch/hermes-agent/pull/4129))
+- **Reject path traversal** in credential file registration ([#4316](https://github.com/NousResearch/hermes-agent/pull/4316))
+- **Validate tar archive member paths** on profile import — blocks zip-slip attacks ([#4318](https://github.com/NousResearch/hermes-agent/pull/4318))
+- **Exclude auth.json and .env** from profile exports ([#4475](https://github.com/NousResearch/hermes-agent/pull/4475))
+
+### Reliability
+- **Prevent compression death spiral** from API disconnects ([#4750](https://github.com/NousResearch/hermes-agent/pull/4750), closes [#2153](https://github.com/NousResearch/hermes-agent/issues/2153))
+- **Handle `is_closed` as method** in OpenAI SDK — prevents false positive client closure detection ([#4416](https://github.com/NousResearch/hermes-agent/pull/4416), closes [#4377](https://github.com/NousResearch/hermes-agent/issues/4377))
+- **Exclude matrix from [all] extras** — python-olm is upstream-broken, prevents install failures ([#4615](https://github.com/NousResearch/hermes-agent/pull/4615), closes [#4178](https://github.com/NousResearch/hermes-agent/issues/4178))
+- **OpenCode model routing** repaired ([#4508](https://github.com/NousResearch/hermes-agent/pull/4508))
+- **Docker container image** optimized ([#4034](https://github.com/NousResearch/hermes-agent/pull/4034)) — @bcross
+
+### Windows & Cross-Platform
+- **Voice mode in WSL** with PulseAudio bridge ([#4317](https://github.com/NousResearch/hermes-agent/pull/4317))
+- **Homebrew packaging** preparation ([#4099](https://github.com/NousResearch/hermes-agent/pull/4099))
+- **CI fork conditionals** to prevent workflow failures on forks ([#4107](https://github.com/NousResearch/hermes-agent/pull/4107))
+
+---
+
+## 🐛 Notable Bug Fixes
+
+- **Gateway approval blocked agent thread** — approval now blocks the agent thread like CLI does, preventing tool result loss ([#4557](https://github.com/NousResearch/hermes-agent/pull/4557), closes [#4542](https://github.com/NousResearch/hermes-agent/issues/4542))
+- **Compression death spiral** from API disconnects — detected and halted instead of looping ([#4750](https://github.com/NousResearch/hermes-agent/pull/4750), closes [#2153](https://github.com/NousResearch/hermes-agent/issues/2153))
+- **Anthropic thinking blocks lost** across tool-use turns ([#4626](https://github.com/NousResearch/hermes-agent/pull/4626))
+- **Profile model config ignored** with `-p` flag — model.model now promoted to model.default correctly ([#4160](https://github.com/NousResearch/hermes-agent/pull/4160), closes [#4486](https://github.com/NousResearch/hermes-agent/issues/4486))
+- **CLI blank space** between response and input area ([#4412](https://github.com/NousResearch/hermes-agent/pull/4412), [#4359](https://github.com/NousResearch/hermes-agent/pull/4359), closes [#4398](https://github.com/NousResearch/hermes-agent/issues/4398))
+- **Dragged file paths** treated as slash commands instead of file references ([#4533](https://github.com/NousResearch/hermes-agent/pull/4533)) — @rolme
+- **Orphaned `</think>` tags** leaking into user-facing responses ([#4311](https://github.com/NousResearch/hermes-agent/pull/4311), closes [#4285](https://github.com/NousResearch/hermes-agent/issues/4285))
+- **OpenAI SDK `is_closed`** is a method not property — false positive client closure ([#4416](https://github.com/NousResearch/hermes-agent/pull/4416), closes [#4377](https://github.com/NousResearch/hermes-agent/issues/4377))
+- **MCP OAuth server** could block Hermes startup instead of degrading gracefully ([#4757](https://github.com/NousResearch/hermes-agent/pull/4757), closes [#4462](https://github.com/NousResearch/hermes-agent/issues/4462))
+- **MCP event loop closed** on shutdown with HTTP servers ([#4757](https://github.com/NousResearch/hermes-agent/pull/4757), closes [#2537](https://github.com/NousResearch/hermes-agent/issues/2537))
+- **Alibaba provider** hardcoded to wrong endpoint ([#4133](https://github.com/NousResearch/hermes-agent/pull/4133), closes [#3912](https://github.com/NousResearch/hermes-agent/issues/3912))
+- **Slack reply_in_thread** missing config option ([#4643](https://github.com/NousResearch/hermes-agent/pull/4643), closes [#2662](https://github.com/NousResearch/hermes-agent/issues/2662))
+- **Quiet mode exit code** — successful `-q` queries no longer exit nonzero ([#4613](https://github.com/NousResearch/hermes-agent/pull/4613), closes [#4601](https://github.com/NousResearch/hermes-agent/issues/4601))
+- **Mobile sidebar** shows only close button due to backdrop-filter issue in docs site ([#4207](https://github.com/NousResearch/hermes-agent/pull/4207)) — @xsmyile
+- **Config restore reverted** by stale-branch squash merge — `_config_version` fixed ([#4440](https://github.com/NousResearch/hermes-agent/pull/4440))
+
+---
+
+## 🧪 Testing
+
+- **Telegram gateway E2E tests** — full integration test suite for the Telegram adapter ([#4497](https://github.com/NousResearch/hermes-agent/pull/4497)) — @pefontana
+- **11 real test failures fixed** plus sys.modules cascade poisoner resolved ([#4570](https://github.com/NousResearch/hermes-agent/pull/4570))
+- **7 CI failures resolved** across hooks, plugins, and skill tests ([#3936](https://github.com/NousResearch/hermes-agent/pull/3936))
+- **Codex 401 refresh tests** updated for CI compatibility ([#4166](https://github.com/NousResearch/hermes-agent/pull/4166))
+- **Stale OPENAI_BASE_URL test** fixed ([#4217](https://github.com/NousResearch/hermes-agent/pull/4217))
+
+---
+
+## 📚 Documentation
+
+- **Comprehensive documentation audit** — 9 HIGH and 20+ MEDIUM gaps fixed across 21 files ([#4087](https://github.com/NousResearch/hermes-agent/pull/4087))
+- **Site navigation restructured** — features and platforms promoted to top-level ([#4116](https://github.com/NousResearch/hermes-agent/pull/4116))
+- **Tool progress streaming** documented for API server and Open WebUI ([#4138](https://github.com/NousResearch/hermes-agent/pull/4138))
+- **Telegram webhook mode** documentation ([#4089](https://github.com/NousResearch/hermes-agent/pull/4089))
+- **Local LLM provider guides** — comprehensive setup guides with context length warnings ([#4294](https://github.com/NousResearch/hermes-agent/pull/4294))
+- **WhatsApp allowlist behavior** clarified with `WHATSAPP_ALLOW_ALL_USERS` documentation ([#4293](https://github.com/NousResearch/hermes-agent/pull/4293))
+- **Slack configuration options** — new config section in Slack docs ([#4644](https://github.com/NousResearch/hermes-agent/pull/4644))
+- **Terminal backends section** expanded + docs build fixes ([#4016](https://github.com/NousResearch/hermes-agent/pull/4016))
+- **Adding-providers guide** updated for unified setup flow ([#4201](https://github.com/NousResearch/hermes-agent/pull/4201))
+- **ACP Zed config** fixed ([#4743](https://github.com/NousResearch/hermes-agent/pull/4743))
+- **Community FAQ** entries for common workflows and troubleshooting ([#4797](https://github.com/NousResearch/hermes-agent/pull/4797))
+- **Skills browse and search page** on docs site ([#4500](https://github.com/NousResearch/hermes-agent/pull/4500)) — @IAvecilla
+
+---
+
+## 👥 Contributors
+
+### Core
+- **@teknium1** — 135 commits across all subsystems
+
+### Top Community Contributors
+- **@kshitijk4poor** — 13 commits: preserve allowed_users during setup ([#4551](https://github.com/NousResearch/hermes-agent/pull/4551)), and various fixes
+- **@erosika** — 12 commits: Honcho full integration parity restored as memory provider plugin ([#4355](https://github.com/NousResearch/hermes-agent/pull/4355))
+- **@pefontana** — 9 commits: Telegram gateway E2E test suite ([#4497](https://github.com/NousResearch/hermes-agent/pull/4497))
+- **@bcross** — 5 commits: Docker container image optimization ([#4034](https://github.com/NousResearch/hermes-agent/pull/4034))
+- **@SHL0MS** — 4 commits: NO_COLOR/TERM=dumb support ([#4079](https://github.com/NousResearch/hermes-agent/pull/4079)), ascii-video skill updates ([#4054](https://github.com/NousResearch/hermes-agent/pull/4054)), research-paper-writing skill ([#4654](https://github.com/NousResearch/hermes-agent/pull/4654))
+
+### All Contributors
+@0xbyt4, @arasovic, @Bartok9, @bcross, @binhnt92, @camden-lowrance, @curtitoo, @Dakota, @Dave Tist, @Dean Kerr, @devorun, @dieutx, @Dilee, @el-analista, @erosika, @Gutslabs, @IAvecilla, @Jack, @Johannnnn506, @kshitijk4poor, @Laura Batalha, @Leegenux, @Lume, @MacroAnarchy, @maymuneth, @memosr, @NexVeridian, @Nick, @nils010485, @pefontana, @Penov, @rolme, @SHL0MS, @txchen, @xsmyile
+
+### Issues Resolved from Community
+@acsezen ([#2537](https://github.com/NousResearch/hermes-agent/issues/2537)), @arasovic ([#4285](https://github.com/NousResearch/hermes-agent/issues/4285)), @camden-lowrance ([#4462](https://github.com/NousResearch/hermes-agent/issues/4462)), @devorun ([#4601](https://github.com/NousResearch/hermes-agent/issues/4601)), @eloklam ([#4486](https://github.com/NousResearch/hermes-agent/issues/4486)), @HenkDz ([#3719](https://github.com/NousResearch/hermes-agent/issues/3719)), @hypotyposis ([#2153](https://github.com/NousResearch/hermes-agent/issues/2153)), @kazamak ([#4178](https://github.com/NousResearch/hermes-agent/issues/4178)), @lstep ([#4366](https://github.com/NousResearch/hermes-agent/issues/4366)), @Mark-Lok ([#4542](https://github.com/NousResearch/hermes-agent/issues/4542)), @NoJster ([#4421](https://github.com/NousResearch/hermes-agent/issues/4421)), @patp ([#2662](https://github.com/NousResearch/hermes-agent/issues/2662)), @pr0n ([#4601](https://github.com/NousResearch/hermes-agent/issues/4601)), @saulmc ([#4377](https://github.com/NousResearch/hermes-agent/issues/4377)), @SHL0MS ([#4060](https://github.com/NousResearch/hermes-agent/issues/4060), [#4061](https://github.com/NousResearch/hermes-agent/issues/4061), [#4066](https://github.com/NousResearch/hermes-agent/issues/4066), [#4172](https://github.com/NousResearch/hermes-agent/issues/4172), [#4277](https://github.com/NousResearch/hermes-agent/issues/4277)), @Z-Mackintosh ([#4398](https://github.com/NousResearch/hermes-agent/issues/4398))
+
+---
+
+**Full Changelog**: [v2026.3.30...v2026.4.3](https://github.com/NousResearch/hermes-agent/compare/v2026.3.30...v2026.4.3)

acp_adapter/entry.py

@@ -18,6 +18,7 @@ import logging
 import os
 import sys
 from pathlib import Path
+from hermes_constants import get_hermes_home
 
 
 def _setup_logging() -> None:
@@ -42,19 +43,16 @@ def _setup_logging() -> None:
 
 def _load_env() -> None:
     """Load .env from HERMES_HOME (default ``~/.hermes``)."""
-    from dotenv import load_dotenv
+    from hermes_cli.env_loader import load_hermes_dotenv
 
-    hermes_home = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
-    env_file = hermes_home / ".env"
-    if env_file.exists():
-        try:
-            load_dotenv(dotenv_path=env_file, encoding="utf-8")
-        except UnicodeDecodeError:
-            load_dotenv(dotenv_path=env_file, encoding="latin-1")
-        logging.getLogger(__name__).info("Loaded env from %s", env_file)
+    hermes_home = get_hermes_home()
+    loaded = load_hermes_dotenv(hermes_home=hermes_home)
+    if loaded:
+        for env_file in loaded:
+            logging.getLogger(__name__).info("Loaded env from %s", env_file)
     else:
         logging.getLogger(__name__).info(
-            "No .env found at %s, using system env", env_file
+            "No .env found at %s, using system env", hermes_home / ".env"
         )
 
 
@@ -76,7 +74,7 @@ def main() -> None:
 
     agent = HermesACPAgent()
     try:
-        asyncio.run(acp.run_agent(agent))
+        asyncio.run(acp.run_agent(agent, use_unstable_protocol=True))
     except KeyboardInterrupt:
         logger.info("Shutting down (KeyboardInterrupt)")
     except Exception:

acp_adapter/events.py

@@ -10,7 +10,7 @@ thread while the event loop lives on the main thread).
 import asyncio
 import json
 import logging
-from collections import defaultdict, deque
+from collections import deque
 from typing import Any, Callable, Deque, Dict
 
 import acp
@@ -54,14 +54,18 @@ def make_tool_progress_cb(
 
     Signature expected by AIAgent::
 
-        tool_progress_callback(name: str, preview: str, args: dict)
+        tool_progress_callback(event_type: str, name: str, preview: str, args: dict, **kwargs)
 
-    Emits ``ToolCallStart`` for each tool invocation and tracks IDs in a FIFO
+    Emits ``ToolCallStart`` for ``tool.started`` events and tracks IDs in a FIFO
     queue per tool name so duplicate/parallel same-name calls still complete
-    against the correct ACP tool call.
+    against the correct ACP tool call.  Other event types (``tool.completed``,
+    ``reasoning.available``) are silently ignored.
     """
 
-    def _tool_progress(name: str, preview: str, args: Any = None) -> None:
+    def _tool_progress(event_type: str, name: str = None, preview: str = None, args: Any = None, **kwargs) -> None:
+        # Only emit ACP ToolCallStart for tool.started; ignore other event types
+        if event_type != "tool.started":
+            return
         if isinstance(args, str):
             try:
                 args = json.loads(args)

acp_adapter/permissions.py

@@ -5,14 +5,11 @@ from __future__ import annotations
 import asyncio
 import logging
 from concurrent.futures import TimeoutError as FutureTimeout
-from typing import Any, Callable, Optional
+from typing import Callable
 
 from acp.schema import (
     AllowedOutcome,
-    DeniedOutcome,
     PermissionOption,
-    RequestPermissionRequest,
-    SelectedPermissionOutcome,
 )
 
 logger = logging.getLogger(__name__)

acp_adapter/server.py

@@ -12,7 +12,8 @@ import acp
 from acp.schema import (
     AgentCapabilities,
     AuthenticateResponse,
-    AuthMethod,
+    AvailableCommand,
+    AvailableCommandsUpdate,
     ClientCapabilities,
     EmbeddedResourceContentBlock,
     ForkSessionResponse,
@@ -22,18 +23,31 @@ from acp.schema import (
     InitializeResponse,
     ListSessionsResponse,
     LoadSessionResponse,
+    McpServerHttp,
+    McpServerSse,
+    McpServerStdio,
     NewSessionResponse,
     PromptResponse,
     ResumeSessionResponse,
+    SetSessionConfigOptionResponse,
+    SetSessionModelResponse,
+    SetSessionModeResponse,
     ResourceContentBlock,
     SessionCapabilities,
     SessionForkCapabilities,
     SessionListCapabilities,
     SessionInfo,
     TextContentBlock,
+    UnstructuredCommandInput,
     Usage,
 )
 
+# AuthMethodAgent was renamed from AuthMethod in agent-client-protocol 0.9.0
+try:
+    from acp.schema import AuthMethodAgent
+except ImportError:
+    from acp.schema import AuthMethod as AuthMethodAgent  # type: ignore[attr-defined]
+
 from acp_adapter.auth import detect_provider, has_provider
 from acp_adapter.events import (
     make_message_cb,
@@ -42,7 +56,7 @@ from acp_adapter.events import (
     make_tool_progress_cb,
 )
 from acp_adapter.permissions import make_approval_callback
-from acp_adapter.session import SessionManager
+from acp_adapter.session import SessionManager, SessionState
 
 logger = logging.getLogger(__name__)
 
@@ -78,6 +92,48 @@ def _extract_text(
 class HermesACPAgent(acp.Agent):
     """ACP Agent implementation wrapping Hermes AIAgent."""
 
+    _SLASH_COMMANDS = {
+        "help": "Show available commands",
+        "model": "Show or change current model",
+        "tools": "List available tools",
+        "context": "Show conversation context info",
+        "reset": "Clear conversation history",
+        "compact": "Compress conversation context",
+        "version": "Show Hermes version",
+    }
+
+    _ADVERTISED_COMMANDS = (
+        {
+            "name": "help",
+            "description": "List available commands",
+        },
+        {
+            "name": "model",
+            "description": "Show current model and provider, or switch models",
+            "input_hint": "model name to switch to",
+        },
+        {
+            "name": "tools",
+            "description": "List available tools with descriptions",
+        },
+        {
+            "name": "context",
+            "description": "Show conversation message counts by role",
+        },
+        {
+            "name": "reset",
+            "description": "Clear conversation history",
+        },
+        {
+            "name": "compact",
+            "description": "Compress conversation context",
+        },
+        {
+            "name": "version",
+            "description": "Show Hermes version",
+        },
+    )
+
     def __init__(self, session_manager: SessionManager | None = None):
         super().__init__()
         self.session_manager = session_manager or SessionManager()
@@ -90,20 +146,88 @@ class HermesACPAgent(acp.Agent):
         self._conn = conn
         logger.info("ACP client connected")
 
+    async def _register_session_mcp_servers(
+        self,
+        state: SessionState,
+        mcp_servers: list[McpServerStdio | McpServerHttp | McpServerSse] | None,
+    ) -> None:
+        """Register ACP-provided MCP servers and refresh the agent tool surface."""
+        if not mcp_servers:
+            return
+
+        try:
+            from tools.mcp_tool import register_mcp_servers
+
+            config_map: dict[str, dict] = {}
+            for server in mcp_servers:
+                name = server.name
+                if isinstance(server, McpServerStdio):
+                    config = {
+                        "command": server.command,
+                        "args": list(server.args),
+                        "env": {item.name: item.value for item in server.env},
+                    }
+                else:
+                    config = {
+                        "url": server.url,
+                        "headers": {item.name: item.value for item in server.headers},
+                    }
+                config_map[name] = config
+
+            await asyncio.to_thread(register_mcp_servers, config_map)
+        except Exception:
+            logger.warning(
+                "Session %s: failed to register ACP MCP servers",
+                state.session_id,
+                exc_info=True,
+            )
+            return
+
+        try:
+            from model_tools import get_tool_definitions
+
+            enabled_toolsets = getattr(state.agent, "enabled_toolsets", None) or ["hermes-acp"]
+            disabled_toolsets = getattr(state.agent, "disabled_toolsets", None)
+            state.agent.tools = get_tool_definitions(
+                enabled_toolsets=enabled_toolsets,
+                disabled_toolsets=disabled_toolsets,
+                quiet_mode=True,
+            )
+            state.agent.valid_tool_names = {
+                tool["function"]["name"] for tool in state.agent.tools or []
+            }
+            invalidate = getattr(state.agent, "_invalidate_system_prompt", None)
+            if callable(invalidate):
+                invalidate()
+            logger.info(
+                "Session %s: refreshed tool surface after ACP MCP registration (%d tools)",
+                state.session_id,
+                len(state.agent.tools or []),
+            )
+        except Exception:
+            logger.warning(
+                "Session %s: failed to refresh tool surface after ACP MCP registration",
+                state.session_id,
+                exc_info=True,
+            )
+
     # ---- ACP lifecycle ------------------------------------------------------
 
     async def initialize(
         self,
-        protocol_version: int,
+        protocol_version: int | None = None,
         client_capabilities: ClientCapabilities | None = None,
         client_info: Implementation | None = None,
         **kwargs: Any,
     ) -> InitializeResponse:
+        resolved_protocol_version = (
+            protocol_version if isinstance(protocol_version, int) else acp.PROTOCOL_VERSION
+        )
         provider = detect_provider()
         auth_methods = None
         if provider:
             auth_methods = [
-                AuthMethod(
+                AuthMethodAgent(
                     id=provider,
                     name=f"{provider} runtime credentials",
                     description=f"Authenticate Hermes using the currently configured {provider} runtime credentials.",
@@ -111,7 +235,11 @@ class HermesACPAgent(acp.Agent):
             ]
 
         client_name = client_info.name if client_info else "unknown"
-        logger.info("Initialize from %s (protocol v%s)", client_name, protocol_version)
+        logger.info(
+            "Initialize from %s (protocol v%s)",
+            client_name,
+            resolved_protocol_version,
+        )
 
         return InitializeResponse(
             protocol_version=acp.PROTOCOL_VERSION,
@@ -139,7 +267,9 @@ class HermesACPAgent(acp.Agent):
         **kwargs: Any,
     ) -> NewSessionResponse:
         state = self.session_manager.create_session(cwd=cwd)
+        await self._register_session_mcp_servers(state, mcp_servers)
         logger.info("New session %s (cwd=%s)", state.session_id, cwd)
+        self._schedule_available_commands_update(state.session_id)
         return NewSessionResponse(session_id=state.session_id)
 
     async def load_session(
@@ -153,7 +283,9 @@ class HermesACPAgent(acp.Agent):
         if state is None:
             logger.warning("load_session: session %s not found", session_id)
             return None
+        await self._register_session_mcp_servers(state, mcp_servers)
         logger.info("Loaded session %s", session_id)
+        self._schedule_available_commands_update(session_id)
         return LoadSessionResponse()
 
     async def resume_session(
@@ -167,7 +299,9 @@ class HermesACPAgent(acp.Agent):
         if state is None:
             logger.warning("resume_session: session %s not found, creating new", session_id)
             state = self.session_manager.create_session(cwd=cwd)
+        await self._register_session_mcp_servers(state, mcp_servers)
         logger.info("Resumed session %s", state.session_id)
+        self._schedule_available_commands_update(state.session_id)
         return ResumeSessionResponse()
 
     async def cancel(self, session_id: str, **kwargs: Any) -> None:
@@ -190,7 +324,11 @@ class HermesACPAgent(acp.Agent):
     ) -> ForkSessionResponse:
         state = self.session_manager.fork_session(session_id, cwd=cwd)
         new_id = state.session_id if state else ""
+        if state is not None:
+            await self._register_session_mcp_servers(state, mcp_servers)
         logger.info("Forked session %s -> %s", session_id, new_id)
+        if new_id:
+            self._schedule_available_commands_update(new_id)
         return ForkSessionResponse(session_id=new_id)
 
     async def list_sessions(
@@ -226,10 +364,19 @@ class HermesACPAgent(acp.Agent):
             logger.error("prompt: session %s not found", session_id)
             return PromptResponse(stop_reason="refusal")
 
-        user_text = _extract_text(prompt)
-        if not user_text.strip():
+        user_text = _extract_text(prompt).strip()
+        if not user_text:
             return PromptResponse(stop_reason="end_turn")
 
+        # Intercept slash commands — handle locally without calling the LLM
+        if user_text.startswith("/"):
+            response_text = self._handle_slash_command(user_text, state)
+            if response_text is not None:
+                if self._conn:
+                    update = acp.update_agent_message_text(response_text)
+                    await self._conn.session_update(session_id, update)
+                return PromptResponse(stop_reason="end_turn")
+
         logger.info("Prompt on session %s: %s", session_id, user_text[:100])
 
         conn = self._conn
@@ -295,6 +442,8 @@ class HermesACPAgent(acp.Agent):
 
         if result.get("messages"):
             state.history = result["messages"]
+            # Persist updated history so sessions survive process restarts.
+            self.session_manager.save_session(session_id)
 
         final_response = result.get("final_response", "")
         if final_response and conn:
@@ -315,19 +464,263 @@ class HermesACPAgent(acp.Agent):
         stop_reason = "cancelled" if state.cancel_event and state.cancel_event.is_set() else "end_turn"
         return PromptResponse(stop_reason=stop_reason, usage=usage)
 
-    # ---- Model switching ----------------------------------------------------
+    # ---- Slash commands (headless) -------------------------------------------
+
+    @classmethod
+    def _available_commands(cls) -> list[AvailableCommand]:
+        commands: list[AvailableCommand] = []
+        for spec in cls._ADVERTISED_COMMANDS:
+            input_hint = spec.get("input_hint")
+            commands.append(
+                AvailableCommand(
+                    name=spec["name"],
+                    description=spec["description"],
+                    input=UnstructuredCommandInput(hint=input_hint)
+                    if input_hint
+                    else None,
+                )
+            )
+        return commands
+
+    async def _send_available_commands_update(self, session_id: str) -> None:
+        """Advertise supported slash commands to the connected ACP client."""
+        if not self._conn:
+            return
+
+        try:
+            await self._conn.session_update(
+                session_id=session_id,
+                update=AvailableCommandsUpdate(
+                    sessionUpdate="available_commands_update",
+                    availableCommands=self._available_commands(),
+                ),
+            )
+        except Exception:
+            logger.warning(
+                "Failed to advertise ACP slash commands for session %s",
+                session_id,
+                exc_info=True,
+            )
+
+    def _schedule_available_commands_update(self, session_id: str) -> None:
+        """Send the command advertisement after the session response is queued."""
+        if not self._conn:
+            return
+        loop = asyncio.get_running_loop()
+        loop.call_soon(
+            asyncio.create_task, self._send_available_commands_update(session_id)
+        )
+
+    def _handle_slash_command(self, text: str, state: SessionState) -> str | None:
+        """Dispatch a slash command and return the response text.
+
+        Returns ``None`` for unrecognized commands so they fall through
+        to the LLM (the user may have typed ``/something`` as prose).
+        """
+        parts = text.split(maxsplit=1)
+        cmd = parts[0].lstrip("/").lower()
+        args = parts[1].strip() if len(parts) > 1 else ""
+
+        handler = {
+            "help": self._cmd_help,
+            "model": self._cmd_model,
+            "tools": self._cmd_tools,
+            "context": self._cmd_context,
+            "reset": self._cmd_reset,
+            "compact": self._cmd_compact,
+            "version": self._cmd_version,
+        }.get(cmd)
+
+        if handler is None:
+            return None  # not a known command — let the LLM handle it
+
+        try:
+            return handler(args, state)
+        except Exception as e:
+            logger.error("Slash command /%s error: %s", cmd, e, exc_info=True)
+            return f"Error executing /{cmd}: {e}"
+
+    def _cmd_help(self, args: str, state: SessionState) -> str:
+        lines = ["Available commands:", ""]
+        for cmd, desc in self._SLASH_COMMANDS.items():
+            lines.append(f"  /{cmd:10s}  {desc}")
+        lines.append("")
+        lines.append("Unrecognized /commands are sent to the model as normal messages.")
+        return "\n".join(lines)
+
+    def _cmd_model(self, args: str, state: SessionState) -> str:
+        if not args:
+            model = state.model or getattr(state.agent, "model", "unknown")
+            provider = getattr(state.agent, "provider", None) or "auto"
+            return f"Current model: {model}\nProvider: {provider}"
+
+        new_model = args.strip()
+        target_provider = None
+        current_provider = getattr(state.agent, "provider", None) or "openrouter"
+
+        # Auto-detect provider for the requested model
+        try:
+            from hermes_cli.models import parse_model_input, detect_provider_for_model
+            target_provider, new_model = parse_model_input(new_model, current_provider)
+            if target_provider == current_provider:
+                detected = detect_provider_for_model(new_model, current_provider)
+                if detected:
+                    target_provider, new_model = detected
+        except Exception:
+            logger.debug("Provider detection failed, using model as-is", exc_info=True)
+
+        state.model = new_model
+        state.agent = self.session_manager._make_agent(
+            session_id=state.session_id,
+            cwd=state.cwd,
+            model=new_model,
+            requested_provider=target_provider or current_provider,
+        )
+        self.session_manager.save_session(state.session_id)
+        provider_label = getattr(state.agent, "provider", None) or target_provider or current_provider
+        logger.info("Session %s: model switched to %s", state.session_id, new_model)
+        return f"Model switched to: {new_model}\nProvider: {provider_label}"
+
+    def _cmd_tools(self, args: str, state: SessionState) -> str:
+        try:
+            from model_tools import get_tool_definitions
+            toolsets = getattr(state.agent, "enabled_toolsets", None) or ["hermes-acp"]
+            tools = get_tool_definitions(enabled_toolsets=toolsets, quiet_mode=True)
+            if not tools:
+                return "No tools available."
+            lines = [f"Available tools ({len(tools)}):"]
+            for t in tools:
+                name = t.get("function", {}).get("name", "?")
+                desc = t.get("function", {}).get("description", "")
+                # Truncate long descriptions
+                if len(desc) > 80:
+                    desc = desc[:77] + "..."
+                lines.append(f"  {name}: {desc}")
+            return "\n".join(lines)
+        except Exception as e:
+            return f"Could not list tools: {e}"
+
+    def _cmd_context(self, args: str, state: SessionState) -> str:
+        n_messages = len(state.history)
+        if n_messages == 0:
+            return "Conversation is empty (no messages yet)."
+        # Count by role
+        roles: dict[str, int] = {}
+        for msg in state.history:
+            role = msg.get("role", "unknown")
+            roles[role] = roles.get(role, 0) + 1
+        lines = [
+            f"Conversation: {n_messages} messages",
+            f"  user: {roles.get('user', 0)}, assistant: {roles.get('assistant', 0)}, "
+            f"tool: {roles.get('tool', 0)}, system: {roles.get('system', 0)}",
+        ]
+        model = state.model or getattr(state.agent, "model", "")
+        if model:
+            lines.append(f"Model: {model}")
+        return "\n".join(lines)
+
+    def _cmd_reset(self, args: str, state: SessionState) -> str:
+        state.history.clear()
+        self.session_manager.save_session(state.session_id)
+        return "Conversation history cleared."
+
+    def _cmd_compact(self, args: str, state: SessionState) -> str:
+        if not state.history:
+            return "Nothing to compress — conversation is empty."
+        try:
+            agent = state.agent
+            if not getattr(agent, "compression_enabled", True):
+                return "Context compression is disabled for this agent."
+            if not hasattr(agent, "_compress_context"):
+                return "Context compression not available for this agent."
+
+            from agent.model_metadata import estimate_messages_tokens_rough
+
+            original_count = len(state.history)
+            approx_tokens = estimate_messages_tokens_rough(state.history)
+            original_session_db = getattr(agent, "_session_db", None)
+
+            try:
+                # ACP sessions must keep a stable session id, so avoid the
+                # SQLite session-splitting side effect inside _compress_context.
+                agent._session_db = None
+                compressed, _ = agent._compress_context(
+                    state.history,
+                    getattr(agent, "_cached_system_prompt", "") or "",
+                    approx_tokens=approx_tokens,
+                    task_id=state.session_id,
+                )
+            finally:
+                agent._session_db = original_session_db
+
+            state.history = compressed
+            self.session_manager.save_session(state.session_id)
+
+            new_count = len(state.history)
+            new_tokens = estimate_messages_tokens_rough(state.history)
+            return (
+                f"Context compressed: {original_count} -> {new_count} messages\n"
+                f"~{approx_tokens:,} -> ~{new_tokens:,} tokens"
+            )
+        except Exception as e:
+            return f"Compression failed: {e}"
+
+    def _cmd_version(self, args: str, state: SessionState) -> str:
+        return f"Hermes Agent v{HERMES_VERSION}"
+
+    # ---- Model switching (ACP protocol method) -------------------------------
 
     async def set_session_model(
         self, model_id: str, session_id: str, **kwargs: Any
-    ):
-        """Switch the model for a session."""
+    ) -> SetSessionModelResponse | None:
+        """Switch the model for a session (called by ACP protocol)."""
         state = self.session_manager.get_session(session_id)
         if state:
             state.model = model_id
+            current_provider = getattr(state.agent, "provider", None)
+            current_base_url = getattr(state.agent, "base_url", None)
+            current_api_mode = getattr(state.agent, "api_mode", None)
             state.agent = self.session_manager._make_agent(
                 session_id=session_id,
                 cwd=state.cwd,
                 model=model_id,
+                requested_provider=current_provider,
+                base_url=current_base_url,
+                api_mode=current_api_mode,
             )
+            self.session_manager.save_session(session_id)
             logger.info("Session %s: model switched to %s", session_id, model_id)
+            return SetSessionModelResponse()
+        logger.warning("Session %s: model switch requested for missing session", session_id)
         return None
+
+    async def set_session_mode(
+        self, mode_id: str, session_id: str, **kwargs: Any
+    ) -> SetSessionModeResponse | None:
+        """Persist the editor-requested mode so ACP clients do not fail on mode switches."""
+        state = self.session_manager.get_session(session_id)
+        if state is None:
+            logger.warning("Session %s: mode switch requested for missing session", session_id)
+            return None
+        setattr(state, "mode", mode_id)
+        self.session_manager.save_session(session_id)
+        logger.info("Session %s: mode switched to %s", session_id, mode_id)
+        return SetSessionModeResponse()
+
+    async def set_config_option(
+        self, config_id: str, session_id: str, value: str, **kwargs: Any
+    ) -> SetSessionConfigOptionResponse | None:
+        """Accept ACP config option updates even when Hermes has no typed ACP config surface yet."""
+        state = self.session_manager.get_session(session_id)
+        if state is None:
+            logger.warning("Session %s: config update requested for missing session", session_id)
+            return None
+
+        options = getattr(state, "config_options", None)
+        if not isinstance(options, dict):
+            options = {}
+        options[str(config_id)] = value
+        setattr(state, "config_options", options)
+        self.session_manager.save_session(session_id)
+        logger.info("Session %s: config option %s updated", session_id, config_id)
+        return SetSessionConfigOptionResponse(config_options=[])

acp_adapter/session.py

@@ -1,8 +1,19 @@
-"""ACP session manager — maps ACP sessions to Hermes AIAgent instances."""
+"""ACP session manager — maps ACP sessions to Hermes AIAgent instances.
+
+Sessions are persisted to the shared SessionDB (``~/.hermes/state.db``) so they
+survive process restarts and appear in ``session_search``.  When the editor
+reconnects after idle/restart, the ``load_session`` / ``resume_session`` calls
+find the persisted session in the database and restore the full conversation
+history.
+"""
 from __future__ import annotations
 
+from hermes_constants import get_hermes_home
+
 import copy
+import json
 import logging
+import sys
 import uuid
 from dataclasses import dataclass, field
 from threading import Lock
@@ -11,6 +22,17 @@ from typing import Any, Dict, List, Optional
 logger = logging.getLogger(__name__)
 
 
+def _acp_stderr_print(*args, **kwargs) -> None:
+    """Best-effort human-readable output sink for ACP stdio sessions.
+
+    ACP reserves stdout for JSON-RPC frames, so any incidental CLI/status output
+    from AIAgent must be redirected away from stdout. Route it to stderr instead.
+    """
+    kwargs = dict(kwargs)
+    kwargs.setdefault("file", sys.stderr)
+    print(*args, **kwargs)
+
+
 def _register_task_cwd(task_id: str, cwd: str) -> None:
     """Bind a task/session id to the editor's working directory for tools."""
     if not task_id:
@@ -46,18 +68,26 @@ class SessionState:
 
 
 class SessionManager:
-    """Thread-safe manager for ACP sessions backed by Hermes AIAgent instances."""
+    """Thread-safe manager for ACP sessions backed by Hermes AIAgent instances.
 
-    def __init__(self, agent_factory=None):
+    Sessions are held in-memory for fast access **and** persisted to the
+    shared SessionDB so they survive process restarts and are searchable
+    via ``session_search``.
+    """
+
+    def __init__(self, agent_factory=None, db=None):
         """
         Args:
             agent_factory: Optional callable that creates an AIAgent-like object.
                            Used by tests. When omitted, a real AIAgent is created
                            using the current Hermes runtime provider configuration.
+            db:            Optional SessionDB instance. When omitted, the default
+                           SessionDB (``~/.hermes/state.db``) is lazily created.
         """
         self._sessions: Dict[str, SessionState] = {}
         self._lock = Lock()
         self._agent_factory = agent_factory
+        self._db_instance = db  # None → lazy-init on first use
 
     # ---- public API ---------------------------------------------------------
 
@@ -77,54 +107,67 @@ class SessionManager:
         with self._lock:
             self._sessions[session_id] = state
         _register_task_cwd(session_id, cwd)
+        self._persist(state)
         logger.info("Created ACP session %s (cwd=%s)", session_id, cwd)
         return state
 
     def get_session(self, session_id: str) -> Optional[SessionState]:
-        """Return the session for *session_id*, or ``None``."""
+        """Return the session for *session_id*, or ``None``.
+
+        If the session is not in memory but exists in the database (e.g. after
+        a process restart), it is transparently restored.
+        """
         with self._lock:
-            return self._sessions.get(session_id)
+            state = self._sessions.get(session_id)
+        if state is not None:
+            return state
+        # Attempt to restore from database.
+        return self._restore(session_id)
 
     def remove_session(self, session_id: str) -> bool:
-        """Remove a session. Returns True if it existed."""
+        """Remove a session from memory and database. Returns True if it existed."""
         with self._lock:
             existed = self._sessions.pop(session_id, None) is not None
-        if existed:
+        db_existed = self._delete_persisted(session_id)
+        if existed or db_existed:
             _clear_task_cwd(session_id)
-        return existed
+        return existed or db_existed
 
     def fork_session(self, session_id: str, cwd: str = ".") -> Optional[SessionState]:
         """Deep-copy a session's history into a new session."""
         import threading
 
-        with self._lock:
-            original = self._sessions.get(session_id)
-            if original is None:
-                return None
+        original = self.get_session(session_id)  # checks DB too
+        if original is None:
+            return None
 
-            new_id = str(uuid.uuid4())
-            agent = self._make_agent(
-                session_id=new_id,
-                cwd=cwd,
-                model=original.model or None,
-            )
-            state = SessionState(
-                session_id=new_id,
-                agent=agent,
-                cwd=cwd,
-                model=getattr(agent, "model", original.model) or original.model,
-                history=copy.deepcopy(original.history),
-                cancel_event=threading.Event(),
-            )
+        new_id = str(uuid.uuid4())
+        agent = self._make_agent(
+            session_id=new_id,
+            cwd=cwd,
+            model=original.model or None,
+        )
+        state = SessionState(
+            session_id=new_id,
+            agent=agent,
+            cwd=cwd,
+            model=getattr(agent, "model", original.model) or original.model,
+            history=copy.deepcopy(original.history),
+            cancel_event=threading.Event(),
+        )
+        with self._lock:
             self._sessions[new_id] = state
         _register_task_cwd(new_id, cwd)
+        self._persist(state)
         logger.info("Forked ACP session %s -> %s", session_id, new_id)
         return state
 
     def list_sessions(self) -> List[Dict[str, Any]]:
-        """Return lightweight info dicts for all sessions."""
+        """Return lightweight info dicts for all sessions (memory + database)."""
+        # Collect in-memory sessions first.
         with self._lock:
-            return [
+            seen_ids = set(self._sessions.keys())
+            results = [
                 {
                     "session_id": s.session_id,
                     "cwd": s.cwd,
@@ -134,23 +177,245 @@ class SessionManager:
                 for s in self._sessions.values()
             ]
 
+        # Merge any persisted sessions not currently in memory.
+        db = self._get_db()
+        if db is not None:
+            try:
+                rows = db.search_sessions(source="acp", limit=1000)
+                for row in rows:
+                    sid = row["id"]
+                    if sid in seen_ids:
+                        continue
+                    # Extract cwd from model_config JSON.
+                    cwd = "."
+                    mc = row.get("model_config")
+                    if mc:
+                        try:
+                            cwd = json.loads(mc).get("cwd", ".")
+                        except (json.JSONDecodeError, TypeError):
+                            pass
+                    results.append({
+                        "session_id": sid,
+                        "cwd": cwd,
+                        "model": row.get("model") or "",
+                        "history_len": row.get("message_count") or 0,
+                    })
+            except Exception:
+                logger.debug("Failed to list ACP sessions from DB", exc_info=True)
+
+        return results
+
     def update_cwd(self, session_id: str, cwd: str) -> Optional[SessionState]:
         """Update the working directory for a session and its tool overrides."""
-        with self._lock:
-            state = self._sessions.get(session_id)
-            if state is None:
-                return None
-            state.cwd = cwd
+        state = self.get_session(session_id)  # checks DB too
+        if state is None:
+            return None
+        state.cwd = cwd
         _register_task_cwd(session_id, cwd)
+        self._persist(state)
         return state
 
     def cleanup(self) -> None:
-        """Remove all sessions and clear task-specific cwd overrides."""
+        """Remove all sessions (memory and database) and clear task-specific cwd overrides."""
         with self._lock:
             session_ids = list(self._sessions.keys())
             self._sessions.clear()
         for session_id in session_ids:
             _clear_task_cwd(session_id)
+            self._delete_persisted(session_id)
+        # Also remove any DB-only ACP sessions not currently in memory.
+        db = self._get_db()
+        if db is not None:
+            try:
+                rows = db.search_sessions(source="acp", limit=10000)
+                for row in rows:
+                    sid = row["id"]
+                    _clear_task_cwd(sid)
+                    db.delete_session(sid)
+            except Exception:
+                logger.debug("Failed to cleanup ACP sessions from DB", exc_info=True)
+
+    def save_session(self, session_id: str) -> None:
+        """Persist the current state of a session to the database.
+
+        Called by the server after prompt completion, slash commands that
+        mutate history, and model switches.
+        """
+        with self._lock:
+            state = self._sessions.get(session_id)
+        if state is not None:
+            self._persist(state)
+
+    # ---- persistence via SessionDB ------------------------------------------
+
+    def _get_db(self):
+        """Lazily initialise and return the SessionDB instance.
+
+        Returns ``None`` if the DB is unavailable (e.g. import error in a
+        minimal test environment).
+
+        Note: we resolve ``HERMES_HOME`` dynamically rather than relying on
+        the module-level ``DEFAULT_DB_PATH`` constant, because that constant
+        is evaluated at import time and won't reflect env-var changes made
+        later (e.g. by the test fixture ``_isolate_hermes_home``).
+        """
+        if self._db_instance is not None:
+            return self._db_instance
+        try:
+            import os
+            from pathlib import Path
+            from hermes_state import SessionDB
+            hermes_home = get_hermes_home()
+            self._db_instance = SessionDB(db_path=hermes_home / "state.db")
+            return self._db_instance
+        except Exception:
+            logger.debug("SessionDB unavailable for ACP persistence", exc_info=True)
+            return None
+
+    def _persist(self, state: SessionState) -> None:
+        """Write session state to the database.
+
+        Creates the session record if it doesn't exist, then replaces all
+        stored messages with the current in-memory history.
+        """
+        db = self._get_db()
+        if db is None:
+            return
+
+        # Ensure model is a plain string (not a MagicMock or other proxy).
+        model_str = str(state.model) if state.model else None
+        session_meta = {"cwd": state.cwd}
+        provider = getattr(state.agent, "provider", None)
+        base_url = getattr(state.agent, "base_url", None)
+        api_mode = getattr(state.agent, "api_mode", None)
+        if isinstance(provider, str) and provider.strip():
+            session_meta["provider"] = provider.strip()
+        if isinstance(base_url, str) and base_url.strip():
+            session_meta["base_url"] = base_url.strip()
+        if isinstance(api_mode, str) and api_mode.strip():
+            session_meta["api_mode"] = api_mode.strip()
+        cwd_json = json.dumps(session_meta)
+
+        try:
+            # Ensure the session record exists.
+            existing = db.get_session(state.session_id)
+            if existing is None:
+                db.create_session(
+                    session_id=state.session_id,
+                    source="acp",
+                    model=model_str,
+                    model_config={"cwd": state.cwd},
+                )
+            else:
+                # Update model_config (contains cwd) if changed.
+                try:
+                    with db._lock:
+                        db._conn.execute(
+                            "UPDATE sessions SET model_config = ?, model = COALESCE(?, model) WHERE id = ?",
+                            (cwd_json, model_str, state.session_id),
+                        )
+                        db._conn.commit()
+                except Exception:
+                    logger.debug("Failed to update ACP session metadata", exc_info=True)
+
+            # Replace stored messages with current history.
+            db.clear_messages(state.session_id)
+            for msg in state.history:
+                db.append_message(
+                    session_id=state.session_id,
+                    role=msg.get("role", "user"),
+                    content=msg.get("content"),
+                    tool_name=msg.get("tool_name") or msg.get("name"),
+                    tool_calls=msg.get("tool_calls"),
+                    tool_call_id=msg.get("tool_call_id"),
+                )
+        except Exception:
+            logger.warning("Failed to persist ACP session %s", state.session_id, exc_info=True)
+
+    def _restore(self, session_id: str) -> Optional[SessionState]:
+        """Load a session from the database into memory, recreating the AIAgent."""
+        import threading
+
+        db = self._get_db()
+        if db is None:
+            return None
+
+        try:
+            row = db.get_session(session_id)
+        except Exception:
+            logger.debug("Failed to query DB for ACP session %s", session_id, exc_info=True)
+            return None
+
+        if row is None:
+            return None
+
+        # Only restore ACP sessions.
+        if row.get("source") != "acp":
+            return None
+
+        # Extract cwd from model_config.
+        cwd = "."
+        requested_provider = row.get("billing_provider")
+        restored_base_url = row.get("billing_base_url")
+        restored_api_mode = None
+        mc = row.get("model_config")
+        if mc:
+            try:
+                meta = json.loads(mc)
+                if isinstance(meta, dict):
+                    cwd = meta.get("cwd", ".")
+                    requested_provider = meta.get("provider") or requested_provider
+                    restored_base_url = meta.get("base_url") or restored_base_url
+                    restored_api_mode = meta.get("api_mode") or restored_api_mode
+            except (json.JSONDecodeError, TypeError):
+                pass
+
+        model = row.get("model") or None
+
+        # Load conversation history.
+        try:
+            history = db.get_messages_as_conversation(session_id)
+        except Exception:
+            logger.warning("Failed to load messages for ACP session %s", session_id, exc_info=True)
+            history = []
+
+        try:
+            agent = self._make_agent(
+                session_id=session_id,
+                cwd=cwd,
+                model=model,
+                requested_provider=requested_provider,
+                base_url=restored_base_url,
+                api_mode=restored_api_mode,
+            )
+        except Exception:
+            logger.warning("Failed to recreate agent for ACP session %s", session_id, exc_info=True)
+            return None
+
+        state = SessionState(
+            session_id=session_id,
+            agent=agent,
+            cwd=cwd,
+            model=model or getattr(agent, "model", "") or "",
+            history=history,
+            cancel_event=threading.Event(),
+        )
+        with self._lock:
+            self._sessions[session_id] = state
+        _register_task_cwd(session_id, cwd)
+        logger.info("Restored ACP session %s from DB (%d messages)", session_id, len(history))
+        return state
+
+    def _delete_persisted(self, session_id: str) -> bool:
+        """Delete a session from the database. Returns True if it existed."""
+        db = self._get_db()
+        if db is None:
+            return False
+        try:
+            return db.delete_session(session_id)
+        except Exception:
+            logger.debug("Failed to delete ACP session %s from DB", session_id, exc_info=True)
+            return False
 
     # ---- internal -----------------------------------------------------------
 
@@ -160,6 +425,9 @@ class SessionManager:
         session_id: str,
         cwd: str,
         model: str | None = None,
+        requested_provider: str | None = None,
+        base_url: str | None = None,
+        api_mode: str | None = None,
     ):
         if self._agent_factory is not None:
             return self._agent_factory()
@@ -170,11 +438,11 @@ class SessionManager:
 
         config = load_config()
         model_cfg = config.get("model")
-        default_model = "anthropic/claude-opus-4.6"
-        requested_provider = None
+        default_model = ""
+        config_provider = None
         if isinstance(model_cfg, dict):
             default_model = str(model_cfg.get("default") or default_model)
-            requested_provider = model_cfg.get("provider")
+            config_provider = model_cfg.get("provider")
         elif isinstance(model_cfg, str) and model_cfg.strip():
             default_model = model_cfg.strip()
 
@@ -187,17 +455,23 @@ class SessionManager:
         }
 
         try:
-            runtime = resolve_runtime_provider(requested=requested_provider)
+            runtime = resolve_runtime_provider(requested=requested_provider or config_provider)
             kwargs.update(
                 {
                     "provider": runtime.get("provider"),
-                    "api_mode": runtime.get("api_mode"),
-                    "base_url": runtime.get("base_url"),
+                    "api_mode": api_mode or runtime.get("api_mode"),
+                    "base_url": base_url or runtime.get("base_url"),
                     "api_key": runtime.get("api_key"),
+                    "command": runtime.get("command"),
+                    "args": list(runtime.get("args") or []),
                 }
             )
         except Exception:
             logger.debug("ACP session falling back to default provider resolution", exc_info=True)
 
         _register_task_cwd(session_id, cwd)
-        return AIAgent(**kwargs)
+        agent = AIAgent(**kwargs)
+        # ACP stdio transport requires stdout to remain protocol-only JSON-RPC.
+        # Route any incidental human-readable agent output to stderr instead.
+        agent._print_fn = _acp_stderr_print
+        return agent

agent/builtin_memory_provider.py

@@ -0,0 +1,113 @@
+"""BuiltinMemoryProvider — wraps MEMORY.md / USER.md as a MemoryProvider.
+
+Always registered as the first provider. Cannot be disabled or removed.
+This is the existing Hermes memory system exposed through the provider
+interface for compatibility with the MemoryManager.
+
+The actual storage logic lives in tools/memory_tool.py (MemoryStore).
+This provider is a thin adapter that delegates to MemoryStore and
+exposes the memory tool schema.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from typing import Any, Dict, List, Optional
+
+from agent.memory_provider import MemoryProvider
+
+logger = logging.getLogger(__name__)
+
+
+class BuiltinMemoryProvider(MemoryProvider):
+    """Built-in file-backed memory (MEMORY.md + USER.md).
+
+    Always active, never disabled by other providers. The `memory` tool
+    is handled by run_agent.py's agent-level tool interception (not through
+    the normal registry), so get_tool_schemas() returns an empty list —
+    the memory tool is already wired separately.
+    """
+
+    def __init__(
+        self,
+        memory_store=None,
+        memory_enabled: bool = False,
+        user_profile_enabled: bool = False,
+    ):
+        self._store = memory_store
+        self._memory_enabled = memory_enabled
+        self._user_profile_enabled = user_profile_enabled
+
+    @property
+    def name(self) -> str:
+        return "builtin"
+
+    def is_available(self) -> bool:
+        """Built-in memory is always available."""
+        return True
+
+    def initialize(self, session_id: str, **kwargs) -> None:
+        """Load memory from disk if not already loaded."""
+        if self._store is not None:
+            self._store.load_from_disk()
+
+    def system_prompt_block(self) -> str:
+        """Return MEMORY.md and USER.md content for the system prompt.
+
+        Uses the frozen snapshot captured at load time. This ensures the
+        system prompt stays stable throughout a session (preserving the
+        prompt cache), even though the live entries may change via tool calls.
+        """
+        if not self._store:
+            return ""
+
+        parts = []
+        if self._memory_enabled:
+            mem_block = self._store.format_for_system_prompt("memory")
+            if mem_block:
+                parts.append(mem_block)
+        if self._user_profile_enabled:
+            user_block = self._store.format_for_system_prompt("user")
+            if user_block:
+                parts.append(user_block)
+
+        return "\n\n".join(parts)
+
+    def prefetch(self, query: str, *, session_id: str = "") -> str:
+        """Built-in memory doesn't do query-based recall — it's injected via system_prompt_block."""
+        return ""
+
+    def sync_turn(self, user_content: str, assistant_content: str, *, session_id: str = "") -> None:
+        """Built-in memory doesn't auto-sync turns — writes happen via the memory tool."""
+
+    def get_tool_schemas(self) -> List[Dict[str, Any]]:
+        """Return empty list.
+
+        The `memory` tool is an agent-level intercepted tool, handled
+        specially in run_agent.py before normal tool dispatch. It's not
+        part of the standard tool registry. We don't duplicate it here.
+        """
+        return []
+
+    def handle_tool_call(self, tool_name: str, args: Dict[str, Any], **kwargs) -> str:
+        """Not used — the memory tool is intercepted in run_agent.py."""
+        return json.dumps({"error": "Built-in memory tool is handled by the agent loop"})
+
+    def shutdown(self) -> None:
+        """No cleanup needed — files are saved on every write."""
+
+    # -- Property access for backward compatibility --------------------------
+
+    @property
+    def store(self):
+        """Access the underlying MemoryStore for legacy code paths."""
+        return self._store
+
+    @property
+    def memory_enabled(self) -> bool:
+        return self._memory_enabled
+
+    @property
+    def user_profile_enabled(self) -> bool:
+        return self._user_profile_enabled

agent/context_compressor.py

@@ -1,12 +1,19 @@
 """Automatic context window compression for long conversations.
 
 Self-contained class with its own OpenAI client for summarization.
-Uses Gemini Flash (cheap/fast) to summarize middle turns while
+Uses auxiliary model (cheap/fast) to summarize middle turns while
 protecting head and tail context.
+
+Improvements over v1:
+  - Structured summary template (Goal, Progress, Decisions, Files, Next Steps)
+  - Iterative summary updates (preserves info across multiple compactions)
+  - Token-budget tail protection instead of fixed message count
+  - Tool output pruning before LLM summarization (cheap pre-pass)
+  - Scaled summary budget (proportional to compressed content)
+  - Richer tool call/result detail in summarizer input
 """
 
 import logging
-import os
 from typing import Any, Dict, List, Optional
 
 from agent.auxiliary_client import call_llm
@@ -27,12 +34,29 @@ SUMMARY_PREFIX = (
 )
 LEGACY_SUMMARY_PREFIX = "[CONTEXT SUMMARY]:"
 
+# Minimum tokens for the summary output
+_MIN_SUMMARY_TOKENS = 2000
+# Proportion of compressed content to allocate for summary
+_SUMMARY_RATIO = 0.20
+# Absolute ceiling for summary tokens (even on very large context windows)
+_SUMMARY_TOKENS_CEILING = 12_000
+
+# Placeholder used when pruning old tool results
+_PRUNED_TOOL_PLACEHOLDER = "[Old tool output cleared to save context space]"
+
+# Chars per token rough estimate
+_CHARS_PER_TOKEN = 4
+
 
 class ContextCompressor:
     """Compresses conversation context when approaching the model's context limit.
 
-    Algorithm: protect first N + last N turns, summarize everything in between.
-    Token tracking uses actual counts from API responses for accuracy.
+    Algorithm:
+      1. Prune old tool results (cheap, no LLM call)
+      2. Protect head messages (system prompt + first exchange)
+      3. Protect tail messages by token budget (most recent ~20K tokens)
+      4. Summarize middle turns with structured LLM prompt
+      5. On subsequent compactions, iteratively update the previous summary
     """
 
     def __init__(
@@ -40,23 +64,50 @@ class ContextCompressor:
         model: str,
         threshold_percent: float = 0.50,
         protect_first_n: int = 3,
-        protect_last_n: int = 4,
-        summary_target_tokens: int = 2500,
+        protect_last_n: int = 20,
+        summary_target_ratio: float = 0.20,
         quiet_mode: bool = False,
         summary_model_override: str = None,
         base_url: str = "",
+        api_key: str = "",
+        config_context_length: int | None = None,
+        provider: str = "",
     ):
         self.model = model
         self.base_url = base_url
+        self.api_key = api_key
+        self.provider = provider
         self.threshold_percent = threshold_percent
         self.protect_first_n = protect_first_n
         self.protect_last_n = protect_last_n
-        self.summary_target_tokens = summary_target_tokens
+        self.summary_target_ratio = max(0.10, min(summary_target_ratio, 0.80))
         self.quiet_mode = quiet_mode
 
-        self.context_length = get_model_context_length(model, base_url=base_url)
+        self.context_length = get_model_context_length(
+            model, base_url=base_url, api_key=api_key,
+            config_context_length=config_context_length,
+            provider=provider,
+        )
         self.threshold_tokens = int(self.context_length * threshold_percent)
         self.compression_count = 0
+
+        # Derive token budgets: ratio is relative to the threshold, not total context
+        target_tokens = int(self.threshold_tokens * self.summary_target_ratio)
+        self.tail_token_budget = target_tokens
+        self.max_summary_tokens = min(
+            int(self.context_length * 0.05), _SUMMARY_TOKENS_CEILING,
+        )
+
+        if not quiet_mode:
+            logger.info(
+                "Context compressor initialized: model=%s context_length=%d "
+                "threshold=%d (%.0f%%) target_ratio=%.0f%% tail_budget=%d "
+                "provider=%s base_url=%s",
+                model, self.context_length, self.threshold_tokens,
+                threshold_percent * 100, self.summary_target_ratio * 100,
+                self.tail_token_budget,
+                provider or "none", base_url or "none",
+            )
         self._context_probed = False  # True after a step-down from context error
 
         self.last_prompt_tokens = 0
@@ -65,6 +116,9 @@ class ContextCompressor:
 
         self.summary_model = summary_model_override or ""
 
+        # Stores the previous compaction summary for iterative updates
+        self._previous_summary: Optional[str] = None
+
     def update_from_response(self, usage: Dict[str, Any]):
         """Update tracked token usage from API response."""
         self.last_prompt_tokens = usage.get("prompt_tokens", 0)
@@ -87,57 +141,213 @@ class ContextCompressor:
             "last_prompt_tokens": self.last_prompt_tokens,
             "threshold_tokens": self.threshold_tokens,
             "context_length": self.context_length,
-            "usage_percent": (self.last_prompt_tokens / self.context_length * 100) if self.context_length else 0,
+            "usage_percent": min(100, (self.last_prompt_tokens / self.context_length * 100)) if self.context_length else 0,
             "compression_count": self.compression_count,
         }
 
-    def _generate_summary(self, turns_to_summarize: List[Dict[str, Any]]) -> Optional[str]:
-        """Generate a concise summary of conversation turns.
+    # ------------------------------------------------------------------
+    # Tool output pruning (cheap pre-pass, no LLM call)
+    # ------------------------------------------------------------------
 
-        Tries the auxiliary model first, then falls back to the user's main
-        model.  Returns None if all attempts fail — the caller should drop
+    def _prune_old_tool_results(
+        self, messages: List[Dict[str, Any]], protect_tail_count: int,
+    ) -> tuple[List[Dict[str, Any]], int]:
+        """Replace old tool result contents with a short placeholder.
+
+        Walks backward from the end, protecting the most recent
+        ``protect_tail_count`` messages. Older tool results get their
+        content replaced with a placeholder string.
+
+        Returns (pruned_messages, pruned_count).
+        """
+        if not messages:
+            return messages, 0
+
+        result = [m.copy() for m in messages]
+        pruned = 0
+        prune_boundary = len(result) - protect_tail_count
+
+        for i in range(prune_boundary):
+            msg = result[i]
+            if msg.get("role") != "tool":
+                continue
+            content = msg.get("content", "")
+            if not content or content == _PRUNED_TOOL_PLACEHOLDER:
+                continue
+            # Only prune if the content is substantial (>200 chars)
+            if len(content) > 200:
+                result[i] = {**msg, "content": _PRUNED_TOOL_PLACEHOLDER}
+                pruned += 1
+
+        return result, pruned
+
+    # ------------------------------------------------------------------
+    # Summarization
+    # ------------------------------------------------------------------
+
+    def _compute_summary_budget(self, turns_to_summarize: List[Dict[str, Any]]) -> int:
+        """Scale summary token budget with the amount of content being compressed.
+
+        The maximum scales with the model's context window (5% of context,
+        capped at ``_SUMMARY_TOKENS_CEILING``) so large-context models get
+        richer summaries instead of being hard-capped at 8K tokens.
+        """
+        content_tokens = estimate_messages_tokens_rough(turns_to_summarize)
+        budget = int(content_tokens * _SUMMARY_RATIO)
+        return max(_MIN_SUMMARY_TOKENS, min(budget, self.max_summary_tokens))
+
+    def _serialize_for_summary(self, turns: List[Dict[str, Any]]) -> str:
+        """Serialize conversation turns into labeled text for the summarizer.
+
+        Includes tool call arguments and result content (up to 3000 chars
+        per message) so the summarizer can preserve specific details like
+        file paths, commands, and outputs.
+        """
+        parts = []
+        for msg in turns:
+            role = msg.get("role", "unknown")
+            content = msg.get("content") or ""
+
+            # Tool results: keep more content than before (3000 chars)
+            if role == "tool":
+                tool_id = msg.get("tool_call_id", "")
+                if len(content) > 3000:
+                    content = content[:2000] + "\n...[truncated]...\n" + content[-800:]
+                parts.append(f"[TOOL RESULT {tool_id}]: {content}")
+                continue
+
+            # Assistant messages: include tool call names AND arguments
+            if role == "assistant":
+                if len(content) > 3000:
+                    content = content[:2000] + "\n...[truncated]...\n" + content[-800:]
+                tool_calls = msg.get("tool_calls", [])
+                if tool_calls:
+                    tc_parts = []
+                    for tc in tool_calls:
+                        if isinstance(tc, dict):
+                            fn = tc.get("function", {})
+                            name = fn.get("name", "?")
+                            args = fn.get("arguments", "")
+                            # Truncate long arguments but keep enough for context
+                            if len(args) > 500:
+                                args = args[:400] + "..."
+                            tc_parts.append(f"  {name}({args})")
+                        else:
+                            fn = getattr(tc, "function", None)
+                            name = getattr(fn, "name", "?") if fn else "?"
+                            tc_parts.append(f"  {name}(...)")
+                    content += "\n[Tool calls:\n" + "\n".join(tc_parts) + "\n]"
+                parts.append(f"[ASSISTANT]: {content}")
+                continue
+
+            # User and other roles
+            if len(content) > 3000:
+                content = content[:2000] + "\n...[truncated]...\n" + content[-800:]
+            parts.append(f"[{role.upper()}]: {content}")
+
+        return "\n\n".join(parts)
+
+    def _generate_summary(self, turns_to_summarize: List[Dict[str, Any]]) -> Optional[str]:
+        """Generate a structured summary of conversation turns.
+
+        Uses a structured template (Goal, Progress, Decisions, Files, Next Steps)
+        inspired by Pi-mono and OpenCode. When a previous summary exists,
+        generates an iterative update instead of summarizing from scratch.
+
+        Returns None if all attempts fail — the caller should drop
         the middle turns without a summary rather than inject a useless
         placeholder.
         """
-        parts = []
-        for msg in turns_to_summarize:
-            role = msg.get("role", "unknown")
-            content = msg.get("content") or ""
-            if len(content) > 2000:
-                content = content[:1000] + "\n...[truncated]...\n" + content[-500:]
-            tool_calls = msg.get("tool_calls", [])
-            if tool_calls:
-                tool_names = [tc.get("function", {}).get("name", "?") for tc in tool_calls if isinstance(tc, dict)]
-                content += f"\n[Tool calls: {', '.join(tool_names)}]"
-            parts.append(f"[{role.upper()}]: {content}")
+        summary_budget = self._compute_summary_budget(turns_to_summarize)
+        content_to_summarize = self._serialize_for_summary(turns_to_summarize)
 
-        content_to_summarize = "\n\n".join(parts)
-        prompt = f"""Create a concise handoff summary for a later assistant that will continue this conversation after earlier turns are compacted.
+        if self._previous_summary:
+            # Iterative update: preserve existing info, add new progress
+            prompt = f"""You are updating a context compaction summary. A previous compaction produced the summary below. New conversation turns have occurred since then and need to be incorporated.
 
-Describe:
-1. What actions were taken (tool calls, searches, file operations)
-2. Key information or results obtained
-3. Important decisions, constraints, or user preferences
-4. Relevant data, file names, outputs, or next steps needed to continue
+PREVIOUS SUMMARY:
+{self._previous_summary}
 
-Keep it factual, concise, and focused on helping the next assistant resume without repeating work. Target ~{self.summary_target_tokens} tokens.
+NEW TURNS TO INCORPORATE:
+{content_to_summarize}
+
+Update the summary using this exact structure. PRESERVE all existing information that is still relevant. ADD new progress. Move items from "In Progress" to "Done" when completed. Remove information only if it is clearly obsolete.
+
+## Goal
+[What the user is trying to accomplish — preserve from previous summary, update if goal evolved]
+
+## Constraints & Preferences
+[User preferences, coding style, constraints, important decisions — accumulate across compactions]
+
+## Progress
+### Done
+[Completed work — include specific file paths, commands run, results obtained]
+### In Progress
+[Work currently underway]
+### Blocked
+[Any blockers or issues encountered]
+
+## Key Decisions
+[Important technical decisions and why they were made]
+
+## Relevant Files
+[Files read, modified, or created — with brief note on each. Accumulate across compactions.]
+
+## Next Steps
+[What needs to happen next to continue the work]
+
+## Critical Context
+[Any specific values, error messages, configuration details, or data that would be lost without explicit preservation]
+
+Target ~{summary_budget} tokens. Be specific — include file paths, command outputs, error messages, and concrete values rather than vague descriptions.
+
+Write only the summary body. Do not include any preamble or prefix."""
+        else:
+            # First compaction: summarize from scratch
+            prompt = f"""Create a structured handoff summary for a later assistant that will continue this conversation after earlier turns are compacted.
 
----
 TURNS TO SUMMARIZE:
 {content_to_summarize}
----
 
-Write only the summary body. Do not include any preamble or prefix; the system will add the handoff wrapper."""
+Use this exact structure:
+
+## Goal
+[What the user is trying to accomplish]
+
+## Constraints & Preferences
+[User preferences, coding style, constraints, important decisions]
+
+## Progress
+### Done
+[Completed work — include specific file paths, commands run, results obtained]
+### In Progress
+[Work currently underway]
+### Blocked
+[Any blockers or issues encountered]
+
+## Key Decisions
+[Important technical decisions and why they were made]
+
+## Relevant Files
+[Files read, modified, or created — with brief note on each]
+
+## Next Steps
+[What needs to happen next to continue the work]
+
+## Critical Context
+[Any specific values, error messages, configuration details, or data that would be lost without explicit preservation]
+
+Target ~{summary_budget} tokens. Be specific — include file paths, command outputs, error messages, and concrete values rather than vague descriptions. The goal is to prevent the next assistant from repeating work or losing important details.
+
+Write only the summary body. Do not include any preamble or prefix."""
 
-        # Use the centralized LLM router — handles provider resolution,
-        # auth, and fallback internally.
         try:
             call_kwargs = {
                 "task": "compression",
                 "messages": [{"role": "user", "content": prompt}],
                 "temperature": 0.3,
-                "max_tokens": self.summary_target_tokens * 2,
-                "timeout": 30.0,
+                "max_tokens": summary_budget * 2,
+                # timeout resolved from auxiliary.compression.timeout config by call_llm
             }
             if self.summary_model:
                 call_kwargs["model"] = self.summary_model
@@ -147,6 +357,8 @@ Write only the summary body. Do not include any preamble or prefix; the system w
             if not isinstance(content, str):
                 content = str(content) if content else ""
             summary = content.strip()
+            # Store for iterative updates on next compaction
+            self._previous_summary = summary
             return self._with_summary_prefix(summary)
         except RuntimeError:
             logging.warning("Context compression: no provider available for "
@@ -251,56 +463,155 @@ Write only the summary body. Do not include any preamble or prefix; the system w
         """Pull a compress-end boundary backward to avoid splitting a
         tool_call / result group.
 
-        If the message just before ``idx`` is an assistant message with
-        tool_calls, those tool results will start at ``idx`` and would be
-        separated from their parent.  Move backwards to include the whole
-        group in the summarised region.
+        If the boundary falls in the middle of a tool-result group (i.e.
+        there are consecutive tool messages before ``idx``), walk backward
+        past all of them to find the parent assistant message.  If found,
+        move the boundary before the assistant so the entire
+        assistant + tool_results group is included in the summarised region
+        rather than being split (which causes silent data loss when
+        ``_sanitize_tool_pairs`` removes the orphaned tail results).
         """
         if idx <= 0 or idx >= len(messages):
             return idx
-        prev = messages[idx - 1]
-        if prev.get("role") == "assistant" and prev.get("tool_calls"):
-            # The results for this assistant turn sit at idx..idx+k.
-            # Include the assistant message in the summarised region too.
-            idx -= 1
+        # Walk backward past consecutive tool results
+        check = idx - 1
+        while check >= 0 and messages[check].get("role") == "tool":
+            check -= 1
+        # If we landed on the parent assistant with tool_calls, pull the
+        # boundary before it so the whole group gets summarised together.
+        if check >= 0 and messages[check].get("role") == "assistant" and messages[check].get("tool_calls"):
+            idx = check
         return idx
 
+    # ------------------------------------------------------------------
+    # Tail protection by token budget
+    # ------------------------------------------------------------------
+
+    def _find_tail_cut_by_tokens(
+        self, messages: List[Dict[str, Any]], head_end: int,
+        token_budget: int | None = None,
+    ) -> int:
+        """Walk backward from the end of messages, accumulating tokens until
+        the budget is reached. Returns the index where the tail starts.
+
+        ``token_budget`` defaults to ``self.tail_token_budget`` which is
+        derived from ``summary_target_ratio * context_length``, so it
+        scales automatically with the model's context window.
+
+        Never cuts inside a tool_call/result group. Falls back to the old
+        ``protect_last_n`` if the budget would protect fewer messages.
+        """
+        if token_budget is None:
+            token_budget = self.tail_token_budget
+        n = len(messages)
+        min_tail = self.protect_last_n
+        accumulated = 0
+        cut_idx = n  # start from beyond the end
+
+        for i in range(n - 1, head_end - 1, -1):
+            msg = messages[i]
+            content = msg.get("content") or ""
+            msg_tokens = len(content) // _CHARS_PER_TOKEN + 10  # +10 for role/metadata
+            # Include tool call arguments in estimate
+            for tc in msg.get("tool_calls") or []:
+                if isinstance(tc, dict):
+                    args = tc.get("function", {}).get("arguments", "")
+                    msg_tokens += len(args) // _CHARS_PER_TOKEN
+            if accumulated + msg_tokens > token_budget and (n - i) >= min_tail:
+                break
+            accumulated += msg_tokens
+            cut_idx = i
+
+        # Ensure we protect at least protect_last_n messages
+        fallback_cut = n - min_tail
+        if cut_idx > fallback_cut:
+            cut_idx = fallback_cut
+
+        # If the token budget would protect everything (small conversations),
+        # fall back to the fixed protect_last_n approach so compression can
+        # still remove middle turns.
+        if cut_idx <= head_end:
+            cut_idx = fallback_cut
+
+        # Align to avoid splitting tool groups
+        cut_idx = self._align_boundary_backward(messages, cut_idx)
+
+        return max(cut_idx, head_end + 1)
+
+    # ------------------------------------------------------------------
+    # Main compression entry point
+    # ------------------------------------------------------------------
+
     def compress(self, messages: List[Dict[str, Any]], current_tokens: int = None) -> List[Dict[str, Any]]:
         """Compress conversation messages by summarizing middle turns.
 
-        Keeps first N + last N turns, summarizes everything in between.
+        Algorithm:
+          1. Prune old tool results (cheap pre-pass, no LLM call)
+          2. Protect head messages (system prompt + first exchange)
+          3. Find tail boundary by token budget (~20K tokens of recent context)
+          4. Summarize middle turns with structured LLM prompt
+          5. On re-compression, iteratively update the previous summary
+
         After compression, orphaned tool_call / tool_result pairs are cleaned
         up so the API never receives mismatched IDs.
         """
         n_messages = len(messages)
         if n_messages <= self.protect_first_n + self.protect_last_n + 1:
             if not self.quiet_mode:
-                print(f"⚠️  Cannot compress: only {n_messages} messages (need > {self.protect_first_n + self.protect_last_n + 1})")
+                logger.warning(
+                    "Cannot compress: only %d messages (need > %d)",
+                    n_messages,
+                    self.protect_first_n + self.protect_last_n + 1,
+                )
             return messages
 
+        display_tokens = current_tokens if current_tokens else self.last_prompt_tokens or estimate_messages_tokens_rough(messages)
+
+        # Phase 1: Prune old tool results (cheap, no LLM call)
+        messages, pruned_count = self._prune_old_tool_results(
+            messages, protect_tail_count=self.protect_last_n * 3,
+        )
+        if pruned_count and not self.quiet_mode:
+            logger.info("Pre-compression: pruned %d old tool result(s)", pruned_count)
+
+        # Phase 2: Determine boundaries
         compress_start = self.protect_first_n
-        compress_end = n_messages - self.protect_last_n
-        if compress_start >= compress_end:
-            return messages
-
-        # Adjust boundaries to avoid splitting tool_call/result groups.
         compress_start = self._align_boundary_forward(messages, compress_start)
-        compress_end = self._align_boundary_backward(messages, compress_end)
+
+        # Use token-budget tail protection instead of fixed message count
+        compress_end = self._find_tail_cut_by_tokens(messages, compress_start)
+
         if compress_start >= compress_end:
             return messages
 
         turns_to_summarize = messages[compress_start:compress_end]
-        display_tokens = current_tokens if current_tokens else self.last_prompt_tokens or estimate_messages_tokens_rough(messages)
 
         if not self.quiet_mode:
-            print(f"\n📦 Context compression triggered ({display_tokens:,} tokens ≥ {self.threshold_tokens:,} threshold)")
-            print(f"   📊 Model context limit: {self.context_length:,} tokens ({self.threshold_percent*100:.0f}% = {self.threshold_tokens:,})")
-
-        if not self.quiet_mode:
-            print(f"   🗜️  Summarizing turns {compress_start+1}-{compress_end} ({len(turns_to_summarize)} turns)")
+            logger.info(
+                "Context compression triggered (%d tokens >= %d threshold)",
+                display_tokens,
+                self.threshold_tokens,
+            )
+            logger.info(
+                "Model context limit: %d tokens (%.0f%% = %d)",
+                self.context_length,
+                self.threshold_percent * 100,
+                self.threshold_tokens,
+            )
+            tail_msgs = n_messages - compress_end
+            logger.info(
+                "Summarizing turns %d-%d (%d turns), protecting %d head + %d tail messages",
+                compress_start + 1,
+                compress_end,
+                len(turns_to_summarize),
+                compress_start,
+                tail_msgs,
+            )
 
+        # Phase 3: Generate structured summary
         summary = self._generate_summary(turns_to_summarize)
 
+        # Phase 4: Assemble compressed message list
         compressed = []
         for i in range(compress_start):
             msg = messages[i].copy()
@@ -311,16 +622,41 @@ Write only the summary body. Do not include any preamble or prefix; the system w
                 )
             compressed.append(msg)
 
+        _merge_summary_into_tail = False
         if summary:
             last_head_role = messages[compress_start - 1].get("role", "user") if compress_start > 0 else "user"
-            summary_role = "user" if last_head_role in ("assistant", "tool") else "assistant"
-            compressed.append({"role": summary_role, "content": summary})
+            first_tail_role = messages[compress_end].get("role", "user") if compress_end < n_messages else "user"
+            # Pick a role that avoids consecutive same-role with both neighbors.
+            # Priority: avoid colliding with head (already committed), then tail.
+            if last_head_role in ("assistant", "tool"):
+                summary_role = "user"
+            else:
+                summary_role = "assistant"
+            # If the chosen role collides with the tail AND flipping wouldn't
+            # collide with the head, flip it.
+            if summary_role == first_tail_role:
+                flipped = "assistant" if summary_role == "user" else "user"
+                if flipped != last_head_role:
+                    summary_role = flipped
+                else:
+                    # Both roles would create consecutive same-role messages
+                    # (e.g. head=assistant, tail=user — neither role works).
+                    # Merge the summary into the first tail message instead
+                    # of inserting a standalone message that breaks alternation.
+                    _merge_summary_into_tail = True
+            if not _merge_summary_into_tail:
+                compressed.append({"role": summary_role, "content": summary})
         else:
             if not self.quiet_mode:
-                print("   ⚠️  No summary model available — middle turns dropped without summary")
+                logger.warning("No summary model available — middle turns dropped without summary")
 
         for i in range(compress_end, n_messages):
-            compressed.append(messages[i].copy())
+            msg = messages[i].copy()
+            if _merge_summary_into_tail and i == compress_end:
+                original = msg.get("content") or ""
+                msg["content"] = summary + "\n\n" + original
+                _merge_summary_into_tail = False
+            compressed.append(msg)
 
         self.compression_count += 1
 
@@ -329,7 +665,12 @@ Write only the summary body. Do not include any preamble or prefix; the system w
         if not self.quiet_mode:
             new_estimate = estimate_messages_tokens_rough(compressed)
             saved_estimate = display_tokens - new_estimate
-            print(f"   ✅ Compressed: {n_messages} → {len(compressed)} messages (~{saved_estimate:,} tokens saved)")
-            print(f"   💡 Compression #{self.compression_count} complete")
+            logger.info(
+                "Compressed: %d -> %d messages (~%d tokens saved)",
+                n_messages,
+                len(compressed),
+                saved_estimate,
+            )
+            logger.info("Compression #%d complete", self.compression_count)
 
         return compressed

agent/context_references.py

@@ -0,0 +1,492 @@
+from __future__ import annotations
+
+import asyncio
+import inspect
+import json
+import mimetypes
+import os
+import re
+import subprocess
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Awaitable, Callable
+
+from agent.model_metadata import estimate_tokens_rough
+
+REFERENCE_PATTERN = re.compile(
+    r"(?<![\w/])@(?:(?P<simple>diff|staged)\b|(?P<kind>file|folder|git|url):(?P<value>\S+))"
+)
+TRAILING_PUNCTUATION = ",.;!?"
+_SENSITIVE_HOME_DIRS = (".ssh", ".aws", ".gnupg", ".kube", ".docker", ".azure", ".config/gh")
+_SENSITIVE_HERMES_DIRS = (Path("skills") / ".hub",)
+_SENSITIVE_HOME_FILES = (
+    Path(".ssh") / "authorized_keys",
+    Path(".ssh") / "id_rsa",
+    Path(".ssh") / "id_ed25519",
+    Path(".ssh") / "config",
+    Path(".bashrc"),
+    Path(".zshrc"),
+    Path(".profile"),
+    Path(".bash_profile"),
+    Path(".zprofile"),
+    Path(".netrc"),
+    Path(".pgpass"),
+    Path(".npmrc"),
+    Path(".pypirc"),
+)
+
+
+@dataclass(frozen=True)
+class ContextReference:
+    raw: str
+    kind: str
+    target: str
+    start: int
+    end: int
+    line_start: int | None = None
+    line_end: int | None = None
+
+
+@dataclass
+class ContextReferenceResult:
+    message: str
+    original_message: str
+    references: list[ContextReference] = field(default_factory=list)
+    warnings: list[str] = field(default_factory=list)
+    injected_tokens: int = 0
+    expanded: bool = False
+    blocked: bool = False
+
+
+def parse_context_references(message: str) -> list[ContextReference]:
+    refs: list[ContextReference] = []
+    if not message:
+        return refs
+
+    for match in REFERENCE_PATTERN.finditer(message):
+        simple = match.group("simple")
+        if simple:
+            refs.append(
+                ContextReference(
+                    raw=match.group(0),
+                    kind=simple,
+                    target="",
+                    start=match.start(),
+                    end=match.end(),
+                )
+            )
+            continue
+
+        kind = match.group("kind")
+        value = _strip_trailing_punctuation(match.group("value") or "")
+        line_start = None
+        line_end = None
+        target = value
+
+        if kind == "file":
+            range_match = re.match(r"^(?P<path>.+?):(?P<start>\d+)(?:-(?P<end>\d+))?$", value)
+            if range_match:
+                target = range_match.group("path")
+                line_start = int(range_match.group("start"))
+                line_end = int(range_match.group("end") or range_match.group("start"))
+
+        refs.append(
+            ContextReference(
+                raw=match.group(0),
+                kind=kind,
+                target=target,
+                start=match.start(),
+                end=match.end(),
+                line_start=line_start,
+                line_end=line_end,
+            )
+        )
+
+    return refs
+
+
+def preprocess_context_references(
+    message: str,
+    *,
+    cwd: str | Path,
+    context_length: int,
+    url_fetcher: Callable[[str], str | Awaitable[str]] | None = None,
+    allowed_root: str | Path | None = None,
+) -> ContextReferenceResult:
+    coro = preprocess_context_references_async(
+        message,
+        cwd=cwd,
+        context_length=context_length,
+        url_fetcher=url_fetcher,
+        allowed_root=allowed_root,
+    )
+    # Safe for both CLI (no loop) and gateway (loop already running).
+    try:
+        loop = asyncio.get_running_loop()
+    except RuntimeError:
+        loop = None
+    if loop and loop.is_running():
+        import concurrent.futures
+        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
+            return pool.submit(asyncio.run, coro).result()
+    return asyncio.run(coro)
+
+
+async def preprocess_context_references_async(
+    message: str,
+    *,
+    cwd: str | Path,
+    context_length: int,
+    url_fetcher: Callable[[str], str | Awaitable[str]] | None = None,
+    allowed_root: str | Path | None = None,
+) -> ContextReferenceResult:
+    refs = parse_context_references(message)
+    if not refs:
+        return ContextReferenceResult(message=message, original_message=message)
+
+    cwd_path = Path(cwd).expanduser().resolve()
+    # Default to the current working directory so @ references cannot escape
+    # the active workspace unless a caller explicitly widens the root.
+    allowed_root_path = (
+        Path(allowed_root).expanduser().resolve() if allowed_root is not None else cwd_path
+    )
+    warnings: list[str] = []
+    blocks: list[str] = []
+    injected_tokens = 0
+
+    for ref in refs:
+        warning, block = await _expand_reference(
+            ref,
+            cwd_path,
+            url_fetcher=url_fetcher,
+            allowed_root=allowed_root_path,
+        )
+        if warning:
+            warnings.append(warning)
+        if block:
+            blocks.append(block)
+            injected_tokens += estimate_tokens_rough(block)
+
+    hard_limit = max(1, int(context_length * 0.50))
+    soft_limit = max(1, int(context_length * 0.25))
+    if injected_tokens > hard_limit:
+        warnings.append(
+            f"@ context injection refused: {injected_tokens} tokens exceeds the 50% hard limit ({hard_limit})."
+        )
+        return ContextReferenceResult(
+            message=message,
+            original_message=message,
+            references=refs,
+            warnings=warnings,
+            injected_tokens=injected_tokens,
+            expanded=False,
+            blocked=True,
+        )
+
+    if injected_tokens > soft_limit:
+        warnings.append(
+            f"@ context injection warning: {injected_tokens} tokens exceeds the 25% soft limit ({soft_limit})."
+        )
+
+    stripped = _remove_reference_tokens(message, refs)
+    final = stripped
+    if warnings:
+        final = f"{final}\n\n--- Context Warnings ---\n" + "\n".join(f"- {warning}" for warning in warnings)
+    if blocks:
+        final = f"{final}\n\n--- Attached Context ---\n\n" + "\n\n".join(blocks)
+
+    return ContextReferenceResult(
+        message=final.strip(),
+        original_message=message,
+        references=refs,
+        warnings=warnings,
+        injected_tokens=injected_tokens,
+        expanded=bool(blocks or warnings),
+        blocked=False,
+    )
+
+
+async def _expand_reference(
+    ref: ContextReference,
+    cwd: Path,
+    *,
+    url_fetcher: Callable[[str], str | Awaitable[str]] | None = None,
+    allowed_root: Path | None = None,
+) -> tuple[str | None, str | None]:
+    try:
+        if ref.kind == "file":
+            return _expand_file_reference(ref, cwd, allowed_root=allowed_root)
+        if ref.kind == "folder":
+            return _expand_folder_reference(ref, cwd, allowed_root=allowed_root)
+        if ref.kind == "diff":
+            return _expand_git_reference(ref, cwd, ["diff"], "git diff")
+        if ref.kind == "staged":
+            return _expand_git_reference(ref, cwd, ["diff", "--staged"], "git diff --staged")
+        if ref.kind == "git":
+            count = max(1, min(int(ref.target or "1"), 10))
+            return _expand_git_reference(ref, cwd, ["log", f"-{count}", "-p"], f"git log -{count} -p")
+        if ref.kind == "url":
+            content = await _fetch_url_content(ref.target, url_fetcher=url_fetcher)
+            if not content:
+                return f"{ref.raw}: no content extracted", None
+            return None, f"🌐 {ref.raw} ({estimate_tokens_rough(content)} tokens)\n{content}"
+    except Exception as exc:
+        return f"{ref.raw}: {exc}", None
+
+    return f"{ref.raw}: unsupported reference type", None
+
+
+def _expand_file_reference(
+    ref: ContextReference,
+    cwd: Path,
+    *,
+    allowed_root: Path | None = None,
+) -> tuple[str | None, str | None]:
+    path = _resolve_path(cwd, ref.target, allowed_root=allowed_root)
+    _ensure_reference_path_allowed(path)
+    if not path.exists():
+        return f"{ref.raw}: file not found", None
+    if not path.is_file():
+        return f"{ref.raw}: path is not a file", None
+    if _is_binary_file(path):
+        return f"{ref.raw}: binary files are not supported", None
+
+    text = path.read_text(encoding="utf-8")
+    if ref.line_start is not None:
+        lines = text.splitlines()
+        start_idx = max(ref.line_start - 1, 0)
+        end_idx = min(ref.line_end or ref.line_start, len(lines))
+        text = "\n".join(lines[start_idx:end_idx])
+
+    lang = _code_fence_language(path)
+    label = ref.raw
+    return None, f"📄 {label} ({estimate_tokens_rough(text)} tokens)\n```{lang}\n{text}\n```"
+
+
+def _expand_folder_reference(
+    ref: ContextReference,
+    cwd: Path,
+    *,
+    allowed_root: Path | None = None,
+) -> tuple[str | None, str | None]:
+    path = _resolve_path(cwd, ref.target, allowed_root=allowed_root)
+    _ensure_reference_path_allowed(path)
+    if not path.exists():
+        return f"{ref.raw}: folder not found", None
+    if not path.is_dir():
+        return f"{ref.raw}: path is not a folder", None
+
+    listing = _build_folder_listing(path, cwd)
+    return None, f"📁 {ref.raw} ({estimate_tokens_rough(listing)} tokens)\n{listing}"
+
+
+def _expand_git_reference(
+    ref: ContextReference,
+    cwd: Path,
+    args: list[str],
+    label: str,
+) -> tuple[str | None, str | None]:
+    try:
+        result = subprocess.run(
+            ["git", *args],
+            cwd=cwd,
+            capture_output=True,
+            text=True,
+            timeout=30,
+        )
+    except subprocess.TimeoutExpired:
+        return f"{ref.raw}: git command timed out (30s)", None
+    if result.returncode != 0:
+        stderr = (result.stderr or "").strip() or "git command failed"
+        return f"{ref.raw}: {stderr}", None
+    content = result.stdout.strip()
+    if not content:
+        content = "(no output)"
+    return None, f"🧾 {label} ({estimate_tokens_rough(content)} tokens)\n```diff\n{content}\n```"
+
+
+async def _fetch_url_content(
+    url: str,
+    *,
+    url_fetcher: Callable[[str], str | Awaitable[str]] | None = None,
+) -> str:
+    fetcher = url_fetcher or _default_url_fetcher
+    content = fetcher(url)
+    if inspect.isawaitable(content):
+        content = await content
+    return str(content or "").strip()
+
+
+async def _default_url_fetcher(url: str) -> str:
+    from tools.web_tools import web_extract_tool
+
+    raw = await web_extract_tool([url], format="markdown", use_llm_processing=True)
+    payload = json.loads(raw)
+    docs = payload.get("data", {}).get("documents", [])
+    if not docs:
+        return ""
+    doc = docs[0]
+    return str(doc.get("content") or doc.get("raw_content") or "").strip()
+
+
+def _resolve_path(cwd: Path, target: str, *, allowed_root: Path | None = None) -> Path:
+    path = Path(os.path.expanduser(target))
+    if not path.is_absolute():
+        path = cwd / path
+    resolved = path.resolve()
+    if allowed_root is not None:
+        try:
+            resolved.relative_to(allowed_root)
+        except ValueError as exc:
+            raise ValueError("path is outside the allowed workspace") from exc
+    return resolved
+
+
+def _ensure_reference_path_allowed(path: Path) -> None:
+    home = Path(os.path.expanduser("~")).resolve()
+    hermes_home = Path(
+        os.getenv("HERMES_HOME", str(home / ".hermes"))
+    ).expanduser().resolve()
+
+    blocked_exact = {home / rel for rel in _SENSITIVE_HOME_FILES}
+    blocked_exact.add(hermes_home / ".env")
+    blocked_dirs = [home / rel for rel in _SENSITIVE_HOME_DIRS]
+    blocked_dirs.extend(hermes_home / rel for rel in _SENSITIVE_HERMES_DIRS)
+
+    if path in blocked_exact:
+        raise ValueError("path is a sensitive credential file and cannot be attached")
+
+    for blocked_dir in blocked_dirs:
+        try:
+            path.relative_to(blocked_dir)
+        except ValueError:
+            continue
+        raise ValueError("path is a sensitive credential or internal Hermes path and cannot be attached")
+
+
+def _strip_trailing_punctuation(value: str) -> str:
+    stripped = value.rstrip(TRAILING_PUNCTUATION)
+    while stripped.endswith((")", "]", "}")):
+        closer = stripped[-1]
+        opener = {")": "(", "]": "[", "}": "{"}[closer]
+        if stripped.count(closer) > stripped.count(opener):
+            stripped = stripped[:-1]
+            continue
+        break
+    return stripped
+
+
+def _remove_reference_tokens(message: str, refs: list[ContextReference]) -> str:
+    pieces: list[str] = []
+    cursor = 0
+    for ref in refs:
+        pieces.append(message[cursor:ref.start])
+        cursor = ref.end
+    pieces.append(message[cursor:])
+    text = "".join(pieces)
+    text = re.sub(r"\s{2,}", " ", text)
+    text = re.sub(r"\s+([,.;:!?])", r"\1", text)
+    return text.strip()
+
+
+def _is_binary_file(path: Path) -> bool:
+    mime, _ = mimetypes.guess_type(path.name)
+    if mime and not mime.startswith("text/") and not any(
+        path.name.endswith(ext) for ext in (".py", ".md", ".txt", ".json", ".yaml", ".yml", ".toml", ".js", ".ts")
+    ):
+        return True
+    chunk = path.read_bytes()[:4096]
+    return b"\x00" in chunk
+
+
+def _build_folder_listing(path: Path, cwd: Path, limit: int = 200) -> str:
+    lines = [f"{path.relative_to(cwd)}/"]
+    entries = _iter_visible_entries(path, cwd, limit=limit)
+    for entry in entries:
+        rel = entry.relative_to(cwd)
+        indent = "  " * max(len(rel.parts) - len(path.relative_to(cwd).parts) - 1, 0)
+        if entry.is_dir():
+            lines.append(f"{indent}- {entry.name}/")
+        else:
+            meta = _file_metadata(entry)
+            lines.append(f"{indent}- {entry.name} ({meta})")
+    if len(entries) >= limit:
+        lines.append("- ...")
+    return "\n".join(lines)
+
+
+def _iter_visible_entries(path: Path, cwd: Path, limit: int) -> list[Path]:
+    rg_entries = _rg_files(path, cwd, limit=limit)
+    if rg_entries is not None:
+        output: list[Path] = []
+        seen_dirs: set[Path] = set()
+        for rel in rg_entries:
+            full = cwd / rel
+            for parent in full.parents:
+                if parent == cwd or parent in seen_dirs or path not in {parent, *parent.parents}:
+                    continue
+                seen_dirs.add(parent)
+                output.append(parent)
+            output.append(full)
+        return sorted({p for p in output if p.exists()}, key=lambda p: (not p.is_dir(), str(p)))
+
+    output = []
+    for root, dirs, files in os.walk(path):
+        dirs[:] = sorted(d for d in dirs if not d.startswith(".") and d != "__pycache__")
+        files = sorted(f for f in files if not f.startswith("."))
+        root_path = Path(root)
+        for d in dirs:
+            output.append(root_path / d)
+            if len(output) >= limit:
+                return output
+        for f in files:
+            output.append(root_path / f)
+            if len(output) >= limit:
+                return output
+    return output
+
+
+def _rg_files(path: Path, cwd: Path, limit: int) -> list[Path] | None:
+    try:
+        result = subprocess.run(
+            ["rg", "--files", str(path.relative_to(cwd))],
+            cwd=cwd,
+            capture_output=True,
+            text=True,
+            timeout=10,
+        )
+    except FileNotFoundError:
+        return None
+    except subprocess.TimeoutExpired:
+        return None
+    if result.returncode != 0:
+        return None
+    files = [Path(line.strip()) for line in result.stdout.splitlines() if line.strip()]
+    return files[:limit]
+
+
+def _file_metadata(path: Path) -> str:
+    if _is_binary_file(path):
+        return f"{path.stat().st_size} bytes"
+    try:
+        line_count = path.read_text(encoding="utf-8").count("\n") + 1
+    except Exception:
+        return f"{path.stat().st_size} bytes"
+    return f"{line_count} lines"
+
+
+def _code_fence_language(path: Path) -> str:
+    mapping = {
+        ".py": "python",
+        ".js": "javascript",
+        ".ts": "typescript",
+        ".tsx": "tsx",
+        ".jsx": "jsx",
+        ".json": "json",
+        ".md": "markdown",
+        ".sh": "bash",
+        ".yml": "yaml",
+        ".yaml": "yaml",
+        ".toml": "toml",
+    }
+    return mapping.get(path.suffix.lower(), "")

agent/copilot_acp_client.py

@@ -0,0 +1,447 @@
+"""OpenAI-compatible shim that forwards Hermes requests to `copilot --acp`.
+
+This adapter lets Hermes treat the GitHub Copilot ACP server as a chat-style
+backend. Each request starts a short-lived ACP session, sends the formatted
+conversation as a single prompt, collects text chunks, and converts the result
+back into the minimal shape Hermes expects from an OpenAI client.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import queue
+import shlex
+import subprocess
+import threading
+import time
+from collections import deque
+from pathlib import Path
+from types import SimpleNamespace
+from typing import Any
+
+ACP_MARKER_BASE_URL = "acp://copilot"
+_DEFAULT_TIMEOUT_SECONDS = 900.0
+
+
+def _resolve_command() -> str:
+    return (
+        os.getenv("HERMES_COPILOT_ACP_COMMAND", "").strip()
+        or os.getenv("COPILOT_CLI_PATH", "").strip()
+        or "copilot"
+    )
+
+
+def _resolve_args() -> list[str]:
+    raw = os.getenv("HERMES_COPILOT_ACP_ARGS", "").strip()
+    if not raw:
+        return ["--acp", "--stdio"]
+    return shlex.split(raw)
+
+
+def _jsonrpc_error(message_id: Any, code: int, message: str) -> dict[str, Any]:
+    return {
+        "jsonrpc": "2.0",
+        "id": message_id,
+        "error": {
+            "code": code,
+            "message": message,
+        },
+    }
+
+
+def _format_messages_as_prompt(messages: list[dict[str, Any]], model: str | None = None) -> str:
+    sections: list[str] = [
+        "You are being used as the active ACP agent backend for Hermes.",
+        "Use your own ACP capabilities and respond directly in natural language.",
+        "Do not emit OpenAI tool-call JSON.",
+    ]
+    if model:
+        sections.append(f"Hermes requested model hint: {model}")
+
+    transcript: list[str] = []
+    for message in messages:
+        if not isinstance(message, dict):
+            continue
+        role = str(message.get("role") or "unknown").strip().lower()
+        if role == "tool":
+            role = "tool"
+        elif role not in {"system", "user", "assistant"}:
+            role = "context"
+
+        content = message.get("content")
+        rendered = _render_message_content(content)
+        if not rendered:
+            continue
+
+        label = {
+            "system": "System",
+            "user": "User",
+            "assistant": "Assistant",
+            "tool": "Tool",
+            "context": "Context",
+        }.get(role, role.title())
+        transcript.append(f"{label}:\n{rendered}")
+
+    if transcript:
+        sections.append("Conversation transcript:\n\n" + "\n\n".join(transcript))
+
+    sections.append("Continue the conversation from the latest user request.")
+    return "\n\n".join(section.strip() for section in sections if section and section.strip())
+
+
+def _render_message_content(content: Any) -> str:
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content.strip()
+    if isinstance(content, dict):
+        if "text" in content:
+            return str(content.get("text") or "").strip()
+        if "content" in content and isinstance(content.get("content"), str):
+            return str(content.get("content") or "").strip()
+        return json.dumps(content, ensure_ascii=True)
+    if isinstance(content, list):
+        parts: list[str] = []
+        for item in content:
+            if isinstance(item, str):
+                parts.append(item)
+            elif isinstance(item, dict):
+                text = item.get("text")
+                if isinstance(text, str) and text.strip():
+                    parts.append(text.strip())
+        return "\n".join(parts).strip()
+    return str(content).strip()
+
+
+def _ensure_path_within_cwd(path_text: str, cwd: str) -> Path:
+    candidate = Path(path_text)
+    if not candidate.is_absolute():
+        raise PermissionError("ACP file-system paths must be absolute.")
+    resolved = candidate.resolve()
+    root = Path(cwd).resolve()
+    try:
+        resolved.relative_to(root)
+    except ValueError as exc:
+        raise PermissionError(f"Path '{resolved}' is outside the session cwd '{root}'.") from exc
+    return resolved
+
+
+class _ACPChatCompletions:
+    def __init__(self, client: "CopilotACPClient"):
+        self._client = client
+
+    def create(self, **kwargs: Any) -> Any:
+        return self._client._create_chat_completion(**kwargs)
+
+
+class _ACPChatNamespace:
+    def __init__(self, client: "CopilotACPClient"):
+        self.completions = _ACPChatCompletions(client)
+
+
+class CopilotACPClient:
+    """Minimal OpenAI-client-compatible facade for Copilot ACP."""
+
+    def __init__(
+        self,
+        *,
+        api_key: str | None = None,
+        base_url: str | None = None,
+        default_headers: dict[str, str] | None = None,
+        acp_command: str | None = None,
+        acp_args: list[str] | None = None,
+        acp_cwd: str | None = None,
+        command: str | None = None,
+        args: list[str] | None = None,
+        **_: Any,
+    ):
+        self.api_key = api_key or "copilot-acp"
+        self.base_url = base_url or ACP_MARKER_BASE_URL
+        self._default_headers = dict(default_headers or {})
+        self._acp_command = acp_command or command or _resolve_command()
+        self._acp_args = list(acp_args or args or _resolve_args())
+        self._acp_cwd = str(Path(acp_cwd or os.getcwd()).resolve())
+        self.chat = _ACPChatNamespace(self)
+        self.is_closed = False
+        self._active_process: subprocess.Popen[str] | None = None
+        self._active_process_lock = threading.Lock()
+
+    def close(self) -> None:
+        proc: subprocess.Popen[str] | None
+        with self._active_process_lock:
+            proc = self._active_process
+            self._active_process = None
+        self.is_closed = True
+        if proc is None:
+            return
+        try:
+            proc.terminate()
+            proc.wait(timeout=2)
+        except Exception:
+            try:
+                proc.kill()
+            except Exception:
+                pass
+
+    def _create_chat_completion(
+        self,
+        *,
+        model: str | None = None,
+        messages: list[dict[str, Any]] | None = None,
+        timeout: float | None = None,
+        **_: Any,
+    ) -> Any:
+        prompt_text = _format_messages_as_prompt(messages or [], model=model)
+        response_text, reasoning_text = self._run_prompt(
+            prompt_text,
+            timeout_seconds=float(timeout or _DEFAULT_TIMEOUT_SECONDS),
+        )
+
+        usage = SimpleNamespace(
+            prompt_tokens=0,
+            completion_tokens=0,
+            total_tokens=0,
+            prompt_tokens_details=SimpleNamespace(cached_tokens=0),
+        )
+        assistant_message = SimpleNamespace(
+            content=response_text,
+            tool_calls=[],
+            reasoning=reasoning_text or None,
+            reasoning_content=reasoning_text or None,
+            reasoning_details=None,
+        )
+        choice = SimpleNamespace(message=assistant_message, finish_reason="stop")
+        return SimpleNamespace(
+            choices=[choice],
+            usage=usage,
+            model=model or "copilot-acp",
+        )
+
+    def _run_prompt(self, prompt_text: str, *, timeout_seconds: float) -> tuple[str, str]:
+        try:
+            proc = subprocess.Popen(
+                [self._acp_command] + self._acp_args,
+                stdin=subprocess.PIPE,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.PIPE,
+                text=True,
+                bufsize=1,
+                cwd=self._acp_cwd,
+            )
+        except FileNotFoundError as exc:
+            raise RuntimeError(
+                f"Could not start Copilot ACP command '{self._acp_command}'. "
+                "Install GitHub Copilot CLI or set HERMES_COPILOT_ACP_COMMAND/COPILOT_CLI_PATH."
+            ) from exc
+
+        if proc.stdin is None or proc.stdout is None:
+            proc.kill()
+            raise RuntimeError("Copilot ACP process did not expose stdin/stdout pipes.")
+
+        self.is_closed = False
+        with self._active_process_lock:
+            self._active_process = proc
+
+        inbox: queue.Queue[dict[str, Any]] = queue.Queue()
+        stderr_tail: deque[str] = deque(maxlen=40)
+
+        def _stdout_reader() -> None:
+            for line in proc.stdout:
+                try:
+                    inbox.put(json.loads(line))
+                except Exception:
+                    inbox.put({"raw": line.rstrip("\n")})
+
+        def _stderr_reader() -> None:
+            if proc.stderr is None:
+                return
+            for line in proc.stderr:
+                stderr_tail.append(line.rstrip("\n"))
+
+        out_thread = threading.Thread(target=_stdout_reader, daemon=True)
+        err_thread = threading.Thread(target=_stderr_reader, daemon=True)
+        out_thread.start()
+        err_thread.start()
+
+        next_id = 0
+
+        def _request(method: str, params: dict[str, Any], *, text_parts: list[str] | None = None, reasoning_parts: list[str] | None = None) -> Any:
+            nonlocal next_id
+            next_id += 1
+            request_id = next_id
+            payload = {
+                "jsonrpc": "2.0",
+                "id": request_id,
+                "method": method,
+                "params": params,
+            }
+            proc.stdin.write(json.dumps(payload) + "\n")
+            proc.stdin.flush()
+
+            deadline = time.time() + timeout_seconds
+            while time.time() < deadline:
+                if proc.poll() is not None:
+                    break
+                try:
+                    msg = inbox.get(timeout=0.1)
+                except queue.Empty:
+                    continue
+
+                if self._handle_server_message(
+                    msg,
+                    process=proc,
+                    cwd=self._acp_cwd,
+                    text_parts=text_parts,
+                    reasoning_parts=reasoning_parts,
+                ):
+                    continue
+
+                if msg.get("id") != request_id:
+                    continue
+                if "error" in msg:
+                    err = msg.get("error") or {}
+                    raise RuntimeError(
+                        f"Copilot ACP {method} failed: {err.get('message') or err}"
+                    )
+                return msg.get("result")
+
+            stderr_text = "\n".join(stderr_tail).strip()
+            if proc.poll() is not None and stderr_text:
+                raise RuntimeError(f"Copilot ACP process exited early: {stderr_text}")
+            raise TimeoutError(f"Timed out waiting for Copilot ACP response to {method}.")
+
+        try:
+            _request(
+                "initialize",
+                {
+                    "protocolVersion": 1,
+                    "clientCapabilities": {
+                        "fs": {
+                            "readTextFile": True,
+                            "writeTextFile": True,
+                        }
+                    },
+                    "clientInfo": {
+                        "name": "hermes-agent",
+                        "title": "Hermes Agent",
+                        "version": "0.0.0",
+                    },
+                },
+            )
+            session = _request(
+                "session/new",
+                {
+                    "cwd": self._acp_cwd,
+                    "mcpServers": [],
+                },
+            ) or {}
+            session_id = str(session.get("sessionId") or "").strip()
+            if not session_id:
+                raise RuntimeError("Copilot ACP did not return a sessionId.")
+
+            text_parts: list[str] = []
+            reasoning_parts: list[str] = []
+            _request(
+                "session/prompt",
+                {
+                    "sessionId": session_id,
+                    "prompt": [
+                        {
+                            "type": "text",
+                            "text": prompt_text,
+                        }
+                    ],
+                },
+                text_parts=text_parts,
+                reasoning_parts=reasoning_parts,
+            )
+            return "".join(text_parts), "".join(reasoning_parts)
+        finally:
+            self.close()
+
+    def _handle_server_message(
+        self,
+        msg: dict[str, Any],
+        *,
+        process: subprocess.Popen[str],
+        cwd: str,
+        text_parts: list[str] | None,
+        reasoning_parts: list[str] | None,
+    ) -> bool:
+        method = msg.get("method")
+        if not isinstance(method, str):
+            return False
+
+        if method == "session/update":
+            params = msg.get("params") or {}
+            update = params.get("update") or {}
+            kind = str(update.get("sessionUpdate") or "").strip()
+            content = update.get("content") or {}
+            chunk_text = ""
+            if isinstance(content, dict):
+                chunk_text = str(content.get("text") or "")
+            if kind == "agent_message_chunk" and chunk_text and text_parts is not None:
+                text_parts.append(chunk_text)
+            elif kind == "agent_thought_chunk" and chunk_text and reasoning_parts is not None:
+                reasoning_parts.append(chunk_text)
+            return True
+
+        if process.stdin is None:
+            return True
+
+        message_id = msg.get("id")
+        params = msg.get("params") or {}
+
+        if method == "session/request_permission":
+            response = {
+                "jsonrpc": "2.0",
+                "id": message_id,
+                "result": {
+                    "outcome": {
+                        "outcome": "allow_once",
+                    }
+                },
+            }
+        elif method == "fs/read_text_file":
+            try:
+                path = _ensure_path_within_cwd(str(params.get("path") or ""), cwd)
+                content = path.read_text() if path.exists() else ""
+                line = params.get("line")
+                limit = params.get("limit")
+                if isinstance(line, int) and line > 1:
+                    lines = content.splitlines(keepends=True)
+                    start = line - 1
+                    end = start + limit if isinstance(limit, int) and limit > 0 else None
+                    content = "".join(lines[start:end])
+                response = {
+                    "jsonrpc": "2.0",
+                    "id": message_id,
+                    "result": {
+                        "content": content,
+                    },
+                }
+            except Exception as exc:
+                response = _jsonrpc_error(message_id, -32602, str(exc))
+        elif method == "fs/write_text_file":
+            try:
+                path = _ensure_path_within_cwd(str(params.get("path") or ""), cwd)
+                path.parent.mkdir(parents=True, exist_ok=True)
+                path.write_text(str(params.get("content") or ""))
+                response = {
+                    "jsonrpc": "2.0",
+                    "id": message_id,
+                    "result": None,
+                }
+            except Exception as exc:
+                response = _jsonrpc_error(message_id, -32602, str(exc))
+        else:
+            response = _jsonrpc_error(
+                message_id,
+                -32601,
+                f"ACP client method '{method}' is not supported by Hermes yet.",
+            )
+
+        process.stdin.write(json.dumps(response) + "\n")
+        process.stdin.flush()
+        return True

agent/display.py

@@ -10,6 +10,9 @@ import os
 import sys
 import threading
 import time
+from dataclasses import dataclass, field
+from difflib import unified_diff
+from pathlib import Path
 
 # ANSI escape codes for coloring tool failure indicators
 _RED = "\033[31m"
@@ -17,6 +20,39 @@ _RESET = "\033[0m"
 
 logger = logging.getLogger(__name__)
 
+_ANSI_RESET = "\033[0m"
+_ANSI_DIM = "\033[38;2;150;150;150m"
+_ANSI_FILE = "\033[38;2;180;160;255m"
+_ANSI_HUNK = "\033[38;2;120;120;140m"
+_ANSI_MINUS = "\033[38;2;255;255;255;48;2;120;20;20m"
+_ANSI_PLUS = "\033[38;2;255;255;255;48;2;20;90;20m"
+_MAX_INLINE_DIFF_FILES = 6
+_MAX_INLINE_DIFF_LINES = 80
+
+
+@dataclass
+class LocalEditSnapshot:
+    """Pre-tool filesystem snapshot used to render diffs locally after writes."""
+    paths: list[Path] = field(default_factory=list)
+    before: dict[str, str | None] = field(default_factory=dict)
+
+# =========================================================================
+# Configurable tool preview length (0 = no limit)
+# Set once at startup by CLI or gateway from display.tool_preview_length config.
+# =========================================================================
+_tool_preview_max_len: int = 0  # 0 = unlimited
+
+
+def set_tool_preview_max_len(n: int) -> None:
+    """Set the global max length for tool call previews. 0 = no limit."""
+    global _tool_preview_max_len
+    _tool_preview_max_len = max(int(n), 0) if n else 0
+
+
+def get_tool_preview_max_len() -> int:
+    """Return the configured max preview length (0 = unlimited)."""
+    return _tool_preview_max_len
+
 
 # =========================================================================
 # Skin-aware helpers (lazy import to avoid circular deps)
@@ -59,6 +95,32 @@ def get_skin_tool_prefix() -> str:
     return "┊"
 
 
+def get_tool_emoji(tool_name: str, default: str = "⚡") -> str:
+    """Get the display emoji for a tool.
+
+    Resolution order:
+    1. Active skin's ``tool_emojis`` overrides (if a skin is loaded)
+    2. Tool registry's per-tool ``emoji`` field
+    3. *default* fallback
+    """
+    # 1. Skin override
+    skin = _get_skin()
+    if skin and skin.tool_emojis:
+        override = skin.tool_emojis.get(tool_name)
+        if override:
+            return override
+    # 2. Registry default
+    try:
+        from tools.registry import registry
+        emoji = registry.get_emoji(tool_name, default="")
+        if emoji:
+            return emoji
+    except Exception:
+        pass
+    # 3. Hardcoded fallback
+    return default
+
+
 # =========================================================================
 # Tool preview (one-line summary of a tool call's primary argument)
 # =========================================================================
@@ -68,8 +130,14 @@ def _oneline(text: str) -> str:
     return " ".join(text.split())
 
 
-def build_tool_preview(tool_name: str, args: dict, max_len: int = 40) -> str | None:
-    """Build a short preview of a tool call's primary argument for display."""
+def build_tool_preview(tool_name: str, args: dict, max_len: int | None = None) -> str | None:
+    """Build a short preview of a tool call's primary argument for display.
+
+    *max_len* controls truncation.  ``None`` (default) defers to the global
+    ``_tool_preview_max_len`` set via config; ``0`` means unlimited.
+    """
+    if max_len is None:
+        max_len = _tool_preview_max_len
     if not args:
         return None
     primary_args = {
@@ -164,11 +232,305 @@ def build_tool_preview(tool_name: str, args: dict, max_len: int = 40) -> str | N
     preview = _oneline(str(value))
     if not preview:
         return None
-    if len(preview) > max_len:
+    if max_len > 0 and len(preview) > max_len:
         preview = preview[:max_len - 3] + "..."
     return preview
 
 
+# =========================================================================
+# Inline diff previews for write actions
+# =========================================================================
+
+def _resolved_path(path: str) -> Path:
+    """Resolve a possibly-relative filesystem path against the current cwd."""
+    candidate = Path(os.path.expanduser(path))
+    if candidate.is_absolute():
+        return candidate
+    return Path.cwd() / candidate
+
+
+def _snapshot_text(path: Path) -> str | None:
+    """Return UTF-8 file content, or None for missing/unreadable files."""
+    try:
+        return path.read_text(encoding="utf-8")
+    except (FileNotFoundError, IsADirectoryError, UnicodeDecodeError, OSError):
+        return None
+
+
+def _display_diff_path(path: Path) -> str:
+    """Prefer cwd-relative paths in diffs when available."""
+    try:
+        return str(path.resolve().relative_to(Path.cwd().resolve()))
+    except Exception:
+        return str(path)
+
+
+def _resolve_skill_manage_paths(args: dict) -> list[Path]:
+    """Resolve skill_manage write targets to filesystem paths."""
+    action = args.get("action")
+    name = args.get("name")
+    if not action or not name:
+        return []
+
+    from tools.skill_manager_tool import _find_skill, _resolve_skill_dir
+
+    if action == "create":
+        skill_dir = _resolve_skill_dir(name, args.get("category"))
+        return [skill_dir / "SKILL.md"]
+
+    existing = _find_skill(name)
+    if not existing:
+        return []
+
+    skill_dir = Path(existing["path"])
+    if action in {"edit", "patch"}:
+        file_path = args.get("file_path")
+        return [skill_dir / file_path] if file_path else [skill_dir / "SKILL.md"]
+    if action in {"write_file", "remove_file"}:
+        file_path = args.get("file_path")
+        return [skill_dir / file_path] if file_path else []
+    if action == "delete":
+        files = [path for path in sorted(skill_dir.rglob("*")) if path.is_file()]
+        return files
+    return []
+
+
+def _resolve_local_edit_paths(tool_name: str, function_args: dict | None) -> list[Path]:
+    """Resolve local filesystem targets for write-capable tools."""
+    if not isinstance(function_args, dict):
+        return []
+
+    if tool_name == "write_file":
+        path = function_args.get("path")
+        return [_resolved_path(path)] if path else []
+
+    if tool_name == "patch":
+        path = function_args.get("path")
+        return [_resolved_path(path)] if path else []
+
+    if tool_name == "skill_manage":
+        return _resolve_skill_manage_paths(function_args)
+
+    return []
+
+
+def capture_local_edit_snapshot(tool_name: str, function_args: dict | None) -> LocalEditSnapshot | None:
+    """Capture before-state for local write previews."""
+    paths = _resolve_local_edit_paths(tool_name, function_args)
+    if not paths:
+        return None
+
+    snapshot = LocalEditSnapshot(paths=paths)
+    for path in paths:
+        snapshot.before[str(path)] = _snapshot_text(path)
+    return snapshot
+
+
+def _result_succeeded(result: str | None) -> bool:
+    """Conservatively detect whether a tool result represents success."""
+    if not result:
+        return False
+    try:
+        data = json.loads(result)
+    except (json.JSONDecodeError, TypeError):
+        return False
+    if not isinstance(data, dict):
+        return False
+    if data.get("error"):
+        return False
+    if "success" in data:
+        return bool(data.get("success"))
+    return True
+
+
+def _diff_from_snapshot(snapshot: LocalEditSnapshot | None) -> str | None:
+    """Generate unified diff text from a stored before-state and current files."""
+    if not snapshot:
+        return None
+
+    chunks: list[str] = []
+    for path in snapshot.paths:
+        before = snapshot.before.get(str(path))
+        after = _snapshot_text(path)
+        if before == after:
+            continue
+
+        display_path = _display_diff_path(path)
+        diff = "".join(
+            unified_diff(
+                [] if before is None else before.splitlines(keepends=True),
+                [] if after is None else after.splitlines(keepends=True),
+                fromfile=f"a/{display_path}",
+                tofile=f"b/{display_path}",
+            )
+        )
+        if diff:
+            chunks.append(diff)
+
+    if not chunks:
+        return None
+    return "".join(chunk if chunk.endswith("\n") else chunk + "\n" for chunk in chunks)
+
+
+def extract_edit_diff(
+    tool_name: str,
+    result: str | None,
+    *,
+    function_args: dict | None = None,
+    snapshot: LocalEditSnapshot | None = None,
+) -> str | None:
+    """Extract a unified diff from a file-edit tool result."""
+    if tool_name == "patch" and result:
+        try:
+            data = json.loads(result)
+        except (json.JSONDecodeError, TypeError):
+            data = None
+        if isinstance(data, dict):
+            diff = data.get("diff")
+            if isinstance(diff, str) and diff.strip():
+                return diff
+
+    if tool_name not in {"write_file", "patch", "skill_manage"}:
+        return None
+    if not _result_succeeded(result):
+        return None
+    return _diff_from_snapshot(snapshot)
+
+
+def _emit_inline_diff(diff_text: str, print_fn) -> bool:
+    """Emit rendered diff text through the CLI's prompt_toolkit-safe printer."""
+    if print_fn is None or not diff_text:
+        return False
+    try:
+        print_fn("  ┊ review diff")
+        for line in diff_text.rstrip("\n").splitlines():
+            print_fn(line)
+        return True
+    except Exception:
+        return False
+
+
+def _render_inline_unified_diff(diff: str) -> list[str]:
+    """Render unified diff lines in Hermes' inline transcript style."""
+    rendered: list[str] = []
+    from_file = None
+    to_file = None
+
+    for raw_line in diff.splitlines():
+        if raw_line.startswith("--- "):
+            from_file = raw_line[4:].strip()
+            continue
+        if raw_line.startswith("+++ "):
+            to_file = raw_line[4:].strip()
+            if from_file or to_file:
+                rendered.append(f"{_ANSI_FILE}{from_file or 'a/?'} → {to_file or 'b/?'}{_ANSI_RESET}")
+            continue
+        if raw_line.startswith("@@"):
+            rendered.append(f"{_ANSI_HUNK}{raw_line}{_ANSI_RESET}")
+            continue
+        if raw_line.startswith("-"):
+            rendered.append(f"{_ANSI_MINUS}{raw_line}{_ANSI_RESET}")
+            continue
+        if raw_line.startswith("+"):
+            rendered.append(f"{_ANSI_PLUS}{raw_line}{_ANSI_RESET}")
+            continue
+        if raw_line.startswith(" "):
+            rendered.append(f"{_ANSI_DIM}{raw_line}{_ANSI_RESET}")
+            continue
+        if raw_line:
+            rendered.append(raw_line)
+
+    return rendered
+
+
+def _split_unified_diff_sections(diff: str) -> list[str]:
+    """Split a unified diff into per-file sections."""
+    sections: list[list[str]] = []
+    current: list[str] = []
+
+    for line in diff.splitlines():
+        if line.startswith("--- ") and current:
+            sections.append(current)
+            current = [line]
+            continue
+        current.append(line)
+
+    if current:
+        sections.append(current)
+
+    return ["\n".join(section) for section in sections if section]
+
+
+def _summarize_rendered_diff_sections(
+    diff: str,
+    *,
+    max_files: int = _MAX_INLINE_DIFF_FILES,
+    max_lines: int = _MAX_INLINE_DIFF_LINES,
+) -> list[str]:
+    """Render diff sections while capping file count and total line count."""
+    sections = _split_unified_diff_sections(diff)
+    rendered: list[str] = []
+    omitted_files = 0
+    omitted_lines = 0
+
+    for idx, section in enumerate(sections):
+        if idx >= max_files:
+            omitted_files += 1
+            omitted_lines += len(_render_inline_unified_diff(section))
+            continue
+
+        section_lines = _render_inline_unified_diff(section)
+        remaining_budget = max_lines - len(rendered)
+        if remaining_budget <= 0:
+            omitted_lines += len(section_lines)
+            omitted_files += 1
+            continue
+
+        if len(section_lines) <= remaining_budget:
+            rendered.extend(section_lines)
+            continue
+
+        rendered.extend(section_lines[:remaining_budget])
+        omitted_lines += len(section_lines) - remaining_budget
+        omitted_files += 1 + max(0, len(sections) - idx - 1)
+        for leftover in sections[idx + 1:]:
+            omitted_lines += len(_render_inline_unified_diff(leftover))
+        break
+
+    if omitted_files or omitted_lines:
+        summary = f"… omitted {omitted_lines} diff line(s)"
+        if omitted_files:
+            summary += f" across {omitted_files} additional file(s)/section(s)"
+        rendered.append(f"{_ANSI_HUNK}{summary}{_ANSI_RESET}")
+
+    return rendered
+
+
+def render_edit_diff_with_delta(
+    tool_name: str,
+    result: str | None,
+    *,
+    function_args: dict | None = None,
+    snapshot: LocalEditSnapshot | None = None,
+    print_fn=None,
+) -> bool:
+    """Render an edit diff inline without taking over the terminal UI."""
+    diff = extract_edit_diff(
+        tool_name,
+        result,
+        function_args=function_args,
+        snapshot=snapshot,
+    )
+    if not diff:
+        return False
+    try:
+        rendered_lines = _summarize_rendered_diff_sections(diff)
+    except Exception as exc:
+        logger.debug("Could not render inline diff: %s", exc)
+        return False
+    return _emit_inline_diff("\n".join(rendered_lines), print_fn)
+
+
 # =========================================================================
 # KawaiiSpinner
 # =========================================================================
@@ -205,7 +567,7 @@ class KawaiiSpinner:
         "analyzing", "computing", "synthesizing", "formulating", "brainstorming",
     ]
 
-    def __init__(self, message: str = "", spinner_type: str = 'dots'):
+    def __init__(self, message: str = "", spinner_type: str = 'dots', print_fn=None):
         self.message = message
         self.spinner_frames = self.SPINNERS.get(spinner_type, self.SPINNERS['dots'])
         self.running = False
@@ -213,13 +575,26 @@ class KawaiiSpinner:
         self.frame_idx = 0
         self.start_time = None
         self.last_line_len = 0
-        self._last_flush_time = 0.0  # Rate-limit flushes for patch_stdout compat
+        # Optional callable to route all output through (e.g. a no-op for silent
+        # background agents).  When set, bypasses self._out entirely so that
+        # agents with _print_fn overridden remain fully silent.
+        self._print_fn = print_fn
         # Capture stdout NOW, before any redirect_stdout(devnull) from
         # child agents can replace sys.stdout with a black hole.
         self._out = sys.stdout
 
     def _write(self, text: str, end: str = '\n', flush: bool = False):
-        """Write to the stdout captured at spinner creation time."""
+        """Write to the stdout captured at spinner creation time.
+
+        If a print_fn was supplied at construction, all output is routed through
+        it instead — allowing callers to silence the spinner with a no-op lambda.
+        """
+        if self._print_fn is not None:
+            try:
+                self._print_fn(text)
+            except Exception:
+                pass
+            return
         try:
             self._out.write(text + end)
             if flush:
@@ -227,7 +602,50 @@ class KawaiiSpinner:
         except (ValueError, OSError):
             pass
 
+    @property
+    def _is_tty(self) -> bool:
+        """Check if output is a real terminal, safe against closed streams."""
+        try:
+            return hasattr(self._out, 'isatty') and self._out.isatty()
+        except (ValueError, OSError):
+            return False
+
+    def _is_patch_stdout_proxy(self) -> bool:
+        """Return True when stdout is prompt_toolkit's StdoutProxy.
+
+        patch_stdout wraps sys.stdout in a StdoutProxy that queues writes and
+        injects newlines around each flush().  The \\r overwrite never lands on
+        the correct line — each spinner frame ends up on its own line.
+
+        The CLI already drives a TUI widget (_spinner_text) for spinner display,
+        so KawaiiSpinner's \\r-based animation is redundant under StdoutProxy.
+        """
+        try:
+            from prompt_toolkit.patch_stdout import StdoutProxy
+            return isinstance(self._out, StdoutProxy)
+        except ImportError:
+            return False
+
     def _animate(self):
+        # When stdout is not a real terminal (e.g. Docker, systemd, pipe),
+        # skip the animation entirely — it creates massive log bloat.
+        # Just log the start once and let stop() log the completion.
+        if not self._is_tty:
+            self._write(f"  [tool] {self.message}", flush=True)
+            while self.running:
+                time.sleep(0.5)
+            return
+
+        # When running inside prompt_toolkit's patch_stdout context the CLI
+        # renders spinner state via a dedicated TUI widget (_spinner_text).
+        # Driving a \r-based animation here too causes visual overdraw: the
+        # StdoutProxy injects newlines around each flush, so every frame lands
+        # on a new line and overwrites the status bar.
+        if self._is_patch_stdout_proxy():
+            while self.running:
+                time.sleep(0.1)
+            return
+
         # Cache skin wings at start (avoid per-frame imports)
         skin = _get_skin()
         wings = skin.get_spinner_wings() if skin else []
@@ -244,18 +662,7 @@ class KawaiiSpinner:
             else:
                 line = f"  {frame} {self.message} ({elapsed:.1f}s)"
             pad = max(self.last_line_len - len(line), 0)
-            # Rate-limit flush() calls to avoid spinner spam under
-            # prompt_toolkit's patch_stdout.  Each flush() pushes a queue
-            # item that may trigger a separate run_in_terminal() call; if
-            # items are processed one-at-a-time the \r overwrite is lost
-            # and every frame appears on its own line.  By flushing at
-            # most every 0.4s we guarantee multiple \r-frames are batched
-            # into a single write, so the terminal collapses them correctly.
-            now = time.time()
-            should_flush = (now - self._last_flush_time) >= 0.4
-            self._write(f"\r{line}{' ' * pad}", end='', flush=should_flush)
-            if should_flush:
-                self._last_flush_time = now
+            self._write(f"\r{line}{' ' * pad}", end='', flush=True)
             self.last_line_len = len(line)
             self.frame_idx += 1
             time.sleep(0.12)
@@ -293,12 +700,19 @@ class KawaiiSpinner:
         self.running = False
         if self.thread:
             self.thread.join(timeout=0.5)
-        # Clear the spinner line with spaces instead of \033[K to avoid
-        # garbled escape codes when prompt_toolkit's patch_stdout is active.
-        blanks = ' ' * max(self.last_line_len + 5, 40)
-        self._write(f"\r{blanks}\r", end='', flush=True)
+
+        is_tty = self._is_tty
+        if is_tty:
+            # Clear the spinner line with spaces instead of \033[K to avoid
+            # garbled escape codes when prompt_toolkit's patch_stdout is active.
+            blanks = ' ' * max(self.last_line_len + 5, 40)
+            self._write(f"\r{blanks}\r", end='', flush=True)
         if final_message:
-            self._write(f"  {final_message}", flush=True)
+            elapsed = f" ({time.time() - self.start_time:.1f}s)" if self.start_time else ""
+            if is_tty:
+                self._write(f"  {final_message}", flush=True)
+            else:
+                self._write(f"  [done] {final_message}{elapsed}", flush=True)
 
     def __enter__(self):
         self.start()
@@ -406,10 +820,14 @@ def get_cute_tool_message(
 
     def _trunc(s, n=40):
         s = str(s)
+        if _tool_preview_max_len == 0:
+            return s  # no limit
         return (s[:n-3] + "...") if len(s) > n else s
 
     def _path(p, n=35):
         p = str(p)
+        if _tool_preview_max_len == 0:
+            return p  # no limit
         return ("..." + p[-(n-3):]) if len(p) > n else p
 
     def _wrap(line: str) -> str:
@@ -586,3 +1004,81 @@ def write_tty(text: str) -> None:
     except OSError:
         sys.stdout.write(text)
         sys.stdout.flush()
+
+
+# =========================================================================
+# Context pressure display (CLI user-facing warnings)
+# =========================================================================
+
+# ANSI color codes for context pressure tiers
+_CYAN = "\033[36m"
+_YELLOW = "\033[33m"
+_BOLD = "\033[1m"
+_DIM_ANSI = "\033[2m"
+
+# Bar characters
+_BAR_FILLED = "▰"
+_BAR_EMPTY = "▱"
+_BAR_WIDTH = 20
+
+
+def format_context_pressure(
+    compaction_progress: float,
+    threshold_tokens: int,
+    threshold_percent: float,
+    compression_enabled: bool = True,
+) -> str:
+    """Build a formatted context pressure line for CLI display.
+
+    The bar and percentage show progress toward the compaction threshold,
+    NOT the raw context window.  100% = compaction fires.
+
+    Args:
+        compaction_progress: How close to compaction (0.0–1.0, 1.0 = fires).
+        threshold_tokens: Compaction threshold in tokens.
+        threshold_percent: Compaction threshold as a fraction of context window.
+        compression_enabled: Whether auto-compression is active.
+    """
+    pct_int = min(int(compaction_progress * 100), 100)
+    filled = min(int(compaction_progress * _BAR_WIDTH), _BAR_WIDTH)
+    bar = _BAR_FILLED * filled + _BAR_EMPTY * (_BAR_WIDTH - filled)
+
+    threshold_k = f"{threshold_tokens // 1000}k" if threshold_tokens >= 1000 else str(threshold_tokens)
+    threshold_pct_int = int(threshold_percent * 100)
+
+    color = f"{_BOLD}{_YELLOW}"
+    icon = "⚠"
+    if compression_enabled:
+        hint = "compaction approaching"
+    else:
+        hint = "no auto-compaction"
+
+    return (
+        f"  {color}{icon} context {bar} {pct_int}% to compaction{_ANSI_RESET}"
+        f"  {_DIM_ANSI}{threshold_k} threshold ({threshold_pct_int}%) · {hint}{_ANSI_RESET}"
+    )
+
+
+def format_context_pressure_gateway(
+    compaction_progress: float,
+    threshold_percent: float,
+    compression_enabled: bool = True,
+) -> str:
+    """Build a plain-text context pressure notification for messaging platforms.
+
+    No ANSI — just Unicode and plain text suitable for Telegram/Discord/etc.
+    The percentage shows progress toward the compaction threshold.
+    """
+    pct_int = min(int(compaction_progress * 100), 100)
+    filled = min(int(compaction_progress * _BAR_WIDTH), _BAR_WIDTH)
+    bar = _BAR_FILLED * filled + _BAR_EMPTY * (_BAR_WIDTH - filled)
+
+    threshold_pct_int = int(threshold_percent * 100)
+
+    icon = "⚠️"
+    if compression_enabled:
+        hint = f"Context compaction approaching (threshold: {threshold_pct_int}% of window)."
+    else:
+        hint = "Auto-compaction is disabled — context may be truncated."
+
+    return f"{icon} Context: {bar} {pct_int}% to compaction\n{hint}"

agent/insights.py

@@ -20,65 +20,23 @@ import json
 import time
 from collections import Counter, defaultdict
 from datetime import datetime
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, List
 
-# =========================================================================
-# Model pricing (USD per million tokens) — approximate as of early 2026
-# =========================================================================
-MODEL_PRICING = {
-    # OpenAI
-    "gpt-4o": {"input": 2.50, "output": 10.00},
-    "gpt-4o-mini": {"input": 0.15, "output": 0.60},
-    "gpt-4.1": {"input": 2.00, "output": 8.00},
-    "gpt-4.1-mini": {"input": 0.40, "output": 1.60},
-    "gpt-4.1-nano": {"input": 0.10, "output": 0.40},
-    "gpt-4.5-preview": {"input": 75.00, "output": 150.00},
-    "gpt-5": {"input": 10.00, "output": 30.00},
-    "gpt-5.4": {"input": 10.00, "output": 30.00},
-    "o3": {"input": 10.00, "output": 40.00},
-    "o3-mini": {"input": 1.10, "output": 4.40},
-    "o4-mini": {"input": 1.10, "output": 4.40},
-    # Anthropic
-    "claude-opus-4-20250514": {"input": 15.00, "output": 75.00},
-    "claude-sonnet-4-20250514": {"input": 3.00, "output": 15.00},
-    "claude-3-5-sonnet-20241022": {"input": 3.00, "output": 15.00},
-    "claude-3-5-haiku-20241022": {"input": 0.80, "output": 4.00},
-    "claude-3-opus-20240229": {"input": 15.00, "output": 75.00},
-    "claude-3-haiku-20240307": {"input": 0.25, "output": 1.25},
-    # DeepSeek
-    "deepseek-chat": {"input": 0.14, "output": 0.28},
-    "deepseek-reasoner": {"input": 0.55, "output": 2.19},
-    # Google
-    "gemini-2.5-pro": {"input": 1.25, "output": 10.00},
-    "gemini-2.5-flash": {"input": 0.15, "output": 0.60},
-    "gemini-2.0-flash": {"input": 0.10, "output": 0.40},
-    # Meta (via providers)
-    "llama-4-maverick": {"input": 0.50, "output": 0.70},
-    "llama-4-scout": {"input": 0.20, "output": 0.30},
-    # Z.AI / GLM (direct provider — pricing not published externally, treat as local)
-    "glm-5": {"input": 0.0, "output": 0.0},
-    "glm-4.7": {"input": 0.0, "output": 0.0},
-    "glm-4.5": {"input": 0.0, "output": 0.0},
-    "glm-4.5-flash": {"input": 0.0, "output": 0.0},
-    # Kimi / Moonshot (direct provider — pricing not published externally, treat as local)
-    "kimi-k2.5": {"input": 0.0, "output": 0.0},
-    "kimi-k2-thinking": {"input": 0.0, "output": 0.0},
-    "kimi-k2-turbo-preview": {"input": 0.0, "output": 0.0},
-    "kimi-k2-0905-preview": {"input": 0.0, "output": 0.0},
-    # MiniMax (direct provider — pricing not published externally, treat as local)
-    "MiniMax-M2.5": {"input": 0.0, "output": 0.0},
-    "MiniMax-M2.5-highspeed": {"input": 0.0, "output": 0.0},
-    "MiniMax-M2.1": {"input": 0.0, "output": 0.0},
-}
+from agent.usage_pricing import (
+    CanonicalUsage,
+    DEFAULT_PRICING,
+    estimate_usage_cost,
+    format_duration_compact,
+    get_pricing,
+    has_known_pricing,
+)
 
-# Fallback: unknown/custom models get zero cost (we can't assume pricing
-# for self-hosted models, custom OAI endpoints, local inference, etc.)
-_DEFAULT_PRICING = {"input": 0.0, "output": 0.0}
+_DEFAULT_PRICING = DEFAULT_PRICING
 
 
-def _has_known_pricing(model_name: str) -> bool:
+def _has_known_pricing(model_name: str, provider: str = None, base_url: str = None) -> bool:
     """Check if a model has known pricing (vs unknown/custom endpoint)."""
-    return _get_pricing(model_name) is not _DEFAULT_PRICING
+    return has_known_pricing(model_name, provider=provider, base_url=base_url)
 
 
 def _get_pricing(model_name: str) -> Dict[str, float]:
@@ -87,67 +45,51 @@ def _get_pricing(model_name: str) -> Dict[str, float]:
     Returns _DEFAULT_PRICING (zero cost) for unknown/custom models —
     we can't assume costs for self-hosted endpoints, local inference, etc.
     """
-    if not model_name:
-        return _DEFAULT_PRICING
-
-    # Strip provider prefix (e.g., "anthropic/claude-..." -> "claude-...")
-    bare = model_name.split("/")[-1].lower()
-
-    # Exact match first
-    if bare in MODEL_PRICING:
-        return MODEL_PRICING[bare]
-
-    # Fuzzy prefix match — prefer the LONGEST matching key to avoid
-    # e.g. "gpt-4o" matching before "gpt-4o-mini" for "gpt-4o-mini-2024-07-18"
-    best_match = None
-    best_len = 0
-    for key, price in MODEL_PRICING.items():
-        if bare.startswith(key) and len(key) > best_len:
-            best_match = price
-            best_len = len(key)
-    if best_match:
-        return best_match
-
-    # Keyword heuristics (checked in most-specific-first order)
-    if "opus" in bare:
-        return {"input": 15.00, "output": 75.00}
-    if "sonnet" in bare:
-        return {"input": 3.00, "output": 15.00}
-    if "haiku" in bare:
-        return {"input": 0.80, "output": 4.00}
-    if "gpt-4o-mini" in bare:
-        return {"input": 0.15, "output": 0.60}
-    if "gpt-4o" in bare:
-        return {"input": 2.50, "output": 10.00}
-    if "gpt-5" in bare:
-        return {"input": 10.00, "output": 30.00}
-    if "deepseek" in bare:
-        return {"input": 0.14, "output": 0.28}
-    if "gemini" in bare:
-        return {"input": 0.15, "output": 0.60}
-
-    return _DEFAULT_PRICING
+    return get_pricing(model_name)
 
 
-def _estimate_cost(model: str, input_tokens: int, output_tokens: int) -> float:
-    """Estimate the USD cost for a given model and token counts."""
-    pricing = _get_pricing(model)
-    return (input_tokens * pricing["input"] + output_tokens * pricing["output"]) / 1_000_000
+def _estimate_cost(
+    session_or_model: Dict[str, Any] | str,
+    input_tokens: int = 0,
+    output_tokens: int = 0,
+    *,
+    cache_read_tokens: int = 0,
+    cache_write_tokens: int = 0,
+    provider: str = None,
+    base_url: str = None,
+) -> tuple[float, str]:
+    """Estimate the USD cost for a session row or a model/token tuple."""
+    if isinstance(session_or_model, dict):
+        session = session_or_model
+        model = session.get("model") or ""
+        usage = CanonicalUsage(
+            input_tokens=session.get("input_tokens") or 0,
+            output_tokens=session.get("output_tokens") or 0,
+            cache_read_tokens=session.get("cache_read_tokens") or 0,
+            cache_write_tokens=session.get("cache_write_tokens") or 0,
+        )
+        provider = session.get("billing_provider")
+        base_url = session.get("billing_base_url")
+    else:
+        model = session_or_model or ""
+        usage = CanonicalUsage(
+            input_tokens=input_tokens,
+            output_tokens=output_tokens,
+            cache_read_tokens=cache_read_tokens,
+            cache_write_tokens=cache_write_tokens,
+        )
+    result = estimate_usage_cost(
+        model,
+        usage,
+        provider=provider,
+        base_url=base_url,
+    )
+    return float(result.amount_usd or 0.0), result.status
 
 
 def _format_duration(seconds: float) -> str:
     """Format seconds into a human-readable duration string."""
-    if seconds < 60:
-        return f"{seconds:.0f}s"
-    minutes = seconds / 60
-    if minutes < 60:
-        return f"{minutes:.0f}m"
-    hours = minutes / 60
-    if hours < 24:
-        remaining_min = int(minutes % 60)
-        return f"{int(hours)}h {remaining_min}m" if remaining_min else f"{int(hours)}h"
-    days = hours / 24
-    return f"{days:.1f}d"
+    return format_duration_compact(seconds)
 
 
 def _bar_chart(values: List[int], max_width: int = 20) -> List[str]:
@@ -234,24 +176,30 @@ class InsightsEngine:
 
     # Columns we actually need (skip system_prompt, model_config blobs)
     _SESSION_COLS = ("id, source, model, started_at, ended_at, "
-                     "message_count, tool_call_count, input_tokens, output_tokens")
+                     "message_count, tool_call_count, input_tokens, output_tokens, "
+                     "cache_read_tokens, cache_write_tokens, billing_provider, "
+                     "billing_base_url, billing_mode, estimated_cost_usd, "
+                     "actual_cost_usd, cost_status, cost_source")
+
+    # Pre-computed query strings — f-string evaluated once at class definition,
+    # not at runtime, so no user-controlled value can alter the query structure.
+    _GET_SESSIONS_WITH_SOURCE = (
+        f"SELECT {_SESSION_COLS} FROM sessions"
+        " WHERE started_at >= ? AND source = ?"
+        " ORDER BY started_at DESC"
+    )
+    _GET_SESSIONS_ALL = (
+        f"SELECT {_SESSION_COLS} FROM sessions"
+        " WHERE started_at >= ?"
+        " ORDER BY started_at DESC"
+    )
 
     def _get_sessions(self, cutoff: float, source: str = None) -> List[Dict]:
         """Fetch sessions within the time window."""
         if source:
-            cursor = self._conn.execute(
-                f"""SELECT {self._SESSION_COLS} FROM sessions
-                    WHERE started_at >= ? AND source = ?
-                    ORDER BY started_at DESC""",
-                (cutoff, source),
-            )
+            cursor = self._conn.execute(self._GET_SESSIONS_WITH_SOURCE, (cutoff, source))
         else:
-            cursor = self._conn.execute(
-                f"""SELECT {self._SESSION_COLS} FROM sessions
-                    WHERE started_at >= ?
-                    ORDER BY started_at DESC""",
-                (cutoff,),
-            )
+            cursor = self._conn.execute(self._GET_SESSIONS_ALL, (cutoff,))
         return [dict(row) for row in cursor.fetchall()]
 
     def _get_tool_usage(self, cutoff: float, source: str = None) -> List[Dict]:
@@ -386,21 +334,30 @@ class InsightsEngine:
         """Compute high-level overview statistics."""
         total_input = sum(s.get("input_tokens") or 0 for s in sessions)
         total_output = sum(s.get("output_tokens") or 0 for s in sessions)
-        total_tokens = total_input + total_output
+        total_cache_read = sum(s.get("cache_read_tokens") or 0 for s in sessions)
+        total_cache_write = sum(s.get("cache_write_tokens") or 0 for s in sessions)
+        total_tokens = total_input + total_output + total_cache_read + total_cache_write
         total_tool_calls = sum(s.get("tool_call_count") or 0 for s in sessions)
         total_messages = sum(s.get("message_count") or 0 for s in sessions)
 
         # Cost estimation (weighted by model)
         total_cost = 0.0
+        actual_cost = 0.0
         models_with_pricing = set()
         models_without_pricing = set()
+        unknown_cost_sessions = 0
+        included_cost_sessions = 0
         for s in sessions:
             model = s.get("model") or ""
-            inp = s.get("input_tokens") or 0
-            out = s.get("output_tokens") or 0
-            total_cost += _estimate_cost(model, inp, out)
+            estimated, status = _estimate_cost(s)
+            total_cost += estimated
+            actual_cost += s.get("actual_cost_usd") or 0.0
             display = model.split("/")[-1] if "/" in model else (model or "unknown")
-            if _has_known_pricing(model):
+            if status == "included":
+                included_cost_sessions += 1
+            elif status == "unknown":
+                unknown_cost_sessions += 1
+            if _has_known_pricing(model, s.get("billing_provider"), s.get("billing_base_url")):
                 models_with_pricing.add(display)
             else:
                 models_without_pricing.add(display)
@@ -427,8 +384,11 @@ class InsightsEngine:
             "total_tool_calls": total_tool_calls,
             "total_input_tokens": total_input,
             "total_output_tokens": total_output,
+            "total_cache_read_tokens": total_cache_read,
+            "total_cache_write_tokens": total_cache_write,
             "total_tokens": total_tokens,
             "estimated_cost": total_cost,
+            "actual_cost": actual_cost,
             "total_hours": total_hours,
             "avg_session_duration": avg_duration,
             "avg_messages_per_session": total_messages / len(sessions) if sessions else 0,
@@ -440,12 +400,15 @@ class InsightsEngine:
             "date_range_end": date_range_end,
             "models_with_pricing": sorted(models_with_pricing),
             "models_without_pricing": sorted(models_without_pricing),
+            "unknown_cost_sessions": unknown_cost_sessions,
+            "included_cost_sessions": included_cost_sessions,
         }
 
     def _compute_model_breakdown(self, sessions: List[Dict]) -> List[Dict]:
         """Break down usage by model."""
         model_data = defaultdict(lambda: {
             "sessions": 0, "input_tokens": 0, "output_tokens": 0,
+            "cache_read_tokens": 0, "cache_write_tokens": 0,
             "total_tokens": 0, "tool_calls": 0, "cost": 0.0,
         })
 
@@ -457,12 +420,18 @@ class InsightsEngine:
             d["sessions"] += 1
             inp = s.get("input_tokens") or 0
             out = s.get("output_tokens") or 0
+            cache_read = s.get("cache_read_tokens") or 0
+            cache_write = s.get("cache_write_tokens") or 0
             d["input_tokens"] += inp
             d["output_tokens"] += out
-            d["total_tokens"] += inp + out
+            d["cache_read_tokens"] += cache_read
+            d["cache_write_tokens"] += cache_write
+            d["total_tokens"] += inp + out + cache_read + cache_write
             d["tool_calls"] += s.get("tool_call_count") or 0
-            d["cost"] += _estimate_cost(model, inp, out)
-            d["has_pricing"] = _has_known_pricing(model)
+            estimate, status = _estimate_cost(s)
+            d["cost"] += estimate
+            d["has_pricing"] = _has_known_pricing(model, s.get("billing_provider"), s.get("billing_base_url"))
+            d["cost_status"] = status
 
         result = [
             {"model": model, **data}
@@ -476,7 +445,8 @@ class InsightsEngine:
         """Break down usage by platform/source."""
         platform_data = defaultdict(lambda: {
             "sessions": 0, "messages": 0, "input_tokens": 0,
-            "output_tokens": 0, "total_tokens": 0, "tool_calls": 0,
+            "output_tokens": 0, "cache_read_tokens": 0,
+            "cache_write_tokens": 0, "total_tokens": 0, "tool_calls": 0,
         })
 
         for s in sessions:
@@ -486,9 +456,13 @@ class InsightsEngine:
             d["messages"] += s.get("message_count") or 0
             inp = s.get("input_tokens") or 0
             out = s.get("output_tokens") or 0
+            cache_read = s.get("cache_read_tokens") or 0
+            cache_write = s.get("cache_write_tokens") or 0
             d["input_tokens"] += inp
             d["output_tokens"] += out
-            d["total_tokens"] += inp + out
+            d["cache_read_tokens"] += cache_read
+            d["cache_write_tokens"] += cache_write
+            d["total_tokens"] += inp + out + cache_read + cache_write
             d["tool_calls"] += s.get("tool_call_count") or 0
 
         result = [
@@ -670,6 +644,9 @@ class InsightsEngine:
         lines.append(f"  Sessions:          {o['total_sessions']:<12}  Messages:        {o['total_messages']:,}")
         lines.append(f"  Tool calls:        {o['total_tool_calls']:<12,}  User messages:   {o['user_messages']:,}")
         lines.append(f"  Input tokens:      {o['total_input_tokens']:<12,}  Output tokens:   {o['total_output_tokens']:,}")
+        cache_total = o.get("total_cache_read_tokens", 0) + o.get("total_cache_write_tokens", 0)
+        if cache_total > 0:
+            lines.append(f"  Cache read:        {o['total_cache_read_tokens']:<12,}  Cache write:     {o['total_cache_write_tokens']:,}")
         cost_str = f"${o['estimated_cost']:.2f}"
         if o.get("models_without_pricing"):
             cost_str += " *"
@@ -692,7 +669,7 @@ class InsightsEngine:
                     cost_cell = "     N/A"
                 lines.append(f"  {model_name:<30} {m['sessions']:>8} {m['total_tokens']:>12,} {cost_cell}")
             if o.get("models_without_pricing"):
-                lines.append(f"  * Cost N/A for custom/self-hosted models")
+                lines.append("  * Cost N/A for custom/self-hosted models")
             lines.append("")
 
         # Platform breakdown
@@ -772,7 +749,11 @@ class InsightsEngine:
 
         # Overview
         lines.append(f"**Sessions:** {o['total_sessions']} | **Messages:** {o['total_messages']:,} | **Tool calls:** {o['total_tool_calls']:,}")
-        lines.append(f"**Tokens:** {o['total_tokens']:,} (in: {o['total_input_tokens']:,} / out: {o['total_output_tokens']:,})")
+        cache_total = o.get("total_cache_read_tokens", 0) + o.get("total_cache_write_tokens", 0)
+        if cache_total > 0:
+            lines.append(f"**Tokens:** {o['total_tokens']:,} (in: {o['total_input_tokens']:,} / out: {o['total_output_tokens']:,} / cache: {cache_total:,})")
+        else:
+            lines.append(f"**Tokens:** {o['total_tokens']:,} (in: {o['total_input_tokens']:,} / out: {o['total_output_tokens']:,})")
         cost_note = ""
         if o.get("models_without_pricing"):
             cost_note = " _(excludes custom/self-hosted models)_"

agent/memory_manager.py

@@ -0,0 +1,335 @@
+"""MemoryManager — orchestrates the built-in memory provider plus at most
+ONE external plugin memory provider.
+
+Single integration point in run_agent.py. Replaces scattered per-backend
+code with one manager that delegates to registered providers.
+
+The BuiltinMemoryProvider is always registered first and cannot be removed.
+Only ONE external (non-builtin) provider is allowed at a time — attempting
+to register a second external provider is rejected with a warning.  This
+prevents tool schema bloat and conflicting memory backends.
+
+Usage in run_agent.py:
+    self._memory_manager = MemoryManager()
+    self._memory_manager.add_provider(BuiltinMemoryProvider(...))
+    # Only ONE of these:
+    self._memory_manager.add_provider(plugin_provider)
+
+    # System prompt
+    prompt_parts.append(self._memory_manager.build_system_prompt())
+
+    # Pre-turn
+    context = self._memory_manager.prefetch_all(user_message)
+
+    # Post-turn
+    self._memory_manager.sync_all(user_msg, assistant_response)
+    self._memory_manager.queue_prefetch_all(user_msg)
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from typing import Any, Dict, List, Optional
+
+from agent.memory_provider import MemoryProvider
+
+logger = logging.getLogger(__name__)
+
+
+class MemoryManager:
+    """Orchestrates the built-in provider plus at most one external provider.
+
+    The builtin provider is always first. Only one non-builtin (external)
+    provider is allowed.  Failures in one provider never block the other.
+    """
+
+    def __init__(self) -> None:
+        self._providers: List[MemoryProvider] = []
+        self._tool_to_provider: Dict[str, MemoryProvider] = {}
+        self._has_external: bool = False  # True once a non-builtin provider is added
+
+    # -- Registration --------------------------------------------------------
+
+    def add_provider(self, provider: MemoryProvider) -> None:
+        """Register a memory provider.
+
+        Built-in provider (name ``"builtin"``) is always accepted.
+        Only **one** external (non-builtin) provider is allowed — a second
+        attempt is rejected with a warning.
+        """
+        is_builtin = provider.name == "builtin"
+
+        if not is_builtin:
+            if self._has_external:
+                existing = next(
+                    (p.name for p in self._providers if p.name != "builtin"), "unknown"
+                )
+                logger.warning(
+                    "Rejected memory provider '%s' — external provider '%s' is "
+                    "already registered. Only one external memory provider is "
+                    "allowed at a time. Configure which one via memory.provider "
+                    "in config.yaml.",
+                    provider.name, existing,
+                )
+                return
+            self._has_external = True
+
+        self._providers.append(provider)
+
+        # Index tool names → provider for routing
+        for schema in provider.get_tool_schemas():
+            tool_name = schema.get("name", "")
+            if tool_name and tool_name not in self._tool_to_provider:
+                self._tool_to_provider[tool_name] = provider
+            elif tool_name in self._tool_to_provider:
+                logger.warning(
+                    "Memory tool name conflict: '%s' already registered by %s, "
+                    "ignoring from %s",
+                    tool_name,
+                    self._tool_to_provider[tool_name].name,
+                    provider.name,
+                )
+
+        logger.info(
+            "Memory provider '%s' registered (%d tools)",
+            provider.name,
+            len(provider.get_tool_schemas()),
+        )
+
+    @property
+    def providers(self) -> List[MemoryProvider]:
+        """All registered providers in order."""
+        return list(self._providers)
+
+    @property
+    def provider_names(self) -> List[str]:
+        """Names of all registered providers."""
+        return [p.name for p in self._providers]
+
+    def get_provider(self, name: str) -> Optional[MemoryProvider]:
+        """Get a provider by name, or None if not registered."""
+        for p in self._providers:
+            if p.name == name:
+                return p
+        return None
+
+    # -- System prompt -------------------------------------------------------
+
+    def build_system_prompt(self) -> str:
+        """Collect system prompt blocks from all providers.
+
+        Returns combined text, or empty string if no providers contribute.
+        Each non-empty block is labeled with the provider name.
+        """
+        blocks = []
+        for provider in self._providers:
+            try:
+                block = provider.system_prompt_block()
+                if block and block.strip():
+                    blocks.append(block)
+            except Exception as e:
+                logger.warning(
+                    "Memory provider '%s' system_prompt_block() failed: %s",
+                    provider.name, e,
+                )
+        return "\n\n".join(blocks)
+
+    # -- Prefetch / recall ---------------------------------------------------
+
+    def prefetch_all(self, query: str, *, session_id: str = "") -> str:
+        """Collect prefetch context from all providers.
+
+        Returns merged context text labeled by provider. Empty providers
+        are skipped. Failures in one provider don't block others.
+        """
+        parts = []
+        for provider in self._providers:
+            try:
+                result = provider.prefetch(query, session_id=session_id)
+                if result and result.strip():
+                    parts.append(result)
+            except Exception as e:
+                logger.debug(
+                    "Memory provider '%s' prefetch failed (non-fatal): %s",
+                    provider.name, e,
+                )
+        return "\n\n".join(parts)
+
+    def queue_prefetch_all(self, query: str, *, session_id: str = "") -> None:
+        """Queue background prefetch on all providers for the next turn."""
+        for provider in self._providers:
+            try:
+                provider.queue_prefetch(query, session_id=session_id)
+            except Exception as e:
+                logger.debug(
+                    "Memory provider '%s' queue_prefetch failed (non-fatal): %s",
+                    provider.name, e,
+                )
+
+    # -- Sync ----------------------------------------------------------------
+
+    def sync_all(self, user_content: str, assistant_content: str, *, session_id: str = "") -> None:
+        """Sync a completed turn to all providers."""
+        for provider in self._providers:
+            try:
+                provider.sync_turn(user_content, assistant_content, session_id=session_id)
+            except Exception as e:
+                logger.warning(
+                    "Memory provider '%s' sync_turn failed: %s",
+                    provider.name, e,
+                )
+
+    # -- Tools ---------------------------------------------------------------
+
+    def get_all_tool_schemas(self) -> List[Dict[str, Any]]:
+        """Collect tool schemas from all providers."""
+        schemas = []
+        seen = set()
+        for provider in self._providers:
+            try:
+                for schema in provider.get_tool_schemas():
+                    name = schema.get("name", "")
+                    if name and name not in seen:
+                        schemas.append(schema)
+                        seen.add(name)
+            except Exception as e:
+                logger.warning(
+                    "Memory provider '%s' get_tool_schemas() failed: %s",
+                    provider.name, e,
+                )
+        return schemas
+
+    def get_all_tool_names(self) -> set:
+        """Return set of all tool names across all providers."""
+        return set(self._tool_to_provider.keys())
+
+    def has_tool(self, tool_name: str) -> bool:
+        """Check if any provider handles this tool."""
+        return tool_name in self._tool_to_provider
+
+    def handle_tool_call(
+        self, tool_name: str, args: Dict[str, Any], **kwargs
+    ) -> str:
+        """Route a tool call to the correct provider.
+
+        Returns JSON string result. Raises ValueError if no provider
+        handles the tool.
+        """
+        provider = self._tool_to_provider.get(tool_name)
+        if provider is None:
+            return json.dumps({"error": f"No memory provider handles tool '{tool_name}'"})
+        try:
+            return provider.handle_tool_call(tool_name, args, **kwargs)
+        except Exception as e:
+            logger.error(
+                "Memory provider '%s' handle_tool_call(%s) failed: %s",
+                provider.name, tool_name, e,
+            )
+            return json.dumps({"error": f"Memory tool '{tool_name}' failed: {e}"})
+
+    # -- Lifecycle hooks -----------------------------------------------------
+
+    def on_turn_start(self, turn_number: int, message: str, **kwargs) -> None:
+        """Notify all providers of a new turn.
+
+        kwargs may include: remaining_tokens, model, platform, tool_count.
+        """
+        for provider in self._providers:
+            try:
+                provider.on_turn_start(turn_number, message, **kwargs)
+            except Exception as e:
+                logger.debug(
+                    "Memory provider '%s' on_turn_start failed: %s",
+                    provider.name, e,
+                )
+
+    def on_session_end(self, messages: List[Dict[str, Any]]) -> None:
+        """Notify all providers of session end."""
+        for provider in self._providers:
+            try:
+                provider.on_session_end(messages)
+            except Exception as e:
+                logger.debug(
+                    "Memory provider '%s' on_session_end failed: %s",
+                    provider.name, e,
+                )
+
+    def on_pre_compress(self, messages: List[Dict[str, Any]]) -> str:
+        """Notify all providers before context compression.
+
+        Returns combined text from providers to include in the compression
+        summary prompt. Empty string if no provider contributes.
+        """
+        parts = []
+        for provider in self._providers:
+            try:
+                result = provider.on_pre_compress(messages)
+                if result and result.strip():
+                    parts.append(result)
+            except Exception as e:
+                logger.debug(
+                    "Memory provider '%s' on_pre_compress failed: %s",
+                    provider.name, e,
+                )
+        return "\n\n".join(parts)
+
+    def on_memory_write(self, action: str, target: str, content: str) -> None:
+        """Notify external providers when the built-in memory tool writes.
+
+        Skips the builtin provider itself (it's the source of the write).
+        """
+        for provider in self._providers:
+            if provider.name == "builtin":
+                continue
+            try:
+                provider.on_memory_write(action, target, content)
+            except Exception as e:
+                logger.debug(
+                    "Memory provider '%s' on_memory_write failed: %s",
+                    provider.name, e,
+                )
+
+    def on_delegation(self, task: str, result: str, *,
+                      child_session_id: str = "", **kwargs) -> None:
+        """Notify all providers that a subagent completed."""
+        for provider in self._providers:
+            try:
+                provider.on_delegation(
+                    task, result, child_session_id=child_session_id, **kwargs
+                )
+            except Exception as e:
+                logger.debug(
+                    "Memory provider '%s' on_delegation failed: %s",
+                    provider.name, e,
+                )
+
+    def shutdown_all(self) -> None:
+        """Shut down all providers (reverse order for clean teardown)."""
+        for provider in reversed(self._providers):
+            try:
+                provider.shutdown()
+            except Exception as e:
+                logger.warning(
+                    "Memory provider '%s' shutdown failed: %s",
+                    provider.name, e,
+                )
+
+    def initialize_all(self, session_id: str, **kwargs) -> None:
+        """Initialize all providers.
+
+        Automatically injects ``hermes_home`` into *kwargs* so that every
+        provider can resolve profile-scoped storage paths without importing
+        ``get_hermes_home()`` themselves.
+        """
+        if "hermes_home" not in kwargs:
+            from hermes_constants import get_hermes_home
+            kwargs["hermes_home"] = str(get_hermes_home())
+        for provider in self._providers:
+            try:
+                provider.initialize(session_id=session_id, **kwargs)
+            except Exception as e:
+                logger.warning(
+                    "Memory provider '%s' initialize failed: %s",
+                    provider.name, e,
+                )

agent/memory_provider.py

@@ -0,0 +1,231 @@
+"""Abstract base class for pluggable memory providers.
+
+Memory providers give the agent persistent recall across sessions. One
+external provider is active at a time alongside the always-on built-in
+memory (MEMORY.md / USER.md). The MemoryManager enforces this limit.
+
+Built-in memory is always active as the first provider and cannot be removed.
+External providers (Honcho, Hindsight, Mem0, etc.) are additive — they never
+disable the built-in store. Only one external provider runs at a time to
+prevent tool schema bloat and conflicting memory backends.
+
+Registration:
+  1. Built-in: BuiltinMemoryProvider — always present, not removable.
+  2. Plugins: Ship in plugins/memory/<name>/, activated by memory.provider config.
+
+Lifecycle (called by MemoryManager, wired in run_agent.py):
+  initialize()          — connect, create resources, warm up
+  system_prompt_block()  — static text for the system prompt
+  prefetch(query)        — background recall before each turn
+  sync_turn(user, asst)  — async write after each turn
+  get_tool_schemas()     — tool schemas to expose to the model
+  handle_tool_call()     — dispatch a tool call
+  shutdown()             — clean exit
+
+Optional hooks (override to opt in):
+  on_turn_start(turn, message, **kwargs) — per-turn tick with runtime context
+  on_session_end(messages)               — end-of-session extraction
+  on_pre_compress(messages) -> str       — extract before context compression
+  on_memory_write(action, target, content) — mirror built-in memory writes
+  on_delegation(task, result, **kwargs)  — parent-side observation of subagent work
+"""
+
+from __future__ import annotations
+
+import logging
+from abc import ABC, abstractmethod
+from typing import Any, Dict, List, Optional
+
+logger = logging.getLogger(__name__)
+
+
+class MemoryProvider(ABC):
+    """Abstract base class for memory providers."""
+
+    @property
+    @abstractmethod
+    def name(self) -> str:
+        """Short identifier for this provider (e.g. 'builtin', 'honcho', 'hindsight')."""
+
+    # -- Core lifecycle (implement these) ------------------------------------
+
+    @abstractmethod
+    def is_available(self) -> bool:
+        """Return True if this provider is configured, has credentials, and is ready.
+
+        Called during agent init to decide whether to activate the provider.
+        Should not make network calls — just check config and installed deps.
+        """
+
+    @abstractmethod
+    def initialize(self, session_id: str, **kwargs) -> None:
+        """Initialize for a session.
+
+        Called once at agent startup. May create resources (banks, tables),
+        establish connections, start background threads, etc.
+
+        kwargs always include:
+          - hermes_home (str): The active HERMES_HOME directory path. Use this
+            for profile-scoped storage instead of hardcoding ``~/.hermes``.
+          - platform (str): "cli", "telegram", "discord", "cron", etc.
+
+        kwargs may also include:
+          - agent_context (str): "primary", "subagent", "cron", or "flush".
+            Providers should skip writes for non-primary contexts (cron system
+            prompts would corrupt user representations).
+          - agent_identity (str): Profile name (e.g. "coder"). Use for
+            per-profile provider identity scoping.
+          - agent_workspace (str): Shared workspace name (e.g. "hermes").
+          - parent_session_id (str): For subagents, the parent's session_id.
+          - user_id (str): Platform user identifier (gateway sessions).
+        """
+
+    def system_prompt_block(self) -> str:
+        """Return text to include in the system prompt.
+
+        Called during system prompt assembly. Return empty string to skip.
+        This is for STATIC provider info (instructions, status). Prefetched
+        recall context is injected separately via prefetch().
+        """
+        return ""
+
+    def prefetch(self, query: str, *, session_id: str = "") -> str:
+        """Recall relevant context for the upcoming turn.
+
+        Called before each API call. Return formatted text to inject as
+        context, or empty string if nothing relevant. Implementations
+        should be fast — use background threads for the actual recall
+        and return cached results here.
+
+        session_id is provided for providers serving concurrent sessions
+        (gateway group chats, cached agents). Providers that don't need
+        per-session scoping can ignore it.
+        """
+        return ""
+
+    def queue_prefetch(self, query: str, *, session_id: str = "") -> None:
+        """Queue a background recall for the NEXT turn.
+
+        Called after each turn completes. The result will be consumed
+        by prefetch() on the next turn. Default is no-op — providers
+        that do background prefetching should override this.
+        """
+
+    def sync_turn(self, user_content: str, assistant_content: str, *, session_id: str = "") -> None:
+        """Persist a completed turn to the backend.
+
+        Called after each turn. Should be non-blocking — queue for
+        background processing if the backend has latency.
+        """
+
+    @abstractmethod
+    def get_tool_schemas(self) -> List[Dict[str, Any]]:
+        """Return tool schemas this provider exposes.
+
+        Each schema follows the OpenAI function calling format:
+        {"name": "...", "description": "...", "parameters": {...}}
+
+        Return empty list if this provider has no tools (context-only).
+        """
+
+    def handle_tool_call(self, tool_name: str, args: Dict[str, Any], **kwargs) -> str:
+        """Handle a tool call for one of this provider's tools.
+
+        Must return a JSON string (the tool result).
+        Only called for tool names returned by get_tool_schemas().
+        """
+        raise NotImplementedError(f"Provider {self.name} does not handle tool {tool_name}")
+
+    def shutdown(self) -> None:
+        """Clean shutdown — flush queues, close connections."""
+
+    # -- Optional hooks (override to opt in) ---------------------------------
+
+    def on_turn_start(self, turn_number: int, message: str, **kwargs) -> None:
+        """Called at the start of each turn with the user message.
+
+        Use for turn-counting, scope management, periodic maintenance.
+
+        kwargs may include: remaining_tokens, model, platform, tool_count.
+        Providers use what they need; extras are ignored.
+        """
+
+    def on_session_end(self, messages: List[Dict[str, Any]]) -> None:
+        """Called when a session ends (explicit exit or timeout).
+
+        Use for end-of-session fact extraction, summarization, etc.
+        messages is the full conversation history.
+
+        NOT called after every turn — only at actual session boundaries
+        (CLI exit, /reset, gateway session expiry).
+        """
+
+    def on_pre_compress(self, messages: List[Dict[str, Any]]) -> str:
+        """Called before context compression discards old messages.
+
+        Use to extract insights from messages about to be compressed.
+        messages is the list that will be summarized/discarded.
+
+        Return text to include in the compression summary prompt so the
+        compressor preserves provider-extracted insights. Return empty
+        string for no contribution (backwards-compatible default).
+        """
+        return ""
+
+    def on_delegation(self, task: str, result: str, *,
+                      child_session_id: str = "", **kwargs) -> None:
+        """Called on the PARENT agent when a subagent completes.
+
+        The parent's memory provider gets the task+result pair as an
+        observation of what was delegated and what came back. The subagent
+        itself has no provider session (skip_memory=True).
+
+        task: the delegation prompt
+        result: the subagent's final response
+        child_session_id: the subagent's session_id
+        """
+
+    def get_config_schema(self) -> List[Dict[str, Any]]:
+        """Return config fields this provider needs for setup.
+
+        Used by 'hermes memory setup' to walk the user through configuration.
+        Each field is a dict with:
+          key:         config key name (e.g. 'api_key', 'mode')
+          description: human-readable description
+          secret:      True if this should go to .env (default: False)
+          required:    True if required (default: False)
+          default:     default value (optional)
+          choices:     list of valid values (optional)
+          url:         URL where user can get this credential (optional)
+          env_var:     explicit env var name for secrets (default: auto-generated)
+
+        Return empty list if no config needed (e.g. local-only providers).
+        """
+        return []
+
+    def save_config(self, values: Dict[str, Any], hermes_home: str) -> None:
+        """Write non-secret config to the provider's native location.
+
+        Called by 'hermes memory setup' after collecting user inputs.
+        ``values`` contains only non-secret fields (secrets go to .env).
+        ``hermes_home`` is the active HERMES_HOME directory path.
+
+        Providers with native config files (JSON, YAML) should override
+        this to write to their expected location. Providers that use only
+        env vars can leave the default (no-op).
+
+        All new memory provider plugins MUST implement either:
+        - save_config() for native config file formats, OR
+        - use only env vars (in which case get_config_schema() fields
+          should all have ``env_var`` set and this method stays no-op).
+        """
+
+    def on_memory_write(self, action: str, target: str, content: str) -> None:
+        """Called when the built-in memory tool writes an entry.
+
+        action: 'add', 'replace', or 'remove'
+        target: 'memory' or 'user'
+        content: the entry content
+
+        Use to mirror built-in memory writes to your backend.
+        """

agent/model_metadata.py

@@ -10,6 +10,7 @@ import re
 import time
 from pathlib import Path
 from typing import Any, Dict, List, Optional
+from urllib.parse import urlparse
 
 import requests
 import yaml
@@ -18,61 +19,361 @@ from hermes_constants import OPENROUTER_MODELS_URL
 
 logger = logging.getLogger(__name__)
 
+# Provider names that can appear as a "provider:" prefix before a model ID.
+# Only these are stripped — Ollama-style "model:tag" colons (e.g. "qwen3.5:27b")
+# are preserved so the full model name reaches cache lookups and server queries.
+_PROVIDER_PREFIXES: frozenset[str] = frozenset({
+    "openrouter", "nous", "openai-codex", "copilot", "copilot-acp",
+    "zai", "kimi-coding", "minimax", "minimax-cn", "anthropic", "deepseek",
+    "opencode-zen", "opencode-go", "ai-gateway", "kilocode", "alibaba",
+    "custom", "local",
+    # Common aliases
+    "glm", "z-ai", "z.ai", "zhipu", "github", "github-copilot",
+    "github-models", "kimi", "moonshot", "claude", "deep-seek",
+    "opencode", "zen", "go", "vercel", "kilo", "dashscope", "aliyun", "qwen",
+})
+
+
+_OLLAMA_TAG_PATTERN = re.compile(
+    r"^(\d+\.?\d*b|latest|stable|q\d|fp?\d|instruct|chat|coder|vision|text)",
+    re.IGNORECASE,
+)
+
+
+def _strip_provider_prefix(model: str) -> str:
+    """Strip a recognised provider prefix from a model string.
+
+    ``"local:my-model"`` → ``"my-model"``
+    ``"qwen3.5:27b"``   → ``"qwen3.5:27b"``  (unchanged — not a provider prefix)
+    ``"qwen:0.5b"``     → ``"qwen:0.5b"``    (unchanged — Ollama model:tag)
+    ``"deepseek:latest"``→ ``"deepseek:latest"``(unchanged — Ollama model:tag)
+    """
+    if ":" not in model or model.startswith("http"):
+        return model
+    prefix, suffix = model.split(":", 1)
+    prefix_lower = prefix.strip().lower()
+    if prefix_lower in _PROVIDER_PREFIXES:
+        # Don't strip if suffix looks like an Ollama tag (e.g. "7b", "latest", "q4_0")
+        if _OLLAMA_TAG_PATTERN.match(suffix.strip()):
+            return model
+        return suffix
+    return model
+
 _model_metadata_cache: Dict[str, Dict[str, Any]] = {}
 _model_metadata_cache_time: float = 0
 _MODEL_CACHE_TTL = 3600
+_endpoint_model_metadata_cache: Dict[str, Dict[str, Dict[str, Any]]] = {}
+_endpoint_model_metadata_cache_time: Dict[str, float] = {}
+_ENDPOINT_MODEL_CACHE_TTL = 300
 
 # Descending tiers for context length probing when the model is unknown.
-# We start high and step down on context-length errors until one works.
+# We start at 128K (a safe default for most modern models) and step down
+# on context-length errors until one works.
 CONTEXT_PROBE_TIERS = [
-    2_000_000,
-    1_000_000,
-    512_000,
-    200_000,
     128_000,
     64_000,
     32_000,
+    16_000,
+    8_000,
 ]
 
+# Default context length when no detection method succeeds.
+DEFAULT_FALLBACK_CONTEXT = CONTEXT_PROBE_TIERS[0]
+
+# Thin fallback defaults — only broad model family patterns.
+# These fire only when provider is unknown AND models.dev/OpenRouter/Anthropic
+# all miss. Replaced the previous 80+ entry dict.
+# For provider-specific context lengths, models.dev is the primary source.
 DEFAULT_CONTEXT_LENGTHS = {
-    "anthropic/claude-opus-4": 200000,
-    "anthropic/claude-opus-4.5": 200000,
-    "anthropic/claude-opus-4.6": 200000,
-    "anthropic/claude-sonnet-4": 200000,
-    "anthropic/claude-sonnet-4-20250514": 200000,
-    "anthropic/claude-haiku-4.5": 200000,
-    # Bare Anthropic model IDs (for native API provider)
-    "claude-opus-4-6": 200000,
-    "claude-sonnet-4-6": 200000,
-    "claude-opus-4-5-20251101": 200000,
-    "claude-sonnet-4-5-20250929": 200000,
-    "claude-opus-4-1-20250805": 200000,
-    "claude-opus-4-20250514": 200000,
-    "claude-sonnet-4-20250514": 200000,
-    "claude-haiku-4-5-20251001": 200000,
-    "openai/gpt-4o": 128000,
-    "openai/gpt-4-turbo": 128000,
-    "openai/gpt-4o-mini": 128000,
-    "google/gemini-2.0-flash": 1048576,
-    "google/gemini-2.5-pro": 1048576,
-    "meta-llama/llama-3.3-70b-instruct": 131072,
-    "deepseek/deepseek-chat-v3": 65536,
-    "qwen/qwen-2.5-72b-instruct": 32768,
-    "glm-4.7": 202752,
-    "glm-5": 202752,
-    "glm-4.5": 131072,
-    "glm-4.5-flash": 131072,
-    "kimi-for-coding": 262144,
-    "kimi-k2.5": 262144,
-    "kimi-k2-thinking": 262144,
-    "kimi-k2-thinking-turbo": 262144,
-    "kimi-k2-turbo-preview": 262144,
-    "kimi-k2-0905-preview": 131072,
-    "MiniMax-M2.5": 204800,
-    "MiniMax-M2.5-highspeed": 204800,
-    "MiniMax-M2.1": 204800,
+    # Anthropic Claude 4.6 (1M context) — bare IDs only to avoid
+    # fuzzy-match collisions (e.g. "anthropic/claude-sonnet-4" is a
+    # substring of "anthropic/claude-sonnet-4.6").
+    # OpenRouter-prefixed models resolve via OpenRouter live API or models.dev.
+    "claude-opus-4-6": 1000000,
+    "claude-sonnet-4-6": 1000000,
+    "claude-opus-4.6": 1000000,
+    "claude-sonnet-4.6": 1000000,
+    # Catch-all for older Claude models (must sort after specific entries)
+    "claude": 200000,
+    # OpenAI
+    "gpt-4.1": 1047576,
+    "gpt-5": 128000,
+    "gpt-4": 128000,
+    # Google
+    "gemini": 1048576,
+    # DeepSeek
+    "deepseek": 128000,
+    # Meta
+    "llama": 131072,
+    # Qwen
+    "qwen": 131072,
+    # MiniMax
+    "minimax": 204800,
+    # GLM
+    "glm": 202752,
+    # Kimi
+    "kimi": 262144,
+    # Arcee
+    "trinity": 262144,
+    # Hugging Face Inference Providers — model IDs use org/name format
+    "Qwen/Qwen3.5-397B-A17B": 131072,
+    "Qwen/Qwen3.5-35B-A3B": 131072,
+    "deepseek-ai/DeepSeek-V3.2": 65536,
+    "moonshotai/Kimi-K2.5": 262144,
+    "moonshotai/Kimi-K2-Thinking": 262144,
+    "MiniMaxAI/MiniMax-M2.5": 204800,
+    "XiaomiMiMo/MiMo-V2-Flash": 32768,
+    "mimo-v2-pro": 1048576,
+    "mimo-v2-omni": 1048576,
+    "zai-org/GLM-5": 202752,
 }
 
+_CONTEXT_LENGTH_KEYS = (
+    "context_length",
+    "context_window",
+    "max_context_length",
+    "max_position_embeddings",
+    "max_model_len",
+    "max_input_tokens",
+    "max_sequence_length",
+    "max_seq_len",
+    "n_ctx_train",
+    "n_ctx",
+)
+
+_MAX_COMPLETION_KEYS = (
+    "max_completion_tokens",
+    "max_output_tokens",
+    "max_tokens",
+)
+
+# Local server hostnames / address patterns
+_LOCAL_HOSTS = ("localhost", "127.0.0.1", "::1", "0.0.0.0")
+
+
+def _normalize_base_url(base_url: str) -> str:
+    return (base_url or "").strip().rstrip("/")
+
+
+def _is_openrouter_base_url(base_url: str) -> bool:
+    return "openrouter.ai" in _normalize_base_url(base_url).lower()
+
+
+def _is_custom_endpoint(base_url: str) -> bool:
+    normalized = _normalize_base_url(base_url)
+    return bool(normalized) and not _is_openrouter_base_url(normalized)
+
+
+_URL_TO_PROVIDER: Dict[str, str] = {
+    "api.openai.com": "openai",
+    "chatgpt.com": "openai",
+    "api.anthropic.com": "anthropic",
+    "api.z.ai": "zai",
+    "api.moonshot.ai": "kimi-coding",
+    "api.kimi.com": "kimi-coding",
+    "api.minimax": "minimax",
+    "dashscope.aliyuncs.com": "alibaba",
+    "dashscope-intl.aliyuncs.com": "alibaba",
+    "openrouter.ai": "openrouter",
+    "generativelanguage.googleapis.com": "google",
+    "inference-api.nousresearch.com": "nous",
+    "api.deepseek.com": "deepseek",
+    "api.githubcopilot.com": "copilot",
+    "models.github.ai": "copilot",
+    "api.fireworks.ai": "fireworks",
+}
+
+
+def _infer_provider_from_url(base_url: str) -> Optional[str]:
+    """Infer the models.dev provider name from a base URL.
+
+    This allows context length resolution via models.dev for custom endpoints
+    like DashScope (Alibaba), Z.AI, Kimi, etc. without requiring the user to
+    explicitly set the provider name in config.
+    """
+    normalized = _normalize_base_url(base_url)
+    if not normalized:
+        return None
+    parsed = urlparse(normalized if "://" in normalized else f"https://{normalized}")
+    host = parsed.netloc.lower() or parsed.path.lower()
+    for url_part, provider in _URL_TO_PROVIDER.items():
+        if url_part in host:
+            return provider
+    return None
+
+
+def _is_known_provider_base_url(base_url: str) -> bool:
+    return _infer_provider_from_url(base_url) is not None
+
+
+def is_local_endpoint(base_url: str) -> bool:
+    """Return True if base_url points to a local machine (localhost / RFC-1918 / WSL)."""
+    normalized = _normalize_base_url(base_url)
+    if not normalized:
+        return False
+    url = normalized if "://" in normalized else f"http://{normalized}"
+    try:
+        parsed = urlparse(url)
+        host = parsed.hostname or ""
+    except Exception:
+        return False
+    if host in _LOCAL_HOSTS:
+        return True
+    # RFC-1918 private ranges and link-local
+    import ipaddress
+    try:
+        addr = ipaddress.ip_address(host)
+        return addr.is_private or addr.is_loopback or addr.is_link_local
+    except ValueError:
+        pass
+    # Bare IP that looks like a private range (e.g. 172.26.x.x for WSL)
+    parts = host.split(".")
+    if len(parts) == 4:
+        try:
+            first, second = int(parts[0]), int(parts[1])
+            if first == 10:
+                return True
+            if first == 172 and 16 <= second <= 31:
+                return True
+            if first == 192 and second == 168:
+                return True
+        except ValueError:
+            pass
+    return False
+
+
+def detect_local_server_type(base_url: str) -> Optional[str]:
+    """Detect which local server is running at base_url by probing known endpoints.
+
+    Returns one of: "ollama", "lm-studio", "vllm", "llamacpp", or None.
+    """
+    import httpx
+
+    normalized = _normalize_base_url(base_url)
+    server_url = normalized
+    if server_url.endswith("/v1"):
+        server_url = server_url[:-3]
+
+    try:
+        with httpx.Client(timeout=2.0) as client:
+            # LM Studio exposes /api/v1/models — check first (most specific)
+            try:
+                r = client.get(f"{server_url}/api/v1/models")
+                if r.status_code == 200:
+                    return "lm-studio"
+            except Exception:
+                pass
+            # Ollama exposes /api/tags and responds with {"models": [...]}
+            # LM Studio returns {"error": "Unexpected endpoint"} with status 200
+            # on this path, so we must verify the response contains "models".
+            try:
+                r = client.get(f"{server_url}/api/tags")
+                if r.status_code == 200:
+                    try:
+                        data = r.json()
+                        if "models" in data:
+                            return "ollama"
+                    except Exception:
+                        pass
+            except Exception:
+                pass
+            # llama.cpp exposes /v1/props (older builds used /props without the /v1 prefix)
+            try:
+                r = client.get(f"{server_url}/v1/props")
+                if r.status_code != 200:
+                    r = client.get(f"{server_url}/props")  # fallback for older builds
+                if r.status_code == 200 and "default_generation_settings" in r.text:
+                    return "llamacpp"
+            except Exception:
+                pass
+            # vLLM: /version
+            try:
+                r = client.get(f"{server_url}/version")
+                if r.status_code == 200:
+                    data = r.json()
+                    if "version" in data:
+                        return "vllm"
+            except Exception:
+                pass
+    except Exception:
+        pass
+
+    return None
+
+
+def _iter_nested_dicts(value: Any):
+    if isinstance(value, dict):
+        yield value
+        for nested in value.values():
+            yield from _iter_nested_dicts(nested)
+    elif isinstance(value, list):
+        for item in value:
+            yield from _iter_nested_dicts(item)
+
+
+def _coerce_reasonable_int(value: Any, minimum: int = 1024, maximum: int = 10_000_000) -> Optional[int]:
+    try:
+        if isinstance(value, bool):
+            return None
+        if isinstance(value, str):
+            value = value.strip().replace(",", "")
+        result = int(value)
+    except (TypeError, ValueError):
+        return None
+    if minimum <= result <= maximum:
+        return result
+    return None
+
+
+def _extract_first_int(payload: Dict[str, Any], keys: tuple[str, ...]) -> Optional[int]:
+    keyset = {key.lower() for key in keys}
+    for mapping in _iter_nested_dicts(payload):
+        for key, value in mapping.items():
+            if str(key).lower() not in keyset:
+                continue
+            coerced = _coerce_reasonable_int(value)
+            if coerced is not None:
+                return coerced
+    return None
+
+
+def _extract_context_length(payload: Dict[str, Any]) -> Optional[int]:
+    return _extract_first_int(payload, _CONTEXT_LENGTH_KEYS)
+
+
+def _extract_max_completion_tokens(payload: Dict[str, Any]) -> Optional[int]:
+    return _extract_first_int(payload, _MAX_COMPLETION_KEYS)
+
+
+def _extract_pricing(payload: Dict[str, Any]) -> Dict[str, Any]:
+    alias_map = {
+        "prompt": ("prompt", "input", "input_cost_per_token", "prompt_token_cost"),
+        "completion": ("completion", "output", "output_cost_per_token", "completion_token_cost"),
+        "request": ("request", "request_cost"),
+        "cache_read": ("cache_read", "cached_prompt", "input_cache_read", "cache_read_cost_per_token"),
+        "cache_write": ("cache_write", "cache_creation", "input_cache_write", "cache_write_cost_per_token"),
+    }
+    for mapping in _iter_nested_dicts(payload):
+        normalized = {str(key).lower(): value for key, value in mapping.items()}
+        if not any(any(alias in normalized for alias in aliases) for aliases in alias_map.values()):
+            continue
+        pricing: Dict[str, Any] = {}
+        for target, aliases in alias_map.items():
+            for alias in aliases:
+                if alias in normalized and normalized[alias] not in (None, ""):
+                    pricing[target] = normalized[alias]
+                    break
+        if pricing:
+            return pricing
+    return {}
+
+
+def _add_model_aliases(cache: Dict[str, Dict[str, Any]], model_id: str, entry: Dict[str, Any]) -> None:
+    cache[model_id] = entry
+    if "/" in model_id:
+        bare_model = model_id.split("/", 1)[1]
+        cache.setdefault(bare_model, entry)
+
 
 def fetch_model_metadata(force_refresh: bool = False) -> Dict[str, Dict[str, Any]]:
     """Fetch model metadata from OpenRouter (cached for 1 hour)."""
@@ -89,15 +390,16 @@ def fetch_model_metadata(force_refresh: bool = False) -> Dict[str, Dict[str, Any
         cache = {}
         for model in data.get("data", []):
             model_id = model.get("id", "")
-            cache[model_id] = {
+            entry = {
                 "context_length": model.get("context_length", 128000),
                 "max_completion_tokens": model.get("top_provider", {}).get("max_completion_tokens", 4096),
                 "name": model.get("name", model_id),
                 "pricing": model.get("pricing", {}),
             }
+            _add_model_aliases(cache, model_id, entry)
             canonical = model.get("canonical_slug", "")
             if canonical and canonical != model_id:
-                cache[canonical] = cache[model_id]
+                _add_model_aliases(cache, canonical, entry)
 
         _model_metadata_cache = cache
         _model_metadata_cache_time = time.time()
@@ -109,6 +411,97 @@ def fetch_model_metadata(force_refresh: bool = False) -> Dict[str, Dict[str, Any
         return _model_metadata_cache or {}
 
 
+def fetch_endpoint_model_metadata(
+    base_url: str,
+    api_key: str = "",
+    force_refresh: bool = False,
+) -> Dict[str, Dict[str, Any]]:
+    """Fetch model metadata from an OpenAI-compatible ``/models`` endpoint.
+
+    This is used for explicit custom endpoints where hardcoded global model-name
+    defaults are unreliable. Results are cached in memory per base URL.
+    """
+    normalized = _normalize_base_url(base_url)
+    if not normalized or _is_openrouter_base_url(normalized):
+        return {}
+
+    if not force_refresh:
+        cached = _endpoint_model_metadata_cache.get(normalized)
+        cached_at = _endpoint_model_metadata_cache_time.get(normalized, 0)
+        if cached is not None and (time.time() - cached_at) < _ENDPOINT_MODEL_CACHE_TTL:
+            return cached
+
+    candidates = [normalized]
+    if normalized.endswith("/v1"):
+        alternate = normalized[:-3].rstrip("/")
+    else:
+        alternate = normalized + "/v1"
+    if alternate and alternate not in candidates:
+        candidates.append(alternate)
+
+    headers = {"Authorization": f"Bearer {api_key}"} if api_key else {}
+    last_error: Optional[Exception] = None
+
+    for candidate in candidates:
+        url = candidate.rstrip("/") + "/models"
+        try:
+            response = requests.get(url, headers=headers, timeout=10)
+            response.raise_for_status()
+            payload = response.json()
+            cache: Dict[str, Dict[str, Any]] = {}
+            for model in payload.get("data", []):
+                if not isinstance(model, dict):
+                    continue
+                model_id = model.get("id")
+                if not model_id:
+                    continue
+                entry: Dict[str, Any] = {"name": model.get("name", model_id)}
+                context_length = _extract_context_length(model)
+                if context_length is not None:
+                    entry["context_length"] = context_length
+                max_completion_tokens = _extract_max_completion_tokens(model)
+                if max_completion_tokens is not None:
+                    entry["max_completion_tokens"] = max_completion_tokens
+                pricing = _extract_pricing(model)
+                if pricing:
+                    entry["pricing"] = pricing
+                _add_model_aliases(cache, model_id, entry)
+
+            # If this is a llama.cpp server, query /props for actual allocated context
+            is_llamacpp = any(
+                m.get("owned_by") == "llamacpp"
+                for m in payload.get("data", []) if isinstance(m, dict)
+            )
+            if is_llamacpp:
+                try:
+                    # Try /v1/props first (current llama.cpp); fall back to /props for older builds
+                    base = candidate.rstrip("/").replace("/v1", "")
+                    props_resp = requests.get(base + "/v1/props", headers=headers, timeout=5)
+                    if not props_resp.ok:
+                        props_resp = requests.get(base + "/props", headers=headers, timeout=5)
+                    if props_resp.ok:
+                        props = props_resp.json()
+                        gen_settings = props.get("default_generation_settings", {})
+                        n_ctx = gen_settings.get("n_ctx")
+                        model_alias = props.get("model_alias", "")
+                        if n_ctx and model_alias and model_alias in cache:
+                            cache[model_alias]["context_length"] = n_ctx
+                except Exception:
+                    pass
+
+            _endpoint_model_metadata_cache[normalized] = cache
+            _endpoint_model_metadata_cache_time[normalized] = time.time()
+            return cache
+        except Exception as exc:
+            last_error = exc
+
+    if last_error:
+        logger.debug("Failed to fetch model metadata from %s/models: %s", normalized, last_error)
+    _endpoint_model_metadata_cache[normalized] = {}
+    _endpoint_model_metadata_cache_time[normalized] = time.time()
+    return {}
+
+
 def _get_context_cache_path() -> Path:
     """Return path to the persistent context length cache file."""
     hermes_home = Path(os.environ.get("HERMES_HOME", Path.home() / ".hermes"))
@@ -116,7 +509,7 @@ def _get_context_cache_path() -> Path:
 
 
 def _load_context_cache() -> Dict[str, int]:
-    """Load the model+provider → context_length cache from disk."""
+    """Load the model+provider -> context_length cache from disk."""
     path = _get_context_cache_path()
     if not path.exists():
         return {}
@@ -145,7 +538,7 @@ def save_context_length(model: str, base_url: str, length: int) -> None:
         path.parent.mkdir(parents=True, exist_ok=True)
         with open(path, "w") as f:
             yaml.dump({"context_lengths": cache}, f, default_flow_style=False)
-        logger.info("Cached context length %s → %s tokens", key, f"{length:,}")
+        logger.info("Cached context length %s -> %s tokens", key, f"{length:,}")
     except Exception as e:
         logger.debug("Failed to save context length cache: %s", e)
 
@@ -193,33 +586,317 @@ def parse_context_limit_from_error(error_msg: str) -> Optional[int]:
     return None
 
 
-def get_model_context_length(model: str, base_url: str = "") -> int:
+def _model_id_matches(candidate_id: str, lookup_model: str) -> bool:
+    """Return True if *candidate_id* (from server) matches *lookup_model* (configured).
+
+    Supports two forms:
+    - Exact match:  "nvidia-nemotron-super-49b-v1" == "nvidia-nemotron-super-49b-v1"
+    - Slug match:   "nvidia/nvidia-nemotron-super-49b-v1" matches "nvidia-nemotron-super-49b-v1"
+                    (the part after the last "/" equals lookup_model)
+
+    This covers LM Studio's native API which stores models as "publisher/slug"
+    while users typically configure only the slug after the "local:" prefix.
+    """
+    if candidate_id == lookup_model:
+        return True
+    # Slug match: basename of candidate equals the lookup name
+    if "/" in candidate_id and candidate_id.rsplit("/", 1)[1] == lookup_model:
+        return True
+    return False
+
+
+def _query_local_context_length(model: str, base_url: str) -> Optional[int]:
+    """Query a local server for the model's context length."""
+    import httpx
+
+    # Strip recognised provider prefix (e.g., "local:model-name" → "model-name").
+    # Ollama "model:tag" colons (e.g. "qwen3.5:27b") are intentionally preserved.
+    model = _strip_provider_prefix(model)
+
+    # Strip /v1 suffix to get the server root
+    server_url = base_url.rstrip("/")
+    if server_url.endswith("/v1"):
+        server_url = server_url[:-3]
+
+    try:
+        server_type = detect_local_server_type(base_url)
+    except Exception:
+        server_type = None
+
+    try:
+        with httpx.Client(timeout=3.0) as client:
+            # Ollama: /api/show returns model details with context info
+            if server_type == "ollama":
+                resp = client.post(f"{server_url}/api/show", json={"name": model})
+                if resp.status_code == 200:
+                    data = resp.json()
+                    # Check model_info for context length
+                    model_info = data.get("model_info", {})
+                    for key, value in model_info.items():
+                        if "context_length" in key and isinstance(value, (int, float)):
+                            return int(value)
+                    # Check parameters string for num_ctx
+                    params = data.get("parameters", "")
+                    if "num_ctx" in params:
+                        for line in params.split("\n"):
+                            if "num_ctx" in line:
+                                parts = line.strip().split()
+                                if len(parts) >= 2:
+                                    try:
+                                        return int(parts[-1])
+                                    except ValueError:
+                                        pass
+
+            # LM Studio native API: /api/v1/models returns max_context_length.
+            # This is more reliable than the OpenAI-compat /v1/models which
+            # doesn't include context window information for LM Studio servers.
+            # Use _model_id_matches for fuzzy matching: LM Studio stores models as
+            # "publisher/slug" but users configure only "slug" after "local:" prefix.
+            if server_type == "lm-studio":
+                resp = client.get(f"{server_url}/api/v1/models")
+                if resp.status_code == 200:
+                    data = resp.json()
+                    for m in data.get("models", []):
+                        if _model_id_matches(m.get("key", ""), model) or _model_id_matches(m.get("id", ""), model):
+                            # Prefer loaded instance context (actual runtime value)
+                            for inst in m.get("loaded_instances", []):
+                                cfg = inst.get("config", {})
+                                ctx = cfg.get("context_length")
+                                if ctx and isinstance(ctx, (int, float)):
+                                    return int(ctx)
+                            # Fall back to max_context_length (theoretical model max)
+                            ctx = m.get("max_context_length") or m.get("context_length")
+                            if ctx and isinstance(ctx, (int, float)):
+                                return int(ctx)
+
+            # LM Studio / vLLM / llama.cpp: try /v1/models/{model}
+            resp = client.get(f"{server_url}/v1/models/{model}")
+            if resp.status_code == 200:
+                data = resp.json()
+                # vLLM returns max_model_len
+                ctx = data.get("max_model_len") or data.get("context_length") or data.get("max_tokens")
+                if ctx and isinstance(ctx, (int, float)):
+                    return int(ctx)
+
+            # Try /v1/models and find the model in the list.
+            # Use _model_id_matches to handle "publisher/slug" vs bare "slug".
+            resp = client.get(f"{server_url}/v1/models")
+            if resp.status_code == 200:
+                data = resp.json()
+                models_list = data.get("data", [])
+                for m in models_list:
+                    if _model_id_matches(m.get("id", ""), model):
+                        ctx = m.get("max_model_len") or m.get("context_length") or m.get("max_tokens")
+                        if ctx and isinstance(ctx, (int, float)):
+                            return int(ctx)
+    except Exception:
+        pass
+
+    return None
+
+
+def _normalize_model_version(model: str) -> str:
+    """Normalize version separators for matching.
+
+    Nous uses dashes: claude-opus-4-6, claude-sonnet-4-5
+    OpenRouter uses dots: claude-opus-4.6, claude-sonnet-4.5
+    Normalize both to dashes for comparison.
+    """
+    return model.replace(".", "-")
+
+
+def _query_anthropic_context_length(model: str, base_url: str, api_key: str) -> Optional[int]:
+    """Query Anthropic's /v1/models endpoint for context length.
+
+    Only works with regular ANTHROPIC_API_KEY (sk-ant-api*).
+    OAuth tokens (sk-ant-oat*) from Claude Code return 401.
+    """
+    if not api_key or api_key.startswith("sk-ant-oat"):
+        return None  # OAuth tokens can't access /v1/models
+    try:
+        base = base_url.rstrip("/")
+        if base.endswith("/v1"):
+            base = base[:-3]
+        url = f"{base}/v1/models?limit=1000"
+        headers = {
+            "x-api-key": api_key,
+            "anthropic-version": "2023-06-01",
+        }
+        resp = requests.get(url, headers=headers, timeout=10)
+        if resp.status_code != 200:
+            return None
+        data = resp.json()
+        for m in data.get("data", []):
+            if m.get("id") == model:
+                ctx = m.get("max_input_tokens")
+                if isinstance(ctx, int) and ctx > 0:
+                    return ctx
+    except Exception as e:
+        logger.debug("Anthropic /v1/models query failed: %s", e)
+    return None
+
+
+def _resolve_nous_context_length(model: str) -> Optional[int]:
+    """Resolve Nous Portal model context length via OpenRouter metadata.
+
+    Nous model IDs are bare (e.g. 'claude-opus-4-6') while OpenRouter uses
+    prefixed IDs (e.g. 'anthropic/claude-opus-4.6'). Try suffix matching
+    with version normalization (dot↔dash).
+    """
+    metadata = fetch_model_metadata()  # OpenRouter cache
+    # Exact match first
+    if model in metadata:
+        return metadata[model].get("context_length")
+
+    normalized = _normalize_model_version(model).lower()
+
+    for or_id, entry in metadata.items():
+        bare = or_id.split("/", 1)[1] if "/" in or_id else or_id
+        if bare.lower() == model.lower() or _normalize_model_version(bare).lower() == normalized:
+            return entry.get("context_length")
+
+    # Partial prefix match for cases like gemini-3-flash → gemini-3-flash-preview
+    # Require match to be at a word boundary (followed by -, :, or end of string)
+    model_lower = model.lower()
+    for or_id, entry in metadata.items():
+        bare = or_id.split("/", 1)[1] if "/" in or_id else or_id
+        for candidate, query in [(bare.lower(), model_lower), (_normalize_model_version(bare).lower(), normalized)]:
+            if candidate.startswith(query) and (
+                len(candidate) == len(query) or candidate[len(query)] in "-:."
+            ):
+                return entry.get("context_length")
+
+    return None
+
+
+def get_model_context_length(
+    model: str,
+    base_url: str = "",
+    api_key: str = "",
+    config_context_length: int | None = None,
+    provider: str = "",
+) -> int:
     """Get the context length for a model.
 
     Resolution order:
+    0. Explicit config override (model.context_length or custom_providers per-model)
     1. Persistent cache (previously discovered via probing)
-    2. OpenRouter API metadata
-    3. Hardcoded DEFAULT_CONTEXT_LENGTHS (fuzzy match)
-    4. First probe tier (2M) — will be narrowed on first context error
+    2. Active endpoint metadata (/models for explicit custom endpoints)
+    3. Local server query (for local endpoints)
+    4. Anthropic /v1/models API (API-key users only, not OAuth)
+    5. OpenRouter live API metadata
+    6. Nous suffix-match via OpenRouter cache
+    7. models.dev registry lookup (provider-aware)
+    8. Thin hardcoded defaults (broad family patterns)
+    9. Default fallback (128K)
     """
+    # 0. Explicit config override — user knows best
+    if config_context_length is not None and isinstance(config_context_length, int) and config_context_length > 0:
+        return config_context_length
+
+    # Normalise provider-prefixed model names (e.g. "local:model-name" →
+    # "model-name") so cache lookups and server queries use the bare ID that
+    # local servers actually know about.  Ollama "model:tag" colons are preserved.
+    model = _strip_provider_prefix(model)
+
     # 1. Check persistent cache (model+provider)
     if base_url:
         cached = get_cached_context_length(model, base_url)
         if cached is not None:
             return cached
 
-    # 2. OpenRouter API metadata
+    # 2. Active endpoint metadata for truly custom/unknown endpoints.
+    # Known providers (Copilot, OpenAI, Anthropic, etc.) skip this — their
+    # /models endpoint may report a provider-imposed limit (e.g. Copilot
+    # returns 128k) instead of the model's full context (400k).  models.dev
+    # has the correct per-provider values and is checked at step 5+.
+    if _is_custom_endpoint(base_url) and not _is_known_provider_base_url(base_url):
+        endpoint_metadata = fetch_endpoint_model_metadata(base_url, api_key=api_key)
+        matched = endpoint_metadata.get(model)
+        if not matched:
+            # Single-model servers: if only one model is loaded, use it
+            if len(endpoint_metadata) == 1:
+                matched = next(iter(endpoint_metadata.values()))
+            else:
+                # Fuzzy match: substring in either direction
+                for key, entry in endpoint_metadata.items():
+                    if model in key or key in model:
+                        matched = entry
+                        break
+        if matched:
+            context_length = matched.get("context_length")
+            if isinstance(context_length, int):
+                return context_length
+        if not _is_known_provider_base_url(base_url):
+            # 3. Try querying local server directly
+            if is_local_endpoint(base_url):
+                local_ctx = _query_local_context_length(model, base_url)
+                if local_ctx and local_ctx > 0:
+                    save_context_length(model, base_url, local_ctx)
+                    return local_ctx
+            logger.info(
+                "Could not detect context length for model %r at %s — "
+                "defaulting to %s tokens (probe-down). Set model.context_length "
+                "in config.yaml to override.",
+                model, base_url, f"{DEFAULT_FALLBACK_CONTEXT:,}",
+            )
+            return DEFAULT_FALLBACK_CONTEXT
+
+    # 4. Anthropic /v1/models API (only for regular API keys, not OAuth)
+    if provider == "anthropic" or (
+        base_url and "api.anthropic.com" in base_url
+    ):
+        ctx = _query_anthropic_context_length(model, base_url or "https://api.anthropic.com", api_key)
+        if ctx:
+            return ctx
+
+    # 5. Provider-aware lookups (before generic OpenRouter cache)
+    # These are provider-specific and take priority over the generic OR cache,
+    # since the same model can have different context limits per provider
+    # (e.g. claude-opus-4.6 is 1M on Anthropic but 128K on GitHub Copilot).
+    # If provider is generic (openrouter/custom/empty), try to infer from URL.
+    effective_provider = provider
+    if not effective_provider or effective_provider in ("openrouter", "custom"):
+        if base_url:
+            inferred = _infer_provider_from_url(base_url)
+            if inferred:
+                effective_provider = inferred
+
+    if effective_provider == "nous":
+        ctx = _resolve_nous_context_length(model)
+        if ctx:
+            return ctx
+    if effective_provider:
+        from agent.models_dev import lookup_models_dev_context
+        ctx = lookup_models_dev_context(effective_provider, model)
+        if ctx:
+            return ctx
+
+    # 6. OpenRouter live API metadata (provider-unaware fallback)
     metadata = fetch_model_metadata()
     if model in metadata:
         return metadata[model].get("context_length", 128000)
 
-    # 3. Hardcoded defaults (fuzzy match)
-    for default_model, length in DEFAULT_CONTEXT_LENGTHS.items():
-        if default_model in model or model in default_model:
+    # 8. Hardcoded defaults (fuzzy match — longest key first for specificity)
+    # Only check `default_model in model` (is the key a substring of the input).
+    # The reverse (`model in default_model`) causes shorter names like
+    # "claude-sonnet-4" to incorrectly match "claude-sonnet-4-6" and return 1M.
+    model_lower = model.lower()
+    for default_model, length in sorted(
+        DEFAULT_CONTEXT_LENGTHS.items(), key=lambda x: len(x[0]), reverse=True
+    ):
+        if default_model in model_lower:
             return length
 
-    # 4. Unknown model — start at highest probe tier
-    return CONTEXT_PROBE_TIERS[0]
+    # 9. Query local server as last resort
+    if base_url and is_local_endpoint(base_url):
+        local_ctx = _query_local_context_length(model, base_url)
+        if local_ctx and local_ctx > 0:
+            save_context_length(model, base_url, local_ctx)
+            return local_ctx
+
+    # 10. Default fallback — 128K
+    return DEFAULT_FALLBACK_CONTEXT
 
 
 def estimate_tokens_rough(text: str) -> int:
@@ -233,3 +910,26 @@ def estimate_messages_tokens_rough(messages: List[Dict[str, Any]]) -> int:
     """Rough token estimate for a message list (pre-flight only)."""
     total_chars = sum(len(str(msg)) for msg in messages)
     return total_chars // 4
+
+
+def estimate_request_tokens_rough(
+    messages: List[Dict[str, Any]],
+    *,
+    system_prompt: str = "",
+    tools: Optional[List[Dict[str, Any]]] = None,
+) -> int:
+    """Rough token estimate for a full chat-completions request.
+
+    Includes the major payload buckets Hermes sends to providers:
+    system prompt, conversation messages, and tool schemas.  With 50+
+    tools enabled, schemas alone can add 20-30K tokens — a significant
+    blind spot when only counting messages.
+    """
+    total_chars = 0
+    if system_prompt:
+        total_chars += len(system_prompt)
+    if messages:
+        total_chars += sum(len(str(msg)) for msg in messages)
+    if tools:
+        total_chars += len(str(tools))
+    return total_chars // 4

agent/models_dev.py

@@ -0,0 +1,747 @@
+"""Models.dev registry integration — primary database for providers and models.
+
+Fetches from https://models.dev/api.json — a community-maintained database
+of 4000+ models across 109+ providers.  Provides:
+
+- **Provider metadata**: name, base URL, env vars, documentation link
+- **Model metadata**: context window, max output, cost/M tokens, capabilities
+  (reasoning, tools, vision, PDF, audio), modalities, knowledge cutoff,
+  open-weights flag, family grouping, deprecation status
+
+Data resolution order (like TypeScript OpenCode):
+  1. Bundled snapshot (ships with the package — offline-first)
+  2. Disk cache (~/.hermes/models_dev_cache.json)
+  3. Network fetch (https://models.dev/api.json)
+  4. Background refresh every 60 minutes
+
+Other modules should import the dataclasses and query functions from here
+rather than parsing the raw JSON themselves.
+"""
+
+import difflib
+import json
+import logging
+import os
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+from utils import atomic_json_write
+
+import requests
+
+logger = logging.getLogger(__name__)
+
+MODELS_DEV_URL = "https://models.dev/api.json"
+_MODELS_DEV_CACHE_TTL = 3600  # 1 hour in-memory
+
+# In-memory cache
+_models_dev_cache: Dict[str, Any] = {}
+_models_dev_cache_time: float = 0
+
+
+# ---------------------------------------------------------------------------
+# Dataclasses — rich metadata for providers and models
+# ---------------------------------------------------------------------------
+
+@dataclass
+class ModelInfo:
+    """Full metadata for a single model from models.dev."""
+
+    id: str
+    name: str
+    family: str
+    provider_id: str        # models.dev provider ID (e.g. "anthropic")
+
+    # Capabilities
+    reasoning: bool = False
+    tool_call: bool = False
+    attachment: bool = False       # supports image/file attachments (vision)
+    temperature: bool = False
+    structured_output: bool = False
+    open_weights: bool = False
+
+    # Modalities
+    input_modalities: Tuple[str, ...] = ()    # ("text", "image", "pdf", ...)
+    output_modalities: Tuple[str, ...] = ()
+
+    # Limits
+    context_window: int = 0
+    max_output: int = 0
+    max_input: Optional[int] = None
+
+    # Cost (per million tokens, USD)
+    cost_input: float = 0.0
+    cost_output: float = 0.0
+    cost_cache_read: Optional[float] = None
+    cost_cache_write: Optional[float] = None
+
+    # Metadata
+    knowledge_cutoff: str = ""
+    release_date: str = ""
+    status: str = ""          # "alpha", "beta", "deprecated", or ""
+    interleaved: Any = False  # True or {"field": "reasoning_content"}
+
+    def has_cost_data(self) -> bool:
+        return self.cost_input > 0 or self.cost_output > 0
+
+    def supports_vision(self) -> bool:
+        return self.attachment or "image" in self.input_modalities
+
+    def supports_pdf(self) -> bool:
+        return "pdf" in self.input_modalities
+
+    def supports_audio_input(self) -> bool:
+        return "audio" in self.input_modalities
+
+    def format_cost(self) -> str:
+        """Human-readable cost string, e.g. '$3.00/M in, $15.00/M out'."""
+        if not self.has_cost_data():
+            return "unknown"
+        parts = [f"${self.cost_input:.2f}/M in", f"${self.cost_output:.2f}/M out"]
+        if self.cost_cache_read is not None:
+            parts.append(f"cache read ${self.cost_cache_read:.2f}/M")
+        return ", ".join(parts)
+
+    def format_capabilities(self) -> str:
+        """Human-readable capabilities, e.g. 'reasoning, tools, vision, PDF'."""
+        caps = []
+        if self.reasoning:
+            caps.append("reasoning")
+        if self.tool_call:
+            caps.append("tools")
+        if self.supports_vision():
+            caps.append("vision")
+        if self.supports_pdf():
+            caps.append("PDF")
+        if self.supports_audio_input():
+            caps.append("audio")
+        if self.structured_output:
+            caps.append("structured output")
+        if self.open_weights:
+            caps.append("open weights")
+        return ", ".join(caps) if caps else "basic"
+
+
+@dataclass
+class ProviderInfo:
+    """Full metadata for a provider from models.dev."""
+
+    id: str                         # models.dev provider ID
+    name: str                       # display name
+    env: Tuple[str, ...]            # env var names for API key
+    api: str                        # base URL
+    doc: str = ""                   # documentation URL
+    model_count: int = 0
+
+    def has_api_url(self) -> bool:
+        return bool(self.api)
+
+
+# ---------------------------------------------------------------------------
+# Provider ID mapping: Hermes ↔ models.dev
+# ---------------------------------------------------------------------------
+
+# Hermes provider names → models.dev provider IDs
+PROVIDER_TO_MODELS_DEV: Dict[str, str] = {
+    "openrouter": "openrouter",
+    "anthropic": "anthropic",
+    "zai": "zai",
+    "kimi-coding": "kimi-for-coding",
+    "minimax": "minimax",
+    "minimax-cn": "minimax-cn",
+    "deepseek": "deepseek",
+    "alibaba": "alibaba",
+    "copilot": "github-copilot",
+    "ai-gateway": "vercel",
+    "opencode-zen": "opencode",
+    "opencode-go": "opencode-go",
+    "kilocode": "kilo",
+    "fireworks": "fireworks-ai",
+    "huggingface": "huggingface",
+    "google": "google",
+    "xai": "xai",
+    "nvidia": "nvidia",
+    "groq": "groq",
+    "mistral": "mistral",
+    "togetherai": "togetherai",
+    "perplexity": "perplexity",
+    "cohere": "cohere",
+}
+
+# Reverse mapping: models.dev → Hermes (built lazily)
+_MODELS_DEV_TO_PROVIDER: Optional[Dict[str, str]] = None
+
+
+def _get_reverse_mapping() -> Dict[str, str]:
+    """Return models.dev ID → Hermes provider ID mapping."""
+    global _MODELS_DEV_TO_PROVIDER
+    if _MODELS_DEV_TO_PROVIDER is None:
+        _MODELS_DEV_TO_PROVIDER = {v: k for k, v in PROVIDER_TO_MODELS_DEV.items()}
+    return _MODELS_DEV_TO_PROVIDER
+
+
+def _get_cache_path() -> Path:
+    """Return path to disk cache file."""
+    env_val = os.environ.get("HERMES_HOME", "")
+    hermes_home = Path(env_val) if env_val else Path.home() / ".hermes"
+    return hermes_home / "models_dev_cache.json"
+
+
+def _load_disk_cache() -> Dict[str, Any]:
+    """Load models.dev data from disk cache."""
+    try:
+        cache_path = _get_cache_path()
+        if cache_path.exists():
+            with open(cache_path, encoding="utf-8") as f:
+                return json.load(f)
+    except Exception as e:
+        logger.debug("Failed to load models.dev disk cache: %s", e)
+    return {}
+
+
+def _save_disk_cache(data: Dict[str, Any]) -> None:
+    """Save models.dev data to disk cache atomically."""
+    try:
+        cache_path = _get_cache_path()
+        atomic_json_write(cache_path, data, indent=None, separators=(",", ":"))
+    except Exception as e:
+        logger.debug("Failed to save models.dev disk cache: %s", e)
+
+
+def fetch_models_dev(force_refresh: bool = False) -> Dict[str, Any]:
+    """Fetch models.dev registry. In-memory cache (1hr) + disk fallback.
+
+    Returns the full registry dict keyed by provider ID, or empty dict on failure.
+    """
+    global _models_dev_cache, _models_dev_cache_time
+
+    # Check in-memory cache
+    if (
+        not force_refresh
+        and _models_dev_cache
+        and (time.time() - _models_dev_cache_time) < _MODELS_DEV_CACHE_TTL
+    ):
+        return _models_dev_cache
+
+    # Try network fetch
+    try:
+        response = requests.get(MODELS_DEV_URL, timeout=15)
+        response.raise_for_status()
+        data = response.json()
+        if isinstance(data, dict) and len(data) > 0:
+            _models_dev_cache = data
+            _models_dev_cache_time = time.time()
+            _save_disk_cache(data)
+            logger.debug(
+                "Fetched models.dev registry: %d providers, %d total models",
+                len(data),
+                sum(len(p.get("models", {})) for p in data.values() if isinstance(p, dict)),
+            )
+            return data
+    except Exception as e:
+        logger.debug("Failed to fetch models.dev: %s", e)
+
+    # Fall back to disk cache — use a short TTL (5 min) so we retry
+    # the network fetch soon instead of serving stale data for a full hour.
+    if not _models_dev_cache:
+        _models_dev_cache = _load_disk_cache()
+        if _models_dev_cache:
+            _models_dev_cache_time = time.time() - _MODELS_DEV_CACHE_TTL + 300
+            logger.debug("Loaded models.dev from disk cache (%d providers)", len(_models_dev_cache))
+
+    return _models_dev_cache
+
+
+def lookup_models_dev_context(provider: str, model: str) -> Optional[int]:
+    """Look up context_length for a provider+model combo in models.dev.
+
+    Returns the context window in tokens, or None if not found.
+    Handles case-insensitive matching and filters out context=0 entries.
+    """
+    mdev_provider_id = PROVIDER_TO_MODELS_DEV.get(provider)
+    if not mdev_provider_id:
+        return None
+
+    data = fetch_models_dev()
+    provider_data = data.get(mdev_provider_id)
+    if not isinstance(provider_data, dict):
+        return None
+
+    models = provider_data.get("models", {})
+    if not isinstance(models, dict):
+        return None
+
+    # Exact match
+    entry = models.get(model)
+    if entry:
+        ctx = _extract_context(entry)
+        if ctx:
+            return ctx
+
+    # Case-insensitive match
+    model_lower = model.lower()
+    for mid, mdata in models.items():
+        if mid.lower() == model_lower:
+            ctx = _extract_context(mdata)
+            if ctx:
+                return ctx
+
+    return None
+
+
+def _extract_context(entry: Dict[str, Any]) -> Optional[int]:
+    """Extract context_length from a models.dev model entry.
+
+    Returns None for invalid/zero values (some audio/image models have context=0).
+    """
+    if not isinstance(entry, dict):
+        return None
+    limit = entry.get("limit")
+    if not isinstance(limit, dict):
+        return None
+    ctx = limit.get("context")
+    if isinstance(ctx, (int, float)) and ctx > 0:
+        return int(ctx)
+    return None
+
+
+# ---------------------------------------------------------------------------
+# Model capability metadata
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class ModelCapabilities:
+    """Structured capability metadata for a model from models.dev."""
+
+    supports_tools: bool = True
+    supports_vision: bool = False
+    supports_reasoning: bool = False
+    context_window: int = 200000
+    max_output_tokens: int = 8192
+    model_family: str = ""
+
+
+def _get_provider_models(provider: str) -> Optional[Dict[str, Any]]:
+    """Resolve a Hermes provider ID to its models dict from models.dev.
+
+    Returns the models dict or None if the provider is unknown or has no data.
+    """
+    mdev_provider_id = PROVIDER_TO_MODELS_DEV.get(provider)
+    if not mdev_provider_id:
+        return None
+
+    data = fetch_models_dev()
+    provider_data = data.get(mdev_provider_id)
+    if not isinstance(provider_data, dict):
+        return None
+
+    models = provider_data.get("models", {})
+    if not isinstance(models, dict):
+        return None
+
+    return models
+
+
+def _find_model_entry(models: Dict[str, Any], model: str) -> Optional[Dict[str, Any]]:
+    """Find a model entry by exact match, then case-insensitive fallback."""
+    # Exact match
+    entry = models.get(model)
+    if isinstance(entry, dict):
+        return entry
+
+    # Case-insensitive match
+    model_lower = model.lower()
+    for mid, mdata in models.items():
+        if mid.lower() == model_lower and isinstance(mdata, dict):
+            return mdata
+
+    return None
+
+
+def get_model_capabilities(provider: str, model: str) -> Optional[ModelCapabilities]:
+    """Look up full capability metadata from models.dev cache.
+
+    Uses the existing fetch_models_dev() and PROVIDER_TO_MODELS_DEV mapping.
+    Returns None if model not found.
+
+    Extracts from model entry fields:
+      - reasoning  (bool)  → supports_reasoning
+      - tool_call  (bool)  → supports_tools
+      - attachment (bool)  → supports_vision
+      - limit.context (int) → context_window
+      - limit.output  (int) → max_output_tokens
+      - family     (str)   → model_family
+    """
+    models = _get_provider_models(provider)
+    if models is None:
+        return None
+
+    entry = _find_model_entry(models, model)
+    if entry is None:
+        return None
+
+    # Extract capability flags (default to False if missing)
+    supports_tools = bool(entry.get("tool_call", False))
+    supports_vision = bool(entry.get("attachment", False))
+    supports_reasoning = bool(entry.get("reasoning", False))
+
+    # Extract limits
+    limit = entry.get("limit", {})
+    if not isinstance(limit, dict):
+        limit = {}
+
+    ctx = limit.get("context")
+    context_window = int(ctx) if isinstance(ctx, (int, float)) and ctx > 0 else 200000
+
+    out = limit.get("output")
+    max_output_tokens = int(out) if isinstance(out, (int, float)) and out > 0 else 8192
+
+    model_family = entry.get("family", "") or ""
+
+    return ModelCapabilities(
+        supports_tools=supports_tools,
+        supports_vision=supports_vision,
+        supports_reasoning=supports_reasoning,
+        context_window=context_window,
+        max_output_tokens=max_output_tokens,
+        model_family=model_family,
+    )
+
+
+def list_provider_models(provider: str) -> List[str]:
+    """Return all model IDs for a provider from models.dev.
+
+    Returns an empty list if the provider is unknown or has no data.
+    """
+    models = _get_provider_models(provider)
+    if models is None:
+        return []
+    return list(models.keys())
+
+
+def search_models_dev(
+    query: str, provider: str = None, limit: int = 5
+) -> List[Dict[str, Any]]:
+    """Fuzzy search across models.dev catalog. Returns matching model entries.
+
+    Args:
+        query: Search string to match against model IDs.
+        provider: Optional Hermes provider ID to restrict search scope.
+                  If None, searches across all providers in PROVIDER_TO_MODELS_DEV.
+        limit: Maximum number of results to return.
+
+    Returns:
+        List of dicts, each containing 'provider', 'model_id', and the full
+        model 'entry' from models.dev.
+    """
+    data = fetch_models_dev()
+    if not data:
+        return []
+
+    # Build list of (provider_id, model_id, entry) candidates
+    candidates: List[tuple] = []
+
+    if provider is not None:
+        # Search only the specified provider
+        mdev_provider_id = PROVIDER_TO_MODELS_DEV.get(provider)
+        if not mdev_provider_id:
+            return []
+        provider_data = data.get(mdev_provider_id, {})
+        if isinstance(provider_data, dict):
+            models = provider_data.get("models", {})
+            if isinstance(models, dict):
+                for mid, mdata in models.items():
+                    candidates.append((provider, mid, mdata))
+    else:
+        # Search across all mapped providers
+        for hermes_prov, mdev_prov in PROVIDER_TO_MODELS_DEV.items():
+            provider_data = data.get(mdev_prov, {})
+            if isinstance(provider_data, dict):
+                models = provider_data.get("models", {})
+                if isinstance(models, dict):
+                    for mid, mdata in models.items():
+                        candidates.append((hermes_prov, mid, mdata))
+
+    if not candidates:
+        return []
+
+    # Use difflib for fuzzy matching — case-insensitive comparison
+    model_ids_lower = [c[1].lower() for c in candidates]
+    query_lower = query.lower()
+
+    # First try exact substring matches (more intuitive than pure edit-distance)
+    substring_matches = []
+    for prov, mid, mdata in candidates:
+        if query_lower in mid.lower():
+            substring_matches.append({"provider": prov, "model_id": mid, "entry": mdata})
+
+    # Then add difflib fuzzy matches for any remaining slots
+    fuzzy_ids = difflib.get_close_matches(
+        query_lower, model_ids_lower, n=limit * 2, cutoff=0.4
+    )
+
+    seen_ids: set = set()
+    results: List[Dict[str, Any]] = []
+
+    # Prioritize substring matches
+    for match in substring_matches:
+        key = (match["provider"], match["model_id"])
+        if key not in seen_ids:
+            seen_ids.add(key)
+            results.append(match)
+            if len(results) >= limit:
+                return results
+
+    # Add fuzzy matches
+    for fid in fuzzy_ids:
+        # Find original-case candidates matching this lowered ID
+        for prov, mid, mdata in candidates:
+            if mid.lower() == fid:
+                key = (prov, mid)
+                if key not in seen_ids:
+                    seen_ids.add(key)
+                    results.append({"provider": prov, "model_id": mid, "entry": mdata})
+                    if len(results) >= limit:
+                        return results
+
+    return results
+
+
+# ---------------------------------------------------------------------------
+# Rich dataclass constructors — parse raw models.dev JSON into dataclasses
+# ---------------------------------------------------------------------------
+
+def _parse_model_info(model_id: str, raw: Dict[str, Any], provider_id: str) -> ModelInfo:
+    """Convert a raw models.dev model entry dict into a ModelInfo dataclass."""
+    limit = raw.get("limit") or {}
+    if not isinstance(limit, dict):
+        limit = {}
+
+    cost = raw.get("cost") or {}
+    if not isinstance(cost, dict):
+        cost = {}
+
+    modalities = raw.get("modalities") or {}
+    if not isinstance(modalities, dict):
+        modalities = {}
+
+    input_mods = modalities.get("input") or []
+    output_mods = modalities.get("output") or []
+
+    ctx = limit.get("context")
+    ctx_int = int(ctx) if isinstance(ctx, (int, float)) and ctx > 0 else 0
+    out = limit.get("output")
+    out_int = int(out) if isinstance(out, (int, float)) and out > 0 else 0
+    inp = limit.get("input")
+    inp_int = int(inp) if isinstance(inp, (int, float)) and inp > 0 else None
+
+    return ModelInfo(
+        id=model_id,
+        name=raw.get("name", "") or model_id,
+        family=raw.get("family", "") or "",
+        provider_id=provider_id,
+        reasoning=bool(raw.get("reasoning", False)),
+        tool_call=bool(raw.get("tool_call", False)),
+        attachment=bool(raw.get("attachment", False)),
+        temperature=bool(raw.get("temperature", False)),
+        structured_output=bool(raw.get("structured_output", False)),
+        open_weights=bool(raw.get("open_weights", False)),
+        input_modalities=tuple(input_mods) if isinstance(input_mods, list) else (),
+        output_modalities=tuple(output_mods) if isinstance(output_mods, list) else (),
+        context_window=ctx_int,
+        max_output=out_int,
+        max_input=inp_int,
+        cost_input=float(cost.get("input", 0) or 0),
+        cost_output=float(cost.get("output", 0) or 0),
+        cost_cache_read=float(cost["cache_read"]) if "cache_read" in cost and cost["cache_read"] is not None else None,
+        cost_cache_write=float(cost["cache_write"]) if "cache_write" in cost and cost["cache_write"] is not None else None,
+        knowledge_cutoff=raw.get("knowledge", "") or "",
+        release_date=raw.get("release_date", "") or "",
+        status=raw.get("status", "") or "",
+        interleaved=raw.get("interleaved", False),
+    )
+
+
+def _parse_provider_info(provider_id: str, raw: Dict[str, Any]) -> ProviderInfo:
+    """Convert a raw models.dev provider entry dict into a ProviderInfo."""
+    env = raw.get("env") or []
+    models = raw.get("models") or {}
+    return ProviderInfo(
+        id=provider_id,
+        name=raw.get("name", "") or provider_id,
+        env=tuple(env) if isinstance(env, list) else (),
+        api=raw.get("api", "") or "",
+        doc=raw.get("doc", "") or "",
+        model_count=len(models) if isinstance(models, dict) else 0,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Provider-level queries
+# ---------------------------------------------------------------------------
+
+def get_provider_info(provider_id: str) -> Optional[ProviderInfo]:
+    """Get full provider metadata from models.dev.
+
+    Accepts either a Hermes provider ID (e.g. "kilocode") or a models.dev
+    ID (e.g. "kilo").  Returns None if the provider is not in the catalog.
+    """
+    # Resolve Hermes ID → models.dev ID
+    mdev_id = PROVIDER_TO_MODELS_DEV.get(provider_id, provider_id)
+
+    data = fetch_models_dev()
+    raw = data.get(mdev_id)
+    if not isinstance(raw, dict):
+        return None
+
+    return _parse_provider_info(mdev_id, raw)
+
+
+def list_all_providers() -> Dict[str, ProviderInfo]:
+    """Return all providers from models.dev as {provider_id: ProviderInfo}.
+
+    Returns the full catalog — 109+ providers.  For providers that have
+    a Hermes alias, both the models.dev ID and the Hermes ID are included.
+    """
+    data = fetch_models_dev()
+    result: Dict[str, ProviderInfo] = {}
+
+    for pid, pdata in data.items():
+        if isinstance(pdata, dict):
+            info = _parse_provider_info(pid, pdata)
+            result[pid] = info
+
+    return result
+
+
+def get_providers_for_env_var(env_var: str) -> List[str]:
+    """Reverse lookup: find all providers that use a given env var.
+
+    Useful for auto-detection: "user has ANTHROPIC_API_KEY set, which
+    providers does that enable?"
+
+    Returns list of models.dev provider IDs.
+    """
+    data = fetch_models_dev()
+    matches: List[str] = []
+
+    for pid, pdata in data.items():
+        if isinstance(pdata, dict):
+            env = pdata.get("env", [])
+            if isinstance(env, list) and env_var in env:
+                matches.append(pid)
+
+    return matches
+
+
+# ---------------------------------------------------------------------------
+# Model-level queries (rich ModelInfo)
+# ---------------------------------------------------------------------------
+
+def get_model_info(
+    provider_id: str, model_id: str
+) -> Optional[ModelInfo]:
+    """Get full model metadata from models.dev.
+
+    Accepts Hermes or models.dev provider ID.  Tries exact match then
+    case-insensitive fallback.  Returns None if not found.
+    """
+    mdev_id = PROVIDER_TO_MODELS_DEV.get(provider_id, provider_id)
+
+    data = fetch_models_dev()
+    pdata = data.get(mdev_id)
+    if not isinstance(pdata, dict):
+        return None
+
+    models = pdata.get("models", {})
+    if not isinstance(models, dict):
+        return None
+
+    # Exact match
+    raw = models.get(model_id)
+    if isinstance(raw, dict):
+        return _parse_model_info(model_id, raw, mdev_id)
+
+    # Case-insensitive fallback
+    model_lower = model_id.lower()
+    for mid, mdata in models.items():
+        if mid.lower() == model_lower and isinstance(mdata, dict):
+            return _parse_model_info(mid, mdata, mdev_id)
+
+    return None
+
+
+def get_model_info_any_provider(model_id: str) -> Optional[ModelInfo]:
+    """Search all providers for a model by ID.
+
+    Useful when you have a full slug like "anthropic/claude-sonnet-4.6" or
+    a bare name and want to find it anywhere.  Checks Hermes-mapped providers
+    first, then falls back to all models.dev providers.
+    """
+    data = fetch_models_dev()
+
+    # Try Hermes-mapped providers first (more likely what the user wants)
+    for hermes_id, mdev_id in PROVIDER_TO_MODELS_DEV.items():
+        pdata = data.get(mdev_id)
+        if not isinstance(pdata, dict):
+            continue
+        models = pdata.get("models", {})
+        if not isinstance(models, dict):
+            continue
+
+        raw = models.get(model_id)
+        if isinstance(raw, dict):
+            return _parse_model_info(model_id, raw, mdev_id)
+
+        # Case-insensitive
+        model_lower = model_id.lower()
+        for mid, mdata in models.items():
+            if mid.lower() == model_lower and isinstance(mdata, dict):
+                return _parse_model_info(mid, mdata, mdev_id)
+
+    # Fall back to ALL providers
+    for pid, pdata in data.items():
+        if pid in _get_reverse_mapping():
+            continue  # already checked
+        if not isinstance(pdata, dict):
+            continue
+        models = pdata.get("models", {})
+        if not isinstance(models, dict):
+            continue
+
+        raw = models.get(model_id)
+        if isinstance(raw, dict):
+            return _parse_model_info(model_id, raw, pid)
+
+    return None
+
+
+def list_provider_model_infos(provider_id: str) -> List[ModelInfo]:
+    """Return all models for a provider as ModelInfo objects.
+
+    Filters out deprecated models by default.
+    """
+    mdev_id = PROVIDER_TO_MODELS_DEV.get(provider_id, provider_id)
+
+    data = fetch_models_dev()
+    pdata = data.get(mdev_id)
+    if not isinstance(pdata, dict):
+        return []
+
+    models = pdata.get("models", {})
+    if not isinstance(models, dict):
+        return []
+
+    result: List[ModelInfo] = []
+    for mid, mdata in models.items():
+        if not isinstance(mdata, dict):
+            continue
+        status = mdata.get("status", "")
+        if status == "deprecated":
+            continue
+        result.append(_parse_model_info(mid, mdata, mdev_id))
+
+    return result

agent/prompt_builder.py

@@ -4,12 +4,28 @@ All functions are stateless. AIAgent._build_system_prompt() calls these to
 assemble pieces, then combines them with memory and ephemeral prompts.
 """
 
+import json
 import logging
 import os
 import re
+import threading
+from collections import OrderedDict
 from pathlib import Path
+
+from hermes_constants import get_hermes_home
 from typing import Optional
 
+from agent.skill_utils import (
+    extract_skill_conditions,
+    extract_skill_description,
+    get_all_skills_dirs,
+    get_disabled_skill_names,
+    iter_skill_index_files,
+    parse_frontmatter,
+    skill_matches_platform,
+)
+from utils import atomic_json_write
+
 logger = logging.getLogger(__name__)
 
 # ---------------------------------------------------------------------------
@@ -56,6 +72,61 @@ def _scan_context_content(content: str, filename: str) -> str:
 
     return content
 
+
+def _find_git_root(start: Path) -> Optional[Path]:
+    """Walk *start* and its parents looking for a ``.git`` directory.
+
+    Returns the directory containing ``.git``, or ``None`` if we hit the
+    filesystem root without finding one.
+    """
+    current = start.resolve()
+    for parent in [current, *current.parents]:
+        if (parent / ".git").exists():
+            return parent
+    return None
+
+
+_HERMES_MD_NAMES = (".hermes.md", "HERMES.md")
+
+
+def _find_hermes_md(cwd: Path) -> Optional[Path]:
+    """Discover the nearest ``.hermes.md`` or ``HERMES.md``.
+
+    Search order: *cwd* first, then each parent directory up to (and
+    including) the git repository root.  Returns the first match, or
+    ``None`` if nothing is found.
+    """
+    stop_at = _find_git_root(cwd)
+    current = cwd.resolve()
+
+    for directory in [current, *current.parents]:
+        for name in _HERMES_MD_NAMES:
+            candidate = directory / name
+            if candidate.is_file():
+                return candidate
+        # Stop walking at the git root (or filesystem root).
+        if stop_at and directory == stop_at:
+            break
+    return None
+
+
+def _strip_yaml_frontmatter(content: str) -> str:
+    """Remove optional YAML frontmatter (``---`` delimited) from *content*.
+
+    The frontmatter may contain structured config (model overrides, tool
+    settings) that will be handled separately in a future PR.  For now we
+    strip it so only the human-readable markdown body is injected into the
+    system prompt.
+    """
+    if content.startswith("---"):
+        end = content.find("\n---", 3)
+        if end != -1:
+            # Skip past the closing --- and any trailing newline
+            body = content[end + 4:].lstrip("\n")
+            return body if body else content
+    return content
+
+
 # =========================================================================
 # Constants
 # =========================================================================
@@ -73,9 +144,15 @@ DEFAULT_AGENT_IDENTITY = (
 MEMORY_GUIDANCE = (
     "You have persistent memory across sessions. Save durable facts using the memory "
     "tool: user preferences, environment details, tool quirks, and stable conventions. "
-    "Memory is injected into every turn, so keep it compact. Do NOT save task progress, "
-    "session outcomes, or completed-work logs to memory; use session_search to recall "
-    "those from past transcripts."
+    "Memory is injected into every turn, so keep it compact and focused on facts that "
+    "will still matter later.\n"
+    "Prioritize what reduces future user steering — the most valuable memory is one "
+    "that prevents the user from having to correct or remind you again. "
+    "User preferences and recurring corrections matter more than procedural task details.\n"
+    "Do NOT save task progress, session outcomes, completed-work logs, or temporary TODO "
+    "state to memory; use session_search to recall those from past transcripts. "
+    "If you've discovered a new way to do something, solved a problem that could be "
+    "necessary later, save it as a skill with the skill tool."
 )
 
 SESSION_SEARCH_GUIDANCE = (
@@ -86,10 +163,61 @@ SESSION_SEARCH_GUIDANCE = (
 
 SKILLS_GUIDANCE = (
     "After completing a complex task (5+ tool calls), fixing a tricky error, "
-    "or discovering a non-trivial workflow, consider saving the approach as a "
-    "skill with skill_manage so you can reuse it next time."
+    "or discovering a non-trivial workflow, save the approach as a "
+    "skill with skill_manage so you can reuse it next time.\n"
+    "When using a skill and finding it outdated, incomplete, or wrong, "
+    "patch it immediately with skill_manage(action='patch') — don't wait to be asked. "
+    "Skills that aren't maintained become liabilities."
 )
 
+TOOL_USE_ENFORCEMENT_GUIDANCE = (
+    "# Tool-use enforcement\n"
+    "You MUST use your tools to take action — do not describe what you would do "
+    "or plan to do without actually doing it. When you say you will perform an "
+    "action (e.g. 'I will run the tests', 'Let me check the file', 'I will create "
+    "the project'), you MUST immediately make the corresponding tool call in the same "
+    "response. Never end your turn with a promise of future action — execute it now.\n"
+    "Keep working until the task is actually complete. Do not stop with a summary of "
+    "what you plan to do next time. If you have tools available that can accomplish "
+    "the task, use them instead of telling the user what you would do.\n"
+    "Every response should either (a) contain tool calls that make progress, or "
+    "(b) deliver a final result to the user. Responses that only describe intentions "
+    "without acting are not acceptable."
+)
+
+# Model name substrings that trigger tool-use enforcement guidance.
+# Add new patterns here when a model family needs explicit steering.
+TOOL_USE_ENFORCEMENT_MODELS = ("gpt", "codex", "gemini", "gemma")
+
+# Gemini/Gemma-specific operational guidance, adapted from OpenCode's gemini.txt.
+# Injected alongside TOOL_USE_ENFORCEMENT_GUIDANCE when the model is Gemini or Gemma.
+GOOGLE_MODEL_OPERATIONAL_GUIDANCE = (
+    "# Google model operational directives\n"
+    "Follow these operational rules strictly:\n"
+    "- **Absolute paths:** Always construct and use absolute file paths for all "
+    "file system operations. Combine the project root with relative paths.\n"
+    "- **Verify first:** Use read_file/search_files to check file contents and "
+    "project structure before making changes. Never guess at file contents.\n"
+    "- **Dependency checks:** Never assume a library is available. Check "
+    "package.json, requirements.txt, Cargo.toml, etc. before importing.\n"
+    "- **Conciseness:** Keep explanatory text brief — a few sentences, not "
+    "paragraphs. Focus on actions and results over narration.\n"
+    "- **Parallel tool calls:** When you need to perform multiple independent "
+    "operations (e.g. reading several files), make all the tool calls in a "
+    "single response rather than sequentially.\n"
+    "- **Non-interactive commands:** Use flags like -y, --yes, --non-interactive "
+    "to prevent CLI tools from hanging on prompts.\n"
+    "- **Keep going:** Work autonomously until the task is fully resolved. "
+    "Don't stop with a plan — execute it.\n"
+)
+
+# Model name substrings that should use the 'developer' role instead of
+# 'system' for the system prompt.  OpenAI's newer models (GPT-5, Codex)
+# give stronger instruction-following weight to the 'developer' role.
+# The swap happens at the API boundary in _build_api_kwargs() so internal
+# message representation stays consistent ("system" everywhere).
+DEVELOPER_ROLE_MODELS = ("gpt-5", "codex")
+
 PLATFORM_HINTS = {
     "whatsapp": (
         "You are on a text messaging communication platform, WhatsApp. "
@@ -142,16 +270,21 @@ PLATFORM_HINTS = {
         "contextually appropriate."
     ),
     "cron": (
-        "You are running as a scheduled cron job. Your final response is automatically "
-        "delivered to the job's configured destination, so do not use send_message to "
-        "send to that same target again. If you want the user to receive something in "
-        "the scheduled destination, put it directly in your final response. Use "
-        "send_message only for additional or different targets."
+        "You are running as a scheduled cron job. There is no user present — you "
+        "cannot ask questions, request clarification, or wait for follow-up. Execute "
+        "the task fully and autonomously, making reasonable decisions where needed. "
+        "Your final response is automatically delivered to the job's configured "
+        "destination — put the primary content directly in your response."
     ),
     "cli": (
         "You are a CLI AI Agent. Try not to use markdown but simple text "
         "renderable inside a terminal."
     ),
+    "sms": (
+        "You are communicating via SMS. Keep responses concise and use plain text "
+        "only — no markdown, no formatting. SMS messages are limited to ~1600 "
+        "characters, so be brief and direct."
+    ),
 }
 
 CONTEXT_FILE_MAX_CHARS = 20_000
@@ -159,6 +292,111 @@ CONTEXT_TRUNCATE_HEAD_RATIO = 0.7
 CONTEXT_TRUNCATE_TAIL_RATIO = 0.2
 
 
+# =========================================================================
+# Skills prompt cache
+# =========================================================================
+
+_SKILLS_PROMPT_CACHE_MAX = 8
+_SKILLS_PROMPT_CACHE: OrderedDict[tuple, str] = OrderedDict()
+_SKILLS_PROMPT_CACHE_LOCK = threading.Lock()
+_SKILLS_SNAPSHOT_VERSION = 1
+
+
+def _skills_prompt_snapshot_path() -> Path:
+    return get_hermes_home() / ".skills_prompt_snapshot.json"
+
+
+def clear_skills_system_prompt_cache(*, clear_snapshot: bool = False) -> None:
+    """Drop the in-process skills prompt cache (and optionally the disk snapshot)."""
+    with _SKILLS_PROMPT_CACHE_LOCK:
+        _SKILLS_PROMPT_CACHE.clear()
+    if clear_snapshot:
+        try:
+            _skills_prompt_snapshot_path().unlink(missing_ok=True)
+        except OSError as e:
+            logger.debug("Could not remove skills prompt snapshot: %s", e)
+
+
+def _build_skills_manifest(skills_dir: Path) -> dict[str, list[int]]:
+    """Build an mtime/size manifest of all SKILL.md and DESCRIPTION.md files."""
+    manifest: dict[str, list[int]] = {}
+    for filename in ("SKILL.md", "DESCRIPTION.md"):
+        for path in iter_skill_index_files(skills_dir, filename):
+            try:
+                st = path.stat()
+            except OSError:
+                continue
+            manifest[str(path.relative_to(skills_dir))] = [st.st_mtime_ns, st.st_size]
+    return manifest
+
+
+def _load_skills_snapshot(skills_dir: Path) -> Optional[dict]:
+    """Load the disk snapshot if it exists and its manifest still matches."""
+    snapshot_path = _skills_prompt_snapshot_path()
+    if not snapshot_path.exists():
+        return None
+    try:
+        snapshot = json.loads(snapshot_path.read_text(encoding="utf-8"))
+    except Exception:
+        return None
+    if not isinstance(snapshot, dict):
+        return None
+    if snapshot.get("version") != _SKILLS_SNAPSHOT_VERSION:
+        return None
+    if snapshot.get("manifest") != _build_skills_manifest(skills_dir):
+        return None
+    return snapshot
+
+
+def _write_skills_snapshot(
+    skills_dir: Path,
+    manifest: dict[str, list[int]],
+    skill_entries: list[dict],
+    category_descriptions: dict[str, str],
+) -> None:
+    """Persist skill metadata to disk for fast cold-start reuse."""
+    payload = {
+        "version": _SKILLS_SNAPSHOT_VERSION,
+        "manifest": manifest,
+        "skills": skill_entries,
+        "category_descriptions": category_descriptions,
+    }
+    try:
+        atomic_json_write(_skills_prompt_snapshot_path(), payload)
+    except Exception as e:
+        logger.debug("Could not write skills prompt snapshot: %s", e)
+
+
+def _build_snapshot_entry(
+    skill_file: Path,
+    skills_dir: Path,
+    frontmatter: dict,
+    description: str,
+) -> dict:
+    """Build a serialisable metadata dict for one skill."""
+    rel_path = skill_file.relative_to(skills_dir)
+    parts = rel_path.parts
+    if len(parts) >= 2:
+        skill_name = parts[-2]
+        category = "/".join(parts[:-2]) if len(parts) > 2 else parts[0]
+    else:
+        category = "general"
+        skill_name = skill_file.parent.name
+
+    platforms = frontmatter.get("platforms") or []
+    if isinstance(platforms, str):
+        platforms = [platforms]
+
+    return {
+        "skill_name": skill_name,
+        "category": category,
+        "frontmatter_name": str(frontmatter.get("name", skill_name)),
+        "description": description,
+        "platforms": [str(p).strip() for p in platforms if str(p).strip()],
+        "conditions": extract_skill_conditions(frontmatter),
+    }
+
+
 # =========================================================================
 # Skills index
 # =========================================================================
@@ -170,22 +408,13 @@ def _parse_skill_file(skill_file: Path) -> tuple[bool, dict, str]:
     (True, {}, "") to err on the side of showing the skill.
     """
     try:
-        from tools.skills_tool import _parse_frontmatter, skill_matches_platform
-
         raw = skill_file.read_text(encoding="utf-8")[:2000]
-        frontmatter, _ = _parse_frontmatter(raw)
+        frontmatter, _ = parse_frontmatter(raw)
 
         if not skill_matches_platform(frontmatter):
-            return False, {}, ""
+            return False, frontmatter, ""
 
-        desc = ""
-        raw_desc = frontmatter.get("description", "")
-        if raw_desc:
-            desc = str(raw_desc).strip().strip("'\"")
-            if len(desc) > 60:
-                desc = desc[:57] + "..."
-
-        return True, frontmatter, desc
+        return True, frontmatter, extract_skill_description(frontmatter)
     except Exception as e:
         logger.debug("Failed to parse skill file %s: %s", skill_file, e)
         return True, {}, ""
@@ -194,16 +423,9 @@ def _parse_skill_file(skill_file: Path) -> tuple[bool, dict, str]:
 def _read_skill_conditions(skill_file: Path) -> dict:
     """Extract conditional activation fields from SKILL.md frontmatter."""
     try:
-        from tools.skills_tool import _parse_frontmatter
         raw = skill_file.read_text(encoding="utf-8")[:2000]
-        frontmatter, _ = _parse_frontmatter(raw)
-        hermes = frontmatter.get("metadata", {}).get("hermes", {})
-        return {
-            "fallback_for_toolsets": hermes.get("fallback_for_toolsets", []),
-            "requires_toolsets": hermes.get("requires_toolsets", []),
-            "fallback_for_tools": hermes.get("fallback_for_tools", []),
-            "requires_tools": hermes.get("requires_tools", []),
-        }
+        frontmatter, _ = parse_frontmatter(raw)
+        return extract_skill_conditions(frontmatter)
     except Exception as e:
         logger.debug("Failed to read skill conditions from %s: %s", skill_file, e)
         return {}
@@ -246,93 +468,285 @@ def build_skills_system_prompt(
 ) -> str:
     """Build a compact skill index for the system prompt.
 
-    Scans ~/.hermes/skills/ for SKILL.md files grouped by category.
-    Includes per-skill descriptions from frontmatter so the model can
-    match skills by meaning, not just name.
-    Filters out skills incompatible with the current OS platform.
+    Two-layer cache:
+      1. In-process LRU dict keyed by (skills_dir, tools, toolsets)
+      2. Disk snapshot (``.skills_prompt_snapshot.json``) validated by
+         mtime/size manifest — survives process restarts
+
+    Falls back to a full filesystem scan when both layers miss.
+
+    External skill directories (``skills.external_dirs`` in config.yaml) are
+    scanned alongside the local ``~/.hermes/skills/`` directory.  External dirs
+    are read-only — they appear in the index but new skills are always created
+    in the local dir.  Local skills take precedence when names collide.
     """
-    hermes_home = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
+    hermes_home = get_hermes_home()
     skills_dir = hermes_home / "skills"
+    external_dirs = get_all_skills_dirs()[1:]  # skip local (index 0)
 
-    if not skills_dir.exists():
+    if not skills_dir.exists() and not external_dirs:
         return ""
 
-    # Collect skills with descriptions, grouped by category.
-    # Each entry: (skill_name, description)
-    # Supports sub-categories: skills/mlops/training/axolotl/SKILL.md
-    # -> category "mlops/training", skill "axolotl"
+    # ── Layer 1: in-process LRU cache ─────────────────────────────────
+    # Include the resolved platform so per-platform disabled-skill lists
+    # produce distinct cache entries (gateway serves multiple platforms).
+    _platform_hint = (
+        os.environ.get("HERMES_PLATFORM")
+        or os.environ.get("HERMES_SESSION_PLATFORM")
+        or ""
+    )
+    cache_key = (
+        str(skills_dir.resolve()),
+        tuple(str(d) for d in external_dirs),
+        tuple(sorted(str(t) for t in (available_tools or set()))),
+        tuple(sorted(str(ts) for ts in (available_toolsets or set()))),
+        _platform_hint,
+    )
+    with _SKILLS_PROMPT_CACHE_LOCK:
+        cached = _SKILLS_PROMPT_CACHE.get(cache_key)
+        if cached is not None:
+            _SKILLS_PROMPT_CACHE.move_to_end(cache_key)
+            return cached
+
+    disabled = get_disabled_skill_names()
+
+    # ── Layer 2: disk snapshot ────────────────────────────────────────
+    snapshot = _load_skills_snapshot(skills_dir)
+
     skills_by_category: dict[str, list[tuple[str, str]]] = {}
-    for skill_file in skills_dir.rglob("SKILL.md"):
-        is_compatible, _, desc = _parse_skill_file(skill_file)
-        if not is_compatible:
-            continue
-        # Skip skills whose conditional activation rules exclude them
-        conditions = _read_skill_conditions(skill_file)
-        if not _skill_should_show(conditions, available_tools, available_toolsets):
-            continue
-        rel_path = skill_file.relative_to(skills_dir)
-        parts = rel_path.parts
-        if len(parts) >= 2:
-            # Category is everything between skills_dir and the skill folder
-            # e.g. parts = ("mlops", "training", "axolotl", "SKILL.md")
-            #   → category = "mlops/training", skill_name = "axolotl"
-            # e.g. parts = ("github", "github-auth", "SKILL.md")
-            #   → category = "github", skill_name = "github-auth"
-            skill_name = parts[-2]
-            category = "/".join(parts[:-2]) if len(parts) > 2 else parts[0]
-        else:
-            category = "general"
-            skill_name = skill_file.parent.name
-        skills_by_category.setdefault(category, []).append((skill_name, desc))
+    category_descriptions: dict[str, str] = {}
 
-    if not skills_by_category:
-        return ""
+    if snapshot is not None:
+        # Fast path: use pre-parsed metadata from disk
+        for entry in snapshot.get("skills", []):
+            if not isinstance(entry, dict):
+                continue
+            skill_name = entry.get("skill_name") or ""
+            category = entry.get("category") or "general"
+            frontmatter_name = entry.get("frontmatter_name") or skill_name
+            platforms = entry.get("platforms") or []
+            if not skill_matches_platform({"platforms": platforms}):
+                continue
+            if frontmatter_name in disabled or skill_name in disabled:
+                continue
+            if not _skill_should_show(
+                entry.get("conditions") or {},
+                available_tools,
+                available_toolsets,
+            ):
+                continue
+            skills_by_category.setdefault(category, []).append(
+                (skill_name, entry.get("description", ""))
+            )
+        category_descriptions = {
+            str(k): str(v)
+            for k, v in (snapshot.get("category_descriptions") or {}).items()
+        }
+    else:
+        # Cold path: full filesystem scan + write snapshot for next time
+        skill_entries: list[dict] = []
+        for skill_file in iter_skill_index_files(skills_dir, "SKILL.md"):
+            is_compatible, frontmatter, desc = _parse_skill_file(skill_file)
+            entry = _build_snapshot_entry(skill_file, skills_dir, frontmatter, desc)
+            skill_entries.append(entry)
+            if not is_compatible:
+                continue
+            skill_name = entry["skill_name"]
+            if entry["frontmatter_name"] in disabled or skill_name in disabled:
+                continue
+            if not _skill_should_show(
+                extract_skill_conditions(frontmatter),
+                available_tools,
+                available_toolsets,
+            ):
+                continue
+            skills_by_category.setdefault(entry["category"], []).append(
+                (skill_name, entry["description"])
+            )
 
-    # Read category-level descriptions from DESCRIPTION.md
-    # Checks both the exact category path and parent directories
-    category_descriptions = {}
-    for category in skills_by_category:
-        cat_path = Path(category)
-        desc_file = skills_dir / cat_path / "DESCRIPTION.md"
-        if desc_file.exists():
+        # Read category-level DESCRIPTION.md files
+        for desc_file in iter_skill_index_files(skills_dir, "DESCRIPTION.md"):
             try:
                 content = desc_file.read_text(encoding="utf-8")
-                match = re.search(r"^---\s*\n.*?description:\s*(.+?)\s*\n.*?^---", content, re.MULTILINE | re.DOTALL)
-                if match:
-                    category_descriptions[category] = match.group(1).strip()
+                fm, _ = parse_frontmatter(content)
+                cat_desc = fm.get("description")
+                if not cat_desc:
+                    continue
+                rel = desc_file.relative_to(skills_dir)
+                cat = "/".join(rel.parts[:-1]) if len(rel.parts) > 1 else "general"
+                category_descriptions[cat] = str(cat_desc).strip().strip("'\"")
             except Exception as e:
                 logger.debug("Could not read skill description %s: %s", desc_file, e)
 
-    index_lines = []
-    for category in sorted(skills_by_category.keys()):
-        cat_desc = category_descriptions.get(category, "")
-        if cat_desc:
-            index_lines.append(f"  {category}: {cat_desc}")
-        else:
-            index_lines.append(f"  {category}:")
-        # Deduplicate and sort skills within each category
-        seen = set()
-        for name, desc in sorted(skills_by_category[category], key=lambda x: x[0]):
-            if name in seen:
-                continue
-            seen.add(name)
-            if desc:
-                index_lines.append(f"    - {name}: {desc}")
-            else:
-                index_lines.append(f"    - {name}")
+        _write_skills_snapshot(
+            skills_dir,
+            _build_skills_manifest(skills_dir),
+            skill_entries,
+            category_descriptions,
+        )
 
-    return (
-        "## Skills (mandatory)\n"
-        "Before replying, scan the skills below. If one clearly matches your task, "
-        "load it with skill_view(name) and follow its instructions. "
-        "If a skill has issues, fix it with skill_manage(action='patch').\n"
-        "\n"
-        "<available_skills>\n"
-        + "\n".join(index_lines) + "\n"
-        "</available_skills>\n"
-        "\n"
-        "If none match, proceed normally without loading a skill."
+    # ── External skill directories ─────────────────────────────────────
+    # Scan external dirs directly (no snapshot caching — they're read-only
+    # and typically small).  Local skills already in skills_by_category take
+    # precedence: we track seen names and skip duplicates from external dirs.
+    seen_skill_names: set[str] = set()
+    for cat_skills in skills_by_category.values():
+        for name, _desc in cat_skills:
+            seen_skill_names.add(name)
+
+    for ext_dir in external_dirs:
+        if not ext_dir.exists():
+            continue
+        for skill_file in iter_skill_index_files(ext_dir, "SKILL.md"):
+            try:
+                is_compatible, frontmatter, desc = _parse_skill_file(skill_file)
+                if not is_compatible:
+                    continue
+                entry = _build_snapshot_entry(skill_file, ext_dir, frontmatter, desc)
+                skill_name = entry["skill_name"]
+                if skill_name in seen_skill_names:
+                    continue
+                if entry["frontmatter_name"] in disabled or skill_name in disabled:
+                    continue
+                if not _skill_should_show(
+                    extract_skill_conditions(frontmatter),
+                    available_tools,
+                    available_toolsets,
+                ):
+                    continue
+                seen_skill_names.add(skill_name)
+                skills_by_category.setdefault(entry["category"], []).append(
+                    (skill_name, entry["description"])
+                )
+            except Exception as e:
+                logger.debug("Error reading external skill %s: %s", skill_file, e)
+
+        # External category descriptions
+        for desc_file in iter_skill_index_files(ext_dir, "DESCRIPTION.md"):
+            try:
+                content = desc_file.read_text(encoding="utf-8")
+                fm, _ = parse_frontmatter(content)
+                cat_desc = fm.get("description")
+                if not cat_desc:
+                    continue
+                rel = desc_file.relative_to(ext_dir)
+                cat = "/".join(rel.parts[:-1]) if len(rel.parts) > 1 else "general"
+                category_descriptions.setdefault(cat, str(cat_desc).strip().strip("'\""))
+            except Exception as e:
+                logger.debug("Could not read external skill description %s: %s", desc_file, e)
+
+    if not skills_by_category:
+        result = ""
+    else:
+        index_lines = []
+        for category in sorted(skills_by_category.keys()):
+            cat_desc = category_descriptions.get(category, "")
+            if cat_desc:
+                index_lines.append(f"  {category}: {cat_desc}")
+            else:
+                index_lines.append(f"  {category}:")
+            # Deduplicate and sort skills within each category
+            seen = set()
+            for name, desc in sorted(skills_by_category[category], key=lambda x: x[0]):
+                if name in seen:
+                    continue
+                seen.add(name)
+                if desc:
+                    index_lines.append(f"    - {name}: {desc}")
+                else:
+                    index_lines.append(f"    - {name}")
+
+        result = (
+            "## Skills (mandatory)\n"
+            "Before replying, scan the skills below. If one clearly matches your task, "
+            "load it with skill_view(name) and follow its instructions. "
+            "If a skill has issues, fix it with skill_manage(action='patch').\n"
+            "After difficult/iterative tasks, offer to save as a skill. "
+            "If a skill you loaded was missing steps, had wrong commands, or needed "
+            "pitfalls you discovered, update it before finishing.\n"
+            "\n"
+            "<available_skills>\n"
+            + "\n".join(index_lines) + "\n"
+            "</available_skills>\n"
+            "\n"
+            "If none match, proceed normally without loading a skill."
+        )
+
+    # ── Store in LRU cache ────────────────────────────────────────────
+    with _SKILLS_PROMPT_CACHE_LOCK:
+        _SKILLS_PROMPT_CACHE[cache_key] = result
+        _SKILLS_PROMPT_CACHE.move_to_end(cache_key)
+        while len(_SKILLS_PROMPT_CACHE) > _SKILLS_PROMPT_CACHE_MAX:
+            _SKILLS_PROMPT_CACHE.popitem(last=False)
+
+    return result
+
+
+def build_nous_subscription_prompt(valid_tool_names: "set[str] | None" = None) -> str:
+    """Build a compact Nous subscription capability block for the system prompt."""
+    try:
+        from hermes_cli.nous_subscription import get_nous_subscription_features
+        from tools.tool_backend_helpers import managed_nous_tools_enabled
+    except Exception as exc:
+        logger.debug("Failed to import Nous subscription helper: %s", exc)
+        return ""
+
+    if not managed_nous_tools_enabled():
+        return ""
+
+    valid_names = set(valid_tool_names or set())
+    relevant_tool_names = {
+        "web_search",
+        "web_extract",
+        "browser_navigate",
+        "browser_snapshot",
+        "browser_click",
+        "browser_type",
+        "browser_scroll",
+        "browser_console",
+        "browser_close",
+        "browser_press",
+        "browser_get_images",
+        "browser_vision",
+        "image_generate",
+        "text_to_speech",
+        "terminal",
+        "process",
+        "execute_code",
+    }
+
+    if valid_names and not (valid_names & relevant_tool_names):
+        return ""
+
+    features = get_nous_subscription_features()
+
+    def _status_line(feature) -> str:
+        if feature.managed_by_nous:
+            return f"- {feature.label}: active via Nous subscription"
+        if feature.active:
+            current = feature.current_provider or "configured provider"
+            return f"- {feature.label}: currently using {current}"
+        if feature.included_by_default and features.nous_auth_present:
+            return f"- {feature.label}: included with Nous subscription, not currently selected"
+        if feature.key == "modal" and features.nous_auth_present:
+            return f"- {feature.label}: optional via Nous subscription"
+        return f"- {feature.label}: not currently available"
+
+    lines = [
+        "# Nous Subscription",
+        "Nous subscription includes managed web tools (Firecrawl), image generation (FAL), OpenAI TTS, and browser automation (Browserbase) by default. Modal execution is optional.",
+        "Current capability status:",
+    ]
+    lines.extend(_status_line(feature) for feature in features.items())
+    lines.extend(
+        [
+            "When a Nous-managed feature is active, do not ask the user for Firecrawl, FAL, OpenAI TTS, or Browserbase API keys.",
+            "If the user is not subscribed and asks for a capability that Nous subscription would unlock or simplify, suggest Nous subscription as one option alongside direct setup or local alternatives.",
+            "Do not mention subscription unless the user asks about it or it directly solves the current missing capability.",
+            "Useful commands: hermes setup, hermes setup tools, hermes setup terminal, hermes status.",
+        ]
     )
+    return "\n".join(lines)
 
 
 # =========================================================================
@@ -351,51 +765,91 @@ def _truncate_content(content: str, filename: str, max_chars: int = CONTEXT_FILE
     return head + marker + tail
 
 
-def build_context_files_prompt(cwd: Optional[str] = None) -> str:
-    """Discover and load context files for the system prompt.
+def load_soul_md() -> Optional[str]:
+    """Load SOUL.md from HERMES_HOME and return its content, or None.
 
-    Discovery: AGENTS.md (recursive), .cursorrules / .cursor/rules/*.mdc,
-    and SOUL.md from HERMES_HOME only. Each capped at 20,000 chars.
+    Used as the agent identity (slot #1 in the system prompt).  When this
+    returns content, ``build_context_files_prompt`` should be called with
+    ``skip_soul=True`` so SOUL.md isn't injected twice.
     """
-    if cwd is None:
-        cwd = os.getcwd()
+    try:
+        from hermes_cli.config import ensure_hermes_home
+        ensure_hermes_home()
+    except Exception as e:
+        logger.debug("Could not ensure HERMES_HOME before loading SOUL.md: %s", e)
 
-    cwd_path = Path(cwd).resolve()
-    sections = []
+    soul_path = get_hermes_home() / "SOUL.md"
+    if not soul_path.exists():
+        return None
+    try:
+        content = soul_path.read_text(encoding="utf-8").strip()
+        if not content:
+            return None
+        content = _scan_context_content(content, "SOUL.md")
+        content = _truncate_content(content, "SOUL.md")
+        return content
+    except Exception as e:
+        logger.debug("Could not read SOUL.md from %s: %s", soul_path, e)
+        return None
 
-    # AGENTS.md (hierarchical, recursive)
-    top_level_agents = None
+
+def _load_hermes_md(cwd_path: Path) -> str:
+    """.hermes.md / HERMES.md — walk to git root."""
+    hermes_md_path = _find_hermes_md(cwd_path)
+    if not hermes_md_path:
+        return ""
+    try:
+        content = hermes_md_path.read_text(encoding="utf-8").strip()
+        if not content:
+            return ""
+        content = _strip_yaml_frontmatter(content)
+        rel = hermes_md_path.name
+        try:
+            rel = str(hermes_md_path.relative_to(cwd_path))
+        except ValueError:
+            pass
+        content = _scan_context_content(content, rel)
+        result = f"## {rel}\n\n{content}"
+        return _truncate_content(result, ".hermes.md")
+    except Exception as e:
+        logger.debug("Could not read %s: %s", hermes_md_path, e)
+        return ""
+
+
+def _load_agents_md(cwd_path: Path) -> str:
+    """AGENTS.md — top-level only (no recursive walk)."""
     for name in ["AGENTS.md", "agents.md"]:
         candidate = cwd_path / name
         if candidate.exists():
-            top_level_agents = candidate
-            break
-
-    if top_level_agents:
-        agents_files = []
-        for root, dirs, files in os.walk(cwd_path):
-            dirs[:] = [d for d in dirs if not d.startswith('.') and d not in ('node_modules', '__pycache__', 'venv', '.venv')]
-            for f in files:
-                if f.lower() == "agents.md":
-                    agents_files.append(Path(root) / f)
-        agents_files.sort(key=lambda p: len(p.parts))
-
-        total_agents_content = ""
-        for agents_path in agents_files:
             try:
-                content = agents_path.read_text(encoding="utf-8").strip()
+                content = candidate.read_text(encoding="utf-8").strip()
                 if content:
-                    rel_path = agents_path.relative_to(cwd_path)
-                    content = _scan_context_content(content, str(rel_path))
-                    total_agents_content += f"## {rel_path}\n\n{content}\n\n"
+                    content = _scan_context_content(content, name)
+                    result = f"## {name}\n\n{content}"
+                    return _truncate_content(result, "AGENTS.md")
             except Exception as e:
-                logger.debug("Could not read %s: %s", agents_path, e)
+                logger.debug("Could not read %s: %s", candidate, e)
+    return ""
 
-        if total_agents_content:
-            total_agents_content = _truncate_content(total_agents_content, "AGENTS.md")
-            sections.append(total_agents_content)
 
-    # .cursorrules
+def _load_claude_md(cwd_path: Path) -> str:
+    """CLAUDE.md / claude.md — cwd only."""
+    for name in ["CLAUDE.md", "claude.md"]:
+        candidate = cwd_path / name
+        if candidate.exists():
+            try:
+                content = candidate.read_text(encoding="utf-8").strip()
+                if content:
+                    content = _scan_context_content(content, name)
+                    result = f"## {name}\n\n{content}"
+                    return _truncate_content(result, "CLAUDE.md")
+            except Exception as e:
+                logger.debug("Could not read %s: %s", candidate, e)
+    return ""
+
+
+def _load_cursorrules(cwd_path: Path) -> str:
+    """.cursorrules + .cursor/rules/*.mdc — cwd only."""
     cursorrules_content = ""
     cursorrules_file = cwd_path / ".cursorrules"
     if cursorrules_file.exists():
@@ -419,27 +873,47 @@ def build_context_files_prompt(cwd: Optional[str] = None) -> str:
             except Exception as e:
                 logger.debug("Could not read %s: %s", mdc_file, e)
 
-    if cursorrules_content:
-        cursorrules_content = _truncate_content(cursorrules_content, ".cursorrules")
-        sections.append(cursorrules_content)
+    if not cursorrules_content:
+        return ""
+    return _truncate_content(cursorrules_content, ".cursorrules")
 
-    # SOUL.md from HERMES_HOME only
-    try:
-        from hermes_cli.config import ensure_hermes_home
-        ensure_hermes_home()
-    except Exception as e:
-        logger.debug("Could not ensure HERMES_HOME before loading SOUL.md: %s", e)
 
-    soul_path = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes")) / "SOUL.md"
-    if soul_path.exists():
-        try:
-            content = soul_path.read_text(encoding="utf-8").strip()
-            if content:
-                content = _scan_context_content(content, "SOUL.md")
-                content = _truncate_content(content, "SOUL.md")
-                sections.append(content)
-        except Exception as e:
-            logger.debug("Could not read SOUL.md from %s: %s", soul_path, e)
+def build_context_files_prompt(cwd: Optional[str] = None, skip_soul: bool = False) -> str:
+    """Discover and load context files for the system prompt.
+
+    Priority (first found wins — only ONE project context type is loaded):
+      1. .hermes.md / HERMES.md  (walk to git root)
+      2. AGENTS.md / agents.md   (cwd only)
+      3. CLAUDE.md / claude.md   (cwd only)
+      4. .cursorrules / .cursor/rules/*.mdc  (cwd only)
+
+    SOUL.md from HERMES_HOME is independent and always included when present.
+    Each context source is capped at 20,000 chars.
+
+    When *skip_soul* is True, SOUL.md is not included here (it was already
+    loaded via ``load_soul_md()`` for the identity slot).
+    """
+    if cwd is None:
+        cwd = os.getcwd()
+
+    cwd_path = Path(cwd).resolve()
+    sections = []
+
+    # Priority-based project context: first match wins
+    project_context = (
+        _load_hermes_md(cwd_path)
+        or _load_agents_md(cwd_path)
+        or _load_claude_md(cwd_path)
+        or _load_cursorrules(cwd_path)
+    )
+    if project_context:
+        sections.append(project_context)
+
+    # SOUL.md from HERMES_HOME only — skip when already loaded as identity
+    if not skip_soul:
+        soul_content = load_soul_md()
+        if soul_content:
+            sections.append(soul_content)
 
     if not sections:
         return ""

agent/prompt_caching.py

@@ -12,13 +12,14 @@ import copy
 from typing import Any, Dict, List
 
 
-def _apply_cache_marker(msg: dict, cache_marker: dict) -> None:
+def _apply_cache_marker(msg: dict, cache_marker: dict, native_anthropic: bool = False) -> None:
     """Add cache_control to a single message, handling all format variations."""
     role = msg.get("role", "")
     content = msg.get("content")
 
     if role == "tool":
-        msg["cache_control"] = cache_marker
+        if native_anthropic:
+            msg["cache_control"] = cache_marker
         return
 
     if content is None or content == "":
@@ -40,6 +41,7 @@ def _apply_cache_marker(msg: dict, cache_marker: dict) -> None:
 def apply_anthropic_cache_control(
     api_messages: List[Dict[str, Any]],
     cache_ttl: str = "5m",
+    native_anthropic: bool = False,
 ) -> List[Dict[str, Any]]:
     """Apply system_and_3 caching strategy to messages for Anthropic models.
 
@@ -59,12 +61,12 @@ def apply_anthropic_cache_control(
     breakpoints_used = 0
 
     if messages[0].get("role") == "system":
-        _apply_cache_marker(messages[0], marker)
+        _apply_cache_marker(messages[0], marker, native_anthropic=native_anthropic)
         breakpoints_used += 1
 
     remaining = 4 - breakpoints_used
     non_sys = [i for i in range(len(messages)) if messages[i].get("role") != "system"]
     for idx in non_sys[-remaining:]:
-        _apply_cache_marker(messages[idx], marker)
+        _apply_cache_marker(messages[idx], marker, native_anthropic=native_anthropic)
 
     return messages

agent/redact.py

@@ -13,11 +13,19 @@ import re
 
 logger = logging.getLogger(__name__)
 
+# Snapshot at import time so runtime env mutations (e.g. LLM-generated
+# `export HERMES_REDACT_SECRETS=false`) cannot disable redaction mid-session.
+_REDACT_ENABLED = os.getenv("HERMES_REDACT_SECRETS", "").lower() not in ("0", "false", "no", "off")
+
 # Known API key prefixes -- match the prefix + contiguous token chars
 _PREFIX_PATTERNS = [
     r"sk-[A-Za-z0-9_-]{10,}",           # OpenAI / OpenRouter / Anthropic (sk-ant-*)
     r"ghp_[A-Za-z0-9]{10,}",            # GitHub PAT (classic)
     r"github_pat_[A-Za-z0-9_]{10,}",    # GitHub PAT (fine-grained)
+    r"gho_[A-Za-z0-9]{10,}",            # GitHub OAuth access token
+    r"ghu_[A-Za-z0-9]{10,}",            # GitHub user-to-server token
+    r"ghs_[A-Za-z0-9]{10,}",            # GitHub server-to-server token
+    r"ghr_[A-Za-z0-9]{10,}",            # GitHub refresh token
     r"xox[baprs]-[A-Za-z0-9-]{10,}",    # Slack tokens
     r"AIza[A-Za-z0-9_-]{30,}",          # Google API keys
     r"pplx-[A-Za-z0-9]{10,}",           # Perplexity
@@ -37,13 +45,15 @@ _PREFIX_PATTERNS = [
     r"dop_v1_[A-Za-z0-9]{10,}",         # DigitalOcean PAT
     r"doo_v1_[A-Za-z0-9]{10,}",         # DigitalOcean OAuth
     r"am_[A-Za-z0-9_-]{10,}",           # AgentMail API key
+    r"sk_[A-Za-z0-9_]{10,}",            # ElevenLabs TTS key (sk_ underscore, not sk- dash)
+    r"tvly-[A-Za-z0-9]{10,}",           # Tavily search API key
+    r"exa_[A-Za-z0-9]{10,}",            # Exa search API key
 ]
 
 # ENV assignment patterns: KEY=value where KEY contains a secret-like name
 _SECRET_ENV_NAMES = r"(?:API_?KEY|TOKEN|SECRET|PASSWORD|PASSWD|CREDENTIAL|AUTH)"
 _ENV_ASSIGN_RE = re.compile(
-    rf"([A-Z_]*{_SECRET_ENV_NAMES}[A-Z_]*)\s*=\s*(['\"]?)(\S+)\2",
-    re.IGNORECASE,
+    rf"([A-Z0-9_]{{0,50}}{_SECRET_ENV_NAMES}[A-Z0-9_]{{0,50}})\s*=\s*(['\"]?)(\S+)\2",
 )
 
 # JSON field patterns: "apiKey": "value", "token": "value", etc.
@@ -100,9 +110,13 @@ def redact_sensitive_text(text: str) -> str:
     Safe to call on any string -- non-matching text passes through unchanged.
     Disabled when security.redact_secrets is false in config.yaml.
     """
+    if text is None:
+        return None
+    if not isinstance(text, str):
+        text = str(text)
     if not text:
         return text
-    if os.getenv("HERMES_REDACT_SECRETS", "").lower() in ("0", "false", "no", "off"):
+    if not _REDACT_ENABLED:
         return text
 
     # Known prefixes (sk-, ghp_, etc.)

agent/skill_commands.py

@@ -1,17 +1,42 @@
-"""Skill slash commands — scan installed skills and build invocation messages.
+"""Shared slash command helpers for skills and built-in prompt-style modes.
 
 Shared between CLI (cli.py) and gateway (gateway/run.py) so both surfaces
-can invoke skills via /skill-name commands.
+can invoke skills via /skill-name commands and prompt-only built-ins like
+/plan.
 """
 
 import json
 import logging
+import re
+from datetime import datetime
 from pathlib import Path
 from typing import Any, Dict, Optional
 
 logger = logging.getLogger(__name__)
 
 _skill_commands: Dict[str, Dict[str, Any]] = {}
+_PLAN_SLUG_RE = re.compile(r"[^a-z0-9]+")
+
+
+def build_plan_path(
+    user_instruction: str = "",
+    *,
+    now: datetime | None = None,
+) -> Path:
+    """Return the default workspace-relative markdown path for a /plan invocation.
+
+    Relative paths are intentional: file tools are task/backend-aware and resolve
+    them against the active working directory for local, docker, ssh, modal,
+    daytona, and similar terminal backends. That keeps the plan with the active
+    workspace instead of the Hermes host's global home directory.
+    """
+    slug_source = (user_instruction or "").strip().splitlines()[0] if user_instruction else ""
+    slug = _PLAN_SLUG_RE.sub("-", slug_source.lower()).strip("-")
+    if slug:
+        slug = "-".join(part for part in slug.split("-")[:8] if part)[:48].strip("-")
+    slug = slug or "conversation-plan"
+    timestamp = (now or datetime.now()).strftime("%Y-%m-%d_%H%M%S")
+    return Path(".hermes") / "plans" / f"{timestamp}-{slug}.md"
 
 
 def _load_skill_payload(skill_identifier: str, task_id: str | None = None) -> tuple[dict[str, Any], Path | None, str] | None:
@@ -56,6 +81,7 @@ def _build_skill_message(
     skill_dir: Path | None,
     activation_note: str,
     user_instruction: str = "",
+    runtime_note: str = "",
 ) -> str:
     """Format a loaded skill into a user/system message payload."""
     from tools.skills_tool import SKILLS_DIR
@@ -102,7 +128,11 @@ def _build_skill_message(
                         supporting.append(rel)
 
     if supporting and skill_dir:
-        skill_view_target = str(skill_dir.relative_to(SKILLS_DIR))
+        try:
+            skill_view_target = str(skill_dir.relative_to(SKILLS_DIR))
+        except ValueError:
+            # Skill is from an external dir — use the skill name instead
+            skill_view_target = skill_dir.name
         parts.append("")
         parts.append("[This skill has supporting files you can load with the skill_view tool:]")
         for sf in supporting:
@@ -115,6 +145,10 @@ def _build_skill_message(
         parts.append("")
         parts.append(f"The user has provided the following instruction alongside the skill invocation: {user_instruction}")
 
+    if runtime_note:
+        parts.append("")
+        parts.append(f"[Runtime note: {runtime_note}]")
+
     return "\n".join(parts)
 
 
@@ -127,35 +161,50 @@ def scan_skill_commands() -> Dict[str, Dict[str, Any]]:
     global _skill_commands
     _skill_commands = {}
     try:
-        from tools.skills_tool import SKILLS_DIR, _parse_frontmatter, skill_matches_platform
-        if not SKILLS_DIR.exists():
-            return _skill_commands
-        for skill_md in SKILLS_DIR.rglob("SKILL.md"):
-            if any(part in ('.git', '.github', '.hub') for part in skill_md.parts):
-                continue
-            try:
-                content = skill_md.read_text(encoding='utf-8')
-                frontmatter, body = _parse_frontmatter(content)
-                # Skip skills incompatible with the current OS platform
-                if not skill_matches_platform(frontmatter):
+        from tools.skills_tool import SKILLS_DIR, _parse_frontmatter, skill_matches_platform, _get_disabled_skill_names
+        from agent.skill_utils import get_external_skills_dirs
+        disabled = _get_disabled_skill_names()
+        seen_names: set = set()
+
+        # Scan local dir first, then external dirs
+        dirs_to_scan = []
+        if SKILLS_DIR.exists():
+            dirs_to_scan.append(SKILLS_DIR)
+        dirs_to_scan.extend(get_external_skills_dirs())
+
+        for scan_dir in dirs_to_scan:
+            for skill_md in scan_dir.rglob("SKILL.md"):
+                if any(part in ('.git', '.github', '.hub') for part in skill_md.parts):
+                    continue
+                try:
+                    content = skill_md.read_text(encoding='utf-8')
+                    frontmatter, body = _parse_frontmatter(content)
+                    # Skip skills incompatible with the current OS platform
+                    if not skill_matches_platform(frontmatter):
+                        continue
+                    name = frontmatter.get('name', skill_md.parent.name)
+                    if name in seen_names:
+                        continue
+                    # Respect user's disabled skills config
+                    if name in disabled:
+                        continue
+                    description = frontmatter.get('description', '')
+                    if not description:
+                        for line in body.strip().split('\n'):
+                            line = line.strip()
+                            if line and not line.startswith('#'):
+                                description = line[:80]
+                                break
+                    seen_names.add(name)
+                    cmd_name = name.lower().replace(' ', '-').replace('_', '-')
+                    _skill_commands[f"/{cmd_name}"] = {
+                        "name": name,
+                        "description": description or f"Invoke the {name} skill",
+                        "skill_md_path": str(skill_md),
+                        "skill_dir": str(skill_md.parent),
+                    }
+                except Exception:
                     continue
-                name = frontmatter.get('name', skill_md.parent.name)
-                description = frontmatter.get('description', '')
-                if not description:
-                    for line in body.strip().split('\n'):
-                        line = line.strip()
-                        if line and not line.startswith('#'):
-                            description = line[:80]
-                            break
-                cmd_name = name.lower().replace(' ', '-').replace('_', '-')
-                _skill_commands[f"/{cmd_name}"] = {
-                    "name": name,
-                    "description": description or f"Invoke the {name} skill",
-                    "skill_md_path": str(skill_md),
-                    "skill_dir": str(skill_md.parent),
-                }
-            except Exception:
-                continue
     except Exception:
         pass
     return _skill_commands
@@ -168,10 +217,30 @@ def get_skill_commands() -> Dict[str, Dict[str, Any]]:
     return _skill_commands
 
 
+def resolve_skill_command_key(command: str) -> Optional[str]:
+    """Resolve a user-typed /command to its canonical skill_cmds key.
+
+    Skills are always stored with hyphens — ``scan_skill_commands`` normalizes
+    spaces and underscores to hyphens when building the key. Hyphens and
+    underscores are treated interchangeably in user input: this matches
+    ``_check_unavailable_skill`` and accommodates Telegram bot-command names
+    (which disallow hyphens, so ``/claude-code`` is registered as
+    ``/claude_code`` and comes back in the underscored form).
+
+    Returns the matching ``/slug`` key from ``get_skill_commands()`` or
+    ``None`` if no match.
+    """
+    if not command:
+        return None
+    cmd_key = f"/{command.replace('_', '-')}"
+    return cmd_key if cmd_key in get_skill_commands() else None
+
+
 def build_skill_invocation_message(
     cmd_key: str,
     user_instruction: str = "",
     task_id: str | None = None,
+    runtime_note: str = "",
 ) -> Optional[str]:
     """Build the user message content for a skill slash command invocation.
 
@@ -201,6 +270,7 @@ def build_skill_invocation_message(
         skill_dir,
         activation_note,
         user_instruction=user_instruction,
+        runtime_note=runtime_note,
     )
 
 

agent/skill_utils.py

@@ -0,0 +1,285 @@
+"""Lightweight skill metadata utilities shared by prompt_builder and skills_tool.
+
+This module intentionally avoids importing the tool registry, CLI config, or any
+heavy dependency chain.  It is safe to import at module level without triggering
+tool registration or provider resolution.
+"""
+
+import logging
+import os
+import re
+import sys
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Set, Tuple
+
+from hermes_constants import get_hermes_home
+
+logger = logging.getLogger(__name__)
+
+# ── Platform mapping ──────────────────────────────────────────────────────
+
+PLATFORM_MAP = {
+    "macos": "darwin",
+    "linux": "linux",
+    "windows": "win32",
+}
+
+EXCLUDED_SKILL_DIRS = frozenset((".git", ".github", ".hub"))
+
+# ── Lazy YAML loader ─────────────────────────────────────────────────────
+
+_yaml_load_fn = None
+
+
+def yaml_load(content: str):
+    """Parse YAML with lazy import and CSafeLoader preference."""
+    global _yaml_load_fn
+    if _yaml_load_fn is None:
+        import yaml
+
+        loader = getattr(yaml, "CSafeLoader", None) or yaml.SafeLoader
+
+        def _load(value: str):
+            return yaml.load(value, Loader=loader)
+
+        _yaml_load_fn = _load
+    return _yaml_load_fn(content)
+
+
+# ── Frontmatter parsing ──────────────────────────────────────────────────
+
+
+def parse_frontmatter(content: str) -> Tuple[Dict[str, Any], str]:
+    """Parse YAML frontmatter from a markdown string.
+
+    Uses yaml with CSafeLoader for full YAML support (nested metadata, lists)
+    with a fallback to simple key:value splitting for robustness.
+
+    Returns:
+        (frontmatter_dict, remaining_body)
+    """
+    frontmatter: Dict[str, Any] = {}
+    body = content
+
+    if not content.startswith("---"):
+        return frontmatter, body
+
+    end_match = re.search(r"\n---\s*\n", content[3:])
+    if not end_match:
+        return frontmatter, body
+
+    yaml_content = content[3 : end_match.start() + 3]
+    body = content[end_match.end() + 3 :]
+
+    try:
+        parsed = yaml_load(yaml_content)
+        if isinstance(parsed, dict):
+            frontmatter = parsed
+    except Exception:
+        # Fallback: simple key:value parsing for malformed YAML
+        for line in yaml_content.strip().split("\n"):
+            if ":" not in line:
+                continue
+            key, value = line.split(":", 1)
+            frontmatter[key.strip()] = value.strip()
+
+    return frontmatter, body
+
+
+# ── Platform matching ─────────────────────────────────────────────────────
+
+
+def skill_matches_platform(frontmatter: Dict[str, Any]) -> bool:
+    """Return True when the skill is compatible with the current OS.
+
+    Skills declare platform requirements via a top-level ``platforms`` list
+    in their YAML frontmatter::
+
+        platforms: [macos]          # macOS only
+        platforms: [macos, linux]   # macOS and Linux
+
+    If the field is absent or empty the skill is compatible with **all**
+    platforms (backward-compatible default).
+    """
+    platforms = frontmatter.get("platforms")
+    if not platforms:
+        return True
+    if not isinstance(platforms, list):
+        platforms = [platforms]
+    current = sys.platform
+    for platform in platforms:
+        normalized = str(platform).lower().strip()
+        mapped = PLATFORM_MAP.get(normalized, normalized)
+        if current.startswith(mapped):
+            return True
+    return False
+
+
+# ── Disabled skills ───────────────────────────────────────────────────────
+
+
+def get_disabled_skill_names(platform: str | None = None) -> Set[str]:
+    """Read disabled skill names from config.yaml.
+
+    Args:
+        platform: Explicit platform name (e.g. ``"telegram"``).  When
+            *None*, resolves from ``HERMES_PLATFORM`` or
+            ``HERMES_SESSION_PLATFORM`` env vars.  Falls back to the
+            global disabled list when no platform is determined.
+
+    Reads the config file directly (no CLI config imports) to stay
+    lightweight.
+    """
+    config_path = get_hermes_home() / "config.yaml"
+    if not config_path.exists():
+        return set()
+    try:
+        parsed = yaml_load(config_path.read_text(encoding="utf-8"))
+    except Exception as e:
+        logger.debug("Could not read skill config %s: %s", config_path, e)
+        return set()
+    if not isinstance(parsed, dict):
+        return set()
+
+    skills_cfg = parsed.get("skills")
+    if not isinstance(skills_cfg, dict):
+        return set()
+
+    resolved_platform = (
+        platform
+        or os.getenv("HERMES_PLATFORM")
+        or os.getenv("HERMES_SESSION_PLATFORM")
+    )
+    if resolved_platform:
+        platform_disabled = (skills_cfg.get("platform_disabled") or {}).get(
+            resolved_platform
+        )
+        if platform_disabled is not None:
+            return _normalize_string_set(platform_disabled)
+    return _normalize_string_set(skills_cfg.get("disabled"))
+
+
+def _normalize_string_set(values) -> Set[str]:
+    if values is None:
+        return set()
+    if isinstance(values, str):
+        values = [values]
+    return {str(v).strip() for v in values if str(v).strip()}
+
+
+# ── External skills directories ──────────────────────────────────────────
+
+
+def get_external_skills_dirs() -> List[Path]:
+    """Read ``skills.external_dirs`` from config.yaml and return validated paths.
+
+    Each entry is expanded (``~`` and ``${VAR}``) and resolved to an absolute
+    path.  Only directories that actually exist are returned.  Duplicates and
+    paths that resolve to the local ``~/.hermes/skills/`` are silently skipped.
+    """
+    config_path = get_hermes_home() / "config.yaml"
+    if not config_path.exists():
+        return []
+    try:
+        parsed = yaml_load(config_path.read_text(encoding="utf-8"))
+    except Exception:
+        return []
+    if not isinstance(parsed, dict):
+        return []
+
+    skills_cfg = parsed.get("skills")
+    if not isinstance(skills_cfg, dict):
+        return []
+
+    raw_dirs = skills_cfg.get("external_dirs")
+    if not raw_dirs:
+        return []
+    if isinstance(raw_dirs, str):
+        raw_dirs = [raw_dirs]
+    if not isinstance(raw_dirs, list):
+        return []
+
+    local_skills = (get_hermes_home() / "skills").resolve()
+    seen: Set[Path] = set()
+    result: List[Path] = []
+
+    for entry in raw_dirs:
+        entry = str(entry).strip()
+        if not entry:
+            continue
+        # Expand ~ and environment variables
+        expanded = os.path.expanduser(os.path.expandvars(entry))
+        p = Path(expanded).resolve()
+        if p == local_skills:
+            continue
+        if p in seen:
+            continue
+        if p.is_dir():
+            seen.add(p)
+            result.append(p)
+        else:
+            logger.debug("External skills dir does not exist, skipping: %s", p)
+
+    return result
+
+
+def get_all_skills_dirs() -> List[Path]:
+    """Return all skill directories: local ``~/.hermes/skills/`` first, then external.
+
+    The local dir is always first (and always included even if it doesn't exist
+    yet — callers handle that).  External dirs follow in config order.
+    """
+    dirs = [get_hermes_home() / "skills"]
+    dirs.extend(get_external_skills_dirs())
+    return dirs
+
+
+# ── Condition extraction ──────────────────────────────────────────────────
+
+
+def extract_skill_conditions(frontmatter: Dict[str, Any]) -> Dict[str, List]:
+    """Extract conditional activation fields from parsed frontmatter."""
+    metadata = frontmatter.get("metadata")
+    # Handle cases where metadata is not a dict (e.g., a string from malformed YAML)
+    if not isinstance(metadata, dict):
+        metadata = {}
+    hermes = metadata.get("hermes") or {}
+    if not isinstance(hermes, dict):
+        hermes = {}
+    return {
+        "fallback_for_toolsets": hermes.get("fallback_for_toolsets", []),
+        "requires_toolsets": hermes.get("requires_toolsets", []),
+        "fallback_for_tools": hermes.get("fallback_for_tools", []),
+        "requires_tools": hermes.get("requires_tools", []),
+    }
+
+
+# ── Description extraction ────────────────────────────────────────────────
+
+
+def extract_skill_description(frontmatter: Dict[str, Any]) -> str:
+    """Extract a truncated description from parsed frontmatter."""
+    raw_desc = frontmatter.get("description", "")
+    if not raw_desc:
+        return ""
+    desc = str(raw_desc).strip().strip("'\"")
+    if len(desc) > 60:
+        return desc[:57] + "..."
+    return desc
+
+
+# ── File iteration ────────────────────────────────────────────────────────
+
+
+def iter_skill_index_files(skills_dir: Path, filename: str):
+    """Walk skills_dir yielding sorted paths matching *filename*.
+
+    Excludes ``.git``, ``.github``, ``.hub`` directories.
+    """
+    matches = []
+    for root, dirs, files in os.walk(skills_dir):
+        dirs[:] = [d for d in dirs if d not in EXCLUDED_SKILL_DIRS]
+        if filename in files:
+            matches.append(Path(root) / filename)
+    for path in sorted(matches, key=lambda p: str(p.relative_to(skills_dir))):
+        yield path

agent/smart_model_routing.py

@@ -0,0 +1,194 @@
+"""Helpers for optional cheap-vs-strong model routing."""
+
+from __future__ import annotations
+
+import os
+import re
+from typing import Any, Dict, Optional
+
+from utils import is_truthy_value
+
+_COMPLEX_KEYWORDS = {
+    "debug",
+    "debugging",
+    "implement",
+    "implementation",
+    "refactor",
+    "patch",
+    "traceback",
+    "stacktrace",
+    "exception",
+    "error",
+    "analyze",
+    "analysis",
+    "investigate",
+    "architecture",
+    "design",
+    "compare",
+    "benchmark",
+    "optimize",
+    "optimise",
+    "review",
+    "terminal",
+    "shell",
+    "tool",
+    "tools",
+    "pytest",
+    "test",
+    "tests",
+    "plan",
+    "planning",
+    "delegate",
+    "subagent",
+    "cron",
+    "docker",
+    "kubernetes",
+}
+
+_URL_RE = re.compile(r"https?://|www\.", re.IGNORECASE)
+
+
+def _coerce_bool(value: Any, default: bool = False) -> bool:
+    return is_truthy_value(value, default=default)
+
+
+def _coerce_int(value: Any, default: int) -> int:
+    try:
+        return int(value)
+    except (TypeError, ValueError):
+        return default
+
+
+def choose_cheap_model_route(user_message: str, routing_config: Optional[Dict[str, Any]]) -> Optional[Dict[str, Any]]:
+    """Return the configured cheap-model route when a message looks simple.
+
+    Conservative by design: if the message has signs of code/tool/debugging/
+    long-form work, keep the primary model.
+    """
+    cfg = routing_config or {}
+    if not _coerce_bool(cfg.get("enabled"), False):
+        return None
+
+    cheap_model = cfg.get("cheap_model") or {}
+    if not isinstance(cheap_model, dict):
+        return None
+    provider = str(cheap_model.get("provider") or "").strip().lower()
+    model = str(cheap_model.get("model") or "").strip()
+    if not provider or not model:
+        return None
+
+    text = (user_message or "").strip()
+    if not text:
+        return None
+
+    max_chars = _coerce_int(cfg.get("max_simple_chars"), 160)
+    max_words = _coerce_int(cfg.get("max_simple_words"), 28)
+
+    if len(text) > max_chars:
+        return None
+    if len(text.split()) > max_words:
+        return None
+    if text.count("\n") > 1:
+        return None
+    if "```" in text or "`" in text:
+        return None
+    if _URL_RE.search(text):
+        return None
+
+    lowered = text.lower()
+    words = {token.strip(".,:;!?()[]{}\"'`") for token in lowered.split()}
+    if words & _COMPLEX_KEYWORDS:
+        return None
+
+    route = dict(cheap_model)
+    route["provider"] = provider
+    route["model"] = model
+    route["routing_reason"] = "simple_turn"
+    return route
+
+
+def resolve_turn_route(user_message: str, routing_config: Optional[Dict[str, Any]], primary: Dict[str, Any]) -> Dict[str, Any]:
+    """Resolve the effective model/runtime for one turn.
+
+    Returns a dict with model/runtime/signature/label fields.
+    """
+    route = choose_cheap_model_route(user_message, routing_config)
+    if not route:
+        return {
+            "model": primary.get("model"),
+            "runtime": {
+                "api_key": primary.get("api_key"),
+                "base_url": primary.get("base_url"),
+                "provider": primary.get("provider"),
+                "api_mode": primary.get("api_mode"),
+                "command": primary.get("command"),
+                "args": list(primary.get("args") or []),
+                "credential_pool": primary.get("credential_pool"),
+            },
+            "label": None,
+            "signature": (
+                primary.get("model"),
+                primary.get("provider"),
+                primary.get("base_url"),
+                primary.get("api_mode"),
+                primary.get("command"),
+                tuple(primary.get("args") or ()),
+            ),
+        }
+
+    from hermes_cli.runtime_provider import resolve_runtime_provider
+
+    explicit_api_key = None
+    api_key_env = str(route.get("api_key_env") or "").strip()
+    if api_key_env:
+        explicit_api_key = os.getenv(api_key_env) or None
+
+    try:
+        runtime = resolve_runtime_provider(
+            requested=route.get("provider"),
+            explicit_api_key=explicit_api_key,
+            explicit_base_url=route.get("base_url"),
+        )
+    except Exception:
+        return {
+            "model": primary.get("model"),
+            "runtime": {
+                "api_key": primary.get("api_key"),
+                "base_url": primary.get("base_url"),
+                "provider": primary.get("provider"),
+                "api_mode": primary.get("api_mode"),
+                "command": primary.get("command"),
+                "args": list(primary.get("args") or []),
+                "credential_pool": primary.get("credential_pool"),
+            },
+            "label": None,
+            "signature": (
+                primary.get("model"),
+                primary.get("provider"),
+                primary.get("base_url"),
+                primary.get("api_mode"),
+                primary.get("command"),
+                tuple(primary.get("args") or ()),
+            ),
+        }
+
+    return {
+        "model": route.get("model"),
+        "runtime": {
+            "api_key": runtime.get("api_key"),
+            "base_url": runtime.get("base_url"),
+            "provider": runtime.get("provider"),
+            "api_mode": runtime.get("api_mode"),
+            "command": runtime.get("command"),
+            "args": list(runtime.get("args") or []),
+        },
+        "label": f"smart route → {route.get('model')} ({runtime.get('provider')})",
+        "signature": (
+            route.get("model"),
+            runtime.get("provider"),
+            runtime.get("base_url"),
+            runtime.get("api_mode"),
+            runtime.get("command"),
+            tuple(runtime.get("args") or ()),
+        ),
+    }

agent/subdirectory_hints.py

@@ -0,0 +1,219 @@
+"""Progressive subdirectory hint discovery.
+
+As the agent navigates into subdirectories via tool calls (read_file, terminal,
+search_files, etc.), this module discovers and loads project context files
+(AGENTS.md, CLAUDE.md, .cursorrules) from those directories.  Discovered hints
+are appended to the tool result so the model gets relevant context at the moment
+it starts working in a new area of the codebase.
+
+This complements the startup context loading in ``prompt_builder.py`` which only
+loads from the CWD.  Subdirectory hints are discovered lazily and injected into
+the conversation without modifying the system prompt (preserving prompt caching).
+
+Inspired by Block/goose's SubdirectoryHintTracker.
+"""
+
+import logging
+import os
+import re
+import shlex
+from pathlib import Path
+from typing import Dict, Any, Optional, Set
+
+from agent.prompt_builder import _scan_context_content
+
+logger = logging.getLogger(__name__)
+
+# Context files to look for in subdirectories, in priority order.
+# Same filenames as prompt_builder.py but we load ALL found (not first-wins)
+# since different subdirectories may use different conventions.
+_HINT_FILENAMES = [
+    "AGENTS.md", "agents.md",
+    "CLAUDE.md", "claude.md",
+    ".cursorrules",
+]
+
+# Maximum chars per hint file to prevent context bloat
+_MAX_HINT_CHARS = 8_000
+
+# Tool argument keys that typically contain file paths
+_PATH_ARG_KEYS = {"path", "file_path", "workdir"}
+
+# Tools that take shell commands where we should extract paths
+_COMMAND_TOOLS = {"terminal"}
+
+# How many parent directories to walk up when looking for hints.
+# Prevents scanning all the way to / for deeply nested paths.
+_MAX_ANCESTOR_WALK = 5
+
+class SubdirectoryHintTracker:
+    """Track which directories the agent visits and load hints on first access.
+
+    Usage::
+
+        tracker = SubdirectoryHintTracker(working_dir="/path/to/project")
+
+        # After each tool call:
+        hints = tracker.check_tool_call("read_file", {"path": "backend/src/main.py"})
+        if hints:
+            tool_result += hints  # append to the tool result string
+    """
+
+    def __init__(self, working_dir: Optional[str] = None):
+        self.working_dir = Path(working_dir or os.getcwd()).resolve()
+        self._loaded_dirs: Set[Path] = set()
+        # Pre-mark the working dir as loaded (startup context handles it)
+        self._loaded_dirs.add(self.working_dir)
+
+    def check_tool_call(
+        self,
+        tool_name: str,
+        tool_args: Dict[str, Any],
+    ) -> Optional[str]:
+        """Check tool call arguments for new directories and load any hint files.
+
+        Returns formatted hint text to append to the tool result, or None.
+        """
+        dirs = self._extract_directories(tool_name, tool_args)
+        if not dirs:
+            return None
+
+        all_hints = []
+        for d in dirs:
+            hints = self._load_hints_for_directory(d)
+            if hints:
+                all_hints.append(hints)
+
+        if not all_hints:
+            return None
+
+        return "\n\n" + "\n\n".join(all_hints)
+
+    def _extract_directories(
+        self, tool_name: str, args: Dict[str, Any]
+    ) -> list:
+        """Extract directory paths from tool call arguments."""
+        candidates: Set[Path] = set()
+
+        # Direct path arguments
+        for key in _PATH_ARG_KEYS:
+            val = args.get(key)
+            if isinstance(val, str) and val.strip():
+                self._add_path_candidate(val, candidates)
+
+        # Shell commands — extract path-like tokens
+        if tool_name in _COMMAND_TOOLS:
+            cmd = args.get("command", "")
+            if isinstance(cmd, str):
+                self._extract_paths_from_command(cmd, candidates)
+
+        return list(candidates)
+
+    def _add_path_candidate(self, raw_path: str, candidates: Set[Path]):
+        """Resolve a raw path and add its directory + ancestors to candidates.
+
+        Walks up from the resolved directory toward the filesystem root,
+        stopping at the first directory already in ``_loaded_dirs`` (or after
+        ``_MAX_ANCESTOR_WALK`` levels).  This ensures that reading
+        ``project/src/main.py`` discovers ``project/AGENTS.md`` even when
+        ``project/src/`` has no hint files of its own.
+        """
+        try:
+            p = Path(raw_path).expanduser()
+            if not p.is_absolute():
+                p = self.working_dir / p
+            p = p.resolve()
+            # Use parent if it's a file path (has extension or doesn't exist as dir)
+            if p.suffix or (p.exists() and p.is_file()):
+                p = p.parent
+            # Walk up ancestors — stop at already-loaded or root
+            for _ in range(_MAX_ANCESTOR_WALK):
+                if p in self._loaded_dirs:
+                    break
+                if self._is_valid_subdir(p):
+                    candidates.add(p)
+                parent = p.parent
+                if parent == p:
+                    break  # filesystem root
+                p = parent
+        except (OSError, ValueError):
+            pass
+
+    def _extract_paths_from_command(self, cmd: str, candidates: Set[Path]):
+        """Extract path-like tokens from a shell command string."""
+        try:
+            tokens = shlex.split(cmd)
+        except ValueError:
+            tokens = cmd.split()
+
+        for token in tokens:
+            # Skip flags
+            if token.startswith("-"):
+                continue
+            # Must look like a path (contains / or .)
+            if "/" not in token and "." not in token:
+                continue
+            # Skip URLs
+            if token.startswith(("http://", "https://", "git@")):
+                continue
+            self._add_path_candidate(token, candidates)
+
+    def _is_valid_subdir(self, path: Path) -> bool:
+        """Check if path is a valid directory to scan for hints."""
+        if not path.is_dir():
+            return False
+        if path in self._loaded_dirs:
+            return False
+        return True
+
+    def _load_hints_for_directory(self, directory: Path) -> Optional[str]:
+        """Load hint files from a directory. Returns formatted text or None."""
+        self._loaded_dirs.add(directory)
+
+        found_hints = []
+        for filename in _HINT_FILENAMES:
+            hint_path = directory / filename
+            if not hint_path.is_file():
+                continue
+            try:
+                content = hint_path.read_text(encoding="utf-8").strip()
+                if not content:
+                    continue
+                # Same security scan as startup context loading
+                content = _scan_context_content(content, filename)
+                if len(content) > _MAX_HINT_CHARS:
+                    content = (
+                        content[:_MAX_HINT_CHARS]
+                        + f"\n\n[...truncated {filename}: {len(content):,} chars total]"
+                    )
+                # Best-effort relative path for display
+                rel_path = str(hint_path)
+                try:
+                    rel_path = str(hint_path.relative_to(self.working_dir))
+                except ValueError:
+                    try:
+                        rel_path = str(hint_path.relative_to(Path.home()))
+                        rel_path = "~/" + rel_path
+                    except ValueError:
+                        pass  # keep absolute
+                found_hints.append((rel_path, content))
+                # First match wins per directory (like startup loading)
+                break
+            except Exception as exc:
+                logger.debug("Could not read %s: %s", hint_path, exc)
+
+        if not found_hints:
+            return None
+
+        sections = []
+        for rel_path, content in found_hints:
+            sections.append(
+                f"[Subdirectory context discovered: {rel_path}]\n{content}"
+            )
+
+        logger.debug(
+            "Loaded subdirectory hints from %s: %s",
+            directory,
+            [h[0] for h in found_hints],
+        )
+        return "\n\n".join(sections)

agent/title_generator.py

@@ -0,0 +1,125 @@
+"""Auto-generate short session titles from the first user/assistant exchange.
+
+Runs asynchronously after the first response is delivered so it never
+adds latency to the user-facing reply.
+"""
+
+import logging
+import threading
+from typing import Optional
+
+from agent.auxiliary_client import call_llm
+
+logger = logging.getLogger(__name__)
+
+_TITLE_PROMPT = (
+    "Generate a short, descriptive title (3-7 words) for a conversation that starts with the "
+    "following exchange. The title should capture the main topic or intent. "
+    "Return ONLY the title text, nothing else. No quotes, no punctuation at the end, no prefixes."
+)
+
+
+def generate_title(user_message: str, assistant_response: str, timeout: float = 30.0) -> Optional[str]:
+    """Generate a session title from the first exchange.
+
+    Uses the auxiliary LLM client (cheapest/fastest available model).
+    Returns the title string or None on failure.
+    """
+    # Truncate long messages to keep the request small
+    user_snippet = user_message[:500] if user_message else ""
+    assistant_snippet = assistant_response[:500] if assistant_response else ""
+
+    messages = [
+        {"role": "system", "content": _TITLE_PROMPT},
+        {"role": "user", "content": f"User: {user_snippet}\n\nAssistant: {assistant_snippet}"},
+    ]
+
+    try:
+        response = call_llm(
+            task="compression",  # reuse compression task config (cheap/fast model)
+            messages=messages,
+            max_tokens=30,
+            temperature=0.3,
+            timeout=timeout,
+        )
+        title = (response.choices[0].message.content or "").strip()
+        # Clean up: remove quotes, trailing punctuation, prefixes like "Title: "
+        title = title.strip('"\'')
+        if title.lower().startswith("title:"):
+            title = title[6:].strip()
+        # Enforce reasonable length
+        if len(title) > 80:
+            title = title[:77] + "..."
+        return title if title else None
+    except Exception as e:
+        logger.debug("Title generation failed: %s", e)
+        return None
+
+
+def auto_title_session(
+    session_db,
+    session_id: str,
+    user_message: str,
+    assistant_response: str,
+) -> None:
+    """Generate and set a session title if one doesn't already exist.
+
+    Called in a background thread after the first exchange completes.
+    Silently skips if:
+    - session_db is None
+    - session already has a title (user-set or previously auto-generated)
+    - title generation fails
+    """
+    if not session_db or not session_id:
+        return
+
+    # Check if title already exists (user may have set one via /title before first response)
+    try:
+        existing = session_db.get_session_title(session_id)
+        if existing:
+            return
+    except Exception:
+        return
+
+    title = generate_title(user_message, assistant_response)
+    if not title:
+        return
+
+    try:
+        session_db.set_session_title(session_id, title)
+        logger.debug("Auto-generated session title: %s", title)
+    except Exception as e:
+        logger.debug("Failed to set auto-generated title: %s", e)
+
+
+def maybe_auto_title(
+    session_db,
+    session_id: str,
+    user_message: str,
+    assistant_response: str,
+    conversation_history: list,
+) -> None:
+    """Fire-and-forget title generation after the first exchange.
+
+    Only generates a title when:
+    - This appears to be the first user→assistant exchange
+    - No title is already set
+    """
+    if not session_db or not session_id or not user_message or not assistant_response:
+        return
+
+    # Count user messages in history to detect first exchange.
+    # conversation_history includes the exchange that just happened,
+    # so for a first exchange we expect exactly 1 user message
+    # (or 2 counting system). Be generous: generate on first 2 exchanges.
+    user_msg_count = sum(1 for m in (conversation_history or []) if m.get("role") == "user")
+    if user_msg_count > 2:
+        return
+
+    thread = threading.Thread(
+        target=auto_title_session,
+        args=(session_db, session_id, user_message, assistant_response),
+        daemon=True,
+        name="auto-title",
+    )
+    thread.start()

agent/usage_pricing.py

@@ -0,0 +1,656 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from decimal import Decimal
+from typing import Any, Dict, Literal, Optional
+
+from agent.model_metadata import fetch_endpoint_model_metadata, fetch_model_metadata
+
+DEFAULT_PRICING = {"input": 0.0, "output": 0.0}
+
+_ZERO = Decimal("0")
+_ONE_MILLION = Decimal("1000000")
+
+CostStatus = Literal["actual", "estimated", "included", "unknown"]
+CostSource = Literal[
+    "provider_cost_api",
+    "provider_generation_api",
+    "provider_models_api",
+    "official_docs_snapshot",
+    "user_override",
+    "custom_contract",
+    "none",
+]
+
+
+@dataclass(frozen=True)
+class CanonicalUsage:
+    input_tokens: int = 0
+    output_tokens: int = 0
+    cache_read_tokens: int = 0
+    cache_write_tokens: int = 0
+    reasoning_tokens: int = 0
+    request_count: int = 1
+    raw_usage: Optional[dict[str, Any]] = None
+
+    @property
+    def prompt_tokens(self) -> int:
+        return self.input_tokens + self.cache_read_tokens + self.cache_write_tokens
+
+    @property
+    def total_tokens(self) -> int:
+        return self.prompt_tokens + self.output_tokens
+
+
+@dataclass(frozen=True)
+class BillingRoute:
+    provider: str
+    model: str
+    base_url: str = ""
+    billing_mode: str = "unknown"
+
+
+@dataclass(frozen=True)
+class PricingEntry:
+    input_cost_per_million: Optional[Decimal] = None
+    output_cost_per_million: Optional[Decimal] = None
+    cache_read_cost_per_million: Optional[Decimal] = None
+    cache_write_cost_per_million: Optional[Decimal] = None
+    request_cost: Optional[Decimal] = None
+    source: CostSource = "none"
+    source_url: Optional[str] = None
+    pricing_version: Optional[str] = None
+    fetched_at: Optional[datetime] = None
+
+
+@dataclass(frozen=True)
+class CostResult:
+    amount_usd: Optional[Decimal]
+    status: CostStatus
+    source: CostSource
+    label: str
+    fetched_at: Optional[datetime] = None
+    pricing_version: Optional[str] = None
+    notes: tuple[str, ...] = ()
+
+
+_UTC_NOW = lambda: datetime.now(timezone.utc)
+
+
+# Official docs snapshot entries. Models whose published pricing and cache
+# semantics are stable enough to encode exactly.
+_OFFICIAL_DOCS_PRICING: Dict[tuple[str, str], PricingEntry] = {
+    (
+        "anthropic",
+        "claude-opus-4-20250514",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("15.00"),
+        output_cost_per_million=Decimal("75.00"),
+        cache_read_cost_per_million=Decimal("1.50"),
+        cache_write_cost_per_million=Decimal("18.75"),
+        source="official_docs_snapshot",
+        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
+        pricing_version="anthropic-prompt-caching-2026-03-16",
+    ),
+    (
+        "anthropic",
+        "claude-sonnet-4-20250514",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("3.00"),
+        output_cost_per_million=Decimal("15.00"),
+        cache_read_cost_per_million=Decimal("0.30"),
+        cache_write_cost_per_million=Decimal("3.75"),
+        source="official_docs_snapshot",
+        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
+        pricing_version="anthropic-prompt-caching-2026-03-16",
+    ),
+    # OpenAI
+    (
+        "openai",
+        "gpt-4o",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("2.50"),
+        output_cost_per_million=Decimal("10.00"),
+        cache_read_cost_per_million=Decimal("1.25"),
+        source="official_docs_snapshot",
+        source_url="https://openai.com/api/pricing/",
+        pricing_version="openai-pricing-2026-03-16",
+    ),
+    (
+        "openai",
+        "gpt-4o-mini",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.15"),
+        output_cost_per_million=Decimal("0.60"),
+        cache_read_cost_per_million=Decimal("0.075"),
+        source="official_docs_snapshot",
+        source_url="https://openai.com/api/pricing/",
+        pricing_version="openai-pricing-2026-03-16",
+    ),
+    (
+        "openai",
+        "gpt-4.1",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("2.00"),
+        output_cost_per_million=Decimal("8.00"),
+        cache_read_cost_per_million=Decimal("0.50"),
+        source="official_docs_snapshot",
+        source_url="https://openai.com/api/pricing/",
+        pricing_version="openai-pricing-2026-03-16",
+    ),
+    (
+        "openai",
+        "gpt-4.1-mini",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.40"),
+        output_cost_per_million=Decimal("1.60"),
+        cache_read_cost_per_million=Decimal("0.10"),
+        source="official_docs_snapshot",
+        source_url="https://openai.com/api/pricing/",
+        pricing_version="openai-pricing-2026-03-16",
+    ),
+    (
+        "openai",
+        "gpt-4.1-nano",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.10"),
+        output_cost_per_million=Decimal("0.40"),
+        cache_read_cost_per_million=Decimal("0.025"),
+        source="official_docs_snapshot",
+        source_url="https://openai.com/api/pricing/",
+        pricing_version="openai-pricing-2026-03-16",
+    ),
+    (
+        "openai",
+        "o3",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("10.00"),
+        output_cost_per_million=Decimal("40.00"),
+        cache_read_cost_per_million=Decimal("2.50"),
+        source="official_docs_snapshot",
+        source_url="https://openai.com/api/pricing/",
+        pricing_version="openai-pricing-2026-03-16",
+    ),
+    (
+        "openai",
+        "o3-mini",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("1.10"),
+        output_cost_per_million=Decimal("4.40"),
+        cache_read_cost_per_million=Decimal("0.55"),
+        source="official_docs_snapshot",
+        source_url="https://openai.com/api/pricing/",
+        pricing_version="openai-pricing-2026-03-16",
+    ),
+    # Anthropic older models (pre-4.6 generation)
+    (
+        "anthropic",
+        "claude-3-5-sonnet-20241022",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("3.00"),
+        output_cost_per_million=Decimal("15.00"),
+        cache_read_cost_per_million=Decimal("0.30"),
+        cache_write_cost_per_million=Decimal("3.75"),
+        source="official_docs_snapshot",
+        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
+        pricing_version="anthropic-pricing-2026-03-16",
+    ),
+    (
+        "anthropic",
+        "claude-3-5-haiku-20241022",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.80"),
+        output_cost_per_million=Decimal("4.00"),
+        cache_read_cost_per_million=Decimal("0.08"),
+        cache_write_cost_per_million=Decimal("1.00"),
+        source="official_docs_snapshot",
+        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
+        pricing_version="anthropic-pricing-2026-03-16",
+    ),
+    (
+        "anthropic",
+        "claude-3-opus-20240229",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("15.00"),
+        output_cost_per_million=Decimal("75.00"),
+        cache_read_cost_per_million=Decimal("1.50"),
+        cache_write_cost_per_million=Decimal("18.75"),
+        source="official_docs_snapshot",
+        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
+        pricing_version="anthropic-pricing-2026-03-16",
+    ),
+    (
+        "anthropic",
+        "claude-3-haiku-20240307",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.25"),
+        output_cost_per_million=Decimal("1.25"),
+        cache_read_cost_per_million=Decimal("0.03"),
+        cache_write_cost_per_million=Decimal("0.30"),
+        source="official_docs_snapshot",
+        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
+        pricing_version="anthropic-pricing-2026-03-16",
+    ),
+    # DeepSeek
+    (
+        "deepseek",
+        "deepseek-chat",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.14"),
+        output_cost_per_million=Decimal("0.28"),
+        source="official_docs_snapshot",
+        source_url="https://api-docs.deepseek.com/quick_start/pricing",
+        pricing_version="deepseek-pricing-2026-03-16",
+    ),
+    (
+        "deepseek",
+        "deepseek-reasoner",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.55"),
+        output_cost_per_million=Decimal("2.19"),
+        source="official_docs_snapshot",
+        source_url="https://api-docs.deepseek.com/quick_start/pricing",
+        pricing_version="deepseek-pricing-2026-03-16",
+    ),
+    # Google Gemini
+    (
+        "google",
+        "gemini-2.5-pro",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("1.25"),
+        output_cost_per_million=Decimal("10.00"),
+        source="official_docs_snapshot",
+        source_url="https://ai.google.dev/pricing",
+        pricing_version="google-pricing-2026-03-16",
+    ),
+    (
+        "google",
+        "gemini-2.5-flash",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.15"),
+        output_cost_per_million=Decimal("0.60"),
+        source="official_docs_snapshot",
+        source_url="https://ai.google.dev/pricing",
+        pricing_version="google-pricing-2026-03-16",
+    ),
+    (
+        "google",
+        "gemini-2.0-flash",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.10"),
+        output_cost_per_million=Decimal("0.40"),
+        source="official_docs_snapshot",
+        source_url="https://ai.google.dev/pricing",
+        pricing_version="google-pricing-2026-03-16",
+    ),
+}
+
+
+def _to_decimal(value: Any) -> Optional[Decimal]:
+    if value is None:
+        return None
+    try:
+        return Decimal(str(value))
+    except Exception:
+        return None
+
+
+def _to_int(value: Any) -> int:
+    try:
+        return int(value or 0)
+    except Exception:
+        return 0
+
+
+def resolve_billing_route(
+    model_name: str,
+    provider: Optional[str] = None,
+    base_url: Optional[str] = None,
+) -> BillingRoute:
+    provider_name = (provider or "").strip().lower()
+    base = (base_url or "").strip().lower()
+    model = (model_name or "").strip()
+    if not provider_name and "/" in model:
+        inferred_provider, bare_model = model.split("/", 1)
+        if inferred_provider in {"anthropic", "openai", "google"}:
+            provider_name = inferred_provider
+            model = bare_model
+
+    if provider_name == "openai-codex":
+        return BillingRoute(provider="openai-codex", model=model, base_url=base_url or "", billing_mode="subscription_included")
+    if provider_name == "openrouter" or "openrouter.ai" in base:
+        return BillingRoute(provider="openrouter", model=model, base_url=base_url or "", billing_mode="official_models_api")
+    if provider_name == "anthropic":
+        return BillingRoute(provider="anthropic", model=model.split("/")[-1], base_url=base_url or "", billing_mode="official_docs_snapshot")
+    if provider_name == "openai":
+        return BillingRoute(provider="openai", model=model.split("/")[-1], base_url=base_url or "", billing_mode="official_docs_snapshot")
+    if provider_name in {"custom", "local"} or (base and "localhost" in base):
+        return BillingRoute(provider=provider_name or "custom", model=model, base_url=base_url or "", billing_mode="unknown")
+    return BillingRoute(provider=provider_name or "unknown", model=model.split("/")[-1] if model else "", base_url=base_url or "", billing_mode="unknown")
+
+
+def _lookup_official_docs_pricing(route: BillingRoute) -> Optional[PricingEntry]:
+    return _OFFICIAL_DOCS_PRICING.get((route.provider, route.model.lower()))
+
+
+def _openrouter_pricing_entry(route: BillingRoute) -> Optional[PricingEntry]:
+    return _pricing_entry_from_metadata(
+        fetch_model_metadata(),
+        route.model,
+        source_url="https://openrouter.ai/docs/api/api-reference/models/get-models",
+        pricing_version="openrouter-models-api",
+    )
+
+
+def _pricing_entry_from_metadata(
+    metadata: Dict[str, Dict[str, Any]],
+    model_id: str,
+    *,
+    source_url: str,
+    pricing_version: str,
+) -> Optional[PricingEntry]:
+    if model_id not in metadata:
+        return None
+    pricing = metadata[model_id].get("pricing") or {}
+    prompt = _to_decimal(pricing.get("prompt"))
+    completion = _to_decimal(pricing.get("completion"))
+    request = _to_decimal(pricing.get("request"))
+    cache_read = _to_decimal(
+        pricing.get("cache_read")
+        or pricing.get("cached_prompt")
+        or pricing.get("input_cache_read")
+    )
+    cache_write = _to_decimal(
+        pricing.get("cache_write")
+        or pricing.get("cache_creation")
+        or pricing.get("input_cache_write")
+    )
+    if prompt is None and completion is None and request is None:
+        return None
+
+    def _per_token_to_per_million(value: Optional[Decimal]) -> Optional[Decimal]:
+        if value is None:
+            return None
+        return value * _ONE_MILLION
+
+    return PricingEntry(
+        input_cost_per_million=_per_token_to_per_million(prompt),
+        output_cost_per_million=_per_token_to_per_million(completion),
+        cache_read_cost_per_million=_per_token_to_per_million(cache_read),
+        cache_write_cost_per_million=_per_token_to_per_million(cache_write),
+        request_cost=request,
+        source="provider_models_api",
+        source_url=source_url,
+        pricing_version=pricing_version,
+        fetched_at=_UTC_NOW(),
+    )
+
+
+def get_pricing_entry(
+    model_name: str,
+    provider: Optional[str] = None,
+    base_url: Optional[str] = None,
+    api_key: Optional[str] = None,
+) -> Optional[PricingEntry]:
+    route = resolve_billing_route(model_name, provider=provider, base_url=base_url)
+    if route.billing_mode == "subscription_included":
+        return PricingEntry(
+            input_cost_per_million=_ZERO,
+            output_cost_per_million=_ZERO,
+            cache_read_cost_per_million=_ZERO,
+            cache_write_cost_per_million=_ZERO,
+            source="none",
+            pricing_version="included-route",
+        )
+    if route.provider == "openrouter":
+        return _openrouter_pricing_entry(route)
+    if route.base_url:
+        entry = _pricing_entry_from_metadata(
+            fetch_endpoint_model_metadata(route.base_url, api_key=api_key or ""),
+            route.model,
+            source_url=f"{route.base_url.rstrip('/')}/models",
+            pricing_version="openai-compatible-models-api",
+        )
+        if entry:
+            return entry
+    return _lookup_official_docs_pricing(route)
+
+
+def normalize_usage(
+    response_usage: Any,
+    *,
+    provider: Optional[str] = None,
+    api_mode: Optional[str] = None,
+) -> CanonicalUsage:
+    """Normalize raw API response usage into canonical token buckets.
+
+    Handles three API shapes:
+    - Anthropic: input_tokens/output_tokens/cache_read_input_tokens/cache_creation_input_tokens
+    - Codex Responses: input_tokens includes cache tokens; input_tokens_details.cached_tokens separates them
+    - OpenAI Chat Completions: prompt_tokens includes cache tokens; prompt_tokens_details.cached_tokens separates them
+
+    In both Codex and OpenAI modes, input_tokens is derived by subtracting cache
+    tokens from the total — the API contract is that input/prompt totals include
+    cached tokens and the details object breaks them out.
+    """
+    if not response_usage:
+        return CanonicalUsage()
+
+    provider_name = (provider or "").strip().lower()
+    mode = (api_mode or "").strip().lower()
+
+    if mode == "anthropic_messages" or provider_name == "anthropic":
+        input_tokens = _to_int(getattr(response_usage, "input_tokens", 0))
+        output_tokens = _to_int(getattr(response_usage, "output_tokens", 0))
+        cache_read_tokens = _to_int(getattr(response_usage, "cache_read_input_tokens", 0))
+        cache_write_tokens = _to_int(getattr(response_usage, "cache_creation_input_tokens", 0))
+    elif mode == "codex_responses":
+        input_total = _to_int(getattr(response_usage, "input_tokens", 0))
+        output_tokens = _to_int(getattr(response_usage, "output_tokens", 0))
+        details = getattr(response_usage, "input_tokens_details", None)
+        cache_read_tokens = _to_int(getattr(details, "cached_tokens", 0) if details else 0)
+        cache_write_tokens = _to_int(
+            getattr(details, "cache_creation_tokens", 0) if details else 0
+        )
+        input_tokens = max(0, input_total - cache_read_tokens - cache_write_tokens)
+    else:
+        prompt_total = _to_int(getattr(response_usage, "prompt_tokens", 0))
+        output_tokens = _to_int(getattr(response_usage, "completion_tokens", 0))
+        details = getattr(response_usage, "prompt_tokens_details", None)
+        cache_read_tokens = _to_int(getattr(details, "cached_tokens", 0) if details else 0)
+        cache_write_tokens = _to_int(
+            getattr(details, "cache_write_tokens", 0) if details else 0
+        )
+        input_tokens = max(0, prompt_total - cache_read_tokens - cache_write_tokens)
+
+    reasoning_tokens = 0
+    output_details = getattr(response_usage, "output_tokens_details", None)
+    if output_details:
+        reasoning_tokens = _to_int(getattr(output_details, "reasoning_tokens", 0))
+
+    return CanonicalUsage(
+        input_tokens=input_tokens,
+        output_tokens=output_tokens,
+        cache_read_tokens=cache_read_tokens,
+        cache_write_tokens=cache_write_tokens,
+        reasoning_tokens=reasoning_tokens,
+    )
+
+
+def estimate_usage_cost(
+    model_name: str,
+    usage: CanonicalUsage,
+    *,
+    provider: Optional[str] = None,
+    base_url: Optional[str] = None,
+    api_key: Optional[str] = None,
+) -> CostResult:
+    route = resolve_billing_route(model_name, provider=provider, base_url=base_url)
+    if route.billing_mode == "subscription_included":
+        return CostResult(
+            amount_usd=_ZERO,
+            status="included",
+            source="none",
+            label="included",
+            pricing_version="included-route",
+        )
+
+    entry = get_pricing_entry(model_name, provider=provider, base_url=base_url, api_key=api_key)
+    if not entry:
+        return CostResult(amount_usd=None, status="unknown", source="none", label="n/a")
+
+    notes: list[str] = []
+    amount = _ZERO
+
+    if usage.input_tokens and entry.input_cost_per_million is None:
+        return CostResult(amount_usd=None, status="unknown", source=entry.source, label="n/a")
+    if usage.output_tokens and entry.output_cost_per_million is None:
+        return CostResult(amount_usd=None, status="unknown", source=entry.source, label="n/a")
+    if usage.cache_read_tokens:
+        if entry.cache_read_cost_per_million is None:
+            return CostResult(
+                amount_usd=None,
+                status="unknown",
+                source=entry.source,
+                label="n/a",
+                notes=("cache-read pricing unavailable for route",),
+            )
+    if usage.cache_write_tokens:
+        if entry.cache_write_cost_per_million is None:
+            return CostResult(
+                amount_usd=None,
+                status="unknown",
+                source=entry.source,
+                label="n/a",
+                notes=("cache-write pricing unavailable for route",),
+            )
+
+    if entry.input_cost_per_million is not None:
+        amount += Decimal(usage.input_tokens) * entry.input_cost_per_million / _ONE_MILLION
+    if entry.output_cost_per_million is not None:
+        amount += Decimal(usage.output_tokens) * entry.output_cost_per_million / _ONE_MILLION
+    if entry.cache_read_cost_per_million is not None:
+        amount += Decimal(usage.cache_read_tokens) * entry.cache_read_cost_per_million / _ONE_MILLION
+    if entry.cache_write_cost_per_million is not None:
+        amount += Decimal(usage.cache_write_tokens) * entry.cache_write_cost_per_million / _ONE_MILLION
+    if entry.request_cost is not None and usage.request_count:
+        amount += Decimal(usage.request_count) * entry.request_cost
+
+    status: CostStatus = "estimated"
+    label = f"~${amount:.2f}"
+    if entry.source == "none" and amount == _ZERO:
+        status = "included"
+        label = "included"
+
+    if route.provider == "openrouter":
+        notes.append("OpenRouter cost is estimated from the models API until reconciled.")
+
+    return CostResult(
+        amount_usd=amount,
+        status=status,
+        source=entry.source,
+        label=label,
+        fetched_at=entry.fetched_at,
+        pricing_version=entry.pricing_version,
+        notes=tuple(notes),
+    )
+
+
+def has_known_pricing(
+    model_name: str,
+    provider: Optional[str] = None,
+    base_url: Optional[str] = None,
+    api_key: Optional[str] = None,
+) -> bool:
+    """Check whether we have pricing data for this model+route.
+
+    Uses direct lookup instead of routing through the full estimation
+    pipeline — avoids creating dummy usage objects just to check status.
+    """
+    route = resolve_billing_route(model_name, provider=provider, base_url=base_url)
+    if route.billing_mode == "subscription_included":
+        return True
+    entry = get_pricing_entry(model_name, provider=provider, base_url=base_url, api_key=api_key)
+    return entry is not None
+
+
+def get_pricing(
+    model_name: str,
+    provider: Optional[str] = None,
+    base_url: Optional[str] = None,
+    api_key: Optional[str] = None,
+) -> Dict[str, float]:
+    """Backward-compatible thin wrapper for legacy callers.
+
+    Returns only non-cache input/output fields when a pricing entry exists.
+    Unknown routes return zeroes.
+    """
+    entry = get_pricing_entry(model_name, provider=provider, base_url=base_url, api_key=api_key)
+    if not entry:
+        return {"input": 0.0, "output": 0.0}
+    return {
+        "input": float(entry.input_cost_per_million or _ZERO),
+        "output": float(entry.output_cost_per_million or _ZERO),
+    }
+
+
+def estimate_cost_usd(
+    model: str,
+    input_tokens: int,
+    output_tokens: int,
+    *,
+    provider: Optional[str] = None,
+    base_url: Optional[str] = None,
+    api_key: Optional[str] = None,
+) -> float:
+    """Backward-compatible helper for legacy callers.
+
+    This uses non-cached input/output only. New code should call
+    `estimate_usage_cost()` with canonical usage buckets.
+    """
+    result = estimate_usage_cost(
+        model,
+        CanonicalUsage(input_tokens=input_tokens, output_tokens=output_tokens),
+        provider=provider,
+        base_url=base_url,
+        api_key=api_key,
+    )
+    return float(result.amount_usd or _ZERO)
+
+
+def format_duration_compact(seconds: float) -> str:
+    if seconds < 60:
+        return f"{seconds:.0f}s"
+    minutes = seconds / 60
+    if minutes < 60:
+        return f"{minutes:.0f}m"
+    hours = minutes / 60
+    if hours < 24:
+        remaining_min = int(minutes % 60)
+        return f"{int(hours)}h {remaining_min}m" if remaining_min else f"{int(hours)}h"
+    days = hours / 24
+    return f"{days:.1f}d"
+
+
+def format_token_count_compact(value: int) -> str:
+    abs_value = abs(int(value))
+    if abs_value < 1_000:
+        return str(int(value))
+
+    sign = "-" if value < 0 else ""
+    units = ((1_000_000_000, "B"), (1_000_000, "M"), (1_000, "K"))
+    for threshold, suffix in units:
+        if abs_value >= threshold:
+            scaled = abs_value / threshold
+            if scaled < 10:
+                text = f"{scaled:.2f}"
+            elif scaled < 100:
+                text = f"{scaled:.1f}"
+            else:
+                text = f"{scaled:.0f}"
+            if "." in text:
+                text = text.rstrip("0").rstrip(".")
+            return f"{sign}{text}{suffix}"
+
+    return f"{value:,}"

batch_runner.py

@@ -128,6 +128,7 @@ def _extract_tool_stats(messages: List[Dict[str, Any]]) -> Dict[str, Dict[str, i
         # Track tool calls from assistant messages
         if msg["role"] == "assistant" and "tool_calls" in msg and msg["tool_calls"]:
             for tool_call in msg["tool_calls"]:
+                if not tool_call or not isinstance(tool_call, dict): continue
                 tool_name = tool_call["function"]["name"]
                 tool_call_id = tool_call["id"]
                 

cli-config.yaml.example

@@ -7,17 +7,39 @@
 # =============================================================================
 model:
   # Default model to use (can be overridden with --model flag)
+  # Both "default" and "model" work as the key name here.
   default: "anthropic/claude-opus-4.6"
   
   # Inference provider selection:
-  #   "auto"       - Use Nous Portal if logged in, otherwise OpenRouter/env vars (default)
-  #   "nous-api"   - Use Nous Portal via API key (requires: NOUS_API_KEY)
-  #   "openrouter" - Always use OpenRouter API key from OPENROUTER_API_KEY
-  #   "nous"       - Always use Nous Portal (requires: hermes login)
-  #   "zai"        - Use z.ai / ZhipuAI GLM models (requires: GLM_API_KEY)
-  #   "kimi-coding"- Use Kimi / Moonshot AI models (requires: KIMI_API_KEY)
-  #   "minimax"    - Use MiniMax global endpoint (requires: MINIMAX_API_KEY)
-  #   "minimax-cn" - Use MiniMax China endpoint (requires: MINIMAX_CN_API_KEY)
+  #   "auto"         - Auto-detect from credentials (default)
+  #   "openrouter"   - OpenRouter (requires: OPENROUTER_API_KEY or OPENAI_API_KEY)
+  #   "nous"         - Nous Portal OAuth (requires: hermes login)
+  #   "nous-api"     - Nous Portal API key (requires: NOUS_API_KEY)
+  #   "anthropic"    - Direct Anthropic API (requires: ANTHROPIC_API_KEY)
+  #   "openai-codex" - OpenAI Codex (requires: hermes login --provider openai-codex)
+  #   "copilot"      - GitHub Copilot / GitHub Models (requires: GITHUB_TOKEN)
+  #   "zai"          - z.ai / ZhipuAI GLM (requires: GLM_API_KEY)
+  #   "kimi-coding"  - Kimi / Moonshot AI (requires: KIMI_API_KEY)
+  #   "minimax"      - MiniMax global (requires: MINIMAX_API_KEY)
+  #   "minimax-cn"   - MiniMax China (requires: MINIMAX_CN_API_KEY)
+  #   "huggingface"  - Hugging Face Inference (requires: HF_TOKEN)
+  #   "kilocode"     - KiloCode gateway (requires: KILOCODE_API_KEY)
+  #   "ai-gateway"   - Vercel AI Gateway (requires: AI_GATEWAY_API_KEY)
+  #
+  # Local servers (LM Studio, Ollama, vLLM, llama.cpp):
+  #   "custom"       - Any OpenAI-compatible endpoint. Set base_url below.
+  #   Aliases: "lmstudio", "ollama", "vllm", "llamacpp" all map to "custom".
+  #   Example for LM Studio:
+  #     provider: "lmstudio"
+  #     base_url: "http://localhost:1234/v1"
+  #   No API key needed — local servers typically ignore auth.
+  #
+  #   For Ollama Cloud (https://ollama.com/pricing):
+  #     provider: "custom"
+  #     base_url: "https://ollama.com/v1"
+  #   Set OLLAMA_API_KEY in .env — automatically picked up when base_url
+  #   points to ollama.com.
+  #
   # Can also be overridden with --provider flag or HERMES_INFERENCE_PROVIDER env var.
   provider: "auto"
   
@@ -51,6 +73,20 @@ model:
 #   # Data policy: "allow" (default) or "deny" to exclude providers that may store data
 #   # data_collection: "deny"
 
+# =============================================================================
+# Smart Model Routing (optional)
+# =============================================================================
+# Use a cheaper model for short/simple turns while keeping your main model for
+# more complex requests. Disabled by default.
+#
+# smart_model_routing:
+#   enabled: true
+#   max_simple_chars: 160
+#   max_simple_words: 28
+#   cheap_model:
+#     provider: openrouter
+#     model: google/gemini-2.5-flash
+
 # =============================================================================
 # Git Worktree Isolation
 # =============================================================================
@@ -76,8 +112,9 @@ model:
 #   - Messaging (Telegram/Discord): Uses MESSAGING_CWD from .env (default: home)
 terminal:
   backend: "local"
-  cwd: "."  # For local backend: "." = current directory. Ignored for remote backends.
+  cwd: "."  # For local backend: "." = current directory. Ignored for remote backends unless a backend documents otherwise.
   timeout: 180
+  docker_mount_cwd_to_workspace: false  # SECURITY: off by default. Opt in to mount the launch cwd into Docker /workspace.
   lifetime_seconds: 300
   # sudo_password: ""  # Enable sudo commands (pipes via sudo -S) - SECURITY WARNING: plaintext!
 
@@ -107,6 +144,13 @@ terminal:
 #   timeout: 180
 #   lifetime_seconds: 300
 #   docker_image: "nikolaik/python-nodejs:python3.11-nodejs20"
+#   docker_mount_cwd_to_workspace: true   # Explicit opt-in: mount your launch cwd into /workspace
+#   # Optional: explicitly forward selected env vars into Docker.
+#   # These values come from your current shell first, then ~/.hermes/.env.
+#   # Warning: anything forwarded here is visible to commands run in the container.
+#   docker_forward_env:
+#     - "GITHUB_TOKEN"
+#     - "NPM_TOKEN"
 
 # -----------------------------------------------------------------------------
 # OPTION 4: Singularity/Apptainer container
@@ -210,19 +254,34 @@ browser:
 # 1. Tracks actual token usage from API responses (not estimates)
 # 2. When prompt_tokens >= threshold% of model's context_length, triggers compression
 # 3. Protects first 3 turns (system prompt, initial request, first response)
-# 4. Protects last 4 turns (recent context is most relevant)
+# 4. Protects last N turns (default 20 messages = ~10 full turns of recent context)
 # 5. Summarizes middle turns using a fast/cheap model
 # 6. Inserts summary as a user message, continues conversation seamlessly
 #
+# Post-compression tail budget is target_ratio × threshold × context_length:
+#   200K context, threshold 0.50, ratio 0.20 → 20K tokens of recent tail preserved
+#   1M   context, threshold 0.50, ratio 0.20 → 100K tokens of recent tail preserved
+#
 compression:
   # Enable automatic context compression (default: true)
   # Set to false if you prefer to manage context manually or want errors on overflow
   enabled: true
   
-  # Trigger compression at this % of model's context limit (default: 0.85 = 85%)
+  # Trigger compression at this % of model's context limit (default: 0.50 = 50%)
   # Lower values = more aggressive compression, higher values = compress later
-  threshold: 0.85
+  threshold: 0.50
   
+  # Fraction of the threshold to preserve as recent tail (default: 0.20 = 20%)
+  # e.g. 20% of 50% threshold = 10% of total context kept as recent messages.
+  # Summary output is separately capped at 12K tokens (Gemini output limit).
+  # Range: 0.10 - 0.80
+  target_ratio: 0.20
+
+  # Number of most-recent messages to always preserve (default: 20 ≈ 10 full turns)
+  # Higher values keep more recent conversation intact at the cost of more aggressive
+  # compression of older turns.
+  protect_last_n: 20
+
   # Model to use for generating summaries (fast/cheap recommended)
   # This model compresses the middle turns into a concise summary.
   # IMPORTANT: it receives the full middle section of the conversation, so it
@@ -271,6 +330,9 @@ compression:
 #   vision:
 #     provider: "auto"
 #     model: ""              # e.g. "google/gemini-2.5-flash", "openai/gpt-4o"
+#     timeout: 30            # LLM API call timeout (seconds)
+#     download_timeout: 30   # Image HTTP download timeout (seconds)
+#                            # Increase for slow connections or self-hosted image servers
 #
 #   # Web page scraping / summarization + browser page text extraction
 #   web_extract:
@@ -333,6 +395,25 @@ session_reset:
   idle_minutes: 1440   # Inactivity timeout in minutes (default: 1440 = 24 hours)
   at_hour: 4           # Daily reset hour, 0-23 local time (default: 4 AM)
 
+# When true, group/channel chats use one session per participant when the platform
+# provides a user ID. This is the secure default and prevents users in the same
+# room from sharing context, interrupts, and token costs. Set false only if you
+# explicitly want one shared "room brain" per group/channel.
+group_sessions_per_user: true
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Gateway Streaming
+# ─────────────────────────────────────────────────────────────────────────────
+# Stream tokens to messaging platforms in real-time. The bot sends a message
+# on first token, then progressively edits it as more tokens arrive.
+# Disabled by default — enable to try the streaming UX on Telegram/Discord/Slack.
+streaming:
+  enabled: false
+  # transport: edit           # "edit" = progressive editMessageText
+  # edit_interval: 0.3        # seconds between message edits
+  # buffer_threshold: 40      # chars before forcing an edit flush
+  # cursor: " ▉"              # cursor shown during streaming
+
 # =============================================================================
 # Skills Configuration
 # =============================================================================
@@ -345,6 +426,15 @@ skills:
   # Set to 0 to disable.
   creation_nudge_interval: 15
 
+  # External skill directories — share skills across tools/agents without
+  # copying them into ~/.hermes/skills/.  Each path is expanded (~ and ${VAR})
+  # and resolved to an absolute path.  External dirs are read-only: skill
+  # creation always writes to ~/.hermes/skills/.  Local skills take precedence
+  # when names collide.
+  # external_dirs:
+  #   - ~/.agents/skills
+  #   - /home/shared/team-skills
+
 # =============================================================================
 # Agent Behavior
 # =============================================================================
@@ -383,7 +473,7 @@ agent:
 # Toolsets
 # =============================================================================
 # Control which tools the agent has access to.
-# Use "all" to enable everything, or specify individual toolsets.
+# Use `hermes tools` to interactively enable/disable tools per platform.
 
 # =============================================================================
 # Platform Toolsets (per-platform tool configuration)
@@ -455,7 +545,7 @@ platform_toolsets:
 #   skills_hub   - skill_hub (search/install/manage from online registries — user-driven only)
 #   moa          - mixture_of_agents  (requires OPENROUTER_API_KEY)
 #   todo         - todo (in-memory task planning, no deps)
-#   tts          - text_to_speech  (Edge TTS free, or ELEVENLABS/OPENAI key)
+#   tts          - text_to_speech  (Edge TTS free, or ELEVENLABS/OPENAI/MINIMAX key)
 #   cronjob      - cronjob (create/list/update/pause/resume/run/remove scheduled tasks)
 #   rl           - rl_list_environments, rl_start_training, etc. (requires TINKER_API_KEY)
 #
@@ -484,7 +574,7 @@ platform_toolsets:
 #   todo         - Task planning and tracking for multi-step work
 #   memory       - Persistent memory across sessions (personal notes + user profile)
 #   session_search - Search and recall past conversations (FTS5 + Gemini Flash summarization)
-#   tts          - Text-to-speech (Edge TTS free, ElevenLabs, OpenAI)
+#   tts          - Text-to-speech (Edge TTS free, ElevenLabs, OpenAI, MiniMax)
 #   cronjob      - Schedule and manage automated tasks (CLI-only)
 #   rl           - RL training tools (Tinker-Atropos)
 #
@@ -492,53 +582,11 @@ platform_toolsets:
 #   debugging    - terminal + web + file (for troubleshooting)
 #   safe         - web + vision + moa (no terminal access)
 
-# -----------------------------------------------------------------------------
-# OPTION 1: Enable all tools (default)
-# -----------------------------------------------------------------------------
-toolsets:
-  - all
-
-# -----------------------------------------------------------------------------
-# OPTION 2: Minimal - just web search and terminal
-# Great for: Simple coding tasks, quick lookups
-# -----------------------------------------------------------------------------
-# toolsets:
-#   - web
-#   - terminal
-
-# -----------------------------------------------------------------------------
-# OPTION 3: Research mode - no execution capabilities
-# Great for: Safe information gathering, research tasks
-# -----------------------------------------------------------------------------
-# toolsets:
-#   - web
-#   - vision
-#   - skills
-
-# -----------------------------------------------------------------------------
-# OPTION 4: Full automation - browser + terminal
-# Great for: Web scraping, automation tasks, testing
-# -----------------------------------------------------------------------------
-# toolsets:
-#   - terminal
-#   - browser
-#   - web
-
-# -----------------------------------------------------------------------------
-# OPTION 5: Creative mode - vision + image generation
-# Great for: Design work, image analysis, creative tasks
-# -----------------------------------------------------------------------------
-# toolsets:
-#   - vision
-#   - image_gen
-#   - web
-
-# -----------------------------------------------------------------------------
-# OPTION 6: Safe mode - no terminal or browser
-# Great for: Restricted environments, untrusted queries
-# -----------------------------------------------------------------------------
-# toolsets:
-#   - safe
+# NOTE: The top-level "toolsets" key is deprecated and ignored.
+# Tool configuration is managed per-platform via platform_toolsets above.
+# Use `hermes tools` to configure interactively, or edit platform_toolsets directly.
+#
+# CLI override: hermes chat --toolsets terminal,web,file
 
 # =============================================================================
 # MCP (Model Context Protocol) Servers
@@ -674,6 +722,12 @@ display:
   # Toggle at runtime with /verbose in the CLI
   tool_progress: all
 
+  # What Enter does when Hermes is already busy in the CLI.
+  #   interrupt: Interrupt the current run and redirect Hermes (default)
+  #   queue:     Queue your message for the next turn
+  # Ctrl+C always interrupts regardless of this setting.
+  busy_input_mode: interrupt
+
   # Background process notifications (gateway/messaging only).
   # Controls how chatty the process watcher is when you use
   # terminal(background=true, check_interval=...) from Telegram/Discord/etc.
@@ -694,6 +748,12 @@ display:
   # Toggle at runtime with /reasoning show or /reasoning hide.
   show_reasoning: false
 
+  # Stream tokens to the terminal as they arrive instead of waiting for the
+  # full response. The response box opens on first token and text appears
+  # line-by-line. Tool calls are still captured silently.
+  # Stream tokens to the terminal in real-time. Disable to wait for full responses.
+  streaming: true
+
   # ───────────────────────────────────────────────────────────────────────────
   # Skin / Theme
   # ───────────────────────────────────────────────────────────────────────────
@@ -734,3 +794,35 @@ display:
   #   tool_prefix: "╎"                       # Tool output line prefix (default: ┊)
   #
   skin: default
+
+# =============================================================================
+# Model Aliases — short names for /model command
+# =============================================================================
+# Map short aliases to exact (model, provider, base_url) tuples.
+# Used by /model tab completion and resolve_alias().
+# Aliases are checked BEFORE the models.dev catalog, so they can route
+# to endpoints not in the catalog (e.g. Ollama Cloud, local servers).
+#
+# model_aliases:
+#   opus:
+#     model: claude-opus-4-6
+#     provider: anthropic
+#   qwen:
+#     model: "qwen3.5:397b"
+#     provider: custom
+#     base_url: "https://ollama.com/v1"
+#   glm:
+#     model: glm-4.7
+#     provider: custom
+#     base_url: "https://ollama.com/v1"
+
+# =============================================================================
+# Privacy
+# =============================================================================
+# privacy:
+#   # Redact PII from the LLM context prompt.
+#   # When true, phone numbers are stripped and user/chat IDs are replaced
+#   # with deterministic hashes before being sent to the model.
+#   # Names and usernames are NOT affected (user-chosen, publicly visible).
+#   # Routing/delivery still uses the original values internally.
+#   redact_pii: false

cron/__init__.py

@@ -7,7 +7,8 @@ This module provides scheduled task execution, allowing the agent to:
 - Execute tasks in isolated sessions (no prior context)
 
 Cron jobs are executed automatically by the gateway daemon:
-    hermes gateway install    # Install as system service (recommended)
+    hermes gateway install    # Install as a user service
+    sudo hermes gateway install --system  # Linux servers: boot-time system service
     hermes gateway            # Or run in foreground
 
 The gateway ticks the scheduler every 60 seconds. A file lock prevents

cron/jobs.py

@@ -5,15 +5,20 @@ Jobs are stored in ~/.hermes/cron/jobs.json
 Output is saved to ~/.hermes/cron/output/{job_id}/{timestamp}.md
 """
 
+import copy
 import json
+import logging
 import tempfile
 import os
 import re
 import uuid
 from datetime import datetime, timedelta
 from pathlib import Path
+from hermes_constants import get_hermes_home
 from typing import Optional, Dict, List, Any
 
+logger = logging.getLogger(__name__)
+
 from hermes_time import now as _hermes_now
 
 try:
@@ -26,10 +31,11 @@ except ImportError:
 # Configuration
 # =============================================================================
 
-HERMES_DIR = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
+HERMES_DIR = get_hermes_home()
 CRON_DIR = HERMES_DIR / "cron"
 JOBS_FILE = CRON_DIR / "jobs.json"
 OUTPUT_DIR = CRON_DIR / "output"
+ONESHOT_GRACE_SECONDS = 120
 
 
 def _normalize_skill_list(skill: Optional[str] = None, skills: Optional[Any] = None) -> List[str]:
@@ -164,6 +170,10 @@ def parse_schedule(schedule: str) -> Dict[str, Any]:
         try:
             # Parse and validate
             dt = datetime.fromisoformat(schedule.replace('Z', '+00:00'))
+            # Make naive timestamps timezone-aware at parse time so the stored
+            # value doesn't depend on the system timezone matching at check time.
+            if dt.tzinfo is None:
+                dt = dt.astimezone()  # Interpret as local timezone
             return {
                 "kind": "once",
                 "run_at": dt.isoformat(),
@@ -212,6 +222,65 @@ def _ensure_aware(dt: datetime) -> datetime:
     return dt.astimezone(target_tz)
 
 
+def _recoverable_oneshot_run_at(
+    schedule: Dict[str, Any],
+    now: datetime,
+    *,
+    last_run_at: Optional[str] = None,
+) -> Optional[str]:
+    """Return a one-shot run time if it is still eligible to fire.
+
+    One-shot jobs get a small grace window so jobs created a few seconds after
+    their requested minute still run on the next tick. Once a one-shot has
+    already run, it is never eligible again.
+    """
+    if schedule.get("kind") != "once":
+        return None
+    if last_run_at:
+        return None
+
+    run_at = schedule.get("run_at")
+    if not run_at:
+        return None
+
+    run_at_dt = _ensure_aware(datetime.fromisoformat(run_at))
+    if run_at_dt >= now - timedelta(seconds=ONESHOT_GRACE_SECONDS):
+        return run_at
+    return None
+
+
+def _compute_grace_seconds(schedule: dict) -> int:
+    """Compute how late a job can be and still catch up instead of fast-forwarding.
+
+    Uses half the schedule period, clamped between 120 seconds and 2 hours.
+    This ensures daily jobs can catch up if missed by up to 2 hours,
+    while frequent jobs (every 5-10 min) still fast-forward quickly.
+    """
+    MIN_GRACE = 120
+    MAX_GRACE = 7200  # 2 hours
+
+    kind = schedule.get("kind")
+
+    if kind == "interval":
+        period_seconds = schedule.get("minutes", 1) * 60
+        grace = period_seconds // 2
+        return max(MIN_GRACE, min(grace, MAX_GRACE))
+
+    if kind == "cron" and HAS_CRONITER:
+        try:
+            now = _hermes_now()
+            cron = croniter(schedule["expr"], now)
+            first = cron.get_next(datetime)
+            second = cron.get_next(datetime)
+            period_seconds = int((second - first).total_seconds())
+            grace = period_seconds // 2
+            return max(MIN_GRACE, min(grace, MAX_GRACE))
+        except Exception:
+            pass
+
+    return MIN_GRACE
+
+
 def compute_next_run(schedule: Dict[str, Any], last_run_at: Optional[str] = None) -> Optional[str]:
     """
     Compute the next run time for a schedule.
@@ -221,9 +290,7 @@ def compute_next_run(schedule: Dict[str, Any], last_run_at: Optional[str] = None
     now = _hermes_now()
 
     if schedule["kind"] == "once":
-        run_at = _ensure_aware(datetime.fromisoformat(schedule["run_at"]))
-        # If in the future, return it; if in the past, no more runs
-        return schedule["run_at"] if run_at > now else None
+        return _recoverable_oneshot_run_at(schedule, now, last_run_at=last_run_at)
 
     elif schedule["kind"] == "interval":
         minutes = schedule["minutes"]
@@ -260,7 +327,20 @@ def load_jobs() -> List[Dict[str, Any]]:
         with open(JOBS_FILE, 'r', encoding='utf-8') as f:
             data = json.load(f)
             return data.get("jobs", [])
-    except (json.JSONDecodeError, IOError):
+    except json.JSONDecodeError:
+        # Retry with strict=False to handle bare control chars in string values
+        try:
+            with open(JOBS_FILE, 'r', encoding='utf-8') as f:
+                data = json.loads(f.read(), strict=False)
+                jobs = data.get("jobs", [])
+                if jobs:
+                    # Auto-repair: rewrite with proper escaping
+                    save_jobs(jobs)
+                    logger.warning("Auto-repaired jobs.json (had invalid control characters)")
+                return jobs
+        except Exception:
+            return []
+    except IOError:
         return []
 
 
@@ -292,6 +372,10 @@ def create_job(
     origin: Optional[Dict[str, Any]] = None,
     skill: Optional[str] = None,
     skills: Optional[List[str]] = None,
+    model: Optional[str] = None,
+    provider: Optional[str] = None,
+    base_url: Optional[str] = None,
+    script: Optional[str] = None,
 ) -> Dict[str, Any]:
     """
     Create a new cron job.
@@ -305,12 +389,22 @@ def create_job(
         origin: Source info where job was created (for "origin" delivery)
         skill: Optional legacy single skill name to load before running the prompt
         skills: Optional ordered list of skills to load before running the prompt
+        model: Optional per-job model override
+        provider: Optional per-job provider override
+        base_url: Optional per-job base URL override
+        script: Optional path to a Python script whose stdout is injected into the
+                prompt each run.  The script runs before the agent turn, and its output
+                is prepended as context.  Useful for data collection / change detection.
 
     Returns:
         The created job dict
     """
     parsed_schedule = parse_schedule(schedule)
 
+    # Normalize repeat: treat 0 or negative values as None (infinite)
+    if repeat is not None and repeat <= 0:
+        repeat = None
+
     # Auto-set repeat=1 for one-shot schedules if not specified
     if parsed_schedule["kind"] == "once" and repeat is None:
         repeat = 1
@@ -323,6 +417,15 @@ def create_job(
     now = _hermes_now().isoformat()
 
     normalized_skills = _normalize_skill_list(skill, skills)
+    normalized_model = str(model).strip() if isinstance(model, str) else None
+    normalized_provider = str(provider).strip() if isinstance(provider, str) else None
+    normalized_base_url = str(base_url).strip().rstrip("/") if isinstance(base_url, str) else None
+    normalized_model = normalized_model or None
+    normalized_provider = normalized_provider or None
+    normalized_base_url = normalized_base_url or None
+    normalized_script = str(script).strip() if isinstance(script, str) else None
+    normalized_script = normalized_script or None
+
     label_source = (prompt or (normalized_skills[0] if normalized_skills else None)) or "cron job"
     job = {
         "id": job_id,
@@ -330,6 +433,10 @@ def create_job(
         "prompt": prompt,
         "skills": normalized_skills,
         "skill": normalized_skills[0] if normalized_skills else None,
+        "model": normalized_model,
+        "provider": normalized_provider,
+        "base_url": normalized_base_url,
+        "script": normalized_script,
         "schedule": parsed_schedule,
         "schedule_display": parsed_schedule.get("display", schedule),
         "repeat": {
@@ -489,7 +596,7 @@ def mark_job_run(job_id: str, success: bool, error: Optional[str] = None):
                 # Check if we've hit the repeat limit
                 times = job["repeat"].get("times")
                 completed = job["repeat"]["completed"]
-                if times is not None and completed >= times:
+                if times is not None and times > 0 and completed >= times:
                     # Remove the job (limit reached)
                     jobs.pop(i)
                     save_jobs(jobs)
@@ -511,11 +618,47 @@ def mark_job_run(job_id: str, success: bool, error: Optional[str] = None):
     save_jobs(jobs)
 
 
+def advance_next_run(job_id: str) -> bool:
+    """Preemptively advance next_run_at for a recurring job before execution.
+
+    Call this BEFORE run_job() so that if the process crashes mid-execution,
+    the job won't re-fire on the next gateway restart.  This converts the
+    scheduler from at-least-once to at-most-once for recurring jobs — missing
+    one run is far better than firing dozens of times in a crash loop.
+
+    One-shot jobs are left unchanged so they can still retry on restart.
+
+    Returns True if next_run_at was advanced, False otherwise.
+    """
+    jobs = load_jobs()
+    for job in jobs:
+        if job["id"] == job_id:
+            kind = job.get("schedule", {}).get("kind")
+            if kind not in ("cron", "interval"):
+                return False
+            now = _hermes_now().isoformat()
+            new_next = compute_next_run(job["schedule"], now)
+            if new_next and new_next != job.get("next_run_at"):
+                job["next_run_at"] = new_next
+                save_jobs(jobs)
+                return True
+            return False
+    return False
+
+
 def get_due_jobs() -> List[Dict[str, Any]]:
-    """Get all jobs that are due to run now."""
+    """Get all jobs that are due to run now.
+
+    For recurring jobs (cron/interval), if the scheduled time is stale
+    (more than one period in the past, e.g. because the gateway was down),
+    the job is fast-forwarded to the next future run instead of firing
+    immediately.  This prevents a burst of missed jobs on gateway restart.
+    """
     now = _hermes_now()
-    jobs = [_apply_skill_fields(j) for j in load_jobs()]
+    raw_jobs = load_jobs()
+    jobs = [_apply_skill_fields(j) for j in copy.deepcopy(raw_jobs)]
     due = []
+    needs_save = False
 
     for job in jobs:
         if not job.get("enabled", True):
@@ -523,12 +666,62 @@ def get_due_jobs() -> List[Dict[str, Any]]:
 
         next_run = job.get("next_run_at")
         if not next_run:
-            continue
+            recovered_next = _recoverable_oneshot_run_at(
+                job.get("schedule", {}),
+                now,
+                last_run_at=job.get("last_run_at"),
+            )
+            if not recovered_next:
+                continue
+
+            job["next_run_at"] = recovered_next
+            next_run = recovered_next
+            logger.info(
+                "Job '%s' had no next_run_at; recovering one-shot run at %s",
+                job.get("name", job["id"]),
+                recovered_next,
+            )
+            for rj in raw_jobs:
+                if rj["id"] == job["id"]:
+                    rj["next_run_at"] = recovered_next
+                    needs_save = True
+                    break
 
         next_run_dt = _ensure_aware(datetime.fromisoformat(next_run))
         if next_run_dt <= now:
+            schedule = job.get("schedule", {})
+            kind = schedule.get("kind")
+
+            # For recurring jobs, check if the scheduled time is stale
+            # (gateway was down and missed the window). Fast-forward to
+            # the next future occurrence instead of firing a stale run.
+            grace = _compute_grace_seconds(schedule)
+            if kind in ("cron", "interval") and (now - next_run_dt).total_seconds() > grace:
+                # Job is past its catch-up grace window — this is a stale missed run.
+                # Grace scales with schedule period: daily=2h, hourly=30m, 10min=5m.
+                new_next = compute_next_run(schedule, now.isoformat())
+                if new_next:
+                    logger.info(
+                        "Job '%s' missed its scheduled time (%s, grace=%ds). "
+                        "Fast-forwarding to next run: %s",
+                        job.get("name", job["id"]),
+                        next_run,
+                        grace,
+                        new_next,
+                    )
+                    # Update the job in storage
+                    for rj in raw_jobs:
+                        if rj["id"] == job["id"]:
+                            rj["next_run_at"] = new_next
+                            needs_save = True
+                            break
+                    continue  # Skip this run
+
             due.append(job)
 
+    if needs_save:
+        save_jobs(raw_jobs)
+
     return due
 
 

cron/scheduler.py

@@ -9,11 +9,12 @@ runs at a time if multiple processes overlap.
 """
 
 import asyncio
+import concurrent.futures
 import json
 import logging
 import os
+import subprocess
 import sys
-import traceback
 
 # fcntl is Unix-only; on Windows use msvcrt for file locking
 try:
@@ -24,21 +25,35 @@ except ImportError:
         import msvcrt
     except ImportError:
         msvcrt = None
-from datetime import datetime
 from pathlib import Path
+from hermes_constants import get_hermes_home
+from hermes_cli.config import load_config
 from typing import Optional
 
 from hermes_time import now as _hermes_now
 
 logger = logging.getLogger(__name__)
 
+# Valid delivery platforms — used to validate user-supplied platform names
+# in cron delivery targets, preventing env var enumeration via crafted names.
+_KNOWN_DELIVERY_PLATFORMS = frozenset({
+    "telegram", "discord", "slack", "whatsapp", "signal",
+    "matrix", "mattermost", "homeassistant", "dingtalk", "feishu",
+    "wecom", "sms", "email", "webhook",
+})
+
 # Add parent directory to path for imports
 sys.path.insert(0, str(Path(__file__).parent.parent))
 
-from cron.jobs import get_due_jobs, mark_job_run, save_job_output
+from cron.jobs import get_due_jobs, mark_job_run, save_job_output, advance_next_run
+
+# Sentinel: when a cron agent has nothing new to report, it can start its
+# response with this marker to suppress delivery.  Output is still saved
+# locally for audit.
+SILENT_MARKER = "[SILENT]"
 
 # Resolve Hermes home directory (respects HERMES_HOME override)
-_hermes_home = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
+_hermes_home = get_hermes_home()
 
 # File-based lock prevents concurrent ticks from gateway + daemon + systemd timer
 _LOCK_DIR = _hermes_home / "cron"
@@ -66,20 +81,58 @@ def _resolve_delivery_target(job: dict) -> Optional[dict]:
         return None
 
     if deliver == "origin":
-        if not origin:
-            return None
-        return {
-            "platform": origin["platform"],
-            "chat_id": str(origin["chat_id"]),
-            "thread_id": origin.get("thread_id"),
-        }
+        if origin:
+            return {
+                "platform": origin["platform"],
+                "chat_id": str(origin["chat_id"]),
+                "thread_id": origin.get("thread_id"),
+            }
+        # Origin missing (e.g. job created via API/script) — try each
+        # platform's home channel as a fallback instead of silently dropping.
+        for platform_name in ("matrix", "telegram", "discord", "slack"):
+            chat_id = os.getenv(f"{platform_name.upper()}_HOME_CHANNEL", "")
+            if chat_id:
+                logger.info(
+                    "Job '%s' has deliver=origin but no origin; falling back to %s home channel",
+                    job.get("name", job.get("id", "?")),
+                    platform_name,
+                )
+                return {
+                    "platform": platform_name,
+                    "chat_id": chat_id,
+                    "thread_id": None,
+                }
+        return None
 
     if ":" in deliver:
-        platform_name, chat_id = deliver.split(":", 1)
+        platform_name, rest = deliver.split(":", 1)
+        platform_key = platform_name.lower()
+
+        from tools.send_message_tool import _parse_target_ref
+
+        parsed_chat_id, parsed_thread_id, is_explicit = _parse_target_ref(platform_key, rest)
+        if is_explicit:
+            chat_id, thread_id = parsed_chat_id, parsed_thread_id
+        else:
+            chat_id, thread_id = rest, None
+
+        # Resolve human-friendly labels like "Alice (dm)" to real IDs.
+        try:
+            from gateway.channel_directory import resolve_channel_name
+            resolved = resolve_channel_name(platform_key, chat_id)
+            if resolved:
+                parsed_chat_id, parsed_thread_id, resolved_is_explicit = _parse_target_ref(platform_key, resolved)
+                if resolved_is_explicit:
+                    chat_id, thread_id = parsed_chat_id, parsed_thread_id
+                else:
+                    chat_id = resolved
+        except Exception:
+            pass
+
         return {
             "platform": platform_name,
             "chat_id": chat_id,
-            "thread_id": None,
+            "thread_id": thread_id,
         }
 
     platform_name = deliver
@@ -90,6 +143,8 @@ def _resolve_delivery_target(job: dict) -> Optional[dict]:
             "thread_id": origin.get("thread_id"),
         }
 
+    if platform_name.lower() not in _KNOWN_DELIVERY_PLATFORMS:
+        return None
     chat_id = os.getenv(f"{platform_name.upper()}_HOME_CHANNEL", "")
     if not chat_id:
         return None
@@ -101,12 +156,14 @@ def _resolve_delivery_target(job: dict) -> Optional[dict]:
     }
 
 
-def _deliver_result(job: dict, content: str) -> None:
+def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> None:
     """
     Deliver job output to the configured target (origin chat, specific platform, etc.).
 
-    Uses the standalone platform send functions from send_message_tool so delivery
-    works whether or not the gateway is running.
+    When ``adapters`` and ``loop`` are provided (gateway is running), tries to
+    use the live adapter first — this supports E2EE rooms (e.g. Matrix) where
+    the standalone HTTP path cannot encrypt.  Falls back to standalone send if
+    the adapter path fails or is unavailable.
     """
     target = _resolve_delivery_target(job)
     if not target:
@@ -131,7 +188,14 @@ def _deliver_result(job: dict, content: str) -> None:
         "slack": Platform.SLACK,
         "whatsapp": Platform.WHATSAPP,
         "signal": Platform.SIGNAL,
+        "matrix": Platform.MATRIX,
+        "mattermost": Platform.MATTERMOST,
+        "homeassistant": Platform.HOMEASSISTANT,
+        "dingtalk": Platform.DINGTALK,
+        "feishu": Platform.FEISHU,
+        "wecom": Platform.WECOM,
         "email": Platform.EMAIL,
+        "sms": Platform.SMS,
     }
     platform = platform_map.get(platform_name.lower())
     if not platform:
@@ -149,15 +213,66 @@ def _deliver_result(job: dict, content: str) -> None:
         logger.warning("Job '%s': platform '%s' not configured/enabled", job["id"], platform_name)
         return
 
-    # Run the async send in a fresh event loop (safe from any thread)
+    # Optionally wrap the content with a header/footer so the user knows this
+    # is a cron delivery.  Wrapping is on by default; set cron.wrap_response: false
+    # in config.yaml for clean output.
+    wrap_response = True
     try:
-        result = asyncio.run(_send_to_platform(platform, pconfig, chat_id, content, thread_id=thread_id))
+        user_cfg = load_config()
+        wrap_response = user_cfg.get("cron", {}).get("wrap_response", True)
+    except Exception:
+        pass
+
+    if wrap_response:
+        task_name = job.get("name", job["id"])
+        delivery_content = (
+            f"Cronjob Response: {task_name}\n"
+            f"-------------\n\n"
+            f"{content}\n\n"
+            f"Note: The agent cannot see this message, and therefore cannot respond to it."
+        )
+    else:
+        delivery_content = content
+
+    # Prefer the live adapter when the gateway is running — this supports E2EE
+    # rooms (e.g. Matrix) where the standalone HTTP path cannot encrypt.
+    runtime_adapter = (adapters or {}).get(platform)
+    if runtime_adapter is not None and loop is not None and getattr(loop, "is_running", lambda: False)():
+        send_metadata = {"thread_id": thread_id} if thread_id else None
+        try:
+            future = asyncio.run_coroutine_threadsafe(
+                runtime_adapter.send(chat_id, delivery_content, metadata=send_metadata),
+                loop,
+            )
+            send_result = future.result(timeout=60)
+            if send_result and not getattr(send_result, "success", True):
+                err = getattr(send_result, "error", "unknown")
+                logger.warning(
+                    "Job '%s': live adapter send to %s:%s failed (%s), falling back to standalone",
+                    job["id"], platform_name, chat_id, err,
+                )
+            else:
+                logger.info("Job '%s': delivered to %s:%s via live adapter", job["id"], platform_name, chat_id)
+                return
+        except Exception as e:
+            logger.warning(
+                "Job '%s': live adapter delivery to %s:%s failed (%s), falling back to standalone",
+                job["id"], platform_name, chat_id, e,
+            )
+
+    # Standalone path: run the async send in a fresh event loop (safe from any thread)
+    coro = _send_to_platform(platform, pconfig, chat_id, delivery_content, thread_id=thread_id)
+    try:
+        result = asyncio.run(coro)
     except RuntimeError:
-        # asyncio.run() fails if there's already a running loop in this thread;
-        # spin up a new thread to avoid that.
+        # asyncio.run() checks for a running loop before awaiting the coroutine;
+        # when it raises, the original coro was never started — close it to
+        # prevent "coroutine was never awaited" RuntimeWarning, then retry in a
+        # fresh thread that has no running loop.
+        coro.close()
         import concurrent.futures
         with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
-            future = pool.submit(asyncio.run, _send_to_platform(platform, pconfig, chat_id, content, thread_id=thread_id))
+            future = pool.submit(asyncio.run, _send_to_platform(platform, pconfig, chat_id, delivery_content, thread_id=thread_id))
             result = future.result(timeout=30)
     except Exception as e:
         logger.error("Job '%s': delivery to %s:%s failed: %s", job["id"], platform_name, chat_id, e)
@@ -167,18 +282,115 @@ def _deliver_result(job: dict, content: str) -> None:
         logger.error("Job '%s': delivery error: %s", job["id"], result["error"])
     else:
         logger.info("Job '%s': delivered to %s:%s", job["id"], platform_name, chat_id)
-        # Mirror the delivered content into the target's gateway session
+
+
+_SCRIPT_TIMEOUT = 120  # seconds
+
+
+def _run_job_script(script_path: str) -> tuple[bool, str]:
+    """Execute a cron job's data-collection script and capture its output.
+
+    Args:
+        script_path: Path to a Python script (resolved via HERMES_HOME/scripts/ or absolute).
+
+    Returns:
+        (success, output) — on failure *output* contains the error message so the
+        LLM can report the problem to the user.
+    """
+    from hermes_constants import get_hermes_home
+
+    path = Path(script_path).expanduser()
+    if not path.is_absolute():
+        # Resolve relative paths against HERMES_HOME/scripts/
+        scripts_dir = get_hermes_home() / "scripts"
+        path = (scripts_dir / path).resolve()
+        # Guard against path traversal (e.g. "../../etc/passwd")
         try:
-            from gateway.mirror import mirror_to_session
-            mirror_to_session(platform_name, chat_id, content, source_label="cron", thread_id=thread_id)
-        except Exception as e:
-            logger.warning("Job '%s': mirror_to_session failed: %s", job["id"], e)
+            path.relative_to(scripts_dir.resolve())
+        except ValueError:
+            return False, f"Script path escapes the scripts directory: {script_path!r}"
+
+    if not path.exists():
+        return False, f"Script not found: {path}"
+    if not path.is_file():
+        return False, f"Script path is not a file: {path}"
+
+    try:
+        result = subprocess.run(
+            [sys.executable, str(path)],
+            capture_output=True,
+            text=True,
+            timeout=_SCRIPT_TIMEOUT,
+            cwd=str(path.parent),
+        )
+        stdout = (result.stdout or "").strip()
+        stderr = (result.stderr or "").strip()
+
+        if result.returncode != 0:
+            parts = [f"Script exited with code {result.returncode}"]
+            if stderr:
+                parts.append(f"stderr:\n{stderr}")
+            if stdout:
+                parts.append(f"stdout:\n{stdout}")
+            return False, "\n".join(parts)
+
+        # Redact any secrets that may appear in script output before
+        # they are injected into the LLM prompt context.
+        try:
+            from agent.redact import redact_sensitive_text
+            stdout = redact_sensitive_text(stdout)
+        except Exception:
+            pass
+        return True, stdout
+
+    except subprocess.TimeoutExpired:
+        return False, f"Script timed out after {_SCRIPT_TIMEOUT}s: {path}"
+    except Exception as exc:
+        return False, f"Script execution failed: {exc}"
 
 
 def _build_job_prompt(job: dict) -> str:
     """Build the effective prompt for a cron job, optionally loading one or more skills first."""
     prompt = job.get("prompt", "")
     skills = job.get("skills")
+
+    # Run data-collection script if configured, inject output as context.
+    script_path = job.get("script")
+    if script_path:
+        success, script_output = _run_job_script(script_path)
+        if success:
+            if script_output:
+                prompt = (
+                    "## Script Output\n"
+                    "The following data was collected by a pre-run script. "
+                    "Use it as context for your analysis.\n\n"
+                    f"```\n{script_output}\n```\n\n"
+                    f"{prompt}"
+                )
+            else:
+                prompt = (
+                    "[Script ran successfully but produced no output.]\n\n"
+                    f"{prompt}"
+                )
+        else:
+            prompt = (
+                "## Script Error\n"
+                "The data-collection script failed. Report this to the user.\n\n"
+                f"```\n{script_output}\n```\n\n"
+                f"{prompt}"
+            )
+
+    # Always prepend [SILENT] guidance so the cron agent can suppress
+    # delivery when it has nothing new or noteworthy to report.
+    silent_hint = (
+        "[SYSTEM: If you have a meaningful status report or findings, "
+        "send them — that is the whole point of this job. Only respond "
+        "with exactly \"[SILENT]\" (nothing else) when there is genuinely "
+        "nothing new to report. [SILENT] suppresses delivery to the user. "
+        "Never combine [SILENT] with content — either report your "
+        "findings normally, or say [SILENT] and nothing more.]\n\n"
+    )
+    prompt = silent_hint + prompt
     if skills is None:
         legacy = job.get("skill")
         skills = [legacy] if legacy else []
@@ -190,11 +402,14 @@ def _build_job_prompt(job: dict) -> str:
     from tools.skills_tool import skill_view
 
     parts = []
+    skipped: list[str] = []
     for skill_name in skill_names:
         loaded = json.loads(skill_view(skill_name))
         if not loaded.get("success"):
             error = loaded.get("error") or f"Failed to load skill '{skill_name}'"
-            raise RuntimeError(error)
+            logger.warning("Cron job '%s': skill not found, skipping — %s", job.get("name", job.get("id")), error)
+            skipped.append(skill_name)
+            continue
 
         content = str(loaded.get("content") or "").strip()
         if parts:
@@ -207,6 +422,15 @@ def _build_job_prompt(job: dict) -> str:
             ]
         )
 
+    if skipped:
+        notice = (
+            f"[SYSTEM: The following skill(s) were listed for this job but could not be found "
+            f"and were skipped: {', '.join(skipped)}. "
+            f"Start your response with a brief notice so the user is aware, e.g.: "
+            f"'⚠️ Skill(s) not found and skipped: {', '.join(skipped)}']"
+        )
+        parts.insert(0, notice)
+
     if prompt:
         parts.extend(["", f"The user has provided the following instruction alongside the skill invocation: {prompt}"])
     return "\n".join(parts)
@@ -234,6 +458,7 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
     job_name = job["name"]
     prompt = _build_job_prompt(job)
     origin = _resolve_origin(job)
+    _cron_session_id = f"cron_{job_id}_{_hermes_now().strftime('%Y%m%d_%H%M%S')}"
 
     logger.info("Running job '%s' (ID: %s)", job_name, job_id)
     logger.info("Prompt: %s", prompt[:100])
@@ -261,7 +486,7 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
             if delivery_target.get("thread_id") is not None:
                 os.environ["HERMES_CRON_AUTO_DELIVER_THREAD_ID"] = str(delivery_target["thread_id"])
 
-        model = os.getenv("HERMES_MODEL") or "anthropic/claude-opus-4.6"
+        model = job.get("model") or os.getenv("HERMES_MODEL") or ""
 
         # Load config.yaml for model, reasoning, prefill, toolsets, provider routing
         _cfg = {}
@@ -272,24 +497,20 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
                 with open(_cfg_path) as _f:
                     _cfg = yaml.safe_load(_f) or {}
                 _model_cfg = _cfg.get("model", {})
-                if isinstance(_model_cfg, str):
-                    model = _model_cfg
-                elif isinstance(_model_cfg, dict):
-                    model = _model_cfg.get("default", model)
+                if not job.get("model"):
+                    if isinstance(_model_cfg, str):
+                        model = _model_cfg
+                    elif isinstance(_model_cfg, dict):
+                        model = _model_cfg.get("default", model)
         except Exception as e:
             logger.warning("Job '%s': failed to load config.yaml, using defaults: %s", job_id, e)
 
         # Reasoning config from env or config.yaml
-        reasoning_config = None
+        from hermes_constants import parse_reasoning_effort
         effort = os.getenv("HERMES_REASONING_EFFORT", "")
         if not effort:
             effort = str(_cfg.get("agent", {}).get("reasoning_effort", "")).strip()
-        if effort and effort.lower() != "none":
-            valid = ("xhigh", "high", "medium", "low", "minimal")
-            if effort.lower() in valid:
-                reasoning_config = {"enabled": True, "effort": effort.lower()}
-        elif effort.lower() == "none":
-            reasoning_config = {"enabled": False}
+        reasoning_config = parse_reasoning_effort(effort)
 
         # Prefill messages from env or config.yaml
         prefill_messages = None
@@ -314,25 +535,46 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
 
         # Provider routing
         pr = _cfg.get("provider_routing", {})
+        smart_routing = _cfg.get("smart_model_routing", {}) or {}
 
         from hermes_cli.runtime_provider import (
             resolve_runtime_provider,
             format_runtime_provider_error,
         )
         try:
-            runtime = resolve_runtime_provider(
-                requested=os.getenv("HERMES_INFERENCE_PROVIDER"),
-            )
+            runtime_kwargs = {
+                "requested": job.get("provider") or os.getenv("HERMES_INFERENCE_PROVIDER"),
+            }
+            if job.get("base_url"):
+                runtime_kwargs["explicit_base_url"] = job.get("base_url")
+            runtime = resolve_runtime_provider(**runtime_kwargs)
         except Exception as exc:
             message = format_runtime_provider_error(exc)
             raise RuntimeError(message) from exc
 
+        from agent.smart_model_routing import resolve_turn_route
+        turn_route = resolve_turn_route(
+            prompt,
+            smart_routing,
+            {
+                "model": model,
+                "api_key": runtime.get("api_key"),
+                "base_url": runtime.get("base_url"),
+                "provider": runtime.get("provider"),
+                "api_mode": runtime.get("api_mode"),
+                "command": runtime.get("command"),
+                "args": list(runtime.get("args") or []),
+            },
+        )
+
         agent = AIAgent(
-            model=model,
-            api_key=runtime.get("api_key"),
-            base_url=runtime.get("base_url"),
-            provider=runtime.get("provider"),
-            api_mode=runtime.get("api_mode"),
+            model=turn_route["model"],
+            api_key=turn_route["runtime"].get("api_key"),
+            base_url=turn_route["runtime"].get("base_url"),
+            provider=turn_route["runtime"].get("provider"),
+            api_mode=turn_route["runtime"].get("api_mode"),
+            acp_command=turn_route["runtime"].get("command"),
+            acp_args=turn_route["runtime"].get("args"),
             max_iterations=max_iterations,
             reasoning_config=reasoning_config,
             prefill_messages=prefill_messages,
@@ -340,18 +582,42 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
             providers_ignored=pr.get("ignore"),
             providers_order=pr.get("order"),
             provider_sort=pr.get("sort"),
-            disabled_toolsets=["cronjob"],
+            disabled_toolsets=["cronjob", "messaging", "clarify"],
             quiet_mode=True,
+            skip_memory=True,  # Cron system prompts would corrupt user representations
             platform="cron",
-            session_id=f"cron_{job_id}_{_hermes_now().strftime('%Y%m%d_%H%M%S')}",
+            session_id=_cron_session_id,
             session_db=_session_db,
         )
         
-        result = agent.run_conversation(prompt)
-        
-        final_response = result.get("final_response", "")
-        if not final_response:
-            final_response = "(No response generated)"
+        # Run the agent with a timeout so a hung API call or tool doesn't
+        # block the cron ticker thread indefinitely.  Default 10 minutes;
+        # override via env var.  Uses a separate thread because
+        # run_conversation is synchronous.
+        _cron_timeout = float(os.getenv("HERMES_CRON_TIMEOUT", 600))
+        _cron_pool = concurrent.futures.ThreadPoolExecutor(max_workers=1)
+        _cron_future = _cron_pool.submit(agent.run_conversation, prompt)
+        try:
+            result = _cron_future.result(timeout=_cron_timeout)
+        except concurrent.futures.TimeoutError:
+            logger.error(
+                "Job '%s' timed out after %.0fs — interrupting agent",
+                job_name, _cron_timeout,
+            )
+            if hasattr(agent, "interrupt"):
+                agent.interrupt("Cron job timed out")
+            _cron_pool.shutdown(wait=False, cancel_futures=True)
+            raise TimeoutError(
+                f"Cron job '{job_name}' timed out after "
+                f"{int(_cron_timeout // 60)} minutes"
+            )
+        finally:
+            _cron_pool.shutdown(wait=False)
+
+        final_response = result.get("final_response", "") or ""
+        # Use a separate variable for log display; keep final_response clean
+        # for delivery logic (empty response = no delivery).
+        logged_response = final_response if final_response else "(No response generated)"
         
         output = f"""# Cron Job: {job_name}
 
@@ -365,7 +631,7 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
 
 ## Response
 
-{final_response}
+{logged_response}
 """
         
         logger.info("Job '%s' completed successfully", job_name)
@@ -373,7 +639,7 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
         
     except Exception as e:
         error_msg = f"{type(e).__name__}: {str(e)}"
-        logger.error("Job '%s' failed: %s", job_name, error_msg)
+        logger.exception("Job '%s' failed: %s", job_name, error_msg)
         
         output = f"""# Cron Job: {job_name} (FAILED)
 
@@ -389,8 +655,6 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
 
 ```
 {error_msg}
-
-{traceback.format_exc()}
 ```
 """
         return False, output, "", error_msg
@@ -407,13 +671,17 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
         ):
             os.environ.pop(key, None)
         if _session_db:
+            try:
+                _session_db.end_session(_cron_session_id, "cron_complete")
+            except (Exception, KeyboardInterrupt) as e:
+                logger.debug("Job '%s': failed to end session: %s", job_id, e)
             try:
                 _session_db.close()
-            except Exception as e:
+            except (Exception, KeyboardInterrupt) as e:
                 logger.debug("Job '%s': failed to close SQLite session store: %s", job_id, e)
 
 
-def tick(verbose: bool = True) -> int:
+def tick(verbose: bool = True, adapters=None, loop=None) -> int:
     """
     Check and run all due jobs.
     
@@ -422,6 +690,8 @@ def tick(verbose: bool = True) -> int:
     
     Args:
         verbose: Whether to print status messages
+        adapters: Optional dict mapping Platform → live adapter (from gateway)
+        loop: Optional asyncio event loop (from gateway) for live adapter sends
     
     Returns:
         Number of jobs executed (0 if another tick is already running)
@@ -455,17 +725,30 @@ def tick(verbose: bool = True) -> int:
         executed = 0
         for job in due_jobs:
             try:
+                # For recurring jobs (cron/interval), advance next_run_at to the
+                # next future occurrence BEFORE execution.  This way, if the
+                # process crashes mid-run, the job won't re-fire on restart.
+                # One-shot jobs are left alone so they can retry on restart.
+                advance_next_run(job["id"])
+
                 success, output, final_response, error = run_job(job)
 
                 output_file = save_job_output(job["id"], output)
                 if verbose:
                     logger.info("Output saved to: %s", output_file)
 
-                # Deliver the final response to the origin/target chat
+                # Deliver the final response to the origin/target chat.
+                # If the agent responded with [SILENT], skip delivery (but
+                # output is already saved above).  Failed jobs always deliver.
                 deliver_content = final_response if success else f"⚠️ Cron job '{job.get('name', job['id'])}' failed:\n{error}"
-                if deliver_content:
+                should_deliver = bool(deliver_content)
+                if should_deliver and success and deliver_content.strip().upper().startswith(SILENT_MARKER):
+                    logger.info("Job '%s': agent returned %s — skipping delivery", job["id"], SILENT_MARKER)
+                    should_deliver = False
+
+                if should_deliver:
                     try:
-                        _deliver_result(job, deliver_content)
+                        _deliver_result(job, deliver_content, adapters=adapters, loop=loop)
                     except Exception as de:
                         logger.error("Delivery failed for job %s: %s", job["id"], de)
 

docker/SOUL.md

@@ -0,0 +1,15 @@
+# Hermes Agent Persona
+
+<!--
+This file defines the agent's personality and tone.
+The agent will embody whatever you write here.
+Edit this to customize how Hermes communicates with you.
+
+Examples:
+  - "You are a warm, playful assistant who uses kaomoji occasionally."
+  - "You are a concise technical expert. No fluff, just facts."
+  - "You speak like a friendly coworker who happens to know everything."
+
+This file is loaded fresh each message -- no restart needed.
+Delete the contents (or this file) to use the default personality.
+-->

docker/entrypoint.sh

@@ -0,0 +1,34 @@
+#!/bin/bash
+# Docker entrypoint: bootstrap config files into the mounted volume, then run hermes.
+set -e
+
+HERMES_HOME="/opt/data"
+INSTALL_DIR="/opt/hermes"
+
+# Create essential directory structure.  Cache and platform directories
+# (cache/images, cache/audio, platforms/whatsapp, etc.) are created on
+# demand by the application — don't pre-create them here so new installs
+# get the consolidated layout from get_hermes_dir().
+mkdir -p "$HERMES_HOME"/{cron,sessions,logs,hooks,memories,skills}
+
+# .env
+if [ ! -f "$HERMES_HOME/.env" ]; then
+    cp "$INSTALL_DIR/.env.example" "$HERMES_HOME/.env"
+fi
+
+# config.yaml
+if [ ! -f "$HERMES_HOME/config.yaml" ]; then
+    cp "$INSTALL_DIR/cli-config.yaml.example" "$HERMES_HOME/config.yaml"
+fi
+
+# SOUL.md
+if [ ! -f "$HERMES_HOME/SOUL.md" ]; then
+    cp "$INSTALL_DIR/docker/SOUL.md" "$HERMES_HOME/SOUL.md"
+fi
+
+# Sync bundled skills (manifest-based so user edits are preserved)
+if [ -d "$INSTALL_DIR/skills" ]; then
+    python3 "$INSTALL_DIR/tools/skills_sync.py"
+fi
+
+exec hermes "$@"

docs/acp-setup.md

@@ -76,14 +76,13 @@ Open Zed settings (`Cmd+,` on macOS or `Ctrl+,` on Linux) and add to your
 
 ```json
 {
-  "acp": {
-    "agents": [
-      {
-        "name": "hermes-agent",
-        "registry_dir": "/path/to/hermes-agent/acp_registry"
-      }
-    ]
-  }
+  "agent_servers": {
+    "hermes-agent": {
+      "type": "custom",
+      "command": "hermes",
+      "args": ["acp"],
+    },
+  },
 }
 ```
 

docs/plans/2026-03-16-pricing-accuracy-architecture-design.md

@@ -0,0 +1,608 @@
+# Pricing Accuracy Architecture
+
+Date: 2026-03-16
+
+## Goal
+
+Hermes should only show dollar costs when they are backed by an official source for the user's actual billing path.
+
+This design replaces the current static, heuristic pricing flow in:
+
+- `run_agent.py`
+- `agent/usage_pricing.py`
+- `agent/insights.py`
+- `cli.py`
+
+with a provider-aware pricing system that:
+
+- handles cache billing correctly
+- distinguishes `actual` vs `estimated` vs `included` vs `unknown`
+- reconciles post-hoc costs when providers expose authoritative billing data
+- supports direct providers, OpenRouter, subscriptions, enterprise pricing, and custom endpoints
+
+## Problems In The Current Design
+
+Current Hermes behavior has four structural issues:
+
+1. It stores only `prompt_tokens` and `completion_tokens`, which is insufficient for providers that bill cache reads and cache writes separately.
+2. It uses a static model price table and fuzzy heuristics, which can drift from current official pricing.
+3. It assumes public API list pricing matches the user's real billing path.
+4. It has no distinction between live estimates and reconciled billed cost.
+
+## Design Principles
+
+1. Normalize usage before pricing.
+2. Never fold cached tokens into plain input cost.
+3. Track certainty explicitly.
+4. Treat the billing path as part of the model identity.
+5. Prefer official machine-readable sources over scraped docs.
+6. Use post-hoc provider cost APIs when available.
+7. Show `n/a` rather than inventing precision.
+
+## High-Level Architecture
+
+The new system has four layers:
+
+1. `usage_normalization`
+   Converts raw provider usage into a canonical usage record.
+2. `pricing_source_resolution`
+   Determines the billing path, source of truth, and applicable pricing source.
+3. `cost_estimation_and_reconciliation`
+   Produces an immediate estimate when possible, then replaces or annotates it with actual billed cost later.
+4. `presentation`
+   `/usage`, `/insights`, and the status bar display cost with certainty metadata.
+
+## Canonical Usage Record
+
+Add a canonical usage model that every provider path maps into before any pricing math happens.
+
+Suggested structure:
+
+```python
+@dataclass
+class CanonicalUsage:
+    provider: str
+    billing_provider: str
+    model: str
+    billing_route: str
+
+    input_tokens: int = 0
+    output_tokens: int = 0
+    cache_read_tokens: int = 0
+    cache_write_tokens: int = 0
+    reasoning_tokens: int = 0
+    request_count: int = 1
+
+    raw_usage: dict[str, Any] | None = None
+    raw_usage_fields: dict[str, str] | None = None
+    computed_fields: set[str] | None = None
+
+    provider_request_id: str | None = None
+    provider_generation_id: str | None = None
+    provider_response_id: str | None = None
+```
+
+Rules:
+
+- `input_tokens` means non-cached input only.
+- `cache_read_tokens` and `cache_write_tokens` are never merged into `input_tokens`.
+- `output_tokens` excludes cache metrics.
+- `reasoning_tokens` is telemetry unless a provider officially bills it separately.
+
+This is the same normalization pattern used by `opencode`, extended with provenance and reconciliation ids.
+
+## Provider Normalization Rules
+
+### OpenAI Direct
+
+Source usage fields:
+
+- `prompt_tokens`
+- `completion_tokens`
+- `prompt_tokens_details.cached_tokens`
+
+Normalization:
+
+- `cache_read_tokens = cached_tokens`
+- `input_tokens = prompt_tokens - cached_tokens`
+- `cache_write_tokens = 0` unless OpenAI exposes it in the relevant route
+- `output_tokens = completion_tokens`
+
+### Anthropic Direct
+
+Source usage fields:
+
+- `input_tokens`
+- `output_tokens`
+- `cache_read_input_tokens`
+- `cache_creation_input_tokens`
+
+Normalization:
+
+- `input_tokens = input_tokens`
+- `output_tokens = output_tokens`
+- `cache_read_tokens = cache_read_input_tokens`
+- `cache_write_tokens = cache_creation_input_tokens`
+
+### OpenRouter
+
+Estimate-time usage normalization should use the response usage payload with the same rules as the underlying provider when possible.
+
+Reconciliation-time records should also store:
+
+- OpenRouter generation id
+- native token fields when available
+- `total_cost`
+- `cache_discount`
+- `upstream_inference_cost`
+- `is_byok`
+
+### Gemini / Vertex
+
+Use official Gemini or Vertex usage fields where available.
+
+If cached content tokens are exposed:
+
+- map them to `cache_read_tokens`
+
+If a route exposes no cache creation metric:
+
+- store `cache_write_tokens = 0`
+- preserve the raw usage payload for later extension
+
+### DeepSeek And Other Direct Providers
+
+Normalize only the fields that are officially exposed.
+
+If a provider does not expose cache buckets:
+
+- do not infer them unless the provider explicitly documents how to derive them
+
+### Subscription / Included-Cost Routes
+
+These still use the canonical usage model.
+
+Tokens are tracked normally. Cost depends on billing mode, not on whether usage exists.
+
+## Billing Route Model
+
+Hermes must stop keying pricing solely by `model`.
+
+Introduce a billing route descriptor:
+
+```python
+@dataclass
+class BillingRoute:
+    provider: str
+    base_url: str | None
+    model: str
+    billing_mode: str
+    organization_hint: str | None = None
+```
+
+`billing_mode` values:
+
+- `official_cost_api`
+- `official_generation_api`
+- `official_models_api`
+- `official_docs_snapshot`
+- `subscription_included`
+- `user_override`
+- `custom_contract`
+- `unknown`
+
+Examples:
+
+- OpenAI direct API with Costs API access: `official_cost_api`
+- Anthropic direct API with Usage & Cost API access: `official_cost_api`
+- OpenRouter request before reconciliation: `official_models_api`
+- OpenRouter request after generation lookup: `official_generation_api`
+- GitHub Copilot style subscription route: `subscription_included`
+- local OpenAI-compatible server: `unknown`
+- enterprise contract with configured rates: `custom_contract`
+
+## Cost Status Model
+
+Every displayed cost should have:
+
+```python
+@dataclass
+class CostResult:
+    amount_usd: Decimal | None
+    status: Literal["actual", "estimated", "included", "unknown"]
+    source: Literal[
+        "provider_cost_api",
+        "provider_generation_api",
+        "provider_models_api",
+        "official_docs_snapshot",
+        "user_override",
+        "custom_contract",
+        "none",
+    ]
+    label: str
+    fetched_at: datetime | None
+    pricing_version: str | None
+    notes: list[str]
+```
+
+Presentation rules:
+
+- `actual`: show dollar amount as final
+- `estimated`: show dollar amount with estimate labeling
+- `included`: show `included` or `$0.00 (included)` depending on UX choice
+- `unknown`: show `n/a`
+
+## Official Source Hierarchy
+
+Resolve cost using this order:
+
+1. Request-level or account-level official billed cost
+2. Official machine-readable model pricing
+3. Official docs snapshot
+4. User override or custom contract
+5. Unknown
+
+The system must never skip to a lower level if a higher-confidence source exists for the current billing route.
+
+## Provider-Specific Truth Rules
+
+### OpenAI Direct
+
+Preferred truth:
+
+1. Costs API for reconciled spend
+2. Official pricing page for live estimate
+
+### Anthropic Direct
+
+Preferred truth:
+
+1. Usage & Cost API for reconciled spend
+2. Official pricing docs for live estimate
+
+### OpenRouter
+
+Preferred truth:
+
+1. `GET /api/v1/generation` for reconciled `total_cost`
+2. `GET /api/v1/models` pricing for live estimate
+
+Do not use underlying provider public pricing as the source of truth for OpenRouter billing.
+
+### Gemini / Vertex
+
+Preferred truth:
+
+1. official billing export or billing API for reconciled spend when available for the route
+2. official pricing docs for estimate
+
+### DeepSeek
+
+Preferred truth:
+
+1. official machine-readable cost source if available in the future
+2. official pricing docs snapshot today
+
+### Subscription-Included Routes
+
+Preferred truth:
+
+1. explicit route config marking the model as included in subscription
+
+These should display `included`, not an API list-price estimate.
+
+### Custom Endpoint / Local Model
+
+Preferred truth:
+
+1. user override
+2. custom contract config
+3. unknown
+
+These should default to `unknown`.
+
+## Pricing Catalog
+
+Replace the current `MODEL_PRICING` dict with a richer pricing catalog.
+
+Suggested record:
+
+```python
+@dataclass
+class PricingEntry:
+    provider: str
+    route_pattern: str
+    model_pattern: str
+
+    input_cost_per_million: Decimal | None = None
+    output_cost_per_million: Decimal | None = None
+    cache_read_cost_per_million: Decimal | None = None
+    cache_write_cost_per_million: Decimal | None = None
+    request_cost: Decimal | None = None
+    image_cost: Decimal | None = None
+
+    source: str = "official_docs_snapshot"
+    source_url: str | None = None
+    fetched_at: datetime | None = None
+    pricing_version: str | None = None
+```
+
+The catalog should be route-aware:
+
+- `openai:gpt-5`
+- `anthropic:claude-opus-4-6`
+- `openrouter:anthropic/claude-opus-4.6`
+- `copilot:gpt-4o`
+
+This avoids conflating direct-provider billing with aggregator billing.
+
+## Pricing Sync Architecture
+
+Introduce a pricing sync subsystem instead of manually maintaining a single hardcoded table.
+
+Suggested modules:
+
+- `agent/pricing/catalog.py`
+- `agent/pricing/sources.py`
+- `agent/pricing/sync.py`
+- `agent/pricing/reconcile.py`
+- `agent/pricing/types.py`
+
+### Sync Sources
+
+- OpenRouter models API
+- official provider docs snapshots where no API exists
+- user overrides from config
+
+### Sync Output
+
+Cache pricing entries locally with:
+
+- source URL
+- fetch timestamp
+- version/hash
+- confidence/source type
+
+### Sync Frequency
+
+- startup warm cache
+- background refresh every 6 to 24 hours depending on source
+- manual `hermes pricing sync`
+
+## Reconciliation Architecture
+
+Live requests may produce only an estimate initially. Hermes should reconcile them later when a provider exposes actual billed cost.
+
+Suggested flow:
+
+1. Agent call completes.
+2. Hermes stores canonical usage plus reconciliation ids.
+3. Hermes computes an immediate estimate if a pricing source exists.
+4. A reconciliation worker fetches actual cost when supported.
+5. Session and message records are updated with `actual` cost.
+
+This can run:
+
+- inline for cheap lookups
+- asynchronously for delayed provider accounting
+
+## Persistence Changes
+
+Session storage should stop storing only aggregate prompt/completion totals.
+
+Add fields for both usage and cost certainty:
+
+- `input_tokens`
+- `output_tokens`
+- `cache_read_tokens`
+- `cache_write_tokens`
+- `reasoning_tokens`
+- `estimated_cost_usd`
+- `actual_cost_usd`
+- `cost_status`
+- `cost_source`
+- `pricing_version`
+- `billing_provider`
+- `billing_mode`
+
+If schema expansion is too large for one PR, add a new pricing events table:
+
+```text
+session_cost_events
+  id
+  session_id
+  request_id
+  provider
+  model
+  billing_mode
+  input_tokens
+  output_tokens
+  cache_read_tokens
+  cache_write_tokens
+  estimated_cost_usd
+  actual_cost_usd
+  cost_status
+  cost_source
+  pricing_version
+  created_at
+  updated_at
+```
+
+## Hermes Touchpoints
+
+### `run_agent.py`
+
+Current responsibility:
+
+- parse raw provider usage
+- update session token counters
+
+New responsibility:
+
+- build `CanonicalUsage`
+- update canonical counters
+- store reconciliation ids
+- emit usage event to pricing subsystem
+
+### `agent/usage_pricing.py`
+
+Current responsibility:
+
+- static lookup table
+- direct cost arithmetic
+
+New responsibility:
+
+- move or replace with pricing catalog facade
+- no fuzzy model-family heuristics
+- no direct pricing without billing-route context
+
+### `cli.py`
+
+Current responsibility:
+
+- compute session cost directly from prompt/completion totals
+
+New responsibility:
+
+- display `CostResult`
+- show status badges:
+  - `actual`
+  - `estimated`
+  - `included`
+  - `n/a`
+
+### `agent/insights.py`
+
+Current responsibility:
+
+- recompute historical estimates from static pricing
+
+New responsibility:
+
+- aggregate stored pricing events
+- prefer actual cost over estimate
+- surface estimates only when reconciliation is unavailable
+
+## UX Rules
+
+### Status Bar
+
+Show one of:
+
+- `$1.42`
+- `~$1.42`
+- `included`
+- `cost n/a`
+
+Where:
+
+- `$1.42` means `actual`
+- `~$1.42` means `estimated`
+- `included` means subscription-backed or explicitly zero-cost route
+- `cost n/a` means unknown
+
+### `/usage`
+
+Show:
+
+- token buckets
+- estimated cost
+- actual cost if available
+- cost status
+- pricing source
+
+### `/insights`
+
+Aggregate:
+
+- actual cost totals
+- estimated-only totals
+- unknown-cost sessions count
+- included-cost sessions count
+
+## Config And Overrides
+
+Add user-configurable pricing overrides in config:
+
+```yaml
+pricing:
+  mode: hybrid
+  sync_on_startup: true
+  sync_interval_hours: 12
+  overrides:
+    - provider: openrouter
+      model: anthropic/claude-opus-4.6
+      billing_mode: custom_contract
+      input_cost_per_million: 4.25
+      output_cost_per_million: 22.0
+      cache_read_cost_per_million: 0.5
+      cache_write_cost_per_million: 6.0
+  included_routes:
+    - provider: copilot
+      model: "*"
+    - provider: codex-subscription
+      model: "*"
+```
+
+Overrides must win over catalog defaults for the matching billing route.
+
+## Rollout Plan
+
+### Phase 1
+
+- add canonical usage model
+- split cache token buckets in `run_agent.py`
+- stop pricing cache-inflated prompt totals
+- preserve current UI with improved backend math
+
+### Phase 2
+
+- add route-aware pricing catalog
+- integrate OpenRouter models API sync
+- add `estimated` vs `included` vs `unknown`
+
+### Phase 3
+
+- add reconciliation for OpenRouter generation cost
+- add actual cost persistence
+- update `/insights` to prefer actual cost
+
+### Phase 4
+
+- add direct OpenAI and Anthropic reconciliation paths
+- add user overrides and contract pricing
+- add pricing sync CLI command
+
+## Testing Strategy
+
+Add tests for:
+
+- OpenAI cached token subtraction
+- Anthropic cache read/write separation
+- OpenRouter estimated vs actual reconciliation
+- subscription-backed models showing `included`
+- custom endpoints showing `n/a`
+- override precedence
+- stale catalog fallback behavior
+
+Current tests that assume heuristic pricing should be replaced with route-aware expectations.
+
+## Non-Goals
+
+- exact enterprise billing reconstruction without an official source or user override
+- backfilling perfect historical cost for old sessions that lack cache bucket data
+- scraping arbitrary provider web pages at request time
+
+## Recommendation
+
+Do not expand the existing `MODEL_PRICING` dict.
+
+That path cannot satisfy the product requirement. Hermes should instead migrate to:
+
+- canonical usage normalization
+- route-aware pricing sources
+- estimate-then-reconcile cost lifecycle
+- explicit certainty states in the UI
+
+This is the minimum architecture that makes the statement "Hermes pricing is backed by official sources where possible, and otherwise clearly labeled" defensible.

environments/README.md

@@ -101,21 +101,11 @@ Available methods:
 
 ### Patches (`patches.py`)
 
-**Problem**: Some hermes-agent tools use `asyncio.run()` internally (e.g., mini-swe-agent's Modal backend via SWE-ReX). This crashes when called from inside Atropos's event loop because `asyncio.run()` cannot be nested.
+**Problem**: Some hermes-agent tools use `asyncio.run()` internally (e.g., the Modal backend). This crashes when called from inside Atropos's event loop because `asyncio.run()` cannot be nested.
 
-**Solution**: `patches.py` monkey-patches `SwerexModalEnvironment` to use a dedicated background thread (`_AsyncWorker`) with its own event loop. The calling code sees the same sync interface, but internally the async work happens on a separate thread that doesn't conflict with Atropos's loop.
+**Solution**: `ModalEnvironment` uses a dedicated `_AsyncWorker` background thread with its own event loop. The calling code sees a sync interface, but internally all async Modal SDK calls happen on the worker thread so they don't conflict with Atropos's loop. This is built directly into `tools/environments/modal.py` — no monkey-patching required.
 
-What gets patched:
-- `SwerexModalEnvironment.__init__` -- creates Modal deployment on a background thread
-- `SwerexModalEnvironment.execute` -- runs commands on the same background thread
-- `SwerexModalEnvironment.stop` -- stops deployment on the background thread
-
-The patches are:
-- **Idempotent** -- calling `apply_patches()` multiple times is safe
-- **Transparent** -- same interface and behavior, only the internal async execution changes
-- **Universal** -- works identically in normal CLI use (no running event loop)
-
-Applied automatically at import time by `hermes_base_env.py`.
+`patches.py` is now a no-op (kept for backward compatibility with imports).
 
 ### Tool Call Parsers (`tool_call_parsers/`)
 

environments/agent_loop.py

@@ -23,7 +23,7 @@ from typing import Any, Dict, List, Optional, Set
 from model_tools import handle_function_call
 
 # Thread pool for running sync tool calls that internally use asyncio.run()
-# (e.g., mini-swe-agent's modal/docker/daytona backends). Running them in a separate
+# (e.g., the Modal/Docker/Daytona terminal backends). Running them in a separate
 # thread gives them a clean event loop so they don't deadlock inside Atropos's loop.
 # Size must be large enough for concurrent eval tasks (e.g., 89 TB2 tasks all
 # making tool calls). Too small = thread pool starvation, tasks queue for minutes.
@@ -346,78 +346,89 @@ class HermesAgentLoop:
                             tool_name, turn + 1,
                         )
                     else:
-                        # Parse arguments and dispatch
+                        # Parse arguments
                         try:
                             args = json.loads(tool_args_raw)
-                        except json.JSONDecodeError:
-                            args = {}
+                        except json.JSONDecodeError as e:
+                            args = None
+                            tool_result = json.dumps(
+                                {"error": f"Invalid JSON in tool arguments: {e}. Please retry with valid JSON."}
+                            )
+                            tool_errors.append(ToolError(
+                                turn=turn + 1, tool_name=tool_name,
+                                arguments=tool_args_raw[:200],
+                                error=f"Invalid JSON: {e}",
+                                tool_result=tool_result,
+                            ))
                             logger.warning(
                                 "Invalid JSON in tool call arguments for '%s': %s",
                                 tool_name, tool_args_raw[:200],
                             )
 
-                        try:
-                            if tool_name == "terminal":
-                                backend = os.getenv("TERMINAL_ENV", "local")
-                                cmd_preview = args.get("command", "")[:80]
-                                logger.info(
-                                    "[%s] $ %s", self.task_id[:8], cmd_preview,
-                                )
+                        # Dispatch tool only if arguments parsed successfully
+                        if args is not None:
+                            try:
+                                if tool_name == "terminal":
+                                    backend = os.getenv("TERMINAL_ENV", "local")
+                                    cmd_preview = args.get("command", "")[:80]
+                                    logger.info(
+                                        "[%s] $ %s", self.task_id[:8], cmd_preview,
+                                    )
 
-                            tool_submit_time = _time.monotonic()
+                                tool_submit_time = _time.monotonic()
 
-                            # Todo tool -- handle locally (needs per-loop TodoStore)
-                            if tool_name == "todo":
-                                tool_result = _todo_tool(
-                                    todos=args.get("todos"),
-                                    merge=args.get("merge", False),
-                                    store=_todo_store,
-                                )
-                                tool_elapsed = _time.monotonic() - tool_submit_time
-                            elif tool_name == "memory":
-                                tool_result = json.dumps({"error": "Memory is not available in RL environments."})
-                                tool_elapsed = _time.monotonic() - tool_submit_time
-                            elif tool_name == "session_search":
-                                tool_result = json.dumps({"error": "Session search is not available in RL environments."})
-                                tool_elapsed = _time.monotonic() - tool_submit_time
-                            else:
-                                # Run tool calls in a thread pool so backends that
-                                # use asyncio.run() internally (modal, docker, daytona) get
-                                # a clean event loop instead of deadlocking.
-                                loop = asyncio.get_event_loop()
-                                # Capture current tool_name/args for the lambda
-                                _tn, _ta, _tid = tool_name, args, self.task_id
-                                tool_result = await loop.run_in_executor(
-                                    _tool_executor,
-                                    lambda: handle_function_call(
-                                        _tn, _ta, task_id=_tid,
-                                        user_task=_user_task,
-                                    ),
-                                )
-                                tool_elapsed = _time.monotonic() - tool_submit_time
+                                # Todo tool -- handle locally (needs per-loop TodoStore)
+                                if tool_name == "todo":
+                                    tool_result = _todo_tool(
+                                        todos=args.get("todos"),
+                                        merge=args.get("merge", False),
+                                        store=_todo_store,
+                                    )
+                                    tool_elapsed = _time.monotonic() - tool_submit_time
+                                elif tool_name == "memory":
+                                    tool_result = json.dumps({"error": "Memory is not available in RL environments."})
+                                    tool_elapsed = _time.monotonic() - tool_submit_time
+                                elif tool_name == "session_search":
+                                    tool_result = json.dumps({"error": "Session search is not available in RL environments."})
+                                    tool_elapsed = _time.monotonic() - tool_submit_time
+                                else:
+                                    # Run tool calls in a thread pool so backends that
+                                    # use asyncio.run() internally (modal, docker, daytona) get
+                                    # a clean event loop instead of deadlocking.
+                                    loop = asyncio.get_event_loop()
+                                    # Capture current tool_name/args for the lambda
+                                    _tn, _ta, _tid = tool_name, args, self.task_id
+                                    tool_result = await loop.run_in_executor(
+                                        _tool_executor,
+                                        lambda: handle_function_call(
+                                            _tn, _ta, task_id=_tid,
+                                            user_task=_user_task,
+                                        ),
+                                    )
+                                    tool_elapsed = _time.monotonic() - tool_submit_time
 
-                            # Log slow tools and thread pool stats for debugging
-                            pool_active = _tool_executor._work_queue.qsize()
-                            if tool_elapsed > 30:
-                                logger.warning(
-                                    "[%s] turn %d: %s took %.1fs (pool queue=%d)",
-                                    self.task_id[:8], turn + 1, tool_name,
-                                    tool_elapsed, pool_active,
+                                # Log slow tools and thread pool stats for debugging
+                                pool_active = _tool_executor._work_queue.qsize()
+                                if tool_elapsed > 30:
+                                    logger.warning(
+                                        "[%s] turn %d: %s took %.1fs (pool queue=%d)",
+                                        self.task_id[:8], turn + 1, tool_name,
+                                        tool_elapsed, pool_active,
+                                    )
+                            except Exception as e:
+                                tool_result = json.dumps(
+                                    {"error": f"Tool execution failed: {type(e).__name__}: {str(e)}"}
+                                )
+                                tool_errors.append(ToolError(
+                                    turn=turn + 1, tool_name=tool_name,
+                                    arguments=tool_args_raw[:200],
+                                    error=f"{type(e).__name__}: {str(e)}",
+                                    tool_result=tool_result,
+                                ))
+                                logger.error(
+                                    "Tool '%s' execution failed on turn %d: %s",
+                                    tool_name, turn + 1, e,
                                 )
-                        except Exception as e:
-                            tool_result = json.dumps(
-                                {"error": f"Tool execution failed: {type(e).__name__}: {str(e)}"}
-                            )
-                            tool_errors.append(ToolError(
-                                turn=turn + 1, tool_name=tool_name,
-                                arguments=tool_args_raw[:200],
-                                error=f"{type(e).__name__}: {str(e)}",
-                                tool_result=tool_result,
-                            ))
-                            logger.error(
-                                "Tool '%s' execution failed on turn %d: %s",
-                                tool_name, turn + 1, e,
-                            )
 
                         # Also check if the tool returned an error in its JSON result
                         try:

environments/benchmarks/terminalbench_2/terminalbench2_env.py

@@ -209,7 +209,7 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
 
             # Agent settings -- TB2 tasks are complex, need many turns
             max_agent_turns=60,
-            max_token_length=***
+            max_token_length=16000,
             agent_temperature=0.6,
             system_prompt=None,
 
@@ -233,7 +233,7 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
             steps_per_eval=1,
             total_steps=1,
 
-            tokenizer_name="NousRe...1-8B",
+            tokenizer_name="NousResearch/Hermes-3-Llama-3.1-8B",
             use_wandb=True,
             wandb_name="terminal-bench-2",
             ensure_scores_are_not_same=False,  # Binary rewards may all be 0 or 1
@@ -245,7 +245,7 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
                 base_url="https://openrouter.ai/api/v1",
                 model_name="anthropic/claude-sonnet-4",
                 server_type="openai",
-                api_key=os.get...EY", ""),
+                api_key=os.getenv("OPENROUTER_API_KEY", ""),
                 health_check=False,
             )
         ]
@@ -513,3 +513,446 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
                 reward = 0.0
             else:
                 # Run tests in a thread so the blocking ctx.terminal() calls
+                # don't freeze the entire event loop (which would stall all
+                # other tasks, tqdm updates, and timeout timers).
+                ctx = ToolContext(task_id)
+                try:
+                    loop = asyncio.get_event_loop()
+                    reward = await loop.run_in_executor(
+                        None,  # default thread pool
+                        self._run_tests, eval_item, ctx, task_name,
+                    )
+                except Exception as e:
+                    logger.error("Task %s: test verification failed: %s", task_name, e)
+                    reward = 0.0
+                finally:
+                    ctx.cleanup()
+
+            passed = reward == 1.0
+            status = "PASS" if passed else "FAIL"
+            elapsed = time.time() - task_start
+            tqdm.write(f"  [{status}] {task_name} (turns={result.turns_used}, {elapsed:.0f}s)")
+            logger.info(
+                "Task %s: reward=%.1f, turns=%d, finished=%s",
+                task_name, reward, result.turns_used, result.finished_naturally,
+            )
+
+            out = {
+                "passed": passed,
+                "reward": reward,
+                "task_name": task_name,
+                "category": category,
+                "turns_used": result.turns_used,
+                "finished_naturally": result.finished_naturally,
+                "messages": result.messages,
+            }
+            self._save_result(out)
+            return out
+
+        except Exception as e:
+            elapsed = time.time() - task_start
+            logger.error("Task %s: rollout failed: %s", task_name, e, exc_info=True)
+            tqdm.write(f"  [ERROR] {task_name}: {e} ({elapsed:.0f}s)")
+            out = {
+                "passed": False, "reward": 0.0,
+                "task_name": task_name, "category": category,
+                "error": str(e),
+            }
+            self._save_result(out)
+            return out
+
+        finally:
+            # --- Cleanup: clear overrides, sandbox, and temp files ---
+            clear_task_env_overrides(task_id)
+            try:
+                cleanup_vm(task_id)
+            except Exception as e:
+                logger.debug("VM cleanup for %s: %s", task_id[:8], e)
+            if task_dir and task_dir.exists():
+                shutil.rmtree(task_dir, ignore_errors=True)
+
+    def _run_tests(
+        self, item: Dict[str, Any], ctx: ToolContext, task_name: str
+    ) -> float:
+        """
+        Upload and execute the test suite in the agent's sandbox, then
+        download the verifier output locally to read the reward.
+
+        Follows Harbor's verification pattern:
+        1. Upload tests/ directory into the sandbox
+        2. Execute test.sh inside the sandbox
+        3. Download /logs/verifier/ directory to a local temp dir
+        4. Read reward.txt locally with native Python I/O
+
+        Downloading locally avoids issues with the file_read tool on
+        the Modal VM and matches how Harbor handles verification.
+
+        TB2 test scripts (test.sh) typically:
+        1. Install pytest via uv/pip
+        2. Run pytest against the test files in /tests/
+        3. Write results to /logs/verifier/reward.txt
+
+        Args:
+            item: The TB2 task dict (contains tests_tar, test_sh)
+            ctx: ToolContext scoped to this task's sandbox
+            task_name: For logging
+
+        Returns:
+            1.0 if tests pass, 0.0 otherwise
+        """
+        tests_tar = item.get("tests_tar", "")
+        test_sh = item.get("test_sh", "")
+
+        if not test_sh:
+            logger.warning("Task %s: no test_sh content, reward=0", task_name)
+            return 0.0
+
+        # Create required directories in the sandbox
+        ctx.terminal("mkdir -p /tests /logs/verifier")
+
+        # Upload test files into the sandbox (binary-safe via base64)
+        if tests_tar:
+            tests_temp = Path(tempfile.mkdtemp(prefix=f"tb2-tests-{task_name}-"))
+            try:
+                _extract_base64_tar(tests_tar, tests_temp)
+                ctx.upload_dir(str(tests_temp), "/tests")
+            except Exception as e:
+                logger.warning("Task %s: failed to upload test files: %s", task_name, e)
+            finally:
+                shutil.rmtree(tests_temp, ignore_errors=True)
+
+        # Write the test runner script (test.sh)
+        ctx.write_file("/tests/test.sh", test_sh)
+        ctx.terminal("chmod +x /tests/test.sh")
+
+        # Execute the test suite
+        logger.info(
+            "Task %s: running test suite (timeout=%ds)",
+            task_name, self.config.test_timeout,
+        )
+        test_result = ctx.terminal(
+            "bash /tests/test.sh",
+            timeout=self.config.test_timeout,
+        )
+
+        exit_code = test_result.get("exit_code", -1)
+        output = test_result.get("output", "")
+
+        # Download the verifier output directory locally, then read reward.txt
+        # with native Python I/O. This avoids issues with file_read on the
+        # Modal VM and matches Harbor's verification pattern.
+        reward = 0.0
+        local_verifier_dir = Path(tempfile.mkdtemp(prefix=f"tb2-verifier-{task_name}-"))
+        try:
+            ctx.download_dir("/logs/verifier", str(local_verifier_dir))
+
+            reward_file = local_verifier_dir / "reward.txt"
+            if reward_file.exists() and reward_file.stat().st_size > 0:
+                content = reward_file.read_text().strip()
+                if content == "1":
+                    reward = 1.0
+                elif content == "0":
+                    reward = 0.0
+                else:
+                    # Unexpected content -- try parsing as float
+                    try:
+                        reward = float(content)
+                    except (ValueError, TypeError):
+                        logger.warning(
+                            "Task %s: reward.txt content unexpected (%r), "
+                            "falling back to exit_code=%d",
+                            task_name, content, exit_code,
+                        )
+                        reward = 1.0 if exit_code == 0 else 0.0
+            else:
+                # reward.txt not written -- fall back to exit code
+                logger.warning(
+                    "Task %s: reward.txt not found after download, "
+                    "falling back to exit_code=%d",
+                    task_name, exit_code,
+                )
+                reward = 1.0 if exit_code == 0 else 0.0
+        except Exception as e:
+            logger.warning(
+                "Task %s: failed to download verifier dir: %s, "
+                "falling back to exit_code=%d",
+                task_name, e, exit_code,
+            )
+            reward = 1.0 if exit_code == 0 else 0.0
+        finally:
+            shutil.rmtree(local_verifier_dir, ignore_errors=True)
+
+        # Log test output for debugging failures
+        if reward == 0.0:
+            output_preview = output[-500:] if output else "(no output)"
+            logger.info(
+                "Task %s: FAIL (exit_code=%d)\n%s",
+                task_name, exit_code, output_preview,
+            )
+
+        return reward
+
+    # =========================================================================
+    # Evaluate -- main entry point for the eval subcommand
+    # =========================================================================
+
+    async def _eval_with_timeout(self, item: Dict[str, Any]) -> Dict:
+        """
+        Wrap rollout_and_score_eval with a per-task wall-clock timeout.
+
+        If the task exceeds task_timeout seconds, it's automatically scored
+        as FAIL. This prevents any single task from hanging indefinitely.
+        """
+        task_name = item.get("task_name", "unknown")
+        category = item.get("category", "unknown")
+        try:
+            return await asyncio.wait_for(
+                self.rollout_and_score_eval(item),
+                timeout=self.config.task_timeout,
+            )
+        except asyncio.TimeoutError:
+            from tqdm import tqdm
+            elapsed = self.config.task_timeout
+            tqdm.write(f"  [TIMEOUT] {task_name} (exceeded {elapsed}s wall-clock limit)")
+            logger.error("Task %s: wall-clock timeout after %ds", task_name, elapsed)
+            out = {
+                "passed": False, "reward": 0.0,
+                "task_name": task_name, "category": category,
+                "error": f"timeout ({elapsed}s)",
+            }
+            self._save_result(out)
+            return out
+
+    async def evaluate(self, *args, **kwargs) -> None:
+        """
+        Run Terminal-Bench 2.0 evaluation over all tasks.
+
+        This is the main entry point when invoked via:
+            python environments/terminalbench2_env.py evaluate
+
+        Runs all tasks through rollout_and_score_eval() via asyncio.gather()
+        (same pattern as GPQA and other Atropos eval envs). Each task is
+        wrapped with a wall-clock timeout so hung tasks auto-fail.
+
+        Suppresses noisy Modal/terminal output (HERMES_QUIET) so the tqdm
+        bar stays visible.
+        """
+        start_time = time.time()
+
+        # Route all logging through tqdm.write() so the progress bar stays
+        # pinned at the bottom while log lines scroll above it.
+        from tqdm import tqdm
+
+        class _TqdmHandler(logging.Handler):
+            def emit(self, record):
+                try:
+                    tqdm.write(self.format(record))
+                except Exception:
+                    self.handleError(record)
+
+        handler = _TqdmHandler()
+        handler.setFormatter(logging.Formatter(
+            "%(asctime)s [%(name)s] %(levelname)s: %(message)s",
+            datefmt="%H:%M:%S",
+        ))
+        root = logging.getLogger()
+        root.handlers = [handler]  # Replace any existing handlers
+        root.setLevel(logging.INFO)
+
+        # Silence noisy third-party loggers that flood the output
+        logging.getLogger("httpx").setLevel(logging.WARNING)      # Every HTTP request
+        logging.getLogger("openai").setLevel(logging.WARNING)     # OpenAI client retries
+        logging.getLogger("rex-deploy").setLevel(logging.WARNING) # Swerex deployment
+        logging.getLogger("rex_image_builder").setLevel(logging.WARNING)  # Image builds
+
+        print(f"\n{'='*60}")
+        print("Starting Terminal-Bench 2.0 Evaluation")
+        print(f"{'='*60}")
+        print(f"  Dataset: {self.config.dataset_name}")
+        print(f"  Total tasks: {len(self.all_eval_items)}")
+        print(f"  Max agent turns: {self.config.max_agent_turns}")
+        print(f"  Task timeout: {self.config.task_timeout}s")
+        print(f"  Terminal backend: {self.config.terminal_backend}")
+        print(f"  Tool thread pool: {self.config.tool_pool_size}")
+        print(f"  Terminal timeout: {self.config.terminal_timeout}s/cmd")
+        print(f"  Terminal lifetime: {self.config.terminal_lifetime}s (auto: task_timeout + 120)")
+        print(f"  Max concurrent tasks: {self.config.max_concurrent_tasks}")
+        print(f"{'='*60}\n")
+
+        # Semaphore to limit concurrent Modal sandbox creations.
+        # Without this, all 86 tasks fire simultaneously, each creating a Modal
+        # sandbox via asyncio.run() inside a thread pool worker. Modal's blocking
+        # calls (App.lookup, etc.) deadlock when too many are created at once.
+        semaphore = asyncio.Semaphore(self.config.max_concurrent_tasks)
+
+        async def _eval_with_semaphore(item):
+            async with semaphore:
+                return await self._eval_with_timeout(item)
+
+        # Fire all tasks with wall-clock timeout, track live accuracy on the bar
+        total_tasks = len(self.all_eval_items)
+        eval_tasks = [
+            asyncio.ensure_future(_eval_with_semaphore(item))
+            for item in self.all_eval_items
+        ]
+
+        results = []
+        passed_count = 0
+        pbar = tqdm(total=total_tasks, desc="Evaluating TB2", dynamic_ncols=True)
+        try:
+            for coro in asyncio.as_completed(eval_tasks):
+                result = await coro
+                results.append(result)
+                if result and result.get("passed"):
+                    passed_count += 1
+                done = len(results)
+                pct = (passed_count / done * 100) if done else 0
+                pbar.set_postfix_str(f"pass={passed_count}/{done} ({pct:.1f}%)")
+                pbar.update(1)
+        except (KeyboardInterrupt, asyncio.CancelledError):
+            pbar.close()
+            print(f"\n\nInterrupted! Cleaning up {len(eval_tasks)} tasks...")
+            # Cancel all pending tasks
+            for task in eval_tasks:
+                task.cancel()
+            # Let cancellations propagate (finally blocks run cleanup_vm)
+            await asyncio.gather(*eval_tasks, return_exceptions=True)
+            # Belt-and-suspenders: clean up any remaining sandboxes
+            from tools.terminal_tool import cleanup_all_environments
+            cleanup_all_environments()
+            print("All sandboxes cleaned up.")
+            return
+        finally:
+            pbar.close()
+
+        end_time = time.time()
+
+        # Filter out None results (shouldn't happen, but be safe)
+        valid_results = [r for r in results if r is not None]
+
+        if not valid_results:
+            print("Warning: No valid evaluation results obtained")
+            return
+
+        # ---- Compute metrics ----
+        total = len(valid_results)
+        passed = sum(1 for r in valid_results if r.get("passed"))
+        overall_pass_rate = passed / total if total > 0 else 0.0
+
+        # Per-category breakdown
+        cat_results: Dict[str, List[Dict]] = defaultdict(list)
+        for r in valid_results:
+            cat_results[r.get("category", "unknown")].append(r)
+
+        # Build metrics dict
+        eval_metrics = {
+            "eval/pass_rate": overall_pass_rate,
+            "eval/total_tasks": total,
+            "eval/passed_tasks": passed,
+            "eval/evaluation_time_seconds": end_time - start_time,
+        }
+
+        # Per-category metrics
+        for category, cat_items in sorted(cat_results.items()):
+            cat_passed = sum(1 for r in cat_items if r.get("passed"))
+            cat_total = len(cat_items)
+            cat_pass_rate = cat_passed / cat_total if cat_total > 0 else 0.0
+            cat_key = category.replace(" ", "_").replace("-", "_").lower()
+            eval_metrics[f"eval/pass_rate_{cat_key}"] = cat_pass_rate
+
+        # Store metrics for wandb_log
+        self.eval_metrics = [(k, v) for k, v in eval_metrics.items()]
+
+        # ---- Print summary ----
+        print(f"\n{'='*60}")
+        print("Terminal-Bench 2.0 Evaluation Results")
+        print(f"{'='*60}")
+        print(f"Overall Pass Rate: {overall_pass_rate:.4f} ({passed}/{total})")
+        print(f"Evaluation Time: {end_time - start_time:.1f} seconds")
+
+        print("\nCategory Breakdown:")
+        for category, cat_items in sorted(cat_results.items()):
+            cat_passed = sum(1 for r in cat_items if r.get("passed"))
+            cat_total = len(cat_items)
+            cat_rate = cat_passed / cat_total if cat_total > 0 else 0.0
+            print(f"  {category}: {cat_rate:.1%} ({cat_passed}/{cat_total})")
+
+        # Print individual task results
+        print("\nTask Results:")
+        for r in sorted(valid_results, key=lambda x: x.get("task_name", "")):
+            status = "PASS" if r.get("passed") else "FAIL"
+            turns = r.get("turns_used", "?")
+            error = r.get("error", "")
+            extra = f" (error: {error})" if error else ""
+            print(f"  [{status}] {r['task_name']} (turns={turns}){extra}")
+
+        print(f"{'='*60}\n")
+
+        # Build sample records for evaluate_log (includes full conversations)
+        samples = [
+            {
+                "task_name": r.get("task_name"),
+                "category": r.get("category"),
+                "passed": r.get("passed"),
+                "reward": r.get("reward"),
+                "turns_used": r.get("turns_used"),
+                "error": r.get("error"),
+                "messages": r.get("messages"),
+            }
+            for r in valid_results
+        ]
+
+        # Log evaluation results
+        try:
+            await self.evaluate_log(
+                metrics=eval_metrics,
+                samples=samples,
+                start_time=start_time,
+                end_time=end_time,
+                generation_parameters={
+                    "temperature": self.config.agent_temperature,
+                    "max_tokens": self.config.max_token_length,
+                    "max_agent_turns": self.config.max_agent_turns,
+                    "terminal_backend": self.config.terminal_backend,
+                },
+            )
+        except Exception as e:
+            print(f"Error logging evaluation results: {e}")
+
+        # Close streaming file
+        if hasattr(self, "_streaming_file") and not self._streaming_file.closed:
+            self._streaming_file.close()
+            print(f"  Live results saved to: {self._streaming_path}")
+
+        # Kill all remaining sandboxes. Timed-out tasks leave orphaned thread
+        # pool workers still executing commands -- cleanup_all stops them.
+        from tools.terminal_tool import cleanup_all_environments
+        print("\nCleaning up all sandboxes...")
+        cleanup_all_environments()
+
+        # Shut down the tool thread pool so orphaned workers from timed-out
+        # tasks are killed immediately instead of retrying against dead
+        # sandboxes and spamming the console with TimeoutError warnings.
+        from environments.agent_loop import _tool_executor
+        _tool_executor.shutdown(wait=False, cancel_futures=True)
+        print("Done.")
+
+    # =========================================================================
+    # Wandb logging
+    # =========================================================================
+
+    async def wandb_log(self, wandb_metrics: Optional[Dict] = None):
+        """Log TB2-specific metrics to wandb."""
+        if wandb_metrics is None:
+            wandb_metrics = {}
+
+        # Add stored eval metrics
+        for metric_name, metric_value in self.eval_metrics:
+            wandb_metrics[metric_name] = metric_value
+        self.eval_metrics = []
+
+        await super().wandb_log(wandb_metrics)
+
+
+if __name__ == "__main__":
+    TerminalBench2EvalEnv.cli()

environments/patches.py

@@ -2,203 +2,34 @@
 Monkey patches for making hermes-agent tools work inside async frameworks (Atropos).
 
 Problem:
-    Some tools use asyncio.run() internally (e.g., mini-swe-agent's Modal backend,
+    Some tools use asyncio.run() internally (e.g., Modal backend via SWE-ReX,
     web_extract). This crashes when called from inside Atropos's event loop because
     asyncio.run() can't be nested.
 
 Solution:
-    Replace the problematic methods with versions that use a dedicated background
-    thread with its own event loop. The calling code sees the same sync interface --
-    call a function, get a result -- but internally the async work happens on a
-    separate thread that doesn't conflict with Atropos's loop.
+    The Modal environment (tools/environments/modal.py) now uses a dedicated
+    _AsyncWorker thread internally, making it safe for both CLI and Atropos use.
+    No monkey-patching is required.
 
-    These patches are safe for normal CLI use too: when there's no running event
-    loop, the behavior is identical (the background thread approach works regardless).
-
-What gets patched:
-    - SwerexModalEnvironment.__init__ -- creates Modal deployment on a background thread
-    - SwerexModalEnvironment.execute -- runs commands on the same background thread
-    - SwerexModalEnvironment.stop -- stops deployment on the background thread
+    This module is kept for backward compatibility. apply_patches() is a no-op.
 
 Usage:
     Call apply_patches() once at import time (done automatically by hermes_base_env.py).
-    This is idempotent -- calling it multiple times is safe.
+    This is idempotent and safe to call multiple times.
 """
 
-import asyncio
 import logging
-import threading
-from typing import Any
 
 logger = logging.getLogger(__name__)
 
 _patches_applied = False
 
 
-class _AsyncWorker:
-    """
-    A dedicated background thread with its own event loop.
-
-    Allows sync code to submit async coroutines and block for results,
-    even when called from inside another running event loop. Used to
-    bridge sync tool interfaces with async backends (Modal, SWE-ReX).
-    """
-
-    def __init__(self):
-        self._loop: asyncio.AbstractEventLoop = None
-        self._thread: threading.Thread = None
-        self._started = threading.Event()
-
-    def start(self):
-        """Start the background event loop thread."""
-        self._thread = threading.Thread(target=self._run_loop, daemon=True)
-        self._thread.start()
-        self._started.wait(timeout=30)
-
-    def _run_loop(self):
-        """Background thread entry point -- runs the event loop forever."""
-        self._loop = asyncio.new_event_loop()
-        asyncio.set_event_loop(self._loop)
-        self._started.set()
-        self._loop.run_forever()
-
-    def run_coroutine(self, coro, timeout=600):
-        """
-        Submit a coroutine to the background loop and block until it completes.
-
-        Safe to call from any thread, including threads that already have
-        a running event loop.
-        """
-        if self._loop is None or self._loop.is_closed():
-            raise RuntimeError("AsyncWorker loop is not running")
-        future = asyncio.run_coroutine_threadsafe(coro, self._loop)
-        return future.result(timeout=timeout)
-
-    def stop(self):
-        """Stop the background event loop and join the thread."""
-        if self._loop and self._loop.is_running():
-            self._loop.call_soon_threadsafe(self._loop.stop)
-        if self._thread:
-            self._thread.join(timeout=10)
-
-
-def _patch_swerex_modal():
-    """
-    Monkey patch SwerexModalEnvironment to use a background thread event loop
-    instead of asyncio.run(). This makes it safe to call from inside Atropos's
-    async event loop.
-
-    The patched methods have the exact same interface and behavior -- the only
-    difference is HOW the async work is executed internally.
-    """
-    try:
-        from minisweagent.environments.extra.swerex_modal import (
-            SwerexModalEnvironment,
-            SwerexModalEnvironmentConfig,
-        )
-        from swerex.deployment.modal import ModalDeployment
-        from swerex.runtime.abstract import Command as RexCommand
-    except ImportError:
-        # mini-swe-agent or swe-rex not installed -- nothing to patch
-        logger.debug("mini-swe-agent Modal backend not available, skipping patch")
-        return
-
-    # Save original methods so we can refer to config handling
-    _original_init = SwerexModalEnvironment.__init__
-
-    def _patched_init(self, **kwargs):
-        """Patched __init__: creates Modal deployment on a background thread."""
-        self.config = SwerexModalEnvironmentConfig(**kwargs)
-
-        # Start a dedicated event loop thread for all Modal async operations
-        self._worker = _AsyncWorker()
-        self._worker.start()
-
-        # Pre-build a modal.Image with pip fix for Modal's legacy image builder.
-        # Modal requires `python -m pip` to work during image build, but some
-        # task images (e.g., TBLite's broken-python) have intentionally broken pip.
-        # Fix: remove stale pip dist-info and reinstall via ensurepip before Modal
-        # tries to use it. This is a no-op for images where pip already works.
-        import modal as _modal
-        image_spec = self.config.image
-        if isinstance(image_spec, str):
-            image_spec = _modal.Image.from_registry(
-                image_spec,
-                setup_dockerfile_commands=[
-                    "RUN rm -rf /usr/local/lib/python*/site-packages/pip* 2>/dev/null; "
-                    "python -m ensurepip --upgrade --default-pip 2>/dev/null || true",
-                ],
-            )
-
-        # Create AND start the deployment entirely on the worker's loop/thread
-        # so all gRPC channels and async state are bound to that loop
-        async def _create_and_start():
-            deployment = ModalDeployment(
-                image=image_spec,
-                startup_timeout=self.config.startup_timeout,
-                runtime_timeout=self.config.runtime_timeout,
-                deployment_timeout=self.config.deployment_timeout,
-                install_pipx=self.config.install_pipx,
-                modal_sandbox_kwargs=self.config.modal_sandbox_kwargs,
-            )
-            await deployment.start()
-            return deployment
-
-        self.deployment = self._worker.run_coroutine(_create_and_start())
-
-    def _patched_execute(self, command: str, cwd: str = "", *, timeout: int | None = None) -> dict[str, Any]:
-        """Patched execute: runs commands on the background thread's loop."""
-        async def _do_execute():
-            return await self.deployment.runtime.execute(
-                RexCommand(
-                    command=command,
-                    shell=True,
-                    check=False,
-                    cwd=cwd or self.config.cwd,
-                    timeout=timeout or self.config.timeout,
-                    merge_output_streams=True,
-                    env=self.config.env if self.config.env else None,
-                )
-            )
-
-        output = self._worker.run_coroutine(_do_execute())
-        return {
-            "output": output.stdout,
-            "returncode": output.exit_code,
-        }
-
-    def _patched_stop(self):
-        """Patched stop: stops deployment on the background thread, then stops the thread."""
-        try:
-            self._worker.run_coroutine(
-                asyncio.wait_for(self.deployment.stop(), timeout=10),
-                timeout=15,
-            )
-        except Exception:
-            pass
-        finally:
-            self._worker.stop()
-
-    # Apply the patches
-    SwerexModalEnvironment.__init__ = _patched_init
-    SwerexModalEnvironment.execute = _patched_execute
-    SwerexModalEnvironment.stop = _patched_stop
-
-    logger.debug("Patched SwerexModalEnvironment for async-safe operation")
-
-
 def apply_patches():
-    """
-    Apply all monkey patches needed for Atropos compatibility.
-
-    Safe to call multiple times -- patches are only applied once.
-    Safe for normal CLI use -- patched code works identically when
-    there is no running event loop.
-    """
+    """Apply all monkey patches needed for Atropos compatibility."""
     global _patches_applied
     if _patches_applied:
         return
 
-    _patch_swerex_modal()
-
+    logger.debug("apply_patches() called; no patches needed (async safety is built-in)")
     _patches_applied = True

environments/tool_call_parsers/deepseek_v3_parser.py

@@ -10,12 +10,13 @@ Format uses special unicode tokens:
     <｜tool▁call▁end｜>
     <｜tool▁calls▁end｜>
 
-Based on VLLM's DeepSeekV3ToolParser.extract_tool_calls()
+Fixes Issue #989: Support for multiple simultaneous tool calls.
 """
 
 import re
 import uuid
-from typing import List, Optional
+import logging
+from typing import List, Optional, Tuple
 
 from openai.types.chat.chat_completion_message_tool_call import (
     ChatCompletionMessageToolCall,
@@ -24,6 +25,7 @@ from openai.types.chat.chat_completion_message_tool_call import (
 
 from environments.tool_call_parsers import ParseResult, ToolCallParser, register_parser
 
+logger = logging.getLogger(__name__)
 
 @register_parser("deepseek_v3")
 class DeepSeekV3ToolCallParser(ToolCallParser):
@@ -32,45 +34,56 @@ class DeepSeekV3ToolCallParser(ToolCallParser):
 
     Uses special unicode tokens with fullwidth angle brackets and block elements.
     Extracts type, function name, and JSON arguments from the structured format.
+    Ensures all tool calls are captured when the model executes multiple actions.
     """
 
     START_TOKEN = "<｜tool▁calls▁begin｜>"
 
-    # Regex captures: type, function_name, function_arguments
+    # Updated PATTERN: Using \s* instead of literal \n for increased robustness
+    # against variations in model formatting (Issue #989).
     PATTERN = re.compile(
-        r"<｜tool▁call▁begin｜>(?P<type>.*?)<｜tool▁sep｜>(?P<function_name>.*?)\n```json\n(?P<function_arguments>.*?)\n```<｜tool▁call▁end｜>",
+        r"<｜tool▁call▁begin｜>(?P<type>.*?)<｜tool▁sep｜>(?P<function_name>.*?)\s*```json\s*(?P<function_arguments>.*?)\s*```\s*<｜tool▁call▁end｜>",
         re.DOTALL,
     )
 
     def parse(self, text: str) -> ParseResult:
+        """
+        Parses the input text and extracts all available tool calls.
+        """
         if self.START_TOKEN not in text:
             return text, None
 
         try:
-            matches = self.PATTERN.findall(text)
+            # Using finditer to capture ALL tool calls in the sequence
+            matches = list(self.PATTERN.finditer(text))
             if not matches:
                 return text, None
 
             tool_calls: List[ChatCompletionMessageToolCall] = []
+            
             for match in matches:
-                tc_type, func_name, func_args = match
+                func_name = match.group("function_name").strip()
+                func_args = match.group("function_arguments").strip()
+                
                 tool_calls.append(
                     ChatCompletionMessageToolCall(
                         id=f"call_{uuid.uuid4().hex[:8]}",
                         type="function",
                         function=Function(
-                            name=func_name.strip(),
-                            arguments=func_args.strip(),
+                            name=func_name,
+                            arguments=func_args,
                         ),
                     )
                 )
 
-            if not tool_calls:
-                return text, None
+            if tool_calls:
+                # Content is text before the first tool call block
+                content_index = text.find(self.START_TOKEN)
+                content = text[:content_index].strip()
+                return content if content else None, tool_calls
 
-            # Content is everything before the tool calls section
-            content = text[: text.find(self.START_TOKEN)].strip()
-            return content if content else None, tool_calls
-
-        except Exception:
+            return text, None
+
+        except Exception as e:
+            logger.error(f"Error parsing DeepSeek V3 tool calls: {e}")
             return text, None

environments/tool_call_parsers/mistral_parser.py

@@ -10,7 +10,6 @@ The [TOOL_CALLS] token is the bot_token used by Mistral models.
 """
 
 import json
-import re
 import uuid
 from typing import List, Optional
 
@@ -42,9 +41,6 @@ class MistralToolCallParser(ToolCallParser):
     # The [TOOL_CALLS] token -- may appear as different strings depending on tokenizer
     BOT_TOKEN = "[TOOL_CALLS]"
 
-    # Fallback regex for pre-v11 format when JSON parsing fails
-    TOOL_CALL_REGEX = re.compile(r"\[?\s*(\{.*?\})\s*\]?", re.DOTALL)
-
     def parse(self, text: str) -> ParseResult:
         if self.BOT_TOKEN not in text:
             return text, None
@@ -71,6 +67,13 @@ class MistralToolCallParser(ToolCallParser):
                     tool_name = raw[:brace_idx].strip()
                     args_str = raw[brace_idx:]
 
+                    # Validate and clean the JSON arguments
+                    try:
+                        parsed_args = json.loads(args_str)
+                        args_str = json.dumps(parsed_args, ensure_ascii=False)
+                    except json.JSONDecodeError:
+                        pass  # Keep raw if parsing fails
+
                     tool_calls.append(
                         ChatCompletionMessageToolCall(
                             id=_generate_mistral_id(),
@@ -100,13 +103,14 @@ class MistralToolCallParser(ToolCallParser):
                             )
                         )
                 except json.JSONDecodeError:
-                    # Fallback regex extraction
-                    match = self.TOOL_CALL_REGEX.findall(first_raw)
-                    if match:
-                        for raw_json in match:
-                            try:
-                                tc = json.loads(raw_json)
-                                args = tc.get("arguments", {})
+                    # Fallback: extract JSON objects using raw_decode
+                    decoder = json.JSONDecoder()
+                    idx = 0
+                    while idx < len(first_raw):
+                        try:
+                            obj, end_idx = decoder.raw_decode(first_raw, idx)
+                            if isinstance(obj, dict) and "name" in obj:
+                                args = obj.get("arguments", {})
                                 if isinstance(args, dict):
                                     args = json.dumps(args, ensure_ascii=False)
                                 tool_calls.append(
@@ -114,12 +118,13 @@ class MistralToolCallParser(ToolCallParser):
                                         id=_generate_mistral_id(),
                                         type="function",
                                         function=Function(
-                                            name=tc["name"], arguments=args
+                                            name=obj["name"], arguments=args
                                         ),
                                     )
                                 )
-                            except (json.JSONDecodeError, KeyError):
-                                continue
+                            idx = end_idx
+                        except json.JSONDecodeError:
+                            idx += 1
 
             if not tool_calls:
                 return text, None

flake.lock

@@ -0,0 +1,181 @@
+{
+  "nodes": {
+    "flake-parts": {
+      "inputs": {
+        "nixpkgs-lib": [
+          "nixpkgs"
+        ]
+      },
+      "locked": {
+        "lastModified": 1772408722,
+        "narHash": "sha256-rHuJtdcOjK7rAHpHphUb1iCvgkU3GpfvicLMwwnfMT0=",
+        "owner": "hercules-ci",
+        "repo": "flake-parts",
+        "rev": "f20dc5d9b8027381c474144ecabc9034d6a839a3",
+        "type": "github"
+      },
+      "original": {
+        "owner": "hercules-ci",
+        "repo": "flake-parts",
+        "type": "github"
+      }
+    },
+    "nixpkgs": {
+      "locked": {
+        "lastModified": 1751274312,
+        "narHash": "sha256-/bVBlRpECLVzjV19t5KMdMFWSwKLtb5RyXdjz3LJT+g=",
+        "owner": "NixOS",
+        "repo": "nixpkgs",
+        "rev": "50ab793786d9de88ee30ec4e4c24fb4236fc2674",
+        "type": "github"
+      },
+      "original": {
+        "owner": "NixOS",
+        "ref": "nixos-24.11",
+        "repo": "nixpkgs",
+        "type": "github"
+      }
+    },
+    "pyproject-build-systems": {
+      "inputs": {
+        "nixpkgs": [
+          "nixpkgs"
+        ],
+        "pyproject-nix": "pyproject-nix",
+        "uv2nix": "uv2nix"
+      },
+      "locked": {
+        "lastModified": 1772555609,
+        "narHash": "sha256-3BA3HnUvJSbHJAlJj6XSy0Jmu7RyP2gyB/0fL7XuEDo=",
+        "owner": "pyproject-nix",
+        "repo": "build-system-pkgs",
+        "rev": "c37f66a953535c394244888598947679af231863",
+        "type": "github"
+      },
+      "original": {
+        "owner": "pyproject-nix",
+        "repo": "build-system-pkgs",
+        "type": "github"
+      }
+    },
+    "pyproject-nix": {
+      "inputs": {
+        "nixpkgs": [
+          "pyproject-build-systems",
+          "nixpkgs"
+        ]
+      },
+      "locked": {
+        "lastModified": 1769936401,
+        "narHash": "sha256-kwCOegKLZJM9v/e/7cqwg1p/YjjTAukKPqmxKnAZRgA=",
+        "owner": "nix-community",
+        "repo": "pyproject.nix",
+        "rev": "b0d513eeeebed6d45b4f2e874f9afba2021f7812",
+        "type": "github"
+      },
+      "original": {
+        "owner": "nix-community",
+        "repo": "pyproject.nix",
+        "type": "github"
+      }
+    },
+    "pyproject-nix_2": {
+      "inputs": {
+        "nixpkgs": [
+          "nixpkgs"
+        ]
+      },
+      "locked": {
+        "lastModified": 1772865871,
+        "narHash": "sha256-/ZTSg97aouL0SlPHaokA4r3iuH9QzHVuWPACD2CUCFY=",
+        "owner": "pyproject-nix",
+        "repo": "pyproject.nix",
+        "rev": "e537db02e72d553cea470976b9733581bcf5b3ed",
+        "type": "github"
+      },
+      "original": {
+        "owner": "pyproject-nix",
+        "repo": "pyproject.nix",
+        "type": "github"
+      }
+    },
+    "pyproject-nix_3": {
+      "inputs": {
+        "nixpkgs": [
+          "uv2nix",
+          "nixpkgs"
+        ]
+      },
+      "locked": {
+        "lastModified": 1771518446,
+        "narHash": "sha256-nFJSfD89vWTu92KyuJWDoTQJuoDuddkJV3TlOl1cOic=",
+        "owner": "pyproject-nix",
+        "repo": "pyproject.nix",
+        "rev": "eb204c6b3335698dec6c7fc1da0ebc3c6df05937",
+        "type": "github"
+      },
+      "original": {
+        "owner": "pyproject-nix",
+        "repo": "pyproject.nix",
+        "type": "github"
+      }
+    },
+    "root": {
+      "inputs": {
+        "flake-parts": "flake-parts",
+        "nixpkgs": "nixpkgs",
+        "pyproject-build-systems": "pyproject-build-systems",
+        "pyproject-nix": "pyproject-nix_2",
+        "uv2nix": "uv2nix_2"
+      }
+    },
+    "uv2nix": {
+      "inputs": {
+        "nixpkgs": [
+          "pyproject-build-systems",
+          "nixpkgs"
+        ],
+        "pyproject-nix": [
+          "pyproject-build-systems",
+          "pyproject-nix"
+        ]
+      },
+      "locked": {
+        "lastModified": 1770770348,
+        "narHash": "sha256-A2GzkmzdYvdgmMEu5yxW+xhossP+txrYb7RuzRaqhlg=",
+        "owner": "pyproject-nix",
+        "repo": "uv2nix",
+        "rev": "5d1b2cb4fe3158043fbafbbe2e46238abbc954b0",
+        "type": "github"
+      },
+      "original": {
+        "owner": "pyproject-nix",
+        "repo": "uv2nix",
+        "type": "github"
+      }
+    },
+    "uv2nix_2": {
+      "inputs": {
+        "nixpkgs": [
+          "nixpkgs"
+        ],
+        "pyproject-nix": "pyproject-nix_3"
+      },
+      "locked": {
+        "lastModified": 1773039484,
+        "narHash": "sha256-+boo33KYkJDw9KItpeEXXv8+65f7hHv/earxpcyzQ0I=",
+        "owner": "pyproject-nix",
+        "repo": "uv2nix",
+        "rev": "b68be7cfeacbed9a3fa38a2b5adc0cfb81d9bb1f",
+        "type": "github"
+      },
+      "original": {
+        "owner": "pyproject-nix",
+        "repo": "uv2nix",
+        "type": "github"
+      }
+    }
+  },
+  "root": "root",
+  "version": 7
+}

flake.nix

@@ -0,0 +1,35 @@
+{
+  description = "Hermes Agent - AI agent framework by Nous Research";
+
+  inputs = {
+    nixpkgs.url = "github:NixOS/nixpkgs/nixos-24.11";
+    flake-parts = {
+      url = "github:hercules-ci/flake-parts";
+      inputs.nixpkgs-lib.follows = "nixpkgs";
+    };
+    pyproject-nix = {
+      url = "github:pyproject-nix/pyproject.nix";
+      inputs.nixpkgs.follows = "nixpkgs";
+    };
+    uv2nix = {
+      url = "github:pyproject-nix/uv2nix";
+      inputs.nixpkgs.follows = "nixpkgs";
+    };
+    pyproject-build-systems = {
+      url = "github:pyproject-nix/build-system-pkgs";
+      inputs.nixpkgs.follows = "nixpkgs";
+    };
+  };
+
+  outputs = inputs:
+    inputs.flake-parts.lib.mkFlake { inherit inputs; } {
+      systems = [ "x86_64-linux" "aarch64-linux" "aarch64-darwin" ];
+
+      imports = [
+        ./nix/packages.nix
+        ./nix/nixosModules.nix
+        ./nix/checks.nix
+        ./nix/devShell.nix
+      ];
+    };
+}

gateway/builtin_hooks/__init__.py

@@ -0,0 +1 @@
+"""Built-in gateway hooks that are always registered."""

gateway/builtin_hooks/boot_md.py

@@ -0,0 +1,86 @@
+"""Built-in boot-md hook — run ~/.hermes/BOOT.md on gateway startup.
+
+This hook is always registered. It silently skips if no BOOT.md exists.
+To activate, create ``~/.hermes/BOOT.md`` with instructions for the
+agent to execute on every gateway restart.
+
+Example BOOT.md::
+
+    # Startup Checklist
+
+    1. Check if any cron jobs failed overnight
+    2. Send a status update to Discord #general
+    3. If there are errors in /opt/app/deploy.log, summarize them
+
+The agent runs in a background thread so it doesn't block gateway
+startup. If nothing needs attention, it replies with [SILENT] to
+suppress delivery.
+"""
+
+import logging
+import os
+import threading
+from pathlib import Path
+
+logger = logging.getLogger("hooks.boot-md")
+
+HERMES_HOME = Path(os.environ.get("HERMES_HOME", Path.home() / ".hermes"))
+BOOT_FILE = HERMES_HOME / "BOOT.md"
+
+
+def _build_boot_prompt(content: str) -> str:
+    """Wrap BOOT.md content in a system-level instruction."""
+    return (
+        "You are running a startup boot checklist. Follow the BOOT.md "
+        "instructions below exactly.\n\n"
+        "---\n"
+        f"{content}\n"
+        "---\n\n"
+        "Execute each instruction. If you need to send a message to a "
+        "platform, use the send_message tool.\n"
+        "If nothing needs attention and there is nothing to report, "
+        "reply with ONLY: [SILENT]"
+    )
+
+
+def _run_boot_agent(content: str) -> None:
+    """Spawn a one-shot agent session to execute the boot instructions."""
+    try:
+        from run_agent import AIAgent
+
+        prompt = _build_boot_prompt(content)
+        agent = AIAgent(
+            quiet_mode=True,
+            skip_context_files=True,
+            skip_memory=True,
+            max_iterations=20,
+        )
+        result = agent.run_conversation(prompt)
+        response = result.get("final_response", "")
+        if response and "[SILENT]" not in response:
+            logger.info("boot-md completed: %s", response[:200])
+        else:
+            logger.info("boot-md completed (nothing to report)")
+    except Exception as e:
+        logger.error("boot-md agent failed: %s", e)
+
+
+async def handle(event_type: str, context: dict) -> None:
+    """Gateway startup handler — run BOOT.md if it exists."""
+    if not BOOT_FILE.exists():
+        return
+
+    content = BOOT_FILE.read_text(encoding="utf-8").strip()
+    if not content:
+        return
+
+    logger.info("Running BOOT.md (%d chars)", len(content))
+
+    # Run in a background thread so we don't block gateway startup.
+    thread = threading.Thread(
+        target=_run_boot_agent,
+        args=(content,),
+        name="boot-md",
+        daemon=True,
+    )
+    thread.start()

gateway/channel_directory.py

@@ -9,7 +9,6 @@ action="list" and for resolving human-friendly channel names to numeric IDs.
 import json
 import logging
 from datetime import datetime
-from pathlib import Path
 from typing import Any, Dict, List, Optional
 
 from hermes_cli.config import get_hermes_home
@@ -19,6 +18,20 @@ logger = logging.getLogger(__name__)
 DIRECTORY_PATH = get_hermes_home() / "channel_directory.json"
 
 
+def _normalize_channel_query(value: str) -> str:
+    return value.lstrip("#").strip().lower()
+
+
+def _channel_target_name(platform_name: str, channel: Dict[str, Any]) -> str:
+    """Return the human-facing target label shown to users for a channel entry."""
+    name = channel["name"]
+    if platform_name == "discord" and channel.get("guild"):
+        return f"#{name}"
+    if platform_name != "discord" and channel.get("type"):
+        return f"{name} ({channel['type']})"
+    return name
+
+
 def _session_entry_id(origin: Dict[str, Any]) -> Optional[str]:
     chat_id = origin.get("chat_id")
     if not chat_id:
@@ -63,7 +76,7 @@ def build_channel_directory(adapters: Dict[Any, Any]) -> Dict[str, Any]:
             logger.warning("Channel directory: failed to build %s: %s", platform.value, e)
 
     # Telegram, WhatsApp & Signal can't enumerate chats -- pull from session history
-    for plat_name in ("telegram", "whatsapp", "signal", "email"):
+    for plat_name in ("telegram", "whatsapp", "signal", "email", "sms"):
         if plat_name not in platforms:
             platforms[plat_name] = _build_from_sessions(plat_name)
 
@@ -90,7 +103,7 @@ def _build_discord(adapter) -> List[Dict[str, str]]:
         return channels
 
     try:
-        import discord as _discord
+        import discord as _discord  # noqa: F401 — SDK presence check
     except ImportError:
         return channels
 
@@ -119,7 +132,6 @@ def _build_slack(adapter) -> List[Dict[str, str]]:
         return _build_from_sessions("slack")
 
     try:
-        import asyncio
         from tools.send_message_tool import _send_slack  # noqa: F401
         # Use the Slack Web API directly if available
     except Exception:
@@ -190,23 +202,25 @@ def resolve_channel_name(platform_name: str, name: str) -> Optional[str]:
     if not channels:
         return None
 
-    query = name.lstrip("#").lower()
+    query = _normalize_channel_query(name)
 
-    # 1. Exact name match
+    # 1. Exact name match, including the display labels shown by send_message(action="list")
     for ch in channels:
-        if ch["name"].lower() == query:
+        if _normalize_channel_query(ch["name"]) == query:
+            return ch["id"]
+        if _normalize_channel_query(_channel_target_name(platform_name, ch)) == query:
             return ch["id"]
 
     # 2. Guild-qualified match for Discord ("GuildName/channel")
     if "/" in query:
         guild_part, ch_part = query.rsplit("/", 1)
         for ch in channels:
-            guild = ch.get("guild", "").lower()
-            if guild == guild_part and ch["name"].lower() == ch_part:
+            guild = ch.get("guild", "").strip().lower()
+            if guild == guild_part and _normalize_channel_query(ch["name"]) == ch_part:
                 return ch["id"]
 
     # 3. Partial prefix match (only if unambiguous)
-    matches = [ch for ch in channels if ch["name"].lower().startswith(query)]
+    matches = [ch for ch in channels if _normalize_channel_query(ch["name"]).startswith(query)]
     if len(matches) == 1:
         return matches[0]["id"]
 
@@ -241,17 +255,16 @@ def format_directory_for_display() -> str:
             for guild_name, guild_channels in sorted(guilds.items()):
                 lines.append(f"Discord ({guild_name}):")
                 for ch in sorted(guild_channels, key=lambda c: c["name"]):
-                    lines.append(f"  discord:#{ch['name']}")
+                    lines.append(f"  discord:{_channel_target_name(plat_name, ch)}")
             if dms:
                 lines.append("Discord (DMs):")
                 for ch in dms:
-                    lines.append(f"  discord:{ch['name']}")
+                    lines.append(f"  discord:{_channel_target_name(plat_name, ch)}")
             lines.append("")
         else:
             lines.append(f"{plat_name.title()}:")
             for ch in channels:
-                type_label = f" ({ch['type']})" if ch.get("type") else ""
-                lines.append(f"  {plat_name}:{ch['name']}{type_label}")
+                lines.append(f"  {plat_name}:{_channel_target_name(plat_name, ch)}")
             lines.append("")
 
     lines.append('Use these as the "target" parameter when sending.')

gateway/config.py

@@ -17,10 +17,34 @@ from typing import Dict, List, Optional, Any
 from enum import Enum
 
 from hermes_cli.config import get_hermes_home
+from utils import is_truthy_value
 
 logger = logging.getLogger(__name__)
 
 
+def _coerce_bool(value: Any, default: bool = True) -> bool:
+    """Coerce bool-ish config values, preserving a caller-provided default."""
+    if value is None:
+        return default
+    if isinstance(value, str):
+        lowered = value.strip().lower()
+        if lowered in ("true", "1", "yes", "on"):
+            return True
+        if lowered in ("false", "0", "no", "off"):
+            return False
+        return default
+    return is_truthy_value(value, default=default)
+
+
+def _normalize_unauthorized_dm_behavior(value: Any, default: str = "pair") -> str:
+    """Normalize unauthorized DM behavior to a supported value."""
+    if isinstance(value, str):
+        normalized = value.strip().lower()
+        if normalized in {"pair", "ignore"}:
+            return normalized
+    return default
+
+
 class Platform(Enum):
     """Supported messaging platforms."""
     LOCAL = "local"
@@ -29,8 +53,16 @@ class Platform(Enum):
     WHATSAPP = "whatsapp"
     SLACK = "slack"
     SIGNAL = "signal"
+    MATTERMOST = "mattermost"
+    MATRIX = "matrix"
     HOMEASSISTANT = "homeassistant"
     EMAIL = "email"
+    SMS = "sms"
+    DINGTALK = "dingtalk"
+    API_SERVER = "api_server"
+    WEBHOOK = "webhook"
+    FEISHU = "feishu"
+    WECOM = "wecom"
 
 
 @dataclass
@@ -75,23 +107,32 @@ class SessionResetPolicy:
     mode: str = "both"  # "daily", "idle", "both", or "none"
     at_hour: int = 4  # Hour for daily reset (0-23, local time)
     idle_minutes: int = 1440  # Minutes of inactivity before reset (24 hours)
+    notify: bool = True  # Send a notification to the user when auto-reset occurs
+    notify_exclude_platforms: tuple = ("api_server", "webhook")  # Platforms that don't get reset notifications
     
     def to_dict(self) -> Dict[str, Any]:
         return {
             "mode": self.mode,
             "at_hour": self.at_hour,
             "idle_minutes": self.idle_minutes,
+            "notify": self.notify,
+            "notify_exclude_platforms": list(self.notify_exclude_platforms),
         }
     
     @classmethod
     def from_dict(cls, data: Dict[str, Any]) -> "SessionResetPolicy":
         # Handle both missing keys and explicit null values (YAML null → None)
+        mode = data.get("mode")
         at_hour = data.get("at_hour")
         idle_minutes = data.get("idle_minutes")
+        notify = data.get("notify")
+        exclude = data.get("notify_exclude_platforms")
         return cls(
-            mode=data.get("mode", "both"),
+            mode=mode if mode is not None else "both",
             at_hour=at_hour if at_hour is not None else 4,
             idle_minutes=idle_minutes if idle_minutes is not None else 1440,
+            notify=notify if notify is not None else True,
+            notify_exclude_platforms=tuple(exclude) if exclude is not None else ("api_server", "webhook"),
         )
 
 
@@ -103,6 +144,12 @@ class PlatformConfig:
     api_key: Optional[str] = None  # API key if different from token
     home_channel: Optional[HomeChannel] = None
     
+    # Reply threading mode (Telegram/Slack)
+    # - "off": Never thread replies to original message
+    # - "first": Only first chunk threads to user's message (default)
+    # - "all": All chunks in multi-part replies thread to user's message
+    reply_to_mode: str = "first"
+    
     # Platform-specific settings
     extra: Dict[str, Any] = field(default_factory=dict)
     
@@ -110,6 +157,7 @@ class PlatformConfig:
         result = {
             "enabled": self.enabled,
             "extra": self.extra,
+            "reply_to_mode": self.reply_to_mode,
         }
         if self.token:
             result["token"] = self.token
@@ -130,10 +178,42 @@ class PlatformConfig:
             token=data.get("token"),
             api_key=data.get("api_key"),
             home_channel=home_channel,
+            reply_to_mode=data.get("reply_to_mode", "first"),
             extra=data.get("extra", {}),
         )
 
 
+@dataclass
+class StreamingConfig:
+    """Configuration for real-time token streaming to messaging platforms."""
+    enabled: bool = False
+    transport: str = "edit"       # "edit" (progressive editMessageText) or "off"
+    edit_interval: float = 0.3    # Seconds between message edits
+    buffer_threshold: int = 40    # Chars before forcing an edit
+    cursor: str = " ▉"           # Cursor shown during streaming
+
+    def to_dict(self) -> Dict[str, Any]:
+        return {
+            "enabled": self.enabled,
+            "transport": self.transport,
+            "edit_interval": self.edit_interval,
+            "buffer_threshold": self.buffer_threshold,
+            "cursor": self.cursor,
+        }
+
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "StreamingConfig":
+        if not data:
+            return cls()
+        return cls(
+            enabled=data.get("enabled", False),
+            transport=data.get("transport", "edit"),
+            edit_interval=float(data.get("edit_interval", 0.3)),
+            buffer_threshold=int(data.get("buffer_threshold", 40)),
+            cursor=data.get("cursor", " ▉"),
+        )
+
+
 @dataclass
 class GatewayConfig:
     """
@@ -160,7 +240,19 @@ class GatewayConfig:
     
     # Delivery settings
     always_log_local: bool = True  # Always save cron outputs to local files
-    
+
+    # STT settings
+    stt_enabled: bool = True  # Whether to auto-transcribe inbound voice messages
+
+    # Session isolation in shared chats
+    group_sessions_per_user: bool = True  # Isolate group/channel sessions per participant when user IDs are available
+
+    # Unauthorized DM policy
+    unauthorized_dm_behavior: str = "pair"  # "pair" or "ignore"
+
+    # Streaming configuration
+    streaming: StreamingConfig = field(default_factory=StreamingConfig)
+
     def get_connected_platforms(self) -> List[Platform]:
         """Return list of platforms that are enabled and configured."""
         connected = []
@@ -179,6 +271,21 @@ class GatewayConfig:
             # Email uses extra dict for config (address + imap_host + smtp_host)
             elif platform == Platform.EMAIL and config.extra.get("address"):
                 connected.append(platform)
+            # SMS uses api_key (Twilio auth token) — SID checked via env
+            elif platform == Platform.SMS and os.getenv("TWILIO_ACCOUNT_SID"):
+                connected.append(platform)
+            # API Server uses enabled flag only (no token needed)
+            elif platform == Platform.API_SERVER:
+                connected.append(platform)
+            # Webhook uses enabled flag only (secrets are per-route)
+            elif platform == Platform.WEBHOOK:
+                connected.append(platform)
+            # Feishu uses extra dict for app credentials
+            elif platform == Platform.FEISHU and config.extra.get("app_id"):
+                connected.append(platform)
+            # WeCom uses extra dict for bot credentials
+            elif platform == Platform.WECOM and config.extra.get("bot_id"):
+                connected.append(platform)
         return connected
     
     def get_home_channel(self, platform: Platform) -> Optional[HomeChannel]:
@@ -224,6 +331,10 @@ class GatewayConfig:
             "quick_commands": self.quick_commands,
             "sessions_dir": str(self.sessions_dir),
             "always_log_local": self.always_log_local,
+            "stt_enabled": self.stt_enabled,
+            "group_sessions_per_user": self.group_sessions_per_user,
+            "unauthorized_dm_behavior": self.unauthorized_dm_behavior,
+            "streaming": self.streaming.to_dict(),
         }
     
     @classmethod
@@ -260,6 +371,16 @@ class GatewayConfig:
         if not isinstance(quick_commands, dict):
             quick_commands = {}
 
+        stt_enabled = data.get("stt_enabled")
+        if stt_enabled is None:
+            stt_enabled = data.get("stt", {}).get("enabled") if isinstance(data.get("stt"), dict) else None
+
+        group_sessions_per_user = data.get("group_sessions_per_user")
+        unauthorized_dm_behavior = _normalize_unauthorized_dm_behavior(
+            data.get("unauthorized_dm_behavior"),
+            "pair",
+        )
+
         return cls(
             platforms=platforms,
             default_reset_policy=default_policy,
@@ -269,57 +390,152 @@ class GatewayConfig:
             quick_commands=quick_commands,
             sessions_dir=sessions_dir,
             always_log_local=data.get("always_log_local", True),
+            stt_enabled=_coerce_bool(stt_enabled, True),
+            group_sessions_per_user=_coerce_bool(group_sessions_per_user, True),
+            unauthorized_dm_behavior=unauthorized_dm_behavior,
+            streaming=StreamingConfig.from_dict(data.get("streaming", {})),
         )
 
+    def get_unauthorized_dm_behavior(self, platform: Optional[Platform] = None) -> str:
+        """Return the effective unauthorized-DM behavior for a platform."""
+        if platform:
+            platform_cfg = self.platforms.get(platform)
+            if platform_cfg and "unauthorized_dm_behavior" in platform_cfg.extra:
+                return _normalize_unauthorized_dm_behavior(
+                    platform_cfg.extra.get("unauthorized_dm_behavior"),
+                    self.unauthorized_dm_behavior,
+                )
+        return self.unauthorized_dm_behavior
+
 
 def load_gateway_config() -> GatewayConfig:
     """
     Load gateway configuration from multiple sources.
-    
+
     Priority (highest to lowest):
     1. Environment variables
-    2. ~/.hermes/gateway.json
-    3. cli-config.yaml gateway section
-    4. Defaults
+    2. ~/.hermes/config.yaml (primary user-facing config)
+    3. ~/.hermes/gateway.json (legacy — provides defaults under config.yaml)
+    4. Built-in defaults
     """
-    config = GatewayConfig()
-    
-    # Try loading from ~/.hermes/gateway.json
     _home = get_hermes_home()
-    gateway_config_path = _home / "gateway.json"
-    if gateway_config_path.exists():
-        try:
-            with open(gateway_config_path, "r", encoding="utf-8") as f:
-                data = json.load(f)
-                config = GatewayConfig.from_dict(data)
-        except Exception as e:
-            print(f"[gateway] Warning: Failed to load {gateway_config_path}: {e}")
+    gw_data: dict = {}
 
-    # Bridge session_reset from config.yaml (the user-facing config file)
-    # into the gateway config. config.yaml takes precedence over gateway.json
-    # for session reset policy since that's where hermes setup writes it.
+    # Legacy fallback: gateway.json provides the base layer.
+    # config.yaml keys always win when both specify the same setting.
+    gateway_json_path = _home / "gateway.json"
+    if gateway_json_path.exists():
+        try:
+            with open(gateway_json_path, "r", encoding="utf-8") as f:
+                gw_data = json.load(f) or {}
+            logger.info(
+                "Loaded legacy %s — consider moving settings to config.yaml",
+                gateway_json_path,
+            )
+        except Exception as e:
+            logger.warning("Failed to load %s: %s", gateway_json_path, e)
+
+    # Primary source: config.yaml
     try:
         import yaml
         config_yaml_path = _home / "config.yaml"
         if config_yaml_path.exists():
             with open(config_yaml_path, encoding="utf-8") as f:
                 yaml_cfg = yaml.safe_load(f) or {}
+
+            # Map config.yaml keys → GatewayConfig.from_dict() schema.
+            # Each key overwrites whatever gateway.json may have set.
             sr = yaml_cfg.get("session_reset")
             if sr and isinstance(sr, dict):
-                config.default_reset_policy = SessionResetPolicy.from_dict(sr)
+                gw_data["default_reset_policy"] = sr
 
-            # Bridge quick commands from config.yaml into gateway runtime config.
-            # config.yaml is the user-facing config source, so when present it
-            # should override gateway.json for this setting.
             qc = yaml_cfg.get("quick_commands")
             if qc is not None:
                 if isinstance(qc, dict):
-                    config.quick_commands = qc
+                    gw_data["quick_commands"] = qc
                 else:
-                    logger.warning("Ignoring invalid quick_commands in config.yaml (expected mapping, got %s)", type(qc).__name__)
+                    logger.warning(
+                        "Ignoring invalid quick_commands in config.yaml "
+                        "(expected mapping, got %s)",
+                        type(qc).__name__,
+                    )
 
-            # Bridge discord settings from config.yaml to env vars
-            # (env vars take precedence — only set if not already defined)
+            stt_cfg = yaml_cfg.get("stt")
+            if isinstance(stt_cfg, dict):
+                gw_data["stt"] = stt_cfg
+
+            if "group_sessions_per_user" in yaml_cfg:
+                gw_data["group_sessions_per_user"] = yaml_cfg["group_sessions_per_user"]
+
+            streaming_cfg = yaml_cfg.get("streaming")
+            if isinstance(streaming_cfg, dict):
+                gw_data["streaming"] = streaming_cfg
+
+            if "reset_triggers" in yaml_cfg:
+                gw_data["reset_triggers"] = yaml_cfg["reset_triggers"]
+
+            if "always_log_local" in yaml_cfg:
+                gw_data["always_log_local"] = yaml_cfg["always_log_local"]
+
+            if "unauthorized_dm_behavior" in yaml_cfg:
+                gw_data["unauthorized_dm_behavior"] = _normalize_unauthorized_dm_behavior(
+                    yaml_cfg.get("unauthorized_dm_behavior"),
+                    "pair",
+                )
+
+            # Merge platforms section from config.yaml into gw_data so that
+            # nested keys like platforms.webhook.extra.routes are loaded.
+            yaml_platforms = yaml_cfg.get("platforms")
+            platforms_data = gw_data.setdefault("platforms", {})
+            if not isinstance(platforms_data, dict):
+                platforms_data = {}
+                gw_data["platforms"] = platforms_data
+            if isinstance(yaml_platforms, dict):
+                for plat_name, plat_block in yaml_platforms.items():
+                    if not isinstance(plat_block, dict):
+                        continue
+                    existing = platforms_data.get(plat_name, {})
+                    if not isinstance(existing, dict):
+                        existing = {}
+                    # Deep-merge extra dicts so gateway.json defaults survive
+                    merged_extra = {**existing.get("extra", {}), **plat_block.get("extra", {})}
+                    merged = {**existing, **plat_block}
+                    if merged_extra:
+                        merged["extra"] = merged_extra
+                    platforms_data[plat_name] = merged
+                gw_data["platforms"] = platforms_data
+            for plat in Platform:
+                if plat == Platform.LOCAL:
+                    continue
+                platform_cfg = yaml_cfg.get(plat.value)
+                if not isinstance(platform_cfg, dict):
+                    continue
+                # Collect bridgeable keys from this platform section
+                bridged = {}
+                if "unauthorized_dm_behavior" in platform_cfg:
+                    bridged["unauthorized_dm_behavior"] = _normalize_unauthorized_dm_behavior(
+                        platform_cfg.get("unauthorized_dm_behavior"),
+                        gw_data.get("unauthorized_dm_behavior", "pair"),
+                    )
+                if "reply_prefix" in platform_cfg:
+                    bridged["reply_prefix"] = platform_cfg["reply_prefix"]
+                if "require_mention" in platform_cfg:
+                    bridged["require_mention"] = platform_cfg["require_mention"]
+                if "mention_patterns" in platform_cfg:
+                    bridged["mention_patterns"] = platform_cfg["mention_patterns"]
+                if not bridged:
+                    continue
+                plat_data = platforms_data.setdefault(plat.value, {})
+                if not isinstance(plat_data, dict):
+                    plat_data = {}
+                    platforms_data[plat.value] = plat_data
+                extra = plat_data.setdefault("extra", {})
+                if not isinstance(extra, dict):
+                    extra = {}
+                    plat_data["extra"] = extra
+                extra.update(bridged)
+
+            # Discord settings → env vars (env vars take precedence)
             discord_cfg = yaml_cfg.get("discord", {})
             if isinstance(discord_cfg, dict):
                 if "require_mention" in discord_cfg and not os.getenv("DISCORD_REQUIRE_MENTION"):
@@ -331,8 +547,57 @@ def load_gateway_config() -> GatewayConfig:
                     os.environ["DISCORD_FREE_RESPONSE_CHANNELS"] = str(frc)
                 if "auto_thread" in discord_cfg and not os.getenv("DISCORD_AUTO_THREAD"):
                     os.environ["DISCORD_AUTO_THREAD"] = str(discord_cfg["auto_thread"]).lower()
-    except Exception:
-        pass
+                if "reactions" in discord_cfg and not os.getenv("DISCORD_REACTIONS"):
+                    os.environ["DISCORD_REACTIONS"] = str(discord_cfg["reactions"]).lower()
+
+            # Telegram settings → env vars (env vars take precedence)
+            telegram_cfg = yaml_cfg.get("telegram", {})
+            if isinstance(telegram_cfg, dict):
+                if "require_mention" in telegram_cfg and not os.getenv("TELEGRAM_REQUIRE_MENTION"):
+                    os.environ["TELEGRAM_REQUIRE_MENTION"] = str(telegram_cfg["require_mention"]).lower()
+                if "mention_patterns" in telegram_cfg and not os.getenv("TELEGRAM_MENTION_PATTERNS"):
+                    import json as _json
+                    os.environ["TELEGRAM_MENTION_PATTERNS"] = _json.dumps(telegram_cfg["mention_patterns"])
+                frc = telegram_cfg.get("free_response_chats")
+                if frc is not None and not os.getenv("TELEGRAM_FREE_RESPONSE_CHATS"):
+                    if isinstance(frc, list):
+                        frc = ",".join(str(v) for v in frc)
+                    os.environ["TELEGRAM_FREE_RESPONSE_CHATS"] = str(frc)
+
+            whatsapp_cfg = yaml_cfg.get("whatsapp", {})
+            if isinstance(whatsapp_cfg, dict):
+                if "require_mention" in whatsapp_cfg and not os.getenv("WHATSAPP_REQUIRE_MENTION"):
+                    os.environ["WHATSAPP_REQUIRE_MENTION"] = str(whatsapp_cfg["require_mention"]).lower()
+                if "mention_patterns" in whatsapp_cfg and not os.getenv("WHATSAPP_MENTION_PATTERNS"):
+                    os.environ["WHATSAPP_MENTION_PATTERNS"] = json.dumps(whatsapp_cfg["mention_patterns"])
+                frc = whatsapp_cfg.get("free_response_chats")
+                if frc is not None and not os.getenv("WHATSAPP_FREE_RESPONSE_CHATS"):
+                    if isinstance(frc, list):
+                        frc = ",".join(str(v) for v in frc)
+                    os.environ["WHATSAPP_FREE_RESPONSE_CHATS"] = str(frc)
+
+            # Matrix settings → env vars (env vars take precedence)
+            matrix_cfg = yaml_cfg.get("matrix", {})
+            if isinstance(matrix_cfg, dict):
+                if "require_mention" in matrix_cfg and not os.getenv("MATRIX_REQUIRE_MENTION"):
+                    os.environ["MATRIX_REQUIRE_MENTION"] = str(matrix_cfg["require_mention"]).lower()
+                frc = matrix_cfg.get("free_response_rooms")
+                if frc is not None and not os.getenv("MATRIX_FREE_RESPONSE_ROOMS"):
+                    if isinstance(frc, list):
+                        frc = ",".join(str(v) for v in frc)
+                    os.environ["MATRIX_FREE_RESPONSE_ROOMS"] = str(frc)
+                if "auto_thread" in matrix_cfg and not os.getenv("MATRIX_AUTO_THREAD"):
+                    os.environ["MATRIX_AUTO_THREAD"] = str(matrix_cfg["auto_thread"]).lower()
+
+    except Exception as e:
+        logger.warning(
+            "Failed to process config.yaml — falling back to .env / gateway.json values. "
+            "Check %s for syntax errors. Error: %s",
+            _home / "config.yaml",
+            e,
+        )
+
+    config = GatewayConfig.from_dict(gw_data)
 
     # Override with environment variables
     _apply_env_overrides(config)
@@ -359,6 +624,8 @@ def load_gateway_config() -> GatewayConfig:
         Platform.TELEGRAM: "TELEGRAM_BOT_TOKEN",
         Platform.DISCORD: "DISCORD_BOT_TOKEN",
         Platform.SLACK: "SLACK_BOT_TOKEN",
+        Platform.MATTERMOST: "MATTERMOST_TOKEN",
+        Platform.MATRIX: "MATRIX_ACCESS_TOKEN",
     }
     for platform, pconfig in config.platforms.items():
         if not pconfig.enabled:
@@ -385,6 +652,21 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
         config.platforms[Platform.TELEGRAM].enabled = True
         config.platforms[Platform.TELEGRAM].token = telegram_token
     
+    # Reply threading mode for Telegram (off/first/all)
+    telegram_reply_mode = os.getenv("TELEGRAM_REPLY_TO_MODE", "").lower()
+    if telegram_reply_mode in ("off", "first", "all"):
+        if Platform.TELEGRAM not in config.platforms:
+            config.platforms[Platform.TELEGRAM] = PlatformConfig()
+        config.platforms[Platform.TELEGRAM].reply_to_mode = telegram_reply_mode
+    
+    telegram_fallback_ips = os.getenv("TELEGRAM_FALLBACK_IPS", "")
+    if telegram_fallback_ips:
+        if Platform.TELEGRAM not in config.platforms:
+            config.platforms[Platform.TELEGRAM] = PlatformConfig()
+        config.platforms[Platform.TELEGRAM].extra["fallback_ips"] = [
+            ip.strip() for ip in telegram_fallback_ips.split(",") if ip.strip()
+        ]
+
     telegram_home = os.getenv("TELEGRAM_HOME_CHANNEL")
     if telegram_home and Platform.TELEGRAM in config.platforms:
         config.platforms[Platform.TELEGRAM].home_channel = HomeChannel(
@@ -423,14 +705,13 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             config.platforms[Platform.SLACK] = PlatformConfig()
         config.platforms[Platform.SLACK].enabled = True
         config.platforms[Platform.SLACK].token = slack_token
-        # Home channel
-        slack_home = os.getenv("SLACK_HOME_CHANNEL")
-        if slack_home:
-            config.platforms[Platform.SLACK].home_channel = HomeChannel(
-                platform=Platform.SLACK,
-                chat_id=slack_home,
-                name=os.getenv("SLACK_HOME_CHANNEL_NAME", ""),
-            )
+    slack_home = os.getenv("SLACK_HOME_CHANNEL")
+    if slack_home and Platform.SLACK in config.platforms:
+        config.platforms[Platform.SLACK].home_channel = HomeChannel(
+            platform=Platform.SLACK,
+            chat_id=slack_home,
+            name=os.getenv("SLACK_HOME_CHANNEL_NAME", ""),
+        )
     
     # Signal
     signal_url = os.getenv("SIGNAL_HTTP_URL")
@@ -444,13 +725,60 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             "account": signal_account,
             "ignore_stories": os.getenv("SIGNAL_IGNORE_STORIES", "true").lower() in ("true", "1", "yes"),
         })
-        signal_home = os.getenv("SIGNAL_HOME_CHANNEL")
-        if signal_home:
-            config.platforms[Platform.SIGNAL].home_channel = HomeChannel(
-                platform=Platform.SIGNAL,
-                chat_id=signal_home,
-                name=os.getenv("SIGNAL_HOME_CHANNEL_NAME", "Home"),
-            )
+    signal_home = os.getenv("SIGNAL_HOME_CHANNEL")
+    if signal_home and Platform.SIGNAL in config.platforms:
+        config.platforms[Platform.SIGNAL].home_channel = HomeChannel(
+            platform=Platform.SIGNAL,
+            chat_id=signal_home,
+            name=os.getenv("SIGNAL_HOME_CHANNEL_NAME", "Home"),
+        )
+
+    # Mattermost
+    mattermost_token = os.getenv("MATTERMOST_TOKEN")
+    if mattermost_token:
+        mattermost_url = os.getenv("MATTERMOST_URL", "")
+        if not mattermost_url:
+            logger.warning("MATTERMOST_TOKEN set but MATTERMOST_URL is missing")
+        if Platform.MATTERMOST not in config.platforms:
+            config.platforms[Platform.MATTERMOST] = PlatformConfig()
+        config.platforms[Platform.MATTERMOST].enabled = True
+        config.platforms[Platform.MATTERMOST].token = mattermost_token
+        config.platforms[Platform.MATTERMOST].extra["url"] = mattermost_url
+    mattermost_home = os.getenv("MATTERMOST_HOME_CHANNEL")
+    if mattermost_home and Platform.MATTERMOST in config.platforms:
+        config.platforms[Platform.MATTERMOST].home_channel = HomeChannel(
+            platform=Platform.MATTERMOST,
+            chat_id=mattermost_home,
+            name=os.getenv("MATTERMOST_HOME_CHANNEL_NAME", "Home"),
+        )
+
+    # Matrix
+    matrix_token = os.getenv("MATRIX_ACCESS_TOKEN")
+    matrix_homeserver = os.getenv("MATRIX_HOMESERVER", "")
+    if matrix_token or os.getenv("MATRIX_PASSWORD"):
+        if not matrix_homeserver:
+            logger.warning("MATRIX_ACCESS_TOKEN/MATRIX_PASSWORD set but MATRIX_HOMESERVER is missing")
+        if Platform.MATRIX not in config.platforms:
+            config.platforms[Platform.MATRIX] = PlatformConfig()
+        config.platforms[Platform.MATRIX].enabled = True
+        if matrix_token:
+            config.platforms[Platform.MATRIX].token = matrix_token
+        config.platforms[Platform.MATRIX].extra["homeserver"] = matrix_homeserver
+        matrix_user = os.getenv("MATRIX_USER_ID", "")
+        if matrix_user:
+            config.platforms[Platform.MATRIX].extra["user_id"] = matrix_user
+        matrix_password = os.getenv("MATRIX_PASSWORD", "")
+        if matrix_password:
+            config.platforms[Platform.MATRIX].extra["password"] = matrix_password
+        matrix_e2ee = os.getenv("MATRIX_ENCRYPTION", "").lower() in ("true", "1", "yes")
+        config.platforms[Platform.MATRIX].extra["encryption"] = matrix_e2ee
+    matrix_home = os.getenv("MATRIX_HOME_ROOM")
+    if matrix_home and Platform.MATRIX in config.platforms:
+        config.platforms[Platform.MATRIX].home_channel = HomeChannel(
+            platform=Platform.MATRIX,
+            chat_id=matrix_home,
+            name=os.getenv("MATRIX_HOME_ROOM_NAME", "Home"),
+        )
 
     # Home Assistant
     hass_token = os.getenv("HASS_TOKEN")
@@ -477,12 +805,116 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             "imap_host": email_imap,
             "smtp_host": email_smtp,
         })
-        email_home = os.getenv("EMAIL_HOME_ADDRESS")
-        if email_home:
-            config.platforms[Platform.EMAIL].home_channel = HomeChannel(
-                platform=Platform.EMAIL,
-                chat_id=email_home,
-                name=os.getenv("EMAIL_HOME_ADDRESS_NAME", "Home"),
+    email_home = os.getenv("EMAIL_HOME_ADDRESS")
+    if email_home and Platform.EMAIL in config.platforms:
+        config.platforms[Platform.EMAIL].home_channel = HomeChannel(
+            platform=Platform.EMAIL,
+            chat_id=email_home,
+            name=os.getenv("EMAIL_HOME_ADDRESS_NAME", "Home"),
+        )
+
+    # SMS (Twilio)
+    twilio_sid = os.getenv("TWILIO_ACCOUNT_SID")
+    if twilio_sid:
+        if Platform.SMS not in config.platforms:
+            config.platforms[Platform.SMS] = PlatformConfig()
+        config.platforms[Platform.SMS].enabled = True
+        config.platforms[Platform.SMS].api_key = os.getenv("TWILIO_AUTH_TOKEN", "")
+    sms_home = os.getenv("SMS_HOME_CHANNEL")
+    if sms_home and Platform.SMS in config.platforms:
+        config.platforms[Platform.SMS].home_channel = HomeChannel(
+            platform=Platform.SMS,
+            chat_id=sms_home,
+            name=os.getenv("SMS_HOME_CHANNEL_NAME", "Home"),
+        )
+
+    # API Server
+    api_server_enabled = os.getenv("API_SERVER_ENABLED", "").lower() in ("true", "1", "yes")
+    api_server_key = os.getenv("API_SERVER_KEY", "")
+    api_server_cors_origins = os.getenv("API_SERVER_CORS_ORIGINS", "")
+    api_server_port = os.getenv("API_SERVER_PORT")
+    api_server_host = os.getenv("API_SERVER_HOST")
+    if api_server_enabled or api_server_key:
+        if Platform.API_SERVER not in config.platforms:
+            config.platforms[Platform.API_SERVER] = PlatformConfig()
+        config.platforms[Platform.API_SERVER].enabled = True
+        if api_server_key:
+            config.platforms[Platform.API_SERVER].extra["key"] = api_server_key
+        if api_server_cors_origins:
+            origins = [origin.strip() for origin in api_server_cors_origins.split(",") if origin.strip()]
+            if origins:
+                config.platforms[Platform.API_SERVER].extra["cors_origins"] = origins
+        if api_server_port:
+            try:
+                config.platforms[Platform.API_SERVER].extra["port"] = int(api_server_port)
+            except ValueError:
+                pass
+        if api_server_host:
+            config.platforms[Platform.API_SERVER].extra["host"] = api_server_host
+
+    # Webhook platform
+    webhook_enabled = os.getenv("WEBHOOK_ENABLED", "").lower() in ("true", "1", "yes")
+    webhook_port = os.getenv("WEBHOOK_PORT")
+    webhook_secret = os.getenv("WEBHOOK_SECRET", "")
+    if webhook_enabled:
+        if Platform.WEBHOOK not in config.platforms:
+            config.platforms[Platform.WEBHOOK] = PlatformConfig()
+        config.platforms[Platform.WEBHOOK].enabled = True
+        if webhook_port:
+            try:
+                config.platforms[Platform.WEBHOOK].extra["port"] = int(webhook_port)
+            except ValueError:
+                pass
+        if webhook_secret:
+            config.platforms[Platform.WEBHOOK].extra["secret"] = webhook_secret
+
+    # Feishu / Lark
+    feishu_app_id = os.getenv("FEISHU_APP_ID")
+    feishu_app_secret = os.getenv("FEISHU_APP_SECRET")
+    if feishu_app_id and feishu_app_secret:
+        if Platform.FEISHU not in config.platforms:
+            config.platforms[Platform.FEISHU] = PlatformConfig()
+        config.platforms[Platform.FEISHU].enabled = True
+        config.platforms[Platform.FEISHU].extra.update({
+            "app_id": feishu_app_id,
+            "app_secret": feishu_app_secret,
+            "domain": os.getenv("FEISHU_DOMAIN", "feishu"),
+            "connection_mode": os.getenv("FEISHU_CONNECTION_MODE", "websocket"),
+        })
+        feishu_encrypt_key = os.getenv("FEISHU_ENCRYPT_KEY", "")
+        if feishu_encrypt_key:
+            config.platforms[Platform.FEISHU].extra["encrypt_key"] = feishu_encrypt_key
+        feishu_verification_token = os.getenv("FEISHU_VERIFICATION_TOKEN", "")
+        if feishu_verification_token:
+            config.platforms[Platform.FEISHU].extra["verification_token"] = feishu_verification_token
+        feishu_home = os.getenv("FEISHU_HOME_CHANNEL")
+        if feishu_home:
+            config.platforms[Platform.FEISHU].home_channel = HomeChannel(
+                platform=Platform.FEISHU,
+                chat_id=feishu_home,
+                name=os.getenv("FEISHU_HOME_CHANNEL_NAME", "Home"),
+            )
+
+    # WeCom (Enterprise WeChat)
+    wecom_bot_id = os.getenv("WECOM_BOT_ID")
+    wecom_secret = os.getenv("WECOM_SECRET")
+    if wecom_bot_id and wecom_secret:
+        if Platform.WECOM not in config.platforms:
+            config.platforms[Platform.WECOM] = PlatformConfig()
+        config.platforms[Platform.WECOM].enabled = True
+        config.platforms[Platform.WECOM].extra.update({
+            "bot_id": wecom_bot_id,
+            "secret": wecom_secret,
+        })
+        wecom_ws_url = os.getenv("WECOM_WEBSOCKET_URL", "")
+        if wecom_ws_url:
+            config.platforms[Platform.WECOM].extra["websocket_url"] = wecom_ws_url
+        wecom_home = os.getenv("WECOM_HOME_CHANNEL")
+        if wecom_home:
+            config.platforms[Platform.WECOM].home_channel = HomeChannel(
+                platform=Platform.WECOM,
+                chat_id=wecom_home,
+                name=os.getenv("WECOM_HOME_CHANNEL_NAME", "Home"),
             )
 
     # Session settings
@@ -499,12 +931,3 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             config.default_reset_policy.at_hour = int(reset_hour)
         except ValueError:
             pass
-
-
-def save_gateway_config(config: GatewayConfig) -> None:
-    """Save gateway configuration to ~/.hermes/gateway.json."""
-    gateway_config_path = get_hermes_home() / "gateway.json"
-    gateway_config_path.parent.mkdir(parents=True, exist_ok=True)
-    
-    with open(gateway_config_path, "w", encoding="utf-8") as f:
-        json.dump(config.to_dict(), f, indent=2)

gateway/delivery.py

@@ -13,7 +13,6 @@ from pathlib import Path
 from datetime import datetime
 from dataclasses import dataclass
 from typing import Dict, List, Optional, Any, Union
-from enum import Enum
 
 from hermes_cli.config import get_hermes_home
 
@@ -71,12 +70,15 @@ class DeliveryTarget:
         if target == "local":
             return cls(platform=Platform.LOCAL)
         
-        # Check for platform:chat_id format
+        # Check for platform:chat_id or platform:chat_id:thread_id format
         if ":" in target:
-            platform_str, chat_id = target.split(":", 1)
+            parts = target.split(":", 2)
+            platform_str = parts[0]
+            chat_id = parts[1] if len(parts) > 1 else None
+            thread_id = parts[2] if len(parts) > 2 else None
             try:
                 platform = Platform(platform_str)
-                return cls(platform=platform, chat_id=chat_id, is_explicit=True)
+                return cls(platform=platform, chat_id=chat_id, thread_id=thread_id, is_explicit=True)
             except ValueError:
                 # Unknown platform, treat as local
                 return cls(platform=Platform.LOCAL)
@@ -95,6 +97,8 @@ class DeliveryTarget:
             return "origin"
         if self.platform == Platform.LOCAL:
             return "local"
+        if self.chat_id and self.thread_id:
+            return f"{self.platform.value}:{self.chat_id}:{self.thread_id}"
         if self.chat_id:
             return f"{self.platform.value}:{self.chat_id}"
         return self.platform.value

gateway/hooks.py

@@ -8,8 +8,9 @@ Hooks are discovered from ~/.hermes/hooks/ directories, each containing:
 
 Events:
   - gateway:startup     -- Gateway process starts
-  - session:start       -- New session created
-  - session:reset       -- User ran /new or /reset
+  - session:start       -- New session created (first message of a new session)
+  - session:end         -- Session ends (user ran /new or /reset)
+  - session:reset       -- Session reset completed (new session entry created)
   - agent:start         -- Agent begins processing a message
   - agent:step          -- Each turn in the tool-calling loop
   - agent:end           -- Agent finishes processing
@@ -20,8 +21,6 @@ Errors in hooks are caught and logged but never block the main pipeline.
 
 import asyncio
 import importlib.util
-import os
-from pathlib import Path
 from typing import Any, Callable, Dict, List, Optional
 
 import yaml
@@ -52,14 +51,33 @@ class HookRegistry:
         """Return metadata about all loaded hooks."""
         return list(self._loaded_hooks)
 
+    def _register_builtin_hooks(self) -> None:
+        """Register built-in hooks that are always active."""
+        try:
+            from gateway.builtin_hooks.boot_md import handle as boot_md_handle
+
+            self._handlers.setdefault("gateway:startup", []).append(boot_md_handle)
+            self._loaded_hooks.append({
+                "name": "boot-md",
+                "description": "Run ~/.hermes/BOOT.md on gateway startup",
+                "events": ["gateway:startup"],
+                "path": "(builtin)",
+            })
+        except Exception as e:
+            print(f"[hooks] Could not load built-in boot-md hook: {e}", flush=True)
+
     def discover_and_load(self) -> None:
         """
         Scan the hooks directory for hook directories and load their handlers.
 
+        Also registers built-in hooks that are always active.
+
         Each hook directory must contain:
           - HOOK.yaml with at least 'name' and 'events' keys
           - handler.py with a top-level 'handle' function (sync or async)
         """
+        self._register_builtin_hooks()
+
         if not HOOKS_DIR.exists():
             return
 

gateway/mirror.py

@@ -12,7 +12,6 @@ the full SessionStore machinery.
 import json
 import logging
 from datetime import datetime
-from pathlib import Path
 from typing import Optional
 
 from hermes_cli.config import get_hermes_home

gateway/pairing.py

@@ -25,7 +25,7 @@ import time
 from pathlib import Path
 from typing import Optional
 
-from hermes_cli.config import get_hermes_home
+from hermes_constants import get_hermes_dir
 
 
 # Unambiguous alphabet -- excludes 0/O, 1/I to prevent confusion
@@ -41,7 +41,7 @@ LOCKOUT_SECONDS = 3600              # Lockout duration after too many failures
 MAX_PENDING_PER_PLATFORM = 3        # Max pending codes per platform
 MAX_FAILED_ATTEMPTS = 5             # Failed approvals before lockout
 
-PAIRING_DIR = get_hermes_home() / "pairing"
+PAIRING_DIR = get_hermes_dir("platforms/pairing", "pairing")
 
 
 def _secure_write(path: Path, data: str) -> None:

gateway/platforms/base.py

@@ -8,6 +8,7 @@ and implement the required methods.
 import asyncio
 import logging
 import os
+import random
 import re
 import uuid
 from abc import ABC, abstractmethod
@@ -26,6 +27,7 @@ sys.path.insert(0, str(_Path(__file__).resolve().parents[2]))
 from gateway.config import Platform, PlatformConfig
 from gateway.session import SessionSource, build_session_key
 from hermes_cli.config import get_hermes_home
+from hermes_constants import get_hermes_dir
 
 
 GATEWAY_SECRET_CAPTURE_UNSUPPORTED_MESSAGE = (
@@ -43,8 +45,8 @@ GATEWAY_SECRET_CAPTURE_UNSUPPORTED_MESSAGE = (
 # (e.g. Telegram file URLs expire after ~1 hour).
 # ---------------------------------------------------------------------------
 
-# Default location: {HERMES_HOME}/image_cache/
-IMAGE_CACHE_DIR = get_hermes_home() / "image_cache"
+# Default location: {HERMES_HOME}/cache/images/ (legacy: image_cache/)
+IMAGE_CACHE_DIR = get_hermes_dir("cache/images", "image_cache")
 
 
 def get_image_cache_dir() -> Path:
@@ -71,31 +73,51 @@ def cache_image_from_bytes(data: bytes, ext: str = ".jpg") -> str:
     return str(filepath)
 
 
-async def cache_image_from_url(url: str, ext: str = ".jpg") -> str:
+async def cache_image_from_url(url: str, ext: str = ".jpg", retries: int = 2) -> str:
     """
     Download an image from a URL and save it to the local cache.
 
-    Uses httpx for async download with a reasonable timeout.
+    Retries on transient failures (timeouts, 429, 5xx) with exponential
+    backoff so a single slow CDN response doesn't lose the media.
 
     Args:
         url: The HTTP/HTTPS URL to download from.
         ext: File extension including the dot (e.g. ".jpg", ".png").
+        retries: Number of retry attempts on transient failures.
 
     Returns:
         Absolute path to the cached image file as a string.
     """
+    import asyncio
     import httpx
+    import logging as _logging
+    _log = _logging.getLogger(__name__)
 
+    last_exc = None
     async with httpx.AsyncClient(timeout=30.0, follow_redirects=True) as client:
-        response = await client.get(
-            url,
-            headers={
-                "User-Agent": "Mozilla/5.0 (compatible; HermesAgent/1.0)",
-                "Accept": "image/*,*/*;q=0.8",
-            },
-        )
-        response.raise_for_status()
-        return cache_image_from_bytes(response.content, ext)
+        for attempt in range(retries + 1):
+            try:
+                response = await client.get(
+                    url,
+                    headers={
+                        "User-Agent": "Mozilla/5.0 (compatible; HermesAgent/1.0)",
+                        "Accept": "image/*,*/*;q=0.8",
+                    },
+                )
+                response.raise_for_status()
+                return cache_image_from_bytes(response.content, ext)
+            except (httpx.TimeoutException, httpx.HTTPStatusError) as exc:
+                last_exc = exc
+                if isinstance(exc, httpx.HTTPStatusError) and exc.response.status_code < 429:
+                    raise
+                if attempt < retries:
+                    wait = 1.5 * (attempt + 1)
+                    _log.debug("Media cache retry %d/%d for %s (%.1fs): %s",
+                               attempt + 1, retries, url[:80], wait, exc)
+                    await asyncio.sleep(wait)
+                    continue
+                raise
+    raise last_exc
 
 
 def cleanup_image_cache(max_age_hours: int = 24) -> int:
@@ -126,7 +148,7 @@ def cleanup_image_cache(max_age_hours: int = 24) -> int:
 # here so the STT tool (OpenAI Whisper) can transcribe them from local files.
 # ---------------------------------------------------------------------------
 
-AUDIO_CACHE_DIR = get_hermes_home() / "audio_cache"
+AUDIO_CACHE_DIR = get_hermes_dir("cache/audio", "audio_cache")
 
 
 def get_audio_cache_dir() -> Path:
@@ -153,29 +175,51 @@ def cache_audio_from_bytes(data: bytes, ext: str = ".ogg") -> str:
     return str(filepath)
 
 
-async def cache_audio_from_url(url: str, ext: str = ".ogg") -> str:
+async def cache_audio_from_url(url: str, ext: str = ".ogg", retries: int = 2) -> str:
     """
     Download an audio file from a URL and save it to the local cache.
 
+    Retries on transient failures (timeouts, 429, 5xx) with exponential
+    backoff so a single slow CDN response doesn't lose the media.
+
     Args:
         url: The HTTP/HTTPS URL to download from.
         ext: File extension including the dot (e.g. ".ogg", ".mp3").
+        retries: Number of retry attempts on transient failures.
 
     Returns:
         Absolute path to the cached audio file as a string.
     """
+    import asyncio
     import httpx
+    import logging as _logging
+    _log = _logging.getLogger(__name__)
 
+    last_exc = None
     async with httpx.AsyncClient(timeout=30.0, follow_redirects=True) as client:
-        response = await client.get(
-            url,
-            headers={
-                "User-Agent": "Mozilla/5.0 (compatible; HermesAgent/1.0)",
-                "Accept": "audio/*,*/*;q=0.8",
-            },
-        )
-        response.raise_for_status()
-        return cache_audio_from_bytes(response.content, ext)
+        for attempt in range(retries + 1):
+            try:
+                response = await client.get(
+                    url,
+                    headers={
+                        "User-Agent": "Mozilla/5.0 (compatible; HermesAgent/1.0)",
+                        "Accept": "audio/*,*/*;q=0.8",
+                    },
+                )
+                response.raise_for_status()
+                return cache_audio_from_bytes(response.content, ext)
+            except (httpx.TimeoutException, httpx.HTTPStatusError) as exc:
+                last_exc = exc
+                if isinstance(exc, httpx.HTTPStatusError) and exc.response.status_code < 429:
+                    raise
+                if attempt < retries:
+                    wait = 1.5 * (attempt + 1)
+                    _log.debug("Audio cache retry %d/%d for %s (%.1fs): %s",
+                               attempt + 1, retries, url[:80], wait, exc)
+                    await asyncio.sleep(wait)
+                    continue
+                raise
+    raise last_exc
 
 
 # ---------------------------------------------------------------------------
@@ -185,12 +229,13 @@ async def cache_audio_from_url(url: str, ext: str = ".ogg") -> str:
 # here so the agent can reference them by local file path.
 # ---------------------------------------------------------------------------
 
-DOCUMENT_CACHE_DIR = get_hermes_home() / "document_cache"
+DOCUMENT_CACHE_DIR = get_hermes_dir("cache/documents", "document_cache")
 
 SUPPORTED_DOCUMENT_TYPES = {
     ".pdf": "application/pdf",
     ".md": "text/markdown",
     ".txt": "text/plain",
+    ".zip": "application/zip",
     ".docx": "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
     ".xlsx": "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
     ".pptx": "application/vnd.openxmlformats-officedocument.presentationml.presentation",
@@ -288,11 +333,16 @@ class MessageEvent:
     message_id: Optional[str] = None
     
     # Media attachments
+    # media_urls: local file paths (for vision tool access)
     media_urls: List[str] = field(default_factory=list)
     media_types: List[str] = field(default_factory=list)
     
     # Reply context
     reply_to_message_id: Optional[str] = None
+    reply_to_text: Optional[str] = None  # Text of the replied-to message (for context injection)
+    
+    # Auto-loaded skill for topic/channel bindings (e.g., Telegram DM Topics)
+    auto_skill: Optional[str] = None
     
     # Timestamps
     timestamp: datetime = field(default_factory=datetime.now)
@@ -307,7 +357,10 @@ class MessageEvent:
             return None
         # Split on space and get first word, strip the /
         parts = self.text.split(maxsplit=1)
-        return parts[0][1:].lower() if parts else None
+        raw = parts[0][1:].lower() if parts else None
+        if raw and "@" in raw:
+            raw = raw.split("@", 1)[0]
+        return raw
     
     def get_command_args(self) -> str:
         """Get the arguments after a command."""
@@ -324,6 +377,27 @@ class SendResult:
     message_id: Optional[str] = None
     error: Optional[str] = None
     raw_response: Any = None
+    retryable: bool = False  # True for transient connection errors — base will retry automatically
+
+
+# Error substrings that indicate a transient *connection* failure worth retrying.
+# "timeout" / "timed out" / "readtimeout" / "writetimeout" are intentionally
+# excluded: a read/write timeout on a non-idempotent call (e.g. send_message)
+# means the request may have reached the server — retrying risks duplicate
+# delivery.  "connecttimeout" is safe because the connection was never
+# established.  Platforms that know a timeout is safe to retry should set
+# SendResult.retryable = True explicitly.
+_RETRYABLE_ERROR_PATTERNS = (
+    "connecterror",
+    "connectionerror",
+    "connectionreset",
+    "connectionrefused",
+    "connecttimeout",
+    "network",
+    "broken pipe",
+    "remotedisconnected",
+    "eoferror",
+)
 
 
 # Type for message handlers
@@ -355,6 +429,10 @@ class BasePlatformAdapter(ABC):
         # Key: session_key (e.g., chat_id), Value: (event, asyncio.Event for interrupt)
         self._active_sessions: Dict[str, asyncio.Event] = {}
         self._pending_messages: Dict[str, MessageEvent] = {}
+        # Background message-processing tasks spawned by handle_message().
+        # Gateway shutdown cancels these so an old gateway instance doesn't keep
+        # working on a task after --replace or manual restarts.
+        self._background_tasks: set[asyncio.Task] = set()
         # Chats where auto-TTS on voice input is disabled (set by /voice off)
         self._auto_tts_disabled_chats: set = set()
 
@@ -498,6 +576,14 @@ class BasePlatformAdapter(ABC):
         metadata: optional dict with platform-specific context (e.g. thread_id for Slack).
         """
         pass
+
+    async def stop_typing(self, chat_id: str) -> None:
+        """Stop a persistent typing indicator (if the platform uses one).
+
+        Override in subclasses that start background typing loops.
+        Default is a no-op for platforms with one-shot typing indicators.
+        """
+        pass
     
     async def send_image(
         self,
@@ -505,6 +591,7 @@ class BasePlatformAdapter(ABC):
         image_url: str,
         caption: Optional[str] = None,
         reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
     ) -> SendResult:
         """
         Send an image natively via the platform API.
@@ -523,6 +610,7 @@ class BasePlatformAdapter(ABC):
         animation_url: str,
         caption: Optional[str] = None,
         reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
     ) -> SendResult:
         """
         Send an animated GIF natively via the platform API.
@@ -531,7 +619,7 @@ class BasePlatformAdapter(ABC):
         (e.g., Telegram send_animation) so they auto-play inline.
         Default falls back to send_image.
         """
-        return await self.send_image(chat_id=chat_id, image_url=animation_url, caption=caption, reply_to=reply_to)
+        return await self.send_image(chat_id=chat_id, image_url=animation_url, caption=caption, reply_to=reply_to, metadata=metadata)
     
     @staticmethod
     def _is_animation_url(url: str) -> bool:
@@ -705,7 +793,7 @@ class BasePlatformAdapter(ABC):
         # Extract MEDIA:<path> tags, allowing optional whitespace after the colon
         # and quoted/backticked paths for LLM-formatted outputs.
         media_pattern = re.compile(
-            r'''[`"']?MEDIA:\s*(?P<path>`[^`\n]+`|"[^"\n]+"|'[^'\n]+'|\S+)[`"']?'''
+            r'''[`"']?MEDIA:\s*(?P<path>`[^`\n]+`|"[^"\n]+"|'[^'\n]+'|(?:~/|/)\S+(?:[^\S\n]+\S+)*?\.(?:png|jpe?g|gif|webp|mp4|mov|avi|mkv|webm|ogg|opus|mp3|wav|m4a)(?=[\s`"',;:)\]}]|$)|\S+)[`"']?'''
         )
         for match in media_pattern.finditer(content):
             path = match.group("path").strip()
@@ -721,7 +809,75 @@ class BasePlatformAdapter(ABC):
             cleaned = re.sub(r'\n{3,}', '\n\n', cleaned).strip()
         
         return media, cleaned
-    
+
+    @staticmethod
+    def extract_local_files(content: str) -> Tuple[List[str], str]:
+        """
+        Detect bare local file paths in response text for native media delivery.
+
+        Matches absolute paths (/...) and tilde paths (~/) ending in common
+        image or video extensions.  Validates each candidate with
+        ``os.path.isfile()`` to avoid false positives from URLs or
+        non-existent paths.
+
+        Paths inside fenced code blocks (``` ... ```) and inline code
+        (`...`) are ignored so that code samples are never mutilated.
+
+        Returns:
+            Tuple of (list of expanded file paths, cleaned text with the
+            raw path strings removed).
+        """
+        _LOCAL_MEDIA_EXTS = (
+            '.png', '.jpg', '.jpeg', '.gif', '.webp',
+            '.mp4', '.mov', '.avi', '.mkv', '.webm',
+        )
+        ext_part = '|'.join(e.lstrip('.') for e in _LOCAL_MEDIA_EXTS)
+
+        # (?<![/:\w.]) prevents matching inside URLs (e.g. https://…/img.png)
+        #             and relative paths (./foo.png)
+        # (?:~/|/)    anchors to absolute or home-relative paths
+        path_re = re.compile(
+            r'(?<![/:\w.])(?:~/|/)(?:[\w.\-]+/)*[\w.\-]+\.(?:' + ext_part + r')\b',
+            re.IGNORECASE,
+        )
+
+        # Build spans covered by fenced code blocks and inline code
+        code_spans: list = []
+        for m in re.finditer(r'```[^\n]*\n.*?```', content, re.DOTALL):
+            code_spans.append((m.start(), m.end()))
+        for m in re.finditer(r'`[^`\n]+`', content):
+            code_spans.append((m.start(), m.end()))
+
+        def _in_code(pos: int) -> bool:
+            return any(s <= pos < e for s, e in code_spans)
+
+        found: list = []  # (raw_match_text, expanded_path)
+        for match in path_re.finditer(content):
+            if _in_code(match.start()):
+                continue
+            raw = match.group(0)
+            expanded = os.path.expanduser(raw)
+            if os.path.isfile(expanded):
+                found.append((raw, expanded))
+
+        # Deduplicate by expanded path, preserving discovery order
+        seen: set = set()
+        unique: list = []
+        for raw, expanded in found:
+            if expanded not in seen:
+                seen.add(expanded)
+                unique.append((raw, expanded))
+
+        paths = [expanded for _, expanded in unique]
+
+        cleaned = content
+        if unique:
+            for raw, _exp in unique:
+                cleaned = cleaned.replace(raw, '')
+            cleaned = re.sub(r'\n{3,}', '\n\n', cleaned).strip()
+
+        return paths, cleaned
+
     async def _keep_typing(self, chat_id: str, interval: float = 2.0, metadata=None) -> None:
         """
         Continuously send typing indicator until cancelled.
@@ -735,7 +891,139 @@ class BasePlatformAdapter(ABC):
                 await asyncio.sleep(interval)
         except asyncio.CancelledError:
             pass  # Normal cancellation when handler completes
+        finally:
+            # Ensure the underlying platform typing loop is stopped.
+            # _keep_typing may have called send_typing() after an outer
+            # stop_typing() cleared the task dict, recreating the loop.
+            # Cancelling _keep_typing alone won't clean that up.
+            if hasattr(self, "stop_typing"):
+                try:
+                    await self.stop_typing(chat_id)
+                except Exception:
+                    pass
     
+    # ── Processing lifecycle hooks ──────────────────────────────────────────
+    # Subclasses override these to react to message processing events
+    # (e.g. Discord adds 👀/✅/❌ reactions).
+
+    async def on_processing_start(self, event: MessageEvent) -> None:
+        """Hook called when background processing begins."""
+
+    async def on_processing_complete(self, event: MessageEvent, success: bool) -> None:
+        """Hook called when background processing completes."""
+
+    async def _run_processing_hook(self, hook_name: str, *args: Any, **kwargs: Any) -> None:
+        """Run a lifecycle hook without letting failures break message flow."""
+        hook = getattr(self, hook_name, None)
+        if not callable(hook):
+            return
+        try:
+            await hook(*args, **kwargs)
+        except Exception as e:
+            logger.warning("[%s] %s hook failed: %s", self.name, hook_name, e)
+
+    @staticmethod
+    def _is_retryable_error(error: Optional[str]) -> bool:
+        """Return True if the error string looks like a transient network failure."""
+        if not error:
+            return False
+        lowered = error.lower()
+        return any(pat in lowered for pat in _RETRYABLE_ERROR_PATTERNS)
+
+    @staticmethod
+    def _is_timeout_error(error: Optional[str]) -> bool:
+        """Return True if the error string indicates a read/write timeout.
+
+        Timeout errors are NOT retryable and should NOT trigger plain-text
+        fallback — the request may have already been delivered.
+        """
+        if not error:
+            return False
+        lowered = error.lower()
+        return "timed out" in lowered or "readtimeout" in lowered or "writetimeout" in lowered
+
+    async def _send_with_retry(
+        self,
+        chat_id: str,
+        content: str,
+        reply_to: Optional[str] = None,
+        metadata: Any = None,
+        max_retries: int = 2,
+        base_delay: float = 2.0,
+    ) -> "SendResult":
+        """
+        Send a message with automatic retry for transient network errors.
+
+        On permanent failures (e.g. formatting / permission errors) falls back
+        to a plain-text version before giving up. If all attempts fail due to
+        network errors, sends the user a brief delivery-failure notice so they
+        know to retry rather than waiting indefinitely.
+        """
+
+        result = await self.send(
+            chat_id=chat_id,
+            content=content,
+            reply_to=reply_to,
+            metadata=metadata,
+        )
+
+        if result.success:
+            return result
+
+        error_str = result.error or ""
+        is_network = result.retryable or self._is_retryable_error(error_str)
+
+        # Timeout errors are not safe to retry (message may have been
+        # delivered) and not formatting errors — return the failure as-is.
+        if not is_network and self._is_timeout_error(error_str):
+            return result
+
+        if is_network:
+            # Retry with exponential backoff for transient errors
+            for attempt in range(1, max_retries + 1):
+                delay = base_delay * (2 ** (attempt - 1)) + random.uniform(0, 1)
+                logger.warning(
+                    "[%s] Send failed (attempt %d/%d, retrying in %.1fs): %s",
+                    self.name, attempt, max_retries, delay, error_str,
+                )
+                await asyncio.sleep(delay)
+                result = await self.send(
+                    chat_id=chat_id,
+                    content=content,
+                    reply_to=reply_to,
+                    metadata=metadata,
+                )
+                if result.success:
+                    logger.info("[%s] Send succeeded on retry %d", self.name, attempt)
+                    return result
+                error_str = result.error or ""
+                if not (result.retryable or self._is_retryable_error(error_str)):
+                    break  # error switched to non-transient — fall through to plain-text fallback
+            else:
+                # All retries exhausted (loop completed without break) — notify user
+                logger.error("[%s] Failed to deliver response after %d retries: %s", self.name, max_retries, error_str)
+                notice = (
+                    "\u26a0\ufe0f Message delivery failed after multiple attempts. "
+                    "Please try again \u2014 your request was processed but the response could not be sent."
+                )
+                try:
+                    await self.send(chat_id=chat_id, content=notice, reply_to=reply_to, metadata=metadata)
+                except Exception as notify_err:
+                    logger.debug("[%s] Could not send delivery-failure notice: %s", self.name, notify_err)
+                return result
+
+        # Non-network / post-retry formatting failure: try plain text as fallback
+        logger.warning("[%s] Send failed: %s — trying plain-text fallback", self.name, error_str)
+        fallback_result = await self.send(
+            chat_id=chat_id,
+            content=f"(Response formatting failed, plain text:)\n\n{content[:3500]}",
+            reply_to=reply_to,
+            metadata=metadata,
+        )
+        if not fallback_result.success:
+            logger.error("[%s] Fallback send also failed: %s", self.name, fallback_result.error)
+        return fallback_result
+
     async def handle_message(self, event: MessageEvent) -> None:
         """
         Process an incoming message.
@@ -747,19 +1035,103 @@ class BasePlatformAdapter(ABC):
         if not self._message_handler:
             return
         
-        session_key = build_session_key(event.source)
+        session_key = build_session_key(
+            event.source,
+            group_sessions_per_user=self.config.extra.get("group_sessions_per_user", True),
+        )
         
         # Check if there's already an active handler for this session
         if session_key in self._active_sessions:
-            # Store this as a pending message - it will interrupt the running agent
-            print(f"[{self.name}] ⚡ New message while session {session_key} is active - triggering interrupt")
+            # /approve and /deny must bypass the active-session guard.
+            # The agent thread is blocked on threading.Event.wait() inside
+            # tools/approval.py — queuing these commands creates a deadlock:
+            # the agent waits for approval, approval waits for agent to finish.
+            # Dispatch directly to the message handler without touching session
+            # lifecycle (no competing background task, no session guard removal).
+            cmd = event.get_command()
+            if cmd in ("approve", "deny"):
+                logger.debug(
+                    "[%s] Approval command '/%s' bypassing active-session guard for %s",
+                    self.name, cmd, session_key,
+                )
+                try:
+                    _thread_meta = {"thread_id": event.source.thread_id} if event.source.thread_id else None
+                    response = await self._message_handler(event)
+                    if response:
+                        await self._send_with_retry(
+                            chat_id=event.source.chat_id,
+                            content=response,
+                            reply_to=event.message_id,
+                            metadata=_thread_meta,
+                        )
+                except Exception as e:
+                    logger.error("[%s] Approval dispatch failed: %s", self.name, e, exc_info=True)
+                return
+
+            # /status must also bypass the active-session guard so it always
+            # returns a system-generated response instead of being queued as
+            # user text and passed to the agent (#5046).
+            if cmd == "status":
+                logger.debug(
+                    "[%s] Status command bypassing active-session guard for %s",
+                    self.name, session_key,
+                )
+                try:
+                    _thread_meta = {"thread_id": event.source.thread_id} if event.source.thread_id else None
+                    response = await self._message_handler(event)
+                    if response:
+                        await self._send_with_retry(
+                            chat_id=event.source.chat_id,
+                            content=response,
+                            reply_to=event.message_id,
+                            metadata=_thread_meta,
+                        )
+                except Exception as e:
+                    logger.error("[%s] Status dispatch failed: %s", self.name, e, exc_info=True)
+                return
+
+            # Special case: photo bursts/albums frequently arrive as multiple near-
+            # simultaneous messages. Queue them without interrupting the active run,
+            # then process them immediately after the current task finishes.
+            if event.message_type == MessageType.PHOTO:
+                logger.debug("[%s] Queuing photo follow-up for session %s without interrupt", self.name, session_key)
+                existing = self._pending_messages.get(session_key)
+                if existing and existing.message_type == MessageType.PHOTO:
+                    existing.media_urls.extend(event.media_urls)
+                    existing.media_types.extend(event.media_types)
+                    if event.text:
+                        if not existing.text:
+                            existing.text = event.text
+                        elif event.text not in existing.text:
+                            existing.text = f"{existing.text}\n\n{event.text}".strip()
+                else:
+                    self._pending_messages[session_key] = event
+                return  # Don't interrupt now - will run after current task completes
+
+            # Default behavior for non-photo follow-ups: interrupt the running agent
+            logger.debug("[%s] New message while session %s is active — triggering interrupt", self.name, session_key)
             self._pending_messages[session_key] = event
             # Signal the interrupt (the processing task checks this)
             self._active_sessions[session_key].set()
             return  # Don't process now - will be handled after current task finishes
         
+        # Mark session as active BEFORE spawning background task to close
+        # the race window where a second message arriving before the task
+        # starts would also pass the _active_sessions check and spawn a
+        # duplicate task.  (grammY sequentialize / aiogram EventIsolation
+        # pattern — set the guard synchronously, not inside the task.)
+        self._active_sessions[session_key] = asyncio.Event()
+
         # Spawn background task to process this message
-        asyncio.create_task(self._process_message_background(event, session_key))
+        task = asyncio.create_task(self._process_message_background(event, session_key))
+        try:
+            self._background_tasks.add(task)
+        except TypeError:
+            # Some tests stub create_task() with lightweight sentinels that are not
+            # hashable and do not support lifecycle callbacks.
+            return
+        if hasattr(task, "add_done_callback"):
+            task.add_done_callback(self._background_tasks.discard)
     
     @staticmethod
     def _get_human_delay() -> float:
@@ -784,8 +1156,22 @@ class BasePlatformAdapter(ABC):
 
     async def _process_message_background(self, event: MessageEvent, session_key: str) -> None:
         """Background task that actually processes the message."""
-        # Create interrupt event for this session
-        interrupt_event = asyncio.Event()
+        # Track delivery outcomes for the processing-complete hook
+        delivery_attempted = False
+        delivery_succeeded = False
+
+        def _record_delivery(result):
+            nonlocal delivery_attempted, delivery_succeeded
+            if result is None:
+                return
+            delivery_attempted = True
+            if getattr(result, "success", False):
+                delivery_succeeded = True
+
+        # Reuse the interrupt event set by handle_message() (which marks
+        # the session active before spawning this task to prevent races).
+        # Fall back to a new Event only if the entry was removed externally.
+        interrupt_event = self._active_sessions.get(session_key) or asyncio.Event()
         self._active_sessions[session_key] = interrupt_event
         
         # Start continuous typing indicator (refreshes every 2 seconds)
@@ -793,20 +1179,34 @@ class BasePlatformAdapter(ABC):
         typing_task = asyncio.create_task(self._keep_typing(event.source.chat_id, metadata=_thread_metadata))
         
         try:
+            await self._run_processing_hook("on_processing_start", event)
+
             # Call the handler (this can take a while with tool calls)
             response = await self._message_handler(event)
             
-            # Send response if any
+            # Send response if any.  A None/empty response is normal when
+            # streaming already delivered the text (already_sent=True) or
+            # when the message was queued behind an active agent.  Log at
+            # DEBUG to avoid noisy warnings for expected behavior.
             if not response:
-                logger.warning("[%s] Handler returned empty/None response for %s", self.name, event.source.chat_id)
+                logger.debug("[%s] Handler returned empty/None response for %s", self.name, event.source.chat_id)
             if response:
                 # Extract MEDIA:<path> tags (from TTS tool) before other processing
                 media_files, response = self.extract_media(response)
                 
                 # Extract image URLs and send them as native platform attachments
                 images, text_content = self.extract_images(response)
+                # Strip any remaining internal directives from message body (fixes #1561)
+                text_content = text_content.replace("[[audio_as_voice]]", "").strip()
+                text_content = re.sub(r"MEDIA:\s*\S+", "", text_content).strip()
                 if images:
                     logger.info("[%s] extract_images found %d image(s) in response (%d chars)", self.name, len(images), len(response))
+
+                # Auto-detect bare local file paths for native media delivery
+                # (helps small models that don't use MEDIA: syntax)
+                local_files, text_content = self.extract_local_files(text_content)
+                if local_files:
+                    logger.info("[%s] extract_local_files found %d file(s) in response", self.name, len(local_files))
                 
                 # Auto-TTS: if voice message, generate audio FIRST (before sending text)
                 # Skipped when the chat has voice mode disabled (/voice off)
@@ -847,25 +1247,13 @@ class BasePlatformAdapter(ABC):
                 # Send the text portion
                 if text_content:
                     logger.info("[%s] Sending response (%d chars) to %s", self.name, len(text_content), event.source.chat_id)
-                    result = await self.send(
+                    result = await self._send_with_retry(
                         chat_id=event.source.chat_id,
                         content=text_content,
                         reply_to=event.message_id,
                         metadata=_thread_metadata,
                     )
-
-                    # Log send failures (don't raise - user already saw tool progress)
-                    if not result.success:
-                        print(f"[{self.name}] Failed to send response: {result.error}")
-                        # Try sending without markdown as fallback
-                        fallback_result = await self.send(
-                            chat_id=event.source.chat_id,
-                            content=f"(Response formatting failed, plain text:)\n\n{text_content[:3500]}",
-                            reply_to=event.message_id,
-                            metadata=_thread_metadata,
-                        )
-                        if not fallback_result.success:
-                            print(f"[{self.name}] Fallback send also failed: {fallback_result.error}")
+                    _record_delivery(result)
 
                 # Human-like pacing delay between text and media
                 human_delay = self._get_human_delay()
@@ -900,7 +1288,7 @@ class BasePlatformAdapter(ABC):
 
                 # Send extracted media files — route by file type
                 _AUDIO_EXTS = {'.ogg', '.opus', '.mp3', '.wav', '.m4a'}
-                _VIDEO_EXTS = {'.mp4', '.mov', '.avi', '.mkv', '.3gp'}
+                _VIDEO_EXTS = {'.mp4', '.mov', '.avi', '.mkv', '.webm', '.3gp'}
                 _IMAGE_EXTS = {'.jpg', '.jpeg', '.png', '.webp', '.gif'}
 
                 for media_path, is_voice in media_files:
@@ -934,14 +1322,45 @@ class BasePlatformAdapter(ABC):
                             )
 
                         if not media_result.success:
-                            print(f"[{self.name}] Failed to send media ({ext}): {media_result.error}")
+                            logger.warning("[%s] Failed to send media (%s): %s", self.name, ext, media_result.error)
                     except Exception as media_err:
-                        print(f"[{self.name}] Error sending media: {media_err}")
-            
+                        logger.warning("[%s] Error sending media: %s", self.name, media_err)
+
+                # Send auto-detected local files as native attachments
+                for file_path in local_files:
+                    if human_delay > 0:
+                        await asyncio.sleep(human_delay)
+                    try:
+                        ext = Path(file_path).suffix.lower()
+                        if ext in _IMAGE_EXTS:
+                            await self.send_image_file(
+                                chat_id=event.source.chat_id,
+                                image_path=file_path,
+                                metadata=_thread_metadata,
+                            )
+                        elif ext in _VIDEO_EXTS:
+                            await self.send_video(
+                                chat_id=event.source.chat_id,
+                                video_path=file_path,
+                                metadata=_thread_metadata,
+                            )
+                        else:
+                            await self.send_document(
+                                chat_id=event.source.chat_id,
+                                file_path=file_path,
+                                metadata=_thread_metadata,
+                            )
+                    except Exception as file_err:
+                        logger.error("[%s] Error sending local file %s: %s", self.name, file_path, file_err)
+
+            # Determine overall success for the processing hook
+            processing_ok = delivery_succeeded if delivery_attempted else not bool(response)
+            await self._run_processing_hook("on_processing_complete", event, processing_ok)
+
             # Check if there's a pending message that was queued during our processing
             if session_key in self._pending_messages:
                 pending_event = self._pending_messages.pop(session_key)
-                print(f"[{self.name}] 📨 Processing queued message from interrupt")
+                logger.debug("[%s] Processing queued message from interrupt", self.name)
                 # Clean up current session before processing pending
                 if session_key in self._active_sessions:
                     del self._active_sessions[session_key]
@@ -954,10 +1373,28 @@ class BasePlatformAdapter(ABC):
                 await self._process_message_background(pending_event, session_key)
                 return  # Already cleaned up
                 
+        except asyncio.CancelledError:
+            await self._run_processing_hook("on_processing_complete", event, False)
+            raise
         except Exception as e:
-            print(f"[{self.name}] Error handling message: {e}")
-            import traceback
-            traceback.print_exc()
+            await self._run_processing_hook("on_processing_complete", event, False)
+            logger.error("[%s] Error handling message: %s", self.name, e, exc_info=True)
+            # Send the error to the user so they aren't left with radio silence
+            try:
+                error_type = type(e).__name__
+                error_detail = str(e)[:300] if str(e) else "no details available"
+                _thread_metadata = {"thread_id": event.source.thread_id} if event.source.thread_id else None
+                await self.send(
+                    chat_id=event.source.chat_id,
+                    content=(
+                        f"Sorry, I encountered an error ({error_type}).\n"
+                        f"{error_detail}\n"
+                        "Try again or use /reset to start a fresh session."
+                    ),
+                    metadata=_thread_metadata,
+                )
+            except Exception:
+                pass  # Last resort — don't let error reporting crash the handler
         finally:
             # Stop typing indicator
             typing_task.cancel()
@@ -965,10 +1402,32 @@ class BasePlatformAdapter(ABC):
                 await typing_task
             except asyncio.CancelledError:
                 pass
+            # Also cancel any platform-level persistent typing tasks (e.g. Discord)
+            # that may have been recreated by _keep_typing after the last stop_typing()
+            try:
+                if hasattr(self, "stop_typing"):
+                    await self.stop_typing(event.source.chat_id)
+            except Exception:
+                pass
             # Clean up session tracking
             if session_key in self._active_sessions:
                 del self._active_sessions[session_key]
     
+    async def cancel_background_tasks(self) -> None:
+        """Cancel any in-flight background message-processing tasks.
+
+        Used during gateway shutdown/replacement so active sessions from the old
+        process do not keep running after adapters are being torn down.
+        """
+        tasks = [task for task in self._background_tasks if not task.done()]
+        for task in tasks:
+            task.cancel()
+        if tasks:
+            await asyncio.gather(*tasks, return_exceptions=True)
+        self._background_tasks.clear()
+        self._pending_messages.clear()
+        self._active_sessions.clear()
+
     def has_pending_interrupt(self, session_key: str) -> bool:
         """Check if there's a pending interrupt for a session."""
         return session_key in self._active_sessions and self._active_sessions[session_key].is_set()
@@ -1028,7 +1487,8 @@ class BasePlatformAdapter(ABC):
         """
         return content
     
-    def truncate_message(self, content: str, max_length: int = 4096) -> List[str]:
+    @staticmethod
+    def truncate_message(content: str, max_length: int = 4096) -> List[str]:
         """
         Split a long message into chunks, preserving code block boundaries.
 
@@ -1080,6 +1540,27 @@ class BasePlatformAdapter(ABC):
             if split_at < 1:
                 split_at = headroom
 
+            # Avoid splitting inside an inline code span (`...`).
+            # If the text before split_at has an odd number of unescaped
+            # backticks, the split falls inside inline code — the resulting
+            # chunk would have an unpaired backtick and any special characters
+            # (like parentheses) inside the broken span would be unescaped,
+            # causing MarkdownV2 parse errors on Telegram.
+            candidate = remaining[:split_at]
+            backtick_count = candidate.count("`") - candidate.count("\\`")
+            if backtick_count % 2 == 1:
+                # Find the last unescaped backtick and split before it
+                last_bt = candidate.rfind("`")
+                while last_bt > 0 and candidate[last_bt - 1] == "\\":
+                    last_bt = candidate.rfind("`", 0, last_bt)
+                if last_bt > 0:
+                    # Try to find a space or newline just before the backtick
+                    safe_split = candidate.rfind(" ", 0, last_bt)
+                    nl_split = candidate.rfind("\n", 0, last_bt)
+                    safe_split = max(safe_split, nl_split)
+                    if safe_split > headroom // 4:
+                        split_at = safe_split
+
             chunk_body = remaining[:split_at]
             remaining = remaining[split_at:].lstrip()
 

gateway/platforms/dingtalk.py

@@ -0,0 +1,340 @@
+"""
+DingTalk platform adapter using Stream Mode.
+
+Uses dingtalk-stream SDK for real-time message reception without webhooks.
+Responses are sent via DingTalk's session webhook (markdown format).
+
+Requires:
+    pip install dingtalk-stream httpx
+    DINGTALK_CLIENT_ID and DINGTALK_CLIENT_SECRET env vars
+
+Configuration in config.yaml:
+    platforms:
+      dingtalk:
+        enabled: true
+        extra:
+          client_id: "your-app-key"      # or DINGTALK_CLIENT_ID env var
+          client_secret: "your-secret"   # or DINGTALK_CLIENT_SECRET env var
+"""
+
+import asyncio
+import logging
+import os
+import time
+import uuid
+from datetime import datetime, timezone
+from typing import Any, Dict, Optional
+
+try:
+    import dingtalk_stream
+    from dingtalk_stream import ChatbotHandler, ChatbotMessage
+    DINGTALK_STREAM_AVAILABLE = True
+except ImportError:
+    DINGTALK_STREAM_AVAILABLE = False
+    dingtalk_stream = None  # type: ignore[assignment]
+
+try:
+    import httpx
+    HTTPX_AVAILABLE = True
+except ImportError:
+    HTTPX_AVAILABLE = False
+    httpx = None  # type: ignore[assignment]
+
+from gateway.config import Platform, PlatformConfig
+from gateway.platforms.base import (
+    BasePlatformAdapter,
+    MessageEvent,
+    MessageType,
+    SendResult,
+)
+
+logger = logging.getLogger(__name__)
+
+MAX_MESSAGE_LENGTH = 20000
+DEDUP_WINDOW_SECONDS = 300
+DEDUP_MAX_SIZE = 1000
+RECONNECT_BACKOFF = [2, 5, 10, 30, 60]
+
+
+def check_dingtalk_requirements() -> bool:
+    """Check if DingTalk dependencies are available and configured."""
+    if not DINGTALK_STREAM_AVAILABLE or not HTTPX_AVAILABLE:
+        return False
+    if not os.getenv("DINGTALK_CLIENT_ID") or not os.getenv("DINGTALK_CLIENT_SECRET"):
+        return False
+    return True
+
+
+class DingTalkAdapter(BasePlatformAdapter):
+    """DingTalk chatbot adapter using Stream Mode.
+
+    The dingtalk-stream SDK maintains a long-lived WebSocket connection.
+    Incoming messages arrive via a ChatbotHandler callback. Replies are
+    sent via the incoming message's session_webhook URL using httpx.
+    """
+
+    MAX_MESSAGE_LENGTH = MAX_MESSAGE_LENGTH
+
+    def __init__(self, config: PlatformConfig):
+        super().__init__(config, Platform.DINGTALK)
+
+        extra = config.extra or {}
+        self._client_id: str = extra.get("client_id") or os.getenv("DINGTALK_CLIENT_ID", "")
+        self._client_secret: str = extra.get("client_secret") or os.getenv("DINGTALK_CLIENT_SECRET", "")
+
+        self._stream_client: Any = None
+        self._stream_task: Optional[asyncio.Task] = None
+        self._http_client: Optional["httpx.AsyncClient"] = None
+
+        # Message deduplication: msg_id -> timestamp
+        self._seen_messages: Dict[str, float] = {}
+        # Map chat_id -> session_webhook for reply routing
+        self._session_webhooks: Dict[str, str] = {}
+
+    # -- Connection lifecycle -----------------------------------------------
+
+    async def connect(self) -> bool:
+        """Connect to DingTalk via Stream Mode."""
+        if not DINGTALK_STREAM_AVAILABLE:
+            logger.warning("[%s] dingtalk-stream not installed. Run: pip install dingtalk-stream", self.name)
+            return False
+        if not HTTPX_AVAILABLE:
+            logger.warning("[%s] httpx not installed. Run: pip install httpx", self.name)
+            return False
+        if not self._client_id or not self._client_secret:
+            logger.warning("[%s] DINGTALK_CLIENT_ID and DINGTALK_CLIENT_SECRET required", self.name)
+            return False
+
+        try:
+            self._http_client = httpx.AsyncClient(timeout=30.0)
+
+            credential = dingtalk_stream.Credential(self._client_id, self._client_secret)
+            self._stream_client = dingtalk_stream.DingTalkStreamClient(credential)
+
+            # Capture the current event loop for cross-thread dispatch
+            loop = asyncio.get_running_loop()
+            handler = _IncomingHandler(self, loop)
+            self._stream_client.register_callback_handler(
+                dingtalk_stream.ChatbotMessage.TOPIC, handler
+            )
+
+            self._stream_task = asyncio.create_task(self._run_stream())
+            self._mark_connected()
+            logger.info("[%s] Connected via Stream Mode", self.name)
+            return True
+        except Exception as e:
+            logger.error("[%s] Failed to connect: %s", self.name, e)
+            return False
+
+    async def _run_stream(self) -> None:
+        """Run the blocking stream client with auto-reconnection."""
+        backoff_idx = 0
+        while self._running:
+            try:
+                logger.debug("[%s] Starting stream client...", self.name)
+                await asyncio.to_thread(self._stream_client.start)
+            except asyncio.CancelledError:
+                return
+            except Exception as e:
+                if not self._running:
+                    return
+                logger.warning("[%s] Stream client error: %s", self.name, e)
+
+            if not self._running:
+                return
+
+            delay = RECONNECT_BACKOFF[min(backoff_idx, len(RECONNECT_BACKOFF) - 1)]
+            logger.info("[%s] Reconnecting in %ds...", self.name, delay)
+            await asyncio.sleep(delay)
+            backoff_idx += 1
+
+    async def disconnect(self) -> None:
+        """Disconnect from DingTalk."""
+        self._running = False
+        self._mark_disconnected()
+
+        if self._stream_task:
+            self._stream_task.cancel()
+            try:
+                await self._stream_task
+            except asyncio.CancelledError:
+                pass
+            self._stream_task = None
+
+        if self._http_client:
+            await self._http_client.aclose()
+            self._http_client = None
+
+        self._stream_client = None
+        self._session_webhooks.clear()
+        self._seen_messages.clear()
+        logger.info("[%s] Disconnected", self.name)
+
+    # -- Inbound message processing -----------------------------------------
+
+    async def _on_message(self, message: "ChatbotMessage") -> None:
+        """Process an incoming DingTalk chatbot message."""
+        msg_id = getattr(message, "message_id", None) or uuid.uuid4().hex
+        if self._is_duplicate(msg_id):
+            logger.debug("[%s] Duplicate message %s, skipping", self.name, msg_id)
+            return
+
+        text = self._extract_text(message)
+        if not text:
+            logger.debug("[%s] Empty message, skipping", self.name)
+            return
+
+        # Chat context
+        conversation_id = getattr(message, "conversation_id", "") or ""
+        conversation_type = getattr(message, "conversation_type", "1")
+        is_group = str(conversation_type) == "2"
+        sender_id = getattr(message, "sender_id", "") or ""
+        sender_nick = getattr(message, "sender_nick", "") or sender_id
+        sender_staff_id = getattr(message, "sender_staff_id", "") or ""
+
+        chat_id = conversation_id or sender_id
+        chat_type = "group" if is_group else "dm"
+
+        # Store session webhook for reply routing
+        session_webhook = getattr(message, "session_webhook", None) or ""
+        if session_webhook and chat_id:
+            self._session_webhooks[chat_id] = session_webhook
+
+        source = self.build_source(
+            chat_id=chat_id,
+            chat_name=getattr(message, "conversation_title", None),
+            chat_type=chat_type,
+            user_id=sender_id,
+            user_name=sender_nick,
+            user_id_alt=sender_staff_id if sender_staff_id else None,
+        )
+
+        # Parse timestamp
+        create_at = getattr(message, "create_at", None)
+        try:
+            timestamp = datetime.fromtimestamp(int(create_at) / 1000, tz=timezone.utc) if create_at else datetime.now(tz=timezone.utc)
+        except (ValueError, OSError, TypeError):
+            timestamp = datetime.now(tz=timezone.utc)
+
+        event = MessageEvent(
+            text=text,
+            message_type=MessageType.TEXT,
+            source=source,
+            message_id=msg_id,
+            raw_message=message,
+            timestamp=timestamp,
+        )
+
+        logger.debug("[%s] Message from %s in %s: %s",
+                      self.name, sender_nick, chat_id[:20] if chat_id else "?", text[:50])
+        await self.handle_message(event)
+
+    @staticmethod
+    def _extract_text(message: "ChatbotMessage") -> str:
+        """Extract plain text from a DingTalk chatbot message."""
+        text = getattr(message, "text", None) or ""
+        if isinstance(text, dict):
+            content = text.get("content", "").strip()
+        else:
+            content = str(text).strip()
+
+        # Fall back to rich text if present
+        if not content:
+            rich_text = getattr(message, "rich_text", None)
+            if rich_text and isinstance(rich_text, list):
+                parts = [item["text"] for item in rich_text
+                         if isinstance(item, dict) and item.get("text")]
+                content = " ".join(parts).strip()
+        return content
+
+    # -- Deduplication ------------------------------------------------------
+
+    def _is_duplicate(self, msg_id: str) -> bool:
+        """Check and record a message ID. Returns True if already seen."""
+        now = time.time()
+        if len(self._seen_messages) > DEDUP_MAX_SIZE:
+            cutoff = now - DEDUP_WINDOW_SECONDS
+            self._seen_messages = {k: v for k, v in self._seen_messages.items() if v > cutoff}
+
+        if msg_id in self._seen_messages:
+            return True
+        self._seen_messages[msg_id] = now
+        return False
+
+    # -- Outbound messaging -------------------------------------------------
+
+    async def send(
+        self,
+        chat_id: str,
+        content: str,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        """Send a markdown reply via DingTalk session webhook."""
+        metadata = metadata or {}
+
+        session_webhook = metadata.get("session_webhook") or self._session_webhooks.get(chat_id)
+        if not session_webhook:
+            return SendResult(success=False,
+                              error="No session_webhook available. Reply must follow an incoming message.")
+
+        if not self._http_client:
+            return SendResult(success=False, error="HTTP client not initialized")
+
+        payload = {
+            "msgtype": "markdown",
+            "markdown": {"title": "Hermes", "text": content[:self.MAX_MESSAGE_LENGTH]},
+        }
+
+        try:
+            resp = await self._http_client.post(session_webhook, json=payload, timeout=15.0)
+            if resp.status_code < 300:
+                return SendResult(success=True, message_id=uuid.uuid4().hex[:12])
+            body = resp.text
+            logger.warning("[%s] Send failed HTTP %d: %s", self.name, resp.status_code, body[:200])
+            return SendResult(success=False, error=f"HTTP {resp.status_code}: {body[:200]}")
+        except httpx.TimeoutException:
+            return SendResult(success=False, error="Timeout sending message to DingTalk")
+        except Exception as e:
+            logger.error("[%s] Send error: %s", self.name, e)
+            return SendResult(success=False, error=str(e))
+
+    async def send_typing(self, chat_id: str, metadata=None) -> None:
+        """DingTalk does not support typing indicators."""
+        pass
+
+    async def get_chat_info(self, chat_id: str) -> Dict[str, Any]:
+        """Return basic info about a DingTalk conversation."""
+        return {"name": chat_id, "type": "group" if "group" in chat_id.lower() else "dm"}
+
+
+# ---------------------------------------------------------------------------
+# Internal stream handler
+# ---------------------------------------------------------------------------
+
+class _IncomingHandler(ChatbotHandler if DINGTALK_STREAM_AVAILABLE else object):
+    """dingtalk-stream ChatbotHandler that forwards messages to the adapter."""
+
+    def __init__(self, adapter: DingTalkAdapter, loop: asyncio.AbstractEventLoop):
+        if DINGTALK_STREAM_AVAILABLE:
+            super().__init__()
+        self._adapter = adapter
+        self._loop = loop
+
+    def process(self, message: "ChatbotMessage"):
+        """Called by dingtalk-stream in its thread when a message arrives.
+
+        Schedules the async handler on the main event loop.
+        """
+        loop = self._loop
+        if loop is None or loop.is_closed():
+            logger.error("[DingTalk] Event loop unavailable, cannot dispatch message")
+            return dingtalk_stream.AckMessage.STATUS_OK, "OK"
+
+        future = asyncio.run_coroutine_threadsafe(self._adapter._on_message(message), loop)
+        try:
+            future.result(timeout=60)
+        except Exception:
+            logger.exception("[DingTalk] Error processing incoming message")
+
+        return dingtalk_stream.AckMessage.STATUS_OK, "OK"

gateway/platforms/email.py

@@ -24,7 +24,6 @@ import re
 import smtplib
 import ssl
 import uuid
-from datetime import datetime
 from email.header import decode_header
 from email.mime.multipart import MIMEMultipart
 from email.mime.text import MIMEText
@@ -44,6 +43,20 @@ from gateway.platforms.base import (
 from gateway.config import Platform, PlatformConfig
 
 logger = logging.getLogger(__name__)
+# Automated sender patterns — emails from these are silently ignored
+_NOREPLY_PATTERNS = (
+    "noreply", "no-reply", "no_reply", "donotreply", "do-not-reply",
+    "mailer-daemon", "postmaster", "bounce", "notifications@",
+    "automated@", "auto-confirm", "auto-reply", "automailer",
+)
+
+# RFC headers that indicate bulk/automated mail
+_AUTOMATED_HEADERS = {
+    "Auto-Submitted": lambda v: v.lower() != "no",
+    "Precedence": lambda v: v.lower() in ("bulk", "list", "junk"),
+    "X-Auto-Response-Suppress": lambda v: bool(v),
+    "List-Unsubscribe": lambda v: bool(v),
+}
 
 # Gmail-safe max length per email body
 MAX_MESSAGE_LENGTH = 50_000
@@ -51,7 +64,17 @@ MAX_MESSAGE_LENGTH = 50_000
 # Supported image extensions for inline detection
 _IMAGE_EXTS = {".jpg", ".jpeg", ".png", ".gif", ".webp"}
 
-
+def _is_automated_sender(address: str, headers: dict) -> bool:
+    """Return True if this email is from an automated/noreply source."""
+    addr = address.lower()
+    if any(pattern in addr for pattern in _NOREPLY_PATTERNS):
+        return True
+    for header, check in _AUTOMATED_HEADERS.items():
+        value = headers.get(header, "")
+        if value and check(value):
+            return True
+    return False
+    
 def check_email_requirements() -> bool:
     """Check if email platform dependencies are available."""
     addr = os.getenv("EMAIL_ADDRESS")
@@ -135,14 +158,23 @@ def _extract_email_address(raw: str) -> str:
     return raw.strip().lower()
 
 
-def _extract_attachments(msg: email_lib.message.Message) -> List[Dict[str, Any]]:
-    """Extract attachment metadata and cache files locally."""
+def _extract_attachments(
+    msg: email_lib.message.Message,
+    skip_attachments: bool = False,
+) -> List[Dict[str, Any]]:
+    """Extract attachment metadata and cache files locally.
+
+    When *skip_attachments* is True, all attachment/inline parts are ignored
+    (useful for malware protection or bandwidth savings).
+    """
     attachments = []
     if not msg.is_multipart():
         return attachments
 
     for part in msg.walk():
         disposition = str(part.get("Content-Disposition", ""))
+        if skip_attachments and ("attachment" in disposition or "inline" in disposition):
+            continue
         if "attachment" not in disposition and "inline" not in disposition:
             continue
         # Skip text/plain and text/html body parts
@@ -196,8 +228,16 @@ class EmailAdapter(BasePlatformAdapter):
         self._smtp_port = int(os.getenv("EMAIL_SMTP_PORT", "587"))
         self._poll_interval = int(os.getenv("EMAIL_POLL_INTERVAL", "15"))
 
+        # Skip attachments — configured via config.yaml:
+        #   platforms:
+        #     email:
+        #       skip_attachments: true
+        extra = config.extra or {}
+        self._skip_attachments = extra.get("skip_attachments", False)
+
         # Track message IDs we've already processed to avoid duplicates
         self._seen_uids: set = set()
+        self._seen_uids_max: int = 2000   # cap to prevent unbounded memory growth
         self._poll_task: Optional[asyncio.Task] = None
 
         # Map chat_id (sender email) -> last subject + message-id for threading
@@ -205,18 +245,40 @@ class EmailAdapter(BasePlatformAdapter):
 
         logger.info("[Email] Adapter initialized for %s", self._address)
 
+    def _trim_seen_uids(self) -> None:
+        """Keep only the most recent UIDs to prevent unbounded memory growth.
+
+        IMAP UIDs are monotonically increasing integers. When the set grows
+        beyond the cap, we keep only the highest half — old UIDs are safe to
+        drop because new messages always have higher UIDs and IMAP's UNSEEN
+        flag prevents re-delivery regardless.
+        """
+        if len(self._seen_uids) <= self._seen_uids_max:
+            return
+        try:
+            # UIDs are bytes like b'1234' — sort numerically and keep top half
+            sorted_uids = sorted(self._seen_uids, key=lambda u: int(u))
+            keep = self._seen_uids_max // 2
+            self._seen_uids = set(sorted_uids[-keep:])
+            logger.debug("[Email] Trimmed seen UIDs to %d entries", len(self._seen_uids))
+        except (ValueError, TypeError):
+            # Fallback: just clear old entries if sort fails
+            self._seen_uids = set(list(self._seen_uids)[-self._seen_uids_max // 2:])
+
     async def connect(self) -> bool:
         """Connect to the IMAP server and start polling for new messages."""
         try:
             # Test IMAP connection
-            imap = imaplib.IMAP4_SSL(self._imap_host, self._imap_port)
+            imap = imaplib.IMAP4_SSL(self._imap_host, self._imap_port, timeout=30)
             imap.login(self._address, self._password)
             # Mark all existing messages as seen so we only process new ones
             imap.select("INBOX")
             status, data = imap.uid("search", None, "ALL")
-            if status == "OK" and data[0]:
+            if status == "OK" and data and data[0]:
                 for uid in data[0].split():
                     self._seen_uids.add(uid)
+            # Keep only the most recent UIDs to prevent unbounded growth
+            self._trim_seen_uids()
             imap.logout()
             logger.info("[Email] IMAP connection test passed. %d existing messages skipped.", len(self._seen_uids))
         except Exception as e:
@@ -225,7 +287,7 @@ class EmailAdapter(BasePlatformAdapter):
 
         try:
             # Test SMTP connection
-            smtp = smtplib.SMTP(self._smtp_host, self._smtp_port)
+            smtp = smtplib.SMTP(self._smtp_host, self._smtp_port, timeout=30)
             smtp.starttls(context=ssl.create_default_context())
             smtp.login(self._address, self._password)
             smtp.quit()
@@ -274,53 +336,64 @@ class EmailAdapter(BasePlatformAdapter):
         """Fetch new (unseen) messages from IMAP. Runs in executor thread."""
         results = []
         try:
-            imap = imaplib.IMAP4_SSL(self._imap_host, self._imap_port)
-            imap.login(self._address, self._password)
-            imap.select("INBOX")
+            imap = imaplib.IMAP4_SSL(self._imap_host, self._imap_port, timeout=30)
+            try:
+                imap.login(self._address, self._password)
+                imap.select("INBOX")
 
-            status, data = imap.uid("search", None, "UNSEEN")
-            if status != "OK" or not data[0]:
-                imap.logout()
-                return results
+                status, data = imap.uid("search", None, "UNSEEN")
+                if status != "OK" or not data or not data[0]:
+                    return results
 
-            for uid in data[0].split():
-                if uid in self._seen_uids:
-                    continue
-                self._seen_uids.add(uid)
+                for uid in data[0].split():
+                    if uid in self._seen_uids:
+                        continue
+                    self._seen_uids.add(uid)
+                    # Trim periodically to prevent unbounded memory growth
+                    if len(self._seen_uids) > self._seen_uids_max:
+                        self._trim_seen_uids()
 
-                status, msg_data = imap.uid("fetch", uid, "(RFC822)")
-                if status != "OK":
-                    continue
+                    status, msg_data = imap.uid("fetch", uid, "(RFC822)")
+                    if status != "OK":
+                        continue
 
-                raw_email = msg_data[0][1]
-                msg = email_lib.message_from_bytes(raw_email)
+                    raw_email = msg_data[0][1]
+                    msg = email_lib.message_from_bytes(raw_email)
 
-                sender_raw = msg.get("From", "")
-                sender_addr = _extract_email_address(sender_raw)
-                sender_name = _decode_header_value(sender_raw)
-                # Remove email from name if present
-                if "<" in sender_name:
-                    sender_name = sender_name.split("<")[0].strip().strip('"')
+                    sender_raw = msg.get("From", "")
+                    sender_addr = _extract_email_address(sender_raw)
+                    sender_name = _decode_header_value(sender_raw)
+                    # Remove email from name if present
+                    if "<" in sender_name:
+                        sender_name = sender_name.split("<")[0].strip().strip('"')
 
-                subject = _decode_header_value(msg.get("Subject", "(no subject)"))
-                message_id = msg.get("Message-ID", "")
-                in_reply_to = msg.get("In-Reply-To", "")
-                body = _extract_text_body(msg)
-                attachments = _extract_attachments(msg)
+                    subject = _decode_header_value(msg.get("Subject", "(no subject)"))
+                    message_id = msg.get("Message-ID", "")
+                    in_reply_to = msg.get("In-Reply-To", "")
+                    # Skip automated/noreply senders before any processing
+                    msg_headers = dict(msg.items())
+                    if _is_automated_sender(sender_addr, msg_headers):
+                        logger.debug("[Email] Skipping automated sender: %s", sender_addr)
+                        continue
+                    body = _extract_text_body(msg)
+                    attachments = _extract_attachments(msg, skip_attachments=self._skip_attachments)
 
-                results.append({
-                    "uid": uid,
-                    "sender_addr": sender_addr,
-                    "sender_name": sender_name,
-                    "subject": subject,
-                    "message_id": message_id,
-                    "in_reply_to": in_reply_to,
-                    "body": body,
-                    "attachments": attachments,
-                    "date": msg.get("Date", ""),
-                })
-
-            imap.logout()
+                    results.append({
+                        "uid": uid,
+                        "sender_addr": sender_addr,
+                        "sender_name": sender_name,
+                        "subject": subject,
+                        "message_id": message_id,
+                        "in_reply_to": in_reply_to,
+                        "body": body,
+                        "attachments": attachments,
+                        "date": msg.get("Date", ""),
+                    })
+            finally:
+                try:
+                    imap.logout()
+                except Exception:
+                    pass
         except Exception as e:
             logger.error("[Email] IMAP fetch error: %s", e)
         return results
@@ -333,6 +406,11 @@ class EmailAdapter(BasePlatformAdapter):
         if sender_addr == self._address.lower():
             return
 
+        # Never reply to automated senders
+        if _is_automated_sender(sender_addr, {}):
+            logger.debug("[Email] Dropping automated sender at dispatch: %s", sender_addr)
+            return
+
         subject = msg_data["subject"]
         body = msg_data["body"].strip()
         attachments = msg_data["attachments"]
@@ -427,18 +505,22 @@ class EmailAdapter(BasePlatformAdapter):
 
         msg.attach(MIMEText(body, "plain", "utf-8"))
 
-        smtp = smtplib.SMTP(self._smtp_host, self._smtp_port)
-        smtp.starttls(context=ssl.create_default_context())
-        smtp.login(self._address, self._password)
-        smtp.send_message(msg)
-        smtp.quit()
+        smtp = smtplib.SMTP(self._smtp_host, self._smtp_port, timeout=30)
+        try:
+            smtp.starttls(context=ssl.create_default_context())
+            smtp.login(self._address, self._password)
+            smtp.send_message(msg)
+        finally:
+            try:
+                smtp.quit()
+            except Exception:
+                smtp.close()
 
         logger.info("[Email] Sent reply to %s (subject: %s)", to_addr, subject)
         return msg_id
 
-    async def send_typing(self, chat_id: str) -> None:
+    async def send_typing(self, chat_id: str, metadata: Optional[Dict[str, Any]] = None) -> None:
         """Email has no typing indicator — no-op."""
-        pass
 
     async def send_image(
         self,
@@ -515,11 +597,16 @@ class EmailAdapter(BasePlatformAdapter):
             part.add_header("Content-Disposition", f"attachment; filename={fname}")
             msg.attach(part)
 
-        smtp = smtplib.SMTP(self._smtp_host, self._smtp_port)
-        smtp.starttls(context=ssl.create_default_context())
-        smtp.login(self._address, self._password)
-        smtp.send_message(msg)
-        smtp.quit()
+        smtp = smtplib.SMTP(self._smtp_host, self._smtp_port, timeout=30)
+        try:
+            smtp.starttls(context=ssl.create_default_context())
+            smtp.login(self._address, self._password)
+            smtp.send_message(msg)
+        finally:
+            try:
+                smtp.quit()
+            except Exception:
+                smtp.close()
 
         return msg_id
 

gateway/platforms/homeassistant.py

@@ -19,7 +19,7 @@ import os
 import time
 import uuid
 from datetime import datetime
-from typing import Any, Dict, List, Optional, Set
+from typing import Any, Dict, Optional, Set
 
 try:
     import aiohttp
@@ -114,7 +114,9 @@ class HomeAssistantAdapter(BasePlatformAdapter):
                 return False
 
             # Dedicated REST session for send() calls
-            self._rest_session = aiohttp.ClientSession()
+            self._rest_session = aiohttp.ClientSession(
+                timeout=aiohttp.ClientTimeout(total=30)
+            )
 
             # Warn if no event filters are configured
             if not self._watch_domains and not self._watch_entities and not self._watch_all:
@@ -140,8 +142,10 @@ class HomeAssistantAdapter(BasePlatformAdapter):
         ws_url = self._hass_url.replace("http://", "ws://").replace("https://", "wss://")
         ws_url = f"{ws_url}/api/websocket"
 
-        self._session = aiohttp.ClientSession()
-        self._ws = await self._session.ws_connect(ws_url, heartbeat=30)
+        self._session = aiohttp.ClientSession(
+            timeout=aiohttp.ClientTimeout(total=30)
+        )
+        self._ws = await self._session.ws_connect(ws_url, heartbeat=30, timeout=30)
 
         # Step 1: Receive auth_required
         msg = await self._ws.receive_json()
@@ -435,7 +439,6 @@ class HomeAssistantAdapter(BasePlatformAdapter):
 
     async def send_typing(self, chat_id: str, metadata=None) -> None:
         """No typing indicator for Home Assistant."""
-        pass
 
     async def get_chat_info(self, chat_id: str) -> Dict[str, Any]:
         """Return basic info about the HA event channel."""

gateway/platforms/mattermost.py

@@ -0,0 +1,733 @@
+"""Mattermost gateway adapter.
+
+Connects to a self-hosted (or cloud) Mattermost instance via its REST API
+(v4) and WebSocket for real-time events.  No external Mattermost library
+required — uses aiohttp which is already a Hermes dependency.
+
+Environment variables:
+    MATTERMOST_URL              Server URL (e.g. https://mm.example.com)
+    MATTERMOST_TOKEN            Bot token or personal-access token
+    MATTERMOST_ALLOWED_USERS    Comma-separated user IDs
+    MATTERMOST_HOME_CHANNEL     Channel ID for cron/notification delivery
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import os
+import re
+import time
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+from gateway.config import Platform, PlatformConfig
+from gateway.platforms.base import (
+    BasePlatformAdapter,
+    MessageEvent,
+    MessageType,
+    SendResult,
+)
+
+logger = logging.getLogger(__name__)
+
+# Mattermost post size limit (server default is 16383, but 4000 is the
+# practical limit for readable messages — matching OpenClaw's choice).
+MAX_POST_LENGTH = 4000
+
+# Channel type codes returned by the Mattermost API.
+_CHANNEL_TYPE_MAP = {
+    "D": "dm",
+    "G": "group",
+    "P": "group",   # private channel → treat as group
+    "O": "channel",
+}
+
+# Reconnect parameters (exponential backoff).
+_RECONNECT_BASE_DELAY = 2.0
+_RECONNECT_MAX_DELAY = 60.0
+_RECONNECT_JITTER = 0.2
+
+
+def check_mattermost_requirements() -> bool:
+    """Return True if the Mattermost adapter can be used."""
+    token = os.getenv("MATTERMOST_TOKEN", "")
+    url = os.getenv("MATTERMOST_URL", "")
+    if not token:
+        logger.debug("Mattermost: MATTERMOST_TOKEN not set")
+        return False
+    if not url:
+        logger.warning("Mattermost: MATTERMOST_URL not set")
+        return False
+    try:
+        import aiohttp  # noqa: F401
+        return True
+    except ImportError:
+        logger.warning("Mattermost: aiohttp not installed")
+        return False
+
+
+class MattermostAdapter(BasePlatformAdapter):
+    """Gateway adapter for Mattermost (self-hosted or cloud)."""
+
+    def __init__(self, config: PlatformConfig):
+        super().__init__(config, Platform.MATTERMOST)
+
+        self._base_url: str = (
+            config.extra.get("url", "")
+            or os.getenv("MATTERMOST_URL", "")
+        ).rstrip("/")
+        self._token: str = config.token or os.getenv("MATTERMOST_TOKEN", "")
+
+        self._bot_user_id: str = ""
+        self._bot_username: str = ""
+
+        # aiohttp session + websocket handle
+        self._session: Any = None  # aiohttp.ClientSession
+        self._ws: Any = None       # aiohttp.ClientWebSocketResponse
+        self._ws_task: Optional[asyncio.Task] = None
+        self._reconnect_task: Optional[asyncio.Task] = None
+        self._closing = False
+
+        # Reply mode: "thread" to nest replies, "off" for flat messages.
+        self._reply_mode: str = (
+            config.extra.get("reply_mode", "")
+            or os.getenv("MATTERMOST_REPLY_MODE", "off")
+        ).lower()
+
+        # Dedup cache: post_id → timestamp (prevent reprocessing)
+        self._seen_posts: Dict[str, float] = {}
+        self._SEEN_MAX = 2000
+        self._SEEN_TTL = 300  # 5 minutes
+
+    # ------------------------------------------------------------------
+    # HTTP helpers
+    # ------------------------------------------------------------------
+
+    def _headers(self) -> Dict[str, str]:
+        return {
+            "Authorization": f"Bearer {self._token}",
+            "Content-Type": "application/json",
+        }
+
+    async def _api_get(self, path: str) -> Dict[str, Any]:
+        """GET /api/v4/{path}."""
+        import aiohttp
+        url = f"{self._base_url}/api/v4/{path.lstrip('/')}"
+        try:
+            async with self._session.get(url, headers=self._headers(), timeout=aiohttp.ClientTimeout(total=30)) as resp:
+                if resp.status >= 400:
+                    body = await resp.text()
+                    logger.error("MM API GET %s → %s: %s", path, resp.status, body[:200])
+                    return {}
+                return await resp.json()
+        except aiohttp.ClientError as exc:
+            logger.error("MM API GET %s network error: %s", path, exc)
+            return {}
+
+    async def _api_post(
+        self, path: str, payload: Dict[str, Any]
+    ) -> Dict[str, Any]:
+        """POST /api/v4/{path} with JSON body."""
+        import aiohttp
+        url = f"{self._base_url}/api/v4/{path.lstrip('/')}"
+        try:
+            async with self._session.post(
+                url, headers=self._headers(), json=payload,
+                timeout=aiohttp.ClientTimeout(total=30)
+            ) as resp:
+                if resp.status >= 400:
+                    body = await resp.text()
+                    logger.error("MM API POST %s → %s: %s", path, resp.status, body[:200])
+                    return {}
+                return await resp.json()
+        except aiohttp.ClientError as exc:
+            logger.error("MM API POST %s network error: %s", path, exc)
+            return {}
+
+    async def _api_put(
+        self, path: str, payload: Dict[str, Any]
+    ) -> Dict[str, Any]:
+        """PUT /api/v4/{path} with JSON body."""
+        import aiohttp
+        url = f"{self._base_url}/api/v4/{path.lstrip('/')}"
+        try:
+            async with self._session.put(
+                url, headers=self._headers(), json=payload
+            ) as resp:
+                if resp.status >= 400:
+                    body = await resp.text()
+                    logger.error("MM API PUT %s → %s: %s", path, resp.status, body[:200])
+                    return {}
+                return await resp.json()
+        except aiohttp.ClientError as exc:
+            logger.error("MM API PUT %s network error: %s", path, exc)
+            return {}
+
+    async def _upload_file(
+        self, channel_id: str, file_data: bytes, filename: str, content_type: str = "application/octet-stream"
+    ) -> Optional[str]:
+        """Upload a file and return its file ID, or None on failure."""
+        import aiohttp
+
+        url = f"{self._base_url}/api/v4/files"
+        form = aiohttp.FormData()
+        form.add_field("channel_id", channel_id)
+        form.add_field(
+            "files",
+            file_data,
+            filename=filename,
+            content_type=content_type,
+        )
+        headers = {"Authorization": f"Bearer {self._token}"}
+        async with self._session.post(url, headers=headers, data=form, timeout=aiohttp.ClientTimeout(total=60)) as resp:
+            if resp.status >= 400:
+                body = await resp.text()
+                logger.error("MM file upload → %s: %s", resp.status, body[:200])
+                return None
+            data = await resp.json()
+            infos = data.get("file_infos", [])
+            return infos[0]["id"] if infos else None
+
+    # ------------------------------------------------------------------
+    # Required overrides
+    # ------------------------------------------------------------------
+
+    async def connect(self) -> bool:
+        """Connect to Mattermost and start the WebSocket listener."""
+        import aiohttp
+
+        if not self._base_url or not self._token:
+            logger.error("Mattermost: URL or token not configured")
+            return False
+
+        self._session = aiohttp.ClientSession(
+            timeout=aiohttp.ClientTimeout(total=30)
+        )
+        self._closing = False
+
+        # Verify credentials and fetch bot identity.
+        me = await self._api_get("users/me")
+        if not me or "id" not in me:
+            logger.error("Mattermost: failed to authenticate — check MATTERMOST_TOKEN and MATTERMOST_URL")
+            await self._session.close()
+            return False
+
+        self._bot_user_id = me["id"]
+        self._bot_username = me.get("username", "")
+        logger.info(
+            "Mattermost: authenticated as @%s (%s) on %s",
+            self._bot_username,
+            self._bot_user_id,
+            self._base_url,
+        )
+
+        # Start WebSocket in background.
+        self._ws_task = asyncio.create_task(self._ws_loop())
+        self._mark_connected()
+        return True
+
+    async def disconnect(self) -> None:
+        """Disconnect from Mattermost."""
+        self._closing = True
+
+        if self._ws_task and not self._ws_task.done():
+            self._ws_task.cancel()
+            try:
+                await self._ws_task
+            except (asyncio.CancelledError, Exception):
+                pass
+
+        if self._reconnect_task and not self._reconnect_task.done():
+            self._reconnect_task.cancel()
+
+        if self._ws:
+            await self._ws.close()
+            self._ws = None
+
+        if self._session and not self._session.closed:
+            await self._session.close()
+
+        logger.info("Mattermost: disconnected")
+
+    async def send(
+        self,
+        chat_id: str,
+        content: str,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        """Send a message (or multiple chunks) to a channel."""
+        if not content:
+            return SendResult(success=True)
+
+        formatted = self.format_message(content)
+        chunks = self.truncate_message(formatted, MAX_POST_LENGTH)
+
+        last_id = None
+        for chunk in chunks:
+            payload: Dict[str, Any] = {
+                "channel_id": chat_id,
+                "message": chunk,
+            }
+            # Thread support: reply_to is the root post ID.
+            if reply_to and self._reply_mode == "thread":
+                payload["root_id"] = reply_to
+
+            data = await self._api_post("posts", payload)
+            if not data or "id" not in data:
+                return SendResult(success=False, error="Failed to create post")
+            last_id = data["id"]
+
+        return SendResult(success=True, message_id=last_id)
+
+    async def get_chat_info(self, chat_id: str) -> Dict[str, Any]:
+        """Return channel name and type."""
+        data = await self._api_get(f"channels/{chat_id}")
+        if not data:
+            return {"name": chat_id, "type": "channel"}
+
+        ch_type = _CHANNEL_TYPE_MAP.get(data.get("type", "O"), "channel")
+        display_name = data.get("display_name") or data.get("name") or chat_id
+        return {"name": display_name, "type": ch_type}
+
+    # ------------------------------------------------------------------
+    # Optional overrides
+    # ------------------------------------------------------------------
+
+    async def send_typing(
+        self, chat_id: str, metadata: Optional[Dict[str, Any]] = None
+    ) -> None:
+        """Send a typing indicator."""
+        await self._api_post(
+            f"users/{self._bot_user_id}/typing",
+            {"channel_id": chat_id},
+        )
+
+    async def edit_message(
+        self, chat_id: str, message_id: str, content: str
+    ) -> SendResult:
+        """Edit an existing post."""
+        formatted = self.format_message(content)
+        data = await self._api_put(
+            f"posts/{message_id}/patch",
+            {"message": formatted},
+        )
+        if not data or "id" not in data:
+            return SendResult(success=False, error="Failed to edit post")
+        return SendResult(success=True, message_id=data["id"])
+
+    async def send_image(
+        self,
+        chat_id: str,
+        image_url: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        """Download an image and upload it as a file attachment."""
+        return await self._send_url_as_file(
+            chat_id, image_url, caption, reply_to, "image"
+        )
+
+    async def send_image_file(
+        self,
+        chat_id: str,
+        image_path: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        """Upload a local image file."""
+        return await self._send_local_file(
+            chat_id, image_path, caption, reply_to
+        )
+
+    async def send_document(
+        self,
+        chat_id: str,
+        file_path: str,
+        caption: Optional[str] = None,
+        file_name: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        """Upload a local file as a document."""
+        return await self._send_local_file(
+            chat_id, file_path, caption, reply_to, file_name
+        )
+
+    async def send_voice(
+        self,
+        chat_id: str,
+        audio_path: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        """Upload an audio file."""
+        return await self._send_local_file(
+            chat_id, audio_path, caption, reply_to
+        )
+
+    async def send_video(
+        self,
+        chat_id: str,
+        video_path: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        """Upload a video file."""
+        return await self._send_local_file(
+            chat_id, video_path, caption, reply_to
+        )
+
+    def format_message(self, content: str) -> str:
+        """Mattermost uses standard Markdown — mostly pass through.
+
+        Strip image markdown into plain links (files are uploaded separately).
+        """
+        # Convert ![alt](url) to just the URL — Mattermost renders
+        # image URLs as inline previews automatically.
+        content = re.sub(r"!\[([^\]]*)\]\(([^)]+)\)", r"\2", content)
+        return content
+
+    # ------------------------------------------------------------------
+    # File helpers
+    # ------------------------------------------------------------------
+
+    async def _send_url_as_file(
+        self,
+        chat_id: str,
+        url: str,
+        caption: Optional[str],
+        reply_to: Optional[str],
+        kind: str = "file",
+    ) -> SendResult:
+        """Download a URL and upload it as a file attachment."""
+        import asyncio
+        import aiohttp
+
+        last_exc = None
+        file_data = None
+        ct = "application/octet-stream"
+        fname = url.rsplit("/", 1)[-1].split("?")[0] or f"{kind}.png"
+
+        for attempt in range(3):
+            try:
+                async with self._session.get(url, timeout=aiohttp.ClientTimeout(total=30)) as resp:
+                    if resp.status >= 500 or resp.status == 429:
+                        if attempt < 2:
+                            logger.debug("Mattermost download retry %d/2 for %s (status %d)",
+                                         attempt + 1, url[:80], resp.status)
+                            await asyncio.sleep(1.5 * (attempt + 1))
+                            continue
+                    if resp.status >= 400:
+                        return await self.send(chat_id, f"{caption or ''}\n{url}".strip(), reply_to)
+                    file_data = await resp.read()
+                    ct = resp.content_type or "application/octet-stream"
+                    break
+            except (aiohttp.ClientError, asyncio.TimeoutError) as exc:
+                last_exc = exc
+                if attempt < 2:
+                    await asyncio.sleep(1.5 * (attempt + 1))
+                    continue
+                logger.warning("Mattermost: failed to download %s after %d attempts: %s", url, attempt + 1, exc)
+                return await self.send(chat_id, f"{caption or ''}\n{url}".strip(), reply_to)
+
+        if file_data is None:
+            logger.warning("Mattermost: download returned no data for %s", url)
+            return await self.send(chat_id, f"{caption or ''}\n{url}".strip(), reply_to)
+
+        file_id = await self._upload_file(chat_id, file_data, fname, ct)
+        if not file_id:
+            return await self.send(chat_id, f"{caption or ''}\n{url}".strip(), reply_to)
+
+        payload: Dict[str, Any] = {
+            "channel_id": chat_id,
+            "message": caption or "",
+            "file_ids": [file_id],
+        }
+        if reply_to and self._reply_mode == "thread":
+            payload["root_id"] = reply_to
+
+        data = await self._api_post("posts", payload)
+        if not data or "id" not in data:
+            return SendResult(success=False, error="Failed to post with file")
+        return SendResult(success=True, message_id=data["id"])
+
+    async def _send_local_file(
+        self,
+        chat_id: str,
+        file_path: str,
+        caption: Optional[str],
+        reply_to: Optional[str],
+        file_name: Optional[str] = None,
+    ) -> SendResult:
+        """Upload a local file and attach it to a post."""
+        import mimetypes
+
+        p = Path(file_path)
+        if not p.exists():
+            return await self.send(
+                chat_id, f"{caption or ''}\n(file not found: {file_path})", reply_to
+            )
+
+        fname = file_name or p.name
+        ct = mimetypes.guess_type(fname)[0] or "application/octet-stream"
+        file_data = p.read_bytes()
+
+        file_id = await self._upload_file(chat_id, file_data, fname, ct)
+        if not file_id:
+            return SendResult(success=False, error="File upload failed")
+
+        payload: Dict[str, Any] = {
+            "channel_id": chat_id,
+            "message": caption or "",
+            "file_ids": [file_id],
+        }
+        if reply_to and self._reply_mode == "thread":
+            payload["root_id"] = reply_to
+
+        data = await self._api_post("posts", payload)
+        if not data or "id" not in data:
+            return SendResult(success=False, error="Failed to post with file")
+        return SendResult(success=True, message_id=data["id"])
+
+    # ------------------------------------------------------------------
+    # WebSocket
+    # ------------------------------------------------------------------
+
+    async def _ws_loop(self) -> None:
+        """Connect to the WebSocket and listen for events, reconnecting on failure."""
+        delay = _RECONNECT_BASE_DELAY
+        while not self._closing:
+            try:
+                await self._ws_connect_and_listen()
+                # Clean disconnect — reset delay.
+                delay = _RECONNECT_BASE_DELAY
+            except asyncio.CancelledError:
+                return
+            except Exception as exc:
+                if self._closing:
+                    return
+                # Detect permanent auth/permission failures that will never
+                # succeed on retry — stop reconnecting instead of looping forever.
+                import aiohttp
+                err_str = str(exc).lower()
+                if isinstance(exc, aiohttp.WSServerHandshakeError) and exc.status in (401, 403):
+                    logger.error("Mattermost WS auth failed (HTTP %d) — stopping reconnect", exc.status)
+                    return
+                if "401" in err_str or "403" in err_str or "unauthorized" in err_str:
+                    logger.error("Mattermost WS permanent error: %s — stopping reconnect", exc)
+                    return
+                logger.warning("Mattermost WS error: %s — reconnecting in %.0fs", exc, delay)
+
+            if self._closing:
+                return
+
+            # Exponential backoff with jitter.
+            import random
+            jitter = delay * _RECONNECT_JITTER * random.random()
+            await asyncio.sleep(delay + jitter)
+            delay = min(delay * 2, _RECONNECT_MAX_DELAY)
+
+    async def _ws_connect_and_listen(self) -> None:
+        """Single WebSocket session: connect, authenticate, process events."""
+        # Build WS URL: https:// → wss://, http:// → ws://
+        ws_url = re.sub(r"^http", "ws", self._base_url) + "/api/v4/websocket"
+        logger.info("Mattermost: connecting to %s", ws_url)
+
+        self._ws = await self._session.ws_connect(ws_url, heartbeat=30.0)
+
+        # Authenticate via the WebSocket.
+        auth_msg = {
+            "seq": 1,
+            "action": "authentication_challenge",
+            "data": {"token": self._token},
+        }
+        await self._ws.send_json(auth_msg)
+        logger.info("Mattermost: WebSocket connected and authenticated")
+
+        async for raw_msg in self._ws:
+            if self._closing:
+                return
+
+            if raw_msg.type in (
+                raw_msg.type.TEXT,
+                raw_msg.type.BINARY,
+            ):
+                try:
+                    event = json.loads(raw_msg.data)
+                except (json.JSONDecodeError, TypeError):
+                    continue
+                await self._handle_ws_event(event)
+            elif raw_msg.type in (
+                raw_msg.type.ERROR,
+                raw_msg.type.CLOSE,
+                raw_msg.type.CLOSING,
+                raw_msg.type.CLOSED,
+            ):
+                logger.info("Mattermost: WebSocket closed (%s)", raw_msg.type)
+                break
+
+    async def _handle_ws_event(self, event: Dict[str, Any]) -> None:
+        """Process a single WebSocket event."""
+        event_type = event.get("event")
+        if event_type != "posted":
+            return
+
+        data = event.get("data", {})
+        raw_post_str = data.get("post")
+        if not raw_post_str:
+            return
+
+        try:
+            post = json.loads(raw_post_str)
+        except (json.JSONDecodeError, TypeError):
+            return
+
+        # Ignore own messages.
+        if post.get("user_id") == self._bot_user_id:
+            return
+
+        # Ignore system posts.
+        if post.get("type"):
+            return
+
+        post_id = post.get("id", "")
+
+        # Dedup.
+        self._prune_seen()
+        if post_id in self._seen_posts:
+            return
+        self._seen_posts[post_id] = time.time()
+
+        # Build message event.
+        channel_id = post.get("channel_id", "")
+        channel_type_raw = data.get("channel_type", "O")
+        chat_type = _CHANNEL_TYPE_MAP.get(channel_type_raw, "channel")
+
+        # For DMs, user_id is sufficient.  For channels, check for @mention.
+        message_text = post.get("message", "")
+
+        # Mention-gating for non-DM channels.
+        # Config (env vars):
+        #   MATTERMOST_REQUIRE_MENTION: Require @mention in channels (default: true)
+        #   MATTERMOST_FREE_RESPONSE_CHANNELS: Channel IDs where bot responds without mention
+        if channel_type_raw != "D":
+            require_mention = os.getenv(
+                "MATTERMOST_REQUIRE_MENTION", "true"
+            ).lower() not in ("false", "0", "no")
+
+            free_channels_raw = os.getenv("MATTERMOST_FREE_RESPONSE_CHANNELS", "")
+            free_channels = {ch.strip() for ch in free_channels_raw.split(",") if ch.strip()}
+            is_free_channel = channel_id in free_channels
+
+            mention_patterns = [
+                f"@{self._bot_username}",
+                f"@{self._bot_user_id}",
+            ]
+            has_mention = any(
+                pattern.lower() in message_text.lower()
+                for pattern in mention_patterns
+            )
+
+            if require_mention and not is_free_channel and not has_mention:
+                logger.debug(
+                    "Mattermost: skipping non-DM message without @mention (channel=%s)",
+                    channel_id,
+                )
+                return
+
+            # Strip @mention from the message text so the agent sees clean input.
+            if has_mention:
+                for pattern in mention_patterns:
+                    message_text = re.sub(
+                        re.escape(pattern), "", message_text, flags=re.IGNORECASE
+                    ).strip()
+
+        # Resolve sender info.
+        sender_id = post.get("user_id", "")
+        sender_name = data.get("sender_name", "").lstrip("@") or sender_id
+
+        # Thread support: if the post is in a thread, use root_id.
+        thread_id = post.get("root_id") or None
+
+        # Determine message type.
+        file_ids = post.get("file_ids") or []
+        msg_type = MessageType.TEXT
+        if message_text.startswith("/"):
+            msg_type = MessageType.COMMAND
+
+        # Download file attachments immediately (URLs require auth headers
+        # that downstream tools won't have).
+        media_urls: List[str] = []
+        media_types: List[str] = []
+        for fid in file_ids:
+            try:
+                file_info = await self._api_get(f"files/{fid}/info")
+                fname = file_info.get("name", f"file_{fid}")
+                ext = Path(fname).suffix or ""
+                mime = file_info.get("mime_type", "application/octet-stream")
+
+                import aiohttp
+                dl_url = f"{self._base_url}/api/v4/files/{fid}"
+                async with self._session.get(
+                    dl_url,
+                    headers={"Authorization": f"Bearer {self._token}"},
+                    timeout=aiohttp.ClientTimeout(total=30),
+                ) as resp:
+                    if resp.status < 400:
+                        file_data = await resp.read()
+                        from gateway.platforms.base import cache_image_from_bytes, cache_document_from_bytes
+                        if mime.startswith("image/"):
+                            local_path = cache_image_from_bytes(file_data, ext or ".png")
+                            media_urls.append(local_path)
+                            media_types.append(mime)
+                        elif mime.startswith("audio/"):
+                            from gateway.platforms.base import cache_audio_from_bytes
+                            local_path = cache_audio_from_bytes(file_data, ext or ".ogg")
+                            media_urls.append(local_path)
+                            media_types.append(mime)
+                        else:
+                            local_path = cache_document_from_bytes(file_data, fname)
+                            media_urls.append(local_path)
+                            media_types.append(mime)
+                    else:
+                        logger.warning("Mattermost: failed to download file %s: HTTP %s", fid, resp.status)
+            except Exception as exc:
+                logger.warning("Mattermost: error downloading file %s: %s", fid, exc)
+
+        source = self.build_source(
+            chat_id=channel_id,
+            chat_type=chat_type,
+            user_id=sender_id,
+            user_name=sender_name,
+            thread_id=thread_id,
+        )
+
+        msg_event = MessageEvent(
+            text=message_text,
+            message_type=msg_type,
+            source=source,
+            raw_message=post,
+            message_id=post_id,
+            media_urls=media_urls if media_urls else None,
+            media_types=media_types if media_types else None,
+        )
+
+        await self.handle_message(msg_event)
+
+    def _prune_seen(self) -> None:
+        """Remove expired entries from the dedup cache."""
+        if len(self._seen_posts) < self._SEEN_MAX:
+            return
+        now = time.time()
+        self._seen_posts = {
+            pid: ts
+            for pid, ts in self._seen_posts.items()
+            if now - ts < self._SEEN_TTL
+        }

gateway/platforms/signal.py

@@ -22,7 +22,7 @@ import time
 from datetime import datetime, timezone
 from pathlib import Path
 from typing import Dict, List, Optional, Any
-from urllib.parse import unquote
+from urllib.parse import quote, unquote
 
 import httpx
 
@@ -179,6 +179,13 @@ class SignalAdapter(BasePlatformAdapter):
         # Normalize account for self-message filtering
         self._account_normalized = self.account.strip()
 
+        # Track recently sent message timestamps to prevent echo-back loops
+        # in Note to Self / self-chat mode (mirrors WhatsApp recentlySentIds)
+        self._recent_sent_timestamps: set = set()
+        self._max_recent_timestamps = 50
+
+        self._phone_lock_identity: Optional[str] = None
+
         logger.info("Signal adapter initialized: url=%s account=%s groups=%s",
                      self.http_url, _redact_phone(self.account),
                      "enabled" if self.group_allow_from else "disabled")
@@ -193,6 +200,29 @@ class SignalAdapter(BasePlatformAdapter):
             logger.error("Signal: SIGNAL_HTTP_URL and SIGNAL_ACCOUNT are required")
             return False
 
+        # Acquire scoped lock to prevent duplicate Signal listeners for the same phone
+        try:
+            from gateway.status import acquire_scoped_lock
+
+            self._phone_lock_identity = self.account
+            acquired, existing = acquire_scoped_lock(
+                "signal-phone",
+                self._phone_lock_identity,
+                metadata={"platform": self.platform.value},
+            )
+            if not acquired:
+                owner_pid = existing.get("pid") if isinstance(existing, dict) else None
+                message = (
+                    "Another local Hermes gateway is already using this Signal account"
+                    + (f" (PID {owner_pid})." if owner_pid else ".")
+                    + " Stop the other gateway before starting a second Signal listener."
+                )
+                logger.error("Signal: %s", message)
+                self._set_fatal_error("signal_phone_lock", message, retryable=False)
+                return False
+        except Exception as e:
+            logger.warning("Signal: Could not acquire phone lock (non-fatal): %s", e)
+
         self.client = httpx.AsyncClient(timeout=30.0)
 
         # Health check — verify signal-cli daemon is reachable
@@ -240,6 +270,14 @@ class SignalAdapter(BasePlatformAdapter):
             await self.client.aclose()
             self.client = None
 
+        if self._phone_lock_identity:
+            try:
+                from gateway.status import release_scoped_lock
+                release_scoped_lock("signal-phone", self._phone_lock_identity)
+            except Exception as e:
+                logger.warning("Signal: Error releasing phone lock: %s", e, exc_info=True)
+            self._phone_lock_identity = None
+
         logger.info("Signal: disconnected")
 
     # ------------------------------------------------------------------
@@ -248,7 +286,7 @@ class SignalAdapter(BasePlatformAdapter):
 
     async def _sse_listener(self) -> None:
         """Listen for SSE events from signal-cli daemon."""
-        url = f"{self.http_url}/api/v1/events?account={self.account}"
+        url = f"{self.http_url}/api/v1/events?account={quote(self.account, safe='')}"
         backoff = SSE_RETRY_DELAY_INITIAL
 
         while self._running:
@@ -274,6 +312,12 @@ class SignalAdapter(BasePlatformAdapter):
                             line = line.strip()
                             if not line:
                                 continue
+                            # SSE keepalive comments (":") prove the connection
+                            # is alive — update activity so the health monitor
+                            # doesn't report false idle warnings.
+                            if line.startswith(":"):
+                                self._last_sse_activity = time.time()
+                                continue
                             # Parse SSE data lines
                             if line.startswith("data:"):
                                 data_str = line[5:].strip()
@@ -339,7 +383,9 @@ class SignalAdapter(BasePlatformAdapter):
         """Force SSE reconnection by closing the current response."""
         if self._sse_response and not self._sse_response.is_stream_consumed:
             try:
-                asyncio.create_task(self._sse_response.aclose())
+                task = asyncio.create_task(self._sse_response.aclose())
+                self._background_tasks.add(task)
+                task.add_done_callback(self._background_tasks.discard)
             except Exception:
                 pass
             self._sse_response = None
@@ -353,10 +399,26 @@ class SignalAdapter(BasePlatformAdapter):
         # Unwrap nested envelope if present
         envelope_data = envelope.get("envelope", envelope)
 
-        # Filter syncMessage envelopes (sent transcripts, read receipts, etc.)
-        # signal-cli may set syncMessage to null vs omitting it, so check key existence
+        # Handle syncMessage: extract "Note to Self" messages (sent to own account)
+        # while still filtering other sync events (read receipts, typing, etc.)
+        is_note_to_self = False
         if "syncMessage" in envelope_data:
-            return
+            sync_msg = envelope_data.get("syncMessage")
+            if sync_msg and isinstance(sync_msg, dict):
+                sent_msg = sync_msg.get("sentMessage")
+                if sent_msg and isinstance(sent_msg, dict):
+                    dest = sent_msg.get("destinationNumber") or sent_msg.get("destination")
+                    sent_ts = sent_msg.get("timestamp")
+                    if dest == self._account_normalized:
+                        # Check if this is an echo of our own outbound reply
+                        if sent_ts and sent_ts in self._recent_sent_timestamps:
+                            self._recent_sent_timestamps.discard(sent_ts)
+                            return
+                        # Genuine user Note to Self — promote to dataMessage
+                        is_note_to_self = True
+                        envelope_data = {**envelope_data, "dataMessage": sent_msg}
+            if not is_note_to_self:
+                return
 
         # Extract sender info
         sender = (
@@ -371,8 +433,8 @@ class SignalAdapter(BasePlatformAdapter):
             logger.debug("Signal: ignoring envelope with no sender")
             return
 
-        # Self-message filtering — prevent reply loops
-        if self._account_normalized and sender == self._account_normalized:
+        # Self-message filtering — prevent reply loops (but allow Note to Self)
+        if self._account_normalized and sender == self._account_normalized and not is_note_to_self:
             return
 
         # Filter stories
@@ -457,7 +519,7 @@ class SignalAdapter(BasePlatformAdapter):
             if any(mt.startswith("audio/") for mt in media_types):
                 msg_type = MessageType.VOICE
             elif any(mt.startswith("image/") for mt in media_types):
-                msg_type = MessageType.IMAGE
+                msg_type = MessageType.PHOTO
 
         # Parse timestamp from envelope data (milliseconds since epoch)
         ts_ms = envelope_data.get("timestamp", 0)
@@ -492,12 +554,19 @@ class SignalAdapter(BasePlatformAdapter):
         """Fetch an attachment via JSON-RPC and cache it. Returns (path, ext)."""
         result = await self._rpc("getAttachment", {
             "account": self.account,
-            "attachmentId": attachment_id,
+            "id": attachment_id,
         })
 
         if not result:
             return None, ""
 
+        # Handle dict response (signal-cli returns {"data": "base64..."})
+        if isinstance(result, dict):
+            result = result.get("data")
+            if not result:
+                logger.warning("Signal: attachment response missing 'data' key")
+                return None, ""
+
         # Result is base64-encoded file content
         raw_data = base64.b64decode(result)
         ext = _guess_extension(raw_data)
@@ -577,9 +646,18 @@ class SignalAdapter(BasePlatformAdapter):
         result = await self._rpc("send", params)
 
         if result is not None:
+            self._track_sent_timestamp(result)
             return SendResult(success=True)
         return SendResult(success=False, error="RPC send failed")
 
+    def _track_sent_timestamp(self, rpc_result) -> None:
+        """Record outbound message timestamp for echo-back filtering."""
+        ts = rpc_result.get("timestamp") if isinstance(rpc_result, dict) else None
+        if ts:
+            self._recent_sent_timestamps.add(ts)
+            if len(self._recent_sent_timestamps) > self._max_recent_timestamps:
+                self._recent_sent_timestamps.pop()
+
     async def send_typing(self, chat_id: str, metadata=None) -> None:
         """Send a typing indicator."""
         params: Dict[str, Any] = {
@@ -635,6 +713,7 @@ class SignalAdapter(BasePlatformAdapter):
 
         result = await self._rpc("send", params)
         if result is not None:
+            self._track_sent_timestamp(result)
             return SendResult(success=True)
         return SendResult(success=False, error="RPC send with attachment failed")
 
@@ -665,6 +744,7 @@ class SignalAdapter(BasePlatformAdapter):
 
         result = await self._rpc("send", params)
         if result is not None:
+            self._track_sent_timestamp(result)
             return SendResult(success=True)
         return SendResult(success=False, error="RPC send document failed")
 

gateway/platforms/slack.py

@@ -9,10 +9,12 @@ Uses slack-bolt (Python) with Socket Mode for:
 """
 
 import asyncio
+import json
 import logging
 import os
 import re
-from typing import Dict, List, Optional, Any
+import time
+from typing import Dict, Optional, Any
 
 try:
     from slack_bolt.async_app import AsyncApp
@@ -37,8 +39,6 @@ from gateway.platforms.base import (
     SendResult,
     SUPPORTED_DOCUMENT_TYPES,
     cache_document_from_bytes,
-    cache_image_from_url,
-    cache_audio_from_url,
 )
 
 
@@ -74,6 +74,16 @@ class SlackAdapter(BasePlatformAdapter):
         self._handler: Optional[AsyncSocketModeHandler] = None
         self._bot_user_id: Optional[str] = None
         self._user_name_cache: Dict[str, str] = {}  # user_id → display name
+        self._socket_mode_task: Optional[asyncio.Task] = None
+        # Multi-workspace support
+        self._team_clients: Dict[str, AsyncWebClient] = {}   # team_id → WebClient
+        self._team_bot_user_ids: Dict[str, str] = {}          # team_id → bot_user_id
+        self._channel_team: Dict[str, str] = {}                # channel_id → team_id
+        # Dedup cache: event_ts → timestamp.  Prevents duplicate bot
+        # responses when Socket Mode reconnects redeliver events.
+        self._seen_messages: Dict[str, float] = {}
+        self._SEEN_TTL = 300   # 5 minutes
+        self._SEEN_MAX = 2000  # prune threshold
 
     async def connect(self) -> bool:
         """Connect to Slack via Socket Mode."""
@@ -83,23 +93,70 @@ class SlackAdapter(BasePlatformAdapter):
             )
             return False
 
-        bot_token = self.config.token
+        raw_token = self.config.token
         app_token = os.getenv("SLACK_APP_TOKEN")
 
-        if not bot_token:
+        if not raw_token:
             logger.error("[Slack] SLACK_BOT_TOKEN not set")
             return False
         if not app_token:
             logger.error("[Slack] SLACK_APP_TOKEN not set")
             return False
 
-        try:
-            self._app = AsyncApp(token=bot_token)
+        # Support comma-separated bot tokens for multi-workspace
+        bot_tokens = [t.strip() for t in raw_token.split(",") if t.strip()]
 
-            # Get our own bot user ID for mention detection
-            auth_response = await self._app.client.auth_test()
-            self._bot_user_id = auth_response.get("user_id")
-            bot_name = auth_response.get("user", "unknown")
+        # Also load tokens from OAuth token file
+        from hermes_constants import get_hermes_home
+        tokens_file = get_hermes_home() / "slack_tokens.json"
+        if tokens_file.exists():
+            try:
+                saved = json.loads(tokens_file.read_text(encoding="utf-8"))
+                for team_id, entry in saved.items():
+                    tok = entry.get("token", "") if isinstance(entry, dict) else ""
+                    if tok and tok not in bot_tokens:
+                        bot_tokens.append(tok)
+                        team_label = entry.get("team_name", team_id) if isinstance(entry, dict) else team_id
+                        logger.info("[Slack] Loaded saved token for workspace %s", team_label)
+            except Exception as e:
+                logger.warning("[Slack] Failed to read %s: %s", tokens_file, e)
+
+        try:
+            # Acquire scoped lock to prevent duplicate app token usage
+            from gateway.status import acquire_scoped_lock
+            self._token_lock_identity = app_token
+            acquired, existing = acquire_scoped_lock('slack-app-token', app_token, metadata={'platform': 'slack'})
+            if not acquired:
+                owner_pid = existing.get('pid') if isinstance(existing, dict) else None
+                message = f'Slack app token already in use' + (f' (PID {owner_pid})' if owner_pid else '') + '. Stop the other gateway first.'
+                logger.error('[%s] %s', self.name, message)
+                self._set_fatal_error('slack_token_lock', message, retryable=False)
+                return False
+
+            # First token is the primary — used for AsyncApp / Socket Mode
+            primary_token = bot_tokens[0]
+            self._app = AsyncApp(token=primary_token)
+
+            # Register each bot token and map team_id → client
+            for token in bot_tokens:
+                client = AsyncWebClient(token=token)
+                auth_response = await client.auth_test()
+                team_id = auth_response.get("team_id", "")
+                bot_user_id = auth_response.get("user_id", "")
+                bot_name = auth_response.get("user", "unknown")
+                team_name = auth_response.get("team", "unknown")
+
+                self._team_clients[team_id] = client
+                self._team_bot_user_ids[team_id] = bot_user_id
+
+                # First token sets the primary bot_user_id (backward compat)
+                if self._bot_user_id is None:
+                    self._bot_user_id = bot_user_id
+
+                logger.info(
+                    "[Slack] Authenticated as @%s in workspace %s (team: %s)",
+                    bot_name, team_name, team_id,
+                )
 
             # Register message event handler
             @self._app.event("message")
@@ -121,10 +178,13 @@ class SlackAdapter(BasePlatformAdapter):
 
             # Start Socket Mode handler in background
             self._handler = AsyncSocketModeHandler(self._app, app_token)
-            asyncio.create_task(self._handler.start_async())
+            self._socket_mode_task = asyncio.create_task(self._handler.start_async())
 
             self._running = True
-            logger.info("[Slack] Connected as @%s (Socket Mode)", bot_name)
+            logger.info(
+                "[Slack] Socket Mode connected (%d workspace(s))",
+                len(self._team_clients),
+            )
             return True
 
         except Exception as e:  # pragma: no cover - defensive logging
@@ -139,8 +199,25 @@ class SlackAdapter(BasePlatformAdapter):
             except Exception as e:  # pragma: no cover - defensive logging
                 logger.warning("[Slack] Error while closing Socket Mode handler: %s", e, exc_info=True)
         self._running = False
+
+        # Release the token lock (use stored identity, not re-read env)
+        try:
+            from gateway.status import release_scoped_lock
+            if getattr(self, '_token_lock_identity', None):
+                release_scoped_lock('slack-app-token', self._token_lock_identity)
+                self._token_lock_identity = None
+        except Exception:
+            pass
+
         logger.info("[Slack] Disconnected")
 
+    def _get_client(self, chat_id: str) -> AsyncWebClient:
+        """Return the workspace-specific WebClient for a channel."""
+        team_id = self._channel_team.get(chat_id)
+        if team_id and team_id in self._team_clients:
+            return self._team_clients[team_id]
+        return self._app.client  # fallback to primary
+
     async def send(
         self,
         chat_id: str,
@@ -177,7 +254,7 @@ class SlackAdapter(BasePlatformAdapter):
                     if broadcast and i == 0:
                         kwargs["reply_broadcast"] = True
 
-                last_result = await self._app.client.chat_postMessage(**kwargs)
+                last_result = await self._get_client(chat_id).chat_postMessage(**kwargs)
 
             return SendResult(
                 success=True,
@@ -199,7 +276,7 @@ class SlackAdapter(BasePlatformAdapter):
         if not self._app:
             return SendResult(success=False, error="Not connected")
         try:
-            await self._app.client.chat_update(
+            await self._get_client(chat_id).chat_update(
                 channel=chat_id,
                 ts=message_id,
                 text=content,
@@ -233,7 +310,7 @@ class SlackAdapter(BasePlatformAdapter):
             return  # Can only set status in a thread context
 
         try:
-            await self._app.client.assistant_threads_setStatus(
+            await self._get_client(chat_id).assistant_threads_setStatus(
                 channel_id=chat_id,
                 thread_ts=thread_ts,
                 status="is thinking...",
@@ -252,7 +329,18 @@ class SlackAdapter(BasePlatformAdapter):
 
         Prefers metadata thread_id (the thread parent's ts, set by the
         gateway) over reply_to (which may be a child message's ts).
+
+        When ``reply_in_thread`` is ``false`` in the platform extra config,
+        top-level channel messages receive direct channel replies instead of
+        thread replies.  Messages that originate inside an existing thread are
+        always replied to in-thread to preserve conversation context.
         """
+        # When reply_in_thread is disabled (default: True for backward compat),
+        # only thread messages that are already part of an existing thread.
+        if not self.config.extra.get("reply_in_thread", True):
+            existing_thread = (metadata or {}).get("thread_id") or (metadata or {}).get("thread_ts")
+            return existing_thread or None
+
         if metadata:
             if metadata.get("thread_id"):
                 return metadata["thread_id"]
@@ -275,7 +363,7 @@ class SlackAdapter(BasePlatformAdapter):
         if not os.path.exists(file_path):
             raise FileNotFoundError(f"File not found: {file_path}")
 
-        result = await self._app.client.files_upload_v2(
+        result = await self._get_client(chat_id).files_upload_v2(
             channel=chat_id,
             file=file_path,
             filename=os.path.basename(file_path),
@@ -377,7 +465,7 @@ class SlackAdapter(BasePlatformAdapter):
         if not self._app:
             return False
         try:
-            await self._app.client.reactions_add(
+            await self._get_client(channel).reactions_add(
                 channel=channel, timestamp=timestamp, name=emoji
             )
             return True
@@ -393,7 +481,7 @@ class SlackAdapter(BasePlatformAdapter):
         if not self._app:
             return False
         try:
-            await self._app.client.reactions_remove(
+            await self._get_client(channel).reactions_remove(
                 channel=channel, timestamp=timestamp, name=emoji
             )
             return True
@@ -403,7 +491,7 @@ class SlackAdapter(BasePlatformAdapter):
 
     # ----- User identity resolution -----
 
-    async def _resolve_user_name(self, user_id: str) -> str:
+    async def _resolve_user_name(self, user_id: str, chat_id: str = "") -> str:
         """Resolve a Slack user ID to a display name, with caching."""
         if not user_id:
             return ""
@@ -414,7 +502,8 @@ class SlackAdapter(BasePlatformAdapter):
             return user_id
 
         try:
-            result = await self._app.client.users_info(user=user_id)
+            client = self._get_client(chat_id) if chat_id else self._app.client
+            result = await client.users_info(user=user_id)
             user = result.get("user", {})
             # Prefer display_name → real_name → user_id
             profile = user.get("profile", {})
@@ -478,7 +567,7 @@ class SlackAdapter(BasePlatformAdapter):
                 response = await client.get(image_url)
                 response.raise_for_status()
 
-            result = await self._app.client.files_upload_v2(
+            result = await self._get_client(chat_id).files_upload_v2(
                 channel=chat_id,
                 content=response.content,
                 filename="image.png",
@@ -538,7 +627,7 @@ class SlackAdapter(BasePlatformAdapter):
             return SendResult(success=False, error=f"Video file not found: {video_path}")
 
         try:
-            result = await self._app.client.files_upload_v2(
+            result = await self._get_client(chat_id).files_upload_v2(
                 channel=chat_id,
                 file=video_path,
                 filename=os.path.basename(video_path),
@@ -579,7 +668,7 @@ class SlackAdapter(BasePlatformAdapter):
         display_name = file_name or os.path.basename(file_path)
 
         try:
-            result = await self._app.client.files_upload_v2(
+            result = await self._get_client(chat_id).files_upload_v2(
                 channel=chat_id,
                 file=file_path,
                 filename=display_name,
@@ -607,7 +696,7 @@ class SlackAdapter(BasePlatformAdapter):
             return {"name": chat_id, "type": "unknown"}
 
         try:
-            result = await self._app.client.conversations_info(channel=chat_id)
+            result = await self._get_client(chat_id).conversations_info(channel=chat_id)
             channel = result.get("channel", {})
             is_dm = channel.get("is_im", False)
             return {
@@ -627,6 +716,20 @@ class SlackAdapter(BasePlatformAdapter):
 
     async def _handle_slack_message(self, event: dict) -> None:
         """Handle an incoming Slack message event."""
+        # Dedup: Slack Socket Mode can redeliver events after reconnects (#4777)
+        event_ts = event.get("ts", "")
+        if event_ts:
+            now = time.time()
+            if event_ts in self._seen_messages:
+                return
+            self._seen_messages[event_ts] = now
+            if len(self._seen_messages) > self._SEEN_MAX:
+                cutoff = now - self._SEEN_TTL
+                self._seen_messages = {
+                    k: v for k, v in self._seen_messages.items()
+                    if v > cutoff
+                }
+
         # Ignore bot messages (including our own)
         if event.get("bot_id") or event.get("subtype") == "bot_message":
             return
@@ -640,6 +743,11 @@ class SlackAdapter(BasePlatformAdapter):
         user_id = event.get("user", "")
         channel_id = event.get("channel", "")
         ts = event.get("ts", "")
+        team_id = event.get("team", "")
+
+        # Track which workspace owns this channel
+        if team_id and channel_id:
+            self._channel_team[channel_id] = team_id
 
         # Determine if this is a DM or channel message
         channel_type = event.get("channel_type", "")
@@ -656,11 +764,12 @@ class SlackAdapter(BasePlatformAdapter):
             thread_ts = event.get("thread_ts") or ts  # ts fallback for channels
 
         # In channels, only respond if bot is mentioned
-        if not is_dm and self._bot_user_id:
-            if f"<@{self._bot_user_id}>" not in text:
+        bot_uid = self._team_bot_user_ids.get(team_id, self._bot_user_id)
+        if not is_dm and bot_uid:
+            if f"<@{bot_uid}>" not in text:
                 return
             # Strip the bot mention from the text
-            text = text.replace(f"<@{self._bot_user_id}>", "").strip()
+            text = text.replace(f"<@{bot_uid}>", "").strip()
 
         # Determine message type
         msg_type = MessageType.TEXT
@@ -680,7 +789,7 @@ class SlackAdapter(BasePlatformAdapter):
                     if ext not in (".jpg", ".jpeg", ".png", ".gif", ".webp"):
                         ext = ".jpg"
                     # Slack private URLs require the bot token as auth header
-                    cached = await self._download_slack_file(url, ext)
+                    cached = await self._download_slack_file(url, ext, team_id=team_id)
                     media_urls.append(cached)
                     media_types.append(mimetype)
                     msg_type = MessageType.PHOTO
@@ -691,7 +800,7 @@ class SlackAdapter(BasePlatformAdapter):
                     ext = "." + mimetype.split("/")[-1].split(";")[0]
                     if ext not in (".ogg", ".mp3", ".wav", ".webm", ".m4a"):
                         ext = ".ogg"
-                    cached = await self._download_slack_file(url, ext, audio=True)
+                    cached = await self._download_slack_file(url, ext, audio=True, team_id=team_id)
                     media_urls.append(cached)
                     media_types.append(mimetype)
                     msg_type = MessageType.VOICE
@@ -722,7 +831,7 @@ class SlackAdapter(BasePlatformAdapter):
                         continue
 
                     # Download and cache
-                    raw_bytes = await self._download_slack_file_bytes(url)
+                    raw_bytes = await self._download_slack_file_bytes(url, team_id=team_id)
                     cached_path = cache_document_from_bytes(
                         raw_bytes, original_filename or f"document{ext}"
                     )
@@ -751,7 +860,7 @@ class SlackAdapter(BasePlatformAdapter):
                     logger.warning("[Slack] Failed to cache document from %s: %s", url, e, exc_info=True)
 
         # Resolve user display name (cached after first lookup)
-        user_name = await self._resolve_user_name(user_id)
+        user_name = await self._resolve_user_name(user_id, chat_id=channel_id)
 
         # Build source
         source = self.build_source(
@@ -788,24 +897,17 @@ class SlackAdapter(BasePlatformAdapter):
         text = command.get("text", "").strip()
         user_id = command.get("user_id", "")
         channel_id = command.get("channel_id", "")
+        team_id = command.get("team_id", "")
 
-        # Map subcommands to gateway commands
-        subcommand_map = {
-            "new": "/reset", "reset": "/reset",
-            "status": "/status", "stop": "/stop",
-            "help": "/help",
-            "model": "/model", "personality": "/personality",
-            "retry": "/retry", "undo": "/undo",
-            "compact": "/compress", "compress": "/compress",
-            "resume": "/resume",
-            "background": "/background",
-            "usage": "/usage",
-            "insights": "/insights",
-            "title": "/title",
-            "reasoning": "/reasoning",
-            "provider": "/provider",
-            "rollback": "/rollback",
-        }
+        # Track which workspace owns this channel
+        if team_id and channel_id:
+            self._channel_team[channel_id] = team_id
+
+        # Map subcommands to gateway commands — derived from central registry.
+        # Also keep "compact" as a Slack-specific alias for /compress.
+        from hermes_cli.commands import slack_subcommand_map
+        subcommand_map = slack_subcommand_map()
+        subcommand_map["compact"] = "/compress"
         first_word = text.split()[0] if text else ""
         if first_word in subcommand_map:
             # Preserve arguments after the subcommand
@@ -831,34 +933,66 @@ class SlackAdapter(BasePlatformAdapter):
 
         await self.handle_message(event)
 
-    async def _download_slack_file(self, url: str, ext: str, audio: bool = False) -> str:
-        """Download a Slack file using the bot token for auth."""
+    async def _download_slack_file(self, url: str, ext: str, audio: bool = False, team_id: str = "") -> str:
+        """Download a Slack file using the bot token for auth, with retry."""
+        import asyncio
         import httpx
 
-        bot_token = self.config.token
+        bot_token = self._team_clients[team_id].token if team_id and team_id in self._team_clients else self.config.token
+        last_exc = None
+
         async with httpx.AsyncClient(timeout=30.0, follow_redirects=True) as client:
-            response = await client.get(
-                url,
-                headers={"Authorization": f"Bearer {bot_token}"},
-            )
-            response.raise_for_status()
+            for attempt in range(3):
+                try:
+                    response = await client.get(
+                        url,
+                        headers={"Authorization": f"Bearer {bot_token}"},
+                    )
+                    response.raise_for_status()
 
-        if audio:
-            from gateway.platforms.base import cache_audio_from_bytes
-            return cache_audio_from_bytes(response.content, ext)
-        else:
-            from gateway.platforms.base import cache_image_from_bytes
-            return cache_image_from_bytes(response.content, ext)
+                    if audio:
+                        from gateway.platforms.base import cache_audio_from_bytes
+                        return cache_audio_from_bytes(response.content, ext)
+                    else:
+                        from gateway.platforms.base import cache_image_from_bytes
+                        return cache_image_from_bytes(response.content, ext)
+                except (httpx.TimeoutException, httpx.HTTPStatusError) as exc:
+                    last_exc = exc
+                    if isinstance(exc, httpx.HTTPStatusError) and exc.response.status_code < 429:
+                        raise
+                    if attempt < 2:
+                        logger.debug("Slack file download retry %d/2 for %s: %s",
+                                     attempt + 1, url[:80], exc)
+                        await asyncio.sleep(1.5 * (attempt + 1))
+                        continue
+                    raise
+        raise last_exc
 
-    async def _download_slack_file_bytes(self, url: str) -> bytes:
-        """Download a Slack file and return raw bytes."""
+    async def _download_slack_file_bytes(self, url: str, team_id: str = "") -> bytes:
+        """Download a Slack file and return raw bytes, with retry."""
+        import asyncio
         import httpx
 
-        bot_token = self.config.token
+        bot_token = self._team_clients[team_id].token if team_id and team_id in self._team_clients else self.config.token
+        last_exc = None
+
         async with httpx.AsyncClient(timeout=30.0, follow_redirects=True) as client:
-            response = await client.get(
-                url,
-                headers={"Authorization": f"Bearer {bot_token}"},
-            )
-            response.raise_for_status()
-        return response.content
+            for attempt in range(3):
+                try:
+                    response = await client.get(
+                        url,
+                        headers={"Authorization": f"Bearer {bot_token}"},
+                    )
+                    response.raise_for_status()
+                    return response.content
+                except (httpx.TimeoutException, httpx.HTTPStatusError) as exc:
+                    last_exc = exc
+                    if isinstance(exc, httpx.HTTPStatusError) and exc.response.status_code < 429:
+                        raise
+                    if attempt < 2:
+                        logger.debug("Slack file download retry %d/2 for %s: %s",
+                                     attempt + 1, url[:80], exc)
+                        await asyncio.sleep(1.5 * (attempt + 1))
+                        continue
+                    raise
+        raise last_exc

gateway/platforms/sms.py

@@ -0,0 +1,276 @@
+"""SMS (Twilio) platform adapter.
+
+Connects to the Twilio REST API for outbound SMS and runs an aiohttp
+webhook server to receive inbound messages.
+
+Shares credentials with the optional telephony skill — same env vars:
+  - TWILIO_ACCOUNT_SID
+  - TWILIO_AUTH_TOKEN
+  - TWILIO_PHONE_NUMBER  (E.164 from-number, e.g. +15551234567)
+
+Gateway-specific env vars:
+  - SMS_WEBHOOK_PORT     (default 8080)
+  - SMS_ALLOWED_USERS    (comma-separated E.164 phone numbers)
+  - SMS_ALLOW_ALL_USERS  (true/false)
+  - SMS_HOME_CHANNEL     (phone number for cron delivery)
+"""
+
+import asyncio
+import base64
+import logging
+import os
+import re
+import urllib.parse
+from typing import Any, Dict, Optional
+
+from gateway.config import Platform, PlatformConfig
+from gateway.platforms.base import (
+    BasePlatformAdapter,
+    MessageEvent,
+    MessageType,
+    SendResult,
+)
+
+logger = logging.getLogger(__name__)
+
+TWILIO_API_BASE = "https://api.twilio.com/2010-04-01/Accounts"
+MAX_SMS_LENGTH = 1600  # ~10 SMS segments
+DEFAULT_WEBHOOK_PORT = 8080
+
+# E.164 phone number pattern for redaction
+_PHONE_RE = re.compile(r"\+[1-9]\d{6,14}")
+
+
+def _redact_phone(phone: str) -> str:
+    """Redact a phone number for logging: +15551234567 -> +1555***4567."""
+    if not phone:
+        return "<none>"
+    if len(phone) <= 8:
+        return phone[:2] + "***" + phone[-2:] if len(phone) > 4 else "****"
+    return phone[:5] + "***" + phone[-4:]
+
+
+def check_sms_requirements() -> bool:
+    """Check if SMS adapter dependencies are available."""
+    try:
+        import aiohttp  # noqa: F401
+    except ImportError:
+        return False
+    return bool(os.getenv("TWILIO_ACCOUNT_SID") and os.getenv("TWILIO_AUTH_TOKEN"))
+
+
+class SmsAdapter(BasePlatformAdapter):
+    """
+    Twilio SMS <-> Hermes gateway adapter.
+
+    Each inbound phone number gets its own Hermes session (multi-tenant).
+    Replies are always sent from the configured TWILIO_PHONE_NUMBER.
+    """
+
+    MAX_MESSAGE_LENGTH = MAX_SMS_LENGTH
+
+    def __init__(self, config: PlatformConfig):
+        super().__init__(config, Platform.SMS)
+        self._account_sid: str = os.environ["TWILIO_ACCOUNT_SID"]
+        self._auth_token: str = os.environ["TWILIO_AUTH_TOKEN"]
+        self._from_number: str = os.getenv("TWILIO_PHONE_NUMBER", "")
+        self._webhook_port: int = int(
+            os.getenv("SMS_WEBHOOK_PORT", str(DEFAULT_WEBHOOK_PORT))
+        )
+        self._runner = None
+        self._http_session: Optional["aiohttp.ClientSession"] = None
+
+    def _basic_auth_header(self) -> str:
+        """Build HTTP Basic auth header value for Twilio."""
+        creds = f"{self._account_sid}:{self._auth_token}"
+        encoded = base64.b64encode(creds.encode("ascii")).decode("ascii")
+        return f"Basic {encoded}"
+
+    # ------------------------------------------------------------------
+    # Required abstract methods
+    # ------------------------------------------------------------------
+
+    async def connect(self) -> bool:
+        import aiohttp
+        from aiohttp import web
+
+        if not self._from_number:
+            logger.error("[sms] TWILIO_PHONE_NUMBER not set — cannot send replies")
+            return False
+
+        app = web.Application()
+        app.router.add_post("/webhooks/twilio", self._handle_webhook)
+        app.router.add_get("/health", lambda _: web.Response(text="ok"))
+
+        self._runner = web.AppRunner(app)
+        await self._runner.setup()
+        site = web.TCPSite(self._runner, "0.0.0.0", self._webhook_port)
+        await site.start()
+        self._http_session = aiohttp.ClientSession(
+            timeout=aiohttp.ClientTimeout(total=30),
+        )
+        self._running = True
+
+        logger.info(
+            "[sms] Twilio webhook server listening on port %d, from: %s",
+            self._webhook_port,
+            _redact_phone(self._from_number),
+        )
+        return True
+
+    async def disconnect(self) -> None:
+        if self._http_session:
+            await self._http_session.close()
+            self._http_session = None
+        if self._runner:
+            await self._runner.cleanup()
+            self._runner = None
+        self._running = False
+        logger.info("[sms] Disconnected")
+
+    async def send(
+        self,
+        chat_id: str,
+        content: str,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        import aiohttp
+
+        formatted = self.format_message(content)
+        chunks = self.truncate_message(formatted)
+        last_result = SendResult(success=True)
+
+        url = f"{TWILIO_API_BASE}/{self._account_sid}/Messages.json"
+        headers = {
+            "Authorization": self._basic_auth_header(),
+        }
+
+        session = self._http_session or aiohttp.ClientSession(
+            timeout=aiohttp.ClientTimeout(total=30),
+        )
+        try:
+            for chunk in chunks:
+                form_data = aiohttp.FormData()
+                form_data.add_field("From", self._from_number)
+                form_data.add_field("To", chat_id)
+                form_data.add_field("Body", chunk)
+
+                try:
+                    async with session.post(url, data=form_data, headers=headers) as resp:
+                        body = await resp.json()
+                        if resp.status >= 400:
+                            error_msg = body.get("message", str(body))
+                            logger.error(
+                                "[sms] send failed to %s: %s %s",
+                                _redact_phone(chat_id),
+                                resp.status,
+                                error_msg,
+                            )
+                            return SendResult(
+                                success=False,
+                                error=f"Twilio {resp.status}: {error_msg}",
+                            )
+                        msg_sid = body.get("sid", "")
+                        last_result = SendResult(success=True, message_id=msg_sid)
+                except Exception as e:
+                    logger.error("[sms] send error to %s: %s", _redact_phone(chat_id), e)
+                    return SendResult(success=False, error=str(e))
+        finally:
+            # Close session only if we created a fallback (no persistent session)
+            if not self._http_session and session:
+                await session.close()
+
+        return last_result
+
+    async def get_chat_info(self, chat_id: str) -> Dict[str, Any]:
+        return {"name": chat_id, "type": "dm"}
+
+    # ------------------------------------------------------------------
+    # SMS-specific formatting
+    # ------------------------------------------------------------------
+
+    def format_message(self, content: str) -> str:
+        """Strip markdown — SMS renders it as literal characters."""
+        content = re.sub(r"\*\*(.+?)\*\*", r"\1", content, flags=re.DOTALL)
+        content = re.sub(r"\*(.+?)\*", r"\1", content, flags=re.DOTALL)
+        content = re.sub(r"__(.+?)__", r"\1", content, flags=re.DOTALL)
+        content = re.sub(r"_(.+?)_", r"\1", content, flags=re.DOTALL)
+        content = re.sub(r"```[a-z]*\n?", "", content)
+        content = re.sub(r"`(.+?)`", r"\1", content)
+        content = re.sub(r"^#{1,6}\s+", "", content, flags=re.MULTILINE)
+        content = re.sub(r"\[([^\]]+)\]\([^\)]+\)", r"\1", content)
+        content = re.sub(r"\n{3,}", "\n\n", content)
+        return content.strip()
+
+    # ------------------------------------------------------------------
+    # Twilio webhook handler
+    # ------------------------------------------------------------------
+
+    async def _handle_webhook(self, request) -> "aiohttp.web.Response":
+        from aiohttp import web
+
+        try:
+            raw = await request.read()
+            # Twilio sends form-encoded data, not JSON
+            form = urllib.parse.parse_qs(raw.decode("utf-8"))
+        except Exception as e:
+            logger.error("[sms] webhook parse error: %s", e)
+            return web.Response(
+                text='<?xml version="1.0" encoding="UTF-8"?><Response></Response>',
+                content_type="application/xml",
+                status=400,
+            )
+
+        # Extract fields (parse_qs returns lists)
+        from_number = (form.get("From", [""]))[0].strip()
+        to_number = (form.get("To", [""]))[0].strip()
+        text = (form.get("Body", [""]))[0].strip()
+        message_sid = (form.get("MessageSid", [""]))[0].strip()
+
+        if not from_number or not text:
+            return web.Response(
+                text='<?xml version="1.0" encoding="UTF-8"?><Response></Response>',
+                content_type="application/xml",
+            )
+
+        # Ignore messages from our own number (echo prevention)
+        if from_number == self._from_number:
+            logger.debug("[sms] ignoring echo from own number %s", _redact_phone(from_number))
+            return web.Response(
+                text='<?xml version="1.0" encoding="UTF-8"?><Response></Response>',
+                content_type="application/xml",
+            )
+
+        logger.info(
+            "[sms] inbound from %s -> %s: %s",
+            _redact_phone(from_number),
+            _redact_phone(to_number),
+            text[:80],
+        )
+
+        source = self.build_source(
+            chat_id=from_number,
+            chat_name=from_number,
+            chat_type="dm",
+            user_id=from_number,
+            user_name=from_number,
+        )
+        event = MessageEvent(
+            text=text,
+            message_type=MessageType.TEXT,
+            source=source,
+            raw_message=form,
+            message_id=message_sid,
+        )
+
+        # Non-blocking: Twilio expects a fast response
+        task = asyncio.create_task(self.handle_message(event))
+        self._background_tasks.add(task)
+        task.add_done_callback(self._background_tasks.discard)
+
+        # Return empty TwiML — we send replies via the REST API, not inline TwiML
+        return web.Response(
+            text='<?xml version="1.0" encoding="UTF-8"?><Response></Response>',
+            content_type="application/xml",
+        )

gateway/platforms/telegram_network.py

@@ -0,0 +1,248 @@
+"""Telegram-specific network helpers.
+
+Provides a hostname-preserving fallback transport for networks where
+api.telegram.org resolves to an endpoint that is unreachable from the current
+host. The transport keeps the logical request host and TLS SNI as
+api.telegram.org while retrying the TCP connection against one or more fallback
+IPv4 addresses.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import ipaddress
+import logging
+import os
+import socket
+from typing import Iterable, Optional
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+_TELEGRAM_API_HOST = "api.telegram.org"
+
+# DNS-over-HTTPS providers used to discover Telegram API IPs that may differ
+# from the (potentially unreachable) IP returned by the local system resolver.
+_DOH_TIMEOUT = 4.0  # seconds — bounded so connect() isn't noticeably delayed
+
+_DOH_PROVIDERS: list[dict] = [
+    {
+        "url": "https://dns.google/resolve",
+        "params": {"name": _TELEGRAM_API_HOST, "type": "A"},
+        "headers": {},
+    },
+    {
+        "url": "https://cloudflare-dns.com/dns-query",
+        "params": {"name": _TELEGRAM_API_HOST, "type": "A"},
+        "headers": {"Accept": "application/dns-json"},
+    },
+]
+
+# Last-resort IPs when DoH is also blocked.  These are stable Telegram Bot API
+# endpoints in the 149.154.160.0/20 block (same seed used by OpenClaw).
+_SEED_FALLBACK_IPS: list[str] = ["149.154.167.220"]
+
+
+def _resolve_proxy_url() -> str | None:
+    for key in ("HTTPS_PROXY", "HTTP_PROXY", "ALL_PROXY", "https_proxy", "http_proxy", "all_proxy"):
+        value = (os.environ.get(key) or "").strip()
+        if value:
+            return value
+    return None
+
+
+class TelegramFallbackTransport(httpx.AsyncBaseTransport):
+    """Retry Telegram Bot API requests via fallback IPs while preserving TLS/SNI.
+
+    Requests continue to target https://api.telegram.org/... logically, but on
+    connect failures the underlying TCP connection is retried against a known
+    reachable IP. This is effectively the programmatic equivalent of
+    ``curl --resolve api.telegram.org:443:<ip>``.
+    """
+
+    def __init__(self, fallback_ips: Iterable[str], **transport_kwargs):
+        self._fallback_ips = [ip for ip in dict.fromkeys(_normalize_fallback_ips(fallback_ips))]
+        proxy_url = _resolve_proxy_url()
+        if proxy_url and "proxy" not in transport_kwargs:
+            transport_kwargs["proxy"] = proxy_url
+        self._primary = httpx.AsyncHTTPTransport(**transport_kwargs)
+        self._fallbacks = {
+            ip: httpx.AsyncHTTPTransport(**transport_kwargs) for ip in self._fallback_ips
+        }
+        self._sticky_ip: Optional[str] = None
+        self._sticky_lock = asyncio.Lock()
+
+    async def handle_async_request(self, request: httpx.Request) -> httpx.Response:
+        if request.url.host != _TELEGRAM_API_HOST or not self._fallback_ips:
+            return await self._primary.handle_async_request(request)
+
+        sticky_ip = self._sticky_ip
+        attempt_order: list[Optional[str]] = [sticky_ip] if sticky_ip else [None]
+        for ip in self._fallback_ips:
+            if ip != sticky_ip:
+                attempt_order.append(ip)
+
+        last_error: Exception | None = None
+        for ip in attempt_order:
+            candidate = request if ip is None else _rewrite_request_for_ip(request, ip)
+            transport = self._primary if ip is None else self._fallbacks[ip]
+            try:
+                response = await transport.handle_async_request(candidate)
+                if ip is not None and self._sticky_ip != ip:
+                    async with self._sticky_lock:
+                        if self._sticky_ip != ip:
+                            self._sticky_ip = ip
+                            logger.warning(
+                                "[Telegram] Primary api.telegram.org path unreachable; using sticky fallback IP %s",
+                                ip,
+                            )
+                return response
+            except Exception as exc:
+                last_error = exc
+                if not _is_retryable_connect_error(exc):
+                    raise
+                if ip is None:
+                    logger.warning(
+                        "[Telegram] Primary api.telegram.org connection failed (%s); trying fallback IPs %s",
+                        exc,
+                        ", ".join(self._fallback_ips),
+                    )
+                    continue
+                logger.warning("[Telegram] Fallback IP %s failed: %s", ip, exc)
+                continue
+
+        assert last_error is not None
+        raise last_error
+
+    async def aclose(self) -> None:
+        await self._primary.aclose()
+        for transport in self._fallbacks.values():
+            await transport.aclose()
+
+
+def _normalize_fallback_ips(values: Iterable[str]) -> list[str]:
+    normalized: list[str] = []
+    for value in values:
+        raw = str(value).strip()
+        if not raw:
+            continue
+        try:
+            addr = ipaddress.ip_address(raw)
+        except ValueError:
+            logger.warning("Ignoring invalid Telegram fallback IP: %r", raw)
+            continue
+        if addr.version != 4:
+            logger.warning("Ignoring non-IPv4 Telegram fallback IP: %s", raw)
+            continue
+        if addr.is_private or addr.is_loopback or addr.is_link_local or addr.is_unspecified:
+            logger.warning("Ignoring private/internal Telegram fallback IP: %s", raw)
+            continue
+        normalized.append(str(addr))
+    return normalized
+
+
+def parse_fallback_ip_env(value: str | None) -> list[str]:
+    if not value:
+        return []
+    parts = [part.strip() for part in value.split(",")]
+    return _normalize_fallback_ips(parts)
+
+
+def _resolve_system_dns() -> set[str]:
+    """Return the IPv4 addresses that the OS resolver gives for api.telegram.org."""
+    try:
+        results = socket.getaddrinfo(_TELEGRAM_API_HOST, 443, socket.AF_INET)
+        return {addr[4][0] for addr in results}
+    except Exception:
+        return set()
+
+
+async def _query_doh_provider(
+    client: httpx.AsyncClient, provider: dict
+) -> list[str]:
+    """Query one DoH provider and return A-record IPs."""
+    try:
+        resp = await client.get(
+            provider["url"], params=provider["params"], headers=provider["headers"]
+        )
+        resp.raise_for_status()
+        data = resp.json()
+        ips: list[str] = []
+        for answer in data.get("Answer", []):
+            if answer.get("type") != 1:  # A record
+                continue
+            raw = answer.get("data", "").strip()
+            try:
+                ipaddress.ip_address(raw)
+                ips.append(raw)
+            except ValueError:
+                continue
+        return ips
+    except Exception as exc:
+        logger.debug("DoH query to %s failed: %s", provider["url"], exc)
+        return []
+
+
+async def discover_fallback_ips() -> list[str]:
+    """Auto-discover Telegram API IPs via DNS-over-HTTPS.
+
+    Resolves api.telegram.org through Google and Cloudflare DoH, collects all
+    unique IPs, and excludes the system-DNS-resolved IP (which is presumably
+    unreachable on this network).  Falls back to a hardcoded seed list when DoH
+    is also unavailable.
+    """
+    async with httpx.AsyncClient(timeout=httpx.Timeout(_DOH_TIMEOUT)) as client:
+        doh_tasks = [_query_doh_provider(client, p) for p in _DOH_PROVIDERS]
+        system_dns_task = asyncio.to_thread(_resolve_system_dns)
+        results = await asyncio.gather(system_dns_task, *doh_tasks, return_exceptions=True)
+
+    # results[0] = system DNS IPs (set), results[1:] = DoH IP lists
+    system_ips: set[str] = results[0] if isinstance(results[0], set) else set()
+
+    doh_ips: list[str] = []
+    for r in results[1:]:
+        if isinstance(r, list):
+            doh_ips.extend(r)
+
+    # Deduplicate preserving order, exclude system-DNS IPs
+    seen: set[str] = set()
+    candidates: list[str] = []
+    for ip in doh_ips:
+        if ip not in seen and ip not in system_ips:
+            seen.add(ip)
+            candidates.append(ip)
+
+    # Validate through existing normalization
+    validated = _normalize_fallback_ips(candidates)
+
+    if validated:
+        logger.debug("Discovered Telegram fallback IPs via DoH: %s", ", ".join(validated))
+        return validated
+
+    logger.info(
+        "DoH discovery yielded no new IPs (system DNS: %s); using seed fallback IPs %s",
+        ", ".join(system_ips) or "unknown",
+        ", ".join(_SEED_FALLBACK_IPS),
+    )
+    return list(_SEED_FALLBACK_IPS)
+
+
+def _rewrite_request_for_ip(request: httpx.Request, ip: str) -> httpx.Request:
+    original_host = request.url.host or _TELEGRAM_API_HOST
+    url = request.url.copy_with(host=ip)
+    headers = request.headers.copy()
+    headers["host"] = original_host
+    extensions = dict(request.extensions)
+    extensions["sni_hostname"] = original_host
+    return httpx.Request(
+        method=request.method,
+        url=url,
+        headers=headers,
+        stream=request.stream,
+        extensions=extensions,
+    )
+
+
+def _is_retryable_connect_error(exc: Exception) -> bool:
+    return isinstance(exc, (httpx.ConnectTimeout, httpx.ConnectError))

gateway/platforms/webhook.py

@@ -0,0 +1,616 @@
+"""Generic webhook platform adapter.
+
+Runs an aiohttp HTTP server that receives webhook POSTs from external
+services (GitHub, GitLab, JIRA, Stripe, etc.), validates HMAC signatures,
+transforms payloads into agent prompts, and routes responses back to the
+source or to another configured platform.
+
+Configuration lives in config.yaml under platforms.webhook.extra.routes.
+Each route defines:
+  - events: which event types to accept (header-based filtering)
+  - secret: HMAC secret for signature validation (REQUIRED)
+  - prompt: template string formatted with the webhook payload
+  - skills: optional list of skills to load for the agent
+  - deliver: where to send the response (github_comment, telegram, etc.)
+  - deliver_extra: additional delivery config (repo, pr_number, chat_id)
+
+Security:
+  - HMAC secret is required per route (validated at startup)
+  - Rate limiting per route (fixed-window, configurable)
+  - Idempotency cache prevents duplicate agent runs on webhook retries
+  - Body size limits checked before reading payload
+  - Set secret to "INSECURE_NO_AUTH" to skip validation (testing only)
+"""
+
+import asyncio
+import hashlib
+import hmac
+import json
+import logging
+import os
+import re
+import subprocess
+import time
+from typing import Any, Dict, List, Optional
+
+try:
+    from aiohttp import web
+
+    AIOHTTP_AVAILABLE = True
+except ImportError:
+    AIOHTTP_AVAILABLE = False
+    web = None  # type: ignore[assignment]
+
+from gateway.config import Platform, PlatformConfig
+from gateway.platforms.base import (
+    BasePlatformAdapter,
+    MessageEvent,
+    MessageType,
+    SendResult,
+)
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_HOST = "0.0.0.0"
+DEFAULT_PORT = 8644
+_INSECURE_NO_AUTH = "INSECURE_NO_AUTH"
+_DYNAMIC_ROUTES_FILENAME = "webhook_subscriptions.json"
+
+
+def check_webhook_requirements() -> bool:
+    """Check if webhook adapter dependencies are available."""
+    return AIOHTTP_AVAILABLE
+
+
+class WebhookAdapter(BasePlatformAdapter):
+    """Generic webhook receiver that triggers agent runs from HTTP POSTs."""
+
+    def __init__(self, config: PlatformConfig):
+        super().__init__(config, Platform.WEBHOOK)
+        self._host: str = config.extra.get("host", DEFAULT_HOST)
+        self._port: int = int(config.extra.get("port", DEFAULT_PORT))
+        self._global_secret: str = config.extra.get("secret", "")
+        self._static_routes: Dict[str, dict] = config.extra.get("routes", {})
+        self._dynamic_routes: Dict[str, dict] = {}
+        self._dynamic_routes_mtime: float = 0.0
+        self._routes: Dict[str, dict] = dict(self._static_routes)
+        self._runner = None
+
+        # Delivery info keyed by session chat_id — consumed by send()
+        self._delivery_info: Dict[str, dict] = {}
+
+        # Reference to gateway runner for cross-platform delivery (set externally)
+        self.gateway_runner = None
+
+        # Idempotency: TTL cache of recently processed delivery IDs.
+        # Prevents duplicate agent runs when webhook providers retry.
+        self._seen_deliveries: Dict[str, float] = {}
+        self._idempotency_ttl: int = 3600  # 1 hour
+
+        # Rate limiting: per-route timestamps in a fixed window.
+        self._rate_counts: Dict[str, List[float]] = {}
+        self._rate_limit: int = int(config.extra.get("rate_limit", 30))  # per minute
+
+        # Body size limit (auth-before-body pattern)
+        self._max_body_bytes: int = int(
+            config.extra.get("max_body_bytes", 1_048_576)
+        )  # 1MB
+
+    # ------------------------------------------------------------------
+    # Lifecycle
+    # ------------------------------------------------------------------
+
+    async def connect(self) -> bool:
+        # Load agent-created subscriptions before validating
+        self._reload_dynamic_routes()
+
+        # Validate routes at startup — secret is required per route
+        for name, route in self._routes.items():
+            secret = route.get("secret", self._global_secret)
+            if not secret:
+                raise ValueError(
+                    f"[webhook] Route '{name}' has no HMAC secret. "
+                    f"Set 'secret' on the route or globally. "
+                    f"For testing without auth, set secret to '{_INSECURE_NO_AUTH}'."
+                )
+
+        app = web.Application()
+        app.router.add_get("/health", self._handle_health)
+        app.router.add_post("/webhooks/{route_name}", self._handle_webhook)
+
+        # Port conflict detection — fail fast if port is already in use
+        import socket as _socket
+        try:
+            with _socket.socket(_socket.AF_INET, _socket.SOCK_STREAM) as _s:
+                _s.settimeout(1)
+                _s.connect(('127.0.0.1', self._port))
+            logger.error('[webhook] Port %d already in use. Set a different port in config.yaml: platforms.webhook.port', self._port)
+            return False
+        except (ConnectionRefusedError, OSError):
+            pass  # port is free
+
+        self._runner = web.AppRunner(app)
+        await self._runner.setup()
+        site = web.TCPSite(self._runner, self._host, self._port)
+        await site.start()
+        self._mark_connected()
+
+        route_names = ", ".join(self._routes.keys()) or "(none configured)"
+        logger.info(
+            "[webhook] Listening on %s:%d — routes: %s",
+            self._host,
+            self._port,
+            route_names,
+        )
+        return True
+
+    async def disconnect(self) -> None:
+        if self._runner:
+            await self._runner.cleanup()
+            self._runner = None
+        self._mark_disconnected()
+        logger.info("[webhook] Disconnected")
+
+    async def send(
+        self,
+        chat_id: str,
+        content: str,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        """Deliver the agent's response to the configured destination.
+
+        chat_id is ``webhook:{route}:{delivery_id}`` — we pop the delivery
+        info stored during webhook receipt so it doesn't leak memory.
+        """
+        delivery = self._delivery_info.pop(chat_id, {})
+        deliver_type = delivery.get("deliver", "log")
+
+        if deliver_type == "log":
+            logger.info("[webhook] Response for %s: %s", chat_id, content[:200])
+            return SendResult(success=True)
+
+        if deliver_type == "github_comment":
+            return await self._deliver_github_comment(content, delivery)
+
+        # Cross-platform delivery (telegram, discord, etc.)
+        if self.gateway_runner and deliver_type in (
+            "telegram",
+            "discord",
+            "slack",
+            "signal",
+            "sms",
+        ):
+            return await self._deliver_cross_platform(
+                deliver_type, content, delivery
+            )
+
+        logger.warning("[webhook] Unknown deliver type: %s", deliver_type)
+        return SendResult(
+            success=False, error=f"Unknown deliver type: {deliver_type}"
+        )
+
+    async def get_chat_info(self, chat_id: str) -> Dict[str, Any]:
+        return {"name": chat_id, "type": "webhook"}
+
+    # ------------------------------------------------------------------
+    # HTTP handlers
+    # ------------------------------------------------------------------
+
+    async def _handle_health(self, request: "web.Request") -> "web.Response":
+        """GET /health — simple health check."""
+        return web.json_response({"status": "ok", "platform": "webhook"})
+
+    def _reload_dynamic_routes(self) -> None:
+        """Reload agent-created subscriptions from disk if the file changed."""
+        from pathlib import Path as _Path
+        hermes_home = _Path(
+            os.getenv("HERMES_HOME", str(_Path.home() / ".hermes"))
+        ).expanduser()
+        subs_path = hermes_home / _DYNAMIC_ROUTES_FILENAME
+        if not subs_path.exists():
+            if self._dynamic_routes:
+                self._dynamic_routes = {}
+                self._routes = dict(self._static_routes)
+                logger.debug("[webhook] Dynamic subscriptions file removed, cleared dynamic routes")
+            return
+        try:
+            mtime = subs_path.stat().st_mtime
+            if mtime <= self._dynamic_routes_mtime:
+                return  # No change
+            data = json.loads(subs_path.read_text(encoding="utf-8"))
+            if not isinstance(data, dict):
+                return
+            # Merge: static routes take precedence over dynamic ones
+            self._dynamic_routes = {
+                k: v for k, v in data.items()
+                if k not in self._static_routes
+            }
+            self._routes = {**self._dynamic_routes, **self._static_routes}
+            self._dynamic_routes_mtime = mtime
+            logger.info(
+                "[webhook] Reloaded %d dynamic route(s): %s",
+                len(self._dynamic_routes),
+                ", ".join(self._dynamic_routes.keys()) or "(none)",
+            )
+        except Exception as e:
+            logger.warning("[webhook] Failed to reload dynamic routes: %s", e)
+
+    async def _handle_webhook(self, request: "web.Request") -> "web.Response":
+        """POST /webhooks/{route_name} — receive and process a webhook event."""
+        # Hot-reload dynamic subscriptions on each request (mtime-gated, cheap)
+        self._reload_dynamic_routes()
+
+        route_name = request.match_info.get("route_name", "")
+        route_config = self._routes.get(route_name)
+
+        if not route_config:
+            return web.json_response(
+                {"error": f"Unknown route: {route_name}"}, status=404
+            )
+
+        # ── Auth-before-body ─────────────────────────────────────
+        # Check Content-Length before reading the full payload.
+        content_length = request.content_length or 0
+        if content_length > self._max_body_bytes:
+            return web.json_response(
+                {"error": "Payload too large"}, status=413
+            )
+
+        # ── Rate limiting ────────────────────────────────────────
+        now = time.time()
+        window = self._rate_counts.setdefault(route_name, [])
+        window[:] = [t for t in window if now - t < 60]
+        if len(window) >= self._rate_limit:
+            return web.json_response(
+                {"error": "Rate limit exceeded"}, status=429
+            )
+        window.append(now)
+
+        # Read body
+        try:
+            raw_body = await request.read()
+        except Exception as e:
+            logger.error("[webhook] Failed to read body: %s", e)
+            return web.json_response({"error": "Bad request"}, status=400)
+
+        # Validate HMAC signature (skip for INSECURE_NO_AUTH testing mode)
+        secret = route_config.get("secret", self._global_secret)
+        if secret and secret != _INSECURE_NO_AUTH:
+            if not self._validate_signature(request, raw_body, secret):
+                logger.warning(
+                    "[webhook] Invalid signature for route %s", route_name
+                )
+                return web.json_response(
+                    {"error": "Invalid signature"}, status=401
+                )
+
+        # Parse payload
+        try:
+            payload = json.loads(raw_body)
+        except json.JSONDecodeError:
+            # Try form-encoded as fallback
+            try:
+                import urllib.parse
+
+                payload = dict(
+                    urllib.parse.parse_qsl(raw_body.decode("utf-8"))
+                )
+            except Exception:
+                return web.json_response(
+                    {"error": "Cannot parse body"}, status=400
+                )
+
+        # Check event type filter
+        event_type = (
+            request.headers.get("X-GitHub-Event", "")
+            or request.headers.get("X-GitLab-Event", "")
+            or payload.get("event_type", "")
+            or "unknown"
+        )
+        allowed_events = route_config.get("events", [])
+        if allowed_events and event_type not in allowed_events:
+            logger.debug(
+                "[webhook] Ignoring event %s for route %s (allowed: %s)",
+                event_type,
+                route_name,
+                allowed_events,
+            )
+            return web.json_response(
+                {"status": "ignored", "event": event_type}
+            )
+
+        # Format prompt from template
+        prompt_template = route_config.get("prompt", "")
+        prompt = self._render_prompt(
+            prompt_template, payload, event_type, route_name
+        )
+
+        # Inject skill content if configured.
+        # We call build_skill_invocation_message() directly rather than
+        # using /skill-name slash commands — the gateway's command parser
+        # would intercept those and break the flow.
+        skills = route_config.get("skills", [])
+        if skills:
+            try:
+                from agent.skill_commands import (
+                    build_skill_invocation_message,
+                    get_skill_commands,
+                )
+
+                skill_cmds = get_skill_commands()
+                for skill_name in skills:
+                    cmd_key = f"/{skill_name}"
+                    if cmd_key in skill_cmds:
+                        skill_content = build_skill_invocation_message(
+                            cmd_key, user_instruction=prompt
+                        )
+                        if skill_content:
+                            prompt = skill_content
+                            break  # Load the first matching skill
+                    else:
+                        logger.warning(
+                            "[webhook] Skill '%s' not found", skill_name
+                        )
+            except Exception as e:
+                logger.warning("[webhook] Skill loading failed: %s", e)
+
+        # Build a unique delivery ID
+        delivery_id = request.headers.get(
+            "X-GitHub-Delivery",
+            request.headers.get("X-Request-ID", str(int(time.time() * 1000))),
+        )
+
+        # ── Idempotency ─────────────────────────────────────────
+        # Skip duplicate deliveries (webhook retries).
+        now = time.time()
+        # Prune expired entries
+        self._seen_deliveries = {
+            k: v
+            for k, v in self._seen_deliveries.items()
+            if now - v < self._idempotency_ttl
+        }
+        if delivery_id in self._seen_deliveries:
+            logger.info(
+                "[webhook] Skipping duplicate delivery %s", delivery_id
+            )
+            return web.json_response(
+                {"status": "duplicate", "delivery_id": delivery_id},
+                status=200,
+            )
+        self._seen_deliveries[delivery_id] = now
+
+        # Use delivery_id in session key so concurrent webhooks on the
+        # same route get independent agent runs (not queued/interrupted).
+        session_chat_id = f"webhook:{route_name}:{delivery_id}"
+
+        # Store delivery info for send() — consumed (popped) on delivery
+        deliver_config = {
+            "deliver": route_config.get("deliver", "log"),
+            "deliver_extra": self._render_delivery_extra(
+                route_config.get("deliver_extra", {}), payload
+            ),
+            "payload": payload,
+        }
+        self._delivery_info[session_chat_id] = deliver_config
+
+        # Build source and event
+        source = self.build_source(
+            chat_id=session_chat_id,
+            chat_name=f"webhook/{route_name}",
+            chat_type="webhook",
+            user_id=f"webhook:{route_name}",
+            user_name=route_name,
+        )
+        event = MessageEvent(
+            text=prompt,
+            message_type=MessageType.TEXT,
+            source=source,
+            raw_message=payload,
+            message_id=delivery_id,
+        )
+
+        logger.info(
+            "[webhook] %s event=%s route=%s prompt_len=%d delivery=%s",
+            request.method,
+            event_type,
+            route_name,
+            len(prompt),
+            delivery_id,
+        )
+
+        # Non-blocking — return 202 Accepted immediately
+        task = asyncio.create_task(self.handle_message(event))
+        self._background_tasks.add(task)
+        task.add_done_callback(self._background_tasks.discard)
+
+        return web.json_response(
+            {
+                "status": "accepted",
+                "route": route_name,
+                "event": event_type,
+                "delivery_id": delivery_id,
+            },
+            status=202,
+        )
+
+    # ------------------------------------------------------------------
+    # Signature validation
+    # ------------------------------------------------------------------
+
+    def _validate_signature(
+        self, request: "web.Request", body: bytes, secret: str
+    ) -> bool:
+        """Validate webhook signature (GitHub, GitLab, generic HMAC-SHA256)."""
+        # GitHub: X-Hub-Signature-256 = sha256=<hex>
+        gh_sig = request.headers.get("X-Hub-Signature-256", "")
+        if gh_sig:
+            expected = "sha256=" + hmac.new(
+                secret.encode(), body, hashlib.sha256
+            ).hexdigest()
+            return hmac.compare_digest(gh_sig, expected)
+
+        # GitLab: X-Gitlab-Token = <plain secret>
+        gl_token = request.headers.get("X-Gitlab-Token", "")
+        if gl_token:
+            return hmac.compare_digest(gl_token, secret)
+
+        # Generic: X-Webhook-Signature = <hex HMAC-SHA256>
+        generic_sig = request.headers.get("X-Webhook-Signature", "")
+        if generic_sig:
+            expected = hmac.new(
+                secret.encode(), body, hashlib.sha256
+            ).hexdigest()
+            return hmac.compare_digest(generic_sig, expected)
+
+        # No recognised signature header but secret is configured → reject
+        logger.debug(
+            "[webhook] Secret configured but no signature header found"
+        )
+        return False
+
+    # ------------------------------------------------------------------
+    # Prompt rendering
+    # ------------------------------------------------------------------
+
+    def _render_prompt(
+        self,
+        template: str,
+        payload: dict,
+        event_type: str,
+        route_name: str,
+    ) -> str:
+        """Render a prompt template with the webhook payload.
+
+        Supports dot-notation access into nested dicts:
+        ``{pull_request.title}`` → ``payload["pull_request"]["title"]``
+        """
+        if not template:
+            truncated = json.dumps(payload, indent=2)[:4000]
+            return (
+                f"Webhook event '{event_type}' on route "
+                f"'{route_name}':\n\n```json\n{truncated}\n```"
+            )
+
+        def _resolve(match: re.Match) -> str:
+            key = match.group(1)
+            value: Any = payload
+            for part in key.split("."):
+                if isinstance(value, dict):
+                    value = value.get(part, f"{{{key}}}")
+                else:
+                    return f"{{{key}}}"
+            if isinstance(value, (dict, list)):
+                return json.dumps(value, indent=2)[:2000]
+            return str(value)
+
+        return re.sub(r"\{([a-zA-Z0-9_.]+)\}", _resolve, template)
+
+    def _render_delivery_extra(
+        self, extra: dict, payload: dict
+    ) -> dict:
+        """Render delivery_extra template values with payload data."""
+        rendered: Dict[str, Any] = {}
+        for key, value in extra.items():
+            if isinstance(value, str):
+                rendered[key] = self._render_prompt(value, payload, "", "")
+            else:
+                rendered[key] = value
+        return rendered
+
+    # ------------------------------------------------------------------
+    # Response delivery
+    # ------------------------------------------------------------------
+
+    async def _deliver_github_comment(
+        self, content: str, delivery: dict
+    ) -> SendResult:
+        """Post agent response as a GitHub PR/issue comment via ``gh`` CLI."""
+        extra = delivery.get("deliver_extra", {})
+        repo = extra.get("repo", "")
+        pr_number = extra.get("pr_number", "")
+
+        if not repo or not pr_number:
+            logger.error(
+                "[webhook] github_comment delivery missing repo or pr_number"
+            )
+            return SendResult(
+                success=False, error="Missing repo or pr_number"
+            )
+
+        try:
+            result = subprocess.run(
+                [
+                    "gh",
+                    "pr",
+                    "comment",
+                    str(pr_number),
+                    "--repo",
+                    repo,
+                    "--body",
+                    content,
+                ],
+                capture_output=True,
+                text=True,
+                timeout=30,
+            )
+            if result.returncode == 0:
+                logger.info(
+                    "[webhook] Posted comment on %s#%s", repo, pr_number
+                )
+                return SendResult(success=True)
+            else:
+                logger.error(
+                    "[webhook] gh pr comment failed: %s", result.stderr
+                )
+                return SendResult(success=False, error=result.stderr)
+        except FileNotFoundError:
+            logger.error(
+                "[webhook] 'gh' CLI not found — install GitHub CLI for "
+                "github_comment delivery"
+            )
+            return SendResult(
+                success=False, error="gh CLI not installed"
+            )
+        except Exception as e:
+            logger.error("[webhook] github_comment delivery error: %s", e)
+            return SendResult(success=False, error=str(e))
+
+    async def _deliver_cross_platform(
+        self, platform_name: str, content: str, delivery: dict
+    ) -> SendResult:
+        """Route response to another platform (telegram, discord, etc.)."""
+        if not self.gateway_runner:
+            return SendResult(
+                success=False,
+                error="No gateway runner for cross-platform delivery",
+            )
+
+        try:
+            target_platform = Platform(platform_name)
+        except ValueError:
+            return SendResult(
+                success=False, error=f"Unknown platform: {platform_name}"
+            )
+
+        adapter = self.gateway_runner.adapters.get(target_platform)
+        if not adapter:
+            return SendResult(
+                success=False,
+                error=f"Platform {platform_name} not connected",
+            )
+
+        # Use home channel if no specific chat_id in deliver_extra
+        extra = delivery.get("deliver_extra", {})
+        chat_id = extra.get("chat_id", "")
+        if not chat_id:
+            home = self.gateway_runner.config.get_home_channel(target_platform)
+            if home:
+                chat_id = home.chat_id
+            else:
+                return SendResult(
+                    success=False,
+                    error=f"No chat_id or home channel for {platform_name}",
+                )
+
+        return await adapter.send(chat_id, content)

gateway/platforms/whatsapp.py

@@ -20,13 +20,15 @@ import json
 import logging
 import os
 import platform
+import re
 import subprocess
 
 _IS_WINDOWS = platform.system() == "Windows"
 from pathlib import Path
-from typing import Dict, List, Optional, Any
+from typing import Dict, Optional, Any
 
 from hermes_cli.config import get_hermes_home
+from hermes_constants import get_hermes_dir
 
 logger = logging.getLogger(__name__)
 
@@ -74,6 +76,7 @@ from gateway.platforms.base import (
     MessageEvent,
     MessageType,
     SendResult,
+    SUPPORTED_DOCUMENT_TYPES,
     cache_image_from_url,
     cache_audio_from_url,
 )
@@ -134,11 +137,140 @@ class WhatsAppAdapter(BasePlatformAdapter):
         )
         self._session_path: Path = Path(config.extra.get(
             "session_path",
-            get_hermes_home() / "whatsapp" / "session"
+            get_hermes_dir("platforms/whatsapp/session", "whatsapp/session")
         ))
+        self._reply_prefix: Optional[str] = config.extra.get("reply_prefix")
+        self._mention_patterns = self._compile_mention_patterns()
         self._message_queue: asyncio.Queue = asyncio.Queue()
         self._bridge_log_fh = None
         self._bridge_log: Optional[Path] = None
+        self._poll_task: Optional[asyncio.Task] = None
+        self._http_session: Optional["aiohttp.ClientSession"] = None
+        self._session_lock_identity: Optional[str] = None
+
+    def _whatsapp_require_mention(self) -> bool:
+        configured = self.config.extra.get("require_mention")
+        if configured is not None:
+            if isinstance(configured, str):
+                return configured.lower() in ("true", "1", "yes", "on")
+            return bool(configured)
+        return os.getenv("WHATSAPP_REQUIRE_MENTION", "false").lower() in ("true", "1", "yes", "on")
+
+    def _whatsapp_free_response_chats(self) -> set[str]:
+        raw = self.config.extra.get("free_response_chats")
+        if raw is None:
+            raw = os.getenv("WHATSAPP_FREE_RESPONSE_CHATS", "")
+        if isinstance(raw, list):
+            return {str(part).strip() for part in raw if str(part).strip()}
+        return {part.strip() for part in str(raw).split(",") if part.strip()}
+
+    def _compile_mention_patterns(self):
+        patterns = self.config.extra.get("mention_patterns")
+        if patterns is None:
+            raw = os.getenv("WHATSAPP_MENTION_PATTERNS", "").strip()
+            if raw:
+                try:
+                    patterns = json.loads(raw)
+                except Exception:
+                    patterns = [part.strip() for part in raw.splitlines() if part.strip()]
+                    if not patterns:
+                        patterns = [part.strip() for part in raw.split(",") if part.strip()]
+        if patterns is None:
+            return []
+        if isinstance(patterns, str):
+            patterns = [patterns]
+        if not isinstance(patterns, list):
+            logger.warning("[%s] whatsapp mention_patterns must be a list or string; got %s", self.name, type(patterns).__name__)
+            return []
+
+        compiled = []
+        for pattern in patterns:
+            if not isinstance(pattern, str) or not pattern.strip():
+                continue
+            try:
+                compiled.append(re.compile(pattern, re.IGNORECASE))
+            except re.error as exc:
+                logger.warning("[%s] Invalid WhatsApp mention pattern %r: %s", self.name, pattern, exc)
+        if compiled:
+            logger.info("[%s] Loaded %d WhatsApp mention pattern(s)", self.name, len(compiled))
+        return compiled
+
+    @staticmethod
+    def _normalize_whatsapp_id(value: Optional[str]) -> str:
+        if not value:
+            return ""
+        normalized = str(value).strip()
+        if ":" in normalized and "@" in normalized:
+            normalized = normalized.replace(":", "@", 1)
+        return normalized
+
+    def _bot_ids_from_message(self, data: Dict[str, Any]) -> set[str]:
+        bot_ids = set()
+        for candidate in data.get("botIds") or []:
+            normalized = self._normalize_whatsapp_id(candidate)
+            if normalized:
+                bot_ids.add(normalized)
+        return bot_ids
+
+    def _message_is_reply_to_bot(self, data: Dict[str, Any]) -> bool:
+        quoted_participant = self._normalize_whatsapp_id(data.get("quotedParticipant"))
+        if not quoted_participant:
+            return False
+        return quoted_participant in self._bot_ids_from_message(data)
+
+    def _message_mentions_bot(self, data: Dict[str, Any]) -> bool:
+        bot_ids = self._bot_ids_from_message(data)
+        if not bot_ids:
+            return False
+        mentioned_ids = {
+            nid
+            for candidate in (data.get("mentionedIds") or [])
+            if (nid := self._normalize_whatsapp_id(candidate))
+        }
+        if mentioned_ids & bot_ids:
+            return True
+
+        body = str(data.get("body") or "")
+        lower_body = body.lower()
+        for bot_id in bot_ids:
+            bare_id = bot_id.split("@", 1)[0].lower()
+            if bare_id and (f"@{bare_id}" in lower_body or bare_id in lower_body):
+                return True
+        return False
+
+    def _message_matches_mention_patterns(self, data: Dict[str, Any]) -> bool:
+        if not self._mention_patterns:
+            return False
+        body = str(data.get("body") or "")
+        return any(pattern.search(body) for pattern in self._mention_patterns)
+
+    def _clean_bot_mention_text(self, text: str, data: Dict[str, Any]) -> str:
+        if not text:
+            return text
+        bot_ids = self._bot_ids_from_message(data)
+        cleaned = text
+        for bot_id in bot_ids:
+            bare_id = bot_id.split("@", 1)[0]
+            if bare_id:
+                cleaned = re.sub(rf"@{re.escape(bare_id)}\b[,:\-]*\s*", "", cleaned)
+        return cleaned.strip() or text
+
+    def _should_process_message(self, data: Dict[str, Any]) -> bool:
+        if not data.get("isGroup"):
+            return True
+        chat_id = str(data.get("chatId") or "")
+        if chat_id in self._whatsapp_free_response_chats():
+            return True
+        if not self._whatsapp_require_mention():
+            return True
+        body = str(data.get("body") or "").strip()
+        if body.startswith("/"):
+            return True
+        if self._message_is_reply_to_bot(data):
+            return True
+        if self._message_mentions_bot(data):
+            return True
+        return self._message_matches_mention_patterns(data)
     
     async def connect(self) -> bool:
         """
@@ -157,6 +289,29 @@ class WhatsAppAdapter(BasePlatformAdapter):
         
         logger.info("[%s] Bridge found at %s", self.name, bridge_path)
         
+        # Acquire scoped lock to prevent duplicate sessions
+        try:
+            from gateway.status import acquire_scoped_lock
+
+            self._session_lock_identity = str(self._session_path)
+            acquired, existing = acquire_scoped_lock(
+                "whatsapp-session",
+                self._session_lock_identity,
+                metadata={"platform": self.platform.value},
+            )
+            if not acquired:
+                owner_pid = existing.get("pid") if isinstance(existing, dict) else None
+                message = (
+                    "Another local Hermes gateway is already using this WhatsApp session"
+                    + (f" (PID {owner_pid})." if owner_pid else ".")
+                    + " Stop the other gateway before starting a second WhatsApp bridge."
+                )
+                logger.error("[%s] %s", self.name, message)
+                self._set_fatal_error("whatsapp_session_lock", message, retryable=False)
+                return False
+        except Exception as e:
+            logger.warning("[%s] Could not acquire session lock (non-fatal): %s", self.name, e)
+
         # Auto-install npm dependencies if node_modules doesn't exist
         bridge_dir = bridge_path.parent
         if not (bridge_dir / "node_modules").exists():
@@ -181,9 +336,32 @@ class WhatsAppAdapter(BasePlatformAdapter):
             # Ensure session directory exists
             self._session_path.mkdir(parents=True, exist_ok=True)
             
+            # Check if bridge is already running and connected
+            import aiohttp
+            import asyncio
+            try:
+                async with aiohttp.ClientSession() as session:
+                    async with session.get(
+                        f"http://127.0.0.1:{self._bridge_port}/health",
+                        timeout=aiohttp.ClientTimeout(total=2)
+                    ) as resp:
+                        if resp.status == 200:
+                            data = await resp.json()
+                            bridge_status = data.get("status", "unknown")
+                            if bridge_status == "connected":
+                                print(f"[{self.name}] Using existing bridge (status: {bridge_status})")
+                                self._mark_connected()
+                                self._bridge_process = None  # Not managed by us
+                                self._http_session = aiohttp.ClientSession()
+                                self._poll_task = asyncio.create_task(self._poll_messages())
+                                return True
+                            else:
+                                print(f"[{self.name}] Bridge found but not connected (status: {bridge_status}), restarting")
+            except Exception:
+                pass  # Bridge not running, start a new one
+            
             # Kill any orphaned bridge from a previous gateway run
             _kill_port_process(self._bridge_port)
-            import asyncio
             await asyncio.sleep(1)
             
             # Start the bridge process in its own process group.
@@ -193,6 +371,14 @@ class WhatsAppAdapter(BasePlatformAdapter):
             self._bridge_log = self._session_path.parent / "bridge.log"
             bridge_log_fh = open(self._bridge_log, "a")
             self._bridge_log_fh = bridge_log_fh
+
+            # Build bridge subprocess environment.
+            # Pass WHATSAPP_REPLY_PREFIX from config.yaml so the Node bridge
+            # can use it without the user needing to set a separate env var.
+            bridge_env = os.environ.copy()
+            if self._reply_prefix is not None:
+                bridge_env["WHATSAPP_REPLY_PREFIX"] = self._reply_prefix
+
             self._bridge_process = subprocess.Popen(
                 [
                     "node",
@@ -204,6 +390,7 @@ class WhatsAppAdapter(BasePlatformAdapter):
                 stdout=bridge_log_fh,
                 stderr=bridge_log_fh,
                 preexec_fn=None if _IS_WINDOWS else os.setsid,
+                env=bridge_env,
             )
             
             # Wait for the bridge to connect to WhatsApp.
@@ -222,7 +409,7 @@ class WhatsAppAdapter(BasePlatformAdapter):
                 try:
                     async with aiohttp.ClientSession() as session:
                         async with session.get(
-                            f"http://localhost:{self._bridge_port}/health",
+                            f"http://127.0.0.1:{self._bridge_port}/health",
                             timeout=aiohttp.ClientTimeout(total=2)
                         ) as resp:
                             if resp.status == 200:
@@ -254,7 +441,7 @@ class WhatsAppAdapter(BasePlatformAdapter):
                     try:
                         async with aiohttp.ClientSession() as session:
                             async with session.get(
-                                f"http://localhost:{self._bridge_port}/health",
+                                f"http://127.0.0.1:{self._bridge_port}/health",
                                 timeout=aiohttp.ClientTimeout(total=2)
                             ) as resp:
                                 if resp.status == 200:
@@ -271,14 +458,23 @@ class WhatsAppAdapter(BasePlatformAdapter):
                     print(f"[{self.name}]   Bridge log: {self._bridge_log}")
                     print(f"[{self.name}]   If session expired, re-pair: hermes whatsapp")
             
+            # Create a persistent HTTP session for all bridge communication
+            self._http_session = aiohttp.ClientSession()
+
             # Start message polling task
-            asyncio.create_task(self._poll_messages())
+            self._poll_task = asyncio.create_task(self._poll_messages())
             
-            self._running = True
+            self._mark_connected()
             print(f"[{self.name}] Bridge started on port {self._bridge_port}")
             return True
             
         except Exception as e:
+            if self._session_lock_identity:
+                try:
+                    from gateway.status import release_scoped_lock
+                    release_scoped_lock("whatsapp-session", self._session_lock_identity)
+                except Exception:
+                    pass
             logger.error("[%s] Failed to start bridge: %s", self.name, e, exc_info=True)
             self._close_bridge_log()
             return False
@@ -292,6 +488,23 @@ class WhatsAppAdapter(BasePlatformAdapter):
                 pass
             self._bridge_log_fh = None
 
+    async def _check_managed_bridge_exit(self) -> Optional[str]:
+        """Return a fatal error message if the managed bridge child exited."""
+        if self._bridge_process is None:
+            return None
+
+        returncode = self._bridge_process.poll()
+        if returncode is None:
+            return None
+
+        message = f"WhatsApp bridge process exited unexpectedly (code {returncode})."
+        if not self.has_fatal_error:
+            logger.error("[%s] %s", self.name, message)
+            self._set_fatal_error("whatsapp_bridge_exited", message, retryable=True)
+            self._close_bridge_log()
+            await self._notify_fatal_error()
+        return self.fatal_error_message or message
+
     async def disconnect(self) -> None:
         """Stop the WhatsApp bridge and clean up any orphaned processes."""
         if self._bridge_process:
@@ -316,13 +529,35 @@ class WhatsAppAdapter(BasePlatformAdapter):
                         self._bridge_process.kill()
             except Exception as e:
                 print(f"[{self.name}] Error stopping bridge: {e}")
-        
-        # Also kill any orphaned bridge processes on our port
-        _kill_port_process(self._bridge_port)
-        
-        self._running = False
+        else:
+            # Bridge was not started by us, don't kill it
+            print(f"[{self.name}] Disconnecting (external bridge left running)")
+
+        # Cancel the poll task explicitly
+        if self._poll_task and not self._poll_task.done():
+            self._poll_task.cancel()
+            try:
+                await self._poll_task
+            except (asyncio.CancelledError, Exception):
+                pass
+        self._poll_task = None
+
+        # Close the persistent HTTP session
+        if self._http_session and not self._http_session.closed:
+            await self._http_session.close()
+        self._http_session = None
+
+        if self._session_lock_identity:
+            try:
+                from gateway.status import release_scoped_lock
+                release_scoped_lock("whatsapp-session", self._session_lock_identity)
+            except Exception as e:
+                logger.warning("[%s] Error releasing WhatsApp session lock: %s", self.name, e, exc_info=True)
+
+        self._mark_disconnected()
         self._bridge_process = None
         self._close_bridge_log()
+        self._session_lock_identity = None
         print(f"[{self.name}] Disconnected")
     
     async def send(
@@ -333,41 +568,37 @@ class WhatsAppAdapter(BasePlatformAdapter):
         metadata: Optional[Dict[str, Any]] = None
     ) -> SendResult:
         """Send a message via the WhatsApp bridge."""
-        if not self._running:
+        if not self._running or not self._http_session:
             return SendResult(success=False, error="Not connected")
+        bridge_exit = await self._check_managed_bridge_exit()
+        if bridge_exit:
+            return SendResult(success=False, error=bridge_exit)
         
         try:
             import aiohttp
+
+            payload = {
+                "chatId": chat_id,
+                "message": content,
+            }
+            if reply_to:
+                payload["replyTo"] = reply_to
             
-            async with aiohttp.ClientSession() as session:
-                payload = {
-                    "chatId": chat_id,
-                    "message": content,
-                }
-                if reply_to:
-                    payload["replyTo"] = reply_to
-                
-                async with session.post(
-                    f"http://localhost:{self._bridge_port}/send",
-                    json=payload,
-                    timeout=aiohttp.ClientTimeout(total=30)
-                ) as resp:
-                    if resp.status == 200:
-                        data = await resp.json()
-                        return SendResult(
-                            success=True,
-                            message_id=data.get("messageId"),
-                            raw_response=data
-                        )
-                    else:
-                        error = await resp.text()
-                        return SendResult(success=False, error=error)
-                        
-        except ImportError:
-            return SendResult(
-                success=False, 
-                error="aiohttp not installed. Run: pip install aiohttp"
-            )
+            async with self._http_session.post(
+                f"http://127.0.0.1:{self._bridge_port}/send",
+                json=payload,
+                timeout=aiohttp.ClientTimeout(total=30)
+            ) as resp:
+                if resp.status == 200:
+                    data = await resp.json()
+                    return SendResult(
+                        success=True,
+                        message_id=data.get("messageId"),
+                        raw_response=data
+                    )
+                else:
+                    error = await resp.text()
+                    return SendResult(success=False, error=error)
         except Exception as e:
             return SendResult(success=False, error=str(e))
 
@@ -378,25 +609,27 @@ class WhatsAppAdapter(BasePlatformAdapter):
         content: str,
     ) -> SendResult:
         """Edit a previously sent message via the WhatsApp bridge."""
-        if not self._running:
+        if not self._running or not self._http_session:
             return SendResult(success=False, error="Not connected")
+        bridge_exit = await self._check_managed_bridge_exit()
+        if bridge_exit:
+            return SendResult(success=False, error=bridge_exit)
         try:
             import aiohttp
-            async with aiohttp.ClientSession() as session:
-                async with session.post(
-                    f"http://localhost:{self._bridge_port}/edit",
-                    json={
-                        "chatId": chat_id,
-                        "messageId": message_id,
-                        "message": content,
-                    },
-                    timeout=aiohttp.ClientTimeout(total=15)
-                ) as resp:
-                    if resp.status == 200:
-                        return SendResult(success=True, message_id=message_id)
-                    else:
-                        error = await resp.text()
-                        return SendResult(success=False, error=error)
+            async with self._http_session.post(
+                f"http://127.0.0.1:{self._bridge_port}/edit",
+                json={
+                    "chatId": chat_id,
+                    "messageId": message_id,
+                    "message": content,
+                },
+                timeout=aiohttp.ClientTimeout(total=15)
+            ) as resp:
+                if resp.status == 200:
+                    return SendResult(success=True, message_id=message_id)
+                else:
+                    error = await resp.text()
+                    return SendResult(success=False, error=error)
         except Exception as e:
             return SendResult(success=False, error=str(e))
 
@@ -409,8 +642,11 @@ class WhatsAppAdapter(BasePlatformAdapter):
         file_name: Optional[str] = None,
     ) -> SendResult:
         """Send any media file via bridge /send-media endpoint."""
-        if not self._running:
+        if not self._running or not self._http_session:
             return SendResult(success=False, error="Not connected")
+        bridge_exit = await self._check_managed_bridge_exit()
+        if bridge_exit:
+            return SendResult(success=False, error=bridge_exit)
         try:
             import aiohttp
 
@@ -427,22 +663,21 @@ class WhatsAppAdapter(BasePlatformAdapter):
             if file_name:
                 payload["fileName"] = file_name
 
-            async with aiohttp.ClientSession() as session:
-                async with session.post(
-                    f"http://localhost:{self._bridge_port}/send-media",
-                    json=payload,
-                    timeout=aiohttp.ClientTimeout(total=120),
-                ) as resp:
-                    if resp.status == 200:
-                        data = await resp.json()
-                        return SendResult(
-                            success=True,
-                            message_id=data.get("messageId"),
-                            raw_response=data,
-                        )
-                    else:
-                        error = await resp.text()
-                        return SendResult(success=False, error=error)
+            async with self._http_session.post(
+                f"http://127.0.0.1:{self._bridge_port}/send-media",
+                json=payload,
+                timeout=aiohttp.ClientTimeout(total=120),
+            ) as resp:
+                if resp.status == 200:
+                    data = await resp.json()
+                    return SendResult(
+                        success=True,
+                        message_id=data.get("messageId"),
+                        raw_response=data,
+                    )
+                else:
+                    error = await resp.text()
+                    return SendResult(success=False, error=error)
 
         except Exception as e:
             return SendResult(success=False, error=str(e))
@@ -467,6 +702,7 @@ class WhatsAppAdapter(BasePlatformAdapter):
         image_path: str,
         caption: Optional[str] = None,
         reply_to: Optional[str] = None,
+        **kwargs,
     ) -> SendResult:
         """Send a local image file natively via bridge."""
         return await self._send_media_to_bridge(chat_id, image_path, "image", caption)
@@ -477,6 +713,7 @@ class WhatsAppAdapter(BasePlatformAdapter):
         video_path: str,
         caption: Optional[str] = None,
         reply_to: Optional[str] = None,
+        **kwargs,
     ) -> SendResult:
         """Send a video natively via bridge — plays inline in WhatsApp."""
         return await self._send_media_to_bridge(chat_id, video_path, "video", caption)
@@ -488,6 +725,7 @@ class WhatsAppAdapter(BasePlatformAdapter):
         caption: Optional[str] = None,
         file_name: Optional[str] = None,
         reply_to: Optional[str] = None,
+        **kwargs,
     ) -> SendResult:
         """Send a document/file as a downloadable attachment via bridge."""
         return await self._send_media_to_bridge(
@@ -497,41 +735,43 @@ class WhatsAppAdapter(BasePlatformAdapter):
 
     async def send_typing(self, chat_id: str, metadata=None) -> None:
         """Send typing indicator via bridge."""
-        if not self._running:
+        if not self._running or not self._http_session:
+            return
+        if await self._check_managed_bridge_exit():
             return
         
         try:
             import aiohttp
-            
-            async with aiohttp.ClientSession() as session:
-                await session.post(
-                    f"http://localhost:{self._bridge_port}/typing",
-                    json={"chatId": chat_id},
-                    timeout=aiohttp.ClientTimeout(total=5)
-                )
+
+            await self._http_session.post(
+                f"http://127.0.0.1:{self._bridge_port}/typing",
+                json={"chatId": chat_id},
+                timeout=aiohttp.ClientTimeout(total=5)
+            )
         except Exception:
             pass  # Ignore typing indicator failures
     
     async def get_chat_info(self, chat_id: str) -> Dict[str, Any]:
         """Get information about a WhatsApp chat."""
-        if not self._running:
+        if not self._running or not self._http_session:
             return {"name": "Unknown", "type": "dm"}
+        if await self._check_managed_bridge_exit():
+            return {"name": chat_id, "type": "dm"}
         
         try:
             import aiohttp
-            
-            async with aiohttp.ClientSession() as session:
-                async with session.get(
-                    f"http://localhost:{self._bridge_port}/chat/{chat_id}",
-                    timeout=aiohttp.ClientTimeout(total=10)
-                ) as resp:
-                    if resp.status == 200:
-                        data = await resp.json()
-                        return {
-                            "name": data.get("name", chat_id),
-                            "type": "group" if data.get("isGroup") else "dm",
-                            "participants": data.get("participants", []),
-                        }
+
+            async with self._http_session.get(
+                f"http://127.0.0.1:{self._bridge_port}/chat/{chat_id}",
+                timeout=aiohttp.ClientTimeout(total=10)
+            ) as resp:
+                if resp.status == 200:
+                    data = await resp.json()
+                    return {
+                        "name": data.get("name", chat_id),
+                        "type": "group" if data.get("isGroup") else "dm",
+                        "participants": data.get("participants", []),
+                    }
         except Exception as e:
             logger.debug("Could not get WhatsApp chat info for %s: %s", chat_id, e)
         
@@ -539,28 +779,33 @@ class WhatsAppAdapter(BasePlatformAdapter):
     
     async def _poll_messages(self) -> None:
         """Poll the bridge for incoming messages."""
-        try:
-            import aiohttp
-        except ImportError:
-            print(f"[{self.name}] aiohttp not installed, message polling disabled")
-            return
-        
+        import aiohttp
+
         while self._running:
+            if not self._http_session:
+                break
+            bridge_exit = await self._check_managed_bridge_exit()
+            if bridge_exit:
+                print(f"[{self.name}] {bridge_exit}")
+                break
             try:
-                async with aiohttp.ClientSession() as session:
-                    async with session.get(
-                        f"http://localhost:{self._bridge_port}/messages",
-                        timeout=aiohttp.ClientTimeout(total=30)
-                    ) as resp:
-                        if resp.status == 200:
-                            messages = await resp.json()
-                            for msg_data in messages:
-                                event = await self._build_message_event(msg_data)
-                                if event:
-                                    await self.handle_message(event)
+                async with self._http_session.get(
+                    f"http://127.0.0.1:{self._bridge_port}/messages",
+                    timeout=aiohttp.ClientTimeout(total=30)
+                ) as resp:
+                    if resp.status == 200:
+                        messages = await resp.json()
+                        for msg_data in messages:
+                            event = await self._build_message_event(msg_data)
+                            if event:
+                                await self.handle_message(event)
             except asyncio.CancelledError:
                 break
             except Exception as e:
+                bridge_exit = await self._check_managed_bridge_exit()
+                if bridge_exit:
+                    print(f"[{self.name}] {bridge_exit}")
+                    break
                 print(f"[{self.name}] Poll error: {e}")
                 await asyncio.sleep(5)
             
@@ -569,6 +814,9 @@ class WhatsAppAdapter(BasePlatformAdapter):
     async def _build_message_event(self, data: Dict[str, Any]) -> Optional[MessageEvent]:
         """Build a MessageEvent from bridge message data, downloading images to cache."""
         try:
+            if not self._should_process_message(data):
+                return None
+
             # Determine message type
             msg_type = MessageType.TEXT
             if data.get("hasMedia"):
@@ -595,7 +843,7 @@ class WhatsAppAdapter(BasePlatformAdapter):
                 user_name=data.get("senderName"),
             )
             
-            # Download image media URLs to the local cache so the vision tool
+            # Download media URLs to the local cache so agent tools
             # can access them reliably regardless of URL expiration.
             raw_urls = data.get("mediaUrls", [])
             cached_urls = []
@@ -611,6 +859,11 @@ class WhatsAppAdapter(BasePlatformAdapter):
                         print(f"[{self.name}] Failed to cache image: {e}", flush=True)
                         cached_urls.append(url)
                         media_types.append("image/jpeg")
+                elif msg_type == MessageType.PHOTO and os.path.isabs(url):
+                    # Local file path — bridge already downloaded the image
+                    cached_urls.append(url)
+                    media_types.append("image/jpeg")
+                    print(f"[{self.name}] Using bridge-cached image: {url}", flush=True)
                 elif msg_type == MessageType.VOICE and url.startswith(("http://", "https://")):
                     try:
                         cached_path = await cache_audio_from_url(url, ext=".ogg")
@@ -621,12 +874,61 @@ class WhatsAppAdapter(BasePlatformAdapter):
                         print(f"[{self.name}] Failed to cache voice: {e}", flush=True)
                         cached_urls.append(url)
                         media_types.append("audio/ogg")
+                elif msg_type == MessageType.VOICE and os.path.isabs(url):
+                    # Local file path — bridge already downloaded the audio
+                    cached_urls.append(url)
+                    media_types.append("audio/ogg")
+                    print(f"[{self.name}] Using bridge-cached audio: {url}", flush=True)
+                elif msg_type == MessageType.DOCUMENT and os.path.isabs(url):
+                    # Local file path — bridge already downloaded the document
+                    cached_urls.append(url)
+                    ext = Path(url).suffix.lower()
+                    mime = SUPPORTED_DOCUMENT_TYPES.get(ext, "application/octet-stream")
+                    media_types.append(mime)
+                    print(f"[{self.name}] Using bridge-cached document: {url}", flush=True)
+                elif msg_type == MessageType.VIDEO and os.path.isabs(url):
+                    cached_urls.append(url)
+                    media_types.append("video/mp4")
+                    print(f"[{self.name}] Using bridge-cached video: {url}", flush=True)
                 else:
                     cached_urls.append(url)
                     media_types.append("unknown")
-            
+
+            # For text-readable documents, inject file content directly into
+            # the message text so the agent can read it inline.
+            # Cap at 100KB to match Telegram/Discord/Slack behaviour.
+            body = data.get("body", "")
+            if data.get("isGroup"):
+                body = self._clean_bot_mention_text(body, data)
+            MAX_TEXT_INJECT_BYTES = 100 * 1024
+            if msg_type == MessageType.DOCUMENT and cached_urls:
+                for doc_path in cached_urls:
+                    ext = Path(doc_path).suffix.lower()
+                    if ext in (".txt", ".md", ".csv", ".json", ".xml", ".yaml", ".yml", ".log", ".py", ".js", ".ts", ".html", ".css"):
+                        try:
+                            file_size = Path(doc_path).stat().st_size
+                            if file_size > MAX_TEXT_INJECT_BYTES:
+                                print(f"[{self.name}] Skipping text injection for {doc_path} ({file_size} bytes > {MAX_TEXT_INJECT_BYTES})", flush=True)
+                                continue
+                            content = Path(doc_path).read_text(errors="replace")
+                            fname = Path(doc_path).name
+                            # Remove the doc_<hex>_ prefix for display
+                            display_name = fname
+                            if "_" in fname:
+                                parts = fname.split("_", 2)
+                                if len(parts) >= 3:
+                                    display_name = parts[2]
+                            injection = f"[Content of {display_name}]:\n{content}"
+                            if body:
+                                body = f"{injection}\n\n{body}"
+                            else:
+                                body = injection
+                            print(f"[{self.name}] Injected text content from: {doc_path}", flush=True)
+                        except Exception as e:
+                            print(f"[{self.name}] Failed to read document text: {e}", flush=True)
+
             return MessageEvent(
-                text=data.get("body", ""),
+                text=body,
                 message_type=msg_type,
                 source=source,
                 raw_message=data,
@@ -637,4 +939,3 @@ class WhatsAppAdapter(BasePlatformAdapter):
         except Exception as e:
             print(f"[{self.name}] Error building event: {e}")
             return None
-

gateway/session.py

@@ -8,21 +8,64 @@ Handles:
 - Dynamic system prompt injection (agent knows its context)
 """
 
+import hashlib
 import logging
 import os
 import json
+import re
+import threading
 import uuid
 from pathlib import Path
 from datetime import datetime, timedelta
-from dataclasses import dataclass, field
+from dataclasses import dataclass
 from typing import Dict, List, Optional, Any
 
 logger = logging.getLogger(__name__)
 
+
+def _now() -> datetime:
+    """Return the current local time."""
+    return datetime.now()
+
+
+# ---------------------------------------------------------------------------
+# PII redaction helpers
+# ---------------------------------------------------------------------------
+
+_PHONE_RE = re.compile(r"^\+?\d[\d\-\s]{6,}$")
+
+
+def _hash_id(value: str) -> str:
+    """Deterministic 12-char hex hash of an identifier."""
+    return hashlib.sha256(value.encode("utf-8")).hexdigest()[:12]
+
+
+def _hash_sender_id(value: str) -> str:
+    """Hash a sender ID to ``user_<12hex>``."""
+    return f"user_{_hash_id(value)}"
+
+
+def _hash_chat_id(value: str) -> str:
+    """Hash the numeric portion of a chat ID, preserving platform prefix.
+
+    ``telegram:12345`` → ``telegram:<hash>``
+    ``12345``          → ``<hash>``
+    """
+    colon = value.find(":")
+    if colon > 0:
+        prefix = value[:colon]
+        return f"{prefix}:{_hash_id(value[colon + 1:])}"
+    return _hash_id(value)
+
+
+def _looks_like_phone(value: str) -> bool:
+    """Return True if *value* looks like a phone number (E.164 or similar)."""
+    return bool(_PHONE_RE.match(value.strip()))
+
 from .config import (
     Platform,
     GatewayConfig,
-    SessionResetPolicy,
+    SessionResetPolicy,  # noqa: F401 — re-exported via gateway/__init__.py
     HomeChannel,
 )
 
@@ -146,7 +189,21 @@ class SessionContext:
         }
 
 
-def build_session_context_prompt(context: SessionContext) -> str:
+_PII_SAFE_PLATFORMS = frozenset({
+    Platform.WHATSAPP,
+    Platform.SIGNAL,
+    Platform.TELEGRAM,
+})
+"""Platforms where user IDs can be safely redacted (no in-message mention system
+that requires raw IDs).  Discord is excluded because mentions use ``<@user_id>``
+and the LLM needs the real ID to tag users."""
+
+
+def build_session_context_prompt(
+    context: SessionContext,
+    *,
+    redact_pii: bool = False,
+) -> str:
     """
     Build the dynamic system prompt section that tells the agent about its context.
     
@@ -154,7 +211,15 @@ def build_session_context_prompt(context: SessionContext) -> str:
     - Where messages are coming from
     - What platforms are connected
     - Where it can deliver scheduled task outputs
+
+    When *redact_pii* is True **and** the source platform is in
+    ``_PII_SAFE_PLATFORMS``, phone numbers are stripped and user/chat IDs
+    are replaced with deterministic hashes before being sent to the LLM.
+    Platforms like Discord are excluded because mentions need real IDs.
+    Routing still uses the original values (they stay in SessionSource).
     """
+    # Only apply redaction on platforms where IDs aren't needed for mentions
+    redact_pii = redact_pii and context.source.platform in _PII_SAFE_PLATFORMS
     lines = [
         "## Current Session Context",
         "",
@@ -165,7 +230,25 @@ def build_session_context_prompt(context: SessionContext) -> str:
     if context.source.platform == Platform.LOCAL:
         lines.append(f"**Source:** {platform_name} (the machine running this agent)")
     else:
-        lines.append(f"**Source:** {platform_name} ({context.source.description})")
+        # Build a description that respects PII redaction
+        src = context.source
+        if redact_pii:
+            # Build a safe description without raw IDs
+            _uname = src.user_name or (
+                _hash_sender_id(src.user_id) if src.user_id else "user"
+            )
+            _cname = src.chat_name or _hash_chat_id(src.chat_id)
+            if src.chat_type == "dm":
+                desc = f"DM with {_uname}"
+            elif src.chat_type == "group":
+                desc = f"group: {_cname}"
+            elif src.chat_type == "channel":
+                desc = f"channel: {_cname}"
+            else:
+                desc = _cname
+        else:
+            desc = src.description
+        lines.append(f"**Source:** {platform_name} ({desc})")
     
     # Channel topic (if available - provides context about the channel's purpose)
     if context.source.chat_topic:
@@ -175,7 +258,10 @@ def build_session_context_prompt(context: SessionContext) -> str:
     if context.source.user_name:
         lines.append(f"**User:** {context.source.user_name}")
     elif context.source.user_id:
-        lines.append(f"**User ID:** {context.source.user_id}")
+        uid = context.source.user_id
+        if redact_pii:
+            uid = _hash_sender_id(uid)
+        lines.append(f"**User ID:** {uid}")
     
     # Platform-specific behavioral notes
     if context.source.platform == Platform.SLACK:
@@ -210,7 +296,8 @@ def build_session_context_prompt(context: SessionContext) -> str:
         lines.append("")
         lines.append("**Home Channels (default destinations):**")
         for platform, home in context.home_channels.items():
-            lines.append(f"  - {platform.value}: {home.name} (ID: {home.chat_id})")
+            hc_id = _hash_chat_id(home.chat_id) if redact_pii else home.chat_id
+            lines.append(f"  - {platform.value}: {home.name} (ID: {hc_id})")
     
     # Delivery options for scheduled tasks
     lines.append("")
@@ -220,7 +307,10 @@ def build_session_context_prompt(context: SessionContext) -> str:
     if context.source.platform == Platform.LOCAL:
         lines.append("- `\"origin\"` → Local output (saved to files)")
     else:
-        lines.append(f"- `\"origin\"` → Back to this chat ({context.source.chat_name or context.source.chat_id})")
+        _origin_label = context.source.chat_name or (
+            _hash_chat_id(context.source.chat_id) if redact_pii else context.source.chat_id
+        )
+        lines.append(f"- `\"origin\"` → Back to this chat ({_origin_label})")
     
     # Local always available
     lines.append("- `\"local\"` → Save to local files only (~/.hermes/cron/output/)")
@@ -259,7 +349,11 @@ class SessionEntry:
     # Token tracking
     input_tokens: int = 0
     output_tokens: int = 0
+    cache_read_tokens: int = 0
+    cache_write_tokens: int = 0
     total_tokens: int = 0
+    estimated_cost_usd: float = 0.0
+    cost_status: str = "unknown"
     
     # Last API-reported prompt tokens (for accurate compression pre-check)
     last_prompt_tokens: int = 0
@@ -267,6 +361,14 @@ class SessionEntry:
     # Set when a session was created because the previous one expired;
     # consumed once by the message handler to inject a notice into context
     was_auto_reset: bool = False
+    auto_reset_reason: Optional[str] = None  # "idle" or "daily"
+    reset_had_activity: bool = False  # whether the expired session had any messages
+    
+    # Set by the background expiry watcher after it successfully flushes
+    # memories for this session.  Persisted to sessions.json so the flag
+    # survives gateway restarts (the old in-memory _pre_flushed_sessions
+    # set was lost on restart, causing redundant re-flushes).
+    memory_flushed: bool = False
     
     def to_dict(self) -> Dict[str, Any]:
         result = {
@@ -279,8 +381,13 @@ class SessionEntry:
             "chat_type": self.chat_type,
             "input_tokens": self.input_tokens,
             "output_tokens": self.output_tokens,
+            "cache_read_tokens": self.cache_read_tokens,
+            "cache_write_tokens": self.cache_write_tokens,
             "total_tokens": self.total_tokens,
             "last_prompt_tokens": self.last_prompt_tokens,
+            "estimated_cost_usd": self.estimated_cost_usd,
+            "cost_status": self.cost_status,
+            "memory_flushed": self.memory_flushed,
         }
         if self.origin:
             result["origin"] = self.origin.to_dict()
@@ -310,36 +417,57 @@ class SessionEntry:
             chat_type=data.get("chat_type", "dm"),
             input_tokens=data.get("input_tokens", 0),
             output_tokens=data.get("output_tokens", 0),
+            cache_read_tokens=data.get("cache_read_tokens", 0),
+            cache_write_tokens=data.get("cache_write_tokens", 0),
             total_tokens=data.get("total_tokens", 0),
             last_prompt_tokens=data.get("last_prompt_tokens", 0),
+            estimated_cost_usd=data.get("estimated_cost_usd", 0.0),
+            cost_status=data.get("cost_status", "unknown"),
+            memory_flushed=data.get("memory_flushed", False),
         )
 
 
-def build_session_key(source: SessionSource) -> str:
+def build_session_key(source: SessionSource, group_sessions_per_user: bool = True) -> str:
     """Build a deterministic session key from a message source.
 
     This is the single source of truth for session key construction.
 
     DM rules:
-      - WhatsApp DMs include chat_id (multi-user support).
-      - Other DMs include thread_id when present (e.g. Slack threaded DMs),
-        so each DM thread gets its own session while top-level DMs share one.
-      - Without thread_id or chat_id, all DMs share a single session.
+      - DMs include chat_id when present, so each private conversation is isolated.
+      - thread_id further differentiates threaded DMs within the same DM chat.
+      - Without chat_id, thread_id is used as a best-effort fallback.
+      - Without thread_id or chat_id, DMs share a single session.
 
     Group/channel rules:
-      - thread_id differentiates threads within a channel.
-      - Without thread_id, all messages in a channel share one session.
+      - chat_id identifies the parent group/channel.
+      - user_id/user_id_alt isolates participants within that parent chat when available when
+        ``group_sessions_per_user`` is enabled.
+      - thread_id differentiates threads within that parent chat.
+      - Without participant identifiers, or when isolation is disabled, messages fall back to one
+        shared session per chat.
+      - Without identifiers, messages fall back to one session per platform/chat_type.
     """
     platform = source.platform.value
     if source.chat_type == "dm":
+        if source.chat_id:
+            if source.thread_id:
+                return f"agent:main:{platform}:dm:{source.chat_id}:{source.thread_id}"
+            return f"agent:main:{platform}:dm:{source.chat_id}"
         if source.thread_id:
             return f"agent:main:{platform}:dm:{source.thread_id}"
-        if platform == "whatsapp" and source.chat_id:
-            return f"agent:main:{platform}:dm:{source.chat_id}"
         return f"agent:main:{platform}:dm"
+
+    participant_id = source.user_id_alt or source.user_id
+    key_parts = ["agent:main", platform, source.chat_type]
+
+    if source.chat_id:
+        key_parts.append(source.chat_id)
     if source.thread_id:
-        return f"agent:main:{platform}:{source.chat_type}:{source.chat_id}:{source.thread_id}"
-    return f"agent:main:{platform}:{source.chat_type}:{source.chat_id}"
+        key_parts.append(source.thread_id)
+    if group_sessions_per_user and participant_id:
+        key_parts.append(str(participant_id))
+
+    return ":".join(key_parts)
 
 
 class SessionStore:
@@ -357,10 +485,8 @@ class SessionStore:
         self.config = config
         self._entries: Dict[str, SessionEntry] = {}
         self._loaded = False
+        self._lock = threading.Lock()
         self._has_active_processes_fn = has_active_processes_fn
-        # on_auto_reset is deprecated — memory flush now runs proactively
-        # via the background session expiry watcher in GatewayRunner.
-        self._pre_flushed_sessions: set = set()  # session_ids already flushed by watcher
         
         # Initialize SQLite session database
         self._db = None
@@ -372,12 +498,17 @@ class SessionStore:
     
     def _ensure_loaded(self) -> None:
         """Load sessions index from disk if not already loaded."""
+        with self._lock:
+            self._ensure_loaded_locked()
+
+    def _ensure_loaded_locked(self) -> None:
+        """Load sessions index from disk. Must be called with self._lock held."""
         if self._loaded:
             return
-        
+
         self.sessions_dir.mkdir(parents=True, exist_ok=True)
         sessions_file = self.sessions_dir / "sessions.json"
-        
+
         if sessions_file.exists():
             try:
                 with open(sessions_file, "r", encoding="utf-8") as f:
@@ -390,7 +521,7 @@ class SessionStore:
                             continue
             except Exception as e:
                 print(f"[gateway] Warning: Failed to load sessions: {e}")
-        
+
         self._loaded = True
     
     def _save(self) -> None:
@@ -418,7 +549,10 @@ class SessionStore:
     
     def _generate_session_key(self, source: SessionSource) -> str:
         """Generate a session key from a source."""
-        return build_session_key(source)
+        return build_session_key(
+            source,
+            group_sessions_per_user=getattr(self.config, "group_sessions_per_user", True),
+        )
     
     def _is_session_expired(self, entry: SessionEntry) -> bool:
         """Check if a session has expired based on its reset policy.
@@ -439,7 +573,7 @@ class SessionStore:
         if policy.mode == "none":
             return False
 
-        now = datetime.now()
+        now = _now()
 
         if policy.mode in ("idle", "both"):
             idle_deadline = entry.updated_at + timedelta(minutes=policy.idle_minutes)
@@ -458,16 +592,19 @@ class SessionStore:
 
         return False
 
-    def _should_reset(self, entry: SessionEntry, source: SessionSource) -> bool:
+    def _should_reset(self, entry: SessionEntry, source: SessionSource) -> Optional[str]:
         """
         Check if a session should be reset based on policy.
         
+        Returns the reset reason ("idle" or "daily") if a reset is needed,
+        or None if the session is still valid.
+        
         Sessions with active background processes are never reset.
         """
         if self._has_active_processes_fn:
             session_key = self._generate_session_key(source)
             if self._has_active_processes_fn(session_key):
-                return False
+                return None
 
         policy = self.config.get_reset_policy(
             platform=source.platform,
@@ -475,14 +612,14 @@ class SessionStore:
         )
         
         if policy.mode == "none":
-            return False
+            return None
         
-        now = datetime.now()
+        now = _now()
         
         if policy.mode in ("idle", "both"):
             idle_deadline = entry.updated_at + timedelta(minutes=policy.idle_minutes)
             if now > idle_deadline:
-                return True
+                return "idle"
         
         if policy.mode in ("daily", "both"):
             today_reset = now.replace(
@@ -495,9 +632,9 @@ class SessionStore:
                 today_reset -= timedelta(days=1)
             
             if entry.updated_at < today_reset:
-                return True
+                return "daily"
         
-        return False
+        return None
     
     def has_any_sessions(self) -> bool:
         """Check if any sessions have ever been created (across all platforms).
@@ -517,151 +654,192 @@ class SessionStore:
                 pass  # fall through to heuristic
         # Fallback: check if sessions.json was loaded with existing data.
         # This covers the rare case where the DB is unavailable.
-        self._ensure_loaded()
-        return len(self._entries) > 1
-    
+        with self._lock:
+            self._ensure_loaded_locked()
+            return len(self._entries) > 1
+
     def get_or_create_session(
-        self, 
+        self,
         source: SessionSource,
         force_new: bool = False
     ) -> SessionEntry:
         """
         Get an existing session or create a new one.
-        
+
         Evaluates reset policy to determine if the existing session is stale.
         Creates a session record in SQLite when a new session starts.
         """
-        self._ensure_loaded()
-        
         session_key = self._generate_session_key(source)
-        now = datetime.now()
-        
-        if session_key in self._entries and not force_new:
-            entry = self._entries[session_key]
-            
-            if not self._should_reset(entry, source):
-                entry.updated_at = now
-                self._save()
-                return entry
+        now = _now()
+
+        # SQLite calls are made outside the lock to avoid holding it during I/O.
+        # All _entries / _loaded mutations are protected by self._lock.
+        db_end_session_id = None
+        db_create_kwargs = None
+
+        with self._lock:
+            self._ensure_loaded_locked()
+
+            if session_key in self._entries and not force_new:
+                entry = self._entries[session_key]
+
+                reset_reason = self._should_reset(entry, source)
+                if not reset_reason:
+                    entry.updated_at = now
+                    self._save()
+                    return entry
+                else:
+                    # Session is being auto-reset.
+                    was_auto_reset = True
+                    auto_reset_reason = reset_reason
+                    # Track whether the expired session had any real conversation
+                    reset_had_activity = entry.total_tokens > 0
+                    db_end_session_id = entry.session_id
             else:
-                # Session is being auto-reset.  The background expiry watcher
-                # should have already flushed memories proactively; discard
-                # the marker so it doesn't accumulate.
-                was_auto_reset = True
-                self._pre_flushed_sessions.discard(entry.session_id)
-                if self._db:
-                    try:
-                        self._db.end_session(entry.session_id, "session_reset")
-                    except Exception as e:
-                        logger.debug("Session DB operation failed: %s", e)
-        else:
-            was_auto_reset = False
-        
-        # Create new session
-        session_id = f"{now.strftime('%Y%m%d_%H%M%S')}_{uuid.uuid4().hex[:8]}"
-        
-        entry = SessionEntry(
-            session_key=session_key,
-            session_id=session_id,
-            created_at=now,
-            updated_at=now,
-            origin=source,
-            display_name=source.chat_name,
-            platform=source.platform,
-            chat_type=source.chat_type,
-            was_auto_reset=was_auto_reset,
-        )
-        
-        self._entries[session_key] = entry
-        self._save()
-        
-        # Create session in SQLite
-        if self._db:
+                was_auto_reset = False
+                auto_reset_reason = None
+                reset_had_activity = False
+
+            # Create new session
+            session_id = f"{now.strftime('%Y%m%d_%H%M%S')}_{uuid.uuid4().hex[:8]}"
+
+            entry = SessionEntry(
+                session_key=session_key,
+                session_id=session_id,
+                created_at=now,
+                updated_at=now,
+                origin=source,
+                display_name=source.chat_name,
+                platform=source.platform,
+                chat_type=source.chat_type,
+                was_auto_reset=was_auto_reset,
+                auto_reset_reason=auto_reset_reason,
+                reset_had_activity=reset_had_activity,
+            )
+
+            self._entries[session_key] = entry
+            self._save()
+            db_create_kwargs = {
+                "session_id": session_id,
+                "source": source.platform.value,
+                "user_id": source.user_id,
+            }
+
+        # SQLite operations outside the lock
+        if self._db and db_end_session_id:
             try:
-                self._db.create_session(
-                    session_id=session_id,
-                    source=source.platform.value,
-                    user_id=source.user_id,
-                )
+                self._db.end_session(db_end_session_id, "session_reset")
+            except Exception as e:
+                logger.debug("Session DB operation failed: %s", e)
+
+        if self._db and db_create_kwargs:
+            try:
+                self._db.create_session(**db_create_kwargs)
             except Exception as e:
                 print(f"[gateway] Warning: Failed to create SQLite session: {e}")
-        
-        return entry
-    
-    def update_session(
-        self, 
-        session_key: str,
-        input_tokens: int = 0,
-        output_tokens: int = 0,
-        last_prompt_tokens: int = None,
-        model: str = None,
-    ) -> None:
-        """Update a session's metadata after an interaction."""
-        self._ensure_loaded()
-        
-        if session_key in self._entries:
-            entry = self._entries[session_key]
-            entry.updated_at = datetime.now()
-            entry.input_tokens += input_tokens
-            entry.output_tokens += output_tokens
-            if last_prompt_tokens is not None:
-                entry.last_prompt_tokens = last_prompt_tokens
-            entry.total_tokens = entry.input_tokens + entry.output_tokens
-            self._save()
-            
-            if self._db:
+
+        # Seed new DM thread sessions with parent DM session history.
+        # When a bot reply creates a Slack thread and the user responds in it,
+        # the thread gets a new session (keyed by thread_ts).  Without seeding,
+        # the thread session starts with zero context — the user's original
+        # question and the bot's answer are invisible.  Fix: copy the parent
+        # DM session's transcript into the new thread session so context carries
+        # over while still keeping threads isolated from each other.
+        if (
+            source.chat_type == "dm"
+            and source.thread_id
+            and entry.created_at == entry.updated_at  # brand-new session
+            and not was_auto_reset
+        ):
+            parent_source = SessionSource(
+                platform=source.platform,
+                chat_id=source.chat_id,
+                chat_type="dm",
+                user_id=source.user_id,
+                # no thread_id — this is the parent DM session
+            )
+            parent_key = self._generate_session_key(parent_source)
+            with self._lock:
+                parent_entry = self._entries.get(parent_key)
+            if parent_entry and parent_entry.session_id != entry.session_id:
                 try:
-                    self._db.update_token_counts(
-                        entry.session_id, input_tokens, output_tokens,
-                        model=model,
-                    )
+                    parent_history = self.load_transcript(parent_entry.session_id)
+                    if parent_history:
+                        self.rewrite_transcript(entry.session_id, parent_history)
+                        logger.info(
+                            "[Session] Seeded DM thread session %s with %d messages from parent %s",
+                            entry.session_id, len(parent_history), parent_entry.session_id,
+                        )
                 except Exception as e:
-                    logger.debug("Session DB operation failed: %s", e)
-    
+                    logger.warning("[Session] Failed to seed thread session: %s", e)
+
+        return entry
+
+    def update_session(
+        self,
+        session_key: str,
+        last_prompt_tokens: int = None,
+    ) -> None:
+        """Update lightweight session metadata after an interaction."""
+        with self._lock:
+            self._ensure_loaded_locked()
+
+            if session_key in self._entries:
+                entry = self._entries[session_key]
+                entry.updated_at = _now()
+                if last_prompt_tokens is not None:
+                    entry.last_prompt_tokens = last_prompt_tokens
+                self._save()
+
     def reset_session(self, session_key: str) -> Optional[SessionEntry]:
         """Force reset a session, creating a new session ID."""
-        self._ensure_loaded()
-        
-        if session_key not in self._entries:
-            return None
-        
-        old_entry = self._entries[session_key]
-        
-        # End old session in SQLite
-        if self._db:
+        db_end_session_id = None
+        db_create_kwargs = None
+        new_entry = None
+
+        with self._lock:
+            self._ensure_loaded_locked()
+
+            if session_key not in self._entries:
+                return None
+
+            old_entry = self._entries[session_key]
+            db_end_session_id = old_entry.session_id
+
+            now = _now()
+            session_id = f"{now.strftime('%Y%m%d_%H%M%S')}_{uuid.uuid4().hex[:8]}"
+
+            new_entry = SessionEntry(
+                session_key=session_key,
+                session_id=session_id,
+                created_at=now,
+                updated_at=now,
+                origin=old_entry.origin,
+                display_name=old_entry.display_name,
+                platform=old_entry.platform,
+                chat_type=old_entry.chat_type,
+            )
+
+            self._entries[session_key] = new_entry
+            self._save()
+            db_create_kwargs = {
+                "session_id": session_id,
+                "source": old_entry.platform.value if old_entry.platform else "unknown",
+                "user_id": old_entry.origin.user_id if old_entry.origin else None,
+            }
+
+        if self._db and db_end_session_id:
             try:
-                self._db.end_session(old_entry.session_id, "session_reset")
+                self._db.end_session(db_end_session_id, "session_reset")
             except Exception as e:
                 logger.debug("Session DB operation failed: %s", e)
-        
-        now = datetime.now()
-        session_id = f"{now.strftime('%Y%m%d_%H%M%S')}_{uuid.uuid4().hex[:8]}"
-        
-        new_entry = SessionEntry(
-            session_key=session_key,
-            session_id=session_id,
-            created_at=now,
-            updated_at=now,
-            origin=old_entry.origin,
-            display_name=old_entry.display_name,
-            platform=old_entry.platform,
-            chat_type=old_entry.chat_type,
-        )
-        
-        self._entries[session_key] = new_entry
-        self._save()
-        
-        # Create new session in SQLite
-        if self._db:
+
+        if self._db and db_create_kwargs:
             try:
-                self._db.create_session(
-                    session_id=session_id,
-                    source=old_entry.platform.value if old_entry.platform else "unknown",
-                    user_id=old_entry.origin.user_id if old_entry.origin else None,
-                )
+                self._db.create_session(**db_create_kwargs)
             except Exception as e:
                 logger.debug("Session DB operation failed: %s", e)
-        
+
         return new_entry
 
     def switch_session(self, session_key: str, target_session_id: str) -> Optional[SessionEntry]:
@@ -672,52 +850,58 @@ class SessionStore:
         generating a fresh session ID, re-uses ``target_session_id`` so the
         old transcript is loaded on the next message.
         """
-        self._ensure_loaded()
+        db_end_session_id = None
+        new_entry = None
 
-        if session_key not in self._entries:
-            return None
+        with self._lock:
+            self._ensure_loaded_locked()
 
-        old_entry = self._entries[session_key]
+            if session_key not in self._entries:
+                return None
 
-        # Don't switch if already on that session
-        if old_entry.session_id == target_session_id:
-            return old_entry
+            old_entry = self._entries[session_key]
 
-        # End the current session in SQLite
-        if self._db:
+            # Don't switch if already on that session
+            if old_entry.session_id == target_session_id:
+                return old_entry
+
+            db_end_session_id = old_entry.session_id
+
+            now = _now()
+            new_entry = SessionEntry(
+                session_key=session_key,
+                session_id=target_session_id,
+                created_at=now,
+                updated_at=now,
+                origin=old_entry.origin,
+                display_name=old_entry.display_name,
+                platform=old_entry.platform,
+                chat_type=old_entry.chat_type,
+            )
+
+            self._entries[session_key] = new_entry
+            self._save()
+
+        if self._db and db_end_session_id:
             try:
-                self._db.end_session(old_entry.session_id, "session_switch")
+                self._db.end_session(db_end_session_id, "session_switch")
             except Exception as e:
                 logger.debug("Session DB end_session failed: %s", e)
 
-        now = datetime.now()
-        new_entry = SessionEntry(
-            session_key=session_key,
-            session_id=target_session_id,
-            created_at=now,
-            updated_at=now,
-            origin=old_entry.origin,
-            display_name=old_entry.display_name,
-            platform=old_entry.platform,
-            chat_type=old_entry.chat_type,
-        )
-
-        self._entries[session_key] = new_entry
-        self._save()
         return new_entry
 
     def list_sessions(self, active_minutes: Optional[int] = None) -> List[SessionEntry]:
         """List all sessions, optionally filtered by activity."""
-        self._ensure_loaded()
-        
-        entries = list(self._entries.values())
-        
+        with self._lock:
+            self._ensure_loaded_locked()
+            entries = list(self._entries.values())
+
         if active_minutes is not None:
-            cutoff = datetime.now() - timedelta(minutes=active_minutes)
+            cutoff = _now() - timedelta(minutes=active_minutes)
             entries = [e for e in entries if e.updated_at >= cutoff]
-        
+
         entries.sort(key=lambda e: e.updated_at, reverse=True)
-        
+
         return entries
     
     def get_transcript_path(self, session_id: str) -> Path:
@@ -763,13 +947,17 @@ class SessionStore:
             try:
                 self._db.clear_messages(session_id)
                 for msg in messages:
+                    role = msg.get("role", "unknown")
                     self._db.append_message(
                         session_id=session_id,
-                        role=msg.get("role", "unknown"),
+                        role=role,
                         content=msg.get("content"),
                         tool_name=msg.get("tool_name"),
                         tool_calls=msg.get("tool_calls"),
                         tool_call_id=msg.get("tool_call_id"),
+                        reasoning=msg.get("reasoning") if role == "assistant" else None,
+                        reasoning_details=msg.get("reasoning_details") if role == "assistant" else None,
+                        codex_reasoning_items=msg.get("codex_reasoning_items") if role == "assistant" else None,
                     )
             except Exception as e:
                 logger.debug("Failed to rewrite transcript in DB: %s", e)
@@ -782,29 +970,51 @@ class SessionStore:
 
     def load_transcript(self, session_id: str) -> List[Dict[str, Any]]:
         """Load all messages from a session's transcript."""
+        db_messages = []
         # Try SQLite first
         if self._db:
             try:
-                messages = self._db.get_messages_as_conversation(session_id)
-                if messages:
-                    return messages
+                db_messages = self._db.get_messages_as_conversation(session_id)
             except Exception as e:
                 logger.debug("Could not load messages from DB: %s", e)
-        
-        # Fall back to legacy JSONL
+
+        # Load legacy JSONL transcript (may contain more history than SQLite
+        # for sessions created before the DB layer was introduced).
         transcript_path = self.get_transcript_path(session_id)
-        
-        if not transcript_path.exists():
-            return []
-        
-        messages = []
-        with open(transcript_path, "r", encoding="utf-8") as f:
-            for line in f:
-                line = line.strip()
-                if line:
-                    messages.append(json.loads(line))
-        
-        return messages
+        jsonl_messages = []
+        if transcript_path.exists():
+            with open(transcript_path, "r", encoding="utf-8") as f:
+                for line in f:
+                    line = line.strip()
+                    if line:
+                        try:
+                            jsonl_messages.append(json.loads(line))
+                        except json.JSONDecodeError:
+                            logger.warning(
+                                "Skipping corrupt line in transcript %s: %s",
+                                session_id, line[:120],
+                            )
+
+        # Prefer whichever source has more messages.
+        #
+        # Background: when a session pre-dates SQLite storage (or when the DB
+        # layer was added while a long-lived session was already active), the
+        # first post-migration turn writes only the *new* messages to SQLite
+        # (because _flush_messages_to_session_db skips messages already in
+        # conversation_history, assuming they're persisted).  On the *next*
+        # turn load_transcript returns those few SQLite rows and ignores the
+        # full JSONL history — the model sees a context of 1-4 messages instead
+        # of hundreds.  Using the longer source prevents this silent truncation.
+        if len(jsonl_messages) > len(db_messages):
+            if db_messages:
+                logger.debug(
+                    "Session %s: JSONL has %d messages vs SQLite %d — "
+                    "using JSONL (legacy session not yet fully migrated)",
+                    session_id, len(jsonl_messages), len(db_messages),
+                )
+            return jsonl_messages
+
+        return db_messages
 
 
 def build_session_context(

gateway/status.py

@@ -17,6 +17,7 @@ import os
 import sys
 from datetime import datetime, timezone
 from pathlib import Path
+from hermes_constants import get_hermes_home
 from typing import Any, Optional
 
 _GATEWAY_KIND = "hermes-gateway"
@@ -26,7 +27,7 @@ _LOCKS_DIRNAME = "gateway-locks"
 
 def _get_pid_path() -> Path:
     """Return the path to the gateway PID file, respecting HERMES_HOME."""
-    home = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
+    home = get_hermes_home()
     return home / "gateway.pid"
 
 
@@ -83,11 +84,30 @@ def _looks_like_gateway_process(pid: int) -> bool:
     """Return True when the live PID still looks like the Hermes gateway."""
     cmdline = _read_process_cmdline(pid)
     if not cmdline:
-        # If we cannot inspect the process, fall back to the liveness check.
-        return True
+        return False
 
     patterns = (
         "hermes_cli.main gateway",
+        "hermes_cli/main.py gateway",
+        "hermes gateway",
+        "gateway/run.py",
+    )
+    return any(pattern in cmdline for pattern in patterns)
+
+
+def _record_looks_like_gateway(record: dict[str, Any]) -> bool:
+    """Validate gateway identity from PID-file metadata when cmdline is unavailable."""
+    if record.get("kind") != _GATEWAY_KIND:
+        return False
+
+    argv = record.get("argv")
+    if not isinstance(argv, list) or not argv:
+        return False
+
+    cmdline = " ".join(str(part) for part in argv)
+    patterns = (
+        "hermes_cli.main gateway",
+        "hermes_cli/main.py gateway",
         "hermes gateway",
         "gateway/run.py",
     )
@@ -178,8 +198,8 @@ def write_runtime_status(
     payload = _read_json_file(path) or _build_runtime_status_record()
     payload.setdefault("platforms", {})
     payload.setdefault("kind", _GATEWAY_KIND)
-    payload.setdefault("pid", os.getpid())
-    payload.setdefault("start_time", _get_process_start_time(os.getpid()))
+    payload["pid"] = os.getpid()
+    payload["start_time"] = _get_process_start_time(os.getpid())
     payload["updated_at"] = _utc_now_iso()
 
     if gateway_state is not None:
@@ -255,6 +275,21 @@ def acquire_scoped_lock(scope: str, identity: str, metadata: Optional[dict[str,
                     and current_start != existing.get("start_time")
                 ):
                     stale = True
+                # Check if process is stopped (Ctrl+Z / SIGTSTP) — stopped
+                # processes still respond to os.kill(pid, 0) but are not
+                # actually running. Treat them as stale so --replace works.
+                if not stale:
+                    try:
+                        _proc_status = Path(f"/proc/{existing_pid}/status")
+                        if _proc_status.exists():
+                            for _line in _proc_status.read_text().splitlines():
+                                if _line.startswith("State:"):
+                                    _state = _line.split()[1]
+                                    if _state in ("T", "t"):  # stopped or tracing stop
+                                        stale = True
+                                    break
+                    except (OSError, PermissionError):
+                        pass
         if stale:
             try:
                 lock_path.unlink(missing_ok=True)
@@ -295,6 +330,25 @@ def release_scoped_lock(scope: str, identity: str) -> None:
         pass
 
 
+def release_all_scoped_locks() -> int:
+    """Remove all scoped lock files in the lock directory.
+
+    Called during --replace to clean up stale locks left by stopped/killed
+    gateway processes that did not release their locks gracefully.
+    Returns the number of lock files removed.
+    """
+    lock_dir = _get_lock_dir()
+    removed = 0
+    if lock_dir.exists():
+        for lock_file in lock_dir.glob("*.lock"):
+            try:
+                lock_file.unlink(missing_ok=True)
+                removed += 1
+            except OSError:
+                pass
+    return removed
+
+
 def get_running_pid() -> Optional[int]:
     """Return the PID of a running gateway instance, or ``None``.
 
@@ -325,8 +379,9 @@ def get_running_pid() -> Optional[int]:
         return None
 
     if not _looks_like_gateway_process(pid):
-        remove_pid_file()
-        return None
+        if not _record_looks_like_gateway(record):
+            remove_pid_file()
+            return None
 
     return pid
 

gateway/sticker_cache.py

@@ -9,9 +9,7 @@ Cache location: ~/.hermes/sticker_cache.json
 """
 
 import json
-import os
 import time
-from pathlib import Path
 from typing import Optional
 
 from hermes_cli.config import get_hermes_home

gateway/stream_consumer.py

@@ -0,0 +1,234 @@
+"""Gateway streaming consumer — bridges sync agent callbacks to async platform delivery.
+
+The agent fires stream_delta_callback(text) synchronously from its worker thread.
+GatewayStreamConsumer:
+  1. Receives deltas via on_delta() (thread-safe, sync)
+  2. Queues them to an asyncio task via queue.Queue
+  3. The async run() task buffers, rate-limits, and progressively edits
+     a single message on the target platform
+
+Design: Uses the edit transport (send initial message, then editMessageText).
+This is universally supported across Telegram, Discord, and Slack.
+
+Credit: jobless0x (#774, #1312), OutThisLife (#798), clicksingh (#697).
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import queue
+import re
+import time
+from dataclasses import dataclass
+from typing import Any, Optional
+
+logger = logging.getLogger("gateway.stream_consumer")
+
+# Sentinel to signal the stream is complete
+_DONE = object()
+
+
+@dataclass
+class StreamConsumerConfig:
+    """Runtime config for a single stream consumer instance."""
+    edit_interval: float = 0.3
+    buffer_threshold: int = 40
+    cursor: str = " ▉"
+
+
+class GatewayStreamConsumer:
+    """Async consumer that progressively edits a platform message with streamed tokens.
+
+    Usage::
+
+        consumer = GatewayStreamConsumer(adapter, chat_id, config, metadata=metadata)
+        # Pass consumer.on_delta as stream_delta_callback to AIAgent
+        agent = AIAgent(..., stream_delta_callback=consumer.on_delta)
+        # Start the consumer as an asyncio task
+        task = asyncio.create_task(consumer.run())
+        # ... run agent in thread pool ...
+        consumer.finish()  # signal completion
+        await task         # wait for final edit
+    """
+
+    def __init__(
+        self,
+        adapter: Any,
+        chat_id: str,
+        config: Optional[StreamConsumerConfig] = None,
+        metadata: Optional[dict] = None,
+    ):
+        self.adapter = adapter
+        self.chat_id = chat_id
+        self.cfg = config or StreamConsumerConfig()
+        self.metadata = metadata
+        self._queue: queue.Queue = queue.Queue()
+        self._accumulated = ""
+        self._message_id: Optional[str] = None
+        self._already_sent = False
+        self._edit_supported = True  # Disabled on first edit failure (Signal/Email/HA)
+        self._last_edit_time = 0.0
+        self._last_sent_text = ""   # Track last-sent text to skip redundant edits
+
+    @property
+    def already_sent(self) -> bool:
+        """True if at least one message was sent/edited — signals the base
+        adapter to skip re-sending the final response."""
+        return self._already_sent
+
+    def on_delta(self, text: str) -> None:
+        """Thread-safe callback — called from the agent's worker thread."""
+        if text:
+            self._queue.put(text)
+
+    def finish(self) -> None:
+        """Signal that the stream is complete."""
+        self._queue.put(_DONE)
+
+    async def run(self) -> None:
+        """Async task that drains the queue and edits the platform message."""
+        # Platform message length limit — leave room for cursor + formatting
+        _raw_limit = getattr(self.adapter, "MAX_MESSAGE_LENGTH", 4096)
+        _safe_limit = max(500, _raw_limit - len(self.cfg.cursor) - 100)
+
+        try:
+            while True:
+                # Drain all available items from the queue
+                got_done = False
+                while True:
+                    try:
+                        item = self._queue.get_nowait()
+                        if item is _DONE:
+                            got_done = True
+                            break
+                        self._accumulated += item
+                    except queue.Empty:
+                        break
+
+                # Decide whether to flush an edit
+                now = time.monotonic()
+                elapsed = now - self._last_edit_time
+                should_edit = (
+                    got_done
+                    or (elapsed >= self.cfg.edit_interval
+                        and len(self._accumulated) > 0)
+                    or len(self._accumulated) >= self.cfg.buffer_threshold
+                )
+
+                if should_edit and self._accumulated:
+                    # Split overflow: if accumulated text exceeds the platform
+                    # limit, finalize the current message and start a new one.
+                    while (
+                        len(self._accumulated) > _safe_limit
+                        and self._message_id is not None
+                    ):
+                        split_at = self._accumulated.rfind("\n", 0, _safe_limit)
+                        if split_at < _safe_limit // 2:
+                            split_at = _safe_limit
+                        chunk = self._accumulated[:split_at]
+                        await self._send_or_edit(chunk)
+                        self._accumulated = self._accumulated[split_at:].lstrip("\n")
+                        self._message_id = None
+                        self._last_sent_text = ""
+
+                    display_text = self._accumulated
+                    if not got_done:
+                        display_text += self.cfg.cursor
+
+                    await self._send_or_edit(display_text)
+                    self._last_edit_time = time.monotonic()
+
+                if got_done:
+                    # Final edit without cursor
+                    if self._accumulated and self._message_id:
+                        await self._send_or_edit(self._accumulated)
+                    return
+
+                await asyncio.sleep(0.05)  # Small yield to not busy-loop
+
+        except asyncio.CancelledError:
+            # Best-effort final edit on cancellation
+            if self._accumulated and self._message_id:
+                try:
+                    await self._send_or_edit(self._accumulated)
+                except Exception:
+                    pass
+        except Exception as e:
+            logger.error("Stream consumer error: %s", e)
+
+    # Pattern to strip MEDIA:<path> tags (including optional surrounding quotes).
+    # Matches the simple cleanup regex used by the non-streaming path in
+    # gateway/platforms/base.py for post-processing.
+    _MEDIA_RE = re.compile(r'''[`"']?MEDIA:\s*\S+[`"']?''')
+
+    @staticmethod
+    def _clean_for_display(text: str) -> str:
+        """Strip MEDIA: directives and internal markers from text before display.
+
+        The streaming path delivers raw text chunks that may include
+        ``MEDIA:<path>`` tags and ``[[audio_as_voice]]`` directives meant for
+        the platform adapter's post-processing.  The actual media files are
+        delivered separately via ``_deliver_media_from_response()`` after the
+        stream finishes — we just need to hide the raw directives from the
+        user.
+        """
+        if "MEDIA:" not in text and "[[audio_as_voice]]" not in text:
+            return text
+        cleaned = text.replace("[[audio_as_voice]]", "")
+        cleaned = GatewayStreamConsumer._MEDIA_RE.sub("", cleaned)
+        # Collapse excessive blank lines left behind by removed tags
+        cleaned = re.sub(r'\n{3,}', '\n\n', cleaned)
+        # Strip trailing whitespace/newlines but preserve leading content
+        return cleaned.rstrip()
+
+    async def _send_or_edit(self, text: str) -> None:
+        """Send or edit the streaming message."""
+        # Strip MEDIA: directives so they don't appear as visible text.
+        # Media files are delivered as native attachments after the stream
+        # finishes (via _deliver_media_from_response in gateway/run.py).
+        text = self._clean_for_display(text)
+        if not text.strip():
+            return
+        try:
+            if self._message_id is not None:
+                if self._edit_supported:
+                    # Skip if text is identical to what we last sent
+                    if text == self._last_sent_text:
+                        return
+                    # Edit existing message
+                    result = await self.adapter.edit_message(
+                        chat_id=self.chat_id,
+                        message_id=self._message_id,
+                        content=text,
+                    )
+                    if result.success:
+                        self._already_sent = True
+                        self._last_sent_text = text
+                    else:
+                        # If an edit fails mid-stream (especially Telegram flood control),
+                        # stop progressive edits and let the normal final send path deliver
+                        # the complete answer instead of leaving the user with a partial.
+                        logger.debug("Edit failed, disabling streaming for this adapter")
+                        self._edit_supported = False
+                        self._already_sent = False
+                else:
+                    # Editing not supported — skip intermediate updates.
+                    # The final response will be sent by the normal path.
+                    pass
+            else:
+                # First message — send new
+                result = await self.adapter.send(
+                    chat_id=self.chat_id,
+                    content=text,
+                    metadata=self.metadata,
+                )
+                if result.success and result.message_id:
+                    self._message_id = result.message_id
+                    self._already_sent = True
+                    self._last_sent_text = text
+                else:
+                    # Initial send failed — disable streaming for this session
+                    self._edit_supported = False
+        except Exception as e:
+            logger.error("Stream send/edit error: %s", e)

hermes

@@ -1,12 +1,11 @@
 #!/usr/bin/env python3
 """
-Hermes Agent CLI Launcher
+Hermes Agent CLI launcher.
 
-This is a convenience wrapper to launch the Hermes CLI.
-Usage: ./hermes [options]
+This wrapper should behave like the installed `hermes` command, including
+subcommands such as `gateway`, `cron`, and `doctor`.
 """
 
 if __name__ == "__main__":
-    from cli import main
-    import fire
-    fire.Fire(main)
+    from hermes_cli.main import main
+    main()

hermes_cli/__init__.py

@@ -11,5 +11,5 @@ Provides subcommands for:
 - hermes cron          - Manage cron jobs
 """
 
-__version__ = "0.2.0"
-__release_date__ = "2026.3.12"
+__version__ = "0.7.0"
+__release_date__ = "2026.4.3"

hermes_cli/auth_commands.py

@@ -0,0 +1,493 @@
+"""Credential-pool auth subcommands."""
+
+from __future__ import annotations
+
+from getpass import getpass
+import math
+import time
+from types import SimpleNamespace
+import uuid
+
+from agent.credential_pool import (
+    AUTH_TYPE_API_KEY,
+    AUTH_TYPE_OAUTH,
+    CUSTOM_POOL_PREFIX,
+    SOURCE_MANUAL,
+    STATUS_EXHAUSTED,
+    STRATEGY_FILL_FIRST,
+    STRATEGY_ROUND_ROBIN,
+    STRATEGY_RANDOM,
+    STRATEGY_LEAST_USED,
+    SUPPORTED_POOL_STRATEGIES,
+    PooledCredential,
+    _exhausted_until,
+    _normalize_custom_pool_name,
+    get_pool_strategy,
+    label_from_token,
+    list_custom_pool_providers,
+    load_pool,
+)
+import hermes_cli.auth as auth_mod
+from hermes_cli.auth import PROVIDER_REGISTRY
+from hermes_constants import OPENROUTER_BASE_URL
+
+
+# Providers that support OAuth login in addition to API keys.
+_OAUTH_CAPABLE_PROVIDERS = {"anthropic", "nous", "openai-codex"}
+
+
+def _get_custom_provider_names() -> list:
+    """Return list of (display_name, pool_key) tuples for custom_providers in config."""
+    try:
+        from hermes_cli.config import load_config
+
+        config = load_config()
+    except Exception:
+        return []
+    custom_providers = config.get("custom_providers")
+    if not isinstance(custom_providers, list):
+        return []
+    result = []
+    for entry in custom_providers:
+        if not isinstance(entry, dict):
+            continue
+        name = entry.get("name")
+        if not isinstance(name, str) or not name.strip():
+            continue
+        pool_key = f"{CUSTOM_POOL_PREFIX}{_normalize_custom_pool_name(name)}"
+        result.append((name.strip(), pool_key))
+    return result
+
+
+def _resolve_custom_provider_input(raw: str) -> str | None:
+    """If raw input matches a custom_providers entry name (case-insensitive), return its pool key."""
+    normalized = (raw or "").strip().lower().replace(" ", "-")
+    if not normalized:
+        return None
+    # Direct match on 'custom:name' format
+    if normalized.startswith(CUSTOM_POOL_PREFIX):
+        return normalized
+    for display_name, pool_key in _get_custom_provider_names():
+        if _normalize_custom_pool_name(display_name) == normalized:
+            return pool_key
+    return None
+
+
+def _normalize_provider(provider: str) -> str:
+    normalized = (provider or "").strip().lower()
+    if normalized in {"or", "open-router"}:
+        return "openrouter"
+    # Check if it matches a custom provider name
+    custom_key = _resolve_custom_provider_input(normalized)
+    if custom_key:
+        return custom_key
+    return normalized
+
+
+def _provider_base_url(provider: str) -> str:
+    if provider == "openrouter":
+        return OPENROUTER_BASE_URL
+    if provider.startswith(CUSTOM_POOL_PREFIX):
+        from agent.credential_pool import _get_custom_provider_config
+
+        cp_config = _get_custom_provider_config(provider)
+        if cp_config:
+            return str(cp_config.get("base_url") or "").strip()
+        return ""
+    pconfig = PROVIDER_REGISTRY.get(provider)
+    return pconfig.inference_base_url if pconfig else ""
+
+
+def _oauth_default_label(provider: str, count: int) -> str:
+    return f"{provider}-oauth-{count}"
+
+
+def _api_key_default_label(count: int) -> str:
+    return f"api-key-{count}"
+
+
+def _display_source(source: str) -> str:
+    return source.split(":", 1)[1] if source.startswith("manual:") else source
+
+
+def _format_exhausted_status(entry) -> str:
+    if entry.last_status != STATUS_EXHAUSTED:
+        return ""
+    reason = getattr(entry, "last_error_reason", None)
+    reason_text = f" {reason}" if isinstance(reason, str) and reason.strip() else ""
+    code = f" ({entry.last_error_code})" if entry.last_error_code else ""
+    exhausted_until = _exhausted_until(entry)
+    if exhausted_until is None:
+        return f" exhausted{reason_text}{code}"
+    remaining = max(0, int(math.ceil(exhausted_until - time.time())))
+    if remaining <= 0:
+        return f" exhausted{reason_text}{code} (ready to retry)"
+    minutes, seconds = divmod(remaining, 60)
+    hours, minutes = divmod(minutes, 60)
+    days, hours = divmod(hours, 24)
+    if days:
+        wait = f"{days}d {hours}h"
+    elif hours:
+        wait = f"{hours}h {minutes}m"
+    elif minutes:
+        wait = f"{minutes}m {seconds}s"
+    else:
+        wait = f"{seconds}s"
+    return f" exhausted{reason_text}{code} ({wait} left)"
+
+
+def auth_add_command(args) -> None:
+    provider = _normalize_provider(getattr(args, "provider", ""))
+    if provider not in PROVIDER_REGISTRY and provider != "openrouter" and not provider.startswith(CUSTOM_POOL_PREFIX):
+        raise SystemExit(f"Unknown provider: {provider}")
+
+    requested_type = str(getattr(args, "auth_type", "") or "").strip().lower()
+    if requested_type in {AUTH_TYPE_API_KEY, "api-key"}:
+        requested_type = AUTH_TYPE_API_KEY
+    if not requested_type:
+        if provider.startswith(CUSTOM_POOL_PREFIX):
+            requested_type = AUTH_TYPE_API_KEY
+        else:
+            requested_type = AUTH_TYPE_OAUTH if provider in {"anthropic", "nous", "openai-codex"} else AUTH_TYPE_API_KEY
+
+    pool = load_pool(provider)
+
+    if requested_type == AUTH_TYPE_API_KEY:
+        token = (getattr(args, "api_key", None) or "").strip()
+        if not token:
+            token = getpass("Paste your API key: ").strip()
+        if not token:
+            raise SystemExit("No API key provided.")
+        default_label = _api_key_default_label(len(pool.entries()) + 1)
+        label = (getattr(args, "label", None) or "").strip()
+        if not label:
+            label = input(f"Label (optional, default: {default_label}): ").strip() or default_label
+        entry = PooledCredential(
+            provider=provider,
+            id=uuid.uuid4().hex[:6],
+            label=label,
+            auth_type=AUTH_TYPE_API_KEY,
+            priority=0,
+            source=SOURCE_MANUAL,
+            access_token=token,
+            base_url=_provider_base_url(provider),
+        )
+        pool.add_entry(entry)
+        print(f'Added {provider} credential #{len(pool.entries())}: "{label}"')
+        return
+
+    if provider == "anthropic":
+        from agent import anthropic_adapter as anthropic_mod
+
+        creds = anthropic_mod.run_hermes_oauth_login_pure()
+        if not creds:
+            raise SystemExit("Anthropic OAuth login did not return credentials.")
+        label = (getattr(args, "label", None) or "").strip() or label_from_token(
+            creds["access_token"],
+            _oauth_default_label(provider, len(pool.entries()) + 1),
+        )
+        entry = PooledCredential(
+            provider=provider,
+            id=uuid.uuid4().hex[:6],
+            label=label,
+            auth_type=AUTH_TYPE_OAUTH,
+            priority=0,
+            source=f"{SOURCE_MANUAL}:hermes_pkce",
+            access_token=creds["access_token"],
+            refresh_token=creds.get("refresh_token"),
+            expires_at_ms=creds.get("expires_at_ms"),
+            base_url=_provider_base_url(provider),
+        )
+        pool.add_entry(entry)
+        print(f'Added {provider} OAuth credential #{len(pool.entries())}: "{entry.label}"')
+        return
+
+    if provider == "nous":
+        creds = auth_mod._nous_device_code_login(
+            portal_base_url=getattr(args, "portal_url", None),
+            inference_base_url=getattr(args, "inference_url", None),
+            client_id=getattr(args, "client_id", None),
+            scope=getattr(args, "scope", None),
+            open_browser=not getattr(args, "no_browser", False),
+            timeout_seconds=getattr(args, "timeout", None) or 15.0,
+            insecure=bool(getattr(args, "insecure", False)),
+            ca_bundle=getattr(args, "ca_bundle", None),
+            min_key_ttl_seconds=max(60, int(getattr(args, "min_key_ttl_seconds", 5 * 60))),
+        )
+        label = (getattr(args, "label", None) or "").strip() or label_from_token(
+            creds.get("access_token", ""),
+            _oauth_default_label(provider, len(pool.entries()) + 1),
+        )
+        entry = PooledCredential.from_dict(provider, {
+            **creds,
+            "label": label,
+            "auth_type": AUTH_TYPE_OAUTH,
+            "source": f"{SOURCE_MANUAL}:device_code",
+            "base_url": creds.get("inference_base_url"),
+        })
+        pool.add_entry(entry)
+        print(f'Added {provider} OAuth credential #{len(pool.entries())}: "{entry.label}"')
+        return
+
+    if provider == "openai-codex":
+        creds = auth_mod._codex_device_code_login()
+        label = (getattr(args, "label", None) or "").strip() or label_from_token(
+            creds["tokens"]["access_token"],
+            _oauth_default_label(provider, len(pool.entries()) + 1),
+        )
+        entry = PooledCredential(
+            provider=provider,
+            id=uuid.uuid4().hex[:6],
+            label=label,
+            auth_type=AUTH_TYPE_OAUTH,
+            priority=0,
+            source=f"{SOURCE_MANUAL}:device_code",
+            access_token=creds["tokens"]["access_token"],
+            refresh_token=creds["tokens"].get("refresh_token"),
+            base_url=creds.get("base_url"),
+            last_refresh=creds.get("last_refresh"),
+        )
+        pool.add_entry(entry)
+        print(f'Added {provider} OAuth credential #{len(pool.entries())}: "{entry.label}"')
+        return
+
+    raise SystemExit(f"`hermes auth add {provider}` is not implemented for auth type {requested_type} yet.")
+
+
+def auth_list_command(args) -> None:
+    provider_filter = _normalize_provider(getattr(args, "provider", "") or "")
+    if provider_filter:
+        providers = [provider_filter]
+    else:
+        providers = sorted({
+            *PROVIDER_REGISTRY.keys(),
+            "openrouter",
+            *list_custom_pool_providers(),
+        })
+    for provider in providers:
+        pool = load_pool(provider)
+        entries = pool.entries()
+        if not entries:
+            continue
+        current = pool.peek()
+        print(f"{provider} ({len(entries)} credentials):")
+        for idx, entry in enumerate(entries, start=1):
+            marker = "  "
+            if current is not None and entry.id == current.id:
+                marker = "← "
+            status = _format_exhausted_status(entry)
+            source = _display_source(entry.source)
+            print(f"  #{idx}  {entry.label:<20} {entry.auth_type:<7} {source}{status} {marker}".rstrip())
+        print()
+
+
+def auth_remove_command(args) -> None:
+    provider = _normalize_provider(getattr(args, "provider", ""))
+    target = getattr(args, "target", None)
+    if target is None:
+        target = getattr(args, "index", None)
+    pool = load_pool(provider)
+    index, matched, error = pool.resolve_target(target)
+    if matched is None or index is None:
+        raise SystemExit(f"{error} Provider: {provider}.")
+    removed = pool.remove_index(index)
+    if removed is None:
+        raise SystemExit(f'No credential matching "{target}" for provider {provider}.')
+    print(f"Removed {provider} credential #{index} ({removed.label})")
+
+    # If this was an env-seeded credential, also clear the env var from .env
+    # so it doesn't get re-seeded on the next load_pool() call.
+    if removed.source.startswith("env:"):
+        env_var = removed.source[len("env:"):]
+        if env_var:
+            from hermes_cli.config import remove_env_value
+            cleared = remove_env_value(env_var)
+            if cleared:
+                print(f"Cleared {env_var} from .env")
+
+
+def auth_reset_command(args) -> None:
+    provider = _normalize_provider(getattr(args, "provider", ""))
+    pool = load_pool(provider)
+    count = pool.reset_statuses()
+    print(f"Reset status on {count} {provider} credentials")
+
+
+def _interactive_auth() -> None:
+    """Interactive credential pool management when `hermes auth` is called bare."""
+    # Show current pool status first
+    print("Credential Pool Status")
+    print("=" * 50)
+
+    auth_list_command(SimpleNamespace(provider=None))
+    print()
+
+    # Main menu
+    choices = [
+        "Add a credential",
+        "Remove a credential",
+        "Reset cooldowns for a provider",
+        "Set rotation strategy for a provider",
+        "Exit",
+    ]
+    print("What would you like to do?")
+    for i, choice in enumerate(choices, 1):
+        print(f"  {i}. {choice}")
+
+    try:
+        raw = input("\nChoice: ").strip()
+    except (EOFError, KeyboardInterrupt):
+        return
+
+    if not raw or raw == str(len(choices)):
+        return
+
+    if raw == "1":
+        _interactive_add()
+    elif raw == "2":
+        _interactive_remove()
+    elif raw == "3":
+        _interactive_reset()
+    elif raw == "4":
+        _interactive_strategy()
+
+
+def _pick_provider(prompt: str = "Provider") -> str:
+    """Prompt for a provider name with auto-complete hints."""
+    known = sorted(set(list(PROVIDER_REGISTRY.keys()) + ["openrouter"]))
+    custom_names = _get_custom_provider_names()
+    if custom_names:
+        custom_display = [name for name, _key in custom_names]
+        print(f"\nKnown providers: {', '.join(known)}")
+        print(f"Custom endpoints: {', '.join(custom_display)}")
+    else:
+        print(f"\nKnown providers: {', '.join(known)}")
+    try:
+        raw = input(f"{prompt}: ").strip()
+    except (EOFError, KeyboardInterrupt):
+        raise SystemExit()
+    return _normalize_provider(raw)
+
+
+def _interactive_add() -> None:
+    provider = _pick_provider("Provider to add credential for")
+    if provider not in PROVIDER_REGISTRY and provider != "openrouter" and not provider.startswith(CUSTOM_POOL_PREFIX):
+        raise SystemExit(f"Unknown provider: {provider}")
+
+    # For OAuth-capable providers, ask which type
+    if provider in _OAUTH_CAPABLE_PROVIDERS:
+        print(f"\n{provider} supports both API keys and OAuth login.")
+        print("  1. API key (paste a key from the provider dashboard)")
+        print("  2. OAuth login (authenticate via browser)")
+        try:
+            type_choice = input("Type [1/2]: ").strip()
+        except (EOFError, KeyboardInterrupt):
+            return
+        if type_choice == "2":
+            auth_type = "oauth"
+        else:
+            auth_type = "api_key"
+    else:
+        auth_type = "api_key"
+
+    label = None
+    try:
+        typed_label = input("Label / account name (optional): ").strip()
+    except (EOFError, KeyboardInterrupt):
+        return
+    if typed_label:
+        label = typed_label
+
+    auth_add_command(SimpleNamespace(
+        provider=provider, auth_type=auth_type, label=label, api_key=None,
+        portal_url=None, inference_url=None, client_id=None, scope=None,
+        no_browser=False, timeout=None, insecure=False, ca_bundle=None,
+    ))
+
+
+def _interactive_remove() -> None:
+    provider = _pick_provider("Provider to remove credential from")
+    pool = load_pool(provider)
+    if not pool.has_credentials():
+        print(f"No credentials for {provider}.")
+        return
+
+    # Show entries with indices
+    for i, e in enumerate(pool.entries(), 1):
+        exhausted = _format_exhausted_status(e)
+        print(f"  #{i}  {e.label:25s} {e.auth_type:10s} {e.source}{exhausted} [id:{e.id}]")
+
+    try:
+        raw = input("Remove #, id, or label (blank to cancel): ").strip()
+    except (EOFError, KeyboardInterrupt):
+        return
+    if not raw:
+        return
+
+    auth_remove_command(SimpleNamespace(provider=provider, target=raw))
+
+
+def _interactive_reset() -> None:
+    provider = _pick_provider("Provider to reset cooldowns for")
+
+    auth_reset_command(SimpleNamespace(provider=provider))
+
+
+def _interactive_strategy() -> None:
+    provider = _pick_provider("Provider to set strategy for")
+    current = get_pool_strategy(provider)
+    strategies = [STRATEGY_FILL_FIRST, STRATEGY_ROUND_ROBIN, STRATEGY_LEAST_USED, STRATEGY_RANDOM]
+
+    print(f"\nCurrent strategy for {provider}: {current}")
+    print()
+    descriptions = {
+        STRATEGY_FILL_FIRST: "Use first key until exhausted, then next",
+        STRATEGY_ROUND_ROBIN: "Cycle through keys evenly",
+        STRATEGY_LEAST_USED: "Always pick the least-used key",
+        STRATEGY_RANDOM: "Random selection",
+    }
+    for i, s in enumerate(strategies, 1):
+        marker = " ←" if s == current else ""
+        print(f"  {i}. {s:15s} — {descriptions.get(s, '')}{marker}")
+
+    try:
+        raw = input("\nStrategy [1-4]: ").strip()
+    except (EOFError, KeyboardInterrupt):
+        return
+    if not raw:
+        return
+
+    try:
+        idx = int(raw) - 1
+        strategy = strategies[idx]
+    except (ValueError, IndexError):
+        print("Invalid choice.")
+        return
+
+    from hermes_cli.config import load_config, save_config
+    cfg = load_config()
+    pool_strategies = cfg.get("credential_pool_strategies") or {}
+    if not isinstance(pool_strategies, dict):
+        pool_strategies = {}
+    pool_strategies[provider] = strategy
+    cfg["credential_pool_strategies"] = pool_strategies
+    save_config(cfg)
+    print(f"Set {provider} strategy to: {strategy}")
+
+
+def auth_command(args) -> None:
+    action = getattr(args, "auth_action", "")
+    if action == "add":
+        auth_add_command(args)
+        return
+    if action == "list":
+        auth_list_command(args)
+        return
+    if action == "remove":
+        auth_remove_command(args)
+        return
+    if action == "reset":
+        auth_reset_command(args)
+        return
+    # No subcommand — launch interactive mode
+    _interactive_auth()