fix: inject plugin context after cache markers to preserve Anthropic prompt cache prefix stability

Move mention stripping outside the `if not is_dm` guard so mentions
are stripped in DMs too. Remove the bare-mention early return so a
message containing only a mention passes through as empty string,
matching Discord's behavior.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.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
@@ -38,11 +39,11 @@ 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
 
 # =============================================================================
@@ -50,7 +51,7 @@ MINIMAX_CN_API_KEY=
 # =============================================================================
 # 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_API_KEY=
 # OPENCODE_ZEN_BASE_URL=https://opencode.ai/zen/v1  # Override default base URL
 
 # =============================================================================
@@ -58,34 +59,47 @@ OPENCODE_ZEN_API_KEY=
 # =============================================================================
 # 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=
+# 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=
+# 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).
@@ -168,10 +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
@@ -203,7 +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
@@ -218,6 +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
@@ -274,11 +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

@@ -27,8 +27,11 @@ jobs:
         with:
           python-version: '3.11'
 
-      - name: Install ascii-guard
-        run: python -m pip install ascii-guard
+      - 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

.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

@@ -38,6 +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)
@@ -172,6 +173,7 @@ if canonical == "mycommand":
 - `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.
 
@@ -208,6 +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.
 
 ---
@@ -356,8 +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.
 
@@ -373,6 +440,19 @@ Tool schema descriptions must not mention tools from other toolsets by name (e.g
 ### 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

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

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

@@ -144,16 +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 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_v0.4.0.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:
@@ -44,7 +45,7 @@ def _load_env() -> None:
     """Load .env from HERMES_HOME (default ``~/.hermes``)."""
     from hermes_cli.env_loader import load_hermes_dotenv
 
-    hermes_home = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
+    hermes_home = get_hermes_home()
     loaded = load_hermes_dotenv(hermes_home=hermes_home)
     if loaded:
         for env_file in loaded:
@@ -73,7 +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

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

@@ -22,9 +22,15 @@ from acp.schema import (
     InitializeResponse,
     ListSessionsResponse,
     LoadSessionResponse,
+    McpServerHttp,
+    McpServerSse,
+    McpServerStdio,
     NewSessionResponse,
     PromptResponse,
     ResumeSessionResponse,
+    SetSessionConfigOptionResponse,
+    SetSessionModelResponse,
+    SetSessionModeResponse,
     ResourceContentBlock,
     SessionCapabilities,
     SessionForkCapabilities,
@@ -90,15 +96,83 @@ 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:
@@ -111,7 +185,11 @@ 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,
@@ -139,6 +217,7 @@ 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)
         return NewSessionResponse(session_id=state.session_id)
 
@@ -153,6 +232,7 @@ 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)
         return LoadSessionResponse()
 
@@ -167,6 +247,7 @@ 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)
         return ResumeSessionResponse()
 
@@ -190,6 +271,8 @@ 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)
         return ForkSessionResponse(session_id=new_id)
 
@@ -383,11 +466,11 @@ class HermesACPAgent(acp.Agent):
 
         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
-            current_provider = getattr(state.agent, "provider", None) or "openrouter"
             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)
@@ -401,9 +484,10 @@ class HermesACPAgent(acp.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 = target_provider or getattr(state.agent, "provider", "auto")
+        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}"
 
@@ -470,16 +554,55 @@ class HermesACPAgent(acp.Agent):
 
     async def set_session_model(
         self, model_id: str, session_id: str, **kwargs: Any
-    ):
+    ) -> 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

@@ -8,6 +8,8 @@ history.
 """
 from __future__ import annotations
 
+from hermes_constants import get_hermes_home
+
 import copy
 import json
 import logging
@@ -251,7 +253,7 @@ class SessionManager:
             import os
             from pathlib import Path
             from hermes_state import SessionDB
-            hermes_home = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
+            hermes_home = get_hermes_home()
             self._db_instance = SessionDB(db_path=hermes_home / "state.db")
             return self._db_instance
         except Exception:
@@ -270,7 +272,17 @@ class SessionManager:
 
         # Ensure model is a plain string (not a MagicMock or other proxy).
         model_str = str(state.model) if state.model else None
-        cwd_json = json.dumps({"cwd": state.cwd})
+        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.
@@ -331,10 +343,18 @@ class SessionManager:
 
         # 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:
-                cwd = json.loads(mc).get("cwd", ".")
+                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
 
@@ -348,7 +368,14 @@ class SessionManager:
             history = []
 
         try:
-            agent = self._make_agent(session_id=session_id, cwd=cwd, model=model)
+            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
@@ -386,6 +413,9 @@ 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()
@@ -396,11 +426,11 @@ 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()
 
@@ -413,12 +443,12 @@ 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 []),

agent/anthropic_adapter.py

@@ -10,10 +10,13 @@ Auth supports:
   - Claude Code credentials (~/.claude.json or ~/.claude/.credentials.json) → Bearer auth
 """
 
+import copy
 import json
 import logging
 import os
 from pathlib import Path
+
+from hermes_constants import get_hermes_home
 from types import SimpleNamespace
 from typing import Any, Dict, List, Optional, Tuple
 
@@ -33,6 +36,54 @@ ADAPTIVE_EFFORT_MAP = {
     "minimal": "low",
 }
 
+# ── Max output token limits per Anthropic model ───────────────────────
+# Source: Anthropic docs + Cline model catalog.  Anthropic's API requires
+# max_tokens as a mandatory field.  Previously we hardcoded 16384, which
+# starves thinking-enabled models (thinking tokens count toward the limit).
+_ANTHROPIC_OUTPUT_LIMITS = {
+    # Claude 4.6
+    "claude-opus-4-6":   128_000,
+    "claude-sonnet-4-6":  64_000,
+    # Claude 4.5
+    "claude-opus-4-5":    64_000,
+    "claude-sonnet-4-5":  64_000,
+    "claude-haiku-4-5":   64_000,
+    # Claude 4
+    "claude-opus-4":      32_000,
+    "claude-sonnet-4":    64_000,
+    # Claude 3.7
+    "claude-3-7-sonnet": 128_000,
+    # Claude 3.5
+    "claude-3-5-sonnet":   8_192,
+    "claude-3-5-haiku":    8_192,
+    # Claude 3
+    "claude-3-opus":       4_096,
+    "claude-3-sonnet":     4_096,
+    "claude-3-haiku":      4_096,
+}
+
+# For any model not in the table, assume the highest current limit.
+# Future Anthropic models are unlikely to have *less* output capacity.
+_ANTHROPIC_DEFAULT_OUTPUT_LIMIT = 128_000
+
+
+def _get_anthropic_max_output(model: str) -> int:
+    """Look up the max output token limit for an Anthropic model.
+
+    Uses substring matching against _ANTHROPIC_OUTPUT_LIMITS so date-stamped
+    model IDs (claude-sonnet-4-5-20250929) and variant suffixes (:1m, :fast)
+    resolve correctly.  Longest-prefix match wins to avoid e.g. "claude-3-5"
+    matching before "claude-3-5-sonnet".
+    """
+    m = model.lower()
+    best_key = ""
+    best_val = _ANTHROPIC_DEFAULT_OUTPUT_LIMIT
+    for key, val in _ANTHROPIC_OUTPUT_LIMITS.items():
+        if key in m and len(key) > len(best_key):
+            best_key = key
+            best_val = val
+    return best_val
+
 
 def _supports_adaptive_thinking(model: str) -> bool:
     """Return True for Claude 4.6 models that support adaptive thinking."""
@@ -57,6 +108,7 @@ _OAUTH_ONLY_BETAS = [
 # The version must stay reasonably current — Anthropic rejects OAuth requests
 # when the spoofed user-agent version is too far behind the actual release.
 _CLAUDE_CODE_VERSION_FALLBACK = "2.1.74"
+_claude_code_version_cache: Optional[str] = None
 
 
 def _detect_claude_code_version() -> str:
@@ -84,11 +136,18 @@ def _detect_claude_code_version() -> str:
     return _CLAUDE_CODE_VERSION_FALLBACK
 
 
-_CLAUDE_CODE_VERSION = _detect_claude_code_version()
 _CLAUDE_CODE_SYSTEM_PREFIX = "You are Claude Code, Anthropic's official CLI for Claude."
 _MCP_TOOL_PREFIX = "mcp_"
 
 
+def _get_claude_code_version() -> str:
+    """Lazily detect the installed Claude Code version when OAuth headers need it."""
+    global _claude_code_version_cache
+    if _claude_code_version_cache is None:
+        _claude_code_version_cache = _detect_claude_code_version()
+    return _claude_code_version_cache
+
+
 def _is_oauth_token(key: str) -> bool:
     """Check if the key is an OAuth/setup token (not a regular Console API key).
 
@@ -104,6 +163,36 @@ def _is_oauth_token(key: str) -> bool:
     return True
 
 
+def _is_third_party_anthropic_endpoint(base_url: str | None) -> bool:
+    """Return True for non-Anthropic endpoints using the Anthropic Messages API.
+
+    Third-party proxies (Azure AI Foundry, AWS Bedrock, self-hosted) authenticate
+    with their own API keys via x-api-key, not Anthropic OAuth tokens. OAuth
+    detection should be skipped for these endpoints.
+    """
+    if not base_url:
+        return False  # No base_url = direct Anthropic API
+    normalized = base_url.rstrip("/").lower()
+    if "anthropic.com" in normalized:
+        return False  # Direct Anthropic API — OAuth applies
+    return True  # Any other endpoint is a third-party proxy
+
+
+def _requires_bearer_auth(base_url: str | None) -> bool:
+    """Return True for Anthropic-compatible providers that require Bearer auth.
+
+    Some third-party /anthropic endpoints implement Anthropic's Messages API but
+    require Authorization: Bearer instead of Anthropic's native x-api-key header.
+    MiniMax's global and China Anthropic-compatible endpoints follow this pattern.
+    """
+    if not base_url:
+        return False
+    normalized = base_url.rstrip("/").lower()
+    return normalized.startswith("https://api.minimax.io/anthropic") or normalized.startswith(
+        "https://api.minimaxi.com/anthropic"
+    )
+
+
 def build_anthropic_client(api_key: str, base_url: str = None):
     """Create an Anthropic client, auto-detecting setup-tokens vs API keys.
 
@@ -122,7 +211,25 @@ def build_anthropic_client(api_key: str, base_url: str = None):
     if base_url:
         kwargs["base_url"] = base_url
 
-    if _is_oauth_token(api_key):
+    if _requires_bearer_auth(base_url):
+        # Some Anthropic-compatible providers (e.g. MiniMax) expect the API key in
+        # Authorization: Bearer even for regular API keys. Route those endpoints
+        # through auth_token so the SDK sends Bearer auth instead of x-api-key.
+        # Check this before OAuth token shape detection because MiniMax secrets do
+        # not use Anthropic's sk-ant-api prefix and would otherwise be misread as
+        # Anthropic OAuth/setup tokens.
+        kwargs["auth_token"] = api_key
+        if _COMMON_BETAS:
+            kwargs["default_headers"] = {"anthropic-beta": ",".join(_COMMON_BETAS)}
+    elif _is_third_party_anthropic_endpoint(base_url):
+        # Third-party proxies (Azure AI Foundry, AWS Bedrock, etc.) use their
+        # own API keys with x-api-key auth. Skip OAuth detection — their keys
+        # don't follow Anthropic's sk-ant-* prefix convention and would be
+        # misclassified as OAuth tokens.
+        kwargs["api_key"] = api_key
+        if _COMMON_BETAS:
+            kwargs["default_headers"] = {"anthropic-beta": ",".join(_COMMON_BETAS)}
+    elif _is_oauth_token(api_key):
         # OAuth access token / setup-token → Bearer auth + Claude Code identity.
         # Anthropic routes OAuth requests based on user-agent and headers;
         # without Claude Code's fingerprint, requests get intermittent 500s.
@@ -130,7 +237,7 @@ def build_anthropic_client(api_key: str, base_url: str = None):
         kwargs["auth_token"] = api_key
         kwargs["default_headers"] = {
             "anthropic-beta": ",".join(all_betas),
-            "user-agent": f"claude-cli/{_CLAUDE_CODE_VERSION} (external, cli)",
+            "user-agent": f"claude-cli/{_get_claude_code_version()} (external, cli)",
             "x-app": "cli",
         }
     else:
@@ -201,64 +308,105 @@ def is_claude_code_token_valid(creds: Dict[str, Any]) -> bool:
     return now_ms < (expires_at - 60_000)
 
 
-def _refresh_oauth_token(creds: Dict[str, Any]) -> Optional[str]:
-    """Attempt to refresh an expired Claude Code OAuth token.
-
-    Uses the same token endpoint and client_id as Claude Code / OpenCode.
-    Only works for credentials that have a refresh token (from claude /login
-    or claude setup-token with OAuth flow).
-
-    Returns the new access token, or None if refresh fails.
-    """
+def refresh_anthropic_oauth_pure(refresh_token: str, *, use_json: bool = False) -> Dict[str, Any]:
+    """Refresh an Anthropic OAuth token without mutating local credential files."""
+    import time
     import urllib.parse
     import urllib.request
 
+    if not refresh_token:
+        raise ValueError("refresh_token is required")
+
+    client_id = "9d1c250a-e61b-44d9-88ed-5944d1962f5e"
+    if use_json:
+        data = json.dumps({
+            "grant_type": "refresh_token",
+            "refresh_token": refresh_token,
+            "client_id": client_id,
+        }).encode()
+        content_type = "application/json"
+    else:
+        data = urllib.parse.urlencode({
+            "grant_type": "refresh_token",
+            "refresh_token": refresh_token,
+            "client_id": client_id,
+        }).encode()
+        content_type = "application/x-www-form-urlencoded"
+
+    token_endpoints = [
+        "https://platform.claude.com/v1/oauth/token",
+        "https://console.anthropic.com/v1/oauth/token",
+    ]
+    last_error = None
+    for endpoint in token_endpoints:
+        req = urllib.request.Request(
+            endpoint,
+            data=data,
+            headers={
+                "Content-Type": content_type,
+                "User-Agent": f"claude-cli/{_get_claude_code_version()} (external, cli)",
+            },
+            method="POST",
+        )
+        try:
+            with urllib.request.urlopen(req, timeout=10) as resp:
+                result = json.loads(resp.read().decode())
+        except Exception as exc:
+            last_error = exc
+            logger.debug("Anthropic token refresh failed at %s: %s", endpoint, exc)
+            continue
+
+        access_token = result.get("access_token", "")
+        if not access_token:
+            raise ValueError("Anthropic refresh response was missing access_token")
+        next_refresh = result.get("refresh_token", refresh_token)
+        expires_in = result.get("expires_in", 3600)
+        return {
+            "access_token": access_token,
+            "refresh_token": next_refresh,
+            "expires_at_ms": int(time.time() * 1000) + (expires_in * 1000),
+        }
+
+    if last_error is not None:
+        raise last_error
+    raise ValueError("Anthropic token refresh failed")
+
+
+def _refresh_oauth_token(creds: Dict[str, Any]) -> Optional[str]:
+    """Attempt to refresh an expired Claude Code OAuth token."""
     refresh_token = creds.get("refreshToken", "")
     if not refresh_token:
         logger.debug("No refresh token available — cannot refresh")
         return None
 
-    # Client ID used by Claude Code's OAuth flow
-    CLIENT_ID = "9d1c250a-e61b-44d9-88ed-5944d1962f5e"
-
-    data = urllib.parse.urlencode({
-        "grant_type": "refresh_token",
-        "refresh_token": refresh_token,
-        "client_id": CLIENT_ID,
-    }).encode()
-
-    req = urllib.request.Request(
-        "https://console.anthropic.com/v1/oauth/token",
-        data=data,
-        headers={
-            "Content-Type": "application/x-www-form-urlencoded",
-            "User-Agent": f"claude-cli/{_CLAUDE_CODE_VERSION} (external, cli)",
-        },
-        method="POST",
-    )
-
     try:
-        with urllib.request.urlopen(req, timeout=10) as resp:
-            result = json.loads(resp.read().decode())
-            new_access = result.get("access_token", "")
-            new_refresh = result.get("refresh_token", refresh_token)
-            expires_in = result.get("expires_in", 3600)  # seconds
-
-            if new_access:
-                import time
-                new_expires_ms = int(time.time() * 1000) + (expires_in * 1000)
-                # Write refreshed credentials back to ~/.claude/.credentials.json
-                _write_claude_code_credentials(new_access, new_refresh, new_expires_ms)
-                logger.debug("Successfully refreshed Claude Code OAuth token")
-                return new_access
+        refreshed = refresh_anthropic_oauth_pure(refresh_token, use_json=False)
+        _write_claude_code_credentials(
+            refreshed["access_token"],
+            refreshed["refresh_token"],
+            refreshed["expires_at_ms"],
+        )
+        logger.debug("Successfully refreshed Claude Code OAuth token")
+        return refreshed["access_token"]
     except Exception as e:
         logger.debug("Failed to refresh Claude Code token: %s", e)
-
-    return None
+        return None
 
 
-def _write_claude_code_credentials(access_token: str, refresh_token: str, expires_at_ms: int) -> None:
-    """Write refreshed credentials back to ~/.claude/.credentials.json."""
+def _write_claude_code_credentials(
+    access_token: str,
+    refresh_token: str,
+    expires_at_ms: int,
+    *,
+    scopes: Optional[list] = None,
+) -> None:
+    """Write refreshed credentials back to ~/.claude/.credentials.json.
+
+    The optional *scopes* list (e.g. ``["user:inference", "user:profile", ...]``)
+    is persisted so that Claude Code's own auth check recognises the credential
+    as valid.  Claude Code >=2.1.81 gates on the presence of ``"user:inference"``
+    in the stored scopes before it will use the token.
+    """
     cred_path = Path.home() / ".claude" / ".credentials.json"
     try:
         # Read existing file to preserve other fields
@@ -266,11 +414,19 @@ def _write_claude_code_credentials(access_token: str, refresh_token: str, expire
         if cred_path.exists():
             existing = json.loads(cred_path.read_text(encoding="utf-8"))
 
-        existing["claudeAiOauth"] = {
+        oauth_data: Dict[str, Any] = {
             "accessToken": access_token,
             "refreshToken": refresh_token,
             "expiresAt": expires_at_ms,
         }
+        if scopes is not None:
+            oauth_data["scopes"] = scopes
+        elif "claudeAiOauth" in existing and "scopes" in existing["claudeAiOauth"]:
+            # Preserve previously-stored scopes when the refresh response
+            # does not include a scope field.
+            oauth_data["scopes"] = existing["claudeAiOauth"]["scopes"]
+
+        existing["claudeAiOauth"] = oauth_data
 
         cred_path.parent.mkdir(parents=True, exist_ok=True)
         cred_path.write_text(json.dumps(existing, indent=2), encoding="utf-8")
@@ -376,24 +532,12 @@ def resolve_anthropic_token() -> Optional[str]:
             return preferred
         return cc_token
 
-    # 3. Hermes-managed OAuth credentials (~/.hermes/.anthropic_oauth.json)
-    hermes_creds = read_hermes_oauth_credentials()
-    if hermes_creds:
-        if is_claude_code_token_valid(hermes_creds):
-            logger.debug("Using Hermes-managed OAuth credentials")
-            return hermes_creds["accessToken"]
-        # Expired — try refresh
-        logger.debug("Hermes OAuth token expired — attempting refresh")
-        refreshed = refresh_hermes_oauth_token()
-        if refreshed:
-            return refreshed
-
-    # 4. Claude Code credential file
+    # 3. Claude Code credential file
     resolved_claude_token = _resolve_claude_code_token_from_credentials(creds)
     if resolved_claude_token:
         return resolved_claude_token
 
-    # 5. Regular API key, or a legacy OAuth token saved in ANTHROPIC_API_KEY.
+    # 4. Regular API key, or a legacy OAuth token saved in ANTHROPIC_API_KEY.
     # This remains as a compatibility fallback for pre-migration Hermes configs.
     api_key = os.getenv("ANTHROPIC_API_KEY", "").strip()
     if api_key:
@@ -450,7 +594,7 @@ _OAUTH_CLIENT_ID = "9d1c250a-e61b-44d9-88ed-5944d1962f5e"
 _OAUTH_TOKEN_URL = "https://console.anthropic.com/v1/oauth/token"
 _OAUTH_REDIRECT_URI = "https://console.anthropic.com/oauth/code/callback"
 _OAUTH_SCOPES = "org:create_api_key user:profile user:inference"
-_HERMES_OAUTH_FILE = Path(os.getenv("HERMES_HOME", str(Path.home() / ".hermes"))) / ".anthropic_oauth.json"
+_HERMES_OAUTH_FILE = get_hermes_home() / ".anthropic_oauth.json"
 
 
 def _generate_pkce() -> tuple:
@@ -466,20 +610,13 @@ def _generate_pkce() -> tuple:
     return verifier, challenge
 
 
-def run_hermes_oauth_login() -> Optional[str]:
-    """Run Hermes-native OAuth PKCE flow for Claude Pro/Max subscription.
-
-    Opens a browser to claude.ai for authorization, prompts for the code,
-    exchanges it for tokens, and stores them in ~/.hermes/.anthropic_oauth.json.
-
-    Returns the access token on success, None on failure.
-    """
+def run_hermes_oauth_login_pure() -> Optional[Dict[str, Any]]:
+    """Run Hermes-native OAuth PKCE flow and return credential state."""
     import time
     import webbrowser
 
     verifier, challenge = _generate_pkce()
 
-    # Build authorization URL
     params = {
         "code": "true",
         "client_id": _OAUTH_CLIENT_ID,
@@ -491,6 +628,7 @@ def run_hermes_oauth_login() -> Optional[str]:
         "state": verifier,
     }
     from urllib.parse import urlencode
+
     auth_url = f"https://claude.ai/oauth/authorize?{urlencode(params)}"
 
     print()
@@ -504,7 +642,6 @@ def run_hermes_oauth_login() -> Optional[str]:
     print(f"  {auth_url}")
     print()
 
-    # Try to open browser automatically (works on desktop, silently fails on headless/SSH)
     try:
         webbrowser.open(auth_url)
         print("  (Browser opened automatically)")
@@ -523,14 +660,13 @@ def run_hermes_oauth_login() -> Optional[str]:
         print("No code entered.")
         return None
 
-    # Split code#state format
     splits = auth_code.split("#")
     code = splits[0]
     state = splits[1] if len(splits) > 1 else ""
 
-    # Exchange code for tokens
     try:
         import urllib.request
+
         exchange_data = json.dumps({
             "grant_type": "authorization_code",
             "client_id": _OAUTH_CLIENT_ID,
@@ -545,7 +681,7 @@ def run_hermes_oauth_login() -> Optional[str]:
             data=exchange_data,
             headers={
                 "Content-Type": "application/json",
-                "User-Agent": f"claude-cli/{_CLAUDE_CODE_VERSION} (external, cli)",
+                "User-Agent": f"claude-cli/{_get_claude_code_version()} (external, cli)",
             },
             method="POST",
         )
@@ -564,11 +700,31 @@ def run_hermes_oauth_login() -> Optional[str]:
         print("No access token in response.")
         return None
 
-    # Store credentials
     expires_at_ms = int(time.time() * 1000) + (expires_in * 1000)
-    _save_hermes_oauth_credentials(access_token, refresh_token, expires_at_ms)
+    return {
+        "access_token": access_token,
+        "refresh_token": refresh_token,
+        "expires_at_ms": expires_at_ms,
+    }
 
-    # Also write to Claude Code's credential file for backward compat
+
+def run_hermes_oauth_login() -> Optional[str]:
+    """Run Hermes-native OAuth PKCE flow for Claude Pro/Max subscription.
+
+    Opens a browser to claude.ai for authorization, prompts for the code,
+    exchanges it for tokens, and stores them in ~/.hermes/.anthropic_oauth.json.
+
+    Returns the access token on success, None on failure.
+    """
+    result = run_hermes_oauth_login_pure()
+    if not result:
+        return None
+
+    access_token = result["access_token"]
+    refresh_token = result["refresh_token"]
+    expires_at_ms = result["expires_at_ms"]
+
+    _save_hermes_oauth_credentials(access_token, refresh_token, expires_at_ms)
     _write_claude_code_credentials(access_token, refresh_token, expires_at_ms)
 
     print("Authentication successful!")
@@ -607,44 +763,27 @@ def refresh_hermes_oauth_token() -> Optional[str]:
 
     Returns the new access token, or None if refresh fails.
     """
-    import time
-    import urllib.request
-
     creds = read_hermes_oauth_credentials()
     if not creds or not creds.get("refreshToken"):
         return None
 
     try:
-        data = json.dumps({
-            "grant_type": "refresh_token",
-            "refresh_token": creds["refreshToken"],
-            "client_id": _OAUTH_CLIENT_ID,
-        }).encode()
-
-        req = urllib.request.Request(
-            _OAUTH_TOKEN_URL,
-            data=data,
-            headers={
-                "Content-Type": "application/json",
-                "User-Agent": f"claude-cli/{_CLAUDE_CODE_VERSION} (external, cli)",
-            },
-            method="POST",
+        refreshed = refresh_anthropic_oauth_pure(
+            creds["refreshToken"],
+            use_json=True,
         )
-
-        with urllib.request.urlopen(req, timeout=10) as resp:
-            result = json.loads(resp.read().decode())
-
-        new_access = result.get("access_token", "")
-        new_refresh = result.get("refresh_token", creds["refreshToken"])
-        expires_in = result.get("expires_in", 3600)
-
-        if new_access:
-            new_expires_ms = int(time.time() * 1000) + (expires_in * 1000)
-            _save_hermes_oauth_credentials(new_access, new_refresh, new_expires_ms)
-            # Also update Claude Code's credential file
-            _write_claude_code_credentials(new_access, new_refresh, new_expires_ms)
-            logger.debug("Successfully refreshed Hermes OAuth token")
-            return new_access
+        _save_hermes_oauth_credentials(
+            refreshed["access_token"],
+            refreshed["refresh_token"],
+            refreshed["expires_at_ms"],
+        )
+        _write_claude_code_credentials(
+            refreshed["access_token"],
+            refreshed["refresh_token"],
+            refreshed["expires_at_ms"],
+        )
+        logger.debug("Successfully refreshed Hermes OAuth token")
+        return refreshed["access_token"]
     except Exception as e:
         logger.debug("Failed to refresh Hermes OAuth token: %s", e)
 
@@ -656,19 +795,21 @@ def refresh_hermes_oauth_token() -> Optional[str]:
 # ---------------------------------------------------------------------------
 
 
-def normalize_model_name(model: str) -> str:
+def normalize_model_name(model: str, preserve_dots: bool = False) -> str:
     """Normalize a model name for the Anthropic API.
 
     - Strips 'anthropic/' prefix (OpenRouter format, case-insensitive)
     - Converts dots to hyphens in version numbers (OpenRouter uses dots,
-      Anthropic uses hyphens: claude-opus-4.6 → claude-opus-4-6)
+      Anthropic uses hyphens: claude-opus-4.6 → claude-opus-4-6), unless
+      preserve_dots is True (e.g. for Alibaba/DashScope: qwen3.5-plus).
     """
     lower = model.lower()
     if lower.startswith("anthropic/"):
         model = model[len("anthropic/"):]
-    # OpenRouter uses dots for version separators (claude-opus-4.6),
-    # Anthropic uses hyphens (claude-opus-4-6). Convert dots to hyphens.
-    model = model.replace(".", "-")
+    if not preserve_dots:
+        # OpenRouter uses dots for version separators (claude-opus-4.6),
+        # Anthropic uses hyphens (claude-opus-4-6). Convert dots to hyphens.
+        model = model.replace(".", "-")
     return model
 
 
@@ -809,6 +950,69 @@ def _convert_content_part_to_anthropic(part: Any) -> Optional[Dict[str, Any]]:
     return block
 
 
+def _to_plain_data(value: Any, *, _depth: int = 0, _path: Optional[set] = None) -> Any:
+    """Recursively convert SDK objects to plain Python data structures.
+
+    Guards against circular references (``_path`` tracks ``id()`` of objects
+    on the *current* recursion path) and runaway depth (capped at 20 levels).
+    Uses path-based tracking so shared (but non-cyclic) objects referenced by
+    multiple siblings are converted correctly rather than being stringified.
+    """
+    _MAX_DEPTH = 20
+    if _depth > _MAX_DEPTH:
+        return str(value)
+
+    if _path is None:
+        _path = set()
+
+    obj_id = id(value)
+    if obj_id in _path:
+        return str(value)
+
+    if hasattr(value, "model_dump"):
+        _path.add(obj_id)
+        result = _to_plain_data(value.model_dump(), _depth=_depth + 1, _path=_path)
+        _path.discard(obj_id)
+        return result
+    if isinstance(value, dict):
+        _path.add(obj_id)
+        result = {k: _to_plain_data(v, _depth=_depth + 1, _path=_path) for k, v in value.items()}
+        _path.discard(obj_id)
+        return result
+    if isinstance(value, (list, tuple)):
+        _path.add(obj_id)
+        result = [_to_plain_data(v, _depth=_depth + 1, _path=_path) for v in value]
+        _path.discard(obj_id)
+        return result
+    if hasattr(value, "__dict__"):
+        _path.add(obj_id)
+        result = {
+            k: _to_plain_data(v, _depth=_depth + 1, _path=_path)
+            for k, v in vars(value).items()
+            if not k.startswith("_")
+        }
+        _path.discard(obj_id)
+        return result
+    return value
+
+
+def _extract_preserved_thinking_blocks(message: Dict[str, Any]) -> List[Dict[str, Any]]:
+    """Return Anthropic thinking blocks previously preserved on the message."""
+    raw_details = message.get("reasoning_details")
+    if not isinstance(raw_details, list):
+        return []
+
+    preserved: List[Dict[str, Any]] = []
+    for detail in raw_details:
+        if not isinstance(detail, dict):
+            continue
+        block_type = str(detail.get("type", "") or "").strip().lower()
+        if block_type not in {"thinking", "redacted_thinking"}:
+            continue
+        preserved.append(copy.deepcopy(detail))
+    return preserved
+
+
 def _convert_content_to_anthropic(content: Any) -> Any:
     """Convert OpenAI-style multimodal content arrays to Anthropic blocks."""
     if not isinstance(content, list):
@@ -855,7 +1059,7 @@ def convert_messages_to_anthropic(
             continue
 
         if role == "assistant":
-            blocks = []
+            blocks = _extract_preserved_thinking_blocks(m)
             if content:
                 if isinstance(content, list):
                     converted_content = _convert_content_to_anthropic(content)
@@ -910,14 +1114,21 @@ def convert_messages_to_anthropic(
                 result.append({"role": "user", "content": [tool_result]})
             continue
 
-        # Regular user message
+        # Regular user message — validate non-empty content (Anthropic rejects empty)
         if isinstance(content, list):
             converted_blocks = _convert_content_to_anthropic(content)
-            result.append({
-                "role": "user",
-                "content": converted_blocks or [{"type": "text", "text": ""}],
-            })
+            # Check if all text blocks are empty
+            if not converted_blocks or all(
+                b.get("text", "").strip() == ""
+                for b in converted_blocks
+                if isinstance(b, dict) and b.get("type") == "text"
+            ):
+                converted_blocks = [{"type": "text", "text": "(empty message)"}]
+            result.append({"role": "user", "content": converted_blocks})
         else:
+            # Validate string content is non-empty
+            if not content or (isinstance(content, str) and not content.strip()):
+                content = "(empty message)"
             result.append({"role": "user", "content": content})
 
     # Strip orphaned tool_use blocks (no matching tool_result follows)
@@ -1006,17 +1217,32 @@ def build_anthropic_kwargs(
     reasoning_config: Optional[Dict[str, Any]],
     tool_choice: Optional[str] = None,
     is_oauth: bool = False,
+    preserve_dots: bool = False,
+    context_length: Optional[int] = None,
 ) -> Dict[str, Any]:
     """Build kwargs for anthropic.messages.create().
 
+    When *max_tokens* is None, the model's native output limit is used
+    (e.g. 128K for Opus 4.6, 64K for Sonnet 4.6).  If *context_length*
+    is provided, the effective limit is clamped so it doesn't exceed
+    the context window.
+
     When *is_oauth* is True, applies Claude Code compatibility transforms:
     system prompt prefix, tool name prefixing, and prompt sanitization.
+
+    When *preserve_dots* is True, model name dots are not converted to hyphens
+    (for Alibaba/DashScope anthropic-compatible endpoints: qwen3.5-plus).
     """
     system, anthropic_messages = convert_messages_to_anthropic(messages)
     anthropic_tools = convert_tools_to_anthropic(tools) if tools else []
 
-    model = normalize_model_name(model)
-    effective_max_tokens = max_tokens or 16384
+    model = normalize_model_name(model, preserve_dots=preserve_dots)
+    effective_max_tokens = max_tokens or _get_anthropic_max_output(model)
+
+    # Clamp to context window if the user set a lower context_length
+    # (e.g. custom endpoint with limited capacity).
+    if context_length and effective_max_tokens > context_length:
+        effective_max_tokens = max(context_length - 1, 1)
 
     # ── OAuth: Claude Code identity ──────────────────────────────────
     if is_oauth:
@@ -1117,6 +1343,7 @@ def normalize_anthropic_response(
     """
     text_parts = []
     reasoning_parts = []
+    reasoning_details = []
     tool_calls = []
 
     for block in response.content:
@@ -1124,6 +1351,9 @@ def normalize_anthropic_response(
             text_parts.append(block.text)
         elif block.type == "thinking":
             reasoning_parts.append(block.thinking)
+            block_dict = _to_plain_data(block)
+            if isinstance(block_dict, dict):
+                reasoning_details.append(block_dict)
         elif block.type == "tool_use":
             name = block.name
             if strip_tool_prefix and name.startswith(_MCP_TOOL_PREFIX):
@@ -1154,7 +1384,7 @@ def normalize_anthropic_response(
             tool_calls=tool_calls or None,
             reasoning="\n\n".join(reasoning_parts) if reasoning_parts else None,
             reasoning_content=None,
-            reasoning_details=None,
+            reasoning_details=reasoning_details or None,
         ),
         finish_reason,
-    )
+    )

agent/auxiliary_client.py

@@ -7,7 +7,7 @@ the best available backend without duplicating fallback logic.
 Resolution order for text tasks (auto mode):
   1. OpenRouter  (OPENROUTER_API_KEY)
   2. Nous Portal (~/.hermes/auth.json active provider)
-  3. Custom endpoint (OPENAI_BASE_URL + OPENAI_API_KEY)
+  3. Custom endpoint (config.yaml model.base_url + OPENAI_API_KEY)
   4. Codex OAuth (Responses API via chatgpt.com with gpt-5.3-codex,
      wrapped to look like a chat.completions client)
   5. Native Anthropic
@@ -40,12 +40,14 @@ import json
 import logging
 import os
 import threading
-from pathlib import Path
+import time
+from pathlib import Path  # noqa: F401 — used by test mocks
 from types import SimpleNamespace
 from typing import Any, Dict, List, Optional, Tuple
 
 from openai import OpenAI
 
+from agent.credential_pool import load_pool
 from hermes_cli.config import get_hermes_home
 from hermes_constants import OPENROUTER_BASE_URL
 
@@ -81,7 +83,7 @@ auxiliary_is_nous: bool = False
 
 # Default auxiliary models per provider
 _OPENROUTER_MODEL = "google/gemini-3-flash-preview"
-_NOUS_MODEL = "gemini-3-flash"
+_NOUS_MODEL = "google/gemini-3-flash-preview"
 _NOUS_DEFAULT_BASE_URL = "https://inference-api.nousresearch.com/v1"
 _ANTHROPIC_DEFAULT_BASE_URL = "https://api.anthropic.com"
 _AUTH_JSON_PATH = get_hermes_home() / "auth.json"
@@ -95,6 +97,45 @@ _CODEX_AUX_MODEL = "gpt-5.2-codex"
 _CODEX_AUX_BASE_URL = "https://chatgpt.com/backend-api/codex"
 
 
+def _select_pool_entry(provider: str) -> Tuple[bool, Optional[Any]]:
+    """Return (pool_exists_for_provider, selected_entry)."""
+    try:
+        pool = load_pool(provider)
+    except Exception as exc:
+        logger.debug("Auxiliary client: could not load pool for %s: %s", provider, exc)
+        return False, None
+    if not pool or not pool.has_credentials():
+        return False, None
+    try:
+        return True, pool.select()
+    except Exception as exc:
+        logger.debug("Auxiliary client: could not select pool entry for %s: %s", provider, exc)
+        return True, None
+
+
+def _pool_runtime_api_key(entry: Any) -> str:
+    if entry is None:
+        return ""
+    # Use the PooledCredential.runtime_api_key property which handles
+    # provider-specific fallback (e.g. agent_key for nous).
+    key = getattr(entry, "runtime_api_key", None) or getattr(entry, "access_token", "")
+    return str(key or "").strip()
+
+
+def _pool_runtime_base_url(entry: Any, fallback: str = "") -> str:
+    if entry is None:
+        return str(fallback or "").strip().rstrip("/")
+    # runtime_base_url handles provider-specific logic (e.g. nous prefers inference_base_url).
+    # Fall back through inference_base_url and base_url for non-PooledCredential entries.
+    url = (
+        getattr(entry, "runtime_base_url", None)
+        or getattr(entry, "inference_base_url", None)
+        or getattr(entry, "base_url", None)
+        or fallback
+    )
+    return str(url or "").strip().rstrip("/")
+
+
 # ── Codex Responses → chat.completions adapter ─────────────────────────────
 # All auxiliary consumers call client.chat.completions.create(**kwargs) and
 # read response.choices[0].message.content. This adapter translates those
@@ -325,9 +366,10 @@ class AsyncCodexAuxiliaryClient:
 class _AnthropicCompletionsAdapter:
     """OpenAI-client-compatible adapter for Anthropic Messages API."""
 
-    def __init__(self, real_client: Any, model: str):
+    def __init__(self, real_client: Any, model: str, is_oauth: bool = False):
         self._client = real_client
         self._model = model
+        self._is_oauth = is_oauth
 
     def create(self, **kwargs) -> Any:
         from agent.anthropic_adapter import build_anthropic_kwargs, normalize_anthropic_response
@@ -356,6 +398,7 @@ class _AnthropicCompletionsAdapter:
             max_tokens=max_tokens,
             reasoning_config=None,
             tool_choice=normalized_tool_choice,
+            is_oauth=self._is_oauth,
         )
         if temperature is not None:
             anthropic_kwargs["temperature"] = temperature
@@ -394,9 +437,9 @@ class _AnthropicChatShim:
 class AnthropicAuxiliaryClient:
     """OpenAI-client-compatible wrapper over a native Anthropic client."""
 
-    def __init__(self, real_client: Any, model: str, api_key: str, base_url: str):
+    def __init__(self, real_client: Any, model: str, api_key: str, base_url: str, is_oauth: bool = False):
         self._real_client = real_client
-        adapter = _AnthropicCompletionsAdapter(real_client, model)
+        adapter = _AnthropicCompletionsAdapter(real_client, model, is_oauth=is_oauth)
         self.chat = _AnthropicChatShim(adapter)
         self.api_key = api_key
         self.base_url = base_url
@@ -436,6 +479,22 @@ def _read_nous_auth() -> Optional[dict]:
     Returns the provider state dict if Nous is active with tokens,
     otherwise None.
     """
+    pool_present, entry = _select_pool_entry("nous")
+    if pool_present:
+        if entry is None:
+            return None
+        return {
+            "access_token": getattr(entry, "access_token", ""),
+            "refresh_token": getattr(entry, "refresh_token", None),
+            "agent_key": getattr(entry, "agent_key", None),
+            "inference_base_url": _pool_runtime_base_url(entry, _NOUS_DEFAULT_BASE_URL),
+            "portal_base_url": getattr(entry, "portal_base_url", None),
+            "client_id": getattr(entry, "client_id", None),
+            "scope": getattr(entry, "scope", None),
+            "token_type": getattr(entry, "token_type", "Bearer"),
+            "source": "pool",
+        }
+
     try:
         if not _AUTH_JSON_PATH.is_file():
             return None
@@ -463,15 +522,35 @@ def _nous_base_url() -> str:
 
 
 def _read_codex_access_token() -> Optional[str]:
-    """Read a valid Codex OAuth access token from Hermes auth store (~/.hermes/auth.json)."""
+    """Read a valid, non-expired Codex OAuth access token from Hermes auth store."""
+    pool_present, entry = _select_pool_entry("openai-codex")
+    if pool_present:
+        token = _pool_runtime_api_key(entry)
+        return token or None
+
     try:
         from hermes_cli.auth import _read_codex_tokens
         data = _read_codex_tokens()
         tokens = data.get("tokens", {})
         access_token = tokens.get("access_token")
-        if isinstance(access_token, str) and access_token.strip():
-            return access_token.strip()
-        return None
+        if not isinstance(access_token, str) or not access_token.strip():
+            return None
+
+        # Check JWT expiry — expired tokens block the auto chain and
+        # prevent fallback to working providers (e.g. Anthropic).
+        try:
+            import base64
+            payload = access_token.split(".")[1]
+            payload += "=" * (-len(payload) % 4)
+            claims = json.loads(base64.urlsafe_b64decode(payload))
+            exp = claims.get("exp", 0)
+            if exp and time.time() > exp:
+                logger.debug("Codex access token expired (exp=%s), skipping", exp)
+                return None
+        except Exception:
+            pass  # Non-JWT token or decode error — use as-is
+
+        return access_token.strip()
     except Exception as exc:
         logger.debug("Could not read Codex auth for auxiliary client: %s", exc)
         return None
@@ -495,6 +574,24 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
         if provider_id == "anthropic":
             return _try_anthropic()
 
+        pool_present, entry = _select_pool_entry(provider_id)
+        if pool_present:
+            api_key = _pool_runtime_api_key(entry)
+            if not api_key:
+                continue
+
+            base_url = _pool_runtime_base_url(entry, pconfig.inference_base_url) or pconfig.inference_base_url
+            model = _API_KEY_PROVIDER_AUX_MODELS.get(provider_id, "default")
+            logger.debug("Auxiliary text client: %s (%s) via pool", pconfig.name, model)
+            extra = {}
+            if "api.kimi.com" in base_url.lower():
+                extra["default_headers"] = {"User-Agent": "KimiCLI/1.0"}
+            elif "api.githubcopilot.com" in base_url.lower():
+                from hermes_cli.models import copilot_default_headers
+
+                extra["default_headers"] = copilot_default_headers()
+            return OpenAI(api_key=api_key, base_url=base_url, **extra), model
+
         creds = resolve_api_key_provider_credentials(provider_id)
         api_key = str(creds.get("api_key", "")).strip()
         if not api_key:
@@ -544,6 +641,16 @@ def _get_auxiliary_env_override(task: str, suffix: str) -> Optional[str]:
 
 
 def _try_openrouter() -> Tuple[Optional[OpenAI], Optional[str]]:
+    pool_present, entry = _select_pool_entry("openrouter")
+    if pool_present:
+        or_key = _pool_runtime_api_key(entry)
+        if not or_key:
+            return None, None
+        base_url = _pool_runtime_base_url(entry, OPENROUTER_BASE_URL) or OPENROUTER_BASE_URL
+        logger.debug("Auxiliary client: OpenRouter via pool")
+        return OpenAI(api_key=or_key, base_url=base_url,
+                       default_headers=_OR_HEADERS), _OPENROUTER_MODEL
+
     or_key = os.getenv("OPENROUTER_API_KEY")
     if not or_key:
         return None, None
@@ -559,22 +666,22 @@ def _try_nous() -> Tuple[Optional[OpenAI], Optional[str]]:
     global auxiliary_is_nous
     auxiliary_is_nous = True
     logger.debug("Auxiliary client: Nous Portal")
+    model = "gemini-3-flash" if nous.get("source") == "pool" else _NOUS_MODEL
     return (
-        OpenAI(api_key=_nous_api_key(nous), base_url=_nous_base_url()),
-        _NOUS_MODEL,
+        OpenAI(
+            api_key=_nous_api_key(nous),
+            base_url=str(nous.get("inference_base_url") or _nous_base_url()).rstrip("/"),
+        ),
+        model,
     )
 
 
 def _read_main_model() -> str:
-    """Read the user's configured main model from config/env.
+    """Read the user's configured main model from config.yaml.
 
-    Falls back through HERMES_MODEL → LLM_MODEL → config.yaml model.default
-    so the auxiliary client can use the same model as the main agent when no
-    dedicated auxiliary model is available.
+    config.yaml model.default is the single source of truth for the active
+    model. Environment variables are no longer consulted.
     """
-    from_env = os.getenv("OPENAI_MODEL") or os.getenv("HERMES_MODEL") or os.getenv("LLM_MODEL")
-    if from_env:
-        return from_env.strip()
     try:
         from hermes_cli.config import load_config
         cfg = load_config()
@@ -590,6 +697,25 @@ def _read_main_model() -> str:
     return ""
 
 
+def _read_main_provider() -> str:
+    """Read the user's configured main provider from config.yaml.
+
+    Returns the lowercase provider id (e.g. "alibaba", "openrouter") or ""
+    if not configured.
+    """
+    try:
+        from hermes_cli.config import load_config
+        cfg = load_config()
+        model_cfg = cfg.get("model", {})
+        if isinstance(model_cfg, dict):
+            provider = model_cfg.get("provider", "")
+            if isinstance(provider, str) and provider.strip():
+                return provider.strip().lower()
+    except Exception:
+        pass
+    return ""
+
+
 def _resolve_custom_runtime() -> Tuple[Optional[str], Optional[str]]:
     """Resolve the active custom/main endpoint the same way the main CLI does.
 
@@ -609,8 +735,6 @@ def _resolve_custom_runtime() -> Tuple[Optional[str], Optional[str]]:
     custom_key = runtime.get("api_key")
     if not isinstance(custom_base, str) or not custom_base.strip():
         return None, None
-    if not isinstance(custom_key, str) or not custom_key.strip():
-        return None, None
 
     custom_base = custom_base.strip().rstrip("/")
     if "openrouter.ai" in custom_base.lower():
@@ -618,6 +742,13 @@ def _resolve_custom_runtime() -> Tuple[Optional[str], Optional[str]]:
         # configured. Treat that as "no custom endpoint" for auxiliary routing.
         return None, None
 
+    # Local servers (Ollama, llama.cpp, vLLM, LM Studio) don't require auth.
+    # Use a placeholder key — the OpenAI SDK requires a non-empty string but
+    # local servers ignore the Authorization header.  Same fix as cli.py
+    # _ensure_runtime_credentials() (PR #2556).
+    if not isinstance(custom_key, str) or not custom_key.strip():
+        custom_key = "no-key-required"
+
     return custom_base, custom_key.strip()
 
 
@@ -636,11 +767,19 @@ def _try_custom_endpoint() -> Tuple[Optional[OpenAI], Optional[str]]:
 
 
 def _try_codex() -> Tuple[Optional[Any], Optional[str]]:
-    codex_token = _read_codex_access_token()
-    if not codex_token:
-        return None, None
+    pool_present, entry = _select_pool_entry("openai-codex")
+    if pool_present:
+        codex_token = _pool_runtime_api_key(entry)
+        if not codex_token:
+            return None, None
+        base_url = _pool_runtime_base_url(entry, _CODEX_AUX_BASE_URL) or _CODEX_AUX_BASE_URL
+    else:
+        codex_token = _read_codex_access_token()
+        if not codex_token:
+            return None, None
+        base_url = _CODEX_AUX_BASE_URL
     logger.debug("Auxiliary client: Codex OAuth (%s via Responses API)", _CODEX_AUX_MODEL)
-    real_client = OpenAI(api_key=codex_token, base_url=_CODEX_AUX_BASE_URL)
+    real_client = OpenAI(api_key=codex_token, base_url=base_url)
     return CodexAuxiliaryClient(real_client, _CODEX_AUX_MODEL), _CODEX_AUX_MODEL
 
 
@@ -650,27 +789,46 @@ def _try_anthropic() -> Tuple[Optional[Any], Optional[str]]:
     except ImportError:
         return None, None
 
-    token = resolve_anthropic_token()
+    pool_present, entry = _select_pool_entry("anthropic")
+    if pool_present:
+        if entry is None:
+            return None, None
+        token = _pool_runtime_api_key(entry)
+    else:
+        entry = None
+        token = resolve_anthropic_token()
     if not token:
         return None, None
 
-    # Allow base URL override from config.yaml model.base_url
-    base_url = _ANTHROPIC_DEFAULT_BASE_URL
+    # Allow base URL override from config.yaml model.base_url, but only
+    # when the configured provider is anthropic — otherwise a non-Anthropic
+    # base_url (e.g. Codex endpoint) would leak into Anthropic requests.
+    base_url = _pool_runtime_base_url(entry, _ANTHROPIC_DEFAULT_BASE_URL) if pool_present else _ANTHROPIC_DEFAULT_BASE_URL
     try:
         from hermes_cli.config import load_config
         cfg = load_config()
         model_cfg = cfg.get("model")
         if isinstance(model_cfg, dict):
-            cfg_base_url = (model_cfg.get("base_url") or "").strip().rstrip("/")
-            if cfg_base_url:
-                base_url = cfg_base_url
+            cfg_provider = str(model_cfg.get("provider") or "").strip().lower()
+            if cfg_provider == "anthropic":
+                cfg_base_url = (model_cfg.get("base_url") or "").strip().rstrip("/")
+                if cfg_base_url:
+                    base_url = cfg_base_url
     except Exception:
         pass
 
+    from agent.anthropic_adapter import _is_oauth_token
+    is_oauth = _is_oauth_token(token)
     model = _API_KEY_PROVIDER_AUX_MODELS.get("anthropic", "claude-haiku-4-5-20251001")
-    logger.debug("Auxiliary client: Anthropic native (%s) at %s", model, base_url)
-    real_client = build_anthropic_client(token, base_url)
-    return AnthropicAuxiliaryClient(real_client, model, token, base_url), model
+    logger.debug("Auxiliary client: Anthropic native (%s) at %s (oauth=%s)", model, base_url, is_oauth)
+    try:
+        real_client = build_anthropic_client(token, base_url)
+    except ImportError:
+        # The anthropic_adapter module imports fine but the SDK itself is
+        # missing — build_anthropic_client raises ImportError at call time
+        # when _anthropic_sdk is None.  Treat as unavailable.
+        return None, None
+    return AnthropicAuxiliaryClient(real_client, model, token, base_url, is_oauth=is_oauth), model
 
 
 def _resolve_forced_provider(forced: str) -> Tuple[Optional[OpenAI], Optional[str]]:
@@ -707,16 +865,62 @@ def _resolve_forced_provider(forced: str) -> Tuple[Optional[OpenAI], Optional[st
     return None, None
 
 
+_AUTO_PROVIDER_LABELS = {
+    "_try_openrouter": "openrouter",
+    "_try_nous": "nous",
+    "_try_custom_endpoint": "local/custom",
+    "_try_codex": "openai-codex",
+    "_resolve_api_key_provider": "api-key",
+}
+
+
+_AGGREGATOR_PROVIDERS = frozenset({"openrouter", "nous"})
+
+
 def _resolve_auto() -> Tuple[Optional[OpenAI], Optional[str]]:
-    """Full auto-detection chain: OpenRouter → Nous → custom → Codex → API-key → None."""
+    """Full auto-detection chain.
+
+    Priority:
+      1. If the user's main provider is NOT an aggregator (OpenRouter / Nous),
+         use their main provider + main model directly.  This ensures users on
+         Alibaba, DeepSeek, ZAI, etc. get auxiliary tasks handled by the same
+         provider they already have credentials for — no OpenRouter key needed.
+      2. OpenRouter → Nous → custom → Codex → API-key providers (original chain).
+    """
     global auxiliary_is_nous
     auxiliary_is_nous = False  # Reset — _try_nous() will set True if it wins
+
+    # ── Step 1: non-aggregator main provider → use main model directly ──
+    main_provider = _read_main_provider()
+    main_model = _read_main_model()
+    if (main_provider and main_model
+            and main_provider not in _AGGREGATOR_PROVIDERS
+            and main_provider not in ("auto", "custom", "")):
+        client, resolved = resolve_provider_client(main_provider, main_model)
+        if client is not None:
+            logger.info("Auxiliary auto-detect: using main provider %s (%s)",
+                        main_provider, resolved or main_model)
+            return client, resolved or main_model
+
+    # ── Step 2: aggregator / fallback chain ──────────────────────────────
+    tried = []
     for try_fn in (_try_openrouter, _try_nous, _try_custom_endpoint,
                    _try_codex, _resolve_api_key_provider):
+        fn_name = getattr(try_fn, "__name__", "unknown")
+        label = _AUTO_PROVIDER_LABELS.get(fn_name, fn_name)
         client, model = try_fn()
         if client is not None:
+            if tried:
+                logger.info("Auxiliary auto-detect: using %s (%s) — skipped: %s",
+                            label, model or "default", ", ".join(tried))
+            else:
+                logger.info("Auxiliary auto-detect: using %s (%s)", label, model or "default")
             return client, model
-    logger.debug("Auxiliary client: none available")
+        tried.append(label)
+    logger.warning("Auxiliary auto-detect: no provider available (tried: %s). "
+                   "Compression, summarization, and memory flush will not work. "
+                   "Set OPENROUTER_API_KEY or configure a local model in config.yaml.",
+                   ", ".join(tried))
     return None, None
 
 
@@ -867,11 +1071,12 @@ def resolve_provider_client(
             custom_key = (
                 (explicit_api_key or "").strip()
                 or os.getenv("OPENAI_API_KEY", "").strip()
+                or "no-key-required"  # local servers don't need auth
             )
-            if not custom_base or not custom_key:
+            if not custom_base:
                 logger.warning(
                     "resolve_provider_client: explicit custom endpoint requested "
-                    "but no API key was found (set explicit_api_key or OPENAI_API_KEY)"
+                    "but base_url is empty"
                 )
                 return None, None
             final_model = model or _read_main_model() or "gpt-4o-mini"
@@ -917,9 +1122,9 @@ def resolve_provider_client(
             tried_sources = list(pconfig.api_key_env_vars)
             if provider == "copilot":
                 tried_sources.append("gh auth token")
-            logger.warning("resolve_provider_client: provider %s has no API "
-                           "key configured (tried: %s)",
-                           provider, ", ".join(tried_sources))
+            logger.debug("resolve_provider_client: provider %s has no API "
+                         "key configured (tried: %s)",
+                         provider, ", ".join(tried_sources))
             return None, None
 
         base_url = str(creds.get("base_url", "")).strip().rstrip("/") or pconfig.inference_base_url
@@ -1107,7 +1312,13 @@ def resolve_vision_provider_client(
         return "custom", client, final_model
 
     if requested == "auto":
-        for candidate in get_available_vision_backends():
+        ordered = list(_VISION_AUTO_PROVIDER_ORDER)
+        preferred = _preferred_main_vision_provider()
+        if preferred in ordered:
+            ordered.remove(preferred)
+            ordered.insert(0, preferred)
+
+        for candidate in ordered:
             sync_client, default_model = _resolve_strict_vision_backend(candidate)
             if sync_client is not None:
                 return _finalize(candidate, sync_client, default_model)
@@ -1180,6 +1391,105 @@ _client_cache: Dict[tuple, tuple] = {}
 _client_cache_lock = threading.Lock()
 
 
+def neuter_async_httpx_del() -> None:
+    """Monkey-patch ``AsyncHttpxClientWrapper.__del__`` to be a no-op.
+
+    The OpenAI SDK's ``AsyncHttpxClientWrapper.__del__`` schedules
+    ``self.aclose()`` via ``asyncio.get_running_loop().create_task()``.
+    When an ``AsyncOpenAI`` client is garbage-collected while
+    prompt_toolkit's event loop is running (the common CLI idle state),
+    the ``aclose()`` task runs on prompt_toolkit's loop but the
+    underlying TCP transport is bound to a *different* loop (the worker
+    thread's loop that the client was originally created on).  If that
+    loop is closed or its thread is dead, the transport's
+    ``self._loop.call_soon()`` raises ``RuntimeError("Event loop is
+    closed")``, which prompt_toolkit surfaces as "Unhandled exception
+    in event loop ... Press ENTER to continue...".
+
+    Neutering ``__del__`` is safe because:
+    - Cached clients are explicitly cleaned via ``_force_close_async_httpx``
+      on stale-loop detection and ``shutdown_cached_clients`` on exit.
+    - Uncached clients' TCP connections are cleaned up by the OS when the
+      process exits.
+    - The OpenAI SDK itself marks this as a TODO (``# TODO(someday):
+      support non asyncio runtimes here``).
+
+    Call this once at CLI startup, before any ``AsyncOpenAI`` clients are
+    created.
+    """
+    try:
+        from openai._base_client import AsyncHttpxClientWrapper
+        AsyncHttpxClientWrapper.__del__ = lambda self: None  # type: ignore[assignment]
+    except (ImportError, AttributeError):
+        pass  # Graceful degradation if the SDK changes its internals
+
+
+def _force_close_async_httpx(client: Any) -> None:
+    """Mark the httpx AsyncClient inside an AsyncOpenAI client as closed.
+
+    This prevents ``AsyncHttpxClientWrapper.__del__`` from scheduling
+    ``aclose()`` on a (potentially closed) event loop, which causes
+    ``RuntimeError: Event loop is closed`` → prompt_toolkit's
+    "Press ENTER to continue..." handler.
+
+    We intentionally do NOT run the full async close path — the
+    connections will be dropped by the OS when the process exits.
+    """
+    try:
+        from httpx._client import ClientState
+        inner = getattr(client, "_client", None)
+        if inner is not None and not getattr(inner, "is_closed", True):
+            inner._state = ClientState.CLOSED
+    except Exception:
+        pass
+
+
+def shutdown_cached_clients() -> None:
+    """Close all cached clients (sync and async) to prevent event-loop errors.
+
+    Call this during CLI shutdown, *before* the event loop is closed, to
+    avoid ``AsyncHttpxClientWrapper.__del__`` raising on a dead loop.
+    """
+    import inspect
+
+    with _client_cache_lock:
+        for key, entry in list(_client_cache.items()):
+            client = entry[0]
+            if client is None:
+                continue
+            # Mark any async httpx transport as closed first (prevents __del__
+            # from scheduling aclose() on a dead event loop).
+            _force_close_async_httpx(client)
+            # Sync clients: close the httpx connection pool cleanly.
+            # Async clients: skip — we already neutered __del__ above.
+            try:
+                close_fn = getattr(client, "close", None)
+                if close_fn and not inspect.iscoroutinefunction(close_fn):
+                    close_fn()
+            except Exception:
+                pass
+        _client_cache.clear()
+
+
+def cleanup_stale_async_clients() -> None:
+    """Force-close cached async clients whose event loop is closed.
+
+    Call this after each agent turn to proactively clean up stale clients
+    before GC can trigger ``AsyncHttpxClientWrapper.__del__`` on them.
+    This is defense-in-depth — the primary fix is ``neuter_async_httpx_del``
+    which disables ``__del__`` entirely.
+    """
+    with _client_cache_lock:
+        stale_keys = []
+        for key, entry in _client_cache.items():
+            client, _default, cached_loop = entry
+            if cached_loop is not None and cached_loop.is_closed():
+                _force_close_async_httpx(client)
+                stale_keys.append(key)
+        for key in stale_keys:
+            del _client_cache[key]
+
+
 def _get_cached_client(
     provider: str,
     model: str = None,
@@ -1187,17 +1497,38 @@ def _get_cached_client(
     base_url: str = None,
     api_key: str = None,
 ) -> Tuple[Optional[Any], Optional[str]]:
-    """Get or create a cached client for the given provider."""
-    cache_key = (provider, async_mode, base_url or "", api_key or "")
+    """Get or create a cached client for the given provider.
+
+    Async clients (AsyncOpenAI) use httpx.AsyncClient internally, which
+    binds to the event loop that was current when the client was created.
+    Using such a client on a *different* loop causes deadlocks or
+    RuntimeError.  To prevent cross-loop issues (especially in gateway
+    mode where _run_async() may spawn fresh loops in worker threads), the
+    cache key for async clients includes the current event loop's identity
+    so each loop gets its own client instance.
+    """
+    # Include loop identity for async clients to prevent cross-loop reuse.
+    # httpx.AsyncClient (inside AsyncOpenAI) is bound to the loop where it
+    # was created — reusing it on a different loop causes deadlocks (#2681).
+    loop_id = 0
+    current_loop = None
+    if async_mode:
+        try:
+            import asyncio as _aio
+            current_loop = _aio.get_event_loop()
+            loop_id = id(current_loop)
+        except RuntimeError:
+            pass
+    cache_key = (provider, async_mode, base_url or "", api_key or "", loop_id)
     with _client_cache_lock:
         if cache_key in _client_cache:
             cached_client, cached_default, cached_loop = _client_cache[cache_key]
             if async_mode:
-                # Async clients are bound to the event loop that created them.
                 # A cached async client whose loop has been closed will raise
                 # "Event loop is closed" when httpx tries to clean up its
                 # transport.  Discard the stale client and create a fresh one.
                 if cached_loop is not None and cached_loop.is_closed():
+                    _force_close_async_httpx(cached_client)
                     del _client_cache[cache_key]
                 else:
                     return cached_client, model or cached_default
@@ -1214,13 +1545,7 @@ def _get_cached_client(
     if client is not None:
         # For async clients, remember which loop they were created on so we
         # can detect stale entries later.
-        bound_loop = None
-        if async_mode:
-            try:
-                import asyncio as _aio
-                bound_loop = _aio.get_event_loop()
-            except RuntimeError:
-                pass
+        bound_loop = current_loop
         with _client_cache_lock:
             if cache_key not in _client_cache:
                 _client_cache[cache_key] = (client, default_model, bound_loop)
@@ -1308,6 +1633,29 @@ def _resolve_task_provider_model(
     return "auto", resolved_model, None, None
 
 
+_DEFAULT_AUX_TIMEOUT = 30.0
+
+
+def _get_task_timeout(task: str, default: float = _DEFAULT_AUX_TIMEOUT) -> float:
+    """Read timeout from auxiliary.{task}.timeout in config, falling back to *default*."""
+    if not task:
+        return default
+    try:
+        from hermes_cli.config import load_config
+        config = load_config()
+    except ImportError:
+        return default
+    aux = config.get("auxiliary", {}) if isinstance(config, dict) else {}
+    task_config = aux.get(task, {}) if isinstance(aux, dict) else {}
+    raw = task_config.get("timeout")
+    if raw is not None:
+        try:
+            return float(raw)
+        except (ValueError, TypeError):
+            pass
+    return default
+
+
 def _build_call_kwargs(
     provider: str,
     model: str,
@@ -1365,7 +1713,7 @@ def call_llm(
     temperature: float = None,
     max_tokens: int = None,
     tools: list = None,
-    timeout: float = 30.0,
+    timeout: float = None,
     extra_body: dict = None,
 ) -> Any:
     """Centralized synchronous LLM call.
@@ -1383,7 +1731,7 @@ def call_llm(
         temperature: Sampling temperature (None = provider default).
         max_tokens: Max output tokens (handles max_tokens vs max_completion_tokens).
         tools: Tool definitions (for function calling).
-        timeout: Request timeout in seconds.
+        timeout: Request timeout in seconds (None = read from auxiliary.{task}.timeout config).
         extra_body: Additional request body fields.
 
     Returns:
@@ -1427,10 +1775,20 @@ def call_llm(
             api_key=resolved_api_key,
         )
         if client is None:
-            # Fallback: try openrouter
-            if resolved_provider != "openrouter" and not resolved_base_url:
-                logger.warning("Provider %s unavailable, falling back to openrouter",
-                               resolved_provider)
+            # When the user explicitly chose a non-OpenRouter provider but no
+            # credentials were found, fail fast instead of silently routing
+            # through OpenRouter (which causes confusing 404s).
+            _explicit = (resolved_provider or "").strip().lower()
+            if _explicit and _explicit not in ("auto", "openrouter", "custom"):
+                raise RuntimeError(
+                    f"Provider '{_explicit}' is set in config.yaml but no API key "
+                    f"was found. Set the {_explicit.upper()}_API_KEY environment "
+                    f"variable, or switch to a different provider with `hermes model`."
+                )
+            # For auto/custom, fall back to OpenRouter
+            if not resolved_base_url:
+                logger.info("Auxiliary %s: provider %s unavailable, falling back to openrouter",
+                            task or "call", resolved_provider)
                 client, final_model = _get_cached_client(
                     "openrouter", resolved_model or _OPENROUTER_MODEL)
         if client is None:
@@ -1438,10 +1796,19 @@ def call_llm(
                 f"No LLM provider configured for task={task} provider={resolved_provider}. "
                 f"Run: hermes setup")
 
+    effective_timeout = timeout if timeout is not None else _get_task_timeout(task)
+
+    # Log what we're about to do — makes auxiliary operations visible
+    _base_info = str(getattr(client, "base_url", resolved_base_url) or "")
+    if task:
+        logger.info("Auxiliary %s: using %s (%s)%s",
+                     task, resolved_provider or "auto", final_model or "default",
+                     f" at {_base_info}" if _base_info and "openrouter" not in _base_info else "")
+
     kwargs = _build_call_kwargs(
         resolved_provider, final_model, messages,
         temperature=temperature, max_tokens=max_tokens,
-        tools=tools, timeout=timeout, extra_body=extra_body,
+        tools=tools, timeout=effective_timeout, extra_body=extra_body,
         base_url=resolved_base_url)
 
     # Handle max_tokens vs max_completion_tokens retry
@@ -1456,6 +1823,62 @@ def call_llm(
         raise
 
 
+def extract_content_or_reasoning(response) -> str:
+    """Extract content from an LLM response, falling back to reasoning fields.
+
+    Mirrors the main agent loop's behavior when a reasoning model (DeepSeek-R1,
+    Qwen-QwQ, etc.) returns ``content=None`` with reasoning in structured fields.
+
+    Resolution order:
+      1. ``message.content`` — strip inline think/reasoning blocks, check for
+         remaining non-whitespace text.
+      2. ``message.reasoning`` / ``message.reasoning_content`` — direct
+         structured reasoning fields (DeepSeek, Moonshot, Novita, etc.).
+      3. ``message.reasoning_details`` — OpenRouter unified array format.
+
+    Returns the best available text, or ``""`` if nothing found.
+    """
+    import re
+
+    msg = response.choices[0].message
+    content = (msg.content or "").strip()
+
+    if content:
+        # Strip inline think/reasoning blocks (mirrors _strip_think_blocks)
+        cleaned = re.sub(
+            r"<(?:think|thinking|reasoning|REASONING_SCRATCHPAD)>"
+            r".*?"
+            r"</(?:think|thinking|reasoning|REASONING_SCRATCHPAD)>",
+            "", content, flags=re.DOTALL | re.IGNORECASE,
+        ).strip()
+        if cleaned:
+            return cleaned
+
+    # Content is empty or reasoning-only — try structured reasoning fields
+    reasoning_parts: list[str] = []
+    for field in ("reasoning", "reasoning_content"):
+        val = getattr(msg, field, None)
+        if val and isinstance(val, str) and val.strip() and val not in reasoning_parts:
+            reasoning_parts.append(val.strip())
+
+    details = getattr(msg, "reasoning_details", None)
+    if details and isinstance(details, list):
+        for detail in details:
+            if isinstance(detail, dict):
+                summary = (
+                    detail.get("summary")
+                    or detail.get("content")
+                    or detail.get("text")
+                )
+                if summary and summary not in reasoning_parts:
+                    reasoning_parts.append(summary.strip() if isinstance(summary, str) else str(summary))
+
+    if reasoning_parts:
+        return "\n\n".join(reasoning_parts)
+
+    return ""
+
+
 async def async_call_llm(
     task: str = None,
     *,
@@ -1467,7 +1890,7 @@ async def async_call_llm(
     temperature: float = None,
     max_tokens: int = None,
     tools: list = None,
-    timeout: float = 30.0,
+    timeout: float = None,
     extra_body: dict = None,
 ) -> Any:
     """Centralized asynchronous LLM call.
@@ -1510,7 +1933,14 @@ async def async_call_llm(
             api_key=resolved_api_key,
         )
         if client is None:
-            if resolved_provider != "openrouter" and not resolved_base_url:
+            _explicit = (resolved_provider or "").strip().lower()
+            if _explicit and _explicit not in ("auto", "openrouter", "custom"):
+                raise RuntimeError(
+                    f"Provider '{_explicit}' is set in config.yaml but no API key "
+                    f"was found. Set the {_explicit.upper()}_API_KEY environment "
+                    f"variable, or switch to a different provider with `hermes model`."
+                )
+            if not resolved_base_url:
                 logger.warning("Provider %s unavailable, falling back to openrouter",
                                resolved_provider)
                 client, final_model = _get_cached_client(
@@ -1521,10 +1951,12 @@ async def async_call_llm(
                 f"No LLM provider configured for task={task} provider={resolved_provider}. "
                 f"Run: hermes setup")
 
+    effective_timeout = timeout if timeout is not None else _get_task_timeout(task)
+
     kwargs = _build_call_kwargs(
         resolved_provider, final_model, messages,
         temperature=temperature, max_tokens=max_tokens,
-        tools=tools, timeout=timeout, extra_body=extra_body,
+        tools=tools, timeout=effective_timeout, extra_body=extra_body,
         base_url=resolved_base_url)
 
     try:

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__(
@@ -40,8 +64,8 @@ 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 = "",
@@ -56,7 +80,7 @@ class ContextCompressor:
         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(
@@ -66,6 +90,24 @@ class ContextCompressor:
         )
         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
@@ -74,6 +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)
@@ -96,57 +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
@@ -156,6 +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 "
@@ -280,10 +483,75 @@ Write only the summary body. Do not include any preamble or prefix; the system w
             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.
         """
@@ -297,19 +565,26 @@ Write only the summary body. Do not include any preamble or prefix; the system w
                 )
             return messages
 
-        compress_start = self.protect_first_n
-        compress_end = n_messages - self.protect_last_n
-        if compress_start >= compress_end:
-            return messages
+        display_tokens = current_tokens if current_tokens else self.last_prompt_tokens or estimate_messages_tokens_rough(messages)
 
-        # Adjust boundaries to avoid splitting tool_call/result groups.
+        # 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_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:
             logger.info(
@@ -323,15 +598,20 @@ Write only the summary body. Do not include any preamble or prefix; the system w
                 self.threshold_percent * 100,
                 self.threshold_tokens,
             )
+            tail_msgs = n_messages - compress_end
             logger.info(
-                "Summarizing turns %d-%d (%d turns)",
+                "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()

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/credential_pool.py

@@ -0,0 +1,947 @@
+"""Persistent multi-credential pool for same-provider failover."""
+
+from __future__ import annotations
+
+import logging
+import random
+import threading
+import time
+import uuid
+import os
+from dataclasses import dataclass, fields, replace
+from typing import Any, Dict, List, Optional, Set, Tuple
+
+from hermes_constants import OPENROUTER_BASE_URL
+import hermes_cli.auth as auth_mod
+from hermes_cli.auth import (
+    ACCESS_TOKEN_REFRESH_SKEW_SECONDS,
+    CODEX_ACCESS_TOKEN_REFRESH_SKEW_SECONDS,
+    DEFAULT_AGENT_KEY_MIN_TTL_SECONDS,
+    PROVIDER_REGISTRY,
+    _agent_key_is_usable,
+    _codex_access_token_is_expiring,
+    _decode_jwt_claims,
+    _is_expiring,
+    _load_auth_store,
+    _load_provider_state,
+    read_credential_pool,
+    write_credential_pool,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def _load_config_safe() -> Optional[dict]:
+    """Load config.yaml, returning None on any error."""
+    try:
+        from hermes_cli.config import load_config
+
+        return load_config()
+    except Exception:
+        return None
+
+
+# --- Status and type constants ---
+
+STATUS_OK = "ok"
+STATUS_EXHAUSTED = "exhausted"
+
+AUTH_TYPE_OAUTH = "oauth"
+AUTH_TYPE_API_KEY = "api_key"
+
+SOURCE_MANUAL = "manual"
+
+STRATEGY_FILL_FIRST = "fill_first"
+STRATEGY_ROUND_ROBIN = "round_robin"
+STRATEGY_RANDOM = "random"
+STRATEGY_LEAST_USED = "least_used"
+SUPPORTED_POOL_STRATEGIES = {
+    STRATEGY_FILL_FIRST,
+    STRATEGY_ROUND_ROBIN,
+    STRATEGY_RANDOM,
+    STRATEGY_LEAST_USED,
+}
+
+# Cooldown before retrying an exhausted credential.
+# 429 (rate-limited) cools down faster since quotas reset frequently.
+# 402 (billing/quota) and other codes use a longer default.
+EXHAUSTED_TTL_429_SECONDS = 60 * 60          # 1 hour
+EXHAUSTED_TTL_DEFAULT_SECONDS = 24 * 60 * 60 # 24 hours
+
+# Pool key prefix for custom OpenAI-compatible endpoints.
+# Custom endpoints all share provider='custom' but are keyed by their
+# custom_providers name: 'custom:<normalized_name>'.
+CUSTOM_POOL_PREFIX = "custom:"
+
+
+# Fields that are only round-tripped through JSON — never used for logic as attributes.
+_EXTRA_KEYS = frozenset({
+    "token_type", "scope", "client_id", "portal_base_url", "obtained_at",
+    "expires_in", "agent_key_id", "agent_key_expires_in", "agent_key_reused",
+    "agent_key_obtained_at", "tls",
+})
+
+
+@dataclass
+class PooledCredential:
+    provider: str
+    id: str
+    label: str
+    auth_type: str
+    priority: int
+    source: str
+    access_token: str
+    refresh_token: Optional[str] = None
+    last_status: Optional[str] = None
+    last_status_at: Optional[float] = None
+    last_error_code: Optional[int] = None
+    base_url: Optional[str] = None
+    expires_at: Optional[str] = None
+    expires_at_ms: Optional[int] = None
+    last_refresh: Optional[str] = None
+    inference_base_url: Optional[str] = None
+    agent_key: Optional[str] = None
+    agent_key_expires_at: Optional[str] = None
+    request_count: int = 0
+    extra: Dict[str, Any] = None  # type: ignore[assignment]
+
+    def __post_init__(self):
+        if self.extra is None:
+            self.extra = {}
+
+    def __getattr__(self, name: str):
+        if name in _EXTRA_KEYS:
+            return self.extra.get(name)
+        raise AttributeError(f"'{type(self).__name__}' object has no attribute {name!r}")
+
+    @classmethod
+    def from_dict(cls, provider: str, payload: Dict[str, Any]) -> "PooledCredential":
+        field_names = {f.name for f in fields(cls) if f.name != "provider"}
+        data = {k: payload.get(k) for k in field_names if k in payload}
+        extra = {k: payload[k] for k in _EXTRA_KEYS if k in payload and payload[k] is not None}
+        data["extra"] = extra
+        data.setdefault("id", uuid.uuid4().hex[:6])
+        data.setdefault("label", payload.get("source", provider))
+        data.setdefault("auth_type", AUTH_TYPE_API_KEY)
+        data.setdefault("priority", 0)
+        data.setdefault("source", SOURCE_MANUAL)
+        data.setdefault("access_token", "")
+        return cls(provider=provider, **data)
+
+    def to_dict(self) -> Dict[str, Any]:
+        _ALWAYS_EMIT = {"last_status", "last_status_at", "last_error_code"}
+        result: Dict[str, Any] = {}
+        for field_def in fields(self):
+            if field_def.name in ("provider", "extra"):
+                continue
+            value = getattr(self, field_def.name)
+            if value is not None or field_def.name in _ALWAYS_EMIT:
+                result[field_def.name] = value
+        for k, v in self.extra.items():
+            if v is not None:
+                result[k] = v
+        return result
+
+    @property
+    def runtime_api_key(self) -> str:
+        if self.provider == "nous":
+            return str(self.agent_key or self.access_token or "")
+        return str(self.access_token or "")
+
+    @property
+    def runtime_base_url(self) -> Optional[str]:
+        if self.provider == "nous":
+            return self.inference_base_url or self.base_url
+        return self.base_url
+
+
+def label_from_token(token: str, fallback: str) -> str:
+    claims = _decode_jwt_claims(token)
+    for key in ("email", "preferred_username", "upn"):
+        value = claims.get(key)
+        if isinstance(value, str) and value.strip():
+            return value.strip()
+    return fallback
+
+
+def _next_priority(entries: List[PooledCredential]) -> int:
+    return max((entry.priority for entry in entries), default=-1) + 1
+
+
+def _is_manual_source(source: str) -> bool:
+    normalized = (source or "").strip().lower()
+    return normalized == SOURCE_MANUAL or normalized.startswith(f"{SOURCE_MANUAL}:")
+
+
+def _exhausted_ttl(error_code: Optional[int]) -> int:
+    """Return cooldown seconds based on the HTTP status that caused exhaustion."""
+    if error_code == 429:
+        return EXHAUSTED_TTL_429_SECONDS
+    return EXHAUSTED_TTL_DEFAULT_SECONDS
+
+
+def _normalize_custom_pool_name(name: str) -> str:
+    """Normalize a custom provider name for use as a pool key suffix."""
+    return name.strip().lower().replace(" ", "-")
+
+
+def _iter_custom_providers(config: Optional[dict] = None):
+    """Yield (normalized_name, entry_dict) for each valid custom_providers entry."""
+    if config is None:
+        config = _load_config_safe()
+    if config is None:
+        return
+    custom_providers = config.get("custom_providers")
+    if not isinstance(custom_providers, list):
+        return
+    for entry in custom_providers:
+        if not isinstance(entry, dict):
+            continue
+        name = entry.get("name")
+        if not isinstance(name, str):
+            continue
+        yield _normalize_custom_pool_name(name), entry
+
+
+def get_custom_provider_pool_key(base_url: str) -> Optional[str]:
+    """Look up the custom_providers list in config.yaml and return 'custom:<name>' for a matching base_url.
+
+    Returns None if no match is found.
+    """
+    if not base_url:
+        return None
+    normalized_url = base_url.strip().rstrip("/")
+    for norm_name, entry in _iter_custom_providers():
+        entry_url = str(entry.get("base_url") or "").strip().rstrip("/")
+        if entry_url and entry_url == normalized_url:
+            return f"{CUSTOM_POOL_PREFIX}{norm_name}"
+    return None
+
+
+def list_custom_pool_providers() -> List[str]:
+    """Return all 'custom:*' pool keys that have entries in auth.json."""
+    pool_data = read_credential_pool(None)
+    return sorted(
+        key for key in pool_data
+        if key.startswith(CUSTOM_POOL_PREFIX)
+        and isinstance(pool_data.get(key), list)
+        and pool_data[key]
+    )
+
+
+def _get_custom_provider_config(pool_key: str) -> Optional[Dict[str, Any]]:
+    """Return the custom_providers config entry matching a pool key like 'custom:together.ai'."""
+    if not pool_key.startswith(CUSTOM_POOL_PREFIX):
+        return None
+    suffix = pool_key[len(CUSTOM_POOL_PREFIX):]
+    for norm_name, entry in _iter_custom_providers():
+        if norm_name == suffix:
+            return entry
+    return None
+
+
+def get_pool_strategy(provider: str) -> str:
+    """Return the configured selection strategy for a provider."""
+    config = _load_config_safe()
+    if config is None:
+        return STRATEGY_FILL_FIRST
+
+    strategies = config.get("credential_pool_strategies")
+    if not isinstance(strategies, dict):
+        return STRATEGY_FILL_FIRST
+
+    strategy = str(strategies.get(provider, "") or "").strip().lower()
+    if strategy in SUPPORTED_POOL_STRATEGIES:
+        return strategy
+    return STRATEGY_FILL_FIRST
+
+
+class CredentialPool:
+    def __init__(self, provider: str, entries: List[PooledCredential]):
+        self.provider = provider
+        self._entries = sorted(entries, key=lambda entry: entry.priority)
+        self._current_id: Optional[str] = None
+        self._strategy = get_pool_strategy(provider)
+        self._lock = threading.Lock()
+
+    def has_credentials(self) -> bool:
+        return bool(self._entries)
+
+    def has_available(self) -> bool:
+        """True if at least one entry is not currently in exhaustion cooldown."""
+        return bool(self._available_entries())
+
+    def entries(self) -> List[PooledCredential]:
+        return list(self._entries)
+
+    def current(self) -> Optional[PooledCredential]:
+        if not self._current_id:
+            return None
+        return next((entry for entry in self._entries if entry.id == self._current_id), None)
+
+    def _replace_entry(self, old: PooledCredential, new: PooledCredential) -> None:
+        """Swap an entry in-place by id, preserving sort order."""
+        for idx, entry in enumerate(self._entries):
+            if entry.id == old.id:
+                self._entries[idx] = new
+                return
+
+    def _persist(self) -> None:
+        write_credential_pool(
+            self.provider,
+            [entry.to_dict() for entry in self._entries],
+        )
+
+    def _mark_exhausted(self, entry: PooledCredential, status_code: Optional[int]) -> PooledCredential:
+        updated = replace(
+            entry,
+            last_status=STATUS_EXHAUSTED,
+            last_status_at=time.time(),
+            last_error_code=status_code,
+        )
+        self._replace_entry(entry, updated)
+        self._persist()
+        return updated
+
+    def _sync_anthropic_entry_from_credentials_file(self, entry: PooledCredential) -> PooledCredential:
+        """Sync a claude_code pool entry from ~/.claude/.credentials.json if tokens differ.
+
+        OAuth refresh tokens are single-use. When something external (e.g.
+        Claude Code CLI, or another profile's pool) refreshes the token, it
+        writes the new pair to ~/.claude/.credentials.json. The pool entry's
+        refresh token becomes stale. This method detects that and syncs.
+        """
+        if self.provider != "anthropic" or entry.source != "claude_code":
+            return entry
+        try:
+            from agent.anthropic_adapter import read_claude_code_credentials
+            creds = read_claude_code_credentials()
+            if not creds:
+                return entry
+            file_refresh = creds.get("refreshToken", "")
+            file_access = creds.get("accessToken", "")
+            file_expires = creds.get("expiresAt", 0)
+            # If the credentials file has a different token pair, sync it
+            if file_refresh and file_refresh != entry.refresh_token:
+                logger.debug("Pool entry %s: syncing tokens from credentials file (refresh token changed)", entry.id)
+                updated = replace(
+                    entry,
+                    access_token=file_access,
+                    refresh_token=file_refresh,
+                    expires_at_ms=file_expires,
+                    last_status=None,
+                    last_status_at=None,
+                    last_error_code=None,
+                )
+                self._replace_entry(entry, updated)
+                self._persist()
+                return updated
+        except Exception as exc:
+            logger.debug("Failed to sync from credentials file: %s", exc)
+        return entry
+
+    def _refresh_entry(self, entry: PooledCredential, *, force: bool) -> Optional[PooledCredential]:
+        if entry.auth_type != AUTH_TYPE_OAUTH or not entry.refresh_token:
+            if force:
+                self._mark_exhausted(entry, None)
+            return None
+
+        try:
+            if self.provider == "anthropic":
+                from agent.anthropic_adapter import refresh_anthropic_oauth_pure
+
+                refreshed = refresh_anthropic_oauth_pure(
+                    entry.refresh_token,
+                    use_json=entry.source.endswith("hermes_pkce"),
+                )
+                updated = replace(
+                    entry,
+                    access_token=refreshed["access_token"],
+                    refresh_token=refreshed["refresh_token"],
+                    expires_at_ms=refreshed["expires_at_ms"],
+                )
+                # Keep ~/.claude/.credentials.json in sync so that the
+                # fallback path (resolve_anthropic_token) and other profiles
+                # see the latest tokens.
+                if entry.source == "claude_code":
+                    try:
+                        from agent.anthropic_adapter import _write_claude_code_credentials
+                        _write_claude_code_credentials(
+                            refreshed["access_token"],
+                            refreshed["refresh_token"],
+                            refreshed["expires_at_ms"],
+                        )
+                    except Exception as wexc:
+                        logger.debug("Failed to write refreshed token to credentials file: %s", wexc)
+            elif self.provider == "openai-codex":
+                refreshed = auth_mod.refresh_codex_oauth_pure(
+                    entry.access_token,
+                    entry.refresh_token,
+                )
+                updated = replace(
+                    entry,
+                    access_token=refreshed["access_token"],
+                    refresh_token=refreshed["refresh_token"],
+                    last_refresh=refreshed.get("last_refresh"),
+                )
+            elif self.provider == "nous":
+                nous_state = {
+                    "access_token": entry.access_token,
+                    "refresh_token": entry.refresh_token,
+                    "client_id": entry.client_id,
+                    "portal_base_url": entry.portal_base_url,
+                    "inference_base_url": entry.inference_base_url,
+                    "token_type": entry.token_type,
+                    "scope": entry.scope,
+                    "obtained_at": entry.obtained_at,
+                    "expires_at": entry.expires_at,
+                    "agent_key": entry.agent_key,
+                    "agent_key_expires_at": entry.agent_key_expires_at,
+                    "tls": entry.tls,
+                }
+                refreshed = auth_mod.refresh_nous_oauth_from_state(
+                    nous_state,
+                    min_key_ttl_seconds=DEFAULT_AGENT_KEY_MIN_TTL_SECONDS,
+                    force_refresh=force,
+                    force_mint=force,
+                )
+                # Apply returned fields: dataclass fields via replace, extras via dict update
+                field_updates = {}
+                extra_updates = dict(entry.extra)
+                _field_names = {f.name for f in fields(entry)}
+                for k, v in refreshed.items():
+                    if k in _field_names:
+                        field_updates[k] = v
+                    elif k in _EXTRA_KEYS:
+                        extra_updates[k] = v
+                updated = replace(entry, extra=extra_updates, **field_updates)
+            else:
+                return entry
+        except Exception as exc:
+            logger.debug("Credential refresh failed for %s/%s: %s", self.provider, entry.id, exc)
+            # For anthropic claude_code entries: the refresh token may have been
+            # consumed by another process. Check if ~/.claude/.credentials.json
+            # has a newer token pair and retry once.
+            if self.provider == "anthropic" and entry.source == "claude_code":
+                synced = self._sync_anthropic_entry_from_credentials_file(entry)
+                if synced.refresh_token != entry.refresh_token:
+                    logger.debug("Retrying refresh with synced token from credentials file")
+                    try:
+                        from agent.anthropic_adapter import refresh_anthropic_oauth_pure
+                        refreshed = refresh_anthropic_oauth_pure(
+                            synced.refresh_token,
+                            use_json=synced.source.endswith("hermes_pkce"),
+                        )
+                        updated = replace(
+                            synced,
+                            access_token=refreshed["access_token"],
+                            refresh_token=refreshed["refresh_token"],
+                            expires_at_ms=refreshed["expires_at_ms"],
+                            last_status=STATUS_OK,
+                            last_status_at=None,
+                            last_error_code=None,
+                        )
+                        self._replace_entry(synced, updated)
+                        self._persist()
+                        try:
+                            from agent.anthropic_adapter import _write_claude_code_credentials
+                            _write_claude_code_credentials(
+                                refreshed["access_token"],
+                                refreshed["refresh_token"],
+                                refreshed["expires_at_ms"],
+                            )
+                        except Exception as wexc:
+                            logger.debug("Failed to write refreshed token to credentials file (retry path): %s", wexc)
+                        return updated
+                    except Exception as retry_exc:
+                        logger.debug("Retry refresh also failed: %s", retry_exc)
+                elif not self._entry_needs_refresh(synced):
+                    # Credentials file had a valid (non-expired) token — use it directly
+                    logger.debug("Credentials file has valid token, using without refresh")
+                    return synced
+            self._mark_exhausted(entry, None)
+            return None
+
+        updated = replace(updated, last_status=STATUS_OK, last_status_at=None, last_error_code=None)
+        self._replace_entry(entry, updated)
+        self._persist()
+        return updated
+
+    def _entry_needs_refresh(self, entry: PooledCredential) -> bool:
+        if entry.auth_type != AUTH_TYPE_OAUTH:
+            return False
+        if self.provider == "anthropic":
+            if entry.expires_at_ms is None:
+                return False
+            return int(entry.expires_at_ms) <= int(time.time() * 1000) + 120_000
+        if self.provider == "openai-codex":
+            return _codex_access_token_is_expiring(
+                entry.access_token,
+                CODEX_ACCESS_TOKEN_REFRESH_SKEW_SECONDS,
+            )
+        if self.provider == "nous":
+            # Nous refresh/mint can require network access and should happen when
+            # runtime credentials are actually resolved, not merely when the pool
+            # is enumerated for listing, migration, or selection.
+            return False
+        return False
+
+    def mark_used(self, entry_id: Optional[str] = None) -> None:
+        """Increment request_count for tracking. Used by least_used strategy."""
+        target_id = entry_id or self._current_id
+        if not target_id:
+            return
+        with self._lock:
+            for idx, entry in enumerate(self._entries):
+                if entry.id == target_id:
+                    self._entries[idx] = replace(entry, request_count=entry.request_count + 1)
+                    return
+
+    def select(self) -> Optional[PooledCredential]:
+        with self._lock:
+            return self._select_unlocked()
+
+    def _available_entries(self, *, clear_expired: bool = False, refresh: bool = False) -> List[PooledCredential]:
+        """Return entries not currently in exhaustion cooldown.
+
+        When *clear_expired* is True, entries whose cooldown has elapsed are
+        reset to STATUS_OK and persisted.  When *refresh* is True, entries
+        that need a token refresh are refreshed (skipped on failure).
+        """
+        now = time.time()
+        cleared_any = False
+        available: List[PooledCredential] = []
+        for entry in self._entries:
+            # For anthropic claude_code entries, sync from the credentials file
+            # before any status/refresh checks. This picks up tokens refreshed
+            # by other processes (Claude Code CLI, other Hermes profiles).
+            if (self.provider == "anthropic" and entry.source == "claude_code"
+                    and entry.last_status == STATUS_EXHAUSTED):
+                synced = self._sync_anthropic_entry_from_credentials_file(entry)
+                if synced is not entry:
+                    entry = synced
+                    cleared_any = True
+            if entry.last_status == STATUS_EXHAUSTED:
+                ttl = _exhausted_ttl(entry.last_error_code)
+                if entry.last_status_at and now - entry.last_status_at < ttl:
+                    continue
+                if clear_expired:
+                    cleared = replace(entry, last_status=STATUS_OK, last_status_at=None, last_error_code=None)
+                    self._replace_entry(entry, cleared)
+                    entry = cleared
+                    cleared_any = True
+            if refresh and self._entry_needs_refresh(entry):
+                refreshed = self._refresh_entry(entry, force=False)
+                if refreshed is None:
+                    continue
+                entry = refreshed
+            available.append(entry)
+        if cleared_any:
+            self._persist()
+        return available
+
+    def _select_unlocked(self) -> Optional[PooledCredential]:
+        available = self._available_entries(clear_expired=True, refresh=True)
+        if not available:
+            self._current_id = None
+            return None
+
+        if self._strategy == STRATEGY_RANDOM:
+            entry = random.choice(available)
+            self._current_id = entry.id
+            return entry
+
+        if self._strategy == STRATEGY_LEAST_USED and len(available) > 1:
+            entry = min(available, key=lambda e: e.request_count)
+            self._current_id = entry.id
+            return entry
+
+        if self._strategy == STRATEGY_ROUND_ROBIN and len(available) > 1:
+            entry = available[0]
+            rotated = [candidate for candidate in self._entries if candidate.id != entry.id]
+            rotated.append(replace(entry, priority=len(self._entries) - 1))
+            self._entries = [replace(candidate, priority=idx) for idx, candidate in enumerate(rotated)]
+            self._persist()
+            self._current_id = entry.id
+            return self.current() or entry
+
+        entry = available[0]
+        self._current_id = entry.id
+        return entry
+
+    def peek(self) -> Optional[PooledCredential]:
+        current = self.current()
+        if current is not None:
+            return current
+        available = self._available_entries()
+        return available[0] if available else None
+
+    def mark_exhausted_and_rotate(self, *, status_code: Optional[int]) -> Optional[PooledCredential]:
+        with self._lock:
+            entry = self.current() or self._select_unlocked()
+            if entry is None:
+                return None
+            self._mark_exhausted(entry, status_code)
+            self._current_id = None
+            return self._select_unlocked()
+
+    def try_refresh_current(self) -> Optional[PooledCredential]:
+        with self._lock:
+            return self._try_refresh_current_unlocked()
+
+    def _try_refresh_current_unlocked(self) -> Optional[PooledCredential]:
+        entry = self.current()
+        if entry is None:
+            return None
+        refreshed = self._refresh_entry(entry, force=True)
+        if refreshed is not None:
+            self._current_id = refreshed.id
+        return refreshed
+
+    def reset_statuses(self) -> int:
+        count = 0
+        new_entries = []
+        for entry in self._entries:
+            if entry.last_status or entry.last_status_at or entry.last_error_code:
+                new_entries.append(replace(entry, last_status=None, last_status_at=None, last_error_code=None))
+                count += 1
+            else:
+                new_entries.append(entry)
+        if count:
+            self._entries = new_entries
+            self._persist()
+        return count
+
+    def remove_index(self, index: int) -> Optional[PooledCredential]:
+        if index < 1 or index > len(self._entries):
+            return None
+        removed = self._entries.pop(index - 1)
+        self._entries = [
+            replace(entry, priority=new_priority)
+            for new_priority, entry in enumerate(self._entries)
+        ]
+        self._persist()
+        if self._current_id == removed.id:
+            self._current_id = None
+        return removed
+
+    def add_entry(self, entry: PooledCredential) -> PooledCredential:
+        entry = replace(entry, priority=_next_priority(self._entries))
+        self._entries.append(entry)
+        self._persist()
+        return entry
+
+
+def _upsert_entry(entries: List[PooledCredential], provider: str, source: str, payload: Dict[str, Any]) -> bool:
+    existing_idx = None
+    for idx, entry in enumerate(entries):
+        if entry.source == source:
+            existing_idx = idx
+            break
+
+    if existing_idx is None:
+        payload.setdefault("id", uuid.uuid4().hex[:6])
+        payload.setdefault("priority", _next_priority(entries))
+        payload.setdefault("label", payload.get("label") or source)
+        entries.append(PooledCredential.from_dict(provider, payload))
+        return True
+
+    existing = entries[existing_idx]
+    field_updates = {}
+    extra_updates = {}
+    _field_names = {f.name for f in fields(existing)}
+    for key, value in payload.items():
+        if key in {"id", "priority"} or value is None:
+            continue
+        if key == "label" and existing.label:
+            continue
+        if key in _field_names:
+            if getattr(existing, key) != value:
+                field_updates[key] = value
+        elif key in _EXTRA_KEYS:
+            if existing.extra.get(key) != value:
+                extra_updates[key] = value
+    if field_updates or extra_updates:
+        if extra_updates:
+            field_updates["extra"] = {**existing.extra, **extra_updates}
+        entries[existing_idx] = replace(existing, **field_updates)
+        return True
+    return False
+
+
+def _normalize_pool_priorities(provider: str, entries: List[PooledCredential]) -> bool:
+    if provider != "anthropic":
+        return False
+
+    source_rank = {
+        "env:ANTHROPIC_TOKEN": 0,
+        "env:CLAUDE_CODE_OAUTH_TOKEN": 1,
+        "hermes_pkce": 2,
+        "claude_code": 3,
+        "env:ANTHROPIC_API_KEY": 4,
+    }
+    manual_entries = sorted(
+        (entry for entry in entries if _is_manual_source(entry.source)),
+        key=lambda entry: entry.priority,
+    )
+    seeded_entries = sorted(
+        (entry for entry in entries if not _is_manual_source(entry.source)),
+        key=lambda entry: (
+            source_rank.get(entry.source, len(source_rank)),
+            entry.priority,
+            entry.label,
+        ),
+    )
+
+    ordered = [*manual_entries, *seeded_entries]
+    id_to_idx = {entry.id: idx for idx, entry in enumerate(entries)}
+    changed = False
+    for new_priority, entry in enumerate(ordered):
+        if entry.priority != new_priority:
+            entries[id_to_idx[entry.id]] = replace(entry, priority=new_priority)
+            changed = True
+    return changed
+
+
+def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tuple[bool, Set[str]]:
+    changed = False
+    active_sources: Set[str] = set()
+    auth_store = _load_auth_store()
+
+    if provider == "anthropic":
+        from agent.anthropic_adapter import read_claude_code_credentials, read_hermes_oauth_credentials
+
+        for source_name, creds in (
+            ("hermes_pkce", read_hermes_oauth_credentials()),
+            ("claude_code", read_claude_code_credentials()),
+        ):
+            if creds and creds.get("accessToken"):
+                active_sources.add(source_name)
+                changed |= _upsert_entry(
+                    entries,
+                    provider,
+                    source_name,
+                    {
+                        "source": source_name,
+                        "auth_type": AUTH_TYPE_OAUTH,
+                        "access_token": creds.get("accessToken", ""),
+                        "refresh_token": creds.get("refreshToken"),
+                        "expires_at_ms": creds.get("expiresAt"),
+                        "label": label_from_token(creds.get("accessToken", ""), source_name),
+                    },
+                )
+
+    elif provider == "nous":
+        state = _load_provider_state(auth_store, "nous")
+        if state:
+            active_sources.add("device_code")
+            changed |= _upsert_entry(
+                entries,
+                provider,
+                "device_code",
+                {
+                    "source": "device_code",
+                    "auth_type": AUTH_TYPE_OAUTH,
+                    "access_token": state.get("access_token", ""),
+                    "refresh_token": state.get("refresh_token"),
+                    "expires_at": state.get("expires_at"),
+                    "token_type": state.get("token_type"),
+                    "scope": state.get("scope"),
+                    "client_id": state.get("client_id"),
+                    "portal_base_url": state.get("portal_base_url"),
+                    "inference_base_url": state.get("inference_base_url"),
+                    "agent_key": state.get("agent_key"),
+                    "agent_key_expires_at": state.get("agent_key_expires_at"),
+                    "tls": state.get("tls") if isinstance(state.get("tls"), dict) else None,
+                    "label": label_from_token(state.get("access_token", ""), "device_code"),
+                },
+            )
+
+    elif provider == "openai-codex":
+        state = _load_provider_state(auth_store, "openai-codex")
+        tokens = state.get("tokens") if isinstance(state, dict) else None
+        if isinstance(tokens, dict) and tokens.get("access_token"):
+            active_sources.add("device_code")
+            changed |= _upsert_entry(
+                entries,
+                provider,
+                "device_code",
+                {
+                    "source": "device_code",
+                    "auth_type": AUTH_TYPE_OAUTH,
+                    "access_token": tokens.get("access_token", ""),
+                    "refresh_token": tokens.get("refresh_token"),
+                    "base_url": "https://chatgpt.com/backend-api/codex",
+                    "last_refresh": state.get("last_refresh"),
+                    "label": label_from_token(tokens.get("access_token", ""), "device_code"),
+                },
+            )
+
+    return changed, active_sources
+
+
+def _seed_from_env(provider: str, entries: List[PooledCredential]) -> Tuple[bool, Set[str]]:
+    changed = False
+    active_sources: Set[str] = set()
+    if provider == "openrouter":
+        token = os.getenv("OPENROUTER_API_KEY", "").strip()
+        if token:
+            source = "env:OPENROUTER_API_KEY"
+            active_sources.add(source)
+            changed |= _upsert_entry(
+                entries,
+                provider,
+                source,
+                {
+                    "source": source,
+                    "auth_type": AUTH_TYPE_API_KEY,
+                    "access_token": token,
+                    "base_url": OPENROUTER_BASE_URL,
+                    "label": "OPENROUTER_API_KEY",
+                },
+            )
+        return changed, active_sources
+
+    pconfig = PROVIDER_REGISTRY.get(provider)
+    if not pconfig or pconfig.auth_type != AUTH_TYPE_API_KEY:
+        return changed, active_sources
+
+    env_url = ""
+    if pconfig.base_url_env_var:
+        env_url = os.getenv(pconfig.base_url_env_var, "").strip().rstrip("/")
+
+    env_vars = list(pconfig.api_key_env_vars)
+    if provider == "anthropic":
+        env_vars = [
+            "ANTHROPIC_TOKEN",
+            "CLAUDE_CODE_OAUTH_TOKEN",
+            "ANTHROPIC_API_KEY",
+        ]
+
+    for env_var in env_vars:
+        token = os.getenv(env_var, "").strip()
+        if not token:
+            continue
+        source = f"env:{env_var}"
+        active_sources.add(source)
+        auth_type = AUTH_TYPE_OAUTH if provider == "anthropic" and not token.startswith("sk-ant-api") else AUTH_TYPE_API_KEY
+        base_url = env_url or pconfig.inference_base_url
+        changed |= _upsert_entry(
+            entries,
+            provider,
+            source,
+            {
+                "source": source,
+                "auth_type": auth_type,
+                "access_token": token,
+                "base_url": base_url,
+                "label": env_var,
+            },
+        )
+    return changed, active_sources
+
+
+def _prune_stale_seeded_entries(entries: List[PooledCredential], active_sources: Set[str]) -> bool:
+    retained = [
+        entry
+        for entry in entries
+        if _is_manual_source(entry.source)
+        or entry.source in active_sources
+        or not (
+            entry.source.startswith("env:")
+            or entry.source in {"claude_code", "hermes_pkce"}
+        )
+    ]
+    if len(retained) == len(entries):
+        return False
+    entries[:] = retained
+    return True
+
+
+def _seed_custom_pool(pool_key: str, entries: List[PooledCredential]) -> Tuple[bool, Set[str]]:
+    """Seed a custom endpoint pool from custom_providers config and model config."""
+    changed = False
+    active_sources: Set[str] = set()
+
+    # Seed from the custom_providers config entry's api_key field
+    cp_config = _get_custom_provider_config(pool_key)
+    if cp_config:
+        api_key = str(cp_config.get("api_key") or "").strip()
+        base_url = str(cp_config.get("base_url") or "").strip().rstrip("/")
+        name = str(cp_config.get("name") or "").strip()
+        if api_key:
+            source = f"config:{name}"
+            active_sources.add(source)
+            changed |= _upsert_entry(
+                entries,
+                pool_key,
+                source,
+                {
+                    "source": source,
+                    "auth_type": AUTH_TYPE_API_KEY,
+                    "access_token": api_key,
+                    "base_url": base_url,
+                    "label": name or source,
+                },
+            )
+
+    # Seed from model.api_key if model.provider=='custom' and model.base_url matches
+    try:
+        config = _load_config_safe()
+        model_cfg = config.get("model") if config else None
+        if isinstance(model_cfg, dict):
+            model_provider = str(model_cfg.get("provider") or "").strip().lower()
+            model_base_url = str(model_cfg.get("base_url") or "").strip().rstrip("/")
+            model_api_key = ""
+            for k in ("api_key", "api"):
+                v = model_cfg.get(k)
+                if isinstance(v, str) and v.strip():
+                    model_api_key = v.strip()
+                    break
+            if model_provider == "custom" and model_base_url and model_api_key:
+                # Check if this model's base_url matches our custom provider
+                matched_key = get_custom_provider_pool_key(model_base_url)
+                if matched_key == pool_key:
+                    source = "model_config"
+                    active_sources.add(source)
+                    changed |= _upsert_entry(
+                        entries,
+                        pool_key,
+                        source,
+                        {
+                            "source": source,
+                            "auth_type": AUTH_TYPE_API_KEY,
+                            "access_token": model_api_key,
+                            "base_url": model_base_url,
+                            "label": "model_config",
+                        },
+                    )
+    except Exception:
+        pass
+
+    return changed, active_sources
+
+
+def load_pool(provider: str) -> CredentialPool:
+    provider = (provider or "").strip().lower()
+    raw_entries = read_credential_pool(provider)
+    entries = [PooledCredential.from_dict(provider, payload) for payload in raw_entries]
+
+    if provider.startswith(CUSTOM_POOL_PREFIX):
+        # Custom endpoint pool — seed from custom_providers config and model config
+        custom_changed, custom_sources = _seed_custom_pool(provider, entries)
+        changed = custom_changed
+        changed |= _prune_stale_seeded_entries(entries, custom_sources)
+    else:
+        singleton_changed, singleton_sources = _seed_from_singletons(provider, entries)
+        env_changed, env_sources = _seed_from_env(provider, entries)
+        changed = singleton_changed or env_changed
+        changed |= _prune_stale_seeded_entries(entries, singleton_sources | env_sources)
+        changed |= _normalize_pool_priorities(provider, entries)
+
+    if changed:
+        write_credential_pool(
+            provider,
+            [entry.to_dict() for entry in sorted(entries, key=lambda item: item.priority)],
+        )
+    return CredentialPool(provider, entries)

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)
@@ -94,8 +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 = {
@@ -190,11 +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
 # =========================================================================
@@ -231,7 +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
@@ -239,13 +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:
@@ -253,16 +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 hasattr(self._out, 'isatty') or not self._out.isatty():
+        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 []
@@ -279,18 +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)
@@ -329,7 +701,7 @@ class KawaiiSpinner:
         if self.thread:
             self.thread.join(timeout=0.5)
 
-        is_tty = hasattr(self._out, 'isatty') and self._out.isatty()
+        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.
@@ -448,10 +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:
@@ -657,35 +1033,25 @@ def format_context_pressure(
     The bar and percentage show progress toward the compaction threshold,
     NOT the raw context window.  100% = compaction fires.
 
-    Uses ANSI colors:
-      - cyan at ~60% to compaction = informational
-      - bold yellow at ~85% to compaction = warning
-
     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 = int(compaction_progress * 100)
+    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)
 
-    # Tier styling
-    if compaction_progress >= 0.85:
-        color = f"{_BOLD}{_YELLOW}"
-        icon = "⚠"
-        if compression_enabled:
-            hint = "compaction imminent"
-        else:
-            hint = "no auto-compaction"
+    color = f"{_BOLD}{_YELLOW}"
+    icon = "⚠"
+    if compression_enabled:
+        hint = "compaction approaching"
     else:
-        color = _CYAN
-        icon = "◐"
-        hint = "approaching compaction"
+        hint = "no auto-compaction"
 
     return (
         f"  {color}{icon} context {bar} {pct_int}% to compaction{_ANSI_RESET}"
@@ -703,20 +1069,16 @@ def format_context_pressure_gateway(
     No ANSI — just Unicode and plain text suitable for Telegram/Discord/etc.
     The percentage shows progress toward the compaction threshold.
     """
-    pct_int = int(compaction_progress * 100)
+    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)
 
-    if compaction_progress >= 0.85:
-        icon = "⚠️"
-        if compression_enabled:
-            hint = f"Context compaction is imminent (threshold: {threshold_pct_int}% of window)."
-        else:
-            hint = "Auto-compaction is disabled — context may be truncated."
+    icon = "⚠️"
+    if compression_enabled:
+        hint = f"Context compaction approaching (threshold: {threshold_pct_int}% of window)."
     else:
-        icon = "ℹ️"
-        hint = f"Compaction threshold is at {threshold_pct_int}% of context window."
+        hint = "Auto-compaction is disabled — context may be truncated."
 
     return f"{icon} Context: {bar} {pct_int}% to compaction\n{hint}"

agent/insights.py

@@ -644,6 +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 += " *"
@@ -666,7 +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
@@ -746,7 +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

@@ -113,6 +113,19 @@ DEFAULT_CONTEXT_LENGTHS = {
     "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 = (
@@ -162,8 +175,12 @@ _URL_TO_PROVIDER: Dict[str, str] = {
     "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",
 }
 
 
@@ -260,9 +277,11 @@ def detect_local_server_type(base_url: str) -> Optional[str]:
                         pass
             except Exception:
                 pass
-            # llama.cpp exposes /props
+            # llama.cpp exposes /v1/props (older builds used /props without the /v1 prefix)
             try:
-                r = client.get(f"{server_url}/props")
+                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:
@@ -455,8 +474,11 @@ def fetch_endpoint_model_metadata(
             )
             if is_llamacpp:
                 try:
-                    props_url = candidate.rstrip("/").replace("/v1", "") + "/props"
-                    props_resp = requests.get(props_url, headers=headers, timeout=5)
+                    # 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", {})
@@ -783,8 +805,12 @@ def get_model_context_length(
         if cached is not None:
             return cached
 
-    # 2. Active endpoint metadata for explicit custom routes
-    if _is_custom_endpoint(base_url):
+    # 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:
@@ -855,10 +881,11 @@ def get_model_context_length(
     # 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:
+        if default_model in model_lower:
             return length
 
     # 9. Query local server as last resort
@@ -883,3 +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

@@ -15,6 +15,8 @@ import time
 from pathlib import Path
 from typing import Any, Dict, Optional
 
+from utils import atomic_json_write
+
 import requests
 
 logger = logging.getLogger(__name__)
@@ -41,6 +43,7 @@ PROVIDER_TO_MODELS_DEV: Dict[str, str] = {
     "opencode-zen": "opencode",
     "opencode-go": "opencode-go",
     "kilocode": "kilo",
+    "fireworks": "fireworks-ai",
 }
 
 
@@ -64,12 +67,10 @@ def _load_disk_cache() -> Dict[str, Any]:
 
 
 def _save_disk_cache(data: Dict[str, Any]) -> None:
-    """Save models.dev data to disk cache."""
+    """Save models.dev data to disk cache atomically."""
     try:
         cache_path = _get_cache_path()
-        cache_path.parent.mkdir(parents=True, exist_ok=True)
-        with open(cache_path, "w", encoding="utf-8") as f:
-            json.dump(data, f, separators=(",", ":"))
+        atomic_json_write(cache_path, data, indent=None, separators=(",", ":"))
     except Exception as e:
         logger.debug("Failed to save models.dev disk cache: %s", e)
 

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__)
 
 # ---------------------------------------------------------------------------
@@ -154,6 +170,54 @@ SKILLS_GUIDANCE = (
     "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. "
@@ -228,6 +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
 # =========================================================================
@@ -239,22 +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, {}, ""
@@ -263,16 +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 {}
@@ -315,102 +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.
-    """
-    hermes_home = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
-    skills_dir = hermes_home / "skills"
+    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
 
-    if not skills_dir.exists():
+    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 = get_hermes_home()
+    skills_dir = hermes_home / "skills"
+    external_dirs = get_all_skills_dirs()[1:]  # skip local (index 0)
+
+    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"
-    # Load disabled skill names once for the entire scan
-    try:
-        from tools.skills_tool import _get_disabled_skill_names
-        disabled = _get_disabled_skill_names()
-    except Exception:
-        disabled = set()
+    # ── 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, frontmatter, desc = _parse_skill_file(skill_file)
-        if not is_compatible:
-            continue
-        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
-        # Respect user's disabled skills config
-        fm_name = frontmatter.get("name", skill_name)
-        if fm_name in disabled or skill_name in disabled:
-            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
-        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"
-        "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."
+    # ── 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)
 
 
 # =========================================================================
@@ -442,7 +778,7 @@ def load_soul_md() -> Optional[str]:
     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"
+    soul_path = get_hermes_home() / "SOUL.md"
     if not soul_path.exists():
         return None
     try:
@@ -457,54 +793,63 @@ def load_soul_md() -> Optional[str]:
         return None
 
 
-def build_context_files_prompt(cwd: Optional[str] = None, skip_soul: bool = False) -> str:
-    """Discover and load context files for the system prompt.
+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 ""
 
-    Discovery: AGENTS.md (recursive), .cursorrules / .cursor/rules/*.mdc,
-    and SOUL.md from HERMES_HOME only. Each 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 = []
-
-    # AGENTS.md (hierarchical, recursive)
-    top_level_agents = None
+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():
@@ -528,31 +873,41 @@ def build_context_files_prompt(cwd: Optional[str] = None, skip_soul: bool = Fals
             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")
 
-    # .hermes.md / HERMES.md — per-project agent config (walk to git root)
-    hermes_md_content = ""
-    hermes_md_path = _find_hermes_md(cwd_path)
-    if hermes_md_path:
-        try:
-            content = hermes_md_path.read_text(encoding="utf-8").strip()
-            if content:
-                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)
-                hermes_md_content = f"## {rel}\n\n{content}"
-        except Exception as e:
-            logger.debug("Could not read %s: %s", hermes_md_path, e)
 
-    if hermes_md_content:
-        hermes_md_content = _truncate_content(hermes_md_content, ".hermes.md")
-        sections.append(hermes_md_content)
+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:

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

@@ -128,7 +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:
@@ -158,38 +162,49 @@ def scan_skill_commands() -> Dict[str, Dict[str, Any]]:
     _skill_commands = {}
     try:
         from tools.skills_tool import SKILLS_DIR, _parse_frontmatter, skill_matches_platform, _get_disabled_skill_names
-        if not SKILLS_DIR.exists():
-            return _skill_commands
+        from agent.skill_utils import get_external_skills_dirs
         disabled = _get_disabled_skill_names()
-        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):
+        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
-                name = frontmatter.get('name', skill_md.parent.name)
-                # Respect user's disabled skills config
-                if name in disabled:
+                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
-                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

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

@@ -6,6 +6,8 @@ import os
 import re
 from typing import Any, Dict, Optional
 
+from utils import is_truthy_value
+
 _COMPLEX_KEYWORDS = {
     "debug",
     "debugging",
@@ -47,13 +49,7 @@ _URL_RE = re.compile(r"https?://|www\.", re.IGNORECASE)
 
 
 def _coerce_bool(value: Any, default: bool = False) -> bool:
-    if value is None:
-        return default
-    if isinstance(value, bool):
-        return value
-    if isinstance(value, str):
-        return value.strip().lower() in {"1", "true", "yes", "on"}
-    return bool(value)
+    return is_truthy_value(value, default=default)
 
 
 def _coerce_int(value: Any, default: int) -> int:
@@ -127,6 +123,7 @@ def resolve_turn_route(user_message: str, routing_config: Optional[Dict[str, Any
                 "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": (
@@ -162,6 +159,7 @@ def resolve_turn_route(user_message: str, routing_config: Optional[Dict[str, Any
                 "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": (

agent/title_generator.py

@@ -19,7 +19,7 @@ _TITLE_PROMPT = (
 )
 
 
-def generate_title(user_message: str, assistant_response: str, timeout: float = 15.0) -> Optional[str]:
+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).

agent/usage_pricing.py

@@ -649,7 +649,8 @@ def format_token_count_compact(value: int) -> str:
                 text = f"{scaled:.1f}"
             else:
                 text = f"{scaled:.0f}"
-            text = text.rstrip("0").rstrip(".")
+            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 +7,33 @@
 # =============================================================================
 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.
+  #
   # Can also be overridden with --provider flag or HERMES_INFERENCE_PROVIDER env var.
   provider: "auto"
   
@@ -232,19 +248,34 @@ 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
@@ -293,6 +324,9 @@ 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:
@@ -386,6 +420,15 @@ 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
 # =============================================================================
@@ -496,7 +539,7 @@ 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)
 #
@@ -525,7 +568,7 @@ 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)
 #
@@ -673,6 +716,12 @@ 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.
@@ -696,8 +745,8 @@ display:
   # 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.
-  # Disabled by default — enable to try the streaming UX.
-  streaming: false
+  # Stream tokens to the terminal in real-time. Disable to wait for full responses.
+  streaming: true
 
   # ───────────────────────────────────────────────────────────────────────────
   # Skin / Theme

cron/jobs.py

@@ -14,6 +14,7 @@ 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__)
@@ -30,7 +31,7 @@ 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"
@@ -248,6 +249,38 @@ def _recoverable_oneshot_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.
@@ -294,7 +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 []
 
 
@@ -329,6 +375,7 @@ def create_job(
     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.
@@ -345,12 +392,19 @@ def create_job(
         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
@@ -369,6 +423,8 @@ def create_job(
     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 = {
@@ -380,6 +436,7 @@ def create_job(
         "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": {
@@ -539,7 +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)
@@ -561,6 +618,34 @@ 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.
 
@@ -610,16 +695,18 @@ def get_due_jobs() -> List[Dict[str, Any]]:
             # 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.
-            if kind in ("cron", "interval") and (now - next_run_dt).total_seconds() > 120:
-                # More than 2 minutes late — this is a missed run, not a current one.
-                # Recompute next_run_at to the next future occurrence.
+            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). "
+                        "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

cron/scheduler.py

@@ -9,9 +9,11 @@ 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
 
@@ -24,8 +26,9 @@ 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
@@ -35,7 +38,7 @@ logger = logging.getLogger(__name__)
 # 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
@@ -43,7 +46,7 @@ from cron.jobs import get_due_jobs, mark_job_run, save_job_output
 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"
@@ -80,11 +83,32 @@ def _resolve_delivery_target(job: dict) -> Optional[dict]:
         }
 
     if ":" in deliver:
-        platform_name, chat_id = deliver.split(":", 1)
+        platform_name, rest = deliver.split(":", 1)
+        # Check for thread_id suffix (e.g. "telegram:-1003724596514:17")
+        if ":" in rest:
+            chat_id, thread_id = rest.split(":", 1)
+        else:
+            chat_id, thread_id = rest, None
+
+        # Resolve human-friendly labels like "Alice (dm)" to real IDs.
+        # send_message(action="list") shows labels with display suffixes
+        # that aren't valid platform IDs (e.g. WhatsApp JIDs).
+        try:
+            from gateway.channel_directory import resolve_channel_name
+            target = chat_id
+            # Strip display suffix like " (dm)" or " (group)"
+            if target.endswith(")") and " (" in target:
+                target = target.rsplit(" (", 1)[0].strip()
+            resolved = resolve_channel_name(platform_name.lower(), target)
+            if resolved:
+                chat_id = resolved
+        except Exception:
+            pass
+
         return {
             "platform": platform_name,
             "chat_id": chat_id,
-            "thread_id": None,
+            "thread_id": thread_id,
         }
 
     platform_name = deliver
@@ -140,6 +164,8 @@ def _deliver_result(job: dict, content: str) -> None:
         "mattermost": Platform.MATTERMOST,
         "homeassistant": Platform.HOMEASSISTANT,
         "dingtalk": Platform.DINGTALK,
+        "feishu": Platform.FEISHU,
+        "wecom": Platform.WECOM,
         "email": Platform.EMAIL,
         "sms": Platform.SMS,
     }
@@ -159,15 +185,40 @@ 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
+
+    # 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)
@@ -177,12 +228,58 @@ 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
-        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)
+
+
+_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/
+        path = get_hermes_home() / "scripts" / path
+
+    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)
+
+        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:
@@ -190,14 +287,41 @@ def _build_job_prompt(job: dict) -> str:
     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 nothing new or noteworthy to report, respond "
-        "with exactly \"[SILENT]\" (optionally followed by a brief internal "
-        "note). This suppresses delivery to the user while still saving "
-        "output locally. Only use [SILENT] when there are genuinely no "
-        "changes worth reporting.]\n\n"
+        "[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:
@@ -267,6 +391,7 @@ 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])
@@ -294,7 +419,7 @@ 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 = job.get("model") or 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 = {}
@@ -314,16 +439,11 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
             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
@@ -397,16 +517,40 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
             provider_sort=pr.get("sort"),
             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}
 
@@ -420,7 +564,7 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
 
 ## Response
 
-{final_response}
+{logged_response}
 """
         
         logger.info("Job '%s' completed successfully", job_name)
@@ -462,9 +606,13 @@ 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)
 
 
@@ -510,6 +658,12 @@ 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)

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"],
+    },
+  },
 }
 ```
 

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/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
@@ -90,7 +89,7 @@ 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 +118,6 @@ 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:

gateway/config.py

@@ -17,6 +17,7 @@ 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__)
 
@@ -25,11 +26,14 @@ 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, bool):
-        return value
     if isinstance(value, str):
-        return value.strip().lower() in ("true", "1", "yes", "on")
-    return bool(value)
+        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:
@@ -57,6 +61,8 @@ class Platform(Enum):
     DINGTALK = "dingtalk"
     API_SERVER = "api_server"
     WEBHOOK = "webhook"
+    FEISHU = "feishu"
+    WECOM = "wecom"
 
 
 @dataclass
@@ -101,12 +107,16 @@ 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
@@ -115,10 +125,14 @@ class SessionResetPolicy:
         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=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"),
         )
 
 
@@ -130,6 +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)
     
@@ -137,6 +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
@@ -157,6 +178,7 @@ 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", {}),
         )
 
@@ -258,6 +280,12 @@ class GatewayConfig:
             # 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]:
@@ -455,11 +483,27 @@ def load_gateway_config() -> GatewayConfig:
                     "pair",
                 )
 
-            # Bridge per-platform settings from config.yaml into gw_data
+            # 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
@@ -475,6 +519,10 @@ def load_gateway_config() -> GatewayConfig:
                     )
                 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, {})
@@ -499,8 +547,55 @@ 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)
 
@@ -557,6 +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(
@@ -595,14 +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")
@@ -616,13 +725,13 @@ 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")
@@ -635,13 +744,13 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
         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:
-            config.platforms[Platform.MATTERMOST].home_channel = HomeChannel(
-                platform=Platform.MATTERMOST,
-                chat_id=mattermost_home,
-                name=os.getenv("MATTERMOST_HOME_CHANNEL_NAME", "Home"),
-            )
+    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")
@@ -663,13 +772,13 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             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:
-            config.platforms[Platform.MATRIX].home_channel = HomeChannel(
-                platform=Platform.MATRIX,
-                chat_id=matrix_home,
-                name=os.getenv("MATRIX_HOME_ROOM_NAME", "Home"),
-            )
+    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")
@@ -696,13 +805,13 @@ 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")
@@ -711,17 +820,18 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             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:
-            config.platforms[Platform.SMS].home_channel = HomeChannel(
-                platform=Platform.SMS,
-                chat_id=sms_home,
-                name=os.getenv("SMS_HOME_CHANNEL_NAME", "Home"),
-            )
+    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:
@@ -730,6 +840,10 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
         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)
@@ -754,6 +868,55 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
         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
     idle_minutes = os.getenv("SESSION_IDLE_MINUTES")
     if idle_minutes:
@@ -768,6 +931,3 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             config.default_reset_policy.at_hour = int(reset_hour)
         except ValueError:
             pass
-
-
-

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

@@ -21,8 +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
@@ -53,14 +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",
@@ -296,6 +341,9 @@ class MessageEvent:
     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)
     
@@ -309,7 +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."""
@@ -326,6 +377,24 @@ class SendResult:
     message_id: Optional[str] = None
     error: Optional[str] = None
     raw_response: Any = None
+    retryable: bool = False  # True for transient errors (network, timeout) — base will retry automatically
+
+
+# Error substrings that indicate a transient network failure worth retrying
+_RETRYABLE_ERROR_PATTERNS = (
+    "connecterror",
+    "connectionerror",
+    "connectionreset",
+    "connectionrefused",
+    "timeout",
+    "timed out",
+    "network",
+    "broken pipe",
+    "remotedisconnected",
+    "eoferror",
+    "readtimeout",
+    "writetimeout",
+)
 
 
 # Type for message handlers
@@ -504,6 +573,14 @@ 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,
@@ -713,7 +790,7 @@ 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()
@@ -811,7 +888,122 @@ 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)
+
+    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)
+
+        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.
@@ -830,11 +1022,37 @@ class BasePlatformAdapter(ABC):
         
         # Check if there's already an active handler for this session
         if session_key in self._active_sessions:
+            # /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
+
             # 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:
-                print(f"[{self.name}] 🖼️ Queuing photo follow-up for session {session_key} without interrupt")
+                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)
@@ -849,12 +1067,19 @@ class BasePlatformAdapter(ABC):
                 return  # Don't interrupt now - will run after current task completes
 
             # Default behavior for non-photo follow-ups: interrupt the running agent
-            print(f"[{self.name}] ⚡ New message while session {session_key} is active - triggering interrupt")
+            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
         task = asyncio.create_task(self._process_message_background(event, session_key))
         try:
@@ -889,8 +1114,22 @@ 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)
@@ -898,12 +1137,17 @@ 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)
@@ -961,25 +1205,13 @@ 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()
@@ -1048,9 +1280,9 @@ 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:
@@ -1079,10 +1311,14 @@ class BasePlatformAdapter(ABC):
                     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]
@@ -1095,10 +1331,12 @@ 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__
@@ -1122,6 +1360,13 @@ 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]

gateway/platforms/discord.py

@@ -20,7 +20,7 @@ import threading
 import time
 from collections import defaultdict
 from pathlib import Path
-from typing import Callable, Dict, List, Optional, Any
+from typing import Callable, Dict, Optional, Any
 
 logger = logging.getLogger(__name__)
 
@@ -43,6 +43,8 @@ from pathlib import Path as _Path
 sys.path.insert(0, str(_Path(__file__).resolve().parents[2]))
 
 from gateway.config import Platform, PlatformConfig
+import re
+
 from gateway.platforms.base import (
     BasePlatformAdapter,
     MessageEvent,
@@ -50,6 +52,8 @@ from gateway.platforms.base import (
     SendResult,
     cache_image_from_url,
     cache_audio_from_url,
+    cache_document_from_bytes,
+    SUPPORTED_DOCUMENT_TYPES,
 )
 
 
@@ -404,7 +408,7 @@ class VoiceReceiver:
 class DiscordAdapter(BasePlatformAdapter):
     """
     Discord bot adapter.
-    
+
     Handles:
     - Receiving messages from servers and DMs
     - Sending responses with Discord markdown
@@ -414,10 +418,10 @@ class DiscordAdapter(BasePlatformAdapter):
     - Auto-threading for long conversations
     - Reaction-based feedback
     """
-    
+
     # Discord message limits
     MAX_MESSAGE_LENGTH = 2000
-    
+
     # Auto-disconnect from voice channel after this many seconds of inactivity
     VOICE_TIMEOUT = 300
 
@@ -439,9 +443,18 @@ class DiscordAdapter(BasePlatformAdapter):
         # in those threads don't require @mention.  Persisted to disk so the
         # set survives gateway restarts.
         self._bot_participated_threads: set = self._load_participated_threads()
+        # Persistent typing indicator loops per channel (DMs don't reliably
+        # show the standard typing gateway event for bots)
+        self._typing_tasks: Dict[str, asyncio.Task] = {}
+        self._bot_task: Optional[asyncio.Task] = None
         # Cap to prevent unbounded growth (Discord threads get archived).
         self._MAX_TRACKED_THREADS = 500
-    
+        # Dedup cache: message_id → timestamp.  Prevents duplicate bot
+        # responses when Discord RESUME replays events after reconnects.
+        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 Discord and start receiving events."""
         if not DISCORD_AVAILABLE:
@@ -472,12 +485,23 @@ class DiscordAdapter(BasePlatformAdapter):
                     logger.warning("Opus codec found at %s but failed to load", opus_path)
             if not discord.opus.is_loaded():
                 logger.warning("Opus codec not found — voice channel playback disabled")
-        
+
         if not self.config.token:
             logger.error("[%s] No bot token configured", self.name)
             return False
-        
+
         try:
+            # Acquire scoped lock to prevent duplicate bot token usage
+            from gateway.status import acquire_scoped_lock
+            self._token_lock_identity = self.config.token
+            acquired, existing = acquire_scoped_lock('discord-bot-token', self._token_lock_identity, metadata={'platform': 'discord'})
+            if not acquired:
+                owner_pid = existing.get('pid') if isinstance(existing, dict) else None
+                message = f'Discord bot 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('discord_token_lock', message, retryable=False)
+                return False
+
             # Set up intents -- members intent needed for username-to-ID resolution
             intents = Intents.default()
             intents.message_content = True
@@ -485,13 +509,13 @@ class DiscordAdapter(BasePlatformAdapter):
             intents.guild_messages = True
             intents.members = True
             intents.voice_states = True
-            
+
             # Create bot
             self._client = commands.Bot(
                 command_prefix="!",  # Not really used, we handle raw messages
                 intents=intents,
             )
-            
+
             # Parse allowed user entries (may contain usernames or IDs)
             allowed_env = os.getenv("DISCORD_ALLOWED_USERS", "")
             if allowed_env:
@@ -499,17 +523,17 @@ class DiscordAdapter(BasePlatformAdapter):
                     _clean_discord_id(uid) for uid in allowed_env.split(",")
                     if uid.strip()
                 }
-            
+
             adapter_self = self  # capture for closure
-            
+
             # Register event handlers
             @self._client.event
             async def on_ready():
                 logger.info("[%s] Connected as %s", adapter_self.name, adapter_self._client.user)
-                
+
                 # Resolve any usernames in the allowed list to numeric IDs
                 await adapter_self._resolve_allowed_usernames()
-                
+
                 # Sync slash commands with Discord
                 try:
                     synced = await adapter_self._client.tree.sync()
@@ -517,13 +541,35 @@ class DiscordAdapter(BasePlatformAdapter):
                 except Exception as e:  # pragma: no cover - defensive logging
                     logger.warning("[%s] Slash command sync failed: %s", adapter_self.name, e, exc_info=True)
                 adapter_self._ready_event.set()
-            
+
             @self._client.event
             async def on_message(message: DiscordMessage):
+                # Dedup: Discord RESUME replays events after reconnects (#4777)
+                msg_id = str(message.id)
+                now = time.time()
+                if msg_id in adapter_self._seen_messages:
+                    return
+                adapter_self._seen_messages[msg_id] = now
+                if len(adapter_self._seen_messages) > adapter_self._SEEN_MAX:
+                    cutoff = now - adapter_self._SEEN_TTL
+                    adapter_self._seen_messages = {
+                        k: v for k, v in adapter_self._seen_messages.items()
+                        if v > cutoff
+                    }
+
                 # Always ignore our own messages
                 if message.author == self._client.user:
                     return
-                
+
+                # Ignore Discord system messages (thread renames, pins, member joins, etc.)
+                # Allow both default and reply types — replies have a distinct MessageType.
+                if message.type not in (discord.MessageType.default, discord.MessageType.reply):
+                    return
+
+                # Check if the message author is in the allowed user list
+                if not self._is_allowed_user(str(message.author.id)):
+                    return
+
                 # Bot message filtering (DISCORD_ALLOW_BOTS):
                 #   "none"     — ignore all other bots (default)
                 #   "mentions" — accept bot messages only when they @mention us
@@ -536,7 +582,23 @@ class DiscordAdapter(BasePlatformAdapter):
                         if not self._client.user or self._client.user not in message.mentions:
                             return
                     # "all" falls through to handle_message
-                
+
+                # If the message @mentions other users but NOT the bot, the
+                # sender is talking to someone else — stay silent.  Only
+                # applies in server channels; in DMs the user is always
+                # talking to the bot (mentions are just references).
+                # Controlled by DISCORD_IGNORE_NO_MENTION (default: true).
+                _ignore_no_mention = os.getenv(
+                    "DISCORD_IGNORE_NO_MENTION", "true"
+                ).lower() in ("true", "1", "yes")
+                if _ignore_no_mention and message.mentions and not isinstance(message.channel, discord.DMChannel):
+                    _bot_mentioned = (
+                        self._client.user is not None
+                        and self._client.user in message.mentions
+                    )
+                    if not _bot_mentioned:
+                        return  # Talking to someone else, don't interrupt
+
                 await self._handle_message(message)
 
             @self._client.event
@@ -574,23 +636,23 @@ class DiscordAdapter(BasePlatformAdapter):
 
             # Register slash commands
             self._register_slash_commands()
-            
+
             # Start the bot in background
-            asyncio.create_task(self._client.start(self.config.token))
-            
+            self._bot_task = asyncio.create_task(self._client.start(self.config.token))
+
             # Wait for ready
             await asyncio.wait_for(self._ready_event.wait(), timeout=30)
-            
+
             self._running = True
             return True
-            
+
         except asyncio.TimeoutError:
             logger.error("[%s] Timeout waiting for connection to Discord", self.name, exc_info=True)
             return False
         except Exception as e:  # pragma: no cover - defensive logging
             logger.error("[%s] Failed to connect to Discord: %s", self.name, e, exc_info=True)
             return False
-    
+
     async def disconnect(self) -> None:
         """Disconnect from Discord."""
         # Clean up all active voice connections before closing the client
@@ -609,8 +671,61 @@ class DiscordAdapter(BasePlatformAdapter):
         self._running = False
         self._client = None
         self._ready_event.clear()
+
+        # Release the token lock
+        try:
+            from gateway.status import release_scoped_lock
+            if getattr(self, '_token_lock_identity', None):
+                release_scoped_lock('discord-bot-token', self._token_lock_identity)
+                self._token_lock_identity = None
+        except Exception:
+            pass
+
         logger.info("[%s] Disconnected", self.name)
-    
+
+    async def _add_reaction(self, message: Any, emoji: str) -> bool:
+        """Add an emoji reaction to a Discord message."""
+        if not message or not hasattr(message, "add_reaction"):
+            return False
+        try:
+            await message.add_reaction(emoji)
+            return True
+        except Exception as e:
+            logger.debug("[%s] add_reaction failed (%s): %s", self.name, emoji, e)
+            return False
+
+    async def _remove_reaction(self, message: Any, emoji: str) -> bool:
+        """Remove the bot's own emoji reaction from a Discord message."""
+        if not message or not hasattr(message, "remove_reaction") or not self._client or not self._client.user:
+            return False
+        try:
+            await message.remove_reaction(emoji, self._client.user)
+            return True
+        except Exception as e:
+            logger.debug("[%s] remove_reaction failed (%s): %s", self.name, emoji, e)
+            return False
+
+    def _reactions_enabled(self) -> bool:
+        """Check if message reactions are enabled via config/env."""
+        return os.getenv("DISCORD_REACTIONS", "true").lower() not in ("false", "0", "no")
+
+    async def on_processing_start(self, event: MessageEvent) -> None:
+        """Add an in-progress reaction for normal Discord message events."""
+        if not self._reactions_enabled():
+            return
+        message = event.raw_message
+        if hasattr(message, "add_reaction"):
+            await self._add_reaction(message, "👀")
+
+    async def on_processing_complete(self, event: MessageEvent, success: bool) -> None:
+        """Swap the in-progress reaction for a final success/failure reaction."""
+        if not self._reactions_enabled():
+            return
+        message = event.raw_message
+        if hasattr(message, "add_reaction"):
+            await self._remove_reaction(message, "👀")
+            await self._add_reaction(message, "✅" if success else "❌")
+
     async def send(
         self,
         chat_id: str,
@@ -627,24 +742,24 @@ class DiscordAdapter(BasePlatformAdapter):
             channel = self._client.get_channel(int(chat_id))
             if not channel:
                 channel = await self._client.fetch_channel(int(chat_id))
-            
+
             if not channel:
                 return SendResult(success=False, error=f"Channel {chat_id} not found")
-            
+
             # Format and split message if needed
             formatted = self.format_message(content)
             chunks = self.truncate_message(formatted, self.MAX_MESSAGE_LENGTH)
-            
+
             message_ids = []
             reference = None
-            
+
             if reply_to:
                 try:
                     ref_msg = await channel.fetch_message(int(reply_to))
                     reference = ref_msg
                 except Exception as e:
                     logger.debug("Could not fetch reply-to message: %s", e)
-            
+
             for i, chunk in enumerate(chunks):
                 chunk_reference = reference if i == 0 else None
                 try:
@@ -671,13 +786,13 @@ class DiscordAdapter(BasePlatformAdapter):
                     else:
                         raise
                 message_ids.append(str(msg.id))
-            
+
             return SendResult(
                 success=True,
                 message_id=message_ids[0] if message_ids else None,
                 raw_response={"message_ids": message_ids}
             )
-            
+
         except Exception as e:  # pragma: no cover - defensive logging
             logger.error("[%s] Failed to send Discord message: %s", self.name, e, exc_info=True)
             return SendResult(success=False, error=str(e))
@@ -1149,25 +1264,25 @@ class DiscordAdapter(BasePlatformAdapter):
         """Send an image natively as a Discord file attachment."""
         if not self._client:
             return SendResult(success=False, error="Not connected")
-        
+
         try:
             import aiohttp
-            
+
             channel = self._client.get_channel(int(chat_id))
             if not channel:
                 channel = await self._client.fetch_channel(int(chat_id))
             if not channel:
                 return SendResult(success=False, error=f"Channel {chat_id} not found")
-            
+
             # Download the image and send as a Discord file attachment
             # (Discord renders attachments inline, unlike plain URLs)
             async with aiohttp.ClientSession() as session:
                 async with session.get(image_url, timeout=aiohttp.ClientTimeout(total=30)) as resp:
                     if resp.status != 200:
                         raise Exception(f"Failed to download image: HTTP {resp.status}")
-                    
+
                     image_data = await resp.read()
-                    
+
                     # Determine filename from URL or content type
                     content_type = resp.headers.get("content-type", "image/png")
                     ext = "png"
@@ -1177,16 +1292,16 @@ class DiscordAdapter(BasePlatformAdapter):
                         ext = "gif"
                     elif "webp" in content_type:
                         ext = "webp"
-                    
+
                     import io
                     file = discord.File(io.BytesIO(image_data), filename=f"image.{ext}")
-                    
+
                     msg = await channel.send(
                         content=caption if caption else None,
                         file=file,
                     )
                     return SendResult(success=True, message_id=str(msg.id))
-        
+
         except ImportError:
             logger.warning(
                 "[%s] aiohttp not installed, falling back to URL. Run: pip install aiohttp",
@@ -1237,30 +1352,64 @@ class DiscordAdapter(BasePlatformAdapter):
         except Exception as e:  # pragma: no cover - defensive logging
             logger.error("[%s] Failed to send document, falling back to base adapter: %s", self.name, e, exc_info=True)
             return await super().send_document(chat_id, file_path, caption, file_name, reply_to, metadata=metadata)
-    
+
     async def send_typing(self, chat_id: str, metadata=None) -> None:
-        """Send typing indicator."""
-        if self._client:
+        """Start a persistent typing indicator for a channel.
+
+        Discord's TYPING_START gateway event is unreliable in DMs for bots.
+        Instead, start a background loop that hits the typing endpoint every
+        8 seconds (typing indicator lasts ~10s).  The loop is cancelled when
+        stop_typing() is called (after the response is sent).
+        """
+        if not self._client:
+            return
+        # Don't start a duplicate loop
+        if chat_id in self._typing_tasks:
+            return
+
+        async def _typing_loop() -> None:
             try:
-                channel = self._client.get_channel(int(chat_id))
-                if channel:
-                    await channel.typing()
-            except Exception:
-                pass  # Ignore typing indicator failures
-    
+                while True:
+                    try:
+                        route = discord.http.Route(
+                            "POST", "/channels/{channel_id}/typing",
+                            channel_id=chat_id,
+                        )
+                        await self._client.http.request(route)
+                    except asyncio.CancelledError:
+                        return
+                    except Exception as e:
+                        logger.debug("Discord typing indicator failed for %s: %s", chat_id, e)
+                        return
+                    await asyncio.sleep(8)
+            except asyncio.CancelledError:
+                pass
+
+        self._typing_tasks[chat_id] = asyncio.create_task(_typing_loop())
+
+    async def stop_typing(self, chat_id: str) -> None:
+        """Stop the persistent typing indicator for a channel."""
+        task = self._typing_tasks.pop(chat_id, None)
+        if task:
+            task.cancel()
+            try:
+                await task
+            except (asyncio.CancelledError, Exception):
+                pass
+
     async def get_chat_info(self, chat_id: str) -> Dict[str, Any]:
         """Get information about a Discord channel."""
         if not self._client:
             return {"name": "Unknown", "type": "dm"}
-        
+
         try:
             channel = self._client.get_channel(int(chat_id))
             if not channel:
                 channel = await self._client.fetch_channel(int(chat_id))
-            
+
             if not channel:
                 return {"name": str(chat_id), "type": "dm"}
-            
+
             # Determine channel type
             if isinstance(channel, discord.DMChannel):
                 chat_type = "dm"
@@ -1276,7 +1425,7 @@ class DiscordAdapter(BasePlatformAdapter):
             else:
                 chat_type = "channel"
                 name = getattr(channel, "name", str(chat_id))
-            
+
             return {
                 "name": name,
                 "type": chat_type,
@@ -1286,7 +1435,7 @@ class DiscordAdapter(BasePlatformAdapter):
         except Exception as e:  # pragma: no cover - defensive logging
             logger.error("[%s] Failed to get chat info for %s: %s", self.name, chat_id, e, exc_info=True)
             return {"name": str(chat_id), "type": "dm", "error": str(e)}
-    
+
     async def _resolve_allowed_usernames(self) -> None:
         """
         Resolve non-numeric entries in DISCORD_ALLOWED_USERS to Discord user IDs.
@@ -1354,7 +1503,7 @@ class DiscordAdapter(BasePlatformAdapter):
     def format_message(self, content: str) -> str:
         """
         Format message for Discord.
-        
+
         Discord uses its own markdown variant.
         """
         # Discord markdown is fairly standard, no special escaping needed
@@ -1366,15 +1515,23 @@ class DiscordAdapter(BasePlatformAdapter):
         command_text: str,
         followup_msg: str | None = None,
     ) -> None:
-        """Common handler for simple slash commands that dispatch a command string."""
+        """Common handler for simple slash commands that dispatch a command string.
+
+        Defers the interaction (shows "thinking..."), dispatches the command,
+        then cleans up the deferred response.  If *followup_msg* is provided
+        the "thinking..." indicator is replaced with that text; otherwise it
+        is deleted so the channel isn't cluttered.
+        """
         await interaction.response.defer(ephemeral=True)
         event = self._build_slash_event(interaction, command_text)
         await self.handle_message(event)
-        if followup_msg:
-            try:
-                await interaction.followup.send(followup_msg, ephemeral=True)
-            except Exception as e:
-                logger.debug("Discord followup failed: %s", e)
+        try:
+            if followup_msg:
+                await interaction.edit_original_response(content=followup_msg)
+            else:
+                await interaction.delete_original_response()
+        except Exception as e:
+            logger.debug("Discord interaction cleanup failed: %s", e)
 
     def _register_slash_commands(self) -> None:
         """Register Discord slash commands on the command tree."""
@@ -1399,9 +1556,7 @@ class DiscordAdapter(BasePlatformAdapter):
         @tree.command(name="reasoning", description="Show or change reasoning effort")
         @discord.app_commands.describe(effort="Reasoning effort: xhigh, high, medium, low, minimal, or none.")
         async def slash_reasoning(interaction: discord.Interaction, effort: str = ""):
-            await interaction.response.defer(ephemeral=True)
-            event = self._build_slash_event(interaction, f"/reasoning {effort}".strip())
-            await self.handle_message(event)
+            await self._run_simple_slash(interaction, f"/reasoning {effort}".strip())
 
         @tree.command(name="personality", description="Set a personality")
         @discord.app_commands.describe(name="Personality name. Leave empty to list available.")
@@ -1474,14 +1629,22 @@ class DiscordAdapter(BasePlatformAdapter):
             discord.app_commands.Choice(name="status — show current mode", value="status"),
         ])
         async def slash_voice(interaction: discord.Interaction, mode: str = ""):
-            await interaction.response.defer(ephemeral=True)
-            event = self._build_slash_event(interaction, f"/voice {mode}".strip())
-            await self.handle_message(event)
+            await self._run_simple_slash(interaction, f"/voice {mode}".strip())
 
         @tree.command(name="update", description="Update Hermes Agent to the latest version")
         async def slash_update(interaction: discord.Interaction):
             await self._run_simple_slash(interaction, "/update", "Update initiated~")
 
+        @tree.command(name="approve", description="Approve a pending dangerous command")
+        @discord.app_commands.describe(scope="Optional: 'all', 'session', 'always', 'all session', 'all always'")
+        async def slash_approve(interaction: discord.Interaction, scope: str = ""):
+            await self._run_simple_slash(interaction, f"/approve {scope}".strip())
+
+        @tree.command(name="deny", description="Deny a pending dangerous command")
+        @discord.app_commands.describe(scope="Optional: 'all' to deny all pending commands")
+        async def slash_deny(interaction: discord.Interaction, scope: str = ""):
+            await self._run_simple_slash(interaction, f"/deny {scope}".strip())
+
         @tree.command(name="thread", description="Create a new thread and start a Hermes session in it")
         @discord.app_commands.describe(
             name="Thread name",
@@ -1500,13 +1663,23 @@ class DiscordAdapter(BasePlatformAdapter):
     def _build_slash_event(self, interaction: discord.Interaction, text: str) -> MessageEvent:
         """Build a MessageEvent from a Discord slash command interaction."""
         is_dm = isinstance(interaction.channel, discord.DMChannel)
-        chat_type = "dm" if is_dm else "group"
+        is_thread = isinstance(interaction.channel, discord.Thread)
+        thread_id = None
+
+        if is_dm:
+            chat_type = "dm"
+        elif is_thread:
+            chat_type = "thread"
+            thread_id = str(interaction.channel_id)
+        else:
+            chat_type = "group"
+
         chat_name = ""
         if not is_dm and hasattr(interaction.channel, "name"):
             chat_name = interaction.channel.name
             if hasattr(interaction.channel, "guild") and interaction.channel.guild:
                 chat_name = f"{interaction.channel.guild.name} / #{chat_name}"
-        
+
         # Get channel topic (if available)
         chat_topic = getattr(interaction.channel, "topic", None)
 
@@ -1516,6 +1689,7 @@ class DiscordAdapter(BasePlatformAdapter):
             chat_type=chat_type,
             user_id=str(interaction.user.id),
             user_name=interaction.user.display_name,
+            thread_id=thread_id,
             chat_topic=chat_topic,
         )
 
@@ -1714,33 +1888,41 @@ class DiscordAdapter(BasePlatformAdapter):
             return None
 
     async def send_exec_approval(
-        self, chat_id: str, command: str, approval_id: str
+        self, chat_id: str, command: str, session_key: str,
+        description: str = "dangerous command",
+        metadata: Optional[dict] = None,
     ) -> SendResult:
         """
         Send a button-based exec approval prompt for a dangerous command.
 
-        Returns SendResult. The approval is resolved when a user clicks a button.
+        The buttons call ``resolve_gateway_approval()`` to unblock the waiting
+        agent thread — this replaces the text-based ``/approve`` flow on Discord.
         """
         if not self._client or not DISCORD_AVAILABLE:
             return SendResult(success=False, error="Not connected")
 
         try:
-            channel = self._client.get_channel(int(chat_id))
+            # Resolve channel — use thread_id from metadata if present
+            target_id = chat_id
+            if metadata and metadata.get("thread_id"):
+                target_id = metadata["thread_id"]
+
+            channel = self._client.get_channel(int(target_id))
             if not channel:
-                channel = await self._client.fetch_channel(int(chat_id))
+                channel = await self._client.fetch_channel(int(target_id))
 
             # Discord embed description limit is 4096; show full command up to that
             max_desc = 4088
             cmd_display = command if len(command) <= max_desc else command[: max_desc - 3] + "..."
             embed = discord.Embed(
-                title="Command Approval Required",
+                title="⚠️ Command Approval Required",
                 description=f"```\n{cmd_display}\n```",
                 color=discord.Color.orange(),
             )
-            embed.set_footer(text=f"Approval ID: {approval_id}")
+            embed.add_field(name="Reason", value=description, inline=False)
 
             view = ExecApprovalView(
-                approval_id=approval_id,
+                session_key=session_key,
                 allowed_user_ids=self._allowed_user_ids,
             )
 
@@ -1902,9 +2084,14 @@ class DiscordAdapter(BasePlatformAdapter):
                     elif att.content_type.startswith("audio/"):
                         msg_type = MessageType.AUDIO
                     else:
-                        msg_type = MessageType.DOCUMENT
+                        doc_ext = ""
+                        if att.filename:
+                            _, doc_ext = os.path.splitext(att.filename)
+                            doc_ext = doc_ext.lower()
+                        if doc_ext in SUPPORTED_DOCUMENT_TYPES:
+                            msg_type = MessageType.DOCUMENT
                     break
-        
+
         # When auto-threading kicked in, route responses to the new thread
         effective_channel = auto_threaded_channel or message.channel
 
@@ -1923,7 +2110,7 @@ class DiscordAdapter(BasePlatformAdapter):
 
         # Get channel topic (if available - TextChannels have topics, DMs/threads don't)
         chat_topic = getattr(message.channel, "topic", None)
-        
+
         # Build source
         source = self.build_source(
             chat_id=str(effective_channel.id),
@@ -1934,11 +2121,12 @@ class DiscordAdapter(BasePlatformAdapter):
             thread_id=thread_id,
             chat_topic=chat_topic,
         )
-        
+
         # Build media URLs -- download image attachments to local cache so the
         # vision tool can access them reliably (Discord CDN URLs can expire).
         media_urls = []
         media_types = []
+        pending_text_injection: Optional[str] = None
         for att in message.attachments:
             content_type = att.content_type or "unknown"
             if content_type.startswith("image/"):
@@ -1970,12 +2158,75 @@ class DiscordAdapter(BasePlatformAdapter):
                     media_urls.append(att.url)
                     media_types.append(content_type)
             else:
-                # Other attachments: keep the original URL
-                media_urls.append(att.url)
-                media_types.append(content_type)
-        
+                # Document attachments: download, cache, and optionally inject text
+                ext = ""
+                if att.filename:
+                    _, ext = os.path.splitext(att.filename)
+                    ext = ext.lower()
+                if not ext and content_type:
+                    mime_to_ext = {v: k for k, v in SUPPORTED_DOCUMENT_TYPES.items()}
+                    ext = mime_to_ext.get(content_type, "")
+                if ext not in SUPPORTED_DOCUMENT_TYPES:
+                    logger.warning(
+                        "[Discord] Unsupported document type '%s' (%s), skipping",
+                        ext or "unknown", content_type,
+                    )
+                else:
+                    MAX_DOC_BYTES = 20 * 1024 * 1024
+                    if att.size and att.size > MAX_DOC_BYTES:
+                        logger.warning(
+                            "[Discord] Document too large (%s bytes), skipping: %s",
+                            att.size, att.filename,
+                        )
+                    else:
+                        try:
+                            import aiohttp
+                            async with aiohttp.ClientSession() as session:
+                                async with session.get(
+                                    att.url,
+                                    timeout=aiohttp.ClientTimeout(total=30),
+                                ) as resp:
+                                    if resp.status != 200:
+                                        raise Exception(f"HTTP {resp.status}")
+                                    raw_bytes = await resp.read()
+                            cached_path = cache_document_from_bytes(
+                                raw_bytes, att.filename or f"document{ext}"
+                            )
+                            doc_mime = SUPPORTED_DOCUMENT_TYPES[ext]
+                            media_urls.append(cached_path)
+                            media_types.append(doc_mime)
+                            logger.info("[Discord] Cached user document: %s", cached_path)
+                            # Inject text content for .txt/.md files (capped at 100 KB)
+                            MAX_TEXT_INJECT_BYTES = 100 * 1024
+                            if ext in (".md", ".txt") and len(raw_bytes) <= MAX_TEXT_INJECT_BYTES:
+                                try:
+                                    text_content = raw_bytes.decode("utf-8")
+                                    display_name = att.filename or f"document{ext}"
+                                    display_name = re.sub(r'[^\w.\- ]', '_', display_name)
+                                    injection = f"[Content of {display_name}]:\n{text_content}"
+                                    if pending_text_injection:
+                                        pending_text_injection = f"{pending_text_injection}\n\n{injection}"
+                                    else:
+                                        pending_text_injection = injection
+                                except UnicodeDecodeError:
+                                    pass
+                        except Exception as e:
+                            logger.warning(
+                                "[Discord] Failed to cache document %s: %s",
+                                att.filename, e, exc_info=True,
+                            )
+
+        event_text = message.content
+        if pending_text_injection:
+            event_text = f"{pending_text_injection}\n\n{event_text}" if event_text else pending_text_injection
+
+        # Defense-in-depth: prevent empty user messages from entering session
+        # (can happen when user sends @mention-only with no other text)
+        if not event_text or not event_text.strip():
+            event_text = "(The user sent a message with no text content)"
+
         event = MessageEvent(
-            text=message.content,
+            text=event_text,
             message_type=msg_type,
             source=source,
             raw_message=message,
@@ -2004,13 +2255,15 @@ if DISCORD_AVAILABLE:
         """
         Interactive button view for exec approval of dangerous commands.
 
-        Shows three buttons: Allow Once (green), Always Allow (blue), Deny (red).
-        Only users in the allowed list can click. The view times out after 5 minutes.
+        Shows four buttons: Allow Once, Allow Session, Always Allow, Deny.
+        Clicking a button calls ``resolve_gateway_approval()`` to unblock the
+        waiting agent thread — the same mechanism as the text ``/approve`` flow.
+        Only users in the allowed list can click.  Times out after 5 minutes.
         """
 
-        def __init__(self, approval_id: str, allowed_user_ids: set):
+        def __init__(self, session_key: str, allowed_user_ids: set):
             super().__init__(timeout=300)  # 5-minute timeout
-            self.approval_id = approval_id
+            self.session_key = session_key
             self.allowed_user_ids = allowed_user_ids
             self.resolved = False
 
@@ -2021,9 +2274,10 @@ if DISCORD_AVAILABLE:
             return str(interaction.user.id) in self.allowed_user_ids
 
         async def _resolve(
-            self, interaction: discord.Interaction, action: str, color: discord.Color
+            self, interaction: discord.Interaction, choice: str,
+            color: discord.Color, label: str,
         ):
-            """Resolve the approval and update the message."""
+            """Resolve the approval via the gateway approval queue and update the embed."""
             if self.resolved:
                 await interaction.response.send_message(
                     "This approval has already been resolved~", ephemeral=True
@@ -2042,7 +2296,7 @@ if DISCORD_AVAILABLE:
             embed = interaction.message.embeds[0] if interaction.message.embeds else None
             if embed:
                 embed.color = color
-                embed.set_footer(text=f"{action} by {interaction.user.display_name}")
+                embed.set_footer(text=f"{label} by {interaction.user.display_name}")
 
             # Disable all buttons
             for child in self.children:
@@ -2050,33 +2304,40 @@ if DISCORD_AVAILABLE:
 
             await interaction.response.edit_message(embed=embed, view=self)
 
-            # Store the approval decision
+            # Unblock the waiting agent thread via the gateway approval queue
             try:
-                from tools.approval import approve_permanent
-                if action == "allow_once":
-                    pass  # One-time approval handled by gateway
-                elif action == "allow_always":
-                    approve_permanent(self.approval_id)
-            except ImportError:
-                pass
+                from tools.approval import resolve_gateway_approval
+                count = resolve_gateway_approval(self.session_key, choice)
+                logger.info(
+                    "Discord button resolved %d approval(s) for session %s (choice=%s, user=%s)",
+                    count, self.session_key, choice, interaction.user.display_name,
+                )
+            except Exception as exc:
+                logger.error("Failed to resolve gateway approval from button: %s", exc)
 
         @discord.ui.button(label="Allow Once", style=discord.ButtonStyle.green)
         async def allow_once(
             self, interaction: discord.Interaction, button: discord.ui.Button
         ):
-            await self._resolve(interaction, "allow_once", discord.Color.green())
+            await self._resolve(interaction, "once", discord.Color.green(), "Approved once")
+
+        @discord.ui.button(label="Allow Session", style=discord.ButtonStyle.grey)
+        async def allow_session(
+            self, interaction: discord.Interaction, button: discord.ui.Button
+        ):
+            await self._resolve(interaction, "session", discord.Color.blue(), "Approved for session")
 
         @discord.ui.button(label="Always Allow", style=discord.ButtonStyle.blurple)
         async def allow_always(
             self, interaction: discord.Interaction, button: discord.ui.Button
         ):
-            await self._resolve(interaction, "allow_always", discord.Color.blue())
+            await self._resolve(interaction, "always", discord.Color.purple(), "Approved permanently")
 
         @discord.ui.button(label="Deny", style=discord.ButtonStyle.red)
         async def deny(
             self, interaction: discord.Interaction, button: discord.ui.Button
         ):
-            await self._resolve(interaction, "deny", discord.Color.red())
+            await self._resolve(interaction, "deny", discord.Color.red(), "Denied")
 
         async def on_timeout(self):
             """Handle view timeout -- disable buttons and mark as expired."""

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")
@@ -214,6 +237,7 @@ class EmailAdapter(BasePlatformAdapter):
 
         # 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
@@ -221,18 +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:
@@ -241,7 +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()
@@ -290,53 +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, skip_attachments=self._skip_attachments)
+                    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
@@ -349,6 +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"]
@@ -443,18 +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, metadata: Optional[Dict[str, Any]] = None) -> None:
         """Email has no typing indicator — no-op."""
-        pass
 
     async def send_image(
         self,
@@ -531,11 +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/matrix.py

@@ -5,18 +5,22 @@ matrix-nio Python SDK.  Supports optional end-to-end encryption (E2EE)
 when installed with ``pip install "matrix-nio[e2e]"``.
 
 Environment variables:
-    MATRIX_HOMESERVER       Homeserver URL (e.g. https://matrix.example.org)
-    MATRIX_ACCESS_TOKEN     Access token (preferred auth method)
-    MATRIX_USER_ID          Full user ID (@bot:server) — required for password login
-    MATRIX_PASSWORD         Password (alternative to access token)
-    MATRIX_ENCRYPTION       Set "true" to enable E2EE
-    MATRIX_ALLOWED_USERS    Comma-separated Matrix user IDs (@user:server)
-    MATRIX_HOME_ROOM        Room ID for cron/notification delivery
+    MATRIX_HOMESERVER           Homeserver URL (e.g. https://matrix.example.org)
+    MATRIX_ACCESS_TOKEN         Access token (preferred auth method)
+    MATRIX_USER_ID              Full user ID (@bot:server) — required for password login
+    MATRIX_PASSWORD             Password (alternative to access token)
+    MATRIX_ENCRYPTION           Set "true" to enable E2EE
+    MATRIX_ALLOWED_USERS        Comma-separated Matrix user IDs (@user:server)
+    MATRIX_HOME_ROOM            Room ID for cron/notification delivery
+    MATRIX_REQUIRE_MENTION      Require @mention in rooms (default: true)
+    MATRIX_FREE_RESPONSE_ROOMS  Comma-separated room IDs exempt from mention requirement
+    MATRIX_AUTO_THREAD          Auto-create threads for room messages (default: true)
 """
 
 from __future__ import annotations
 
 import asyncio
+import io
 import json
 import logging
 import mimetypes
@@ -24,7 +28,7 @@ import os
 import re
 import time
 from pathlib import Path
-from typing import Any, Dict, List, Optional, Set
+from typing import Any, Dict, Optional, Set
 
 from gateway.config import Platform, PlatformConfig
 from gateway.platforms.base import (
@@ -41,11 +45,21 @@ logger = logging.getLogger(__name__)
 MAX_MESSAGE_LENGTH = 4000
 
 # Store directory for E2EE keys and sync state.
-_STORE_DIR = Path.home() / ".hermes" / "matrix" / "store"
+# Uses get_hermes_home() so each profile gets its own Matrix store.
+from hermes_constants import get_hermes_dir as _get_hermes_dir
+_STORE_DIR = _get_hermes_dir("platforms/matrix/store", "matrix/store")
 
 # Grace period: ignore messages older than this many seconds before startup.
 _STARTUP_GRACE_SECONDS = 5
 
+# E2EE key export file for persistence across restarts.
+_KEY_EXPORT_FILE = _STORE_DIR / "exported_keys.txt"
+_KEY_EXPORT_PASSPHRASE = "hermes-matrix-e2ee-keys"
+
+# Pending undecrypted events: cap and TTL for retry buffer.
+_MAX_PENDING_EVENTS = 100
+_PENDING_EVENT_TTL = 300  # seconds — stop retrying after 5 min
+
 
 def check_matrix_requirements() -> bool:
     """Return True if the Matrix adapter can be used."""
@@ -103,6 +117,31 @@ class MatrixAdapter(BasePlatformAdapter):
         self._dm_rooms: Dict[str, bool] = {}
         # Set of room IDs we've joined
         self._joined_rooms: Set[str] = set()
+        # Event deduplication (bounded deque keeps newest entries)
+        from collections import deque
+        self._processed_events: deque = deque(maxlen=1000)
+        self._processed_events_set: set = set()
+
+        # Buffer for undecrypted events pending key receipt.
+        # Each entry: (room, event, timestamp)
+        self._pending_megolm: list = []
+
+        # Thread participation tracking (for require_mention bypass)
+        self._bot_participated_threads: set = self._load_participated_threads()
+        self._MAX_TRACKED_THREADS = 500
+
+    def _is_duplicate_event(self, event_id) -> bool:
+        """Return True if this event was already processed. Tracks the ID otherwise."""
+        if not event_id:
+            return False
+        if event_id in self._processed_events_set:
+            return True
+        if len(self._processed_events) == self._processed_events.maxlen:
+            evicted = self._processed_events[0]
+            self._processed_events_set.discard(evicted)
+        self._processed_events.append(event_id)
+        self._processed_events_set.add(event_id)
+        return False
 
     # ------------------------------------------------------------------
     # Required overrides
@@ -145,22 +184,49 @@ class MatrixAdapter(BasePlatformAdapter):
         # Authenticate.
         if self._access_token:
             client.access_token = self._access_token
-            # Resolve user_id if not set.
-            if not self._user_id:
-                resp = await client.whoami()
-                if isinstance(resp, nio.WhoamiResponse):
-                    self._user_id = resp.user_id
-                    client.user_id = resp.user_id
-                    logger.info("Matrix: authenticated as %s", self._user_id)
-                else:
-                    logger.error(
-                        "Matrix: whoami failed — check MATRIX_ACCESS_TOKEN and MATRIX_HOMESERVER"
+
+            # With access-token auth, always resolve whoami so we validate the
+            # token and learn the device_id. The device_id matters for E2EE:
+            # without it, matrix-nio can send plain messages but may fail to
+            # decrypt inbound encrypted events or encrypt outbound room sends.
+            resp = await client.whoami()
+            if isinstance(resp, nio.WhoamiResponse):
+                resolved_user_id = getattr(resp, "user_id", "") or self._user_id
+                resolved_device_id = getattr(resp, "device_id", "")
+                if resolved_user_id:
+                    self._user_id = resolved_user_id
+
+                # restore_login() is the matrix-nio path that binds the access
+                # token to a specific device and loads the crypto store.
+                if resolved_device_id and hasattr(client, "restore_login"):
+                    client.restore_login(
+                        self._user_id or resolved_user_id,
+                        resolved_device_id,
+                        self._access_token,
                     )
-                    await client.close()
-                    return False
+                else:
+                    if self._user_id:
+                        client.user_id = self._user_id
+                    if resolved_device_id:
+                        client.device_id = resolved_device_id
+                    client.access_token = self._access_token
+                    if self._encryption:
+                        logger.warning(
+                            "Matrix: access-token login did not restore E2EE state; "
+                            "encrypted rooms may fail until a device_id is available"
+                        )
+
+                logger.info(
+                    "Matrix: using access token for %s%s",
+                    self._user_id or "(unknown user)",
+                    f" (device {resolved_device_id})" if resolved_device_id else "",
+                )
             else:
-                client.user_id = self._user_id
-                logger.info("Matrix: using access token for %s", self._user_id)
+                logger.error(
+                    "Matrix: whoami failed — check MATRIX_ACCESS_TOKEN and MATRIX_HOMESERVER"
+                )
+                await client.close()
+                return False
         elif self._password and self._user_id:
             resp = await client.login(
                 self._password,
@@ -178,7 +244,7 @@ class MatrixAdapter(BasePlatformAdapter):
             return False
 
         # If E2EE is enabled, load the crypto store.
-        if self._encryption and hasattr(client, "olm"):
+        if self._encryption and getattr(client, "olm", None):
             try:
                 if client.should_upload_keys:
                     await client.keys_upload()
@@ -186,9 +252,23 @@ class MatrixAdapter(BasePlatformAdapter):
             except Exception as exc:
                 logger.warning("Matrix: crypto init issue: %s", exc)
 
+            # Import previously exported Megolm keys (survives restarts).
+            if _KEY_EXPORT_FILE.exists():
+                try:
+                    await client.import_keys(
+                        str(_KEY_EXPORT_FILE), _KEY_EXPORT_PASSPHRASE,
+                    )
+                    logger.info("Matrix: imported Megolm keys from backup")
+                except Exception as exc:
+                    logger.debug("Matrix: could not import keys: %s", exc)
+        elif self._encryption:
+            logger.warning(
+                "Matrix: E2EE requested but crypto store is not loaded; "
+                "encrypted rooms may fail"
+            )
+
         # Register event callbacks.
         client.add_event_callback(self._on_room_message, nio.RoomMessageText)
-        client.add_event_callback(self._on_room_message_media, nio.RoomMessageMedia)
         client.add_event_callback(self._on_room_message_media, nio.RoomMessageImage)
         client.add_event_callback(self._on_room_message_media, nio.RoomMessageAudio)
         client.add_event_callback(self._on_room_message_media, nio.RoomMessageVideo)
@@ -215,6 +295,7 @@ class MatrixAdapter(BasePlatformAdapter):
             )
             # Build DM room cache from m.direct account data.
             await self._refresh_dm_cache()
+            await self._run_e2ee_maintenance()
         else:
             logger.warning("Matrix: initial sync returned %s", type(resp).__name__)
 
@@ -234,6 +315,18 @@ class MatrixAdapter(BasePlatformAdapter):
             except (asyncio.CancelledError, Exception):
                 pass
 
+        # Export Megolm keys before closing so the next restart can decrypt
+        # events that used sessions from this run.
+        if self._client and self._encryption and getattr(self._client, "olm", None):
+            try:
+                _STORE_DIR.mkdir(parents=True, exist_ok=True)
+                await self._client.export_keys(
+                    str(_KEY_EXPORT_FILE), _KEY_EXPORT_PASSPHRASE,
+                )
+                logger.info("Matrix: exported Megolm keys for next restart")
+            except Exception as exc:
+                logger.debug("Matrix: could not export keys on disconnect: %s", exc)
+
         if self._client:
             await self._client.close()
             self._client = None
@@ -286,13 +379,48 @@ class MatrixAdapter(BasePlatformAdapter):
                     relates_to["m.in_reply_to"] = {"event_id": reply_to}
                 msg_content["m.relates_to"] = relates_to
 
-            resp = await self._client.room_send(
-                chat_id,
-                "m.room.message",
-                msg_content,
-            )
+            async def _room_send_once(*, ignore_unverified_devices: bool = False):
+                return await asyncio.wait_for(
+                    self._client.room_send(
+                        chat_id,
+                        "m.room.message",
+                        msg_content,
+                        ignore_unverified_devices=ignore_unverified_devices,
+                    ),
+                    timeout=45,
+                )
+
+            try:
+                resp = await _room_send_once(ignore_unverified_devices=False)
+            except Exception as exc:
+                retryable = isinstance(exc, asyncio.TimeoutError)
+                olm_unverified = getattr(nio, "OlmUnverifiedDeviceError", None)
+                send_retry = getattr(nio, "SendRetryError", None)
+                if isinstance(olm_unverified, type) and isinstance(exc, olm_unverified):
+                    retryable = True
+                if isinstance(send_retry, type) and isinstance(exc, send_retry):
+                    retryable = True
+
+                if not retryable:
+                    logger.error("Matrix: failed to send to %s: %s", chat_id, exc)
+                    return SendResult(success=False, error=str(exc))
+
+                logger.warning(
+                    "Matrix: initial encrypted send to %s failed (%s); "
+                    "retrying after E2EE maintenance with ignored unverified devices",
+                    chat_id,
+                    exc,
+                )
+                await self._run_e2ee_maintenance()
+                try:
+                    resp = await _room_send_once(ignore_unverified_devices=True)
+                except Exception as retry_exc:
+                    logger.error("Matrix: failed to send to %s after retry: %s", chat_id, retry_exc)
+                    return SendResult(success=False, error=str(retry_exc))
+
             if isinstance(resp, nio.RoomSendResponse):
                 last_event_id = resp.event_id
+                logger.info("Matrix: sent event %s to %s", last_event_id, chat_id)
             else:
                 err = getattr(resp, "message", str(resp))
                 logger.error("Matrix: failed to send to %s: %s", chat_id, err)
@@ -427,8 +555,11 @@ class MatrixAdapter(BasePlatformAdapter):
         reply_to: Optional[str] = None,
         metadata: Optional[Dict[str, Any]] = None,
     ) -> SendResult:
-        """Upload an audio file as a voice message."""
-        return await self._send_local_file(chat_id, audio_path, "m.audio", caption, reply_to, metadata=metadata)
+        """Upload an audio file as a voice message (MSC3245 native voice)."""
+        return await self._send_local_file(
+            chat_id, audio_path, "m.audio", caption, reply_to, 
+            metadata=metadata, is_voice=True
+        )
 
     async def send_video(
         self,
@@ -461,13 +592,16 @@ class MatrixAdapter(BasePlatformAdapter):
         caption: Optional[str] = None,
         reply_to: Optional[str] = None,
         metadata: Optional[Dict[str, Any]] = None,
+        is_voice: bool = False,
     ) -> SendResult:
         """Upload bytes to Matrix and send as a media message."""
         import nio
 
         # Upload to homeserver.
-        resp = await self._client.upload(
-            data,
+        # nio expects a DataProvider (callable) or file-like object, not raw bytes.
+        # nio.upload() returns a tuple (UploadResponse|UploadError, Optional[Dict])
+        resp, maybe_encryption_info = await self._client.upload(
+            io.BytesIO(data),
             content_type=content_type,
             filename=filename,
         )
@@ -489,6 +623,10 @@ class MatrixAdapter(BasePlatformAdapter):
             },
         }
 
+        # Add MSC3245 voice flag for native voice messages.
+        if is_voice:
+            msg_content["org.matrix.msc3245.voice"] = {}
+
         if reply_to:
             msg_content["m.relates_to"] = {
                 "m.in_reply_to": {"event_id": reply_to}
@@ -516,6 +654,7 @@ class MatrixAdapter(BasePlatformAdapter):
         reply_to: Optional[str] = None,
         file_name: Optional[str] = None,
         metadata: Optional[Dict[str, Any]] = None,
+        is_voice: bool = False,
     ) -> SendResult:
         """Read a local file and upload it."""
         p = Path(file_path)
@@ -528,7 +667,7 @@ class MatrixAdapter(BasePlatformAdapter):
         ct = mimetypes.guess_type(fname)[0] or "application/octet-stream"
         data = p.read_bytes()
 
-        return await self._upload_and_send(room_id, data, fname, ct, msgtype, caption, reply_to, metadata)
+        return await self._upload_and_send(room_id, data, fname, ct, msgtype, caption, reply_to, metadata, is_voice)
 
     # ------------------------------------------------------------------
     # Sync loop
@@ -536,9 +675,23 @@ class MatrixAdapter(BasePlatformAdapter):
 
     async def _sync_loop(self) -> None:
         """Continuously sync with the homeserver."""
+        import nio
+
         while not self._closing:
             try:
-                await self._client.sync(timeout=30000)
+                resp = await self._client.sync(timeout=30000)
+                if isinstance(resp, nio.SyncError):
+                    if self._closing:
+                        return
+                    logger.warning(
+                        "Matrix: sync returned %s: %s — retrying in 5s",
+                        type(resp).__name__,
+                        getattr(resp, "message", resp),
+                    )
+                    await asyncio.sleep(5)
+                    continue
+
+                await self._run_e2ee_maintenance()
             except asyncio.CancelledError:
                 return
             except Exception as exc:
@@ -547,6 +700,148 @@ class MatrixAdapter(BasePlatformAdapter):
                 logger.warning("Matrix: sync error: %s — retrying in 5s", exc)
                 await asyncio.sleep(5)
 
+    async def _run_e2ee_maintenance(self) -> None:
+        """Run matrix-nio E2EE housekeeping between syncs.
+
+        Hermes uses a custom sync loop instead of matrix-nio's sync_forever(),
+        so we need to explicitly drive the key management work that sync_forever()
+        normally handles for encrypted rooms.
+
+        Also auto-trusts all devices (so senders share session keys with us)
+        and retries decryption for any buffered MegolmEvents.
+        """
+        client = self._client
+        if not client or not self._encryption or not getattr(client, "olm", None):
+            return
+
+        did_query_keys = client.should_query_keys
+
+        tasks = [asyncio.create_task(client.send_to_device_messages())]
+
+        if client.should_upload_keys:
+            tasks.append(asyncio.create_task(client.keys_upload()))
+
+        if did_query_keys:
+            tasks.append(asyncio.create_task(client.keys_query()))
+
+        if client.should_claim_keys:
+            users = client.get_users_for_key_claiming()
+            if users:
+                tasks.append(asyncio.create_task(client.keys_claim(users)))
+
+        for task in asyncio.as_completed(tasks):
+            try:
+                await task
+            except asyncio.CancelledError:
+                raise
+            except Exception as exc:
+                logger.warning("Matrix: E2EE maintenance task failed: %s", exc)
+
+        # After key queries, auto-trust all devices so senders share keys with
+        # us.  For a bot this is the right default — we want to decrypt
+        # everything, not enforce manual verification.
+        if did_query_keys:
+            self._auto_trust_devices()
+
+        # Retry any buffered undecrypted events now that new keys may have
+        # arrived (from key requests, key queries, or to-device forwarding).
+        if self._pending_megolm:
+            await self._retry_pending_decryptions()
+
+    def _auto_trust_devices(self) -> None:
+        """Trust/verify all unverified devices we know about.
+
+        When other clients see our device as verified, they proactively share
+        Megolm session keys with us.  Without this, many clients will refuse
+        to include an unverified device in key distributions.
+        """
+        client = self._client
+        if not client:
+            return
+
+        device_store = getattr(client, "device_store", None)
+        if not device_store:
+            return
+
+        own_device = getattr(client, "device_id", None)
+        trusted_count = 0
+
+        try:
+            # DeviceStore.__iter__ yields OlmDevice objects directly.
+            for device in device_store:
+                if getattr(device, "device_id", None) == own_device:
+                    continue
+                if not getattr(device, "verified", False):
+                    client.verify_device(device)
+                    trusted_count += 1
+        except Exception as exc:
+            logger.debug("Matrix: auto-trust error: %s", exc)
+
+        if trusted_count:
+            logger.info("Matrix: auto-trusted %d new device(s)", trusted_count)
+
+    async def _retry_pending_decryptions(self) -> None:
+        """Retry decrypting buffered MegolmEvents after new keys arrive."""
+        import nio
+
+        client = self._client
+        if not client or not self._pending_megolm:
+            return
+
+        now = time.time()
+        still_pending: list = []
+
+        for room, event, ts in self._pending_megolm:
+            # Drop events that have aged past the TTL.
+            if now - ts > _PENDING_EVENT_TTL:
+                logger.debug(
+                    "Matrix: dropping expired pending event %s (age %.0fs)",
+                    getattr(event, "event_id", "?"), now - ts,
+                )
+                continue
+
+            try:
+                decrypted = client.decrypt_event(event)
+            except Exception:
+                # Still missing the key — keep in buffer.
+                still_pending.append((room, event, ts))
+                continue
+
+            if isinstance(decrypted, nio.MegolmEvent):
+                # decrypt_event returned the same undecryptable event.
+                still_pending.append((room, event, ts))
+                continue
+
+            logger.info(
+                "Matrix: decrypted buffered event %s (%s)",
+                getattr(event, "event_id", "?"),
+                type(decrypted).__name__,
+            )
+
+            # Route to the appropriate handler based on decrypted type.
+            try:
+                if isinstance(decrypted, nio.RoomMessageText):
+                    await self._on_room_message(room, decrypted)
+                elif isinstance(
+                    decrypted,
+                    (nio.RoomMessageImage, nio.RoomMessageAudio,
+                     nio.RoomMessageVideo, nio.RoomMessageFile),
+                ):
+                    await self._on_room_message_media(room, decrypted)
+                else:
+                    logger.debug(
+                        "Matrix: decrypted event %s has unhandled type %s",
+                        getattr(event, "event_id", "?"),
+                        type(decrypted).__name__,
+                    )
+            except Exception as exc:
+                logger.warning(
+                    "Matrix: error processing decrypted event %s: %s",
+                    getattr(event, "event_id", "?"), exc,
+                )
+
+        self._pending_megolm = still_pending
+
     # ------------------------------------------------------------------
     # Event callbacks
     # ------------------------------------------------------------------
@@ -559,18 +854,38 @@ class MatrixAdapter(BasePlatformAdapter):
         if event.sender == self._user_id:
             return
 
+        # Deduplicate by event ID (nio can fire the same event more than once).
+        if self._is_duplicate_event(getattr(event, "event_id", None)):
+            return
+
         # Startup grace: ignore old messages from initial sync.
         event_ts = getattr(event, "server_timestamp", 0) / 1000.0
         if event_ts and event_ts < self._startup_ts - _STARTUP_GRACE_SECONDS:
             return
 
-        # Handle decrypted MegolmEvents — extract the inner event.
+        # Handle undecryptable MegolmEvents: request the missing session key
+        # and buffer the event for retry once the key arrives.
         if isinstance(event, nio.MegolmEvent):
-            # Failed to decrypt.
             logger.warning(
-                "Matrix: could not decrypt event %s in %s",
+                "Matrix: could not decrypt event %s in %s — requesting key",
                 event.event_id, room.room_id,
             )
+
+            # Ask other devices in the room to forward the session key.
+            try:
+                resp = await self._client.request_room_key(event)
+                if hasattr(resp, "event_id") or not isinstance(resp, Exception):
+                    logger.debug(
+                        "Matrix: room key request sent for session %s",
+                        getattr(event, "session_id", "?"),
+                    )
+            except Exception as exc:
+                logger.debug("Matrix: room key request failed: %s", exc)
+
+            # Buffer for retry on next maintenance cycle.
+            self._pending_megolm.append((room, event, time.time()))
+            if len(self._pending_megolm) > _MAX_PENDING_EVENTS:
+                self._pending_megolm = self._pending_megolm[-_MAX_PENDING_EVENTS:]
             return
 
         # Skip edits (m.replace relation).
@@ -594,6 +909,30 @@ class MatrixAdapter(BasePlatformAdapter):
         if relates_to.get("rel_type") == "m.thread":
             thread_id = relates_to.get("event_id")
 
+        # Require-mention gating.
+        if not is_dm:
+            free_rooms_raw = os.getenv("MATRIX_FREE_RESPONSE_ROOMS", "")
+            free_rooms = {r.strip() for r in free_rooms_raw.split(",") if r.strip()}
+            require_mention = os.getenv("MATRIX_REQUIRE_MENTION", "true").lower() not in ("false", "0", "no")
+            is_free_room = room.room_id in free_rooms
+            in_bot_thread = bool(thread_id and thread_id in self._bot_participated_threads)
+
+            formatted_body = source_content.get("formatted_body")
+            if require_mention and not is_free_room and not in_bot_thread:
+                if not self._is_bot_mentioned(body, formatted_body):
+                    return
+
+        # Strip mention from body when present (including in DMs).
+        if self._is_bot_mentioned(body, source_content.get("formatted_body")):
+            body = self._strip_mention(body)
+
+        # Auto-thread: create a thread for non-DM, non-threaded messages.
+        if not is_dm and not thread_id:
+            auto_thread = os.getenv("MATRIX_AUTO_THREAD", "true").lower() in ("true", "1", "yes")
+            if auto_thread:
+                thread_id = event.event_id
+                self._track_thread(thread_id)
+
         # Reply-to detection.
         reply_to = None
         in_reply_to = relates_to.get("m.in_reply_to", {})
@@ -638,6 +977,9 @@ class MatrixAdapter(BasePlatformAdapter):
             reply_to_message_id=reply_to,
         )
 
+        if thread_id:
+            self._track_thread(thread_id)
+
         await self.handle_message(msg_event)
 
     async def _on_room_message_media(self, room: Any, event: Any) -> None:
@@ -648,6 +990,10 @@ class MatrixAdapter(BasePlatformAdapter):
         if event.sender == self._user_id:
             return
 
+        # Deduplicate by event ID.
+        if self._is_duplicate_event(getattr(event, "event_id", None)):
+            return
+
         # Startup grace.
         event_ts = getattr(event, "server_timestamp", 0) / 1000.0
         if event_ts and event_ts < self._startup_ts - _STARTUP_GRACE_SECONDS:
@@ -669,11 +1015,19 @@ class MatrixAdapter(BasePlatformAdapter):
         event_mimetype = (content_info.get("info") or {}).get("mimetype", "")
         media_type = "application/octet-stream"
         msg_type = MessageType.DOCUMENT
+        is_voice_message = False
+        
         if isinstance(event, nio.RoomMessageImage):
             msg_type = MessageType.PHOTO
             media_type = event_mimetype or "image/png"
         elif isinstance(event, nio.RoomMessageAudio):
-            msg_type = MessageType.AUDIO
+            # Check for MSC3245 voice flag: org.matrix.msc3245.voice: {}
+            source_content = getattr(event, "source", {}).get("content", {})
+            if source_content.get("org.matrix.msc3245.voice") is not None:
+                is_voice_message = True
+                msg_type = MessageType.VOICE
+            else:
+                msg_type = MessageType.AUDIO
             media_type = event_mimetype or "audio/ogg"
         elif isinstance(event, nio.RoomMessageVideo):
             msg_type = MessageType.VIDEO
@@ -681,6 +1035,24 @@ class MatrixAdapter(BasePlatformAdapter):
         elif event_mimetype:
             media_type = event_mimetype
 
+        # For images, download and cache locally so vision tools can access them.
+        # Matrix MXC URLs require authentication, so direct URL access fails.
+        cached_path = None
+        if msg_type == MessageType.PHOTO and url:
+            try:
+                ext_map = {
+                    "image/jpeg": ".jpg", "image/png": ".png",
+                    "image/gif": ".gif", "image/webp": ".webp",
+                }
+                ext = ext_map.get(event_mimetype, ".jpg")
+                download_resp = await self._client.download(url)
+                if isinstance(download_resp, nio.DownloadResponse):
+                    from gateway.platforms.base import cache_image_from_bytes
+                    cached_path = cache_image_from_bytes(download_resp.body, ext=ext)
+                    logger.info("[Matrix] Cached user image at %s", cached_path)
+            except Exception as e:
+                logger.warning("[Matrix] Failed to cache image: %s", e)
+
         is_dm = self._dm_rooms.get(room.room_id, False)
         if not is_dm and room.member_count == 2:
             is_dm = True
@@ -693,6 +1065,55 @@ class MatrixAdapter(BasePlatformAdapter):
         if relates_to.get("rel_type") == "m.thread":
             thread_id = relates_to.get("event_id")
 
+        # Require-mention gating (media messages).
+        if not is_dm:
+            free_rooms_raw = os.getenv("MATRIX_FREE_RESPONSE_ROOMS", "")
+            free_rooms = {r.strip() for r in free_rooms_raw.split(",") if r.strip()}
+            require_mention = os.getenv("MATRIX_REQUIRE_MENTION", "true").lower() not in ("false", "0", "no")
+            is_free_room = room.room_id in free_rooms
+            in_bot_thread = bool(thread_id and thread_id in self._bot_participated_threads)
+
+            if require_mention and not is_free_room and not in_bot_thread:
+                formatted_body = source_content.get("formatted_body")
+                if not self._is_bot_mentioned(body, formatted_body):
+                    return
+
+        # Strip mention from body when present (including in DMs).
+        if self._is_bot_mentioned(body, source_content.get("formatted_body")):
+            body = self._strip_mention(body)
+
+        # Auto-thread: create a thread for non-DM, non-threaded messages.
+        if not is_dm and not thread_id:
+            auto_thread = os.getenv("MATRIX_AUTO_THREAD", "true").lower() in ("true", "1", "yes")
+            if auto_thread:
+                thread_id = event.event_id
+                self._track_thread(thread_id)
+
+        # For voice messages, cache audio locally for transcription tools.
+        # Use the authenticated nio client to download (Matrix requires auth for media).
+        media_urls = [http_url] if http_url else None
+        media_types = [media_type] if http_url else None
+        
+        if is_voice_message and url and url.startswith("mxc://"):
+            try:
+                import nio
+                from gateway.platforms.base import cache_audio_from_bytes
+                
+                resp = await self._client.download(mxc=url)
+                if isinstance(resp, nio.MemoryDownloadResponse):
+                    # Extract extension from mimetype or default to .ogg
+                    ext = ".ogg"
+                    if media_type and "/" in media_type:
+                        subtype = media_type.split("/")[1]
+                        ext = f".{subtype}" if subtype else ".ogg"
+                    local_path = cache_audio_from_bytes(resp.body, ext)
+                    media_urls = [local_path]
+                    logger.debug("Matrix: cached voice message to %s", local_path)
+                else:
+                    logger.warning("Matrix: failed to download voice: %s", getattr(resp, "message", resp))
+            except Exception as e:
+                logger.warning("Matrix: failed to cache voice message, using HTTP URL: %s", e)
+
         source = self.build_source(
             chat_id=room.room_id,
             chat_type=chat_type,
@@ -701,16 +1122,24 @@ class MatrixAdapter(BasePlatformAdapter):
             thread_id=thread_id,
         )
 
+        # Use cached local path for images (voice messages already handled above).
+        if cached_path:
+            media_urls = [cached_path]
+        media_types = [media_type] if media_urls else None
+
         msg_event = MessageEvent(
             text=body,
             message_type=msg_type,
             source=source,
             raw_message=getattr(event, "source", {}),
             message_id=event.event_id,
-            media_urls=[http_url] if http_url else None,
-            media_types=[media_type] if http_url else None,
+            media_urls=media_urls,
+            media_types=media_types,
         )
 
+        if thread_id:
+            self._track_thread(thread_id)
+
         await self.handle_message(msg_event)
 
     async def _on_invite(self, room: Any, event: Any) -> None:
@@ -798,6 +1227,82 @@ class MatrixAdapter(BasePlatformAdapter):
             for rid in self._joined_rooms
         }
 
+    # ------------------------------------------------------------------
+    # Thread participation tracking
+    # ------------------------------------------------------------------
+
+    @staticmethod
+    def _thread_state_path() -> Path:
+        """Path to the persisted thread participation set."""
+        from hermes_cli.config import get_hermes_home
+        return get_hermes_home() / "matrix_threads.json"
+
+    @classmethod
+    def _load_participated_threads(cls) -> set:
+        """Load persisted thread IDs from disk."""
+        path = cls._thread_state_path()
+        try:
+            if path.exists():
+                data = json.loads(path.read_text(encoding="utf-8"))
+                if isinstance(data, list):
+                    return set(data)
+        except Exception as e:
+            logger.debug("Could not load matrix thread state: %s", e)
+        return set()
+
+    def _save_participated_threads(self) -> None:
+        """Persist the current thread set to disk (best-effort)."""
+        path = self._thread_state_path()
+        try:
+            thread_list = list(self._bot_participated_threads)
+            if len(thread_list) > self._MAX_TRACKED_THREADS:
+                thread_list = thread_list[-self._MAX_TRACKED_THREADS:]
+                self._bot_participated_threads = set(thread_list)
+            path.parent.mkdir(parents=True, exist_ok=True)
+            path.write_text(json.dumps(thread_list), encoding="utf-8")
+        except Exception as e:
+            logger.debug("Could not save matrix thread state: %s", e)
+
+    def _track_thread(self, thread_id: str) -> None:
+        """Add a thread to the participation set and persist."""
+        if thread_id not in self._bot_participated_threads:
+            self._bot_participated_threads.add(thread_id)
+            self._save_participated_threads()
+
+    # ------------------------------------------------------------------
+    # Mention detection helpers
+    # ------------------------------------------------------------------
+
+    def _is_bot_mentioned(self, body: str, formatted_body: Optional[str] = None) -> bool:
+        """Return True if the bot is mentioned in the message."""
+        if not body and not formatted_body:
+            return False
+        # Check for full @user:server in body
+        if self._user_id and self._user_id in body:
+            return True
+        # Check for localpart with word boundaries (case-insensitive)
+        if self._user_id and ":" in self._user_id:
+            localpart = self._user_id.split(":")[0].lstrip("@")
+            if localpart and re.search(r'\b' + re.escape(localpart) + r'\b', body, re.IGNORECASE):
+                return True
+        # Check formatted_body for Matrix pill
+        if formatted_body and self._user_id:
+            if f"matrix.to/#/{self._user_id}" in formatted_body:
+                return True
+        return False
+
+    def _strip_mention(self, body: str) -> str:
+        """Remove bot mention from message body."""
+        # Remove full @user:server
+        if self._user_id:
+            body = body.replace(self._user_id, "")
+        # If still contains localpart mention, remove it
+        if self._user_id and ":" in self._user_id:
+            localpart = self._user_id.split(":")[0].lstrip("@")
+            if localpart:
+                body = re.sub(r'\b' + re.escape(localpart) + r'\b', '', body, flags=re.IGNORECASE)
+        return body.strip()
+
     def _get_display_name(self, room: Any, user_id: str) -> str:
         """Get a user's display name in a room, falling back to user_id."""
         if room and hasattr(room, "users"):

gateway/platforms/mattermost.py

@@ -20,7 +20,7 @@ import os
 import re
 import time
 from pathlib import Path
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any, Dict, List, Optional
 
 from gateway.config import Platform, PlatformConfig
 from gateway.platforms.base import (
@@ -116,7 +116,7 @@ class MattermostAdapter(BasePlatformAdapter):
         import aiohttp
         url = f"{self._base_url}/api/v4/{path.lstrip('/')}"
         try:
-            async with self._session.get(url, headers=self._headers()) as resp:
+            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])
@@ -134,7 +134,8 @@ class MattermostAdapter(BasePlatformAdapter):
         url = f"{self._base_url}/api/v4/{path.lstrip('/')}"
         try:
             async with self._session.post(
-                url, headers=self._headers(), json=payload
+                url, headers=self._headers(), json=payload,
+                timeout=aiohttp.ClientTimeout(total=30)
             ) as resp:
                 if resp.status >= 400:
                     body = await resp.text()
@@ -180,7 +181,7 @@ class MattermostAdapter(BasePlatformAdapter):
             content_type=content_type,
         )
         headers = {"Authorization": f"Bearer {self._token}"}
-        async with self._session.post(url, headers=headers, data=form) as resp:
+        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])
@@ -201,7 +202,9 @@ class MattermostAdapter(BasePlatformAdapter):
             logger.error("Mattermost: URL or token not configured")
             return False
 
-        self._session = aiohttp.ClientSession()
+        self._session = aiohttp.ClientSession(
+            timeout=aiohttp.ClientTimeout(total=30)
+        )
         self._closing = False
 
         # Verify credentials and fetch bot identity.
@@ -404,18 +407,38 @@ class MattermostAdapter(BasePlatformAdapter):
         kind: str = "file",
     ) -> SendResult:
         """Download a URL and upload it as a file attachment."""
+        import asyncio
         import aiohttp
-        try:
-            async with self._session.get(url, timeout=aiohttp.ClientTimeout(total=30)) as resp:
-                if resp.status >= 400:
-                    # Fall back to sending the URL as text.
-                    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"
-                # Derive filename from URL.
-                fname = url.rsplit("/", 1)[-1].split("?")[0] or f"{kind}.png"
-        except Exception as exc:
-            logger.warning("Mattermost: failed to download %s: %s", url, exc)
+
+        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)
@@ -580,6 +603,42 @@ class MattermostAdapter(BasePlatformAdapter):
         # 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
@@ -617,16 +676,16 @@ class MattermostAdapter(BasePlatformAdapter):
                         if mime.startswith("image/"):
                             local_path = cache_image_from_bytes(file_data, ext or ".png")
                             media_urls.append(local_path)
-                            media_types.append("image")
+                            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("audio")
+                            media_types.append(mime)
                         else:
                             local_path = cache_document_from_bytes(file_data, fname)
                             media_urls.append(local_path)
-                            media_types.append("document")
+                            media_types.append(mime)
                     else:
                         logger.warning("Mattermost: failed to download file %s: HTTP %s", fid, resp.status)
             except Exception as exc:

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
 
@@ -184,6 +184,8 @@ class SignalAdapter(BasePlatformAdapter):
         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")
@@ -198,6 +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
@@ -245,6 +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")
 
     # ------------------------------------------------------------------
@@ -253,7 +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:
@@ -279,6 +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()
@@ -344,7 +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
@@ -478,7 +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)
@@ -513,12 +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)

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(
@@ -788,6 +897,11 @@ 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", "")
+
+        # 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.
@@ -819,34 +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

@@ -17,12 +17,11 @@ Gateway-specific env vars:
 
 import asyncio
 import base64
-import json
 import logging
 import os
 import re
 import urllib.parse
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, Optional
 
 from gateway.config import Platform, PlatformConfig
 from gateway.platforms.base import (
@@ -107,7 +106,9 @@ class SmsAdapter(BasePlatformAdapter):
         await self._runner.setup()
         site = web.TCPSite(self._runner, "0.0.0.0", self._webhook_port)
         await site.start()
-        self._http_session = aiohttp.ClientSession()
+        self._http_session = aiohttp.ClientSession(
+            timeout=aiohttp.ClientTimeout(total=30),
+        )
         self._running = True
 
         logger.info(
@@ -145,7 +146,9 @@ class SmsAdapter(BasePlatformAdapter):
             "Authorization": self._basic_auth_header(),
         }
 
-        session = self._http_session or aiohttp.ClientSession()
+        session = self._http_session or aiohttp.ClientSession(
+            timeout=aiohttp.ClientTimeout(total=30),
+        )
         try:
             for chunk in chunks:
                 form_data = aiohttp.FormData()
@@ -262,7 +265,9 @@ class SmsAdapter(BasePlatformAdapter):
         )
 
         # Non-blocking: Twilio expects a fast response
-        asyncio.create_task(self.handle_message(event))
+        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(

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

@@ -27,6 +27,7 @@ import hashlib
 import hmac
 import json
 import logging
+import os
 import re
 import subprocess
 import time
@@ -53,6 +54,7 @@ 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:
@@ -68,7 +70,10 @@ class WebhookAdapter(BasePlatformAdapter):
         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._routes: Dict[str, dict] = config.extra.get("routes", {})
+        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()
@@ -96,6 +101,9 @@ class WebhookAdapter(BasePlatformAdapter):
     # ------------------------------------------------------------------
 
     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)
@@ -110,6 +118,17 @@ class WebhookAdapter(BasePlatformAdapter):
         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)
@@ -182,8 +201,46 @@ class WebhookAdapter(BasePlatformAdapter):
         """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)
 
@@ -363,7 +420,9 @@ class WebhookAdapter(BasePlatformAdapter):
         )
 
         # Non-blocking — return 202 Accepted immediately
-        asyncio.create_task(self.handle_message(event))
+        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(
             {

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,
 )
@@ -134,12 +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:
         """
@@ -158,6 +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():
@@ -196,9 +350,10 @@ class WhatsAppAdapter(BasePlatformAdapter):
                             bridge_status = data.get("status", "unknown")
                             if bridge_status == "connected":
                                 print(f"[{self.name}] Using existing bridge (status: {bridge_status})")
-                                self._running = True
+                                self._mark_connected()
                                 self._bridge_process = None  # Not managed by us
-                                asyncio.create_task(self._poll_messages())
+                                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")
@@ -303,14 +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
@@ -324,6 +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:
@@ -351,10 +532,32 @@ class WhatsAppAdapter(BasePlatformAdapter):
         else:
             # Bridge was not started by us, don't kill it
             print(f"[{self.name}] Disconnecting (external bridge left running)")
-        
-        self._running = False
+
+        # 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(
@@ -365,41 +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://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 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))
 
@@ -410,25 +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://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)
+            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))
 
@@ -441,8 +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
 
@@ -459,22 +663,21 @@ class WhatsAppAdapter(BasePlatformAdapter):
             if file_name:
                 payload["fileName"] = file_name
 
-            async with aiohttp.ClientSession() as session:
-                async with session.post(
-                    f"http://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)
+            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))
@@ -499,6 +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)
@@ -509,6 +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)
@@ -520,6 +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(
@@ -529,41 +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://127.0.0.1:{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://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", []),
-                        }
+
+            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)
         
@@ -571,28 +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://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)
+                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)
             
@@ -601,6 +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"):
@@ -627,7 +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 = []
@@ -658,12 +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,
@@ -674,4 +939,3 @@ class WhatsAppAdapter(BasePlatformAdapter):
         except Exception as e:
             print(f"[{self.name}] Error building event: {e}")
             return None
-

gateway/session.py

@@ -13,15 +13,21 @@ 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
 # ---------------------------------------------------------------------------
@@ -59,7 +65,7 @@ def _looks_like_phone(value: str) -> bool:
 from .config import (
     Platform,
     GatewayConfig,
-    SessionResetPolicy,
+    SessionResetPolicy,  # noqa: F401 — re-exported via gateway/__init__.py
     HomeChannel,
 )
 
@@ -355,6 +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 = {
@@ -373,6 +387,7 @@ class SessionEntry:
             "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()
@@ -408,6 +423,7 @@ class SessionEntry:
             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),
         )
 
 
@@ -469,10 +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
@@ -484,12 +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:
@@ -502,7 +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:
@@ -554,7 +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)
@@ -573,16 +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,
@@ -590,14 +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(
@@ -610,9 +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).
@@ -632,178 +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,
-        cache_read_tokens: int = 0,
-        cache_write_tokens: int = 0,
-        last_prompt_tokens: int = None,
-        model: str = None,
-        estimated_cost_usd: Optional[float] = None,
-        cost_status: Optional[str] = None,
-        cost_source: Optional[str] = None,
-        provider: Optional[str] = None,
-        base_url: Optional[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
-            entry.cache_read_tokens += cache_read_tokens
-            entry.cache_write_tokens += cache_write_tokens
-            if last_prompt_tokens is not None:
-                entry.last_prompt_tokens = last_prompt_tokens
-            if estimated_cost_usd is not None:
-                entry.estimated_cost_usd += estimated_cost_usd
-            if cost_status:
-                entry.cost_status = cost_status
-            entry.total_tokens = (
-                entry.input_tokens
-                + entry.output_tokens
-                + entry.cache_read_tokens
-                + entry.cache_write_tokens
+
+        # 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
             )
-            self._save()
-            
-            if self._db:
+            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=input_tokens,
-                        output_tokens=output_tokens,
-                        cache_read_tokens=cache_read_tokens,
-                        cache_write_tokens=cache_write_tokens,
-                        estimated_cost_usd=estimated_cost_usd,
-                        cost_status=cost_status,
-                        cost_source=cost_source,
-                        billing_provider=provider,
-                        billing_base_url=base_url,
-                        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]:
@@ -814,52 +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:
@@ -905,13 +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)
@@ -924,35 +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:
-                    try:
-                        messages.append(json.loads(line))
-                    except json.JSONDecodeError:
-                        logger.warning(
-                            "Skipping corrupt line in transcript %s: %s",
-                            session_id, line[:120],
-                        )
-        
-        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"
 
 
@@ -274,6 +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)
@@ -314,6 +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``.
 

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

@@ -174,12 +174,12 @@ class GatewayStreamConsumer:
                         self._already_sent = True
                         self._last_sent_text = text
                     else:
-                        # Edit not supported by this adapter — stop streaming,
-                        # let the normal send path handle the final response.
-                        # Without this guard, adapters like Signal/Email would
-                        # flood the chat with a new message every edit_interval.
+                        # 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.

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__ = "0.4.0"
-__release_date__ = "2026.3.18"
+__version__ = "0.7.0"
+__release_date__ = "2026.4.3"

hermes_cli/auth.py

@@ -160,7 +160,7 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
         id="alibaba",
         name="Alibaba Cloud (DashScope)",
         auth_type="api_key",
-        inference_base_url="https://dashscope-intl.aliyuncs.com/apps/anthropic",
+        inference_base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
         api_key_env_vars=("DASHSCOPE_API_KEY",),
         base_url_env_var="DASHSCOPE_BASE_URL",
     ),
@@ -199,9 +199,13 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
     "opencode-go": ProviderConfig(
         id="opencode-go",
         name="OpenCode Go",
-        auth_type="***",
+        auth_type="api_key",
+        # OpenCode Go mixes API surfaces by model:
+        # - GLM / Kimi use OpenAI-compatible chat completions under /v1
+        # - MiniMax models use Anthropic Messages under /v1/messages
+        # Keep the provider base at /v1 and select api_mode per-model.
         inference_base_url="https://opencode.ai/zen/go/v1",
-        api_key_env_vars=("OPEN...",),
+        api_key_env_vars=("OPENCODE_GO_API_KEY",),
         base_url_env_var="OPENCODE_GO_BASE_URL",
     ),
     "kilocode": ProviderConfig(
@@ -212,6 +216,14 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
         api_key_env_vars=("KILOCODE_API_KEY",),
         base_url_env_var="KILOCODE_BASE_URL",
     ),
+    "huggingface": ProviderConfig(
+        id="huggingface",
+        name="Hugging Face",
+        auth_type="api_key",
+        inference_base_url="https://router.huggingface.co/v1",
+        api_key_env_vars=("HF_TOKEN",),
+        base_url_env_var="HF_BASE_URL",
+    ),
 }
 
 
@@ -278,6 +290,33 @@ def _try_gh_cli_token() -> Optional[str]:
     return None
 
 
+_PLACEHOLDER_SECRET_VALUES = {
+    "*",
+    "**",
+    "***",
+    "changeme",
+    "your_api_key",
+    "your-api-key",
+    "placeholder",
+    "example",
+    "dummy",
+    "null",
+    "none",
+}
+
+
+def has_usable_secret(value: Any, *, min_length: int = 4) -> bool:
+    """Return True when a configured secret looks usable, not empty/placeholder."""
+    if not isinstance(value, str):
+        return False
+    cleaned = value.strip()
+    if len(cleaned) < min_length:
+        return False
+    if cleaned.lower() in _PLACEHOLDER_SECRET_VALUES:
+        return False
+    return True
+
+
 def _resolve_api_key_provider_secret(
     provider_id: str, pconfig: ProviderConfig
 ) -> tuple[str, str]:
@@ -297,7 +336,7 @@ def _resolve_api_key_provider_secret(
 
     for env_var in pconfig.api_key_env_vars:
         val = os.getenv(env_var, "").strip()
-        if val:
+        if has_usable_secret(val):
             return val, env_var
 
     return "", ""
@@ -510,7 +549,11 @@ def _load_auth_store(auth_file: Optional[Path] = None) -> Dict[str, Any]:
     except Exception:
         return {"version": AUTH_STORE_VERSION, "providers": {}}
 
-    if isinstance(raw, dict) and isinstance(raw.get("providers"), dict):
+    if isinstance(raw, dict) and (
+        isinstance(raw.get("providers"), dict)
+        or isinstance(raw.get("credential_pool"), dict)
+    ):
+        raw.setdefault("providers", {})
         return raw
 
     # Migrate from PR's "systems" format if present
@@ -578,6 +621,30 @@ def _save_provider_state(auth_store: Dict[str, Any], provider_id: str, state: Di
     auth_store["active_provider"] = provider_id
 
 
+def read_credential_pool(provider_id: Optional[str] = None) -> Dict[str, Any]:
+    """Return the persisted credential pool, or one provider slice."""
+    auth_store = _load_auth_store()
+    pool = auth_store.get("credential_pool")
+    if not isinstance(pool, dict):
+        pool = {}
+    if provider_id is None:
+        return dict(pool)
+    provider_entries = pool.get(provider_id)
+    return list(provider_entries) if isinstance(provider_entries, list) else []
+
+
+def write_credential_pool(provider_id: str, entries: List[Dict[str, Any]]) -> Path:
+    """Persist one provider's credential pool under auth.json."""
+    with _auth_store_lock():
+        auth_store = _load_auth_store()
+        pool = auth_store.get("credential_pool")
+        if not isinstance(pool, dict):
+            pool = {}
+            auth_store["credential_pool"] = pool
+        pool[provider_id] = list(entries)
+        return _save_auth_store(auth_store)
+
+
 def get_provider_auth_state(provider_id: str) -> Optional[Dict[str, Any]]:
     """Return persisted auth state for a provider, or None."""
     auth_store = _load_auth_store()
@@ -603,10 +670,25 @@ def clear_provider_auth(provider_id: Optional[str] = None) -> bool:
             return False
 
         providers = auth_store.get("providers", {})
-        if target not in providers:
-            return False
+        if not isinstance(providers, dict):
+            providers = {}
+            auth_store["providers"] = providers
 
-        del providers[target]
+        pool = auth_store.get("credential_pool")
+        if not isinstance(pool, dict):
+            pool = {}
+            auth_store["credential_pool"] = pool
+
+        cleared = False
+        if target in providers:
+            del providers[target]
+            cleared = True
+        if target in pool:
+            del pool[target]
+            cleared = True
+
+        if not cleared:
+            return False
         if auth_store.get("active_provider") == target:
             auth_store["active_provider"] = None
         _save_auth_store(auth_store)
@@ -658,13 +740,20 @@ def resolve_provider(
         "github-copilot-acp": "copilot-acp", "copilot-acp-agent": "copilot-acp",
         "aigateway": "ai-gateway", "vercel": "ai-gateway", "vercel-ai-gateway": "ai-gateway",
         "opencode": "opencode-zen", "zen": "opencode-zen",
+        "hf": "huggingface", "hugging-face": "huggingface", "huggingface-hub": "huggingface",
         "go": "opencode-go", "opencode-go-sub": "opencode-go",
         "kilo": "kilocode", "kilo-code": "kilocode", "kilo-gateway": "kilocode",
+        # Local server aliases — route through the generic custom provider
+        "lmstudio": "custom", "lm-studio": "custom", "lm_studio": "custom",
+        "ollama": "custom", "vllm": "custom", "llamacpp": "custom",
+        "llama.cpp": "custom", "llama-cpp": "custom",
     }
     normalized = _PROVIDER_ALIASES.get(normalized, normalized)
 
-    if normalized in {"openrouter", "custom"}:
+    if normalized == "openrouter":
         return "openrouter"
+    if normalized == "custom":
+        return "custom"
     if normalized in PROVIDER_REGISTRY:
         return normalized
     if normalized != "auto":
@@ -688,7 +777,7 @@ def resolve_provider(
     except Exception as e:
         logger.debug("Could not detect active auth provider: %s", e)
 
-    if os.getenv("OPENAI_API_KEY") or os.getenv("OPENROUTER_API_KEY"):
+    if has_usable_secret(os.getenv("OPENAI_API_KEY")) or has_usable_secret(os.getenv("OPENROUTER_API_KEY")):
         return "openrouter"
 
     # Auto-detect API-key providers by checking their env vars
@@ -701,10 +790,15 @@ def resolve_provider(
         if pid == "copilot":
             continue
         for env_var in pconfig.api_key_env_vars:
-            if os.getenv(env_var, "").strip():
+            if has_usable_secret(os.getenv(env_var, "")):
                 return pid
 
-    return "openrouter"
+    raise AuthError(
+        "No inference provider configured. Run 'hermes model' to choose a "
+        "provider and model, or set an API key (OPENROUTER_API_KEY, "
+        "OPENAI_API_KEY, etc.) in ~/.hermes/.env.",
+        code="no_provider_configured",
+    )
 
 
 # =============================================================================
@@ -851,15 +945,14 @@ def _save_codex_tokens(tokens: Dict[str, str], last_refresh: str = None) -> None
         _save_auth_store(auth_store)
 
 
-def _refresh_codex_auth_tokens(
-    tokens: Dict[str, str],
-    timeout_seconds: float,
-) -> Dict[str, str]:
-    """Refresh Codex access token using the refresh token.
-    
-    Saves the new tokens to Hermes auth store automatically.
-    """
-    refresh_token = tokens.get("refresh_token")
+def refresh_codex_oauth_pure(
+    access_token: str,
+    refresh_token: str,
+    *,
+    timeout_seconds: float = 20.0,
+) -> Dict[str, Any]:
+    """Refresh Codex OAuth tokens without mutating Hermes auth state."""
+    del access_token  # Access token is only used by callers to decide whether to refresh.
     if not isinstance(refresh_token, str) or not refresh_token.strip():
         raise AuthError(
             "Codex auth is missing refresh_token. Run `hermes login` to re-authenticate.",
@@ -914,8 +1007,8 @@ def _refresh_codex_auth_tokens(
             relogin_required=True,
         ) from exc
 
-    access_token = refresh_payload.get("access_token")
-    if not isinstance(access_token, str) or not access_token.strip():
+    refreshed_access = refresh_payload.get("access_token")
+    if not isinstance(refreshed_access, str) or not refreshed_access.strip():
         raise AuthError(
             "Codex token refresh response was missing access_token.",
             provider="openai-codex",
@@ -923,11 +1016,33 @@ def _refresh_codex_auth_tokens(
             relogin_required=True,
         )
 
-    updated_tokens = dict(tokens)
-    updated_tokens["access_token"] = access_token.strip()
+    updated = {
+        "access_token": refreshed_access.strip(),
+        "refresh_token": refresh_token.strip(),
+        "last_refresh": datetime.now(timezone.utc).isoformat().replace("+00:00", "Z"),
+    }
     next_refresh = refresh_payload.get("refresh_token")
     if isinstance(next_refresh, str) and next_refresh.strip():
-        updated_tokens["refresh_token"] = next_refresh.strip()
+        updated["refresh_token"] = next_refresh.strip()
+    return updated
+
+
+def _refresh_codex_auth_tokens(
+    tokens: Dict[str, str],
+    timeout_seconds: float,
+) -> Dict[str, str]:
+    """Refresh Codex access token using the refresh token.
+    
+    Saves the new tokens to Hermes auth store automatically.
+    """
+    refreshed = refresh_codex_oauth_pure(
+        str(tokens.get("access_token", "") or ""),
+        str(tokens.get("refresh_token", "") or ""),
+        timeout_seconds=timeout_seconds,
+    )
+    updated_tokens = dict(tokens)
+    updated_tokens["access_token"] = refreshed["access_token"]
+    updated_tokens["refresh_token"] = refreshed["refresh_token"]
 
     _save_codex_tokens(updated_tokens)
     return updated_tokens
@@ -1266,6 +1381,205 @@ def _agent_key_is_usable(state: Dict[str, Any], min_ttl_seconds: int) -> bool:
     return not _is_expiring(state.get("agent_key_expires_at"), min_ttl_seconds)
 
 
+def resolve_nous_access_token(
+    *,
+    timeout_seconds: float = 15.0,
+    insecure: Optional[bool] = None,
+    ca_bundle: Optional[str] = None,
+    refresh_skew_seconds: int = ACCESS_TOKEN_REFRESH_SKEW_SECONDS,
+) -> str:
+    """Resolve a refresh-aware Nous Portal access token for managed tool gateways."""
+    with _auth_store_lock():
+        auth_store = _load_auth_store()
+        state = _load_provider_state(auth_store, "nous")
+
+        if not state:
+            raise AuthError(
+                "Hermes is not logged into Nous Portal.",
+                provider="nous",
+                relogin_required=True,
+            )
+
+        portal_base_url = (
+            _optional_base_url(state.get("portal_base_url"))
+            or os.getenv("HERMES_PORTAL_BASE_URL")
+            or os.getenv("NOUS_PORTAL_BASE_URL")
+            or DEFAULT_NOUS_PORTAL_URL
+        ).rstrip("/")
+        client_id = str(state.get("client_id") or DEFAULT_NOUS_CLIENT_ID)
+        verify = _resolve_verify(insecure=insecure, ca_bundle=ca_bundle, auth_state=state)
+
+        access_token = state.get("access_token")
+        refresh_token = state.get("refresh_token")
+        if not isinstance(access_token, str) or not access_token:
+            raise AuthError(
+                "No access token found for Nous Portal login.",
+                provider="nous",
+                relogin_required=True,
+            )
+
+        if not _is_expiring(state.get("expires_at"), refresh_skew_seconds):
+            return access_token
+
+        if not isinstance(refresh_token, str) or not refresh_token:
+            raise AuthError(
+                "Session expired and no refresh token is available.",
+                provider="nous",
+                relogin_required=True,
+            )
+
+        timeout = httpx.Timeout(timeout_seconds if timeout_seconds else 15.0)
+        with httpx.Client(
+            timeout=timeout,
+            headers={"Accept": "application/json"},
+            verify=verify,
+        ) as client:
+            refreshed = _refresh_access_token(
+                client=client,
+                portal_base_url=portal_base_url,
+                client_id=client_id,
+                refresh_token=refresh_token,
+            )
+
+        now = datetime.now(timezone.utc)
+        access_ttl = _coerce_ttl_seconds(refreshed.get("expires_in"))
+        state["access_token"] = refreshed["access_token"]
+        state["refresh_token"] = refreshed.get("refresh_token") or refresh_token
+        state["token_type"] = refreshed.get("token_type") or state.get("token_type") or "Bearer"
+        state["scope"] = refreshed.get("scope") or state.get("scope")
+        state["obtained_at"] = now.isoformat()
+        state["expires_in"] = access_ttl
+        state["expires_at"] = datetime.fromtimestamp(
+            now.timestamp() + access_ttl,
+            tz=timezone.utc,
+        ).isoformat()
+        state["portal_base_url"] = portal_base_url
+        state["client_id"] = client_id
+        state["tls"] = {
+            "insecure": verify is False,
+            "ca_bundle": verify if isinstance(verify, str) else None,
+        }
+        _save_provider_state(auth_store, "nous", state)
+        _save_auth_store(auth_store)
+        return state["access_token"]
+
+
+def refresh_nous_oauth_pure(
+    access_token: str,
+    refresh_token: str,
+    client_id: str,
+    portal_base_url: str,
+    inference_base_url: str,
+    *,
+    token_type: str = "Bearer",
+    scope: str = DEFAULT_NOUS_SCOPE,
+    obtained_at: Optional[str] = None,
+    expires_at: Optional[str] = None,
+    agent_key: Optional[str] = None,
+    agent_key_expires_at: Optional[str] = None,
+    min_key_ttl_seconds: int = DEFAULT_AGENT_KEY_MIN_TTL_SECONDS,
+    timeout_seconds: float = 15.0,
+    insecure: Optional[bool] = None,
+    ca_bundle: Optional[str] = None,
+    force_refresh: bool = False,
+    force_mint: bool = False,
+) -> Dict[str, Any]:
+    """Refresh Nous OAuth state without mutating auth.json."""
+    state: Dict[str, Any] = {
+        "access_token": access_token,
+        "refresh_token": refresh_token,
+        "client_id": client_id or DEFAULT_NOUS_CLIENT_ID,
+        "portal_base_url": (portal_base_url or DEFAULT_NOUS_PORTAL_URL).rstrip("/"),
+        "inference_base_url": (inference_base_url or DEFAULT_NOUS_INFERENCE_URL).rstrip("/"),
+        "token_type": token_type or "Bearer",
+        "scope": scope or DEFAULT_NOUS_SCOPE,
+        "obtained_at": obtained_at,
+        "expires_at": expires_at,
+        "agent_key": agent_key,
+        "agent_key_expires_at": agent_key_expires_at,
+        "tls": {
+            "insecure": bool(insecure),
+            "ca_bundle": ca_bundle,
+        },
+    }
+    verify = _resolve_verify(insecure=insecure, ca_bundle=ca_bundle, auth_state=state)
+    timeout = httpx.Timeout(timeout_seconds if timeout_seconds else 15.0)
+
+    with httpx.Client(timeout=timeout, headers={"Accept": "application/json"}, verify=verify) as client:
+        if force_refresh or _is_expiring(state.get("expires_at"), ACCESS_TOKEN_REFRESH_SKEW_SECONDS):
+            refreshed = _refresh_access_token(
+                client=client,
+                portal_base_url=state["portal_base_url"],
+                client_id=state["client_id"],
+                refresh_token=state["refresh_token"],
+            )
+            now = datetime.now(timezone.utc)
+            access_ttl = _coerce_ttl_seconds(refreshed.get("expires_in"))
+            state["access_token"] = refreshed["access_token"]
+            state["refresh_token"] = refreshed.get("refresh_token") or state["refresh_token"]
+            state["token_type"] = refreshed.get("token_type") or state.get("token_type") or "Bearer"
+            state["scope"] = refreshed.get("scope") or state.get("scope")
+            refreshed_url = _optional_base_url(refreshed.get("inference_base_url"))
+            if refreshed_url:
+                state["inference_base_url"] = refreshed_url
+            state["obtained_at"] = now.isoformat()
+            state["expires_in"] = access_ttl
+            state["expires_at"] = datetime.fromtimestamp(
+                now.timestamp() + access_ttl, tz=timezone.utc
+            ).isoformat()
+
+        if force_mint or not _agent_key_is_usable(state, max(60, int(min_key_ttl_seconds))):
+            mint_payload = _mint_agent_key(
+                client=client,
+                portal_base_url=state["portal_base_url"],
+                access_token=state["access_token"],
+                min_ttl_seconds=min_key_ttl_seconds,
+            )
+            now = datetime.now(timezone.utc)
+            state["agent_key"] = mint_payload.get("api_key")
+            state["agent_key_id"] = mint_payload.get("key_id")
+            state["agent_key_expires_at"] = mint_payload.get("expires_at")
+            state["agent_key_expires_in"] = mint_payload.get("expires_in")
+            state["agent_key_reused"] = bool(mint_payload.get("reused", False))
+            state["agent_key_obtained_at"] = now.isoformat()
+            minted_url = _optional_base_url(mint_payload.get("inference_base_url"))
+            if minted_url:
+                state["inference_base_url"] = minted_url
+
+    return state
+
+
+def refresh_nous_oauth_from_state(
+    state: Dict[str, Any],
+    *,
+    min_key_ttl_seconds: int = DEFAULT_AGENT_KEY_MIN_TTL_SECONDS,
+    timeout_seconds: float = 15.0,
+    force_refresh: bool = False,
+    force_mint: bool = False,
+) -> Dict[str, Any]:
+    """Refresh Nous OAuth from a state dict. Thin wrapper around refresh_nous_oauth_pure."""
+    tls = state.get("tls") or {}
+    return refresh_nous_oauth_pure(
+        state.get("access_token", ""),
+        state.get("refresh_token", ""),
+        state.get("client_id", "hermes-cli"),
+        state.get("portal_base_url", DEFAULT_NOUS_PORTAL_URL),
+        state.get("inference_base_url", DEFAULT_NOUS_INFERENCE_URL),
+        token_type=state.get("token_type", "Bearer"),
+        scope=state.get("scope", DEFAULT_NOUS_SCOPE),
+        obtained_at=state.get("obtained_at"),
+        expires_at=state.get("expires_at"),
+        agent_key=state.get("agent_key"),
+        agent_key_expires_at=state.get("agent_key_expires_at"),
+        min_key_ttl_seconds=min_key_ttl_seconds,
+        timeout_seconds=timeout_seconds,
+        insecure=tls.get("insecure"),
+        ca_bundle=tls.get("ca_bundle"),
+        force_refresh=force_refresh,
+        force_mint=force_mint,
+    )
+
+
 def resolve_nous_runtime_credentials(
     *,
     min_key_ttl_seconds: int = DEFAULT_AGENT_KEY_MIN_TTL_SECONDS,
@@ -1983,7 +2297,8 @@ def _login_openai_codex(args, pconfig: ProviderConfig) -> None:
     config_path = _update_config_for_provider("openai-codex", creds.get("base_url", DEFAULT_CODEX_BASE_URL))
     print()
     print("Login successful!")
-    print(f"  Auth state: ~/.hermes/auth.json")
+    from hermes_constants import display_hermes_home as _dhh
+    print(f"  Auth state: {_dhh()}/auth.json")
     print(f"  Config updated: {config_path} (model.provider=openai-codex)")
 
 
@@ -2027,9 +2342,9 @@ def _codex_device_code_login() -> Dict[str, Any]:
 
     # Step 2: Show user the code
     print("To continue, follow these steps:\n")
-    print(f"  1. Open this URL in your browser:")
+    print("  1. Open this URL in your browser:")
     print(f"     \033[94m{issuer}/codex/device\033[0m\n")
-    print(f"  2. Enter this code:")
+    print("  2. Enter this code:")
     print(f"     \033[94m{user_code}\033[0m\n")
     print("Waiting for sign-in... (press Ctrl+C to cancel)")
 
@@ -2132,34 +2447,36 @@ def _codex_device_code_login() -> Dict[str, Any]:
     }
 
 
-def _login_nous(args, pconfig: ProviderConfig) -> None:
-    """Nous Portal device authorization flow."""
+def _nous_device_code_login(
+    *,
+    portal_base_url: Optional[str] = None,
+    inference_base_url: Optional[str] = None,
+    client_id: Optional[str] = None,
+    scope: Optional[str] = None,
+    open_browser: bool = True,
+    timeout_seconds: float = 15.0,
+    insecure: bool = False,
+    ca_bundle: Optional[str] = None,
+    min_key_ttl_seconds: int = 5 * 60,
+) -> Dict[str, Any]:
+    """Run the Nous device-code flow and return full OAuth state without persisting."""
+    pconfig = PROVIDER_REGISTRY["nous"]
     portal_base_url = (
-        getattr(args, "portal_url", None)
+        portal_base_url
         or os.getenv("HERMES_PORTAL_BASE_URL")
         or os.getenv("NOUS_PORTAL_BASE_URL")
         or pconfig.portal_base_url
     ).rstrip("/")
     requested_inference_url = (
-        getattr(args, "inference_url", None)
+        inference_base_url
         or os.getenv("NOUS_INFERENCE_BASE_URL")
         or pconfig.inference_base_url
     ).rstrip("/")
-    client_id = getattr(args, "client_id", None) or pconfig.client_id
-    scope = getattr(args, "scope", None) or pconfig.scope
-    open_browser = not getattr(args, "no_browser", False)
-    timeout_seconds = getattr(args, "timeout", None) or 15.0
+    client_id = client_id or pconfig.client_id
+    scope = scope or pconfig.scope
     timeout = httpx.Timeout(timeout_seconds)
-
-    insecure = bool(getattr(args, "insecure", False))
-    ca_bundle = (
-        getattr(args, "ca_bundle", None)
-        or os.getenv("HERMES_CA_BUNDLE")
-        or os.getenv("SSL_CERT_FILE")
-    )
     verify: bool | str = False if insecure else (ca_bundle if ca_bundle else True)
 
-    # Skip browser open in SSH sessions
     if _is_remote_session():
         open_browser = False
 
@@ -2170,74 +2487,109 @@ def _login_nous(args, pconfig: ProviderConfig) -> None:
     elif ca_bundle:
         print(f"TLS verification: custom CA bundle ({ca_bundle})")
 
-    try:
-        with httpx.Client(timeout=timeout, headers={"Accept": "application/json"}, verify=verify) as client:
-            device_data = _request_device_code(
-                client=client, portal_base_url=portal_base_url,
-                client_id=client_id, scope=scope,
-            )
-
-            verification_url = str(device_data["verification_uri_complete"])
-            user_code = str(device_data["user_code"])
-            expires_in = int(device_data["expires_in"])
-            interval = int(device_data["interval"])
-
-            print()
-            print("To continue:")
-            print(f"  1. Open: {verification_url}")
-            print(f"  2. If prompted, enter code: {user_code}")
-
-            if open_browser:
-                opened = webbrowser.open(verification_url)
-                if opened:
-                    print("  (Opened browser for verification)")
-                else:
-                    print("  Could not open browser automatically — use the URL above.")
-
-            effective_interval = max(1, min(interval, DEVICE_AUTH_POLL_INTERVAL_CAP_SECONDS))
-            print(f"Waiting for approval (polling every {effective_interval}s)...")
-
-            token_data = _poll_for_token(
-                client=client, portal_base_url=portal_base_url,
-                client_id=client_id, device_code=str(device_data["device_code"]),
-                expires_in=expires_in, poll_interval=interval,
-            )
-
-        # Process token response
-        now = datetime.now(timezone.utc)
-        token_expires_in = _coerce_ttl_seconds(token_data.get("expires_in", 0))
-        expires_at = now.timestamp() + token_expires_in
-        inference_base_url = (
-            _optional_base_url(token_data.get("inference_base_url"))
-            or requested_inference_url
+    with httpx.Client(timeout=timeout, headers={"Accept": "application/json"}, verify=verify) as client:
+        device_data = _request_device_code(
+            client=client,
+            portal_base_url=portal_base_url,
+            client_id=client_id,
+            scope=scope,
         )
-        if inference_base_url != requested_inference_url:
-            print(f"Using portal-provided inference URL: {inference_base_url}")
 
-        auth_state = {
-            "portal_base_url": portal_base_url,
-            "inference_base_url": inference_base_url,
-            "client_id": client_id,
-            "scope": token_data.get("scope") or scope,
-            "token_type": token_data.get("token_type", "Bearer"),
-            "access_token": token_data["access_token"],
-            "refresh_token": token_data.get("refresh_token"),
-            "obtained_at": now.isoformat(),
-            "expires_at": datetime.fromtimestamp(expires_at, tz=timezone.utc).isoformat(),
-            "expires_in": token_expires_in,
-            "tls": {
-                "insecure": verify is False,
-                "ca_bundle": verify if isinstance(verify, str) else None,
-            },
-            "agent_key": None,
-            "agent_key_id": None,
-            "agent_key_expires_at": None,
-            "agent_key_expires_in": None,
-            "agent_key_reused": None,
-            "agent_key_obtained_at": None,
-        }
+        verification_url = str(device_data["verification_uri_complete"])
+        user_code = str(device_data["user_code"])
+        expires_in = int(device_data["expires_in"])
+        interval = int(device_data["interval"])
+
+        print()
+        print("To continue:")
+        print(f"  1. Open: {verification_url}")
+        print(f"  2. If prompted, enter code: {user_code}")
+
+        if open_browser:
+            opened = webbrowser.open(verification_url)
+            if opened:
+                print("  (Opened browser for verification)")
+            else:
+                print("  Could not open browser automatically — use the URL above.")
+
+        effective_interval = max(1, min(interval, DEVICE_AUTH_POLL_INTERVAL_CAP_SECONDS))
+        print(f"Waiting for approval (polling every {effective_interval}s)...")
+
+        token_data = _poll_for_token(
+            client=client,
+            portal_base_url=portal_base_url,
+            client_id=client_id,
+            device_code=str(device_data["device_code"]),
+            expires_in=expires_in,
+            poll_interval=interval,
+        )
+
+    now = datetime.now(timezone.utc)
+    token_expires_in = _coerce_ttl_seconds(token_data.get("expires_in", 0))
+    expires_at = now.timestamp() + token_expires_in
+    resolved_inference_url = (
+        _optional_base_url(token_data.get("inference_base_url"))
+        or requested_inference_url
+    )
+    if resolved_inference_url != requested_inference_url:
+        print(f"Using portal-provided inference URL: {resolved_inference_url}")
+
+    auth_state = {
+        "portal_base_url": portal_base_url,
+        "inference_base_url": resolved_inference_url,
+        "client_id": client_id,
+        "scope": token_data.get("scope") or scope,
+        "token_type": token_data.get("token_type", "Bearer"),
+        "access_token": token_data["access_token"],
+        "refresh_token": token_data.get("refresh_token"),
+        "obtained_at": now.isoformat(),
+        "expires_at": datetime.fromtimestamp(expires_at, tz=timezone.utc).isoformat(),
+        "expires_in": token_expires_in,
+        "tls": {
+            "insecure": verify is False,
+            "ca_bundle": verify if isinstance(verify, str) else None,
+        },
+        "agent_key": None,
+        "agent_key_id": None,
+        "agent_key_expires_at": None,
+        "agent_key_expires_in": None,
+        "agent_key_reused": None,
+        "agent_key_obtained_at": None,
+    }
+    return refresh_nous_oauth_from_state(
+        auth_state,
+        min_key_ttl_seconds=min_key_ttl_seconds,
+        timeout_seconds=timeout_seconds,
+        force_refresh=False,
+        force_mint=True,
+    )
+
+
+def _login_nous(args, pconfig: ProviderConfig) -> None:
+    """Nous Portal device authorization flow."""
+    timeout_seconds = getattr(args, "timeout", None) or 15.0
+    insecure = bool(getattr(args, "insecure", False))
+    ca_bundle = (
+        getattr(args, "ca_bundle", None)
+        or os.getenv("HERMES_CA_BUNDLE")
+        or os.getenv("SSL_CERT_FILE")
+    )
+
+    try:
+        auth_state = _nous_device_code_login(
+            portal_base_url=getattr(args, "portal_url", None) or pconfig.portal_base_url,
+            inference_base_url=getattr(args, "inference_url", None) or pconfig.inference_base_url,
+            client_id=getattr(args, "client_id", None) or pconfig.client_id,
+            scope=getattr(args, "scope", None) or pconfig.scope,
+            open_browser=not getattr(args, "no_browser", False),
+            timeout_seconds=timeout_seconds,
+            insecure=insecure,
+            ca_bundle=ca_bundle,
+            min_key_ttl_seconds=5 * 60,
+        )
+        inference_base_url = auth_state["inference_base_url"]
+        verify: bool | str = False if insecure else (ca_bundle if ca_bundle else True)
 
-        # Save auth state
         with _auth_store_lock():
             auth_store = _load_auth_store()
             _save_provider_state(auth_store, "nous", auth_state)
@@ -2249,34 +2601,29 @@ def _login_nous(args, pconfig: ProviderConfig) -> None:
         print(f"  Auth state: {saved_to}")
         print(f"  Config updated: {config_path} (model.provider=nous)")
 
-        # Mint an initial agent key and list available models
         try:
-            runtime_creds = resolve_nous_runtime_credentials(
-                min_key_ttl_seconds=5 * 60,
-                timeout_seconds=timeout_seconds,
-                insecure=insecure, ca_bundle=ca_bundle,
-            )
-            runtime_key = runtime_creds.get("api_key")
-            runtime_base_url = runtime_creds.get("base_url") or inference_base_url
+            runtime_key = auth_state.get("agent_key") or auth_state.get("access_token")
             if not isinstance(runtime_key, str) or not runtime_key:
-                raise AuthError("No runtime API key available to fetch models",
-                                provider="nous", code="invalid_token")
+                raise AuthError(
+                    "No runtime API key available to fetch models",
+                    provider="nous",
+                    code="invalid_token",
+                )
 
-            model_ids = fetch_nous_models(
-                inference_base_url=runtime_base_url,
-                api_key=runtime_key,
-                timeout_seconds=timeout_seconds,
-                verify=verify,
-            )
+            # Use curated model list (same as OpenRouter defaults) instead
+            # of the full /models dump which returns hundreds of models.
+            from hermes_cli.models import _PROVIDER_MODELS
+            model_ids = _PROVIDER_MODELS.get("nous", [])
 
             print()
             if model_ids:
+                print(f"Showing {len(model_ids)} curated models — use \"Enter custom model name\" for others.")
                 selected_model = _prompt_model_selection(model_ids)
                 if selected_model:
                     _save_model_choice(selected_model)
                     print(f"Default model set to: {selected_model}")
             else:
-                print("No models were returned by the inference API.")
+                print("No curated models available for Nous Portal.")
         except Exception as exc:
             message = format_auth_error(exc) if isinstance(exc, AuthError) else str(exc)
             print()

hermes_cli/auth_commands.py

@@ -0,0 +1,470 @@
+"""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,
+    _normalize_custom_pool_name,
+    get_pool_strategy,
+    label_from_token,
+    list_custom_pool_providers,
+    load_pool,
+    _exhausted_ttl,
+)
+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 ""
+    code = f" ({entry.last_error_code})" if entry.last_error_code else ""
+    if not entry.last_status_at:
+        return f" exhausted{code}"
+    remaining = max(0, int(math.ceil((entry.last_status_at + _exhausted_ttl(entry.last_error_code)) - time.time())))
+    if remaining <= 0:
+        return f" exhausted{code} (ready to retry)"
+    minutes, seconds = divmod(remaining, 60)
+    hours, minutes = divmod(minutes, 60)
+    if hours:
+        wait = f"{hours}h {minutes}m"
+    elif minutes:
+        wait = f"{minutes}m {seconds}s"
+    else:
+        wait = f"{seconds}s"
+    return f" exhausted{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", ""))
+    index = int(getattr(args, "index"))
+    pool = load_pool(provider)
+    removed = pool.remove_index(index)
+    if removed is None:
+        raise SystemExit(f"No credential #{index} for provider {provider}.")
+    print(f"Removed {provider} credential #{index} ({removed.label})")
+
+
+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"
+
+    auth_add_command(SimpleNamespace(
+        provider=provider, auth_type=auth_type, label=None, 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}")
+
+    try:
+        raw = input("Remove # (or blank to cancel): ").strip()
+    except (EOFError, KeyboardInterrupt):
+        return
+    if not raw:
+        return
+
+    try:
+        index = int(raw)
+    except ValueError:
+        print("Invalid number.")
+        return
+
+    auth_remove_command(SimpleNamespace(provider=provider, index=index))
+
+
+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()

hermes_cli/banner.py

@@ -11,7 +11,8 @@ import subprocess
 import threading
 import time
 from pathlib import Path
-from typing import Dict, List, Any, Optional
+from hermes_constants import get_hermes_home
+from typing import Dict, List, Optional
 
 from rich.console import Console
 from rich.panel import Panel
@@ -136,7 +137,7 @@ def check_for_updates() -> Optional[int]:
     ``~/.hermes/.update_check``).  Returns the number of commits behind,
     or ``None`` if the check fails or isn't applicable.
     """
-    hermes_home = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
+    hermes_home = get_hermes_home()
     repo_dir = hermes_home / "hermes-agent"
     cache_file = hermes_home / ".update_check"
 
@@ -266,8 +267,18 @@ def build_welcome_banner(console: Console, model: str, cwd: str,
 
     _, unavailable_toolsets = check_tool_availability(quiet=True)
     disabled_tools = set()
+    # Tools whose toolset has a check_fn are lazy-initialized (e.g. honcho,
+    # homeassistant) — they show as unavailable at banner time because the
+    # check hasn't run yet, but they aren't misconfigured.
+    lazy_tools = set()
     for item in unavailable_toolsets:
-        disabled_tools.update(item.get("tools", []))
+        toolset_name = item.get("name", "")
+        ts_req = TOOLSET_REQUIREMENTS.get(toolset_name, {})
+        tools_in_ts = item.get("tools", [])
+        if ts_req.get("check_fn"):
+            lazy_tools.update(tools_in_ts)
+        else:
+            disabled_tools.update(tools_in_ts)
 
     layout_table = Table.grid(padding=(0, 2))
     layout_table.add_column("left", justify="center")
@@ -327,6 +338,8 @@ def build_welcome_banner(console: Console, model: str, cwd: str,
         for name in sorted(tool_names):
             if name in disabled_tools:
                 colored_names.append(f"[red]{name}[/]")
+            elif name in lazy_tools:
+                colored_names.append(f"[yellow]{name}[/]")
             else:
                 colored_names.append(f"[{text}]{name}[/]")
 
@@ -346,6 +359,8 @@ def build_welcome_banner(console: Console, model: str, cwd: str,
                     colored_names.append("[dim]...[/]")
                 elif name in disabled_tools:
                     colored_names.append(f"[red]{name}[/]")
+                elif name in lazy_tools:
+                    colored_names.append(f"[yellow]{name}[/]")
                 else:
                     colored_names.append(f"[{text}]{name}[/]")
             tools_str = ", ".join(colored_names)
@@ -402,16 +417,26 @@ def build_welcome_banner(console: Console, model: str, cwd: str,
     if mcp_connected:
         summary_parts.append(f"{mcp_connected} MCP servers")
     summary_parts.append("/help for commands")
+    # Show active profile name when not 'default'
+    try:
+        from hermes_cli.profiles import get_active_profile_name
+        _profile_name = get_active_profile_name()
+        if _profile_name and _profile_name != "default":
+            right_lines.append(f"[bold {accent}]Profile:[/] [{text}]{_profile_name}[/]")
+    except Exception:
+        pass  # Never break the banner over a profiles.py bug
+
     right_lines.append(f"[dim {dim}]{' · '.join(summary_parts)}[/]")
 
     # Update check — use prefetched result if available
     try:
         behind = get_update_result(timeout=0.5)
         if behind and behind > 0:
+            from hermes_cli.config import recommended_update_command
             commits_word = "commit" if behind == 1 else "commits"
             right_lines.append(
                 f"[bold yellow]⚠ {behind} {commits_word} behind[/]"
-                f"[dim yellow] — run [bold]hermes update[/bold] to update[/]"
+                f"[dim yellow] — run [bold]{recommended_update_command()}[/bold] to update[/]"
             )
     except Exception:
         pass  # Never break the banner over an update check

hermes_cli/callbacks.py

@@ -12,6 +12,7 @@ import getpass
 
 from hermes_cli.banner import cprint, _DIM, _RST
 from hermes_cli.config import save_env_value_secure
+from hermes_constants import display_hermes_home
 
 
 def clarify_callback(cli, question, choices):
@@ -131,7 +132,8 @@ def prompt_for_secret(cli, var_name: str, prompt: str, metadata=None) -> dict:
             }
 
         stored = save_env_value_secure(var_name, value)
-        cprint(f"\n{_DIM}  ✓ Stored secret in ~/.hermes/.env as {var_name}{_RST}")
+        _dhh = display_hermes_home()
+        cprint(f"\n{_DIM}  ✓ Stored secret in {_dhh}/.env as {var_name}{_RST}")
         return {
             **stored,
             "skipped": False,
@@ -183,7 +185,8 @@ def prompt_for_secret(cli, var_name: str, prompt: str, metadata=None) -> dict:
                 }
 
             stored = save_env_value_secure(var_name, value)
-            cprint(f"\n{_DIM}  ✓ Stored secret in ~/.hermes/.env as {var_name}{_RST}")
+            _dhh = display_hermes_home()
+            cprint(f"\n{_DIM}  ✓ Stored secret in {_dhh}/.env as {var_name}{_RST}")
             return {
                 **stored,
                 "skipped": False,
@@ -238,7 +241,8 @@ def approval_callback(cli, command: str, description: str) -> str:
         lock = cli._approval_lock
 
     with lock:
-        timeout = 60
+        from cli import CLI_CONFIG
+        timeout = CLI_CONFIG.get("approvals", {}).get("timeout", 60)
         response_queue = queue.Queue()
         choices = ["once", "session", "always", "deny"]
         if len(command) > 70:

hermes_cli/checklist.py

@@ -5,6 +5,7 @@ toggleable list of items.  Falls back to a numbered text UI when
 curses is unavailable (Windows without curses, piped stdin, etc.).
 """
 
+import sys
 from typing import List, Set
 
 from hermes_cli.colors import Colors, color
@@ -26,6 +27,10 @@ def curses_checklist(
         The indices the user confirmed as checked.  On cancel (ESC/q),
         returns ``pre_selected`` unchanged.
     """
+    # Safety: return defaults when stdin is not a terminal.
+    if not sys.stdin.isatty():
+        return set(pre_selected)
+
     try:
         import curses
         selected = set(pre_selected)

hermes_cli/claw.py

@@ -4,24 +4,28 @@ Usage:
     hermes claw migrate              # Interactive migration from ~/.openclaw
     hermes claw migrate --dry-run    # Preview what would be migrated
     hermes claw migrate --preset full --overwrite  # Full migration, overwrite conflicts
+    hermes claw cleanup              # Archive leftover OpenClaw directories
+    hermes claw cleanup --dry-run    # Preview what would be archived
 """
 
 import importlib.util
 import logging
+import shutil
 import sys
+from datetime import datetime
 from pathlib import Path
 
 from hermes_cli.config import get_hermes_home, get_config_path, load_config, save_config
+from hermes_constants import get_optional_skills_dir
 from hermes_cli.setup import (
     Colors,
     color,
     print_header,
     print_info,
     print_success,
-    print_warning,
     print_error,
+    print_warning,
     prompt_yes_no,
-    prompt_choice,
 )
 
 logger = logging.getLogger(__name__)
@@ -29,8 +33,7 @@ logger = logging.getLogger(__name__)
 PROJECT_ROOT = Path(__file__).parent.parent.resolve()
 
 _OPENCLAW_SCRIPT = (
-    PROJECT_ROOT
-    / "optional-skills"
+    get_optional_skills_dir(PROJECT_ROOT / "optional-skills")
     / "migration"
     / "openclaw-migration"
     / "scripts"
@@ -47,6 +50,18 @@ _OPENCLAW_SCRIPT_INSTALLED = (
     / "openclaw_to_hermes.py"
 )
 
+# Known OpenClaw directory names (current + legacy)
+_OPENCLAW_DIR_NAMES = (".openclaw", ".clawdbot", ".moldbot")
+
+# State files commonly found in OpenClaw workspace directories that cause
+# confusion after migration (the agent discovers them and writes to them)
+_WORKSPACE_STATE_GLOBS = (
+    "*/todo.json",
+    "*/sessions/*",
+    "*/memory/*.json",
+    "*/logs/*",
+)
+
 
 def _find_migration_script() -> Path | None:
     """Find the openclaw_to_hermes.py script in known locations."""
@@ -73,24 +88,105 @@ def _load_migration_module(script_path: Path):
     return mod
 
 
+def _find_openclaw_dirs() -> list[Path]:
+    """Find all OpenClaw directories on disk."""
+    found = []
+    for name in _OPENCLAW_DIR_NAMES:
+        candidate = Path.home() / name
+        if candidate.is_dir():
+            found.append(candidate)
+    return found
+
+
+def _scan_workspace_state(source_dir: Path) -> list[tuple[Path, str]]:
+    """Scan an OpenClaw directory for workspace state files that cause confusion.
+
+    Returns a list of (path, description) tuples.
+    """
+    findings: list[tuple[Path, str]] = []
+
+    # Direct state files in the root
+    for name in ("todo.json", "sessions", "logs"):
+        candidate = source_dir / name
+        if candidate.exists():
+            kind = "directory" if candidate.is_dir() else "file"
+            findings.append((candidate, f"Root {kind}: {name}"))
+
+    # State files inside workspace directories
+    for child in sorted(source_dir.iterdir()):
+        if not child.is_dir() or child.name.startswith("."):
+            continue
+        # Check for workspace-like subdirectories
+        for state_name in ("todo.json", "sessions", "logs", "memory"):
+            state_path = child / state_name
+            if state_path.exists():
+                kind = "directory" if state_path.is_dir() else "file"
+                rel = state_path.relative_to(source_dir)
+                findings.append((state_path, f"Workspace {kind}: {rel}"))
+
+    return findings
+
+
+def _archive_directory(source_dir: Path, dry_run: bool = False) -> Path:
+    """Rename an OpenClaw directory to .pre-migration.
+
+    Returns the archive path.
+    """
+    timestamp = datetime.now().strftime("%Y%m%d")
+    archive_name = f"{source_dir.name}.pre-migration"
+    archive_path = source_dir.parent / archive_name
+
+    # If archive already exists, add timestamp
+    if archive_path.exists():
+        archive_name = f"{source_dir.name}.pre-migration-{timestamp}"
+        archive_path = source_dir.parent / archive_name
+
+    # If still exists (multiple runs same day), add counter
+    counter = 2
+    while archive_path.exists():
+        archive_name = f"{source_dir.name}.pre-migration-{timestamp}-{counter}"
+        archive_path = source_dir.parent / archive_name
+        counter += 1
+
+    if not dry_run:
+        source_dir.rename(archive_path)
+
+    return archive_path
+
+
 def claw_command(args):
     """Route hermes claw subcommands."""
     action = getattr(args, "claw_action", None)
 
     if action == "migrate":
         _cmd_migrate(args)
+    elif action in ("cleanup", "clean"):
+        _cmd_cleanup(args)
     else:
-        print("Usage: hermes claw migrate [options]")
+        print("Usage: hermes claw <command> [options]")
         print()
         print("Commands:")
         print("  migrate          Migrate settings from OpenClaw to Hermes")
+        print("  cleanup          Archive leftover OpenClaw directories after migration")
         print()
-        print("Run 'hermes claw migrate --help' for migration options.")
+        print("Run 'hermes claw <command> --help' for options.")
 
 
 def _cmd_migrate(args):
     """Run the OpenClaw → Hermes migration."""
-    source_dir = Path(getattr(args, "source", None) or Path.home() / ".openclaw")
+    # Check current and legacy OpenClaw directories
+    explicit_source = getattr(args, "source", None)
+    if explicit_source:
+        source_dir = Path(explicit_source)
+    else:
+        source_dir = Path.home() / ".openclaw"
+        if not source_dir.is_dir():
+            # Try legacy directory names
+            for legacy in (".clawdbot", ".moldbot"):
+                candidate = Path.home() / legacy
+                if candidate.is_dir():
+                    source_dir = candidate
+                    break
     dry_run = getattr(args, "dry_run", False)
     preset = getattr(args, "preset", "full")
     overwrite = getattr(args, "overwrite", False)
@@ -127,7 +223,7 @@ def _cmd_migrate(args):
         print()
         print_error(f"OpenClaw directory not found: {source_dir}")
         print_info("Make sure your OpenClaw installation is at the expected path.")
-        print_info(f"You can specify a custom path: hermes claw migrate --source /path/to/.openclaw")
+        print_info("You can specify a custom path: hermes claw migrate --source /path/to/.openclaw")
         return
 
     # Find the migration script
@@ -200,6 +296,168 @@ def _cmd_migrate(args):
     # Print results
     _print_migration_report(report, dry_run)
 
+    # After successful non-dry-run migration, offer to archive the source directory
+    if not dry_run and report.get("summary", {}).get("migrated", 0) > 0:
+        _offer_source_archival(source_dir, getattr(args, "yes", False))
+
+
+def _offer_source_archival(source_dir: Path, auto_yes: bool = False):
+    """After migration, offer to rename the source directory to prevent state fragmentation.
+
+    OpenClaw workspace directories contain state files (todo.json, sessions, etc.)
+    that the agent may discover and write to, causing confusion. Renaming the
+    directory prevents this.
+    """
+    if not source_dir.is_dir():
+        return
+
+    # Scan for state files that could cause problems
+    state_files = _scan_workspace_state(source_dir)
+
+    print()
+    print_header("Post-Migration Cleanup")
+    print_info("The OpenClaw directory still exists and contains workspace state files")
+    print_info("that can confuse the agent (todo lists, sessions, logs).")
+    if state_files:
+        print()
+        print(color("  Found state files:", Colors.YELLOW))
+        # Show up to 10 most relevant findings
+        for path, desc in state_files[:10]:
+            print(f"      {desc}")
+        if len(state_files) > 10:
+            print(f"      ... and {len(state_files) - 10} more")
+    print()
+    print_info(f"Recommend: rename {source_dir.name}/ to {source_dir.name}.pre-migration/")
+    print_info("This prevents the agent from discovering old workspace directories.")
+    print_info("You can always rename it back if needed.")
+    print()
+
+    if auto_yes or prompt_yes_no(f"Archive {source_dir} now?", default=True):
+        try:
+            archive_path = _archive_directory(source_dir)
+            print_success(f"Archived: {source_dir} → {archive_path}")
+            print_info("The original directory has been renamed, not deleted.")
+            print_info(f"To undo: mv {archive_path} {source_dir}")
+        except OSError as e:
+            print_error(f"Could not archive: {e}")
+            print_info(f"You can do it manually: mv {source_dir} {source_dir}.pre-migration")
+    else:
+        print_info("Skipped. You can archive later with: hermes claw cleanup")
+
+
+def _cmd_cleanup(args):
+    """Archive leftover OpenClaw directories after migration.
+
+    Scans for OpenClaw directories that still exist after migration and offers
+    to rename them to .pre-migration to prevent state fragmentation.
+    """
+    dry_run = getattr(args, "dry_run", False)
+    auto_yes = getattr(args, "yes", False)
+    explicit_source = getattr(args, "source", None)
+
+    print()
+    print(
+        color(
+            "┌─────────────────────────────────────────────────────────┐",
+            Colors.MAGENTA,
+        )
+    )
+    print(
+        color(
+            "│          ⚕ Hermes — OpenClaw Cleanup                   │",
+            Colors.MAGENTA,
+        )
+    )
+    print(
+        color(
+            "└─────────────────────────────────────────────────────────┘",
+            Colors.MAGENTA,
+        )
+    )
+
+    # Find OpenClaw directories
+    if explicit_source:
+        dirs_to_check = [Path(explicit_source)]
+    else:
+        dirs_to_check = _find_openclaw_dirs()
+
+    if not dirs_to_check:
+        print()
+        print_success("No OpenClaw directories found. Nothing to clean up.")
+        return
+
+    total_archived = 0
+
+    for source_dir in dirs_to_check:
+        print()
+        print_header(f"Found: {source_dir}")
+
+        # Scan for state files
+        state_files = _scan_workspace_state(source_dir)
+
+        # Show directory stats
+        try:
+            workspace_dirs = [
+                d for d in source_dir.iterdir()
+                if d.is_dir() and not d.name.startswith(".")
+                and any((d / name).exists() for name in ("todo.json", "SOUL.md", "MEMORY.md", "USER.md"))
+            ]
+        except OSError:
+            workspace_dirs = []
+
+        if workspace_dirs:
+            print_info(f"Workspace directories: {len(workspace_dirs)}")
+            for ws in workspace_dirs[:5]:
+                items = []
+                if (ws / "todo.json").exists():
+                    items.append("todo.json")
+                if (ws / "sessions").is_dir():
+                    items.append("sessions/")
+                if (ws / "SOUL.md").exists():
+                    items.append("SOUL.md")
+                if (ws / "MEMORY.md").exists():
+                    items.append("MEMORY.md")
+                detail = ", ".join(items) if items else "empty"
+                print(f"      {ws.name}/  ({detail})")
+            if len(workspace_dirs) > 5:
+                print(f"      ... and {len(workspace_dirs) - 5} more")
+
+        if state_files:
+            print()
+            print(color(f"  {len(state_files)} state file(s) that could cause confusion:", Colors.YELLOW))
+            for path, desc in state_files[:8]:
+                print(f"      {desc}")
+            if len(state_files) > 8:
+                print(f"      ... and {len(state_files) - 8} more")
+
+        print()
+
+        if dry_run:
+            archive_path = _archive_directory(source_dir, dry_run=True)
+            print_info(f"Would archive: {source_dir} → {archive_path}")
+        else:
+            if auto_yes or prompt_yes_no(f"Archive {source_dir}?", default=True):
+                try:
+                    archive_path = _archive_directory(source_dir)
+                    print_success(f"Archived: {source_dir} → {archive_path}")
+                    total_archived += 1
+                except OSError as e:
+                    print_error(f"Could not archive: {e}")
+                    print_info(f"Try manually: mv {source_dir} {source_dir}.pre-migration")
+            else:
+                print_info("Skipped.")
+
+    # Summary
+    print()
+    if dry_run:
+        print_info(f"Dry run complete. {len(dirs_to_check)} directory(ies) would be archived.")
+        print_info("Run without --dry-run to archive them.")
+    elif total_archived:
+        print_success(f"Cleaned up {total_archived} OpenClaw directory(ies).")
+        print_info("Directories were renamed, not deleted. You can undo by renaming them back.")
+    else:
+        print_info("No directories were archived.")
+
 
 def _print_migration_report(report: dict, dry_run: bool):
     """Print a formatted migration report."""
@@ -208,7 +466,6 @@ def _print_migration_report(report: dict, dry_run: bool):
     skipped = summary.get("skipped", 0)
     conflicts = summary.get("conflict", 0)
     errors = summary.get("error", 0)
-    total = migrated + skipped + conflicts + errors
 
     print()
     if dry_run:
@@ -242,7 +499,7 @@ def _print_migration_report(report: dict, dry_run: bool):
             print()
 
         if conflict_items:
-            print(color(f"  ⚠ Conflicts (skipped — use --overwrite to force):", Colors.YELLOW))
+            print(color("  ⚠ Conflicts (skipped — use --overwrite to force):", Colors.YELLOW))
             for item in conflict_items:
                 kind = item.get("kind", "unknown")
                 reason = item.get("reason", "already exists")
@@ -250,7 +507,7 @@ def _print_migration_report(report: dict, dry_run: bool):
             print()
 
         if skipped_items:
-            print(color(f"  ─ Skipped:", Colors.DIM))
+            print(color("  ─ Skipped:", Colors.DIM))
             for item in skipped_items:
                 kind = item.get("kind", "unknown")
                 reason = item.get("reason", "")
@@ -258,7 +515,7 @@ def _print_migration_report(report: dict, dry_run: bool):
             print()
 
         if error_items:
-            print(color(f"  ✗ Errors:", Colors.RED))
+            print(color("  ✗ Errors:", Colors.RED))
             for item in error_items:
                 kind = item.get("kind", "unknown")
                 reason = item.get("reason", "unknown error")

hermes_cli/codex_models.py

@@ -12,6 +12,8 @@ import os
 logger = logging.getLogger(__name__)
 
 DEFAULT_CODEX_MODELS: List[str] = [
+    "gpt-5.4-mini",
+    "gpt-5.4",
     "gpt-5.3-codex",
     "gpt-5.2-codex",
     "gpt-5.1-codex-max",
@@ -19,8 +21,9 @@ DEFAULT_CODEX_MODELS: List[str] = [
 ]
 
 _FORWARD_COMPAT_TEMPLATE_MODELS: List[tuple[str, tuple[str, ...]]] = [
-    ("gpt-5.3-codex", ("gpt-5.2-codex",)),
+    ("gpt-5.4-mini", ("gpt-5.3-codex", "gpt-5.2-codex")),
     ("gpt-5.4", ("gpt-5.3-codex", "gpt-5.2-codex")),
+    ("gpt-5.3-codex", ("gpt-5.2-codex",)),
     ("gpt-5.3-codex-spark", ("gpt-5.3-codex", "gpt-5.2-codex")),
 ]
 

hermes_cli/colors.py

@@ -1,8 +1,24 @@
 """Shared ANSI color utilities for Hermes CLI modules."""
 
+import os
 import sys
 
 
+def should_use_color() -> bool:
+    """Return True when colored output is appropriate.
+
+    Respects the NO_COLOR environment variable (https://no-color.org/)
+    and TERM=dumb, in addition to the existing TTY check.
+    """
+    if os.environ.get("NO_COLOR") is not None:
+        return False
+    if os.environ.get("TERM") == "dumb":
+        return False
+    if not sys.stdout.isatty():
+        return False
+    return True
+
+
 class Colors:
     RESET = "\033[0m"
     BOLD = "\033[1m"
@@ -16,7 +32,7 @@ class Colors:
 
 
 def color(text: str, *codes) -> str:
-    """Apply color codes to text (only when output is a TTY)."""
-    if not sys.stdout.isatty():
+    """Apply color codes to text (only when color output is appropriate)."""
+    if not should_use_color():
         return text
     return "".join(codes) + text + Colors.RESET

hermes_cli/commands.py

@@ -13,8 +13,7 @@ from __future__ import annotations
 import os
 import re
 from collections.abc import Callable, Mapping
-from dataclasses import dataclass, field
-from pathlib import Path
+from dataclasses import dataclass
 from typing import Any
 
 from prompt_toolkit.auto_suggest import AutoSuggest, Suggestion
@@ -37,6 +36,7 @@ class CommandDef:
     subcommands: tuple[str, ...] = ()  # tab-completable subcommands
     cli_only: bool = False             # only available in CLI
     gateway_only: bool = False         # only available in gateway/messaging
+    gateway_config_gate: str | None = None  # config dotpath; when truthy, overrides cli_only for gateway
 
 
 # ---------------------------------------------------------------------------
@@ -57,6 +57,8 @@ COMMAND_REGISTRY: list[CommandDef] = [
     CommandDef("undo", "Remove the last user/assistant exchange", "Session"),
     CommandDef("title", "Set a title for the current session", "Session",
                args_hint="[name]"),
+    CommandDef("branch", "Branch the current session (explore a different path)", "Session",
+               aliases=("fork",), args_hint="[name]"),
     CommandDef("compress", "Manually compress conversation context", "Session"),
     CommandDef("rollback", "List or restore filesystem checkpoints", "Session",
                args_hint="[number]"),
@@ -67,10 +69,13 @@ COMMAND_REGISTRY: list[CommandDef] = [
                gateway_only=True),
     CommandDef("background", "Run a prompt in the background", "Session",
                aliases=("bg",), args_hint="<prompt>"),
+    CommandDef("btw", "Ephemeral side question using session context (no tools, not persisted)", "Session",
+               args_hint="<question>"),
     CommandDef("queue", "Queue a prompt for the next turn (doesn't interrupt)", "Session",
                aliases=("q",), args_hint="<prompt>"),
     CommandDef("status", "Show session info", "Session",
                gateway_only=True),
+    CommandDef("profile", "Show active profile name and home directory", "Info"),
     CommandDef("sethome", "Set this chat as the home channel", "Session",
                gateway_only=True, aliases=("set-home",)),
     CommandDef("resume", "Resume a previously-named session", "Session",
@@ -79,8 +84,6 @@ COMMAND_REGISTRY: list[CommandDef] = [
     # Configuration
     CommandDef("config", "Show current configuration", "Configuration",
                cli_only=True),
-    CommandDef("model", "Show or change the current model", "Configuration",
-               args_hint="[name]"),
     CommandDef("provider", "Show available providers and current provider",
                "Configuration"),
     CommandDef("prompt", "View/set custom system prompt", "Configuration",
@@ -90,7 +93,10 @@ COMMAND_REGISTRY: list[CommandDef] = [
     CommandDef("statusbar", "Toggle the context/model status bar", "Configuration",
                cli_only=True, aliases=("sb",)),
     CommandDef("verbose", "Cycle tool progress display: off -> new -> all -> verbose",
-               "Configuration", cli_only=True),
+               "Configuration", cli_only=True,
+               gateway_config_gate="display.tool_progress_command"),
+    CommandDef("yolo", "Toggle YOLO mode (skip all dangerous command approvals)",
+               "Configuration"),
     CommandDef("reasoning", "Manage reasoning effort and display", "Configuration",
                args_hint="[level|show|hide]",
                subcommands=("none", "low", "minimal", "medium", "high", "xhigh", "show", "hide", "on", "off")),
@@ -119,6 +125,8 @@ COMMAND_REGISTRY: list[CommandDef] = [
                "Tools & Skills", cli_only=True),
 
     # Info
+    CommandDef("commands", "Browse all commands and skills (paginated)", "Info",
+               gateway_only=True, args_hint="[page]"),
     CommandDef("help", "Show available commands", "Info"),
     CommandDef("usage", "Show token usage for the current session", "Info"),
     CommandDef("insights", "Show usage insights and analytics", "Info",
@@ -137,7 +145,7 @@ COMMAND_REGISTRY: list[CommandDef] = [
 
 
 # ---------------------------------------------------------------------------
-# Derived lookups -- rebuilt once at import time
+# Derived lookups -- rebuilt once at import time, refreshed by rebuild_lookups()
 # ---------------------------------------------------------------------------
 
 def _build_command_lookup() -> dict[str, CommandDef]:
@@ -161,6 +169,58 @@ def resolve_command(name: str) -> CommandDef | None:
     return _COMMAND_LOOKUP.get(name.lower().lstrip("/"))
 
 
+def register_plugin_command(cmd: CommandDef) -> None:
+    """Append a plugin-defined command to the registry and refresh lookups."""
+    COMMAND_REGISTRY.append(cmd)
+    rebuild_lookups()
+
+
+def rebuild_lookups() -> None:
+    """Rebuild all derived lookup dicts from the current COMMAND_REGISTRY.
+
+    Called after plugin commands are registered so they appear in help,
+    autocomplete, gateway dispatch, Telegram menu, and Slack mapping.
+    """
+    global GATEWAY_KNOWN_COMMANDS
+
+    _COMMAND_LOOKUP.clear()
+    _COMMAND_LOOKUP.update(_build_command_lookup())
+
+    COMMANDS.clear()
+    for cmd in COMMAND_REGISTRY:
+        if not cmd.gateway_only:
+            COMMANDS[f"/{cmd.name}"] = _build_description(cmd)
+            for alias in cmd.aliases:
+                COMMANDS[f"/{alias}"] = f"{cmd.description} (alias for /{cmd.name})"
+
+    COMMANDS_BY_CATEGORY.clear()
+    for cmd in COMMAND_REGISTRY:
+        if not cmd.gateway_only:
+            cat = COMMANDS_BY_CATEGORY.setdefault(cmd.category, {})
+            cat[f"/{cmd.name}"] = COMMANDS[f"/{cmd.name}"]
+            for alias in cmd.aliases:
+                cat[f"/{alias}"] = COMMANDS[f"/{alias}"]
+
+    SUBCOMMANDS.clear()
+    for cmd in COMMAND_REGISTRY:
+        if cmd.subcommands:
+            SUBCOMMANDS[f"/{cmd.name}"] = list(cmd.subcommands)
+    for cmd in COMMAND_REGISTRY:
+        key = f"/{cmd.name}"
+        if key in SUBCOMMANDS or not cmd.args_hint:
+            continue
+        m = _PIPE_SUBS_RE.search(cmd.args_hint)
+        if m:
+            SUBCOMMANDS[key] = m.group(0).split("|")
+
+    GATEWAY_KNOWN_COMMANDS = frozenset(
+        name
+        for cmd in COMMAND_REGISTRY
+        if not cmd.cli_only or cmd.gateway_config_gate
+        for name in (cmd.name, *cmd.aliases)
+    )
+
+
 def _build_description(cmd: CommandDef) -> str:
     """Build a CLI-facing description string including usage hint."""
     if cmd.args_hint:
@@ -210,20 +270,76 @@ for _cmd in COMMAND_REGISTRY:
 # Gateway helpers
 # ---------------------------------------------------------------------------
 
-# Set of all command names + aliases recognized by the gateway
+# Set of all command names + aliases recognized by the gateway.
+# Includes config-gated commands so the gateway can dispatch them
+# (the handler checks the config gate at runtime).
 GATEWAY_KNOWN_COMMANDS: frozenset[str] = frozenset(
     name
     for cmd in COMMAND_REGISTRY
-    if not cmd.cli_only
+    if not cmd.cli_only or cmd.gateway_config_gate
     for name in (cmd.name, *cmd.aliases)
 )
 
 
+def _resolve_config_gates() -> set[str]:
+    """Return canonical names of commands whose ``gateway_config_gate`` is truthy.
+
+    Reads ``config.yaml`` and walks the dot-separated key path for each
+    config-gated command.  Returns an empty set on any error so callers
+    degrade gracefully.
+    """
+    gated = [c for c in COMMAND_REGISTRY if c.gateway_config_gate]
+    if not gated:
+        return set()
+    try:
+        import yaml
+        config_path = os.path.join(
+            os.getenv("HERMES_HOME", os.path.expanduser("~/.hermes")),
+            "config.yaml",
+        )
+        if os.path.exists(config_path):
+            with open(config_path, encoding="utf-8") as f:
+                cfg = yaml.safe_load(f) or {}
+        else:
+            cfg = {}
+    except Exception:
+        return set()
+    result: set[str] = set()
+    for cmd in gated:
+        val: Any = cfg
+        for key in cmd.gateway_config_gate.split("."):
+            if isinstance(val, dict):
+                val = val.get(key)
+            else:
+                val = None
+                break
+        if val:
+            result.add(cmd.name)
+    return result
+
+
+def _is_gateway_available(cmd: CommandDef, config_overrides: set[str] | None = None) -> bool:
+    """Check if *cmd* should appear in gateway surfaces (help, menus, mappings).
+
+    Unconditionally available when ``cli_only`` is False.  When ``cli_only``
+    is True but ``gateway_config_gate`` is set, the command is available only
+    when the config value is truthy.  Pass *config_overrides* (from
+    ``_resolve_config_gates()``) to avoid re-reading config for every command.
+    """
+    if not cmd.cli_only:
+        return True
+    if cmd.gateway_config_gate:
+        overrides = config_overrides if config_overrides is not None else _resolve_config_gates()
+        return cmd.name in overrides
+    return False
+
+
 def gateway_help_lines() -> list[str]:
     """Generate gateway help text lines from the registry."""
+    overrides = _resolve_config_gates()
     lines: list[str] = []
     for cmd in COMMAND_REGISTRY:
-        if cmd.cli_only:
+        if not _is_gateway_available(cmd, overrides):
             continue
         args = f" {cmd.args_hint}" if cmd.args_hint else ""
         alias_parts: list[str] = []
@@ -244,24 +360,154 @@ def telegram_bot_commands() -> list[tuple[str, str]]:
     underscores.  Aliases are skipped -- Telegram shows one menu entry per
     canonical command.
     """
+    overrides = _resolve_config_gates()
     result: list[tuple[str, str]] = []
     for cmd in COMMAND_REGISTRY:
-        if cmd.cli_only:
+        if not _is_gateway_available(cmd, overrides):
             continue
         tg_name = cmd.name.replace("-", "_")
         result.append((tg_name, cmd.description))
     return result
 
 
+_TG_NAME_LIMIT = 32
+
+
+def _clamp_telegram_names(
+    entries: list[tuple[str, str]],
+    reserved: set[str],
+) -> list[tuple[str, str]]:
+    """Enforce Telegram's 32-char command name limit with collision avoidance.
+
+    Names exceeding 32 chars are truncated.  If truncation creates a duplicate
+    (against *reserved* names or earlier entries in the same batch), the name is
+    shortened to 31 chars and a digit ``0``-``9`` is appended to differentiate.
+    If all 10 digit slots are taken the entry is silently dropped.
+    """
+    used: set[str] = set(reserved)
+    result: list[tuple[str, str]] = []
+    for name, desc in entries:
+        if len(name) > _TG_NAME_LIMIT:
+            candidate = name[:_TG_NAME_LIMIT]
+            if candidate in used:
+                prefix = name[:_TG_NAME_LIMIT - 1]
+                for digit in range(10):
+                    candidate = f"{prefix}{digit}"
+                    if candidate not in used:
+                        break
+                else:
+                    # All 10 digit slots exhausted — skip entry
+                    continue
+            name = candidate
+        if name in used:
+            continue
+        used.add(name)
+        result.append((name, desc))
+    return result
+
+
+def telegram_menu_commands(max_commands: int = 100) -> tuple[list[tuple[str, str]], int]:
+    """Return Telegram menu commands capped to the Bot API limit.
+
+    Priority order (higher priority = never bumped by overflow):
+      1. Core CommandDef commands (always included)
+      2. Plugin slash commands (take precedence over skills)
+      3. Built-in skill commands (fill remaining slots, alphabetical)
+
+    Skills are the only tier that gets trimmed when the cap is hit.
+    User-installed hub skills are excluded — accessible via /skills.
+    Skills disabled for the ``"telegram"`` platform (via ``hermes skills
+    config``) are excluded from the menu entirely.
+
+    Returns:
+        (menu_commands, hidden_count) where hidden_count is the number of
+        skill commands omitted due to the cap.
+    """
+    core_commands = list(telegram_bot_commands())
+    # Reserve core names so plugin/skill truncation can't collide with them
+    reserved_names = {n for n, _ in core_commands}
+    all_commands = list(core_commands)
+
+    # Plugin slash commands get priority over skills
+    plugin_entries: list[tuple[str, str]] = []
+    try:
+        from hermes_cli.plugins import get_plugin_manager
+        pm = get_plugin_manager()
+        plugin_cmds = getattr(pm, "_plugin_commands", {})
+        for cmd_name in sorted(plugin_cmds):
+            tg_name = cmd_name.replace("-", "_")
+            desc = "Plugin command"
+            if len(desc) > 40:
+                desc = desc[:37] + "..."
+            plugin_entries.append((tg_name, desc))
+    except Exception:
+        pass
+
+    # Clamp plugin names to 32 chars with collision avoidance
+    plugin_entries = _clamp_telegram_names(plugin_entries, reserved_names)
+    reserved_names.update(n for n, _ in plugin_entries)
+    all_commands.extend(plugin_entries)
+
+    # Load per-platform disabled skills so they don't consume menu slots.
+    # get_skill_commands() already filters the *global* disabled list, but
+    # per-platform overrides (skills.platform_disabled.telegram) were never
+    # applied here — that's what this block fixes.
+    _platform_disabled: set[str] = set()
+    try:
+        from agent.skill_utils import get_disabled_skill_names
+        _platform_disabled = get_disabled_skill_names(platform="telegram")
+    except Exception:
+        pass
+
+    # Remaining slots go to built-in skill commands (not hub-installed).
+    skill_entries: list[tuple[str, str]] = []
+    try:
+        from agent.skill_commands import get_skill_commands
+        from tools.skills_tool import SKILLS_DIR
+        _skills_dir = str(SKILLS_DIR.resolve())
+        _hub_dir = str((SKILLS_DIR / ".hub").resolve())
+        skill_cmds = get_skill_commands()
+        for cmd_key in sorted(skill_cmds):
+            info = skill_cmds[cmd_key]
+            skill_path = info.get("skill_md_path", "")
+            if not skill_path.startswith(_skills_dir):
+                continue
+            if skill_path.startswith(_hub_dir):
+                continue
+            # Skip skills disabled for telegram
+            skill_name = info.get("name", "")
+            if skill_name in _platform_disabled:
+                continue
+            name = cmd_key.lstrip("/").replace("-", "_")
+            desc = info.get("description", "")
+            # Keep descriptions short — setMyCommands has an undocumented
+            # total payload limit.  40 chars fits 100 commands safely.
+            if len(desc) > 40:
+                desc = desc[:37] + "..."
+            skill_entries.append((name, desc))
+    except Exception:
+        pass
+
+    # Clamp skill names to 32 chars with collision avoidance
+    skill_entries = _clamp_telegram_names(skill_entries, reserved_names)
+
+    # Skills fill remaining slots — they're the only tier that gets trimmed
+    remaining_slots = max(0, max_commands - len(all_commands))
+    hidden_count = max(0, len(skill_entries) - remaining_slots)
+    all_commands.extend(skill_entries[:remaining_slots])
+    return all_commands[:max_commands], hidden_count
+
+
 def slack_subcommand_map() -> dict[str, str]:
     """Return subcommand -> /command mapping for Slack /hermes handler.
 
     Maps both canonical names and aliases so /hermes bg do stuff works
     the same as /hermes background do stuff.
     """
+    overrides = _resolve_config_gates()
     mapping: dict[str, str] = {}
     for cmd in COMMAND_REGISTRY:
-        if cmd.cli_only:
+        if not _is_gateway_available(cmd, overrides):
             continue
         mapping[cmd.name] = f"/{cmd.name}"
         for alias in cmd.aliases:
@@ -279,29 +525,8 @@ class SlashCommandCompleter(Completer):
     def __init__(
         self,
         skill_commands_provider: Callable[[], Mapping[str, dict[str, Any]]] | None = None,
-        model_completer_provider: Callable[[], dict[str, Any]] | None = None,
     ) -> None:
         self._skill_commands_provider = skill_commands_provider
-        # model_completer_provider returns {"current_provider": str,
-        #   "providers": {id: label, ...}, "models_for": callable(provider) -> list[str]}
-        self._model_completer_provider = model_completer_provider
-        self._model_info_cache: dict[str, Any] | None = None
-        self._model_info_cache_time: float = 0
-
-    def _get_model_info(self) -> dict[str, Any]:
-        """Get cached model/provider info for /model autocomplete."""
-        import time
-        now = time.monotonic()
-        if self._model_info_cache is not None and now - self._model_info_cache_time < 60:
-            return self._model_info_cache
-        if self._model_completer_provider is None:
-            return {}
-        try:
-            self._model_info_cache = self._model_completer_provider() or {}
-            self._model_info_cache_time = now
-        except Exception:
-            self._model_info_cache = self._model_info_cache or {}
-        return self._model_info_cache
 
     def _iter_skill_commands(self) -> Mapping[str, dict[str, Any]]:
         if self._skill_commands_provider is None:
@@ -397,9 +622,136 @@ class SlashCommandCompleter(Completer):
             )
             count += 1
 
+    @staticmethod
+    def _extract_context_word(text: str) -> str | None:
+        """Extract a bare ``@`` token for context reference completions."""
+        if not text:
+            return None
+        # Walk backwards to find the start of the current word
+        i = len(text) - 1
+        while i >= 0 and text[i] != " ":
+            i -= 1
+        word = text[i + 1:]
+        if not word.startswith("@"):
+            return None
+        return word
+
+    @staticmethod
+    def _context_completions(word: str, limit: int = 30):
+        """Yield Claude Code-style @ context completions.
+
+        Bare ``@`` or ``@partial`` shows static references and matching
+        files/folders.  ``@file:path`` and ``@folder:path`` are handled
+        by the existing path completion path.
+        """
+        lowered = word.lower()
+
+        # Static context references
+        _STATIC_REFS = (
+            ("@diff", "Git working tree diff"),
+            ("@staged", "Git staged diff"),
+            ("@file:", "Attach a file"),
+            ("@folder:", "Attach a folder"),
+            ("@git:", "Git log with diffs (e.g. @git:5)"),
+            ("@url:", "Fetch web content"),
+        )
+        for candidate, meta in _STATIC_REFS:
+            if candidate.lower().startswith(lowered) and candidate.lower() != lowered:
+                yield Completion(
+                    candidate,
+                    start_position=-len(word),
+                    display=candidate,
+                    display_meta=meta,
+                )
+
+        # If the user typed @file: or @folder:, delegate to path completions
+        for prefix in ("@file:", "@folder:"):
+            if word.startswith(prefix):
+                path_part = word[len(prefix):] or "."
+                expanded = os.path.expanduser(path_part)
+                if expanded.endswith("/"):
+                    search_dir, match_prefix = expanded, ""
+                else:
+                    search_dir = os.path.dirname(expanded) or "."
+                    match_prefix = os.path.basename(expanded)
+
+                try:
+                    entries = os.listdir(search_dir)
+                except OSError:
+                    return
+
+                count = 0
+                prefix_lower = match_prefix.lower()
+                for entry in sorted(entries):
+                    if match_prefix and not entry.lower().startswith(prefix_lower):
+                        continue
+                    if count >= limit:
+                        break
+                    full_path = os.path.join(search_dir, entry)
+                    is_dir = os.path.isdir(full_path)
+                    display_path = os.path.relpath(full_path)
+                    suffix = "/" if is_dir else ""
+                    kind = "folder" if is_dir else "file"
+                    meta = "dir" if is_dir else _file_size_label(full_path)
+                    completion = f"@{kind}:{display_path}{suffix}"
+                    yield Completion(
+                        completion,
+                        start_position=-len(word),
+                        display=entry + suffix,
+                        display_meta=meta,
+                    )
+                    count += 1
+                return
+
+        # Bare @ or @partial — show matching files/folders from cwd
+        query = word[1:]  # strip the @
+        if not query:
+            search_dir, match_prefix = ".", ""
+        else:
+            expanded = os.path.expanduser(query)
+            if expanded.endswith("/"):
+                search_dir, match_prefix = expanded, ""
+            else:
+                search_dir = os.path.dirname(expanded) or "."
+                match_prefix = os.path.basename(expanded)
+
+        try:
+            entries = os.listdir(search_dir)
+        except OSError:
+            return
+
+        count = 0
+        prefix_lower = match_prefix.lower()
+        for entry in sorted(entries):
+            if match_prefix and not entry.lower().startswith(prefix_lower):
+                continue
+            if entry.startswith("."):
+                continue  # skip hidden files in bare @ mode
+            if count >= limit:
+                break
+            full_path = os.path.join(search_dir, entry)
+            is_dir = os.path.isdir(full_path)
+            display_path = os.path.relpath(full_path)
+            suffix = "/" if is_dir else ""
+            kind = "folder" if is_dir else "file"
+            meta = "dir" if is_dir else _file_size_label(full_path)
+            completion = f"@{kind}:{display_path}{suffix}"
+            yield Completion(
+                completion,
+                start_position=-len(word),
+                display=entry + suffix,
+                display_meta=meta,
+            )
+            count += 1
+
     def get_completions(self, document, complete_event):
         text = document.text_before_cursor
         if not text.startswith("/"):
+            # Try @ context completion (Claude Code-style)
+            ctx_word = self._extract_context_word(text)
+            if ctx_word is not None:
+                yield from self._context_completions(ctx_word)
+                return
             # Try file path completion for non-slash input
             path_word = self._extract_path_word(text)
             if path_word is not None:
@@ -413,52 +765,6 @@ class SlashCommandCompleter(Completer):
             sub_text = parts[1] if len(parts) > 1 else ""
             sub_lower = sub_text.lower()
 
-            # /model gets two-stage completion:
-            #   Stage 1: provider names (with : suffix)
-            #   Stage 2: after "provider:", list that provider's models
-            if base_cmd == "/model" and " " not in sub_text:
-                info = self._get_model_info()
-                if info:
-                    current_prov = info.get("current_provider", "")
-                    providers = info.get("providers", {})
-                    models_for = info.get("models_for")
-
-                    if ":" in sub_text:
-                        # Stage 2: "anthropic:cl" → models for anthropic
-                        prov_part, model_part = sub_text.split(":", 1)
-                        model_lower = model_part.lower()
-                        if models_for:
-                            try:
-                                prov_models = models_for(prov_part)
-                            except Exception:
-                                prov_models = []
-                            for mid in prov_models:
-                                if mid.lower().startswith(model_lower) and mid.lower() != model_lower:
-                                    full = f"{prov_part}:{mid}"
-                                    yield Completion(
-                                        full,
-                                        start_position=-len(sub_text),
-                                        display=mid,
-                                    )
-                    else:
-                        # Stage 1: providers sorted: non-current first, current last
-                        for pid, plabel in sorted(
-                            providers.items(),
-                            key=lambda kv: (kv[0] == current_prov, kv[0]),
-                        ):
-                            display_name = f"{pid}:"
-                            if display_name.lower().startswith(sub_lower):
-                                meta = f"({plabel})" if plabel != pid else ""
-                                if pid == current_prov:
-                                    meta = f"(current — {plabel})" if plabel != pid else "(current)"
-                                yield Completion(
-                                    display_name,
-                                    start_position=-len(sub_text),
-                                    display=display_name,
-                                    display_meta=meta,
-                                )
-                return
-
             # Static subcommand completions
             if " " not in sub_text and base_cmd in SUBCOMMANDS:
                 for sub in SUBCOMMANDS[base_cmd]:
@@ -540,32 +846,6 @@ class SlashCommandAutoSuggest(AutoSuggest):
         sub_text = parts[1] if len(parts) > 1 else ""
         sub_lower = sub_text.lower()
 
-        # /model gets two-stage ghost text
-        if base_cmd == "/model" and " " not in sub_text and self._completer:
-            info = self._completer._get_model_info()
-            if info:
-                providers = info.get("providers", {})
-                models_for = info.get("models_for")
-                current_prov = info.get("current_provider", "")
-
-                if ":" in sub_text:
-                    # Stage 2: after provider:, suggest model
-                    prov_part, model_part = sub_text.split(":", 1)
-                    model_lower = model_part.lower()
-                    if models_for:
-                        try:
-                            for mid in models_for(prov_part):
-                                if mid.lower().startswith(model_lower) and mid.lower() != model_lower:
-                                    return Suggestion(mid[len(model_part):])
-                        except Exception:
-                            pass
-                else:
-                    # Stage 1: suggest provider name with :
-                    for pid in sorted(providers, key=lambda p: (p == current_prov, p)):
-                        candidate = f"{pid}:"
-                        if candidate.lower().startswith(sub_lower) and candidate.lower() != sub_lower:
-                            return Suggestion(candidate[len(sub_text):])
-
         # Static subcommands
         if base_cmd in SUBCOMMANDS and SUBCOMMANDS[base_cmd]:
             if " " not in sub_text: