refactor: extract codex_responses logic into dedicated adapter

Move 10 Responses API format-conversion and normalization functions from
run_agent.py into agent/codex_responses_adapter.py. All functions are now
stateless module-level functions with zero self references.

The AIAgent methods remain as thin one-line wrappers that delegate to the
adapter, so all callers (tests, gateway, CLI) are unchanged.

Functions extracted:
- _deterministic_call_id: deterministic tool call ID generation
- _split_responses_tool_id: composite ID splitting
- _derive_responses_function_call_id: call_ to fc_ prefix conversion
- _responses_tools: chat completions tool schema → Responses format
- _chat_messages_to_responses_input: message format conversion
- _preflight_codex_input_items: input item normalization
- _preflight_codex_api_kwargs: API kwargs validation/cleaning
- _extract_responses_message_text: text extraction from response items
- _extract_responses_reasoning_text: reasoning extraction
- _normalize_codex_response: full response normalization

This brings codex_responses in line with anthropic_adapter.py and
bedrock_adapter.py which already have their own adapter files.

run_agent.py: 12410 → 11845 lines (-565 net)

.env.example

@@ -24,6 +24,15 @@
 # Optional base URL override (default: Google's OpenAI-compatible endpoint)
 # GEMINI_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai
 
+# =============================================================================
+# LLM PROVIDER (Ollama Cloud)
+# =============================================================================
+# Cloud-hosted open models via Ollama's OpenAI-compatible endpoint.
+# Get your key at: https://ollama.com/settings
+# OLLAMA_API_KEY=your_ollama_key_here
+# Optional base URL override (default: https://ollama.com/v1)
+# OLLAMA_BASE_URL=https://ollama.com/v1
+
 # =============================================================================
 # LLM PROVIDER (z.ai / GLM)
 # =============================================================================

.envrc

@@ -1 +1,5 @@
+watch_file pyproject.toml uv.lock
+watch_file ui-tui/package-lock.json ui-tui/package.json
+watch_file flake.nix flake.lock nix/devShell.nix nix/tui.nix nix/package.nix nix/python.nix
+
 use flake

.github/workflows/deploy-site.yml

@@ -1,11 +1,12 @@
 name: Deploy Site
 
 on:
+  release:
+    types: [published]
   push:
     branches: [main]
     paths:
       - 'website/**'
-      - 'landingpage/**'
       - 'skills/**'
       - 'optional-skills/**'
       - '.github/workflows/deploy-site.yml'
@@ -20,8 +21,14 @@ concurrency:
   cancel-in-progress: false
 
 jobs:
-  build-and-deploy:
-    # Only run on the upstream repository, not on forks
+  deploy-vercel:
+    if: github.event_name == 'release'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Trigger Vercel Deploy
+        run: curl -X POST "${{ secrets.VERCEL_DEPLOY_HOOK }}"
+
+  deploy-docs:
     if: github.repository == 'NousResearch/hermes-agent'
     runs-on: ubuntu-latest
     environment:
@@ -65,12 +72,7 @@ jobs:
       - name: Stage deployment
         run: |
           mkdir -p _site/docs
-          # Landing page at root
-          cp -r landingpage/* _site/
-          # Docusaurus at /docs/
           cp -r website/build/* _site/docs/
-          # CNAME so GitHub Pages keeps the custom domain between deploys
-          echo "hermes-agent.nousresearch.com" > _site/CNAME
 
       - name: Upload artifact
         uses: actions/upload-pages-artifact@56afc609e74202658d3ffba0e8f6dda462b719fa  # v3

.github/workflows/docker-publish.yml

@@ -3,8 +3,13 @@ name: Docker Build and Publish
 on:
   push:
     branches: [main]
-  pull_request:
-    branches: [main]
+    paths:
+      - '**/*.py'
+      - 'pyproject.toml'
+      - 'uv.lock'
+      - 'Dockerfile'
+      - 'docker/**'
+      - '.github/workflows/docker-publish.yml'
   release:
     types: [published]
 
@@ -49,6 +54,14 @@ jobs:
 
       - name: Test image starts
         run: |
+          # The image runs as the hermes user (UID 10000).  GitHub Actions
+          # creates /tmp/hermes-test root-owned by default, which hermes
+          # can't write to — chown it to match the in-container UID before
+          # bind-mounting.  Real users doing `docker run -v ~/.hermes:...`
+          # with their own UID hit the same issue and have their own
+          # remediations (HERMES_UID env var, or chown locally).
+          mkdir -p /tmp/hermes-test
+          sudo chown -R 10000:10000 /tmp/hermes-test
           docker run --rm \
             -v /tmp/hermes-test:/opt/data \
             --entrypoint /opt/hermes/docker/entrypoint.sh \

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

@@ -3,14 +3,31 @@ name: Supply Chain Audit
 on:
   pull_request:
     types: [opened, synchronize, reopened]
+    paths:
+      - '**/*.py'
+      - '**/*.pth'
+      - '**/setup.py'
+      - '**/setup.cfg'
+      - '**/sitecustomize.py'
+      - '**/usercustomize.py'
+      - '**/__init__.pth'
 
 permissions:
   pull-requests: write
   contents: read
 
+# Narrow, high-signal scanner. Only fires on critical indicators of supply
+# chain attacks (e.g. the litellm-style payloads). Low-signal heuristics
+# (plain base64, plain exec/eval, dependency/Dockerfile/workflow edits,
+# Actions version unpinning, outbound POST/PUT) were intentionally
+# removed — they fired on nearly every PR and trained reviewers to ignore
+# the scanner. Keep this file's checks ruthlessly narrow: if you find
+# yourself adding WARNING-tier patterns here again, make a separate
+# advisory-only workflow instead.
+
 jobs:
   scan:
-    name: Scan PR for supply chain risks
+    name: Scan PR for critical supply chain risks
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
@@ -18,7 +35,7 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Scan diff for suspicious patterns
+      - name: Scan diff for critical patterns
         id: scan
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -28,19 +45,19 @@ jobs:
           BASE="${{ github.event.pull_request.base.sha }}"
           HEAD="${{ github.event.pull_request.head.sha }}"
 
-          # Get the full diff (added lines only)
+          # Added lines only, excluding lockfiles.
           DIFF=$(git diff "$BASE".."$HEAD" -- . ':!uv.lock' ':!*.lock' ':!package-lock.json' ':!yarn.lock' || true)
 
           FINDINGS=""
-          CRITICAL=false
 
           # --- .pth files (auto-execute on Python startup) ---
+          # The exact mechanism used in the litellm supply chain attack:
+          # https://github.com/BerriAI/litellm/issues/24512
           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).
+          Python \`.pth\` files in \`site-packages/\` execute automatically when the interpreter starts — no import required.
 
           **Files:**
           \`\`\`
@@ -49,13 +66,12 @@ jobs:
           "
           fi
 
-          # --- base64 + exec/eval combo (the litellm attack pattern) ---
+          # --- base64 decode + exec/eval on the same line (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.
+          Base64-decoded strings passed directly to exec/eval — the signature of hidden credential-stealing payloads.
 
           **Matches:**
           \`\`\`
@@ -64,41 +80,12 @@ jobs:
           "
           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)
+          # --- subprocess with encoded/obfuscated command argument ---
+          PROC_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -E 'subprocess\.(Popen|call|run)\s*\(' | grep -iE 'base64|\\x[0-9a-f]{2}|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.
+          Subprocess calls whose command strings are base64- or hex-encoded are a strong indicator of payload execution.
 
           **Matches:**
           \`\`\`
@@ -107,25 +94,12 @@ jobs:
           "
           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)$' || true)
+          # --- Install-hook files (setup.py/sitecustomize/usercustomize/__init__.pth) ---
+          # These execute during pip install or interpreter startup.
+          SETUP_HITS=$(git diff --name-only "$BASE".."$HEAD" | grep -E '(^|/)(setup\.py|setup\.cfg|sitecustomize\.py|usercustomize\.py|__init__\.pth)$' || true)
           if [ -n "$SETUP_HITS" ]; then
             FINDINGS="${FINDINGS}
-          ### ⚠️ WARNING: Install hook files modified
+          ### 🚨 CRITICAL: Install-hook file added or modified
           These files can execute code during package installation or interpreter startup.
 
           **Files:**
@@ -135,114 +109,31 @@ jobs:
           "
           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:**
-          \`\`\`
-          ${MARSHAL_HITS}
-          \`\`\`
-          "
-          fi
-
-          # --- CI/CD workflow files modified ---
-          WORKFLOW_HITS=$(git diff --name-only "$BASE".."$HEAD" | grep -E '\.github/workflows/.*\.ya?ml$' || true)
-          if [ -n "$WORKFLOW_HITS" ]; then
-            FINDINGS="${FINDINGS}
-          ### ⚠️ WARNING: CI/CD workflow files modified
-          Changes to workflow files can alter build pipelines, inject steps, or modify permissions. Verify no unauthorized actions or secrets access were added.
-
-          **Files:**
-          \`\`\`
-          ${WORKFLOW_HITS}
-          \`\`\`
-          "
-          fi
-
-          # --- Dockerfile / container build files modified ---
-          DOCKER_HITS=$(git diff --name-only "$BASE".."$HEAD" | grep -iE '(Dockerfile|\.dockerignore|docker-compose)' || true)
-          if [ -n "$DOCKER_HITS" ]; then
-            FINDINGS="${FINDINGS}
-          ### ⚠️ WARNING: Container build files modified
-          Changes to Dockerfiles or compose files can alter base images, add build steps, or expose ports. Verify base image pins and build commands.
-
-          **Files:**
-          \`\`\`
-          ${DOCKER_HITS}
-          \`\`\`
-          "
-          fi
-
-          # --- Dependency manifest files modified ---
-          DEP_HITS=$(git diff --name-only "$BASE".."$HEAD" | grep -E '(pyproject\.toml|requirements.*\.txt|package\.json|Gemfile|go\.mod|Cargo\.toml)$' || true)
-          if [ -n "$DEP_HITS" ]; then
-            FINDINGS="${FINDINGS}
-          ### ⚠️ WARNING: Dependency manifest files modified
-          Changes to dependency files can introduce new packages or change version pins. Verify all dependency changes are intentional and from trusted sources.
-
-          **Files:**
-          \`\`\`
-          ${DEP_HITS}
-          \`\`\`
-          "
-          fi
-
-          # --- GitHub Actions version unpinning (mutable tags instead of SHAs) ---
-          ACTIONS_UNPIN=$(echo "$DIFF" | grep -n '^\+' | grep 'uses:' | grep -v '#' | grep -E '@v[0-9]' | head -10 || true)
-          if [ -n "$ACTIONS_UNPIN" ]; then
-            FINDINGS="${FINDINGS}
-          ### ⚠️ WARNING: GitHub Actions with mutable version tags
-          Actions should be pinned to full commit SHAs (not \`@v4\`, \`@v5\`). Mutable tags can be retargeted silently if a maintainer account is compromised.
-
-          **Matches:**
-          \`\`\`
-          ${ACTIONS_UNPIN}
-          \`\`\`
-          "
-          fi
-
-          # --- Output results ---
           if [ -n "$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
+      - name: Post critical finding 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="## 🚨 CRITICAL Supply Chain Risk Detected
 
-          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.
+          This PR contains a pattern that has been used in real supply chain attacks. A maintainer must review the flagged code carefully 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.*"
+          *Scanner only fires on high-signal indicators: .pth files, base64+exec/eval combos, subprocess with encoded commands, or install-hook files. Low-signal warnings were removed intentionally — if you're seeing this comment, the finding is worth inspecting.*"
 
           gh pr comment "${{ github.event.pull_request.number }}" --body "$BODY" || echo "::warning::Could not post PR comment (expected for fork PRs — GITHUB_TOKEN is read-only)"
 
       - name: Fail on critical findings
-        if: steps.scan.outputs.critical == 'true'
+        if: steps.scan.outputs.found == '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

@@ -3,8 +3,14 @@ name: Tests
 on:
   push:
     branches: [main]
+    paths-ignore:
+      - '**/*.md'
+      - 'docs/**'
   pull_request:
     branches: [main]
+    paths-ignore:
+      - '**/*.md'
+      - 'docs/**'
 
 permissions:
   contents: read
@@ -17,7 +23,7 @@ concurrency:
 jobs:
   test:
     runs-on: ubuntu-latest
-    timeout-minutes: 10
+    timeout-minutes: 20
     steps:
       - name: Checkout code
         uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5  # v4

.gitignore

@@ -54,11 +54,17 @@ environments/benchmarks/evals/
 # Web UI build output
 hermes_cli/web_dist/
 
+# Web UI assets — synced from @nous-research/ui at build time via
+# `npm run sync-assets` (see web/package.json).
+web/public/fonts/
+web/public/ds-assets/
+
 # Release script temp files
 .release_notes.md
 mini-swe-agent/
 
 # Nix
 .direnv/
+.nix-stamps/
 result
 website/static/api/skills-index.json

.mailmap

@@ -105,3 +105,4 @@ tesseracttars-creator <tesseracttars@gmail.com> <tesseracttars@gmail.com>
 xinbenlv <zzn+pa@zzn.im> <zzn+pa@zzn.im>
 SaulJWu <saul.jj.wu@gmail.com> <saul.jj.wu@gmail.com>
 angelos <angelos@oikos.lan.home.malaiwah.com> <angelos@oikos.lan.home.malaiwah.com>
+MestreY0d4-Uninter <241404605+MestreY0d4-Uninter@users.noreply.github.com> <MestreY0d4-Uninter@users.noreply.github.com>

AGENTS.md

@@ -56,6 +56,19 @@ hermes-agent/
 │   ├── run.py            # Main loop, slash commands, message dispatch
 │   ├── session.py        # SessionStore — conversation persistence
 │   └── platforms/        # Adapters: telegram, discord, slack, whatsapp, homeassistant, signal, qqbot
+├── ui-tui/               # Ink (React) terminal UI — `hermes --tui`
+│   ├── src/entry.tsx        # TTY gate + render()
+│   ├── src/app.tsx          # Main state machine and UI
+│   ├── src/gatewayClient.ts # Child process + JSON-RPC bridge
+│   ├── src/app/             # Decomposed app logic (event handler, slash handler, stores, hooks)
+│   ├── src/components/      # Ink components (branding, markdown, prompts, pickers, etc.)
+│   ├── src/hooks/           # useCompletion, useInputHistory, useQueue, useVirtualHistory
+│   └── src/lib/             # Pure helpers (history, osc52, text, rpc, messages)
+├── tui_gateway/          # Python JSON-RPC backend for the TUI
+│   ├── entry.py             # stdio entrypoint
+│   ├── server.py            # RPC handlers and session logic
+│   ├── render.py            # Optional rich/ANSI bridge
+│   └── slash_worker.py      # Persistent HermesCLI subprocess for slash commands
 ├── acp_adapter/          # ACP server (VS Code / Zed / JetBrains integration)
 ├── cron/                 # Scheduler (jobs.py, scheduler.py)
 ├── environments/         # RL training environments (Atropos)
@@ -179,6 +192,59 @@ if canonical == "mycommand":
 
 ---
 
+## TUI Architecture (ui-tui + tui_gateway)
+
+The TUI is a full replacement for the classic (prompt_toolkit) CLI, activated via `hermes --tui` or `HERMES_TUI=1`.
+
+### Process Model
+
+```
+hermes --tui
+  └─ Node (Ink)  ──stdio JSON-RPC──  Python (tui_gateway)
+       │                                  └─ AIAgent + tools + sessions
+       └─ renders transcript, composer, prompts, activity
+```
+
+TypeScript owns the screen. Python owns sessions, tools, model calls, and slash command logic.
+
+### Transport
+
+Newline-delimited JSON-RPC over stdio. Requests from Ink, events from Python. See `tui_gateway/server.py` for the full method/event catalog.
+
+### Key Surfaces
+
+| Surface | Ink component | Gateway method |
+|---------|---------------|----------------|
+| Chat streaming | `app.tsx` + `messageLine.tsx` | `prompt.submit` → `message.delta/complete` |
+| Tool activity | `thinking.tsx` | `tool.start/progress/complete` |
+| Approvals | `prompts.tsx` | `approval.respond` ← `approval.request` |
+| Clarify/sudo/secret | `prompts.tsx`, `maskedPrompt.tsx` | `clarify/sudo/secret.respond` |
+| Session picker | `sessionPicker.tsx` | `session.list/resume` |
+| Slash commands | Local handler + fallthrough | `slash.exec` → `_SlashWorker`, `command.dispatch` |
+| Completions | `useCompletion` hook | `complete.slash`, `complete.path` |
+| Theming | `theme.ts` + `branding.tsx` | `gateway.ready` with skin data |
+
+### Slash Command Flow
+
+1. Built-in client commands (`/help`, `/quit`, `/clear`, `/resume`, `/copy`, `/paste`, etc.) handled locally in `app.tsx`
+2. Everything else → `slash.exec` (runs in persistent `_SlashWorker` subprocess) → `command.dispatch` fallback
+
+### Dev Commands
+
+```bash
+cd ui-tui
+npm install       # first time
+npm run dev       # watch mode (rebuilds hermes-ink + tsx --watch)
+npm start         # production
+npm run build     # full build (hermes-ink + tsc)
+npm run type-check # typecheck only (tsc --noEmit)
+npm run lint      # eslint
+npm run fmt       # prettier
+npm test          # vitest
+```
+
+---
+
 ## Adding New Tools
 
 Requires changes in **2 files**:
@@ -458,13 +524,45 @@ def profile_env(tmp_path, monkeypatch):
 
 ## Testing
 
+**ALWAYS use `scripts/run_tests.sh`** — do not call `pytest` directly. The script enforces
+hermetic environment parity with CI (unset credential vars, TZ=UTC, LANG=C.UTF-8,
+4 xdist workers matching GHA ubuntu-latest). Direct `pytest` on a 16+ core
+developer machine with API keys set diverges from CI in ways that have caused
+multiple "works locally, fails in CI" incidents (and the reverse).
+
 ```bash
-source venv/bin/activate
-python -m pytest tests/ -q          # Full suite (~3000 tests, ~3 min)
-python -m pytest tests/test_model_tools.py -q   # Toolset resolution
-python -m pytest tests/test_cli_init.py -q       # CLI config loading
-python -m pytest tests/gateway/ -q               # Gateway tests
-python -m pytest tests/tools/ -q                 # Tool-level tests
+scripts/run_tests.sh                                  # full suite, CI-parity
+scripts/run_tests.sh tests/gateway/                   # one directory
+scripts/run_tests.sh tests/agent/test_foo.py::test_x  # one test
+scripts/run_tests.sh -v --tb=long                     # pass-through pytest flags
 ```
 
+### Why the wrapper (and why the old "just call pytest" doesn't work)
+
+Five real sources of local-vs-CI drift the script closes:
+
+| | Without wrapper | With wrapper |
+|---|---|---|
+| Provider API keys | Whatever is in your env (auto-detects pool) | All `*_API_KEY`/`*_TOKEN`/etc. unset |
+| HOME / `~/.hermes/` | Your real config+auth.json | Temp dir per test |
+| Timezone | Local TZ (PDT etc.) | UTC |
+| Locale | Whatever is set | C.UTF-8 |
+| xdist workers | `-n auto` = all cores (20+ on a workstation) | `-n 4` matching CI |
+
+`tests/conftest.py` also enforces points 1-4 as an autouse fixture so ANY pytest
+invocation (including IDE integrations) gets hermetic behavior — but the wrapper
+is belt-and-suspenders.
+
+### Running without the wrapper (only if you must)
+
+If you can't use the wrapper (e.g. on Windows or inside an IDE that shells
+pytest directly), at minimum activate the venv and pass `-n 4`:
+
+```bash
+source venv/bin/activate
+python -m pytest tests/ -q -n 4
+```
+
+Worker count above 4 will surface test-ordering flakes that CI never sees.
+
 Always run the full suite before pushing changes.

Dockerfile

@@ -21,26 +21,36 @@ RUN useradd -u 10000 -m -d /opt/data hermes
 COPY --chmod=0755 --from=gosu_source /gosu /usr/local/bin/
 COPY --chmod=0755 --from=uv_source /usr/local/bin/uv /usr/local/bin/uvx /usr/local/bin/
 
-COPY . /opt/hermes
 WORKDIR /opt/hermes
 
-# Install Node dependencies and Playwright as root (--with-deps needs apt)
+# ---------- Layer-cached dependency install ----------
+# Copy only package manifests first so npm install + Playwright are cached
+# unless the lockfiles themselves change.
+COPY package.json package-lock.json ./
+COPY scripts/whatsapp-bridge/package.json scripts/whatsapp-bridge/package-lock.json scripts/whatsapp-bridge/
+COPY web/package.json web/package-lock.json web/
+
 RUN 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 && \
+    (cd scripts/whatsapp-bridge && npm install --prefer-offline --no-audit) && \
+    (cd web && npm install --prefer-offline --no-audit) && \
     npm cache clean --force
 
-# Hand ownership to hermes user, then install Python deps in a virtualenv
-RUN chown -R hermes:hermes /opt/hermes
-USER hermes
+# ---------- Source code ----------
+# .dockerignore excludes node_modules, so the installs above survive.
+COPY --chown=hermes:hermes . .
 
+# Build web dashboard (Vite outputs to hermes_cli/web_dist/)
+RUN cd web && npm run build
+
+# ---------- Python virtualenv ----------
+RUN chown hermes:hermes /opt/hermes
+USER hermes
 RUN uv venv && \
     uv pip install --no-cache-dir -e ".[all]"
 
-USER root
-RUN chmod +x /opt/hermes/docker/entrypoint.sh
-
+# ---------- Runtime ----------
+ENV HERMES_WEB_DIST=/opt/hermes/hermes_cli/web_dist
 ENV HERMES_HOME=/opt/data
 VOLUME [ "/opt/data" ]
 ENTRYPOINT [ "/opt/hermes/docker/entrypoint.sh" ]

README.md

@@ -13,7 +13,7 @@
 
 **The self-improving AI agent built by [Nous Research](https://nousresearch.com).** It's the only agent with a built-in learning loop — it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions. Run it on a $5 VPS, a GPU cluster, or serverless infrastructure that costs nearly nothing when idle. It's not tied to your laptop — talk to it from Telegram while it works on a cloud VM.
 
-Use any model you want — [Nous Portal](https://portal.nousresearch.com), [OpenRouter](https://openrouter.ai) (200+ models), [Xiaomi MiMo](https://platform.xiaomimimo.com), [z.ai/GLM](https://z.ai), [Kimi/Moonshot](https://platform.moonshot.ai), [MiniMax](https://www.minimax.io), [Hugging Face](https://huggingface.co), OpenAI, or your own endpoint. Switch with `hermes model` — no code changes, no lock-in.
+Use any model you want — [Nous Portal](https://portal.nousresearch.com), [OpenRouter](https://openrouter.ai) (200+ models), [NVIDIA NIM](https://build.nvidia.com) (Nemotron), [Xiaomi MiMo](https://platform.xiaomimimo.com), [z.ai/GLM](https://z.ai), [Kimi/Moonshot](https://platform.moonshot.ai), [MiniMax](https://www.minimax.io), [Hugging Face](https://huggingface.co), OpenAI, or your own endpoint. Switch with `hermes model` — no code changes, no lock-in.
 
 <table>
 <tr><td><b>A real terminal interface</b></td><td>Full TUI with multiline editing, slash-command autocomplete, conversation history, interrupt-and-redirect, and streaming tool output.</td></tr>
@@ -141,11 +141,18 @@ See `hermes claw migrate --help` for all options, or use the `openclaw-migration
 
 We welcome contributions! See the [Contributing Guide](https://hermes-agent.nousresearch.com/docs/developer-guide/contributing) for development setup, code style, and PR process.
 
-Quick start for contributors:
+Quick start for contributors — clone and go with `setup-hermes.sh`:
 
 ```bash
 git clone https://github.com/NousResearch/hermes-agent.git
 cd hermes-agent
+./setup-hermes.sh     # installs uv, creates venv, installs .[all], symlinks ~/.local/bin/hermes
+./hermes              # auto-detects the venv, no need to `source` first
+```
+
+Manual path (equivalent to the above):
+
+```bash
 curl -LsSf https://astral.sh/uv/install.sh | sh
 uv venv venv --python 3.11
 source venv/bin/activate

RELEASE_v0.10.0.md

@@ -0,0 +1,27 @@
+# Hermes Agent v0.10.0 (v2026.4.16)
+
+**Release Date:** April 16, 2026
+
+> The Tool Gateway release — paid Nous Portal subscribers can now use web search, image generation, text-to-speech, and browser automation through their existing subscription with zero additional API keys.
+
+---
+
+## ✨ Highlights
+
+- **Nous Tool Gateway** — Paid [Nous Portal](https://portal.nousresearch.com) subscribers now get automatic access to **web search** (Firecrawl), **image generation** (FAL / FLUX 2 Pro), **text-to-speech** (OpenAI TTS), and **browser automation** (Browser Use) through their existing subscription. No separate API keys needed — just run `hermes model`, select Nous Portal, and pick which tools to enable. Per-tool opt-in via `use_gateway` config, full integration with `hermes tools` and `hermes status`, and the runtime correctly prefers the gateway even when direct API keys exist. Replaces the old hidden `HERMES_ENABLE_NOUS_MANAGED_TOOLS` env var with clean subscription-based detection. ([#11206](https://github.com/NousResearch/hermes-agent/pull/11206), based on work by @jquesnelle; docs: [#11208](https://github.com/NousResearch/hermes-agent/pull/11208))
+
+---
+
+## 🐛 Bug Fixes & Improvements
+
+This release includes 180+ commits with numerous bug fixes, platform improvements, and reliability enhancements across the agent core, gateway, CLI, and tool system. Full details will be published in the v0.11.0 changelog.
+
+---
+
+## 👥 Contributors
+
+- **@jquesnelle** (emozilla) — Original Tool Gateway implementation ([#10799](https://github.com/NousResearch/hermes-agent/pull/10799)), salvaged and shipped in this release
+
+---
+
+**Full Changelog**: [v2026.4.13...v2026.4.16](https://github.com/NousResearch/hermes-agent/compare/v2026.4.13...v2026.4.16)

SECURITY.md

@@ -0,0 +1,84 @@
+# Hermes Agent Security Policy
+
+This document outlines the security protocols, trust model, and deployment hardening guidelines for the **Hermes Agent** project.
+
+## 1. Vulnerability Reporting
+
+Hermes Agent does **not** operate a bug bounty program. Security issues should be reported via [GitHub Security Advisories (GHSA)](https://github.com/NousResearch/hermes-agent/security/advisories/new) or by emailing **security@nousresearch.com**. Do not open public issues for security vulnerabilities.
+
+### Required Submission Details
+- **Title & Severity:** Concise description and CVSS score/rating.
+- **Affected Component:** Exact file path and line range (e.g., `tools/approval.py:120-145`).
+- **Environment:** Output of `hermes version`, commit SHA, OS, and Python version.
+- **Reproduction:** Step-by-step Proof-of-Concept (PoC) against `main` or the latest release.
+- **Impact:** Explanation of what trust boundary was crossed.
+
+---
+
+## 2. Trust Model
+
+The core assumption is that Hermes is a **personal agent** with one trusted operator.
+
+### Operator & Session Trust
+- **Single Tenant:** The system protects the operator from LLM actions, not from malicious co-tenants. Multi-user isolation must happen at the OS/host level.
+- **Gateway Security:** Authorized callers (Telegram, Discord, Slack, etc.) receive equal trust. Session keys are used for routing, not as authorization boundaries.
+- **Execution:** Defaults to `terminal.backend: local` (direct host execution). Container isolation (Docker, Modal, Daytona) is opt-in for sandboxing.
+
+### Dangerous Command Approval
+The approval system (`tools/approval.py`) is a core security boundary. Terminal commands, file operations, and other potentially destructive actions are gated behind explicit user confirmation before execution. The approval mode is configurable via `approvals.mode` in `config.yaml`:
+- `"on"` (default) — prompts the user to approve dangerous commands.
+- `"auto"` — auto-approves after a configurable delay.
+- `"off"` — disables the gate entirely (break-glass; see Section 3).
+
+### Output Redaction
+`agent/redact.py` strips secret-like patterns (API keys, tokens, credentials) from all display output before it reaches the terminal or gateway platform. This prevents accidental credential leakage in chat logs, tool previews, and response text. Redaction operates on the display layer only — underlying values remain intact for internal agent operations.
+
+### Skills vs. MCP Servers
+- **Installed Skills:** High trust. Equivalent to local host code; skills can read environment variables and run arbitrary commands.
+- **MCP Servers:** Lower trust. MCP subprocesses receive a filtered environment (`_build_safe_env()` in `tools/mcp_tool.py`) — only safe baseline variables (`PATH`, `HOME`, `XDG_*`) plus variables explicitly declared in the server's `env` config block are passed through. Host credentials are stripped by default. Additionally, packages invoked via `npx`/`uvx` are checked against the OSV malware database before spawning.
+
+### Code Execution Sandbox
+The `execute_code` tool (`tools/code_execution_tool.py`) runs LLM-generated Python scripts in a child process with API keys and tokens stripped from the environment to prevent credential exfiltration. Only environment variables explicitly declared by loaded skills (via `env_passthrough`) or by the user in `config.yaml` (`terminal.env_passthrough`) are passed through. The child accesses Hermes tools via RPC, not direct API calls.
+
+### Subagents
+- **No recursive delegation:** The `delegate_task` tool is disabled for child agents.
+- **Depth limit:** `MAX_DEPTH = 2` — parent (depth 0) can spawn a child (depth 1); grandchildren are rejected.
+- **Memory isolation:** Subagents run with `skip_memory=True` and do not have access to the parent's persistent memory provider. The parent receives only the task prompt and final response as an observation.
+
+---
+
+## 3. Out of Scope (Non-Vulnerabilities)
+
+The following scenarios are **not** considered security breaches:
+- **Prompt Injection:** Unless it results in a concrete bypass of the approval system, toolset restrictions, or container sandbox.
+- **Public Exposure:** Deploying the gateway to the public internet without external authentication or network protection.
+- **Trusted State Access:** Reports that require pre-existing write access to `~/.hermes/`, `.env`, or `config.yaml` (these are operator-owned files).
+- **Default Behavior:** Host-level command execution when `terminal.backend` is set to `local` — this is the documented default, not a vulnerability.
+- **Configuration Trade-offs:** Intentional break-glass settings such as `approvals.mode: "off"` or `terminal.backend: local` in production.
+- **Tool-level read/access restrictions:** The agent has unrestricted shell access via the `terminal` tool by design. Reports that a specific tool (e.g., `read_file`) can access a resource are not vulnerabilities if the same access is available through `terminal`. Tool-level deny lists only constitute a meaningful security boundary when paired with equivalent restrictions on the terminal side (as with write operations, where `WRITE_DENIED_PATHS` is paired with the dangerous command approval system).
+
+---
+
+## 4. Deployment Hardening & Best Practices
+
+### Filesystem & Network
+- **Production sandboxing:** Use container backends (`docker`, `modal`, `daytona`) instead of `local` for untrusted workloads.
+- **File permissions:** Run as non-root (the Docker image uses UID 10000); protect credentials with `chmod 600 ~/.hermes/.env` on local installs.
+- **Network exposure:** Do not expose the gateway or API server to the public internet without VPN, Tailscale, or firewall protection. SSRF protection is enabled by default across all gateway platform adapters (Telegram, Discord, Slack, Matrix, Mattermost, etc.) with redirect validation. Note: the local terminal backend does not apply SSRF filtering, as it operates within the trusted operator's environment.
+
+### Skills & Supply Chain
+- **Skill installation:** Review Skills Guard reports (`tools/skills_guard.py`) before installing third-party skills. The audit log at `~/.hermes/skills/.hub/audit.log` tracks every install and removal.
+- **MCP safety:** OSV malware checking runs automatically for `npx`/`uvx` packages before MCP server processes are spawned.
+- **CI/CD:** GitHub Actions are pinned to full commit SHAs. The `supply-chain-audit.yml` workflow blocks PRs containing `.pth` files or suspicious `base64`+`exec` patterns.
+
+### Credential Storage
+- API keys and tokens belong exclusively in `~/.hermes/.env` — never in `config.yaml` or checked into version control.
+- The credential pool system (`agent/credential_pool.py`) handles key rotation and fallback. Credentials are resolved from environment variables, not stored in plaintext databases.
+
+---
+
+## 5. Disclosure Process
+
+- **Coordinated Disclosure:** 90-day window or until a fix is released, whichever comes first.
+- **Communication:** All updates occur via the GHSA thread or email correspondence with security@nousresearch.com.
+- **Credits:** Reporters are credited in release notes unless anonymity is requested.

acp_adapter/entry.py

@@ -20,6 +20,46 @@ from pathlib import Path
 from hermes_constants import get_hermes_home
 
 
+# Methods clients send as periodic liveness probes. They are not part of the
+# ACP schema, so the acp router correctly returns JSON-RPC -32601 to the
+# caller — but the supervisor task that dispatches the request then surfaces
+# the raised RequestError via ``logging.exception("Background task failed")``,
+# which dumps a traceback to stderr every probe interval. Clients like
+# acp-bridge already treat the -32601 response as "agent alive", so the
+# traceback is pure noise. We keep the protocol response intact and only
+# silence the stderr noise for this specific benign case.
+_BENIGN_PROBE_METHODS = frozenset({"ping", "health", "healthcheck"})
+
+
+class _BenignProbeMethodFilter(logging.Filter):
+    """Suppress acp 'Background task failed' tracebacks caused by unknown
+    liveness-probe methods (e.g. ``ping``) while leaving every other
+    background-task error — including method_not_found for any non-probe
+    method — visible in stderr.
+    """
+
+    def filter(self, record: logging.LogRecord) -> bool:
+        if record.getMessage() != "Background task failed":
+            return True
+        exc_info = record.exc_info
+        if not exc_info:
+            return True
+        exc = exc_info[1]
+        # Imported lazily so this module stays importable when the optional
+        # ``agent-client-protocol`` dependency is not installed.
+        try:
+            from acp.exceptions import RequestError
+        except ImportError:
+            return True
+        if not isinstance(exc, RequestError):
+            return True
+        if getattr(exc, "code", None) != -32601:
+            return True
+        data = getattr(exc, "data", None)
+        method = data.get("method") if isinstance(data, dict) else None
+        return method not in _BENIGN_PROBE_METHODS
+
+
 def _setup_logging() -> None:
     """Route all logging to stderr so stdout stays clean for ACP stdio."""
     handler = logging.StreamHandler(sys.stderr)
@@ -29,6 +69,7 @@ def _setup_logging() -> None:
             datefmt="%Y-%m-%d %H:%M:%S",
         )
     )
+    handler.addFilter(_BenignProbeMethodFilter())
     root = logging.getLogger()
     root.handlers.clear()
     root.addHandler(handler)

acp_adapter/events.py

@@ -49,6 +49,7 @@ def make_tool_progress_cb(
     session_id: str,
     loop: asyncio.AbstractEventLoop,
     tool_call_ids: Dict[str, Deque[str]],
+    tool_call_meta: Dict[str, Dict[str, Any]],
 ) -> Callable:
     """Create a ``tool_progress_callback`` for AIAgent.
 
@@ -84,6 +85,16 @@ def make_tool_progress_cb(
             tool_call_ids[name] = queue
         queue.append(tc_id)
 
+        snapshot = None
+        if name in {"write_file", "patch", "skill_manage"}:
+            try:
+                from agent.display import capture_local_edit_snapshot
+
+                snapshot = capture_local_edit_snapshot(name, args)
+            except Exception:
+                logger.debug("Failed to capture ACP edit snapshot for %s", name, exc_info=True)
+        tool_call_meta[tc_id] = {"args": args, "snapshot": snapshot}
+
         update = build_tool_start(tc_id, name, args)
         _send_update(conn, session_id, loop, update)
 
@@ -119,6 +130,7 @@ def make_step_cb(
     session_id: str,
     loop: asyncio.AbstractEventLoop,
     tool_call_ids: Dict[str, Deque[str]],
+    tool_call_meta: Dict[str, Dict[str, Any]],
 ) -> Callable:
     """Create a ``step_callback`` for AIAgent.
 
@@ -132,10 +144,12 @@ def make_step_cb(
             for tool_info in prev_tools:
                 tool_name = None
                 result = None
+                function_args = None
 
                 if isinstance(tool_info, dict):
                     tool_name = tool_info.get("name") or tool_info.get("function_name")
                     result = tool_info.get("result") or tool_info.get("output")
+                    function_args = tool_info.get("arguments") or tool_info.get("args")
                 elif isinstance(tool_info, str):
                     tool_name = tool_info
 
@@ -145,8 +159,13 @@ def make_step_cb(
                     tool_call_ids[tool_name] = queue
                 if tool_name and queue:
                     tc_id = queue.popleft()
+                    meta = tool_call_meta.pop(tc_id, {})
                     update = build_tool_complete(
-                        tc_id, tool_name, result=str(result) if result is not None else None
+                        tc_id,
+                        tool_name,
+                        result=str(result) if result is not None else None,
+                        function_args=function_args or meta.get("args"),
+                        snapshot=meta.get("snapshot"),
                     )
                     _send_update(conn, session_id, loop, update)
                     if not queue:

acp_adapter/server.py

@@ -26,6 +26,7 @@ from acp.schema import (
     McpServerHttp,
     McpServerSse,
     McpServerStdio,
+    ModelInfo,
     NewSessionResponse,
     PromptResponse,
     ResumeSessionResponse,
@@ -36,6 +37,7 @@ from acp.schema import (
     SessionCapabilities,
     SessionForkCapabilities,
     SessionListCapabilities,
+    SessionModelState,
     SessionResumeCapabilities,
     SessionInfo,
     TextContentBlock,
@@ -147,6 +149,98 @@ class HermesACPAgent(acp.Agent):
         self._conn = conn
         logger.info("ACP client connected")
 
+    @staticmethod
+    def _encode_model_choice(provider: str | None, model: str | None) -> str:
+        """Encode a model selection so ACP clients can keep provider context."""
+        raw_model = str(model or "").strip()
+        if not raw_model:
+            return ""
+        raw_provider = str(provider or "").strip().lower()
+        if not raw_provider:
+            return raw_model
+        return f"{raw_provider}:{raw_model}"
+
+    def _build_model_state(self, state: SessionState) -> SessionModelState | None:
+        """Return the ACP model selector payload for editors like Zed."""
+        model = str(state.model or getattr(state.agent, "model", "") or "").strip()
+        provider = getattr(state.agent, "provider", None) or detect_provider() or "openrouter"
+
+        try:
+            from hermes_cli.models import curated_models_for_provider, normalize_provider, provider_label
+
+            normalized_provider = normalize_provider(provider)
+            provider_name = provider_label(normalized_provider)
+            available_models: list[ModelInfo] = []
+            seen_ids: set[str] = set()
+
+            for model_id, description in curated_models_for_provider(normalized_provider):
+                rendered_model = str(model_id or "").strip()
+                if not rendered_model:
+                    continue
+                choice_id = self._encode_model_choice(normalized_provider, rendered_model)
+                if choice_id in seen_ids:
+                    continue
+                desc_parts = [f"Provider: {provider_name}"]
+                if description:
+                    desc_parts.append(str(description).strip())
+                if rendered_model == model:
+                    desc_parts.append("current")
+                available_models.append(
+                    ModelInfo(
+                        model_id=choice_id,
+                        name=rendered_model,
+                        description=" • ".join(part for part in desc_parts if part),
+                    )
+                )
+                seen_ids.add(choice_id)
+
+            current_model_id = self._encode_model_choice(normalized_provider, model)
+            if current_model_id and current_model_id not in seen_ids:
+                available_models.insert(
+                    0,
+                    ModelInfo(
+                        model_id=current_model_id,
+                        name=model,
+                        description=f"Provider: {provider_name} • current",
+                    ),
+                )
+
+            if available_models:
+                return SessionModelState(
+                    available_models=available_models,
+                    current_model_id=current_model_id or available_models[0].model_id,
+                )
+        except Exception:
+            logger.debug("Could not build ACP model state", exc_info=True)
+
+        if not model:
+            return None
+
+        fallback_choice = self._encode_model_choice(provider, model)
+        return SessionModelState(
+            available_models=[ModelInfo(model_id=fallback_choice, name=model)],
+            current_model_id=fallback_choice,
+        )
+
+    @staticmethod
+    def _resolve_model_selection(raw_model: str, current_provider: str) -> tuple[str, str]:
+        """Resolve ``provider:model`` input into the provider and normalized model id."""
+        target_provider = current_provider
+        new_model = raw_model.strip()
+
+        try:
+            from hermes_cli.models import detect_provider_for_model, parse_model_input
+
+            target_provider, new_model = parse_model_input(new_model, current_provider)
+            if target_provider == current_provider:
+                detected = detect_provider_for_model(new_model, current_provider)
+                if detected:
+                    target_provider, new_model = detected
+        except Exception:
+            logger.debug("Provider detection failed, using model as-is", exc_info=True)
+
+        return target_provider, new_model
+
     async def _register_session_mcp_servers(
         self,
         state: SessionState,
@@ -273,7 +367,10 @@ class HermesACPAgent(acp.Agent):
         await self._register_session_mcp_servers(state, mcp_servers)
         logger.info("New session %s (cwd=%s)", state.session_id, cwd)
         self._schedule_available_commands_update(state.session_id)
-        return NewSessionResponse(session_id=state.session_id)
+        return NewSessionResponse(
+            session_id=state.session_id,
+            models=self._build_model_state(state),
+        )
 
     async def load_session(
         self,
@@ -289,7 +386,7 @@ class HermesACPAgent(acp.Agent):
         await self._register_session_mcp_servers(state, mcp_servers)
         logger.info("Loaded session %s", session_id)
         self._schedule_available_commands_update(session_id)
-        return LoadSessionResponse()
+        return LoadSessionResponse(models=self._build_model_state(state))
 
     async def resume_session(
         self,
@@ -305,7 +402,7 @@ class HermesACPAgent(acp.Agent):
         await self._register_session_mcp_servers(state, mcp_servers)
         logger.info("Resumed session %s", state.session_id)
         self._schedule_available_commands_update(state.session_id)
-        return ResumeSessionResponse()
+        return ResumeSessionResponse(models=self._build_model_state(state))
 
     async def cancel(self, session_id: str, **kwargs: Any) -> None:
         state = self.session_manager.get_session(session_id)
@@ -340,11 +437,20 @@ class HermesACPAgent(acp.Agent):
         cwd: str | None = None,
         **kwargs: Any,
     ) -> ListSessionsResponse:
-        infos = self.session_manager.list_sessions()
-        sessions = [
-            SessionInfo(session_id=s["session_id"], cwd=s["cwd"])
-            for s in infos
-        ]
+        infos = self.session_manager.list_sessions(cwd=cwd)
+        sessions = []
+        for s in infos:
+            updated_at = s.get("updated_at")
+            if updated_at is not None and not isinstance(updated_at, str):
+                updated_at = str(updated_at)
+            sessions.append(
+                SessionInfo(
+                    session_id=s["session_id"],
+                    cwd=s["cwd"],
+                    title=s.get("title"),
+                    updated_at=updated_at,
+                )
+            )
         return ListSessionsResponse(sessions=sessions)
 
     # ---- Prompt (core) ------------------------------------------------------
@@ -389,12 +495,13 @@ class HermesACPAgent(acp.Agent):
             state.cancel_event.clear()
 
         tool_call_ids: dict[str, Deque[str]] = defaultdict(deque)
+        tool_call_meta: dict[str, dict[str, Any]] = {}
         previous_approval_cb = None
 
         if conn:
-            tool_progress_cb = make_tool_progress_cb(conn, session_id, loop, tool_call_ids)
+            tool_progress_cb = make_tool_progress_cb(conn, session_id, loop, tool_call_ids, tool_call_meta)
             thinking_cb = make_thinking_cb(conn, session_id, loop)
-            step_cb = make_step_cb(conn, session_id, loop, tool_call_ids)
+            step_cb = make_step_cb(conn, session_id, loop, tool_call_ids, tool_call_meta)
             message_cb = make_message_cb(conn, session_id, loop)
             approval_cb = make_approval_callback(conn.request_permission, loop, session_id)
         else:
@@ -449,6 +556,19 @@ class HermesACPAgent(acp.Agent):
             self.session_manager.save_session(session_id)
 
         final_response = result.get("final_response", "")
+        if final_response:
+            try:
+                from agent.title_generator import maybe_auto_title
+
+                maybe_auto_title(
+                    self.session_manager._get_db(),
+                    session_id,
+                    user_text,
+                    final_response,
+                    state.history,
+                )
+            except Exception:
+                logger.debug("Failed to auto-title ACP session %s", session_id, exc_info=True)
         if final_response and conn:
             update = acp.update_agent_message_text(final_response)
             await conn.session_update(session_id, update)
@@ -556,27 +676,15 @@ class HermesACPAgent(acp.Agent):
             provider = getattr(state.agent, "provider", None) or "auto"
             return f"Current model: {model}\nProvider: {provider}"
 
-        new_model = args.strip()
-        target_provider = None
         current_provider = getattr(state.agent, "provider", None) or "openrouter"
-
-        # Auto-detect provider for the requested model
-        try:
-            from hermes_cli.models import parse_model_input, detect_provider_for_model
-            target_provider, new_model = parse_model_input(new_model, current_provider)
-            if target_provider == current_provider:
-                detected = detect_provider_for_model(new_model, current_provider)
-                if detected:
-                    target_provider, new_model = detected
-        except Exception:
-            logger.debug("Provider detection failed, using model as-is", exc_info=True)
+        target_provider, new_model = self._resolve_model_selection(args, current_provider)
 
         state.model = new_model
         state.agent = self.session_manager._make_agent(
             session_id=state.session_id,
             cwd=state.cwd,
             model=new_model,
-            requested_provider=target_provider or current_provider,
+            requested_provider=target_provider,
         )
         self.session_manager.save_session(state.session_id)
         provider_label = getattr(state.agent, "provider", None) or target_provider or current_provider
@@ -678,20 +786,30 @@ class HermesACPAgent(acp.Agent):
         """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)
+            requested_provider, resolved_model = self._resolve_model_selection(
+                model_id,
+                current_provider or "openrouter",
+            )
+            state.model = resolved_model
+            provider_changed = bool(current_provider and requested_provider != current_provider)
+            current_base_url = None if provider_changed else getattr(state.agent, "base_url", None)
+            current_api_mode = None if provider_changed else 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,
+                model=resolved_model,
+                requested_provider=requested_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)
+            logger.info(
+                "Session %s: model switched to %s via provider %s",
+                session_id,
+                resolved_model,
+                requested_provider,
+            )
             return SetSessionModelResponse()
         logger.warning("Session %s: model switch requested for missing session", session_id)
         return None

acp_adapter/session.py

@@ -13,8 +13,12 @@ from hermes_constants import get_hermes_home
 import copy
 import json
 import logging
+import os
+import re
 import sys
+import time
 import uuid
+from datetime import datetime, timezone
 from dataclasses import dataclass, field
 from threading import Lock
 from typing import Any, Dict, List, Optional
@@ -22,6 +26,64 @@ from typing import Any, Dict, List, Optional
 logger = logging.getLogger(__name__)
 
 
+def _normalize_cwd_for_compare(cwd: str | None) -> str:
+    raw = str(cwd or ".").strip()
+    if not raw:
+        raw = "."
+    expanded = os.path.expanduser(raw)
+
+    # Normalize Windows drive paths into the equivalent WSL mount form so
+    # ACP history filters match the same workspace across Windows and WSL.
+    match = re.match(r"^([A-Za-z]):[\\/](.*)$", expanded)
+    if match:
+        drive = match.group(1).lower()
+        tail = match.group(2).replace("\\", "/")
+        expanded = f"/mnt/{drive}/{tail}"
+    elif re.match(r"^/mnt/[A-Za-z]/", expanded):
+        expanded = f"/mnt/{expanded[5].lower()}/{expanded[7:]}"
+
+    return os.path.normpath(expanded)
+
+
+def _build_session_title(title: Any, preview: Any, cwd: str | None) -> str:
+    explicit = str(title or "").strip()
+    if explicit:
+        return explicit
+    preview_text = str(preview or "").strip()
+    if preview_text:
+        return preview_text
+    leaf = os.path.basename(str(cwd or "").rstrip("/\\"))
+    return leaf or "New thread"
+
+
+def _format_updated_at(value: Any) -> str | None:
+    if value is None:
+        return None
+    if isinstance(value, str) and value.strip():
+        return value
+    try:
+        return datetime.fromtimestamp(float(value), tz=timezone.utc).isoformat()
+    except Exception:
+        return None
+
+
+def _updated_at_sort_key(value: Any) -> float:
+    if value is None:
+        return float("-inf")
+    if isinstance(value, (int, float)):
+        return float(value)
+    raw = str(value).strip()
+    if not raw:
+        return float("-inf")
+    try:
+        return datetime.fromisoformat(raw.replace("Z", "+00:00")).timestamp()
+    except Exception:
+        try:
+            return float(raw)
+        except Exception:
+            return float("-inf")
+
+
 def _acp_stderr_print(*args, **kwargs) -> None:
     """Best-effort human-readable output sink for ACP stdio sessions.
 
@@ -162,47 +224,78 @@ class SessionManager:
         logger.info("Forked ACP session %s -> %s", session_id, new_id)
         return state
 
-    def list_sessions(self) -> List[Dict[str, Any]]:
+    def list_sessions(self, cwd: str | None = None) -> List[Dict[str, Any]]:
         """Return lightweight info dicts for all sessions (memory + database)."""
+        normalized_cwd = _normalize_cwd_for_compare(cwd) if cwd else None
+        db = self._get_db()
+        persisted_rows: dict[str, dict[str, Any]] = {}
+
+        if db is not None:
+            try:
+                for row in db.list_sessions_rich(source="acp", limit=1000):
+                    persisted_rows[str(row["id"])] = dict(row)
+            except Exception:
+                logger.debug("Failed to load ACP sessions from DB", exc_info=True)
+
         # Collect in-memory sessions first.
         with self._lock:
             seen_ids = set(self._sessions.keys())
-            results = [
-                {
-                    "session_id": s.session_id,
-                    "cwd": s.cwd,
-                    "model": s.model,
-                    "history_len": len(s.history),
-                }
-                for s in self._sessions.values()
-            ]
+            results = []
+            for s in self._sessions.values():
+                history_len = len(s.history)
+                if history_len <= 0:
+                    continue
+                if normalized_cwd and _normalize_cwd_for_compare(s.cwd) != normalized_cwd:
+                    continue
+                persisted = persisted_rows.get(s.session_id, {})
+                preview = next(
+                    (
+                        str(msg.get("content") or "").strip()
+                        for msg in s.history
+                        if msg.get("role") == "user" and str(msg.get("content") or "").strip()
+                    ),
+                    persisted.get("preview") or "",
+                )
+                results.append(
+                    {
+                        "session_id": s.session_id,
+                        "cwd": s.cwd,
+                        "model": s.model,
+                        "history_len": history_len,
+                        "title": _build_session_title(persisted.get("title"), preview, s.cwd),
+                        "updated_at": _format_updated_at(
+                            persisted.get("last_active") or persisted.get("started_at") or time.time()
+                        ),
+                    }
+                )
 
         # Merge any persisted sessions not currently in memory.
-        db = self._get_db()
-        if db is not None:
-            try:
-                rows = db.search_sessions(source="acp", limit=1000)
-                for row in rows:
-                    sid = row["id"]
-                    if sid in seen_ids:
-                        continue
-                    # Extract cwd from model_config JSON.
-                    cwd = "."
-                    mc = row.get("model_config")
-                    if mc:
-                        try:
-                            cwd = json.loads(mc).get("cwd", ".")
-                        except (json.JSONDecodeError, TypeError):
-                            pass
-                    results.append({
-                        "session_id": sid,
-                        "cwd": cwd,
-                        "model": row.get("model") or "",
-                        "history_len": row.get("message_count") or 0,
-                    })
-            except Exception:
-                logger.debug("Failed to list ACP sessions from DB", exc_info=True)
+        for sid, row in persisted_rows.items():
+            if sid in seen_ids:
+                continue
+            message_count = int(row.get("message_count") or 0)
+            if message_count <= 0:
+                continue
+            # Extract cwd from model_config JSON.
+            session_cwd = "."
+            mc = row.get("model_config")
+            if mc:
+                try:
+                    session_cwd = json.loads(mc).get("cwd", ".")
+                except (json.JSONDecodeError, TypeError):
+                    pass
+            if normalized_cwd and _normalize_cwd_for_compare(session_cwd) != normalized_cwd:
+                continue
+            results.append({
+                "session_id": sid,
+                "cwd": session_cwd,
+                "model": row.get("model") or "",
+                "history_len": message_count,
+                "title": _build_session_title(row.get("title"), row.get("preview"), session_cwd),
+                "updated_at": _format_updated_at(row.get("last_active") or row.get("started_at")),
+            })
 
+        results.sort(key=lambda item: _updated_at_sort_key(item.get("updated_at")), reverse=True)
         return results
 
     def update_cwd(self, session_id: str, cwd: str) -> Optional[SessionState]:

acp_adapter/tools.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import json
 import uuid
 from typing import Any, Dict, List, Optional
 
@@ -96,6 +97,170 @@ def build_tool_title(tool_name: str, args: Dict[str, Any]) -> str:
     return tool_name
 
 
+def _build_patch_mode_content(patch_text: str) -> List[Any]:
+    """Parse V4A patch mode input into ACP diff blocks when possible."""
+    if not patch_text:
+        return [acp.tool_content(acp.text_block(""))]
+
+    try:
+        from tools.patch_parser import OperationType, parse_v4a_patch
+
+        operations, error = parse_v4a_patch(patch_text)
+        if error or not operations:
+            return [acp.tool_content(acp.text_block(patch_text))]
+
+        content: List[Any] = []
+        for op in operations:
+            if op.operation == OperationType.UPDATE:
+                old_chunks: list[str] = []
+                new_chunks: list[str] = []
+                for hunk in op.hunks:
+                    old_lines = [line.content for line in hunk.lines if line.prefix in (" ", "-")]
+                    new_lines = [line.content for line in hunk.lines if line.prefix in (" ", "+")]
+                    if old_lines or new_lines:
+                        old_chunks.append("\n".join(old_lines))
+                        new_chunks.append("\n".join(new_lines))
+
+                old_text = "\n...\n".join(chunk for chunk in old_chunks if chunk)
+                new_text = "\n...\n".join(chunk for chunk in new_chunks if chunk)
+                if old_text or new_text:
+                    content.append(
+                        acp.tool_diff_content(
+                            path=op.file_path,
+                            old_text=old_text or None,
+                            new_text=new_text or "",
+                        )
+                    )
+                continue
+
+            if op.operation == OperationType.ADD:
+                added_lines = [line.content for hunk in op.hunks for line in hunk.lines if line.prefix == "+"]
+                content.append(
+                    acp.tool_diff_content(
+                        path=op.file_path,
+                        new_text="\n".join(added_lines),
+                    )
+                )
+                continue
+
+            if op.operation == OperationType.DELETE:
+                content.append(
+                    acp.tool_diff_content(
+                        path=op.file_path,
+                        old_text=f"Delete file: {op.file_path}",
+                        new_text="",
+                    )
+                )
+                continue
+
+            if op.operation == OperationType.MOVE:
+                content.append(
+                    acp.tool_content(acp.text_block(f"Move file: {op.file_path} -> {op.new_path}"))
+                )
+
+        return content or [acp.tool_content(acp.text_block(patch_text))]
+    except Exception:
+        return [acp.tool_content(acp.text_block(patch_text))]
+
+
+def _strip_diff_prefix(path: str) -> str:
+    raw = str(path or "").strip()
+    if raw.startswith(("a/", "b/")):
+        return raw[2:]
+    return raw
+
+
+def _parse_unified_diff_content(diff_text: str) -> List[Any]:
+    """Convert unified diff text into ACP diff content blocks."""
+    if not diff_text:
+        return []
+
+    content: List[Any] = []
+    current_old_path: Optional[str] = None
+    current_new_path: Optional[str] = None
+    old_lines: list[str] = []
+    new_lines: list[str] = []
+
+    def _flush() -> None:
+        nonlocal current_old_path, current_new_path, old_lines, new_lines
+        if current_old_path is None and current_new_path is None:
+            return
+        path = current_new_path if current_new_path and current_new_path != "/dev/null" else current_old_path
+        if not path or path == "/dev/null":
+            current_old_path = None
+            current_new_path = None
+            old_lines = []
+            new_lines = []
+            return
+        content.append(
+            acp.tool_diff_content(
+                path=_strip_diff_prefix(path),
+                old_text="\n".join(old_lines) if old_lines else None,
+                new_text="\n".join(new_lines),
+            )
+        )
+        current_old_path = None
+        current_new_path = None
+        old_lines = []
+        new_lines = []
+
+    for line in diff_text.splitlines():
+        if line.startswith("--- "):
+            _flush()
+            current_old_path = line[4:].strip()
+            continue
+        if line.startswith("+++ "):
+            current_new_path = line[4:].strip()
+            continue
+        if line.startswith("@@"):
+            continue
+        if current_old_path is None and current_new_path is None:
+            continue
+        if line.startswith("+"):
+            new_lines.append(line[1:])
+        elif line.startswith("-"):
+            old_lines.append(line[1:])
+        elif line.startswith(" "):
+            shared = line[1:]
+            old_lines.append(shared)
+            new_lines.append(shared)
+
+    _flush()
+    return content
+
+
+def _build_tool_complete_content(
+    tool_name: str,
+    result: Optional[str],
+    *,
+    function_args: Optional[Dict[str, Any]] = None,
+    snapshot: Any = None,
+) -> List[Any]:
+    """Build structured ACP completion content, falling back to plain text."""
+    display_result = result or ""
+    if len(display_result) > 5000:
+        display_result = display_result[:4900] + f"\n... ({len(result)} chars total, truncated)"
+
+    if tool_name in {"write_file", "patch", "skill_manage"}:
+        try:
+            from agent.display import extract_edit_diff
+
+            diff_text = extract_edit_diff(
+                tool_name,
+                result,
+                function_args=function_args,
+                snapshot=snapshot,
+            )
+            if isinstance(diff_text, str) and diff_text.strip():
+                diff_content = _parse_unified_diff_content(diff_text)
+                if diff_content:
+                    return diff_content
+        except Exception:
+            pass
+
+    return [acp.tool_content(acp.text_block(display_result))]
+
+
 # ---------------------------------------------------------------------------
 # Build ACP content objects for tool-call events
 # ---------------------------------------------------------------------------
@@ -119,9 +284,8 @@ def build_tool_start(
             new = arguments.get("new_string", "")
             content = [acp.tool_diff_content(path=path, new_text=new, old_text=old)]
         else:
-            # Patch mode — show the patch content as text
             patch_text = arguments.get("patch", "")
-            content = [acp.tool_content(acp.text_block(patch_text))]
+            content = _build_patch_mode_content(patch_text)
         return acp.start_tool_call(
             tool_call_id, title, kind=kind, content=content, locations=locations,
             raw_input=arguments,
@@ -178,16 +342,17 @@ def build_tool_complete(
     tool_call_id: str,
     tool_name: str,
     result: Optional[str] = None,
+    function_args: Optional[Dict[str, Any]] = None,
+    snapshot: Any = None,
 ) -> ToolCallProgress:
     """Create a ToolCallUpdate (progress) event for a completed tool call."""
     kind = get_tool_kind(tool_name)
-
-    # Truncate very large results for the UI
-    display_result = result or ""
-    if len(display_result) > 5000:
-        display_result = display_result[:4900] + f"\n... ({len(result)} chars total, truncated)"
-
-    content = [acp.tool_content(acp.text_block(display_result))]
+    content = _build_tool_complete_content(
+        tool_name,
+        result,
+        function_args=function_args,
+        snapshot=snapshot,
+    )
     return acp.update_tool_call(
         tool_call_id,
         kind=kind,

agent/anthropic_adapter.py

@@ -28,19 +28,45 @@ except ImportError:
 logger = logging.getLogger(__name__)
 
 THINKING_BUDGET = {"xhigh": 32000, "high": 16000, "medium": 8000, "low": 4000}
+# Hermes effort → Anthropic adaptive-thinking effort (output_config.effort).
+# Anthropic exposes 5 levels on 4.7+: low, medium, high, xhigh, max.
+# Opus/Sonnet 4.6 only expose 4 levels: low, medium, high, max — no xhigh.
+# We preserve xhigh as xhigh on 4.7+ (the recommended default for coding/
+# agentic work) and downgrade it to max on pre-4.7 adaptive models (which
+# is the strongest level they accept).  "minimal" is a legacy alias that
+# maps to low on every model.  See:
+# https://platform.claude.com/docs/en/about-claude/models/migration-guide
 ADAPTIVE_EFFORT_MAP = {
-    "xhigh": "max",
-    "high": "high",
-    "medium": "medium",
-    "low": "low",
+    "max":     "max",
+    "xhigh":   "xhigh",
+    "high":    "high",
+    "medium":  "medium",
+    "low":     "low",
     "minimal": "low",
 }
 
+# Models that accept the "xhigh" output_config.effort level.  Opus 4.7 added
+# xhigh as a distinct level between high and max; older adaptive-thinking
+# models (4.6) reject it with a 400.  Keep this substring list in sync with
+# the Anthropic migration guide as new model families ship.
+_XHIGH_EFFORT_SUBSTRINGS = ("4-7", "4.7")
+
+# Models where extended thinking is deprecated/removed (4.6+ behavior: adaptive
+# is the only supported mode; 4.7 additionally forbids manual thinking entirely
+# and drops temperature/top_p/top_k).
+_ADAPTIVE_THINKING_SUBSTRINGS = ("4-6", "4.6", "4-7", "4.7")
+
+# Models where temperature/top_p/top_k return 400 if set to non-default values.
+# This is the Opus 4.7 contract; future 4.x+ models are expected to follow it.
+_NO_SAMPLING_PARAMS_SUBSTRINGS = ("4-7", "4.7")
+
 # ── 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 4.7
+    "claude-opus-4-7":   128_000,
     # Claude 4.6
     "claude-opus-4-6":   128_000,
     "claude-sonnet-4-6":  64_000,
@@ -91,11 +117,37 @@ def _get_anthropic_max_output(model: str) -> int:
 
 
 def _supports_adaptive_thinking(model: str) -> bool:
-    """Return True for Claude 4.6 models that support adaptive thinking."""
-    return any(v in model for v in ("4-6", "4.6"))
+    """Return True for Claude 4.6+ models that support adaptive thinking."""
+    return any(v in model for v in _ADAPTIVE_THINKING_SUBSTRINGS)
 
 
-# Beta headers for enhanced features (sent with ALL auth types)
+def _supports_xhigh_effort(model: str) -> bool:
+    """Return True for models that accept the 'xhigh' adaptive effort level.
+
+    Opus 4.7 introduced xhigh as a distinct level between high and max.
+    Pre-4.7 adaptive models (Opus/Sonnet 4.6) only accept low/medium/high/max
+    and reject xhigh with an HTTP 400. Callers should downgrade xhigh→max
+    when this returns False.
+    """
+    return any(v in model for v in _XHIGH_EFFORT_SUBSTRINGS)
+
+
+def _forbids_sampling_params(model: str) -> bool:
+    """Return True for models that 400 on any non-default temperature/top_p/top_k.
+
+    Opus 4.7 explicitly rejects sampling parameters; later Claude releases are
+    expected to follow suit.  Callers should omit these fields entirely rather
+    than passing zero/default values (the API rejects anything non-null).
+    """
+    return any(v in model for v in _NO_SAMPLING_PARAMS_SUBSTRINGS)
+
+
+# Beta headers for enhanced features (sent with ALL auth types).
+# As of Opus 4.7 (2026-04-16), both of these are GA on Claude 4.6+ — the
+# beta headers are still accepted (harmless no-op) but not required. Kept
+# here so older Claude (4.5, 4.1) + third-party Anthropic-compat endpoints
+# that still gate on the headers continue to get the enhanced features.
+# Migration guide: remove these if you no longer support ≤4.5 models.
 _COMMON_BETAS = [
     "interleaved-thinking-2025-05-14",
     "fine-grained-tool-streaming-2025-05-14",
@@ -240,9 +292,15 @@ def _common_betas_for_base_url(base_url: str | None) -> list[str]:
     return _COMMON_BETAS
 
 
-def build_anthropic_client(api_key: str, base_url: str = None):
+def build_anthropic_client(api_key: str, base_url: str = None, timeout: float = None):
     """Create an Anthropic client, auto-detecting setup-tokens vs API keys.
 
+    If *timeout* is provided it overrides the default 900s read timeout.  The
+    connect timeout stays at 10s.  Callers pass this from the per-provider /
+    per-model ``request_timeout_seconds`` config so Anthropic-native and
+    Anthropic-compatible providers respect the same knob as OpenAI-wire
+    providers.
+
     Returns an anthropic.Anthropic instance.
     """
     if _anthropic_sdk is None:
@@ -253,8 +311,9 @@ def build_anthropic_client(api_key: str, base_url: str = None):
     from httpx import Timeout
 
     normalized_base_url = _normalize_base_url_text(base_url)
+    _read_timeout = timeout if (isinstance(timeout, (int, float)) and timeout > 0) else 900.0
     kwargs = {
-        "timeout": Timeout(timeout=900.0, connect=10.0),
+        "timeout": Timeout(timeout=float(_read_timeout), connect=10.0),
     }
     if normalized_base_url:
         kwargs["base_url"] = normalized_base_url
@@ -298,6 +357,33 @@ def build_anthropic_client(api_key: str, base_url: str = None):
     return _anthropic_sdk.Anthropic(**kwargs)
 
 
+def build_anthropic_bedrock_client(region: str):
+    """Create an AnthropicBedrock client for Bedrock Claude models.
+
+    Uses the Anthropic SDK's native Bedrock adapter, which provides full
+    Claude feature parity: prompt caching, thinking budgets, adaptive
+    thinking, fast mode — features not available via the Converse API.
+
+    Auth uses the boto3 default credential chain (IAM roles, SSO, env vars).
+    """
+    if _anthropic_sdk is None:
+        raise ImportError(
+            "The 'anthropic' package is required for the Bedrock provider. "
+            "Install it with: pip install 'anthropic>=0.39.0'"
+        )
+    if not hasattr(_anthropic_sdk, "AnthropicBedrock"):
+        raise ImportError(
+            "anthropic.AnthropicBedrock not available. "
+            "Upgrade with: pip install 'anthropic>=0.39.0'"
+        )
+    from httpx import Timeout
+
+    return _anthropic_sdk.AnthropicBedrock(
+        aws_region=region,
+        timeout=Timeout(timeout=900.0, connect=10.0),
+    )
+
+
 def read_claude_code_credentials() -> Optional[Dict[str, Any]]:
     """Read refreshable Claude Code OAuth credentials from ~/.claude/.credentials.json.
 
@@ -1314,18 +1400,31 @@ def build_anthropic_kwargs(
             kwargs["tool_choice"] = {"type": "tool", "name": tool_choice}
 
     # Map reasoning_config to Anthropic's thinking parameter.
-    # Claude 4.6 models use adaptive thinking + output_config.effort.
+    # Claude 4.6+ models use adaptive thinking + output_config.effort.
     # Older models use manual thinking with budget_tokens.
     # MiniMax Anthropic-compat endpoints support thinking (manual mode only,
     # not adaptive).  Haiku does NOT support extended thinking — skip entirely.
+    #
+    # On 4.7+ the `thinking.display` field defaults to "omitted", which
+    # silently hides reasoning text that Hermes surfaces in its CLI. We
+    # request "summarized" so the reasoning blocks stay populated — matching
+    # 4.6 behavior and preserving the activity-feed UX during long tool runs.
     if reasoning_config and isinstance(reasoning_config, dict):
         if reasoning_config.get("enabled") is not False and "haiku" not in model.lower():
             effort = str(reasoning_config.get("effort", "medium")).lower()
             budget = THINKING_BUDGET.get(effort, 8000)
             if _supports_adaptive_thinking(model):
-                kwargs["thinking"] = {"type": "adaptive"}
+                kwargs["thinking"] = {
+                    "type": "adaptive",
+                    "display": "summarized",
+                }
+                adaptive_effort = ADAPTIVE_EFFORT_MAP.get(effort, "medium")
+                # Downgrade xhigh→max on models that don't list xhigh as a
+                # supported level (Opus/Sonnet 4.6). Opus 4.7+ keeps xhigh.
+                if adaptive_effort == "xhigh" and not _supports_xhigh_effort(model):
+                    adaptive_effort = "max"
                 kwargs["output_config"] = {
-                    "effort": ADAPTIVE_EFFORT_MAP.get(effort, "medium")
+                    "effort": adaptive_effort,
                 }
             else:
                 kwargs["thinking"] = {"type": "enabled", "budget_tokens": budget}
@@ -1333,6 +1432,15 @@ def build_anthropic_kwargs(
                 kwargs["temperature"] = 1
                 kwargs["max_tokens"] = max(effective_max_tokens, budget + 4096)
 
+    # ── Strip sampling params on 4.7+ ─────────────────────────────────
+    # Opus 4.7 rejects any non-default temperature/top_p/top_k with a 400.
+    # Callers (auxiliary_client, flush_memories, etc.) may set these for
+    # older models; drop them here as a safety net so upstream 4.6 → 4.7
+    # migrations don't require coordinated edits everywhere.
+    if _forbids_sampling_params(model):
+        for _sampling_key in ("temperature", "top_p", "top_k"):
+            kwargs.pop(_sampling_key, None)
+
     # ── Fast mode (Opus 4.6 only) ────────────────────────────────────
     # Adds extra_body.speed="fast" + the fast-mode beta header for ~2.5x
     # output speed. Only for native Anthropic endpoints — third-party
@@ -1390,12 +1498,20 @@ def normalize_anthropic_response(
                 )
             )
 
-    # Map Anthropic stop_reason to OpenAI finish_reason
+    # Map Anthropic stop_reason to OpenAI finish_reason.
+    # Newer stop reasons added in Claude 4.5+ / 4.7:
+    #   - refusal: the model declined to answer (cyber safeguards, CSAM, etc.)
+    #   - model_context_window_exceeded: hit context limit (not max_tokens)
+    # Both need distinct handling upstream — a refusal should surface to the
+    # user with a clear message, and a context-window overflow should trigger
+    # compression/truncation rather than be treated as normal end-of-turn.
     stop_reason_map = {
         "end_turn": "stop",
         "tool_use": "tool_calls",
         "max_tokens": "length",
         "stop_sequence": "stop",
+        "refusal": "content_filter",
+        "model_context_window_exceeded": "length",
     }
     finish_reason = stop_reason_map.get(response.stop_reason, "stop")
 

agent/auxiliary_client.py

@@ -58,6 +58,9 @@ _PROVIDER_ALIASES = {
     "google": "gemini",
     "google-gemini": "gemini",
     "google-ai-studio": "gemini",
+    "x-ai": "xai",
+    "x.ai": "xai",
+    "grok": "xai",
     "glm": "zai",
     "z-ai": "zai",
     "z.ai": "zai",
@@ -91,6 +94,87 @@ def _normalize_aux_provider(provider: Optional[str]) -> str:
         return "custom"
     return _PROVIDER_ALIASES.get(normalized, normalized)
 
+
+_FIXED_TEMPERATURE_MODELS: Dict[str, float] = {
+    "kimi-for-coding": 0.6,
+}
+
+# Moonshot's kimi-for-coding endpoint (api.kimi.com/coding) documents:
+# "k2.5 model will use a fixed value 1.0, non-thinking mode will use a fixed
+# value 0.6.  Any other value will result in an error."  The same lock applies
+# to the other k2.* models served on that endpoint.  Enumerated explicitly so
+# non-coding siblings like `kimi-k2-instruct` (variable temperature, served on
+# the standard chat API and third parties) are NOT clamped.
+# Source: https://platform.kimi.ai/docs/guide/kimi-k2-5-quickstart
+_KIMI_INSTANT_MODELS: frozenset = frozenset({
+    "kimi-k2.5",
+    "kimi-k2-turbo-preview",
+    "kimi-k2-0905-preview",
+})
+_KIMI_THINKING_MODELS: frozenset = frozenset({
+    "kimi-k2-thinking",
+    "kimi-k2-thinking-turbo",
+})
+
+# Moonshot's public chat endpoint (api.moonshot.ai/v1) enforces a different
+# temperature contract than the Coding Plan endpoint above.  Empirically,
+# `kimi-k2.5` on the public API rejects 0.6 with HTTP 400
+# "invalid temperature: only 1 is allowed for this model" — the Coding Plan
+# lock (0.6 for non-thinking) does not apply.  `kimi-k2-turbo-preview` and the
+# thinking variants already match the Coding Plan contract on the public
+# endpoint, so we only override the models that diverge.
+# Users hit this endpoint when `KIMI_API_KEY` is a legacy `sk-*` key (the
+# `sk-kimi-*` prefix routes to api.kimi.com/coding/v1 instead — see
+# hermes_cli/auth.py:_kimi_base_url_for_key).
+_KIMI_PUBLIC_API_OVERRIDES: Dict[str, float] = {
+    "kimi-k2.5": 1.0,
+}
+
+
+def _fixed_temperature_for_model(
+    model: Optional[str],
+    base_url: Optional[str] = None,
+) -> Optional[float]:
+    """Return a required temperature override for models with strict contracts.
+
+    Moonshot's kimi-for-coding endpoint rejects any non-approved temperature on
+    the k2.5 family.  Non-thinking variants require exactly 0.6; thinking
+    variants require 1.0.  An optional ``vendor/`` prefix (e.g.
+    ``moonshotai/kimi-k2.5``) is tolerated for aggregator routings.
+
+    When ``base_url`` points to Moonshot's public chat endpoint
+    (``api.moonshot.ai``), the contract changes for ``kimi-k2.5``: the public
+    API only accepts ``temperature=1``, not 0.6.  That override takes precedence
+    over the Coding Plan defaults above.
+
+    Returns ``None`` for every other model, including ``kimi-k2-instruct*``
+    which is the separate non-coding K2 family with variable temperature.
+    """
+    normalized = (model or "").strip().lower()
+    bare = normalized.rsplit("/", 1)[-1]
+
+    # Public Moonshot API has a stricter contract for some models than the
+    # Coding Plan endpoint — check it first so it wins on conflict.
+    if base_url and ("api.moonshot.ai" in base_url.lower() or "api.moonshot.cn" in base_url.lower()):
+        public = _KIMI_PUBLIC_API_OVERRIDES.get(bare)
+        if public is not None:
+            logger.debug(
+                "Forcing temperature=%s for %r on public Moonshot API", public, model
+            )
+            return public
+
+    fixed = _FIXED_TEMPERATURE_MODELS.get(normalized)
+    if fixed is not None:
+        logger.debug("Forcing temperature=%s for model %r (fixed map)", fixed, model)
+        return fixed
+    if bare in _KIMI_THINKING_MODELS:
+        logger.debug("Forcing temperature=1.0 for kimi thinking model %r", model)
+        return 1.0
+    if bare in _KIMI_INSTANT_MODELS:
+        logger.debug("Forcing temperature=0.6 for kimi instant model %r", model)
+        return 0.6
+    return None
+
 # Default auxiliary models for direct API-key providers (cheap/fast for side tasks)
 _API_KEY_PROVIDER_AUX_MODELS: Dict[str, str] = {
     "gemini": "gemini-3-flash-preview",
@@ -104,6 +188,7 @@ _API_KEY_PROVIDER_AUX_MODELS: Dict[str, str] = {
     "opencode-zen": "gemini-3-flash",
     "opencode-go": "glm-5",
     "kilocode": "google/gemini-3-flash-preview",
+    "ollama-cloud": "nemotron-3-nano:30b",
 }
 
 # Vision-specific model overrides for direct providers.
@@ -148,6 +233,45 @@ _CODEX_AUX_MODEL = "gpt-5.2-codex"
 _CODEX_AUX_BASE_URL = "https://chatgpt.com/backend-api/codex"
 
 
+def _codex_cloudflare_headers(access_token: str) -> Dict[str, str]:
+    """Headers required to avoid Cloudflare 403s on chatgpt.com/backend-api/codex.
+
+    The Cloudflare layer in front of the Codex endpoint whitelists a small set of
+    first-party originators (``codex_cli_rs``, ``codex_vscode``, ``codex_sdk_ts``,
+    anything starting with ``Codex``). Requests from non-residential IPs (VPS,
+    server-hosted agents) that don't advertise an allowed originator are served
+    a 403 with ``cf-mitigated: challenge`` regardless of auth correctness.
+
+    We pin ``originator: codex_cli_rs`` to match the upstream codex-rs CLI, set
+    ``User-Agent`` to a codex_cli_rs-shaped string (beats SDK fingerprinting),
+    and extract ``ChatGPT-Account-ID`` (canonical casing, from codex-rs
+    ``auth.rs``) out of the OAuth JWT's ``chatgpt_account_id`` claim.
+
+    Malformed tokens are tolerated — we drop the account-ID header rather than
+    raise, so a bad token still surfaces as an auth error (401) instead of a
+    crash at client construction.
+    """
+    headers = {
+        "User-Agent": "codex_cli_rs/0.0.0 (Hermes Agent)",
+        "originator": "codex_cli_rs",
+    }
+    if not isinstance(access_token, str) or not access_token.strip():
+        return headers
+    try:
+        import base64
+        parts = access_token.split(".")
+        if len(parts) < 2:
+            return headers
+        payload_b64 = parts[1] + "=" * (-len(parts[1]) % 4)
+        claims = json.loads(base64.urlsafe_b64decode(payload_b64))
+        acct_id = claims.get("https://api.openai.com/auth", {}).get("chatgpt_account_id")
+        if isinstance(acct_id, str) and acct_id:
+            headers["ChatGPT-Account-ID"] = acct_id
+    except Exception:
+        pass
+    return headers
+
+
 def _to_openai_base_url(base_url: str) -> str:
     """Normalize an Anthropic-style base URL to OpenAI-compatible format.
 
@@ -514,8 +638,13 @@ class _AnthropicCompletionsAdapter:
             tool_choice=normalized_tool_choice,
             is_oauth=self._is_oauth,
         )
+        # Opus 4.7+ rejects any non-default temperature/top_p/top_k; only set
+        # temperature for models that still accept it. build_anthropic_kwargs
+        # additionally strips these keys as a safety net — keep both layers.
         if temperature is not None:
-            anthropic_kwargs["temperature"] = temperature
+            from agent.anthropic_adapter import _forbids_sampling_params
+            if not _forbids_sampling_params(model):
+                anthropic_kwargs["temperature"] = temperature
 
         response = self._client.messages.create(**anthropic_kwargs)
         assistant_message, finish_reason = normalize_anthropic_response(response)
@@ -718,6 +847,11 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
             if model is None:
                 continue  # skip provider if we don't know a valid aux model
             logger.debug("Auxiliary text client: %s (%s) via pool", pconfig.name, model)
+            if provider_id == "gemini":
+                from agent.gemini_native_adapter import GeminiNativeClient, is_native_gemini_base_url
+
+                if is_native_gemini_base_url(base_url):
+                    return GeminiNativeClient(api_key=api_key, base_url=base_url), model
             extra = {}
             if "api.kimi.com" in base_url.lower():
                 extra["default_headers"] = {"User-Agent": "KimiCLI/1.30.0"}
@@ -739,6 +873,11 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
         if model is None:
             continue  # skip provider if we don't know a valid aux model
         logger.debug("Auxiliary text client: %s (%s)", pconfig.name, model)
+        if provider_id == "gemini":
+            from agent.gemini_native_adapter import GeminiNativeClient, is_native_gemini_base_url
+
+            if is_native_gemini_base_url(base_url):
+                return GeminiNativeClient(api_key=api_key, base_url=base_url), model
         extra = {}
         if "api.kimi.com" in base_url.lower():
             extra["default_headers"] = {"User-Agent": "KimiCLI/1.30.0"}
@@ -775,6 +914,21 @@ def _try_openrouter() -> Tuple[Optional[OpenAI], Optional[str]]:
 
 
 def _try_nous(vision: bool = False) -> Tuple[Optional[OpenAI], Optional[str]]:
+    # Check cross-session rate limit guard before attempting Nous —
+    # if another session already recorded a 429, skip Nous entirely
+    # to avoid piling more requests onto the tapped RPH bucket.
+    try:
+        from agent.nous_rate_guard import nous_rate_limit_remaining
+        _remaining = nous_rate_limit_remaining()
+        if _remaining is not None and _remaining > 0:
+            logger.debug(
+                "Auxiliary: skipping Nous Portal (rate-limited, resets in %.0fs)",
+                _remaining,
+            )
+            return None, None
+    except Exception:
+        pass
+
     nous = _read_nous_auth()
     if not nous:
         return None, None
@@ -899,7 +1053,52 @@ def _current_custom_base_url() -> str:
     return custom_base or ""
 
 
-def _try_custom_endpoint() -> Tuple[Optional[OpenAI], Optional[str]]:
+def _validate_proxy_env_urls() -> None:
+    """Fail fast with a clear error when proxy env vars have malformed URLs.
+
+    Common cause: shell config (e.g. .zshrc) with a typo like
+    ``export HTTP_PROXY=http://127.0.0.1:6153export NEXT_VAR=...``
+    which concatenates 'export' into the port number.  Without this
+    check the OpenAI/httpx client raises a cryptic ``Invalid port``
+    error that doesn't name the offending env var.
+    """
+    from urllib.parse import urlparse
+
+    for key in ("HTTPS_PROXY", "HTTP_PROXY", "ALL_PROXY",
+                "https_proxy", "http_proxy", "all_proxy"):
+        value = str(os.environ.get(key) or "").strip()
+        if not value:
+            continue
+        try:
+            parsed = urlparse(value)
+            if parsed.scheme:
+                _ = parsed.port          # raises ValueError for e.g. '6153export'
+        except ValueError as exc:
+            raise RuntimeError(
+                f"Malformed proxy environment variable {key}={value!r}. "
+                "Fix or unset your proxy settings and try again."
+            ) from exc
+
+
+def _validate_base_url(base_url: str) -> None:
+    """Reject obviously broken custom endpoint URLs before they reach httpx."""
+    from urllib.parse import urlparse
+
+    candidate = str(base_url or "").strip()
+    if not candidate or candidate.startswith("acp://"):
+        return
+    try:
+        parsed = urlparse(candidate)
+        if parsed.scheme in {"http", "https"}:
+            _ = parsed.port              # raises ValueError for malformed ports
+    except ValueError as exc:
+        raise RuntimeError(
+            f"Malformed custom endpoint URL: {candidate!r}. "
+            "Run `hermes setup` or `hermes model` and enter a valid http(s) base URL."
+        ) from exc
+
+
+def _try_custom_endpoint() -> Tuple[Optional[Any], Optional[str]]:
     runtime = _resolve_custom_runtime()
     if len(runtime) == 2:
         custom_base, custom_key = runtime
@@ -915,6 +1114,23 @@ def _try_custom_endpoint() -> Tuple[Optional[OpenAI], Optional[str]]:
     if custom_mode == "codex_responses":
         real_client = OpenAI(api_key=custom_key, base_url=custom_base)
         return CodexAuxiliaryClient(real_client, model), model
+    if custom_mode == "anthropic_messages":
+        # Third-party Anthropic-compatible gateway (MiniMax, Zhipu GLM,
+        # LiteLLM proxies, etc.).  Must NEVER be treated as OAuth —
+        # Anthropic OAuth claims only apply to api.anthropic.com.
+        try:
+            from agent.anthropic_adapter import build_anthropic_client
+            real_client = build_anthropic_client(custom_key, custom_base)
+        except ImportError:
+            logger.warning(
+                "Custom endpoint declares api_mode=anthropic_messages but the "
+                "anthropic SDK is not installed — falling back to OpenAI-wire."
+            )
+            return OpenAI(api_key=custom_key, base_url=custom_base), model
+        return (
+            AnthropicAuxiliaryClient(real_client, model, custom_key, custom_base, is_oauth=False),
+            model,
+        )
     return OpenAI(api_key=custom_key, base_url=custom_base), model
 
 
@@ -935,7 +1151,11 @@ def _try_codex() -> Tuple[Optional[Any], Optional[str]]:
             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=base_url)
+    real_client = OpenAI(
+        api_key=codex_token,
+        base_url=base_url,
+        default_headers=_codex_cloudflare_headers(codex_token),
+    )
     return CodexAuxiliaryClient(real_client, _CODEX_AUX_MODEL), _CODEX_AUX_MODEL
 
 
@@ -995,8 +1215,6 @@ _AUTO_PROVIDER_LABELS = {
     "_resolve_api_key_provider": "api-key",
 }
 
-_AGGREGATOR_PROVIDERS = frozenset({"openrouter", "nous"})
-
 _MAIN_RUNTIME_FIELDS = ("provider", "model", "base_url", "api_key", "api_mode")
 
 
@@ -1127,11 +1345,15 @@ def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Option
     """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).
+      1. User's main provider + main model, regardless of provider type.
+         This means auxiliary tasks (compression, vision, web extraction,
+         session search, etc.) use the same model the user configured for
+         chat.  Users on OpenRouter/Nous get their chosen chat model; users
+         on DeepSeek/ZAI/Alibaba get theirs; etc.  Running aux tasks on the
+         user's picked model keeps behavior predictable — no surprise
+         switches to a cheap fallback model for side tasks.
+      2. OpenRouter → Nous → custom → Codex → API-key providers (fallback
+         chain, only used when the main provider has no working client).
     """
     global auxiliary_is_nous, _stale_base_url_warned
     auxiliary_is_nous = False  # Reset — _try_nous() will set True if it wins
@@ -1161,11 +1383,16 @@ def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Option
             )
             _stale_base_url_warned = True
 
-    # ── Step 1: non-aggregator main provider → use main model directly ──
+    # ── Step 1: main provider + main model → use them directly ──
+    #
+    # This is the primary aux backend for every user.  "auto" means
+    # "use my main chat model for side tasks as well" — including users
+    # on aggregators (OpenRouter, Nous) who previously got routed to a
+    # cheap provider-side default.  Explicit per-task overrides set via
+    # config.yaml (auxiliary.<task>.provider) still win over this.
     main_provider = runtime_provider or _read_main_provider()
     main_model = runtime_model or _read_main_model()
     if (main_provider and main_model
-            and main_provider not in _AGGREGATOR_PROVIDERS
             and main_provider not in ("auto", "")):
         resolved_provider = main_provider
         explicit_base_url = None
@@ -1224,6 +1451,13 @@ def _to_async_client(sync_client, model: str):
         return AsyncCodexAuxiliaryClient(sync_client), model
     if isinstance(sync_client, AnthropicAuxiliaryClient):
         return AsyncAnthropicAuxiliaryClient(sync_client), model
+    try:
+        from agent.gemini_native_adapter import GeminiNativeClient, AsyncGeminiNativeClient
+
+        if isinstance(sync_client, GeminiNativeClient):
+            return AsyncGeminiNativeClient(sync_client), model
+    except ImportError:
+        pass
     try:
         from agent.copilot_acp_client import CopilotACPClient
         if isinstance(sync_client, CopilotACPClient):
@@ -1299,6 +1533,7 @@ def resolve_provider_client(
     Returns:
         (client, resolved_model) or (None, None) if auth is unavailable.
     """
+    _validate_proxy_env_urls()
     # Normalise aliases
     provider = _normalize_aux_provider(provider)
 
@@ -1387,7 +1622,11 @@ def resolve_provider_client(
                                "but no Codex OAuth token found (run: hermes model)")
                 return None, None
             final_model = _normalize_resolved_model(model or _CODEX_AUX_MODEL, provider)
-            raw_client = OpenAI(api_key=codex_token, base_url=_CODEX_AUX_BASE_URL)
+            raw_client = OpenAI(
+                api_key=codex_token,
+                base_url=_CODEX_AUX_BASE_URL,
+                default_headers=_codex_cloudflare_headers(codex_token),
+            )
             return (raw_client, final_model)
         # Standard path: wrap in CodexAuxiliaryClient adapter
         client, default = _try_codex()
@@ -1515,6 +1754,15 @@ def resolve_provider_client(
         default_model = _API_KEY_PROVIDER_AUX_MODELS.get(provider, "")
         final_model = _normalize_resolved_model(model or default_model, provider)
 
+        if provider == "gemini":
+            from agent.gemini_native_adapter import GeminiNativeClient, is_native_gemini_base_url
+
+            if is_native_gemini_base_url(base_url):
+                client = GeminiNativeClient(api_key=api_key, base_url=base_url)
+                logger.debug("resolve_provider_client: %s (%s)", provider, final_model)
+                return (_to_async_client(client, final_model) if async_mode
+                        else (client, final_model))
+
         # Provider-specific headers
         headers = {}
         if "api.kimi.com" in base_url.lower():
@@ -1523,7 +1771,6 @@ def resolve_provider_client(
             from hermes_cli.models import copilot_default_headers
 
             headers.update(copilot_default_headers())
-
         client = OpenAI(api_key=api_key, base_url=base_url,
                         **({"default_headers": headers} if headers else {}))
 
@@ -1747,34 +1994,31 @@ def resolve_vision_provider_client(
 
     if requested == "auto":
         # Vision auto-detection order:
-        #   1. Active provider + model (user's main chat config)
-        #   2. OpenRouter  (known vision-capable default model)
-        #   3. Nous Portal (known vision-capable default model)
+        #   1. User's main provider + main model (including aggregators).
+        #      _PROVIDER_VISION_MODELS provides per-provider vision model
+        #      overrides when the provider has a dedicated multimodal model
+        #      that differs from the chat model (e.g. xiaomi → mimo-v2-omni,
+        #      zai → glm-5v-turbo).
+        #   2. OpenRouter  (vision-capable aggregator fallback)
+        #   3. Nous Portal (vision-capable aggregator fallback)
         #   4. Stop
         main_provider = _read_main_provider()
         main_model = _read_main_model()
         if main_provider and main_provider not in ("auto", ""):
-            if main_provider in _VISION_AUTO_PROVIDER_ORDER:
-                # Known strict backend — use its defaults.
-                sync_client, default_model = _resolve_strict_vision_backend(main_provider)
-                if sync_client is not None:
-                    return _finalize(main_provider, sync_client, default_model)
-            else:
-                # Exotic provider (DeepSeek, Alibaba, Xiaomi, named custom, etc.)
-                # Use provider-specific vision model if available, otherwise main model.
-                vision_model = _PROVIDER_VISION_MODELS.get(main_provider, main_model)
-                rpc_client, rpc_model = resolve_provider_client(
-                    main_provider, vision_model,
-                    api_mode=resolved_api_mode)
-                if rpc_client is not None:
-                    logger.info(
-                        "Vision auto-detect: using active provider %s (%s)",
-                        main_provider, rpc_model or vision_model,
-                    )
-                    return _finalize(
-                        main_provider, rpc_client, rpc_model or vision_model)
+            vision_model = _PROVIDER_VISION_MODELS.get(main_provider, main_model)
+            rpc_client, rpc_model = resolve_provider_client(
+                main_provider, vision_model,
+                api_mode=resolved_api_mode)
+            if rpc_client is not None:
+                logger.info(
+                    "Vision auto-detect: using main provider %s (%s)",
+                    main_provider, rpc_model or vision_model,
+                )
+                return _finalize(
+                    main_provider, rpc_client, rpc_model or vision_model)
 
-        # Fall back through aggregators.
+        # Fall back through aggregators (uses their dedicated vision model,
+        # not the user's main model) when main provider has no client.
         for candidate in _VISION_AUTO_PROVIDER_ORDER:
             if candidate == main_provider:
                 continue  # already tried above
@@ -1835,9 +2079,15 @@ def auxiliary_max_tokens_param(value: int) -> dict:
 # Every auxiliary LLM consumer should use these instead of manually
 # constructing clients and calling .chat.completions.create().
 
-# Client cache: (provider, async_mode, base_url, api_key) -> (client, default_model)
+# Client cache: (provider, async_mode, base_url, api_key, api_mode, runtime_key) -> (client, default_model, loop)
+# NOTE: loop identity is NOT part of the key.  On async cache hits we check
+# whether the cached loop is the *current* loop; if not, the stale entry is
+# replaced in-place.  This bounds cache growth to one entry per unique
+# provider config rather than one per (config × event-loop), which previously
+# caused unbounded fd accumulation in long-running gateway processes (#10200).
 _client_cache: Dict[tuple, tuple] = {}
 _client_cache_lock = threading.Lock()
+_CLIENT_CACHE_MAX_SIZE = 64  # safety belt — evict oldest when exceeded
 
 
 def neuter_async_httpx_del() -> None:
@@ -1970,39 +2220,49 @@ def _get_cached_client(
     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.
+    RuntimeError.  To prevent cross-loop issues, the cache validates on
+    every async hit that the cached loop is the *current, open* loop.
+    If the loop changed (e.g. a new gateway worker-thread loop), the stale
+    entry is replaced in-place rather than creating an additional entry.
+
+    This keeps cache size bounded to one entry per unique provider config,
+    preventing the fd-exhaustion that previously occurred in long-running
+    gateways where recycled worker threads created unbounded entries (#10200).
     """
-    # 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
+    # Resolve the current event loop for async clients so we can validate
+    # cached entries.  Loop identity is NOT in the cache key — instead we
+    # check at hit time whether the cached loop is still current and open.
+    # This prevents unbounded cache growth from recycled worker-thread loops
+    # while still guaranteeing we never reuse a client on the wrong loop
+    # (which causes deadlocks, see #2681).
     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
     runtime = _normalize_main_runtime(main_runtime)
     runtime_key = tuple(runtime.get(field, "") for field in _MAIN_RUNTIME_FIELDS) if provider == "auto" else ()
-    cache_key = (provider, async_mode, base_url or "", api_key or "", api_mode or "", loop_id, runtime_key)
+    cache_key = (provider, async_mode, base_url or "", api_key or "", api_mode or "", runtime_key)
     with _client_cache_lock:
         if cache_key in _client_cache:
             cached_client, cached_default, cached_loop = _client_cache[cache_key]
             if async_mode:
-                # 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:
+                # Validate: the cached client must be bound to the CURRENT,
+                # OPEN loop.  If the loop changed or was closed, the httpx
+                # transport inside is dead — force-close and replace.
+                loop_ok = (
+                    cached_loop is not None
+                    and cached_loop is current_loop
+                    and not cached_loop.is_closed()
+                )
+                if loop_ok:
                     effective = _compat_model(cached_client, model, cached_default)
                     return cached_client, effective
+                # Stale — evict and fall through to create a new client.
+                _force_close_async_httpx(cached_client)
+                del _client_cache[cache_key]
             else:
                 effective = _compat_model(cached_client, model, cached_default)
                 return cached_client, effective
@@ -2022,6 +2282,12 @@ def _get_cached_client(
         bound_loop = current_loop
         with _client_cache_lock:
             if cache_key not in _client_cache:
+                # Safety belt: if the cache has grown beyond the max, evict
+                # the oldest entries (FIFO — dict preserves insertion order).
+                while len(_client_cache) >= _CLIENT_CACHE_MAX_SIZE:
+                    evict_key, evict_entry = next(iter(_client_cache.items()))
+                    _force_close_async_httpx(evict_entry[0])
+                    del _client_cache[evict_key]
                 _client_cache[cache_key] = (client, default_model, bound_loop)
             else:
                 client, default_model, _ = _client_cache[cache_key]
@@ -2047,7 +2313,6 @@ def _resolve_task_provider_model(
     to "custom" and the task uses that direct endpoint. api_mode is one of
     "chat_completions", "codex_responses", or None (auto-detect).
     """
-    config = {}
     cfg_provider = None
     cfg_model = None
     cfg_base_url = None
@@ -2055,16 +2320,7 @@ def _resolve_task_provider_model(
     cfg_api_mode = None
 
     if task:
-        try:
-            from hermes_cli.config import load_config
-            config = load_config()
-        except ImportError:
-            config = {}
-
-        aux = config.get("auxiliary", {}) if isinstance(config, dict) else {}
-        task_config = aux.get(task, {}) if isinstance(aux, dict) else {}
-        if not isinstance(task_config, dict):
-            task_config = {}
+        task_config = _get_auxiliary_task_config(task)
         cfg_provider = str(task_config.get("provider", "")).strip() or None
         cfg_model = str(task_config.get("model", "")).strip() or None
         cfg_base_url = str(task_config.get("base_url", "")).strip() or None
@@ -2094,17 +2350,25 @@ def _resolve_task_provider_model(
 _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*."""
+def _get_auxiliary_task_config(task: str) -> Dict[str, Any]:
+    """Return the config dict for auxiliary.<task>, or {} when unavailable."""
     if not task:
-        return default
+        return {}
     try:
         from hermes_cli.config import load_config
         config = load_config()
     except ImportError:
-        return default
+        return {}
     aux = config.get("auxiliary", {}) if isinstance(config, dict) else {}
     task_config = aux.get(task, {}) if isinstance(aux, dict) else {}
+    return task_config if isinstance(task_config, dict) else {}
+
+
+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
+    task_config = _get_auxiliary_task_config(task)
     raw = task_config.get("timeout")
     if raw is not None:
         try:
@@ -2114,6 +2378,15 @@ def _get_task_timeout(task: str, default: float = _DEFAULT_AUX_TIMEOUT) -> float
     return default
 
 
+def _get_task_extra_body(task: str) -> Dict[str, Any]:
+    """Read auxiliary.<task>.extra_body and return a shallow copy when valid."""
+    task_config = _get_auxiliary_task_config(task)
+    raw = task_config.get("extra_body")
+    if isinstance(raw, dict):
+        return dict(raw)
+    return {}
+
+
 # ---------------------------------------------------------------------------
 # Anthropic-compatible endpoint detection + image block conversion
 # ---------------------------------------------------------------------------
@@ -2201,6 +2474,19 @@ def _build_call_kwargs(
         "timeout": timeout,
     }
 
+    fixed_temperature = _fixed_temperature_for_model(model, base_url)
+    if fixed_temperature is not None:
+        temperature = fixed_temperature
+
+    # Opus 4.7+ rejects any non-default temperature/top_p/top_k — silently
+    # drop here so auxiliary callers that hardcode temperature (e.g. 0.3 on
+    # flush_memories, 0 on structured-JSON extraction) don't 400 the moment
+    # the aux model is flipped to 4.7.
+    if temperature is not None:
+        from agent.anthropic_adapter import _forbids_sampling_params
+        if _forbids_sampling_params(model):
+            temperature = None
+
     if temperature is not None:
         kwargs["temperature"] = temperature
 
@@ -2301,13 +2587,15 @@ def call_llm(
     """
     resolved_provider, resolved_model, resolved_base_url, resolved_api_key, resolved_api_mode = _resolve_task_provider_model(
         task, provider, model, base_url, api_key)
+    effective_extra_body = _get_task_extra_body(task)
+    effective_extra_body.update(extra_body or {})
 
     if task == "vision":
         effective_provider, client, final_model = resolve_vision_provider_client(
-            provider=provider,
-            model=model,
-            base_url=base_url,
-            api_key=api_key,
+            provider=resolved_provider if resolved_provider != "auto" else provider,
+            model=resolved_model or model,
+            base_url=resolved_base_url or base_url,
+            api_key=resolved_api_key or api_key,
             async_mode=False,
         )
         if client is None and resolved_provider != "auto" and not resolved_base_url:
@@ -2369,11 +2657,14 @@ def call_llm(
                      task, resolved_provider or "auto", final_model or "default",
                      f" at {_base_info}" if _base_info and "openrouter" not in _base_info else "")
 
+    # Pass the client's actual base_url (not just resolved_base_url) so
+    # endpoint-specific temperature overrides can distinguish
+    # api.moonshot.ai vs api.kimi.com/coding even on auto-detected routes.
     kwargs = _build_call_kwargs(
         resolved_provider, final_model, messages,
         temperature=temperature, max_tokens=max_tokens,
-        tools=tools, timeout=effective_timeout, extra_body=extra_body,
-        base_url=resolved_base_url)
+        tools=tools, timeout=effective_timeout, extra_body=effective_extra_body,
+        base_url=_base_info or resolved_base_url)
 
     # Convert image blocks for Anthropic-compatible endpoints (e.g. MiniMax)
     _client_base = str(getattr(client, "base_url", "") or "")
@@ -2427,7 +2718,8 @@ def call_llm(
                     fb_label, fb_model, messages,
                     temperature=temperature, max_tokens=max_tokens,
                     tools=tools, timeout=effective_timeout,
-                    extra_body=extra_body)
+                    extra_body=effective_extra_body,
+                    base_url=str(getattr(fb_client, "base_url", "") or ""))
                 return _validate_llm_response(
                     fb_client.chat.completions.create(**fb_kwargs), task)
         raise
@@ -2509,13 +2801,15 @@ async def async_call_llm(
     """
     resolved_provider, resolved_model, resolved_base_url, resolved_api_key, resolved_api_mode = _resolve_task_provider_model(
         task, provider, model, base_url, api_key)
+    effective_extra_body = _get_task_extra_body(task)
+    effective_extra_body.update(extra_body or {})
 
     if task == "vision":
         effective_provider, client, final_model = resolve_vision_provider_client(
-            provider=provider,
-            model=model,
-            base_url=base_url,
-            api_key=api_key,
+            provider=resolved_provider if resolved_provider != "auto" else provider,
+            model=resolved_model or model,
+            base_url=resolved_base_url or base_url,
+            api_key=resolved_api_key or api_key,
             async_mode=True,
         )
         if client is None and resolved_provider != "auto" and not resolved_base_url:
@@ -2562,14 +2856,17 @@ async def async_call_llm(
 
     effective_timeout = timeout if timeout is not None else _get_task_timeout(task)
 
+    # Pass the client's actual base_url (not just resolved_base_url) so
+    # endpoint-specific temperature overrides can distinguish
+    # api.moonshot.ai vs api.kimi.com/coding even on auto-detected routes.
+    _client_base = str(getattr(client, "base_url", "") or "")
     kwargs = _build_call_kwargs(
         resolved_provider, final_model, messages,
         temperature=temperature, max_tokens=max_tokens,
-        tools=tools, timeout=effective_timeout, extra_body=extra_body,
-        base_url=resolved_base_url)
+        tools=tools, timeout=effective_timeout, extra_body=effective_extra_body,
+        base_url=_client_base or resolved_base_url)
 
     # Convert image blocks for Anthropic-compatible endpoints (e.g. MiniMax)
-    _client_base = str(getattr(client, "base_url", "") or "")
     if _is_anthropic_compat_endpoint(resolved_provider, _client_base):
         kwargs["messages"] = _convert_openai_images_to_anthropic(kwargs["messages"])
 
@@ -2605,7 +2902,8 @@ async def async_call_llm(
                     fb_label, fb_model, messages,
                     temperature=temperature, max_tokens=max_tokens,
                     tools=tools, timeout=effective_timeout,
-                    extra_body=extra_body)
+                    extra_body=effective_extra_body,
+                    base_url=str(getattr(fb_client, "base_url", "") or ""))
                 # Convert sync fallback client to async
                 async_fb, async_fb_model = _to_async_client(fb_client, fb_model or "")
                 if async_fb_model and async_fb_model != fb_kwargs.get("model"):

agent/codex_responses_adapter.py

@@ -0,0 +1,650 @@
+"""Codex Responses API adapter.
+
+Pure format-conversion and normalization logic for the OpenAI Responses API
+(used by OpenAI Codex, xAI, GitHub Models, and other Responses-compatible endpoints).
+
+Extracted from run_agent.py to isolate Responses API-specific logic from the
+core agent loop. All functions are stateless — they operate on the data passed
+in and return transformed results.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import logging
+import re
+import uuid
+from types import SimpleNamespace
+from typing import Any, Dict, List, Optional
+
+from agent.prompt_builder import DEFAULT_AGENT_IDENTITY
+
+logger = logging.getLogger(__name__)
+
+def _deterministic_call_id(fn_name: str, arguments: str, index: int = 0) -> str:
+    """Generate a deterministic call_id from tool call content.
+
+    Used as a fallback when the API doesn't provide a call_id.
+    Deterministic IDs prevent cache invalidation — random UUIDs would
+    make every API call's prefix unique, breaking OpenAI's prompt cache.
+    """
+    seed = f"{fn_name}:{arguments}:{index}"
+    digest = hashlib.sha256(seed.encode("utf-8", errors="replace")).hexdigest()[:12]
+    return f"call_{digest}"
+
+
+def _split_responses_tool_id(raw_id: Any) -> tuple[Optional[str], Optional[str]]:
+    """Split a stored tool id into (call_id, response_item_id)."""
+    if not isinstance(raw_id, str):
+        return None, None
+    value = raw_id.strip()
+    if not value:
+        return None, None
+    if "|" in value:
+        call_id, response_item_id = value.split("|", 1)
+        call_id = call_id.strip() or None
+        response_item_id = response_item_id.strip() or None
+        return call_id, response_item_id
+    if value.startswith("fc_"):
+        return None, value
+    return value, None
+
+
+def _derive_responses_function_call_id(
+    call_id: str,
+    response_item_id: Optional[str] = None,
+) -> str:
+    """Build a valid Responses `function_call.id` (must start with `fc_`)."""
+    if isinstance(response_item_id, str):
+        candidate = response_item_id.strip()
+        if candidate.startswith("fc_"):
+            return candidate
+
+    source = (call_id or "").strip()
+    if source.startswith("fc_"):
+        return source
+    if source.startswith("call_") and len(source) > len("call_"):
+        return f"fc_{source[len('call_'):]}"
+
+    sanitized = re.sub(r"[^A-Za-z0-9_-]", "", source)
+    if sanitized.startswith("fc_"):
+        return sanitized
+    if sanitized.startswith("call_") and len(sanitized) > len("call_"):
+        return f"fc_{sanitized[len('call_'):]}"
+    if sanitized:
+        return f"fc_{sanitized[:48]}"
+
+    seed = source or str(response_item_id or "") or uuid.uuid4().hex
+    digest = hashlib.sha1(seed.encode("utf-8")).hexdigest()[:24]
+    return f"fc_{digest}"
+
+
+def _responses_tools(tools: Optional[List[Dict[str, Any]]] = None) -> Optional[List[Dict[str, Any]]]:
+    """Convert chat-completions tool schemas to Responses function-tool schemas."""
+    source_tools = tools
+    if not source_tools:
+        return None
+
+    converted: List[Dict[str, Any]] = []
+    for item in source_tools:
+        fn = item.get("function", {}) if isinstance(item, dict) else {}
+        name = fn.get("name")
+        if not isinstance(name, str) or not name.strip():
+            continue
+        converted.append({
+            "type": "function",
+            "name": name,
+            "description": fn.get("description", ""),
+            "strict": False,
+            "parameters": fn.get("parameters", {"type": "object", "properties": {}}),
+        })
+    return converted or None
+
+
+def _chat_messages_to_responses_input(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+    """Convert internal chat-style messages to Responses input items."""
+    items: List[Dict[str, Any]] = []
+    seen_item_ids: set = set()
+
+    for msg in messages:
+        if not isinstance(msg, dict):
+            continue
+        role = msg.get("role")
+        if role == "system":
+            continue
+
+        if role in {"user", "assistant"}:
+            content = msg.get("content", "")
+            content_text = str(content) if content is not None else ""
+
+            if role == "assistant":
+                # Replay encrypted reasoning items from previous turns
+                # so the API can maintain coherent reasoning chains.
+                codex_reasoning = msg.get("codex_reasoning_items")
+                has_codex_reasoning = False
+                if isinstance(codex_reasoning, list):
+                    for ri in codex_reasoning:
+                        if isinstance(ri, dict) and ri.get("encrypted_content"):
+                            item_id = ri.get("id")
+                            if item_id and item_id in seen_item_ids:
+                                continue
+                            # Strip the "id" field — with store=False the
+                            # Responses API cannot look up items by ID and
+                            # returns 404.  The encrypted_content blob is
+                            # self-contained for reasoning chain continuity.
+                            replay_item = {k: v for k, v in ri.items() if k != "id"}
+                            items.append(replay_item)
+                            if item_id:
+                                seen_item_ids.add(item_id)
+                            has_codex_reasoning = True
+
+                if content_text.strip():
+                    items.append({"role": "assistant", "content": content_text})
+                elif has_codex_reasoning:
+                    # The Responses API requires a following item after each
+                    # reasoning item (otherwise: missing_following_item error).
+                    # When the assistant produced only reasoning with no visible
+                    # content, emit an empty assistant message as the required
+                    # following item.
+                    items.append({"role": "assistant", "content": ""})
+
+                tool_calls = msg.get("tool_calls")
+                if isinstance(tool_calls, list):
+                    for tc in tool_calls:
+                        if not isinstance(tc, dict):
+                            continue
+                        fn = tc.get("function", {})
+                        fn_name = fn.get("name")
+                        if not isinstance(fn_name, str) or not fn_name.strip():
+                            continue
+
+                        embedded_call_id, embedded_response_item_id = _split_responses_tool_id(
+                            tc.get("id")
+                        )
+                        call_id = tc.get("call_id")
+                        if not isinstance(call_id, str) or not call_id.strip():
+                            call_id = embedded_call_id
+                        if not isinstance(call_id, str) or not call_id.strip():
+                            if (
+                                isinstance(embedded_response_item_id, str)
+                                and embedded_response_item_id.startswith("fc_")
+                                and len(embedded_response_item_id) > len("fc_")
+                            ):
+                                call_id = f"call_{embedded_response_item_id[len('fc_'):]}"
+                            else:
+                                _raw_args = str(fn.get("arguments", "{}"))
+                                call_id = _deterministic_call_id(fn_name, _raw_args, len(items))
+                        call_id = call_id.strip()
+
+                        arguments = fn.get("arguments", "{}")
+                        if isinstance(arguments, dict):
+                            arguments = json.dumps(arguments, ensure_ascii=False)
+                        elif not isinstance(arguments, str):
+                            arguments = str(arguments)
+                        arguments = arguments.strip() or "{}"
+
+                        items.append({
+                            "type": "function_call",
+                            "call_id": call_id,
+                            "name": fn_name,
+                            "arguments": arguments,
+                        })
+                continue
+
+            items.append({"role": role, "content": content_text})
+            continue
+
+        if role == "tool":
+            raw_tool_call_id = msg.get("tool_call_id")
+            call_id, _ = _split_responses_tool_id(raw_tool_call_id)
+            if not isinstance(call_id, str) or not call_id.strip():
+                if isinstance(raw_tool_call_id, str) and raw_tool_call_id.strip():
+                    call_id = raw_tool_call_id.strip()
+            if not isinstance(call_id, str) or not call_id.strip():
+                continue
+            items.append({
+                "type": "function_call_output",
+                "call_id": call_id,
+                "output": str(msg.get("content", "") or ""),
+            })
+
+    return items
+
+
+def _preflight_codex_input_items(raw_items: Any) -> List[Dict[str, Any]]:
+    if not isinstance(raw_items, list):
+        raise ValueError("Codex Responses input must be a list of input items.")
+
+    normalized: List[Dict[str, Any]] = []
+    seen_ids: set = set()
+    for idx, item in enumerate(raw_items):
+        if not isinstance(item, dict):
+            raise ValueError(f"Codex Responses input[{idx}] must be an object.")
+
+        item_type = item.get("type")
+        if item_type == "function_call":
+            call_id = item.get("call_id")
+            name = item.get("name")
+            if not isinstance(call_id, str) or not call_id.strip():
+                raise ValueError(f"Codex Responses input[{idx}] function_call is missing call_id.")
+            if not isinstance(name, str) or not name.strip():
+                raise ValueError(f"Codex Responses input[{idx}] function_call is missing name.")
+
+            arguments = item.get("arguments", "{}")
+            if isinstance(arguments, dict):
+                arguments = json.dumps(arguments, ensure_ascii=False)
+            elif not isinstance(arguments, str):
+                arguments = str(arguments)
+            arguments = arguments.strip() or "{}"
+
+            normalized.append(
+                {
+                    "type": "function_call",
+                    "call_id": call_id.strip(),
+                    "name": name.strip(),
+                    "arguments": arguments,
+                }
+            )
+            continue
+
+        if item_type == "function_call_output":
+            call_id = item.get("call_id")
+            if not isinstance(call_id, str) or not call_id.strip():
+                raise ValueError(f"Codex Responses input[{idx}] function_call_output is missing call_id.")
+            output = item.get("output", "")
+            if output is None:
+                output = ""
+            if not isinstance(output, str):
+                output = str(output)
+
+            normalized.append(
+                {
+                    "type": "function_call_output",
+                    "call_id": call_id.strip(),
+                    "output": output,
+                }
+            )
+            continue
+
+        if item_type == "reasoning":
+            encrypted = item.get("encrypted_content")
+            if isinstance(encrypted, str) and encrypted:
+                item_id = item.get("id")
+                if isinstance(item_id, str) and item_id:
+                    if item_id in seen_ids:
+                        continue
+                    seen_ids.add(item_id)
+                reasoning_item = {"type": "reasoning", "encrypted_content": encrypted}
+                # Do NOT include the "id" in the outgoing item — with
+                # store=False (our default) the API tries to resolve the
+                # id server-side and returns 404.  The id is still used
+                # above for local deduplication via seen_ids.
+                summary = item.get("summary")
+                if isinstance(summary, list):
+                    reasoning_item["summary"] = summary
+                else:
+                    reasoning_item["summary"] = []
+                normalized.append(reasoning_item)
+            continue
+
+        role = item.get("role")
+        if role in {"user", "assistant"}:
+            content = item.get("content", "")
+            if content is None:
+                content = ""
+            if not isinstance(content, str):
+                content = str(content)
+
+            normalized.append({"role": role, "content": content})
+            continue
+
+        raise ValueError(
+            f"Codex Responses input[{idx}] has unsupported item shape (type={item_type!r}, role={role!r})."
+        )
+
+    return normalized
+
+
+def _preflight_codex_api_kwargs(
+    api_kwargs: Any,
+    *,
+    allow_stream: bool = False,
+) -> Dict[str, Any]:
+    if not isinstance(api_kwargs, dict):
+        raise ValueError("Codex Responses request must be a dict.")
+
+    required = {"model", "instructions", "input"}
+    missing = [key for key in required if key not in api_kwargs]
+    if missing:
+        raise ValueError(f"Codex Responses request missing required field(s): {', '.join(sorted(missing))}.")
+
+    model = api_kwargs.get("model")
+    if not isinstance(model, str) or not model.strip():
+        raise ValueError("Codex Responses request 'model' must be a non-empty string.")
+    model = model.strip()
+
+    instructions = api_kwargs.get("instructions")
+    if instructions is None:
+        instructions = ""
+    if not isinstance(instructions, str):
+        instructions = str(instructions)
+    instructions = instructions.strip() or DEFAULT_AGENT_IDENTITY
+
+    normalized_input = _preflight_codex_input_items(api_kwargs.get("input"))
+
+    tools = api_kwargs.get("tools")
+    normalized_tools = None
+    if tools is not None:
+        if not isinstance(tools, list):
+            raise ValueError("Codex Responses request 'tools' must be a list when provided.")
+        normalized_tools = []
+        for idx, tool in enumerate(tools):
+            if not isinstance(tool, dict):
+                raise ValueError(f"Codex Responses tools[{idx}] must be an object.")
+            if tool.get("type") != "function":
+                raise ValueError(f"Codex Responses tools[{idx}] has unsupported type {tool.get('type')!r}.")
+
+            name = tool.get("name")
+            parameters = tool.get("parameters")
+            if not isinstance(name, str) or not name.strip():
+                raise ValueError(f"Codex Responses tools[{idx}] is missing a valid name.")
+            if not isinstance(parameters, dict):
+                raise ValueError(f"Codex Responses tools[{idx}] is missing valid parameters.")
+
+            description = tool.get("description", "")
+            if description is None:
+                description = ""
+            if not isinstance(description, str):
+                description = str(description)
+
+            strict = tool.get("strict", False)
+            if not isinstance(strict, bool):
+                strict = bool(strict)
+
+            normalized_tools.append(
+                {
+                    "type": "function",
+                    "name": name.strip(),
+                    "description": description,
+                    "strict": strict,
+                    "parameters": parameters,
+                }
+            )
+
+    store = api_kwargs.get("store", False)
+    if store is not False:
+        raise ValueError("Codex Responses contract requires 'store' to be false.")
+
+    allowed_keys = {
+        "model", "instructions", "input", "tools", "store",
+        "reasoning", "include", "max_output_tokens", "temperature",
+        "tool_choice", "parallel_tool_calls", "prompt_cache_key", "service_tier",
+        "extra_headers",
+    }
+    normalized: Dict[str, Any] = {
+        "model": model,
+        "instructions": instructions,
+        "input": normalized_input,
+        "store": False,
+    }
+    if normalized_tools is not None:
+        normalized["tools"] = normalized_tools
+
+    # Pass through reasoning config
+    reasoning = api_kwargs.get("reasoning")
+    if isinstance(reasoning, dict):
+        normalized["reasoning"] = reasoning
+    include = api_kwargs.get("include")
+    if isinstance(include, list):
+        normalized["include"] = include
+    service_tier = api_kwargs.get("service_tier")
+    if isinstance(service_tier, str) and service_tier.strip():
+        normalized["service_tier"] = service_tier.strip()
+
+    # Pass through max_output_tokens and temperature
+    max_output_tokens = api_kwargs.get("max_output_tokens")
+    if isinstance(max_output_tokens, (int, float)) and max_output_tokens > 0:
+        normalized["max_output_tokens"] = int(max_output_tokens)
+    temperature = api_kwargs.get("temperature")
+    if isinstance(temperature, (int, float)):
+        normalized["temperature"] = float(temperature)
+
+    # Pass through tool_choice, parallel_tool_calls, prompt_cache_key
+    for passthrough_key in ("tool_choice", "parallel_tool_calls", "prompt_cache_key"):
+        val = api_kwargs.get(passthrough_key)
+        if val is not None:
+            normalized[passthrough_key] = val
+
+    extra_headers = api_kwargs.get("extra_headers")
+    if extra_headers is not None:
+        if not isinstance(extra_headers, dict):
+            raise ValueError("Codex Responses request 'extra_headers' must be an object.")
+        normalized_headers: Dict[str, str] = {}
+        for key, value in extra_headers.items():
+            if not isinstance(key, str) or not key.strip():
+                raise ValueError("Codex Responses request 'extra_headers' keys must be non-empty strings.")
+            if value is None:
+                continue
+            normalized_headers[key.strip()] = str(value)
+        if normalized_headers:
+            normalized["extra_headers"] = normalized_headers
+
+    if allow_stream:
+        stream = api_kwargs.get("stream")
+        if stream is not None and stream is not True:
+            raise ValueError("Codex Responses 'stream' must be true when set.")
+        if stream is True:
+            normalized["stream"] = True
+        allowed_keys.add("stream")
+    elif "stream" in api_kwargs:
+        raise ValueError("Codex Responses stream flag is only allowed in fallback streaming requests.")
+
+    unexpected = sorted(key for key in api_kwargs if key not in allowed_keys)
+    if unexpected:
+        raise ValueError(
+            f"Codex Responses request has unsupported field(s): {', '.join(unexpected)}."
+        )
+
+    return normalized
+
+
+def _extract_responses_message_text(item: Any) -> str:
+    """Extract assistant text from a Responses message output item."""
+    content = getattr(item, "content", None)
+    if not isinstance(content, list):
+        return ""
+
+    chunks: List[str] = []
+    for part in content:
+        ptype = getattr(part, "type", None)
+        if ptype not in {"output_text", "text"}:
+            continue
+        text = getattr(part, "text", None)
+        if isinstance(text, str) and text:
+            chunks.append(text)
+    return "".join(chunks).strip()
+
+
+def _extract_responses_reasoning_text(item: Any) -> str:
+    """Extract a compact reasoning text from a Responses reasoning item."""
+    summary = getattr(item, "summary", None)
+    if isinstance(summary, list):
+        chunks: List[str] = []
+        for part in summary:
+            text = getattr(part, "text", None)
+            if isinstance(text, str) and text:
+                chunks.append(text)
+        if chunks:
+            return "\n".join(chunks).strip()
+    text = getattr(item, "text", None)
+    if isinstance(text, str) and text:
+        return text.strip()
+    return ""
+
+
+def _normalize_codex_response(response: Any) -> tuple[Any, str]:
+    """Normalize a Responses API object to an assistant_message-like object."""
+    output = getattr(response, "output", None)
+    if not isinstance(output, list) or not output:
+        # The Codex backend can return empty output when the answer was
+        # delivered entirely via stream events. Check output_text as a
+        # last-resort fallback before raising.
+        out_text = getattr(response, "output_text", None)
+        if isinstance(out_text, str) and out_text.strip():
+            logger.debug(
+                "Codex response has empty output but output_text is present (%d chars); "
+                "synthesizing output item.", len(out_text.strip()),
+            )
+            output = [SimpleNamespace(
+                type="message", role="assistant", status="completed",
+                content=[SimpleNamespace(type="output_text", text=out_text.strip())],
+            )]
+            response.output = output
+        else:
+            raise RuntimeError("Responses API returned no output items")
+
+    response_status = getattr(response, "status", None)
+    if isinstance(response_status, str):
+        response_status = response_status.strip().lower()
+    else:
+        response_status = None
+
+    if response_status in {"failed", "cancelled"}:
+        error_obj = getattr(response, "error", None)
+        if isinstance(error_obj, dict):
+            error_msg = error_obj.get("message") or str(error_obj)
+        else:
+            error_msg = str(error_obj) if error_obj else f"Responses API returned status '{response_status}'"
+        raise RuntimeError(error_msg)
+
+    content_parts: List[str] = []
+    reasoning_parts: List[str] = []
+    reasoning_items_raw: List[Dict[str, Any]] = []
+    tool_calls: List[Any] = []
+    has_incomplete_items = response_status in {"queued", "in_progress", "incomplete"}
+    saw_commentary_phase = False
+    saw_final_answer_phase = False
+
+    for item in output:
+        item_type = getattr(item, "type", None)
+        item_status = getattr(item, "status", None)
+        if isinstance(item_status, str):
+            item_status = item_status.strip().lower()
+        else:
+            item_status = None
+
+        if item_status in {"queued", "in_progress", "incomplete"}:
+            has_incomplete_items = True
+
+        if item_type == "message":
+            item_phase = getattr(item, "phase", None)
+            if isinstance(item_phase, str):
+                normalized_phase = item_phase.strip().lower()
+                if normalized_phase in {"commentary", "analysis"}:
+                    saw_commentary_phase = True
+                elif normalized_phase in {"final_answer", "final"}:
+                    saw_final_answer_phase = True
+            message_text = _extract_responses_message_text(item)
+            if message_text:
+                content_parts.append(message_text)
+        elif item_type == "reasoning":
+            reasoning_text = _extract_responses_reasoning_text(item)
+            if reasoning_text:
+                reasoning_parts.append(reasoning_text)
+            # Capture the full reasoning item for multi-turn continuity.
+            # encrypted_content is an opaque blob the API needs back on
+            # subsequent turns to maintain coherent reasoning chains.
+            encrypted = getattr(item, "encrypted_content", None)
+            if isinstance(encrypted, str) and encrypted:
+                raw_item = {"type": "reasoning", "encrypted_content": encrypted}
+                item_id = getattr(item, "id", None)
+                if isinstance(item_id, str) and item_id:
+                    raw_item["id"] = item_id
+                # Capture summary — required by the API when replaying reasoning items
+                summary = getattr(item, "summary", None)
+                if isinstance(summary, list):
+                    raw_summary = []
+                    for part in summary:
+                        text = getattr(part, "text", None)
+                        if isinstance(text, str):
+                            raw_summary.append({"type": "summary_text", "text": text})
+                    raw_item["summary"] = raw_summary
+                reasoning_items_raw.append(raw_item)
+        elif item_type == "function_call":
+            if item_status in {"queued", "in_progress", "incomplete"}:
+                continue
+            fn_name = getattr(item, "name", "") or ""
+            arguments = getattr(item, "arguments", "{}")
+            if not isinstance(arguments, str):
+                arguments = json.dumps(arguments, ensure_ascii=False)
+            raw_call_id = getattr(item, "call_id", None)
+            raw_item_id = getattr(item, "id", None)
+            embedded_call_id, _ = _split_responses_tool_id(raw_item_id)
+            call_id = raw_call_id if isinstance(raw_call_id, str) and raw_call_id.strip() else embedded_call_id
+            if not isinstance(call_id, str) or not call_id.strip():
+                call_id = _deterministic_call_id(fn_name, arguments, len(tool_calls))
+            call_id = call_id.strip()
+            response_item_id = raw_item_id if isinstance(raw_item_id, str) else None
+            response_item_id = _derive_responses_function_call_id(call_id, response_item_id)
+            tool_calls.append(SimpleNamespace(
+                id=call_id,
+                call_id=call_id,
+                response_item_id=response_item_id,
+                type="function",
+                function=SimpleNamespace(name=fn_name, arguments=arguments),
+            ))
+        elif item_type == "custom_tool_call":
+            fn_name = getattr(item, "name", "") or ""
+            arguments = getattr(item, "input", "{}")
+            if not isinstance(arguments, str):
+                arguments = json.dumps(arguments, ensure_ascii=False)
+            raw_call_id = getattr(item, "call_id", None)
+            raw_item_id = getattr(item, "id", None)
+            embedded_call_id, _ = _split_responses_tool_id(raw_item_id)
+            call_id = raw_call_id if isinstance(raw_call_id, str) and raw_call_id.strip() else embedded_call_id
+            if not isinstance(call_id, str) or not call_id.strip():
+                call_id = _deterministic_call_id(fn_name, arguments, len(tool_calls))
+            call_id = call_id.strip()
+            response_item_id = raw_item_id if isinstance(raw_item_id, str) else None
+            response_item_id = _derive_responses_function_call_id(call_id, response_item_id)
+            tool_calls.append(SimpleNamespace(
+                id=call_id,
+                call_id=call_id,
+                response_item_id=response_item_id,
+                type="function",
+                function=SimpleNamespace(name=fn_name, arguments=arguments),
+            ))
+
+    final_text = "\n".join([p for p in content_parts if p]).strip()
+    if not final_text and hasattr(response, "output_text"):
+        out_text = getattr(response, "output_text", "")
+        if isinstance(out_text, str):
+            final_text = out_text.strip()
+
+    assistant_message = SimpleNamespace(
+        content=final_text,
+        tool_calls=tool_calls,
+        reasoning="\n\n".join(reasoning_parts).strip() if reasoning_parts else None,
+        reasoning_content=None,
+        reasoning_details=None,
+        codex_reasoning_items=reasoning_items_raw or None,
+    )
+
+    if tool_calls:
+        finish_reason = "tool_calls"
+    elif has_incomplete_items or (saw_commentary_phase and not saw_final_answer_phase):
+        finish_reason = "incomplete"
+    elif reasoning_items_raw and not final_text:
+        # Response contains only reasoning (encrypted thinking state) with
+        # no visible content or tool calls.  The model is still thinking and
+        # needs another turn to produce the actual answer.  Marking this as
+        # "stop" would send it into the empty-content retry loop which burns
+        # 3 retries then fails — treat it as incomplete instead so the Codex
+        # continuation path handles it correctly.
+        finish_reason = "incomplete"
+    else:
+        finish_reason = "stop"
+    return assistant_message, finish_reason
+
+

agent/context_compressor.py

@@ -39,7 +39,10 @@ SUMMARY_PREFIX = (
     "into the summary below. This is a handoff from a previous context "
     "window — treat it as background reference, NOT as active instructions. "
     "Do NOT answer questions or fulfill requests mentioned in this summary; "
-    "they were already addressed. Respond ONLY to the latest user message "
+    "they were already addressed. "
+    "Your current task is identified in the '## Active Task' section of the "
+    "summary — resume exactly from there. "
+    "Respond ONLY to the latest user message "
     "that appears AFTER this summary. The current session state (files, "
     "config, etc.) may reflect work described here — avoid repeating it:"
 )
@@ -60,6 +63,52 @@ _CHARS_PER_TOKEN = 4
 _SUMMARY_FAILURE_COOLDOWN_SECONDS = 600
 
 
+def _truncate_tool_call_args_json(args: str, head_chars: int = 200) -> str:
+    """Shrink long string values inside a tool-call arguments JSON blob while
+    preserving JSON validity.
+
+    The ``function.arguments`` field on a tool call is a JSON-encoded string
+    passed through to the LLM provider; downstream providers strictly
+    validate it and return a non-retryable 400 when it is not well-formed.
+    An earlier implementation sliced the raw JSON at a fixed byte offset and
+    appended ``...[truncated]`` — which routinely produced strings like::
+
+        {"path": "/foo/bar", "content": "# long markdown
+        ...[truncated]
+
+    i.e. an unterminated string and a missing closing brace. MiniMax, for
+    example, rejects this with ``invalid function arguments json string``
+    and the session gets stuck re-sending the same broken history on every
+    turn. See issue #11762 for the observed loop.
+
+    This helper parses the arguments, shrinks long string leaves inside the
+    parsed structure, and re-serialises. Non-string values (paths, ints,
+    booleans) are preserved intact. If the arguments are not valid JSON
+    to begin with — some model backends use non-JSON tool arguments — the
+    original string is returned unchanged rather than replaced with
+    something neither we nor the backend can parse.
+    """
+    try:
+        parsed = json.loads(args)
+    except (ValueError, TypeError):
+        return args
+
+    def _shrink(obj: Any) -> Any:
+        if isinstance(obj, str):
+            if len(obj) > head_chars:
+                return obj[:head_chars] + "...[truncated]"
+            return obj
+        if isinstance(obj, dict):
+            return {k: _shrink(v) for k, v in obj.items()}
+        if isinstance(obj, list):
+            return [_shrink(v) for v in obj]
+        return obj
+
+    shrunken = _shrink(parsed)
+    # ensure_ascii=False preserves CJK/emoji instead of bloating with \uXXXX
+    return json.dumps(shrunken, ensure_ascii=False)
+
+
 def _summarize_tool_result(tool_name: str, tool_args: str, tool_content: str) -> str:
     """Create an informative 1-line summary of a tool call + result.
 
@@ -446,6 +495,11 @@ class ContextCompressor(ContextEngine):
         # Pass 3: Truncate large tool_call arguments in assistant messages
         # outside the protected tail. write_file with 50KB content, for
         # example, survives pruning entirely without this.
+        #
+        # The shrinking is done inside the parsed JSON structure so the
+        # result remains valid JSON — otherwise downstream providers 400
+        # on every subsequent turn until the broken call falls out of
+        # the window. See ``_truncate_tool_call_args_json`` docstring.
         for i in range(prune_boundary):
             msg = result[i]
             if msg.get("role") != "assistant" or not msg.get("tool_calls"):
@@ -456,8 +510,10 @@ class ContextCompressor(ContextEngine):
                 if isinstance(tc, dict):
                     args = tc.get("function", {}).get("arguments", "")
                     if len(args) > 500:
-                        tc = {**tc, "function": {**tc["function"], "arguments": args[:200] + "...[truncated]"}}
-                        modified = True
+                        new_args = _truncate_tool_call_args_json(args)
+                        if new_args != args:
+                            tc = {**tc, "function": {**tc["function"], "arguments": new_args}}
+                            modified = True
                 new_tcs.append(tc)
             if modified:
                 result[i] = {**msg, "tool_calls": new_tcs}
@@ -577,12 +633,22 @@ class ContextCompressor(ContextEngine):
             "assistant that continues the conversation. "
             "Do NOT respond to any questions or requests in the conversation — "
             "only output the structured summary. "
-            "Do NOT include any preamble, greeting, or prefix."
+            "Do NOT include any preamble, greeting, or prefix. "
+            "Write the summary in the same language the user was using in the "
+            "conversation — do not translate or switch to English."
         )
 
         # Shared structured template (used by both paths).
-        _template_sections = f"""## Goal
-[What the user is trying to accomplish]
+        _template_sections = f"""## Active Task
+[THE SINGLE MOST IMPORTANT FIELD. Copy the user's most recent request or
+task assignment verbatim — the exact words they used. If multiple tasks
+were requested and only some are done, list only the ones NOT yet completed.
+The next assistant must pick up exactly here. Example:
+"User asked: 'Now refactor the auth module to use JWT instead of sessions'"
+If no outstanding task exists, write "None."]
+
+## Goal
+[What the user is trying to accomplish overall]
 
 ## Constraints & Preferences
 [User preferences, coding style, constraints, important decisions]
@@ -644,7 +710,7 @@ PREVIOUS SUMMARY:
 NEW TURNS TO INCORPORATE:
 {content_to_summarize}
 
-Update the summary using this exact structure. PRESERVE all existing information that is still relevant. ADD new completed actions to the numbered list (continue numbering). Move items from "In Progress" to "Completed Actions" when done. Move answered questions to "Resolved Questions". Update "Active State" to reflect current state. Remove information only if it is clearly obsolete.
+Update the summary using this exact structure. PRESERVE all existing information that is still relevant. ADD new completed actions to the numbered list (continue numbering). Move items from "In Progress" to "Completed Actions" when done. Move answered questions to "Resolved Questions". Update "Active State" to reflect current state. Remove information only if it is clearly obsolete. CRITICAL: Update "## Active Task" to reflect the user's most recent unfulfilled request — this is the most important field for task continuity.
 
 {_template_sections}"""
         else:
@@ -862,6 +928,62 @@ The user has requested that this compaction PRIORITISE preserving all informatio
     # Tail protection by token budget
     # ------------------------------------------------------------------
 
+    def _find_last_user_message_idx(
+        self, messages: List[Dict[str, Any]], head_end: int
+    ) -> int:
+        """Return the index of the last user-role message at or after *head_end*, or -1."""
+        for i in range(len(messages) - 1, head_end - 1, -1):
+            if messages[i].get("role") == "user":
+                return i
+        return -1
+
+    def _ensure_last_user_message_in_tail(
+        self,
+        messages: List[Dict[str, Any]],
+        cut_idx: int,
+        head_end: int,
+    ) -> int:
+        """Guarantee the most recent user message is in the protected tail.
+
+        Context compressor bug (#10896): ``_align_boundary_backward`` can pull
+        ``cut_idx`` past a user message when it tries to keep tool_call/result
+        groups together.  If the last user message ends up in the *compressed*
+        middle region the LLM summariser writes it into "Pending User Asks",
+        but ``SUMMARY_PREFIX`` tells the next model to respond only to user
+        messages *after* the summary — so the task effectively disappears from
+        the active context, causing the agent to stall, repeat completed work,
+        or silently drop the user's latest request.
+
+        Fix: if the last user-role message is not already in the tail
+        (``messages[cut_idx:]``), walk ``cut_idx`` back to include it.  We
+        then re-align backward one more time to avoid splitting any
+        tool_call/result group that immediately precedes the user message.
+        """
+        last_user_idx = self._find_last_user_message_idx(messages, head_end)
+        if last_user_idx < 0:
+            # No user message found beyond head — nothing to anchor.
+            return cut_idx
+
+        if last_user_idx >= cut_idx:
+            # Already in the tail; nothing to do.
+            return cut_idx
+
+        # The last user message is in the middle (compressed) region.
+        # Pull cut_idx back to it directly — a user message is already a
+        # clean boundary (no tool_call/result splitting risk), so there is no
+        # need to call _align_boundary_backward here; doing so would
+        # unnecessarily pull the cut further back into the preceding
+        # assistant + tool_calls group.
+        if not self.quiet_mode:
+            logger.debug(
+                "Anchoring tail cut to last user message at index %d "
+                "(was %d) to prevent active-task loss after compression",
+                last_user_idx,
+                cut_idx,
+            )
+        # Safety: never go back into the head region.
+        return max(last_user_idx, head_end + 1)
+
     def _find_tail_cut_by_tokens(
         self, messages: List[Dict[str, Any]], head_end: int,
         token_budget: int | None = None,
@@ -879,7 +1001,8 @@ The user has requested that this compaction PRIORITISE preserving all informatio
         read, etc.).  If even the minimum 3 messages exceed 1.5x the budget
         the cut is placed right after the head so compression still runs.
 
-        Never cuts inside a tool_call/result group.
+        Never cuts inside a tool_call/result group.  Always ensures the most
+        recent user message is in the tail (see ``_ensure_last_user_message_in_tail``).
         """
         if token_budget is None:
             token_budget = self.tail_token_budget
@@ -918,6 +1041,10 @@ The user has requested that this compaction PRIORITISE preserving all informatio
         # Align to avoid splitting tool groups
         cut_idx = self._align_boundary_backward(messages, cut_idx)
 
+        # Ensure the most recent user message is always in the tail so the
+        # active task is never lost to compression (fixes #10896).
+        cut_idx = self._ensure_last_user_message_in_tail(messages, cut_idx, head_end)
+
         return max(cut_idx, head_end + 1)
 
     # ------------------------------------------------------------------

agent/context_references.py

@@ -483,9 +483,7 @@ def _rg_files(path: Path, cwd: Path, limit: int) -> list[Path] | None:
             text=True,
             timeout=10,
         )
-    except FileNotFoundError:
-        return None
-    except subprocess.TimeoutExpired:
+    except (FileNotFoundError, OSError, subprocess.TimeoutExpired):
         return None
     if result.returncode != 0:
         return None

agent/copilot_acp_client.py

@@ -313,9 +313,25 @@ class CopilotACPClient:
             tools=tools,
             tool_choice=tool_choice,
         )
+        # Normalise timeout: run_agent.py may pass an httpx.Timeout object
+        # (used natively by the OpenAI SDK) rather than a plain float.
+        if timeout is None:
+            _effective_timeout = _DEFAULT_TIMEOUT_SECONDS
+        elif isinstance(timeout, (int, float)):
+            _effective_timeout = float(timeout)
+        else:
+            # httpx.Timeout or similar — pick the largest component so the
+            # subprocess has enough wall-clock time for the full response.
+            _candidates = [
+                getattr(timeout, attr, None)
+                for attr in ("read", "write", "connect", "pool", "timeout")
+            ]
+            _numeric = [float(v) for v in _candidates if isinstance(v, (int, float))]
+            _effective_timeout = max(_numeric) if _numeric else _DEFAULT_TIMEOUT_SECONDS
+
         response_text, reasoning_text = self._run_prompt(
             prompt_text,
-            timeout_seconds=float(timeout or _DEFAULT_TIMEOUT_SECONDS),
+            timeout_seconds=_effective_timeout,
         )
 
         tool_calls, cleaned_text = _extract_tool_calls_from_text(response_text)

agent/credential_pool.py

@@ -22,8 +22,6 @@ from hermes_cli.auth import (
     _auth_store_lock,
     _codex_access_token_is_expiring,
     _decode_jwt_claims,
-    _import_codex_cli_tokens,
-    _write_codex_cli_tokens,
     _load_auth_store,
     _load_provider_state,
     _resolve_kimi_base_url,
@@ -457,39 +455,6 @@ class CredentialPool:
             logger.debug("Failed to sync from credentials file: %s", exc)
         return entry
 
-    def _sync_codex_entry_from_cli(self, entry: PooledCredential) -> PooledCredential:
-        """Sync an openai-codex pool entry from ~/.codex/auth.json if tokens differ.
-
-        OpenAI OAuth refresh tokens are single-use and rotate on every refresh.
-        When the Codex CLI (or another Hermes profile) refreshes its token,
-        the pool entry's refresh_token becomes stale.  This method detects that
-        by comparing against ~/.codex/auth.json and syncing the fresh pair.
-        """
-        if self.provider != "openai-codex":
-            return entry
-        try:
-            cli_tokens = _import_codex_cli_tokens()
-            if not cli_tokens:
-                return entry
-            cli_refresh = cli_tokens.get("refresh_token", "")
-            cli_access = cli_tokens.get("access_token", "")
-            if cli_refresh and cli_refresh != entry.refresh_token:
-                logger.debug("Pool entry %s: syncing tokens from ~/.codex/auth.json (refresh token changed)", entry.id)
-                updated = replace(
-                    entry,
-                    access_token=cli_access,
-                    refresh_token=cli_refresh,
-                    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 ~/.codex/auth.json: %s", exc)
-        return entry
-
     def _sync_device_code_entry_to_auth_store(self, entry: PooledCredential) -> None:
         """Write refreshed pool entry tokens back to auth.json providers.
 
@@ -585,13 +550,6 @@ class CredentialPool:
                     except Exception as wexc:
                         logger.debug("Failed to write refreshed token to credentials file: %s", wexc)
             elif self.provider == "openai-codex":
-                # Proactively sync from ~/.codex/auth.json before refresh.
-                # The Codex CLI (or another Hermes profile) may have already
-                # consumed our refresh_token.  Syncing first avoids a
-                # "refresh_token_reused" error when the CLI has a newer pair.
-                synced = self._sync_codex_entry_from_cli(entry)
-                if synced is not entry:
-                    entry = synced
                 refreshed = auth_mod.refresh_codex_oauth_pure(
                     entry.access_token,
                     entry.refresh_token,
@@ -677,45 +635,6 @@ class CredentialPool:
                     # Credentials file had a valid (non-expired) token — use it directly
                     logger.debug("Credentials file has valid token, using without refresh")
                     return synced
-            # For openai-codex: the refresh_token may have been consumed by
-            # the Codex CLI between our proactive sync and the refresh call.
-            # Re-sync and retry once.
-            if self.provider == "openai-codex":
-                synced = self._sync_codex_entry_from_cli(entry)
-                if synced.refresh_token != entry.refresh_token:
-                    logger.debug("Retrying Codex refresh with synced token from ~/.codex/auth.json")
-                    try:
-                        refreshed = auth_mod.refresh_codex_oauth_pure(
-                            synced.access_token,
-                            synced.refresh_token,
-                        )
-                        updated = replace(
-                            synced,
-                            access_token=refreshed["access_token"],
-                            refresh_token=refreshed["refresh_token"],
-                            last_refresh=refreshed.get("last_refresh"),
-                            last_status=STATUS_OK,
-                            last_status_at=None,
-                            last_error_code=None,
-                        )
-                        self._replace_entry(synced, updated)
-                        self._persist()
-                        self._sync_device_code_entry_to_auth_store(updated)
-                        try:
-                            _write_codex_cli_tokens(
-                                updated.access_token,
-                                updated.refresh_token,
-                                last_refresh=updated.last_refresh,
-                            )
-                        except Exception as wexc:
-                            logger.debug("Failed to write refreshed Codex tokens to CLI file (retry): %s", wexc)
-                        return updated
-                    except Exception as retry_exc:
-                        logger.debug("Codex retry refresh also failed: %s", retry_exc)
-                elif not self._entry_needs_refresh(synced):
-                    logger.debug("Codex CLI has valid token, using without refresh")
-                    self._sync_device_code_entry_to_auth_store(synced)
-                    return synced
             self._mark_exhausted(entry, None)
             return None
 
@@ -734,17 +653,6 @@ class CredentialPool:
         # _seed_from_singletons() on the next load_pool() sees fresh state
         # instead of re-seeding stale/consumed tokens.
         self._sync_device_code_entry_to_auth_store(updated)
-        # Write refreshed tokens back to ~/.codex/auth.json so Codex CLI
-        # and VS Code don't hit "refresh_token_reused" on their next refresh.
-        if self.provider == "openai-codex":
-            try:
-                _write_codex_cli_tokens(
-                    updated.access_token,
-                    updated.refresh_token,
-                    last_refresh=updated.last_refresh,
-                )
-            except Exception as wexc:
-                logger.debug("Failed to write refreshed Codex tokens to CLI file: %s", wexc)
         return updated
 
     def _entry_needs_refresh(self, entry: PooledCredential) -> bool:
@@ -790,16 +698,6 @@ class CredentialPool:
                 if synced is not entry:
                     entry = synced
                     cleared_any = True
-            # For openai-codex entries, sync from ~/.codex/auth.json before
-            # any status/refresh checks.  This picks up tokens refreshed by
-            # the Codex CLI or another Hermes profile.
-            if (self.provider == "openai-codex"
-                    and entry.last_status == STATUS_EXHAUSTED
-                    and entry.refresh_token):
-                synced = self._sync_codex_entry_from_cli(entry)
-                if synced is not entry:
-                    entry = synced
-                    cleared_any = True
             if entry.last_status == STATUS_EXHAUSTED:
                 exhausted_until = _exhausted_until(entry)
                 if exhausted_until is not None and now < exhausted_until:
@@ -1130,6 +1028,14 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
         state = _load_provider_state(auth_store, "nous")
         if state:
             active_sources.add("device_code")
+            # Prefer a user-supplied label embedded in the singleton state
+            # (set by persist_nous_credentials(label=...) when the user ran
+            # `hermes auth add nous --label <name>`).  Fall back to the
+            # auto-derived token fingerprint for logins that didn't supply one.
+            custom_label = str(state.get("label") or "").strip()
+            seeded_label = custom_label or label_from_token(
+                state.get("access_token", ""), "device_code"
+            )
             changed |= _upsert_entry(
                 entries,
                 provider,
@@ -1148,7 +1054,7 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
                     "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"),
+                    "label": seeded_label,
                 },
             )
 
@@ -1162,6 +1068,7 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
             if token:
                 source_name = "gh_cli" if "gh" in source.lower() else f"env:{source}"
                 active_sources.add(source_name)
+                pconfig = PROVIDER_REGISTRY.get(provider)
                 changed |= _upsert_entry(
                     entries,
                     provider,
@@ -1170,6 +1077,7 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
                         "source": source_name,
                         "auth_type": AUTH_TYPE_API_KEY,
                         "access_token": token,
+                        "base_url": pconfig.inference_base_url if pconfig else "",
                         "label": source,
                     },
                 )
@@ -1206,25 +1114,27 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
             logger.debug("Qwen OAuth token seed failed: %s", exc)
 
     elif provider == "openai-codex":
+        # Respect user suppression — `hermes auth remove openai-codex` marks
+        # the device_code source as suppressed so it won't be re-seeded from
+        # the Hermes auth store.  Without this gate the removal is instantly
+        # undone on the next load_pool() call.
+        codex_suppressed = False
+        try:
+            from hermes_cli.auth import is_source_suppressed
+            codex_suppressed = is_source_suppressed(provider, "device_code")
+        except ImportError:
+            pass
+        if codex_suppressed:
+            return changed, active_sources
+
         state = _load_provider_state(auth_store, "openai-codex")
         tokens = state.get("tokens") if isinstance(state, dict) else None
-        # Fallback: import from Codex CLI (~/.codex/auth.json) if Hermes auth
-        # store has no tokens.  This mirrors resolve_codex_runtime_credentials()
-        # so that load_pool() and list_authenticated_providers() detect tokens
-        # that only exist in the Codex CLI shared file.
-        if not (isinstance(tokens, dict) and tokens.get("access_token")):
-            try:
-                from hermes_cli.auth import _import_codex_cli_tokens, _save_codex_tokens
-                cli_tokens = _import_codex_cli_tokens()
-                if cli_tokens:
-                    logger.info("Importing Codex CLI tokens into Hermes auth store.")
-                    _save_codex_tokens(cli_tokens)
-                    # Re-read state after import
-                    auth_store = _load_auth_store()
-                    state = _load_provider_state(auth_store, "openai-codex")
-                    tokens = state.get("tokens") if isinstance(state, dict) else None
-            except Exception as exc:
-                logger.debug("Codex CLI token import failed: %s", exc)
+        # Hermes owns its own Codex auth state — we do NOT auto-import from
+        # ~/.codex/auth.json at pool-load time.  OAuth refresh tokens are
+        # single-use, so sharing them with Codex CLI / VS Code causes
+        # refresh_token_reused race failures.  Users who want to adopt
+        # existing Codex CLI credentials get a one-time, explicit prompt
+        # via `hermes auth openai-codex`.
         if isinstance(tokens, dict) and tokens.get("access_token"):
             active_sources.add("device_code")
             changed |= _upsert_entry(

agent/display.py

@@ -225,9 +225,11 @@ def build_tool_preview(tool_name: str, args: dict, max_len: int | None = None) -
             content = _oneline(args.get("content", ""))
             return f"+{target}: \"{content[:25]}{'...' if len(content) > 25 else ''}\""
         elif action == "replace":
-            return f"~{target}: \"{_oneline(args.get('old_text', '')[:20])}\""
+            old = _oneline(args.get("old_text") or "") or "<missing old_text>"
+            return f"~{target}: \"{old[:20]}\""
         elif action == "remove":
-            return f"-{target}: \"{_oneline(args.get('old_text', '')[:20])}\""
+            old = _oneline(args.get("old_text") or "") or "<missing old_text>"
+            return f"-{target}: \"{old[:20]}\""
         return action
 
     if tool_name == "send_message":
@@ -600,6 +602,45 @@ class KawaiiSpinner:
         "analyzing", "computing", "synthesizing", "formulating", "brainstorming",
     ]
 
+    @classmethod
+    def get_waiting_faces(cls) -> list:
+        """Return waiting faces from the active skin, falling back to KAWAII_WAITING."""
+        try:
+            skin = _get_skin()
+            if skin:
+                faces = skin.spinner.get("waiting_faces", [])
+                if faces:
+                    return faces
+        except Exception:
+            pass
+        return cls.KAWAII_WAITING
+
+    @classmethod
+    def get_thinking_faces(cls) -> list:
+        """Return thinking faces from the active skin, falling back to KAWAII_THINKING."""
+        try:
+            skin = _get_skin()
+            if skin:
+                faces = skin.spinner.get("thinking_faces", [])
+                if faces:
+                    return faces
+        except Exception:
+            pass
+        return cls.KAWAII_THINKING
+
+    @classmethod
+    def get_thinking_verbs(cls) -> list:
+        """Return thinking verbs from the active skin, falling back to THINKING_VERBS."""
+        try:
+            skin = _get_skin()
+            if skin:
+                verbs = skin.spinner.get("thinking_verbs", [])
+                if verbs:
+                    return verbs
+        except Exception:
+            pass
+        return cls.THINKING_VERBS
+
     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'])
@@ -900,9 +941,13 @@ def get_cute_tool_message(
         if action == "add":
             return _wrap(f"┊ 🧠 memory    +{target}: \"{_trunc(args.get('content', ''), 30)}\"  {dur}")
         elif action == "replace":
-            return _wrap(f"┊ 🧠 memory    ~{target}: \"{_trunc(args.get('old_text', ''), 20)}\"  {dur}")
+            old = args.get("old_text") or ""
+            old = old if old else "<missing old_text>"
+            return _wrap(f"┊ 🧠 memory    ~{target}: \"{_trunc(old, 20)}\"  {dur}")
         elif action == "remove":
-            return _wrap(f"┊ 🧠 memory    -{target}: \"{_trunc(args.get('old_text', ''), 20)}\"  {dur}")
+            old = args.get("old_text") or ""
+            old = old if old else "<missing old_text>"
+            return _wrap(f"┊ 🧠 memory    -{target}: \"{_trunc(old, 20)}\"  {dur}")
         return _wrap(f"┊ 🧠 memory    {action}  {dur}")
     if tool_name == "skills_list":
         return _wrap(f"┊ 📚 skills    list {args.get('category', 'all')}  {dur}")
@@ -954,84 +999,4 @@ def get_cute_tool_message(
 # Honcho session line (one-liner with clickable OSC 8 hyperlink)
 # =========================================================================
 
-_DIM = "\033[2m"
-_SKY_BLUE = "\033[38;5;117m"
-_ANSI_RESET = "\033[0m"
 
-
-# =========================================================================
-# Context pressure display (CLI user-facing warnings)
-# =========================================================================
-
-# ANSI color codes for context pressure tiers
-_CYAN = "\033[36m"
-_YELLOW = "\033[33m"
-_BOLD = "\033[1m"
-_DIM_ANSI = "\033[2m"
-
-# Bar characters
-_BAR_FILLED = "▰"
-_BAR_EMPTY = "▱"
-_BAR_WIDTH = 20
-
-
-def format_context_pressure(
-    compaction_progress: float,
-    threshold_tokens: int,
-    threshold_percent: float,
-    compression_enabled: bool = True,
-) -> str:
-    """Build a formatted context pressure line for CLI display.
-
-    The bar and percentage show progress toward the compaction threshold,
-    NOT the raw context window.  100% = compaction fires.
-
-    Args:
-        compaction_progress: How close to compaction (0.0–1.0, 1.0 = fires).
-        threshold_tokens: Compaction threshold in tokens.
-        threshold_percent: Compaction threshold as a fraction of context window.
-        compression_enabled: Whether auto-compression is active.
-    """
-    pct_int = min(int(compaction_progress * 100), 100)
-    filled = min(int(compaction_progress * _BAR_WIDTH), _BAR_WIDTH)
-    bar = _BAR_FILLED * filled + _BAR_EMPTY * (_BAR_WIDTH - filled)
-
-    threshold_k = f"{threshold_tokens // 1000}k" if threshold_tokens >= 1000 else str(threshold_tokens)
-    threshold_pct_int = int(threshold_percent * 100)
-
-    color = f"{_BOLD}{_YELLOW}"
-    icon = "⚠"
-    if compression_enabled:
-        hint = "compaction approaching"
-    else:
-        hint = "no auto-compaction"
-
-    return (
-        f"  {color}{icon} context {bar} {pct_int}% to compaction{_ANSI_RESET}"
-        f"  {_DIM_ANSI}{threshold_k} threshold ({threshold_pct_int}%) · {hint}{_ANSI_RESET}"
-    )
-
-
-def format_context_pressure_gateway(
-    compaction_progress: float,
-    threshold_percent: float,
-    compression_enabled: bool = True,
-) -> str:
-    """Build a plain-text context pressure notification for messaging platforms.
-
-    No ANSI — just Unicode and plain text suitable for Telegram/Discord/etc.
-    The percentage shows progress toward the compaction threshold.
-    """
-    pct_int = min(int(compaction_progress * 100), 100)
-    filled = min(int(compaction_progress * _BAR_WIDTH), _BAR_WIDTH)
-    bar = _BAR_FILLED * filled + _BAR_EMPTY * (_BAR_WIDTH - filled)
-
-    threshold_pct_int = int(threshold_percent * 100)
-
-    icon = "⚠️"
-    if compression_enabled:
-        hint = f"Context compaction approaching (threshold: {threshold_pct_int}% of window)."
-    else:
-        hint = "Auto-compaction is disabled — context may be truncated."
-
-    return f"{icon} Context: {bar} {pct_int}% to compaction\n{hint}"

agent/error_classifier.py

@@ -112,6 +112,10 @@ _RATE_LIMIT_PATTERNS = [
     "please retry after",
     "resource_exhausted",
     "rate increased too quickly",  # Alibaba/DashScope throttling
+    # AWS Bedrock throttling
+    "throttlingexception",
+    "too many concurrent requests",
+    "servicequotaexceededexception",
 ]
 
 # Usage-limit patterns that need disambiguation (could be billing OR rate_limit)
@@ -171,6 +175,11 @@ _CONTEXT_OVERFLOW_PATTERNS = [
     # Chinese error messages (some providers return these)
     "超过最大长度",
     "上下文长度",
+    # AWS Bedrock Converse API error patterns
+    "input is too long",
+    "max input token",
+    "input token",
+    "exceeds the maximum number of input tokens",
 ]
 
 # Model not found patterns
@@ -281,7 +290,7 @@ def classify_api_error(
     if isinstance(body, dict):
         _err_obj = body.get("error", {})
         if isinstance(_err_obj, dict):
-            _body_msg = (_err_obj.get("message") or "").lower()
+            _body_msg = str(_err_obj.get("message") or "").lower()
             # Parse metadata.raw for wrapped provider errors
             _metadata = _err_obj.get("metadata", {})
             if isinstance(_metadata, dict):
@@ -293,11 +302,11 @@ def classify_api_error(
                         if isinstance(_inner, dict):
                             _inner_err = _inner.get("error", {})
                             if isinstance(_inner_err, dict):
-                                _metadata_msg = (_inner_err.get("message") or "").lower()
+                                _metadata_msg = str(_inner_err.get("message") or "").lower()
                     except (json.JSONDecodeError, TypeError):
                         pass
         if not _body_msg:
-            _body_msg = (body.get("message") or "").lower()
+            _body_msg = str(body.get("message") or "").lower()
     # Combine all message sources for pattern matching
     parts = [_raw_msg]
     if _body_msg and _body_msg not in _raw_msg:
@@ -597,10 +606,10 @@ def _classify_400(
     if isinstance(body, dict):
         err_obj = body.get("error", {})
         if isinstance(err_obj, dict):
-            err_body_msg = (err_obj.get("message") or "").strip().lower()
+            err_body_msg = str(err_obj.get("message") or "").strip().lower()
         # Responses API (and some providers) use flat body: {"message": "..."}
         if not err_body_msg:
-            err_body_msg = (body.get("message") or "").strip().lower()
+            err_body_msg = str(body.get("message") or "").strip().lower()
     is_generic = len(err_body_msg) < 30 or err_body_msg in ("error", "")
     is_large = approx_tokens > context_length * 0.4 or approx_tokens > 80000 or num_messages > 80
 

agent/gemini_cloudcode_adapter.py

@@ -0,0 +1,904 @@
+"""OpenAI-compatible facade that talks to Google's Cloud Code Assist backend.
+
+This adapter lets Hermes use the ``google-gemini-cli`` provider as if it were
+a standard OpenAI-shaped chat completion endpoint, while the underlying HTTP
+traffic goes to ``cloudcode-pa.googleapis.com/v1internal:{generateContent,
+streamGenerateContent}`` with a Bearer access token obtained via OAuth PKCE.
+
+Architecture
+------------
+- ``GeminiCloudCodeClient`` exposes ``.chat.completions.create(**kwargs)``
+  mirroring the subset of the OpenAI SDK that ``run_agent.py`` uses.
+- Incoming OpenAI ``messages[]`` / ``tools[]`` / ``tool_choice`` are translated
+  to Gemini's native ``contents[]`` / ``tools[].functionDeclarations`` /
+  ``toolConfig`` / ``systemInstruction`` shape.
+- The request body is wrapped ``{project, model, user_prompt_id, request}``
+  per Code Assist API expectations.
+- Responses (``candidates[].content.parts[]``) are converted back to
+  OpenAI ``choices[0].message`` shape with ``content`` + ``tool_calls``.
+- Streaming uses SSE (``?alt=sse``) and yields OpenAI-shaped delta chunks.
+
+Attribution
+-----------
+Translation semantics follow jenslys/opencode-gemini-auth (MIT) and the public
+Gemini API docs. Request envelope shape
+(``{project, model, user_prompt_id, request}``) is documented nowhere; it is
+reverse-engineered from the opencode-gemini-auth and clawdbot implementations.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import time
+import uuid
+from types import SimpleNamespace
+from typing import Any, Dict, Iterator, List, Optional
+
+import httpx
+
+from agent import google_oauth
+from agent.gemini_schema import sanitize_gemini_tool_parameters
+from agent.google_code_assist import (
+    CODE_ASSIST_ENDPOINT,
+    FREE_TIER_ID,
+    CodeAssistError,
+    ProjectContext,
+    resolve_project_context,
+)
+
+logger = logging.getLogger(__name__)
+
+
+# =============================================================================
+# Request translation: OpenAI → Gemini
+# =============================================================================
+
+_ROLE_MAP_OPENAI_TO_GEMINI = {
+    "user": "user",
+    "assistant": "model",
+    "system": "user",   # handled separately via systemInstruction
+    "tool": "user",     # functionResponse is wrapped in a user-role turn
+    "function": "user",
+}
+
+
+def _coerce_content_to_text(content: Any) -> str:
+    """OpenAI content may be str or a list of parts; reduce to plain text."""
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        pieces: List[str] = []
+        for p in content:
+            if isinstance(p, str):
+                pieces.append(p)
+            elif isinstance(p, dict):
+                if p.get("type") == "text" and isinstance(p.get("text"), str):
+                    pieces.append(p["text"])
+                # Multimodal (image_url, etc.) — stub for now; log and skip
+                elif p.get("type") in ("image_url", "input_audio"):
+                    logger.debug("Dropping multimodal part (not yet supported): %s", p.get("type"))
+        return "\n".join(pieces)
+    return str(content)
+
+
+def _translate_tool_call_to_gemini(tool_call: Dict[str, Any]) -> Dict[str, Any]:
+    """OpenAI tool_call -> Gemini functionCall part."""
+    fn = tool_call.get("function") or {}
+    args_raw = fn.get("arguments", "")
+    try:
+        args = json.loads(args_raw) if isinstance(args_raw, str) and args_raw else {}
+    except json.JSONDecodeError:
+        args = {"_raw": args_raw}
+    if not isinstance(args, dict):
+        args = {"_value": args}
+    return {
+        "functionCall": {
+            "name": fn.get("name") or "",
+            "args": args,
+        },
+        # Sentinel signature — matches opencode-gemini-auth's approach.
+        # Without this, Code Assist rejects function calls that originated
+        # outside its own chain.
+        "thoughtSignature": "skip_thought_signature_validator",
+    }
+
+
+def _translate_tool_result_to_gemini(message: Dict[str, Any]) -> Dict[str, Any]:
+    """OpenAI tool-role message -> Gemini functionResponse part.
+
+    The function name isn't in the OpenAI tool message directly; it must be
+    passed via the assistant message that issued the call. For simplicity we
+    look up ``name`` on the message (OpenAI SDK copies it there) or on the
+    ``tool_call_id`` cross-reference.
+    """
+    name = str(message.get("name") or message.get("tool_call_id") or "tool")
+    content = _coerce_content_to_text(message.get("content"))
+    # Gemini expects the response as a dict under `response`. We wrap plain
+    # text in {"output": "..."}.
+    try:
+        parsed = json.loads(content) if content.strip().startswith(("{", "[")) else None
+    except json.JSONDecodeError:
+        parsed = None
+    response = parsed if isinstance(parsed, dict) else {"output": content}
+    return {
+        "functionResponse": {
+            "name": name,
+            "response": response,
+        },
+    }
+
+
+def _build_gemini_contents(
+    messages: List[Dict[str, Any]],
+) -> tuple[List[Dict[str, Any]], Optional[Dict[str, Any]]]:
+    """Convert OpenAI messages[] to Gemini contents[] + systemInstruction."""
+    system_text_parts: List[str] = []
+    contents: List[Dict[str, Any]] = []
+
+    for msg in messages:
+        if not isinstance(msg, dict):
+            continue
+        role = str(msg.get("role") or "user")
+
+        if role == "system":
+            system_text_parts.append(_coerce_content_to_text(msg.get("content")))
+            continue
+
+        # Tool result message — emit a user-role turn with functionResponse
+        if role == "tool" or role == "function":
+            contents.append({
+                "role": "user",
+                "parts": [_translate_tool_result_to_gemini(msg)],
+            })
+            continue
+
+        gemini_role = _ROLE_MAP_OPENAI_TO_GEMINI.get(role, "user")
+        parts: List[Dict[str, Any]] = []
+
+        text = _coerce_content_to_text(msg.get("content"))
+        if text:
+            parts.append({"text": text})
+
+        # Assistant messages can carry tool_calls
+        tool_calls = msg.get("tool_calls") or []
+        if isinstance(tool_calls, list):
+            for tc in tool_calls:
+                if isinstance(tc, dict):
+                    parts.append(_translate_tool_call_to_gemini(tc))
+
+        if not parts:
+            # Gemini rejects empty parts; skip the turn entirely
+            continue
+
+        contents.append({"role": gemini_role, "parts": parts})
+
+    system_instruction: Optional[Dict[str, Any]] = None
+    joined_system = "\n".join(p for p in system_text_parts if p).strip()
+    if joined_system:
+        system_instruction = {
+            "role": "system",
+            "parts": [{"text": joined_system}],
+        }
+
+    return contents, system_instruction
+
+
+def _translate_tools_to_gemini(tools: Any) -> List[Dict[str, Any]]:
+    """OpenAI tools[] -> Gemini tools[].functionDeclarations[]."""
+    if not isinstance(tools, list) or not tools:
+        return []
+    declarations: List[Dict[str, Any]] = []
+    for t in tools:
+        if not isinstance(t, dict):
+            continue
+        fn = t.get("function") or {}
+        if not isinstance(fn, dict):
+            continue
+        name = fn.get("name")
+        if not name:
+            continue
+        decl = {"name": str(name)}
+        if fn.get("description"):
+            decl["description"] = str(fn["description"])
+        params = fn.get("parameters")
+        if isinstance(params, dict):
+            decl["parameters"] = sanitize_gemini_tool_parameters(params)
+        declarations.append(decl)
+    if not declarations:
+        return []
+    return [{"functionDeclarations": declarations}]
+
+
+def _translate_tool_choice_to_gemini(tool_choice: Any) -> Optional[Dict[str, Any]]:
+    """OpenAI tool_choice -> Gemini toolConfig.functionCallingConfig."""
+    if tool_choice is None:
+        return None
+    if isinstance(tool_choice, str):
+        if tool_choice == "auto":
+            return {"functionCallingConfig": {"mode": "AUTO"}}
+        if tool_choice == "required":
+            return {"functionCallingConfig": {"mode": "ANY"}}
+        if tool_choice == "none":
+            return {"functionCallingConfig": {"mode": "NONE"}}
+    if isinstance(tool_choice, dict):
+        fn = tool_choice.get("function") or {}
+        name = fn.get("name")
+        if name:
+            return {
+                "functionCallingConfig": {
+                    "mode": "ANY",
+                    "allowedFunctionNames": [str(name)],
+                },
+            }
+    return None
+
+
+def _normalize_thinking_config(config: Any) -> Optional[Dict[str, Any]]:
+    """Accept thinkingBudget / thinkingLevel / includeThoughts (+ snake_case)."""
+    if not isinstance(config, dict) or not config:
+        return None
+    budget = config.get("thinkingBudget", config.get("thinking_budget"))
+    level = config.get("thinkingLevel", config.get("thinking_level"))
+    include = config.get("includeThoughts", config.get("include_thoughts"))
+    normalized: Dict[str, Any] = {}
+    if isinstance(budget, (int, float)):
+        normalized["thinkingBudget"] = int(budget)
+    if isinstance(level, str) and level.strip():
+        normalized["thinkingLevel"] = level.strip().lower()
+    if isinstance(include, bool):
+        normalized["includeThoughts"] = include
+    return normalized or None
+
+
+def build_gemini_request(
+    *,
+    messages: List[Dict[str, Any]],
+    tools: Any = None,
+    tool_choice: Any = None,
+    temperature: Optional[float] = None,
+    max_tokens: Optional[int] = None,
+    top_p: Optional[float] = None,
+    stop: Any = None,
+    thinking_config: Any = None,
+) -> Dict[str, Any]:
+    """Build the inner Gemini request body (goes inside ``request`` wrapper)."""
+    contents, system_instruction = _build_gemini_contents(messages)
+
+    body: Dict[str, Any] = {"contents": contents}
+    if system_instruction is not None:
+        body["systemInstruction"] = system_instruction
+
+    gemini_tools = _translate_tools_to_gemini(tools)
+    if gemini_tools:
+        body["tools"] = gemini_tools
+    tool_cfg = _translate_tool_choice_to_gemini(tool_choice)
+    if tool_cfg is not None:
+        body["toolConfig"] = tool_cfg
+
+    generation_config: Dict[str, Any] = {}
+    if isinstance(temperature, (int, float)):
+        generation_config["temperature"] = float(temperature)
+    if isinstance(max_tokens, int) and max_tokens > 0:
+        generation_config["maxOutputTokens"] = max_tokens
+    if isinstance(top_p, (int, float)):
+        generation_config["topP"] = float(top_p)
+    if isinstance(stop, str) and stop:
+        generation_config["stopSequences"] = [stop]
+    elif isinstance(stop, list) and stop:
+        generation_config["stopSequences"] = [str(s) for s in stop if s]
+    normalized_thinking = _normalize_thinking_config(thinking_config)
+    if normalized_thinking:
+        generation_config["thinkingConfig"] = normalized_thinking
+    if generation_config:
+        body["generationConfig"] = generation_config
+
+    return body
+
+
+def wrap_code_assist_request(
+    *,
+    project_id: str,
+    model: str,
+    inner_request: Dict[str, Any],
+    user_prompt_id: Optional[str] = None,
+) -> Dict[str, Any]:
+    """Wrap the inner Gemini request in the Code Assist envelope."""
+    return {
+        "project": project_id,
+        "model": model,
+        "user_prompt_id": user_prompt_id or str(uuid.uuid4()),
+        "request": inner_request,
+    }
+
+
+# =============================================================================
+# Response translation: Gemini → OpenAI
+# =============================================================================
+
+def _translate_gemini_response(
+    resp: Dict[str, Any],
+    model: str,
+) -> SimpleNamespace:
+    """Non-streaming Gemini response -> OpenAI-shaped SimpleNamespace.
+
+    Code Assist wraps the actual Gemini response inside ``response``, so we
+    unwrap it first if present.
+    """
+    inner = resp.get("response") if isinstance(resp.get("response"), dict) else resp
+
+    candidates = inner.get("candidates") or []
+    if not isinstance(candidates, list) or not candidates:
+        return _empty_response(model)
+
+    cand = candidates[0]
+    content_obj = cand.get("content") if isinstance(cand, dict) else {}
+    parts = content_obj.get("parts") if isinstance(content_obj, dict) else []
+
+    text_pieces: List[str] = []
+    reasoning_pieces: List[str] = []
+    tool_calls: List[SimpleNamespace] = []
+
+    for i, part in enumerate(parts or []):
+        if not isinstance(part, dict):
+            continue
+        # Thought parts are model's internal reasoning — surface as reasoning,
+        # don't mix into content.
+        if part.get("thought") is True:
+            if isinstance(part.get("text"), str):
+                reasoning_pieces.append(part["text"])
+            continue
+        if isinstance(part.get("text"), str):
+            text_pieces.append(part["text"])
+            continue
+        fc = part.get("functionCall")
+        if isinstance(fc, dict) and fc.get("name"):
+            try:
+                args_str = json.dumps(fc.get("args") or {}, ensure_ascii=False)
+            except (TypeError, ValueError):
+                args_str = "{}"
+            tool_calls.append(SimpleNamespace(
+                id=f"call_{uuid.uuid4().hex[:12]}",
+                type="function",
+                index=i,
+                function=SimpleNamespace(name=str(fc["name"]), arguments=args_str),
+            ))
+
+    finish_reason = "tool_calls" if tool_calls else _map_gemini_finish_reason(
+        str(cand.get("finishReason") or "")
+    )
+
+    usage_meta = inner.get("usageMetadata") or {}
+    usage = SimpleNamespace(
+        prompt_tokens=int(usage_meta.get("promptTokenCount") or 0),
+        completion_tokens=int(usage_meta.get("candidatesTokenCount") or 0),
+        total_tokens=int(usage_meta.get("totalTokenCount") or 0),
+        prompt_tokens_details=SimpleNamespace(
+            cached_tokens=int(usage_meta.get("cachedContentTokenCount") or 0),
+        ),
+    )
+
+    message = SimpleNamespace(
+        role="assistant",
+        content="".join(text_pieces) if text_pieces else None,
+        tool_calls=tool_calls or None,
+        reasoning="".join(reasoning_pieces) or None,
+        reasoning_content="".join(reasoning_pieces) or None,
+        reasoning_details=None,
+    )
+    choice = SimpleNamespace(
+        index=0,
+        message=message,
+        finish_reason=finish_reason,
+    )
+    return SimpleNamespace(
+        id=f"chatcmpl-{uuid.uuid4().hex[:12]}",
+        object="chat.completion",
+        created=int(time.time()),
+        model=model,
+        choices=[choice],
+        usage=usage,
+    )
+
+
+def _empty_response(model: str) -> SimpleNamespace:
+    message = SimpleNamespace(
+        role="assistant", content="", tool_calls=None,
+        reasoning=None, reasoning_content=None, reasoning_details=None,
+    )
+    choice = SimpleNamespace(index=0, message=message, finish_reason="stop")
+    usage = SimpleNamespace(
+        prompt_tokens=0, completion_tokens=0, total_tokens=0,
+        prompt_tokens_details=SimpleNamespace(cached_tokens=0),
+    )
+    return SimpleNamespace(
+        id=f"chatcmpl-{uuid.uuid4().hex[:12]}",
+        object="chat.completion",
+        created=int(time.time()),
+        model=model,
+        choices=[choice],
+        usage=usage,
+    )
+
+
+def _map_gemini_finish_reason(reason: str) -> str:
+    mapping = {
+        "STOP": "stop",
+        "MAX_TOKENS": "length",
+        "SAFETY": "content_filter",
+        "RECITATION": "content_filter",
+        "OTHER": "stop",
+    }
+    return mapping.get(reason.upper(), "stop")
+
+
+# =============================================================================
+# Streaming SSE iterator
+# =============================================================================
+
+class _GeminiStreamChunk(SimpleNamespace):
+    """Mimics an OpenAI ChatCompletionChunk with .choices[0].delta."""
+    pass
+
+
+def _make_stream_chunk(
+    *,
+    model: str,
+    content: str = "",
+    tool_call_delta: Optional[Dict[str, Any]] = None,
+    finish_reason: Optional[str] = None,
+    reasoning: str = "",
+) -> _GeminiStreamChunk:
+    delta_kwargs: Dict[str, Any] = {"role": "assistant"}
+    if content:
+        delta_kwargs["content"] = content
+    if tool_call_delta is not None:
+        delta_kwargs["tool_calls"] = [SimpleNamespace(
+            index=tool_call_delta.get("index", 0),
+            id=tool_call_delta.get("id") or f"call_{uuid.uuid4().hex[:12]}",
+            type="function",
+            function=SimpleNamespace(
+                name=tool_call_delta.get("name") or "",
+                arguments=tool_call_delta.get("arguments") or "",
+            ),
+        )]
+    if reasoning:
+        delta_kwargs["reasoning"] = reasoning
+        delta_kwargs["reasoning_content"] = reasoning
+    delta = SimpleNamespace(**delta_kwargs)
+    choice = SimpleNamespace(index=0, delta=delta, finish_reason=finish_reason)
+    return _GeminiStreamChunk(
+        id=f"chatcmpl-{uuid.uuid4().hex[:12]}",
+        object="chat.completion.chunk",
+        created=int(time.time()),
+        model=model,
+        choices=[choice],
+        usage=None,
+    )
+
+
+def _iter_sse_events(response: httpx.Response) -> Iterator[Dict[str, Any]]:
+    """Parse Server-Sent Events from an httpx streaming response."""
+    buffer = ""
+    for chunk in response.iter_text():
+        if not chunk:
+            continue
+        buffer += chunk
+        while "\n" in buffer:
+            line, buffer = buffer.split("\n", 1)
+            line = line.rstrip("\r")
+            if not line:
+                continue
+            if line.startswith("data: "):
+                data = line[6:]
+                if data == "[DONE]":
+                    return
+                try:
+                    yield json.loads(data)
+                except json.JSONDecodeError:
+                    logger.debug("Non-JSON SSE line: %s", data[:200])
+
+
+def _translate_stream_event(
+    event: Dict[str, Any],
+    model: str,
+    tool_call_counter: List[int],
+) -> List[_GeminiStreamChunk]:
+    """Unwrap Code Assist envelope and emit OpenAI-shaped chunk(s).
+
+    ``tool_call_counter`` is a single-element list used as a mutable counter
+    across events in the same stream. Each ``functionCall`` part gets a
+    fresh, unique OpenAI ``index`` — keying by function name would collide
+    whenever the model issues parallel calls to the same tool (e.g. reading
+    three files in one turn).
+    """
+    inner = event.get("response") if isinstance(event.get("response"), dict) else event
+    candidates = inner.get("candidates") or []
+    if not candidates:
+        return []
+    cand = candidates[0]
+    if not isinstance(cand, dict):
+        return []
+
+    chunks: List[_GeminiStreamChunk] = []
+
+    content = cand.get("content") or {}
+    parts = content.get("parts") if isinstance(content, dict) else []
+    for part in parts or []:
+        if not isinstance(part, dict):
+            continue
+        if part.get("thought") is True and isinstance(part.get("text"), str):
+            chunks.append(_make_stream_chunk(
+                model=model, reasoning=part["text"],
+            ))
+            continue
+        if isinstance(part.get("text"), str) and part["text"]:
+            chunks.append(_make_stream_chunk(model=model, content=part["text"]))
+        fc = part.get("functionCall")
+        if isinstance(fc, dict) and fc.get("name"):
+            name = str(fc["name"])
+            idx = tool_call_counter[0]
+            tool_call_counter[0] += 1
+            try:
+                args_str = json.dumps(fc.get("args") or {}, ensure_ascii=False)
+            except (TypeError, ValueError):
+                args_str = "{}"
+            chunks.append(_make_stream_chunk(
+                model=model,
+                tool_call_delta={
+                    "index": idx,
+                    "name": name,
+                    "arguments": args_str,
+                },
+            ))
+
+    finish_reason_raw = str(cand.get("finishReason") or "")
+    if finish_reason_raw:
+        mapped = _map_gemini_finish_reason(finish_reason_raw)
+        if tool_call_counter[0] > 0:
+            mapped = "tool_calls"
+        chunks.append(_make_stream_chunk(model=model, finish_reason=mapped))
+    return chunks
+
+
+# =============================================================================
+# GeminiCloudCodeClient — OpenAI-compatible facade
+# =============================================================================
+
+MARKER_BASE_URL = "cloudcode-pa://google"
+
+
+class _GeminiChatCompletions:
+    def __init__(self, client: "GeminiCloudCodeClient"):
+        self._client = client
+
+    def create(self, **kwargs: Any) -> Any:
+        return self._client._create_chat_completion(**kwargs)
+
+
+class _GeminiChatNamespace:
+    def __init__(self, client: "GeminiCloudCodeClient"):
+        self.completions = _GeminiChatCompletions(client)
+
+
+class GeminiCloudCodeClient:
+    """Minimal OpenAI-SDK-compatible facade over Code Assist v1internal."""
+
+    def __init__(
+        self,
+        *,
+        api_key: Optional[str] = None,
+        base_url: Optional[str] = None,
+        default_headers: Optional[Dict[str, str]] = None,
+        project_id: str = "",
+        **_: Any,
+    ):
+        # `api_key` here is a dummy — real auth is the OAuth access token
+        # fetched on every call via agent.google_oauth.get_valid_access_token().
+        # We accept the kwarg for openai.OpenAI interface parity.
+        self.api_key = api_key or "google-oauth"
+        self.base_url = base_url or MARKER_BASE_URL
+        self._default_headers = dict(default_headers or {})
+        self._configured_project_id = project_id
+        self._project_context: Optional[ProjectContext] = None
+        self._project_context_lock = False  # simple single-thread guard
+        self.chat = _GeminiChatNamespace(self)
+        self.is_closed = False
+        self._http = httpx.Client(timeout=httpx.Timeout(connect=15.0, read=600.0, write=30.0, pool=30.0))
+
+    def close(self) -> None:
+        self.is_closed = True
+        try:
+            self._http.close()
+        except Exception:
+            pass
+
+    # Implement the OpenAI SDK's context-manager-ish closure check
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        self.close()
+
+    def _ensure_project_context(self, access_token: str, model: str) -> ProjectContext:
+        """Lazily resolve and cache the project context for this client."""
+        if self._project_context is not None:
+            return self._project_context
+
+        env_project = google_oauth.resolve_project_id_from_env()
+        creds = google_oauth.load_credentials()
+        stored_project = creds.project_id if creds else ""
+
+        # Prefer what's already baked into the creds
+        if stored_project:
+            self._project_context = ProjectContext(
+                project_id=stored_project,
+                managed_project_id=creds.managed_project_id if creds else "",
+                tier_id="",
+                source="stored",
+            )
+            return self._project_context
+
+        ctx = resolve_project_context(
+            access_token,
+            configured_project_id=self._configured_project_id,
+            env_project_id=env_project,
+            user_agent_model=model,
+        )
+        # Persist discovered project back to the creds file so the next
+        # session doesn't re-run the discovery.
+        if ctx.project_id or ctx.managed_project_id:
+            google_oauth.update_project_ids(
+                project_id=ctx.project_id,
+                managed_project_id=ctx.managed_project_id,
+            )
+        self._project_context = ctx
+        return ctx
+
+    def _create_chat_completion(
+        self,
+        *,
+        model: str = "gemini-2.5-flash",
+        messages: Optional[List[Dict[str, Any]]] = None,
+        stream: bool = False,
+        tools: Any = None,
+        tool_choice: Any = None,
+        temperature: Optional[float] = None,
+        max_tokens: Optional[int] = None,
+        top_p: Optional[float] = None,
+        stop: Any = None,
+        extra_body: Optional[Dict[str, Any]] = None,
+        timeout: Any = None,
+        **_: Any,
+    ) -> Any:
+        access_token = google_oauth.get_valid_access_token()
+        ctx = self._ensure_project_context(access_token, model)
+
+        thinking_config = None
+        if isinstance(extra_body, dict):
+            thinking_config = extra_body.get("thinking_config") or extra_body.get("thinkingConfig")
+
+        inner = build_gemini_request(
+            messages=messages or [],
+            tools=tools,
+            tool_choice=tool_choice,
+            temperature=temperature,
+            max_tokens=max_tokens,
+            top_p=top_p,
+            stop=stop,
+            thinking_config=thinking_config,
+        )
+        wrapped = wrap_code_assist_request(
+            project_id=ctx.project_id,
+            model=model,
+            inner_request=inner,
+        )
+
+        headers = {
+            "Content-Type": "application/json",
+            "Accept": "application/json",
+            "Authorization": f"Bearer {access_token}",
+            "User-Agent": "hermes-agent (gemini-cli-compat)",
+            "X-Goog-Api-Client": "gl-python/hermes",
+            "x-activity-request-id": str(uuid.uuid4()),
+        }
+        headers.update(self._default_headers)
+
+        if stream:
+            return self._stream_completion(model=model, wrapped=wrapped, headers=headers)
+
+        url = f"{CODE_ASSIST_ENDPOINT}/v1internal:generateContent"
+        response = self._http.post(url, json=wrapped, headers=headers)
+        if response.status_code != 200:
+            raise _gemini_http_error(response)
+        try:
+            payload = response.json()
+        except ValueError as exc:
+            raise CodeAssistError(
+                f"Invalid JSON from Code Assist: {exc}",
+                code="code_assist_invalid_json",
+            ) from exc
+        return _translate_gemini_response(payload, model=model)
+
+    def _stream_completion(
+        self,
+        *,
+        model: str,
+        wrapped: Dict[str, Any],
+        headers: Dict[str, str],
+    ) -> Iterator[_GeminiStreamChunk]:
+        """Generator that yields OpenAI-shaped streaming chunks."""
+        url = f"{CODE_ASSIST_ENDPOINT}/v1internal:streamGenerateContent?alt=sse"
+        stream_headers = dict(headers)
+        stream_headers["Accept"] = "text/event-stream"
+
+        def _generator() -> Iterator[_GeminiStreamChunk]:
+            try:
+                with self._http.stream("POST", url, json=wrapped, headers=stream_headers) as response:
+                    if response.status_code != 200:
+                        # Materialize error body for better diagnostics
+                        response.read()
+                        raise _gemini_http_error(response)
+                    tool_call_counter: List[int] = [0]
+                    for event in _iter_sse_events(response):
+                        for chunk in _translate_stream_event(event, model, tool_call_counter):
+                            yield chunk
+            except httpx.HTTPError as exc:
+                raise CodeAssistError(
+                    f"Streaming request failed: {exc}",
+                    code="code_assist_stream_error",
+                ) from exc
+
+        return _generator()
+
+
+def _gemini_http_error(response: httpx.Response) -> CodeAssistError:
+    """Translate an httpx response into a CodeAssistError with rich metadata.
+
+    Parses Google's error envelope (``{"error": {"code", "message", "status",
+    "details": [...]}}``) so the agent's error classifier can reason about
+    the failure — ``status_code`` enables the rate_limit / auth classification
+    paths, and ``response`` lets the main loop honor ``Retry-After`` just
+    like it does for OpenAI SDK exceptions.
+
+    Also lifts a few recognizable Google conditions into human-readable
+    messages so the user sees something better than a 500-char JSON dump:
+
+        MODEL_CAPACITY_EXHAUSTED → "Gemini model capacity exhausted for
+            <model>. This is a Google-side throttle..."
+        RESOURCE_EXHAUSTED w/o reason → quota-style message
+        404 → "Model <name> not found at cloudcode-pa..."
+    """
+    status = response.status_code
+
+    # Parse the body once, surviving any weird encodings.
+    body_text = ""
+    body_json: Dict[str, Any] = {}
+    try:
+        body_text = response.text
+    except Exception:
+        body_text = ""
+    if body_text:
+        try:
+            parsed = json.loads(body_text)
+            if isinstance(parsed, dict):
+                body_json = parsed
+        except (ValueError, TypeError):
+            body_json = {}
+
+    # Dig into Google's error envelope.  Shape is:
+    #   {"error": {"code": 429, "message": "...", "status": "RESOURCE_EXHAUSTED",
+    #              "details": [{"@type": ".../ErrorInfo", "reason": "MODEL_CAPACITY_EXHAUSTED",
+    #                           "metadata": {...}},
+    #                          {"@type": ".../RetryInfo", "retryDelay": "30s"}]}}
+    err_obj = body_json.get("error") if isinstance(body_json, dict) else None
+    if not isinstance(err_obj, dict):
+        err_obj = {}
+    err_status = str(err_obj.get("status") or "").strip()
+    err_message = str(err_obj.get("message") or "").strip()
+    err_details_list = err_obj.get("details") if isinstance(err_obj.get("details"), list) else []
+
+    # Extract google.rpc.ErrorInfo reason + metadata.  There may be more
+    # than one ErrorInfo (rare), so we pick the first one with a reason.
+    error_reason = ""
+    error_metadata: Dict[str, Any] = {}
+    retry_delay_seconds: Optional[float] = None
+    for detail in err_details_list:
+        if not isinstance(detail, dict):
+            continue
+        type_url = str(detail.get("@type") or "")
+        if not error_reason and type_url.endswith("/google.rpc.ErrorInfo"):
+            reason = detail.get("reason")
+            if isinstance(reason, str) and reason:
+                error_reason = reason
+            md = detail.get("metadata")
+            if isinstance(md, dict):
+                error_metadata = md
+        elif retry_delay_seconds is None and type_url.endswith("/google.rpc.RetryInfo"):
+            # retryDelay is a google.protobuf.Duration string like "30s" or "1.5s".
+            delay_raw = detail.get("retryDelay")
+            if isinstance(delay_raw, str) and delay_raw.endswith("s"):
+                try:
+                    retry_delay_seconds = float(delay_raw[:-1])
+                except ValueError:
+                    pass
+            elif isinstance(delay_raw, (int, float)):
+                retry_delay_seconds = float(delay_raw)
+
+    # Fall back to the Retry-After header if the body didn't include RetryInfo.
+    if retry_delay_seconds is None:
+        try:
+            header_val = response.headers.get("Retry-After") or response.headers.get("retry-after")
+        except Exception:
+            header_val = None
+        if header_val:
+            try:
+                retry_delay_seconds = float(header_val)
+            except (TypeError, ValueError):
+                retry_delay_seconds = None
+
+    # Classify the error code.  ``code_assist_rate_limited`` stays the default
+    # for 429s; a more specific reason tag helps downstream callers (e.g. tests,
+    # logs) without changing the rate_limit classification path.
+    code = f"code_assist_http_{status}"
+    if status == 401:
+        code = "code_assist_unauthorized"
+    elif status == 429:
+        code = "code_assist_rate_limited"
+        if error_reason == "MODEL_CAPACITY_EXHAUSTED":
+            code = "code_assist_capacity_exhausted"
+
+    # Build a human-readable message.  Keep the status + a raw-body tail for
+    # debugging, but lead with a friendlier summary when we recognize the
+    # Google signal.
+    model_hint = ""
+    if isinstance(error_metadata, dict):
+        model_hint = str(error_metadata.get("model") or error_metadata.get("modelId") or "").strip()
+
+    if status == 429 and error_reason == "MODEL_CAPACITY_EXHAUSTED":
+        target = model_hint or "this Gemini model"
+        message = (
+            f"Gemini capacity exhausted for {target} (Google-side throttle, "
+            f"not a Hermes issue). Try a different Gemini model or set a "
+            f"fallback_providers entry to a non-Gemini provider."
+        )
+        if retry_delay_seconds is not None:
+            message += f" Google suggests retrying in {retry_delay_seconds:g}s."
+    elif status == 429 and err_status == "RESOURCE_EXHAUSTED":
+        message = (
+            f"Gemini quota exhausted ({err_message or 'RESOURCE_EXHAUSTED'}). "
+            f"Check /gquota for remaining daily requests."
+        )
+        if retry_delay_seconds is not None:
+            message += f" Retry suggested in {retry_delay_seconds:g}s."
+    elif status == 404:
+        # Google returns 404 when a model has been retired or renamed.
+        target = model_hint or (err_message or "model")
+        message = (
+            f"Code Assist 404: {target} is not available at "
+            f"cloudcode-pa.googleapis.com. It may have been renamed or "
+            f"retired. Check hermes_cli/models.py for the current list."
+        )
+    elif err_message:
+        # Generic fallback with the parsed message.
+        message = f"Code Assist HTTP {status} ({err_status or 'error'}): {err_message}"
+    else:
+        # Last-ditch fallback — raw body snippet.
+        message = f"Code Assist returned HTTP {status}: {body_text[:500]}"
+
+    return CodeAssistError(
+        message,
+        code=code,
+        status_code=status,
+        response=response,
+        retry_after=retry_delay_seconds,
+        details={
+            "status": err_status,
+            "reason": error_reason,
+            "metadata": error_metadata,
+            "message": err_message,
+        },
+    )

agent/gemini_native_adapter.py

@@ -0,0 +1,846 @@
+"""OpenAI-compatible facade over Google AI Studio's native Gemini API.
+
+Hermes keeps ``api_mode='chat_completions'`` for the ``gemini`` provider so the
+main agent loop can keep using its existing OpenAI-shaped message flow.
+This adapter is the transport shim that converts those OpenAI-style
+``messages[]`` / ``tools[]`` requests into Gemini's native
+``models/{model}:generateContent`` schema and converts the responses back.
+
+Why this exists
+---------------
+Google's OpenAI-compatible endpoint has been brittle for Hermes's multi-turn
+agent/tool loop (auth churn, tool-call replay quirks, thought-signature
+requirements).  The native Gemini API is the canonical path and avoids the
+OpenAI-compat layer entirely.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import base64
+import json
+import logging
+import time
+import uuid
+from types import SimpleNamespace
+from typing import Any, Dict, Iterator, List, Optional
+
+import httpx
+
+from agent.gemini_schema import sanitize_gemini_tool_parameters
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_GEMINI_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"
+
+
+def is_native_gemini_base_url(base_url: str) -> bool:
+    """Return True when the endpoint speaks Gemini's native REST API."""
+    normalized = str(base_url or "").strip().rstrip("/").lower()
+    if not normalized:
+        return False
+    if "generativelanguage.googleapis.com" not in normalized:
+        return False
+    return not normalized.endswith("/openai")
+
+
+class GeminiAPIError(Exception):
+    """Error shape compatible with Hermes retry/error classification."""
+
+    def __init__(
+        self,
+        message: str,
+        *,
+        code: str = "gemini_api_error",
+        status_code: Optional[int] = None,
+        response: Optional[httpx.Response] = None,
+        retry_after: Optional[float] = None,
+        details: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        super().__init__(message)
+        self.code = code
+        self.status_code = status_code
+        self.response = response
+        self.retry_after = retry_after
+        self.details = details or {}
+
+
+def _coerce_content_to_text(content: Any) -> str:
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        pieces: List[str] = []
+        for part in content:
+            if isinstance(part, str):
+                pieces.append(part)
+            elif isinstance(part, dict) and part.get("type") == "text":
+                text = part.get("text")
+                if isinstance(text, str):
+                    pieces.append(text)
+        return "\n".join(pieces)
+    return str(content)
+
+
+def _extract_multimodal_parts(content: Any) -> List[Dict[str, Any]]:
+    if not isinstance(content, list):
+        text = _coerce_content_to_text(content)
+        return [{"text": text}] if text else []
+
+    parts: List[Dict[str, Any]] = []
+    for item in content:
+        if isinstance(item, str):
+            parts.append({"text": item})
+            continue
+        if not isinstance(item, dict):
+            continue
+        ptype = item.get("type")
+        if ptype == "text":
+            text = item.get("text")
+            if isinstance(text, str) and text:
+                parts.append({"text": text})
+        elif ptype == "image_url":
+            url = ((item.get("image_url") or {}).get("url") or "")
+            if not isinstance(url, str) or not url.startswith("data:"):
+                continue
+            try:
+                header, encoded = url.split(",", 1)
+                mime = header.split(":", 1)[1].split(";", 1)[0]
+                raw = base64.b64decode(encoded)
+            except Exception:
+                continue
+            parts.append(
+                {
+                    "inlineData": {
+                        "mimeType": mime,
+                        "data": base64.b64encode(raw).decode("ascii"),
+                    }
+                }
+            )
+    return parts
+
+
+def _tool_call_extra_signature(tool_call: Dict[str, Any]) -> Optional[str]:
+    extra = tool_call.get("extra_content") or {}
+    if not isinstance(extra, dict):
+        return None
+    google = extra.get("google") or extra.get("thought_signature")
+    if isinstance(google, dict):
+        sig = google.get("thought_signature") or google.get("thoughtSignature")
+        return str(sig) if isinstance(sig, str) and sig else None
+    if isinstance(google, str) and google:
+        return google
+    return None
+
+
+def _translate_tool_call_to_gemini(tool_call: Dict[str, Any]) -> Dict[str, Any]:
+    fn = tool_call.get("function") or {}
+    args_raw = fn.get("arguments", "")
+    try:
+        args = json.loads(args_raw) if isinstance(args_raw, str) and args_raw else {}
+    except json.JSONDecodeError:
+        args = {"_raw": args_raw}
+    if not isinstance(args, dict):
+        args = {"_value": args}
+
+    part: Dict[str, Any] = {
+        "functionCall": {
+            "name": str(fn.get("name") or ""),
+            "args": args,
+        }
+    }
+    thought_signature = _tool_call_extra_signature(tool_call)
+    if thought_signature:
+        part["thoughtSignature"] = thought_signature
+    return part
+
+
+def _translate_tool_result_to_gemini(
+    message: Dict[str, Any],
+    tool_name_by_call_id: Optional[Dict[str, str]] = None,
+) -> Dict[str, Any]:
+    tool_name_by_call_id = tool_name_by_call_id or {}
+    tool_call_id = str(message.get("tool_call_id") or "")
+    name = str(
+        message.get("name")
+        or tool_name_by_call_id.get(tool_call_id)
+        or tool_call_id
+        or "tool"
+    )
+    content = _coerce_content_to_text(message.get("content"))
+    try:
+        parsed = json.loads(content) if content.strip().startswith(("{", "[")) else None
+    except json.JSONDecodeError:
+        parsed = None
+    response = parsed if isinstance(parsed, dict) else {"output": content}
+    return {
+        "functionResponse": {
+            "name": name,
+            "response": response,
+        }
+    }
+
+
+def _build_gemini_contents(messages: List[Dict[str, Any]]) -> tuple[List[Dict[str, Any]], Optional[Dict[str, Any]]]:
+    system_text_parts: List[str] = []
+    contents: List[Dict[str, Any]] = []
+    tool_name_by_call_id: Dict[str, str] = {}
+
+    for msg in messages:
+        if not isinstance(msg, dict):
+            continue
+        role = str(msg.get("role") or "user")
+
+        if role == "system":
+            system_text_parts.append(_coerce_content_to_text(msg.get("content")))
+            continue
+
+        if role in {"tool", "function"}:
+            contents.append(
+                {
+                    "role": "user",
+                    "parts": [
+                        _translate_tool_result_to_gemini(
+                            msg,
+                            tool_name_by_call_id=tool_name_by_call_id,
+                        )
+                    ],
+                }
+            )
+            continue
+
+        gemini_role = "model" if role == "assistant" else "user"
+        parts: List[Dict[str, Any]] = []
+
+        content_parts = _extract_multimodal_parts(msg.get("content"))
+        parts.extend(content_parts)
+
+        tool_calls = msg.get("tool_calls") or []
+        if isinstance(tool_calls, list):
+            for tool_call in tool_calls:
+                if isinstance(tool_call, dict):
+                    tool_call_id = str(tool_call.get("id") or tool_call.get("call_id") or "")
+                    tool_name = str(((tool_call.get("function") or {}).get("name") or ""))
+                    if tool_call_id and tool_name:
+                        tool_name_by_call_id[tool_call_id] = tool_name
+                    parts.append(_translate_tool_call_to_gemini(tool_call))
+
+        if parts:
+            contents.append({"role": gemini_role, "parts": parts})
+
+    system_instruction = None
+    joined_system = "\n".join(part for part in system_text_parts if part).strip()
+    if joined_system:
+        system_instruction = {"parts": [{"text": joined_system}]}
+    return contents, system_instruction
+
+
+def _translate_tools_to_gemini(tools: Any) -> List[Dict[str, Any]]:
+    if not isinstance(tools, list):
+        return []
+    declarations: List[Dict[str, Any]] = []
+    for tool in tools:
+        if not isinstance(tool, dict):
+            continue
+        fn = tool.get("function") or {}
+        if not isinstance(fn, dict):
+            continue
+        name = fn.get("name")
+        if not isinstance(name, str) or not name:
+            continue
+        decl: Dict[str, Any] = {"name": name}
+        description = fn.get("description")
+        if isinstance(description, str) and description:
+            decl["description"] = description
+        parameters = fn.get("parameters")
+        if isinstance(parameters, dict):
+            decl["parameters"] = sanitize_gemini_tool_parameters(parameters)
+        declarations.append(decl)
+    return [{"functionDeclarations": declarations}] if declarations else []
+
+
+def _translate_tool_choice_to_gemini(tool_choice: Any) -> Optional[Dict[str, Any]]:
+    if tool_choice is None:
+        return None
+    if isinstance(tool_choice, str):
+        if tool_choice == "auto":
+            return {"functionCallingConfig": {"mode": "AUTO"}}
+        if tool_choice == "required":
+            return {"functionCallingConfig": {"mode": "ANY"}}
+        if tool_choice == "none":
+            return {"functionCallingConfig": {"mode": "NONE"}}
+    if isinstance(tool_choice, dict):
+        fn = tool_choice.get("function") or {}
+        name = fn.get("name")
+        if isinstance(name, str) and name:
+            return {"functionCallingConfig": {"mode": "ANY", "allowedFunctionNames": [name]}}
+    return None
+
+
+def _normalize_thinking_config(config: Any) -> Optional[Dict[str, Any]]:
+    if not isinstance(config, dict) or not config:
+        return None
+    budget = config.get("thinkingBudget", config.get("thinking_budget"))
+    include = config.get("includeThoughts", config.get("include_thoughts"))
+    level = config.get("thinkingLevel", config.get("thinking_level"))
+    normalized: Dict[str, Any] = {}
+    if isinstance(budget, (int, float)):
+        normalized["thinkingBudget"] = int(budget)
+    if isinstance(include, bool):
+        normalized["includeThoughts"] = include
+    if isinstance(level, str) and level.strip():
+        normalized["thinkingLevel"] = level.strip().lower()
+    return normalized or None
+
+
+def build_gemini_request(
+    *,
+    messages: List[Dict[str, Any]],
+    tools: Any = None,
+    tool_choice: Any = None,
+    temperature: Optional[float] = None,
+    max_tokens: Optional[int] = None,
+    top_p: Optional[float] = None,
+    stop: Any = None,
+    thinking_config: Any = None,
+) -> Dict[str, Any]:
+    contents, system_instruction = _build_gemini_contents(messages)
+    request: Dict[str, Any] = {"contents": contents}
+    if system_instruction:
+        request["systemInstruction"] = system_instruction
+
+    gemini_tools = _translate_tools_to_gemini(tools)
+    if gemini_tools:
+        request["tools"] = gemini_tools
+
+    tool_config = _translate_tool_choice_to_gemini(tool_choice)
+    if tool_config:
+        request["toolConfig"] = tool_config
+
+    generation_config: Dict[str, Any] = {}
+    if temperature is not None:
+        generation_config["temperature"] = temperature
+    if max_tokens is not None:
+        generation_config["maxOutputTokens"] = max_tokens
+    if top_p is not None:
+        generation_config["topP"] = top_p
+    if stop:
+        generation_config["stopSequences"] = stop if isinstance(stop, list) else [str(stop)]
+    normalized_thinking = _normalize_thinking_config(thinking_config)
+    if normalized_thinking:
+        generation_config["thinkingConfig"] = normalized_thinking
+    if generation_config:
+        request["generationConfig"] = generation_config
+
+    return request
+
+
+def _map_gemini_finish_reason(reason: str) -> str:
+    mapping = {
+        "STOP": "stop",
+        "MAX_TOKENS": "length",
+        "SAFETY": "content_filter",
+        "RECITATION": "content_filter",
+        "OTHER": "stop",
+    }
+    return mapping.get(str(reason or "").upper(), "stop")
+
+
+def _tool_call_extra_from_part(part: Dict[str, Any]) -> Optional[Dict[str, Any]]:
+    sig = part.get("thoughtSignature")
+    if isinstance(sig, str) and sig:
+        return {"google": {"thought_signature": sig}}
+    return None
+
+
+def _empty_response(model: str) -> SimpleNamespace:
+    message = SimpleNamespace(
+        role="assistant",
+        content="",
+        tool_calls=None,
+        reasoning=None,
+        reasoning_content=None,
+        reasoning_details=None,
+    )
+    choice = SimpleNamespace(index=0, message=message, finish_reason="stop")
+    usage = SimpleNamespace(
+        prompt_tokens=0,
+        completion_tokens=0,
+        total_tokens=0,
+        prompt_tokens_details=SimpleNamespace(cached_tokens=0),
+    )
+    return SimpleNamespace(
+        id=f"chatcmpl-{uuid.uuid4().hex[:12]}",
+        object="chat.completion",
+        created=int(time.time()),
+        model=model,
+        choices=[choice],
+        usage=usage,
+    )
+
+
+def translate_gemini_response(resp: Dict[str, Any], model: str) -> SimpleNamespace:
+    candidates = resp.get("candidates") or []
+    if not isinstance(candidates, list) or not candidates:
+        return _empty_response(model)
+
+    cand = candidates[0] if isinstance(candidates[0], dict) else {}
+    content_obj = cand.get("content") if isinstance(cand, dict) else {}
+    parts = content_obj.get("parts") if isinstance(content_obj, dict) else []
+
+    text_pieces: List[str] = []
+    reasoning_pieces: List[str] = []
+    tool_calls: List[SimpleNamespace] = []
+
+    for index, part in enumerate(parts or []):
+        if not isinstance(part, dict):
+            continue
+        if part.get("thought") is True and isinstance(part.get("text"), str):
+            reasoning_pieces.append(part["text"])
+            continue
+        if isinstance(part.get("text"), str):
+            text_pieces.append(part["text"])
+            continue
+        fc = part.get("functionCall")
+        if isinstance(fc, dict) and fc.get("name"):
+            try:
+                args_str = json.dumps(fc.get("args") or {}, ensure_ascii=False)
+            except (TypeError, ValueError):
+                args_str = "{}"
+            tool_call = SimpleNamespace(
+                id=f"call_{uuid.uuid4().hex[:12]}",
+                type="function",
+                index=index,
+                function=SimpleNamespace(name=str(fc["name"]), arguments=args_str),
+            )
+            extra_content = _tool_call_extra_from_part(part)
+            if extra_content:
+                tool_call.extra_content = extra_content
+            tool_calls.append(tool_call)
+
+    finish_reason = "tool_calls" if tool_calls else _map_gemini_finish_reason(str(cand.get("finishReason") or ""))
+    usage_meta = resp.get("usageMetadata") or {}
+    usage = SimpleNamespace(
+        prompt_tokens=int(usage_meta.get("promptTokenCount") or 0),
+        completion_tokens=int(usage_meta.get("candidatesTokenCount") or 0),
+        total_tokens=int(usage_meta.get("totalTokenCount") or 0),
+        prompt_tokens_details=SimpleNamespace(
+            cached_tokens=int(usage_meta.get("cachedContentTokenCount") or 0),
+        ),
+    )
+    reasoning = "".join(reasoning_pieces) or None
+    message = SimpleNamespace(
+        role="assistant",
+        content="".join(text_pieces) if text_pieces else None,
+        tool_calls=tool_calls or None,
+        reasoning=reasoning,
+        reasoning_content=reasoning,
+        reasoning_details=None,
+    )
+    choice = SimpleNamespace(index=0, message=message, finish_reason=finish_reason)
+    return SimpleNamespace(
+        id=f"chatcmpl-{uuid.uuid4().hex[:12]}",
+        object="chat.completion",
+        created=int(time.time()),
+        model=model,
+        choices=[choice],
+        usage=usage,
+    )
+
+
+class _GeminiStreamChunk(SimpleNamespace):
+    pass
+
+
+def _make_stream_chunk(
+    *,
+    model: str,
+    content: str = "",
+    tool_call_delta: Optional[Dict[str, Any]] = None,
+    finish_reason: Optional[str] = None,
+    reasoning: str = "",
+) -> _GeminiStreamChunk:
+    delta_kwargs: Dict[str, Any] = {
+        "role": "assistant",
+        "content": None,
+        "tool_calls": None,
+        "reasoning": None,
+        "reasoning_content": None,
+    }
+    if content:
+        delta_kwargs["content"] = content
+    if tool_call_delta is not None:
+        tool_delta = SimpleNamespace(
+            index=tool_call_delta.get("index", 0),
+            id=tool_call_delta.get("id") or f"call_{uuid.uuid4().hex[:12]}",
+            type="function",
+            function=SimpleNamespace(
+                name=tool_call_delta.get("name") or "",
+                arguments=tool_call_delta.get("arguments") or "",
+            ),
+        )
+        extra_content = tool_call_delta.get("extra_content")
+        if isinstance(extra_content, dict):
+            tool_delta.extra_content = extra_content
+        delta_kwargs["tool_calls"] = [tool_delta]
+    if reasoning:
+        delta_kwargs["reasoning"] = reasoning
+        delta_kwargs["reasoning_content"] = reasoning
+    delta = SimpleNamespace(**delta_kwargs)
+    choice = SimpleNamespace(index=0, delta=delta, finish_reason=finish_reason)
+    return _GeminiStreamChunk(
+        id=f"chatcmpl-{uuid.uuid4().hex[:12]}",
+        object="chat.completion.chunk",
+        created=int(time.time()),
+        model=model,
+        choices=[choice],
+        usage=None,
+    )
+
+
+def _iter_sse_events(response: httpx.Response) -> Iterator[Dict[str, Any]]:
+    buffer = ""
+    for chunk in response.iter_text():
+        if not chunk:
+            continue
+        buffer += chunk
+        while "\n" in buffer:
+            line, buffer = buffer.split("\n", 1)
+            line = line.rstrip("\r")
+            if not line:
+                continue
+            if not line.startswith("data: "):
+                continue
+            data = line[6:]
+            if data == "[DONE]":
+                return
+            try:
+                payload = json.loads(data)
+            except json.JSONDecodeError:
+                logger.debug("Non-JSON Gemini SSE line: %s", data[:200])
+                continue
+            if isinstance(payload, dict):
+                yield payload
+
+
+def translate_stream_event(event: Dict[str, Any], model: str, tool_call_indices: Dict[str, Dict[str, Any]]) -> List[_GeminiStreamChunk]:
+    candidates = event.get("candidates") or []
+    if not candidates:
+        return []
+    cand = candidates[0] if isinstance(candidates[0], dict) else {}
+    parts = ((cand.get("content") or {}).get("parts") or []) if isinstance(cand, dict) else []
+    chunks: List[_GeminiStreamChunk] = []
+
+    for part_index, part in enumerate(parts):
+        if not isinstance(part, dict):
+            continue
+        if part.get("thought") is True and isinstance(part.get("text"), str):
+            chunks.append(_make_stream_chunk(model=model, reasoning=part["text"]))
+            continue
+        if isinstance(part.get("text"), str) and part["text"]:
+            chunks.append(_make_stream_chunk(model=model, content=part["text"]))
+        fc = part.get("functionCall")
+        if isinstance(fc, dict) and fc.get("name"):
+            name = str(fc["name"])
+            try:
+                args_str = json.dumps(fc.get("args") or {}, ensure_ascii=False, sort_keys=True)
+            except (TypeError, ValueError):
+                args_str = "{}"
+            thought_signature = part.get("thoughtSignature") if isinstance(part.get("thoughtSignature"), str) else ""
+            call_key = json.dumps(
+                {
+                    "part_index": part_index,
+                    "name": name,
+                    "thought_signature": thought_signature,
+                },
+                sort_keys=True,
+            )
+            slot = tool_call_indices.get(call_key)
+            if slot is None:
+                slot = {
+                    "index": len(tool_call_indices),
+                    "id": f"call_{uuid.uuid4().hex[:12]}",
+                    "last_arguments": "",
+                }
+                tool_call_indices[call_key] = slot
+            emitted_arguments = args_str
+            last_arguments = str(slot.get("last_arguments") or "")
+            if last_arguments:
+                if args_str == last_arguments:
+                    emitted_arguments = ""
+                elif args_str.startswith(last_arguments):
+                    emitted_arguments = args_str[len(last_arguments):]
+            slot["last_arguments"] = args_str
+            chunks.append(
+                _make_stream_chunk(
+                    model=model,
+                    tool_call_delta={
+                        "index": slot["index"],
+                        "id": slot["id"],
+                        "name": name,
+                        "arguments": emitted_arguments,
+                        "extra_content": _tool_call_extra_from_part(part),
+                    },
+                )
+            )
+
+    finish_reason_raw = str(cand.get("finishReason") or "")
+    if finish_reason_raw:
+        mapped = "tool_calls" if tool_call_indices else _map_gemini_finish_reason(finish_reason_raw)
+        chunks.append(_make_stream_chunk(model=model, finish_reason=mapped))
+    return chunks
+
+
+def gemini_http_error(response: httpx.Response) -> GeminiAPIError:
+    status = response.status_code
+    body_text = ""
+    body_json: Dict[str, Any] = {}
+    try:
+        body_text = response.text
+    except Exception:
+        body_text = ""
+    if body_text:
+        try:
+            parsed = json.loads(body_text)
+            if isinstance(parsed, dict):
+                body_json = parsed
+        except (ValueError, TypeError):
+            body_json = {}
+
+    err_obj = body_json.get("error") if isinstance(body_json, dict) else None
+    if not isinstance(err_obj, dict):
+        err_obj = {}
+    err_status = str(err_obj.get("status") or "").strip()
+    err_message = str(err_obj.get("message") or "").strip()
+    details_list = err_obj.get("details") if isinstance(err_obj.get("details"), list) else []
+
+    reason = ""
+    retry_after: Optional[float] = None
+    metadata: Dict[str, Any] = {}
+    for detail in details_list:
+        if not isinstance(detail, dict):
+            continue
+        type_url = str(detail.get("@type") or "")
+        if not reason and type_url.endswith("/google.rpc.ErrorInfo"):
+            reason_value = detail.get("reason")
+            if isinstance(reason_value, str):
+                reason = reason_value
+            md = detail.get("metadata")
+            if isinstance(md, dict):
+                metadata = md
+    header_retry = response.headers.get("Retry-After") or response.headers.get("retry-after")
+    if header_retry:
+        try:
+            retry_after = float(header_retry)
+        except (TypeError, ValueError):
+            retry_after = None
+
+    code = f"gemini_http_{status}"
+    if status == 401:
+        code = "gemini_unauthorized"
+    elif status == 429:
+        code = "gemini_rate_limited"
+    elif status == 404:
+        code = "gemini_model_not_found"
+
+    if err_message:
+        message = f"Gemini HTTP {status} ({err_status or 'error'}): {err_message}"
+    else:
+        message = f"Gemini returned HTTP {status}: {body_text[:500]}"
+
+    return GeminiAPIError(
+        message,
+        code=code,
+        status_code=status,
+        response=response,
+        retry_after=retry_after,
+        details={
+            "status": err_status,
+            "reason": reason,
+            "metadata": metadata,
+            "message": err_message,
+        },
+    )
+
+
+class _GeminiChatCompletions:
+    def __init__(self, client: "GeminiNativeClient"):
+        self._client = client
+
+    def create(self, **kwargs: Any) -> Any:
+        return self._client._create_chat_completion(**kwargs)
+
+
+class _AsyncGeminiChatCompletions:
+    def __init__(self, client: "AsyncGeminiNativeClient"):
+        self._client = client
+
+    async def create(self, **kwargs: Any) -> Any:
+        return await self._client._create_chat_completion(**kwargs)
+
+
+class _GeminiChatNamespace:
+    def __init__(self, client: "GeminiNativeClient"):
+        self.completions = _GeminiChatCompletions(client)
+
+
+class _AsyncGeminiChatNamespace:
+    def __init__(self, client: "AsyncGeminiNativeClient"):
+        self.completions = _AsyncGeminiChatCompletions(client)
+
+
+class GeminiNativeClient:
+    """Minimal OpenAI-SDK-compatible facade over Gemini's native REST API."""
+
+    def __init__(
+        self,
+        *,
+        api_key: str,
+        base_url: Optional[str] = None,
+        default_headers: Optional[Dict[str, str]] = None,
+        timeout: Any = None,
+        http_client: Optional[httpx.Client] = None,
+        **_: Any,
+    ) -> None:
+        self.api_key = api_key
+        normalized_base = (base_url or DEFAULT_GEMINI_BASE_URL).rstrip("/")
+        if normalized_base.endswith("/openai"):
+            normalized_base = normalized_base[: -len("/openai")]
+        self.base_url = normalized_base
+        self._default_headers = dict(default_headers or {})
+        self.chat = _GeminiChatNamespace(self)
+        self.is_closed = False
+        self._http = http_client or httpx.Client(
+            timeout=timeout or httpx.Timeout(connect=15.0, read=600.0, write=30.0, pool=30.0)
+        )
+
+    def close(self) -> None:
+        self.is_closed = True
+        try:
+            self._http.close()
+        except Exception:
+            pass
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        self.close()
+
+    def _headers(self) -> Dict[str, str]:
+        headers = {
+            "Content-Type": "application/json",
+            "Accept": "application/json",
+            "x-goog-api-key": self.api_key,
+            "User-Agent": "hermes-agent (gemini-native)",
+        }
+        headers.update(self._default_headers)
+        return headers
+
+    @staticmethod
+    def _advance_stream_iterator(iterator: Iterator[_GeminiStreamChunk]) -> tuple[bool, Optional[_GeminiStreamChunk]]:
+        try:
+            return False, next(iterator)
+        except StopIteration:
+            return True, None
+
+    def _create_chat_completion(
+        self,
+        *,
+        model: str = "gemini-2.5-flash",
+        messages: Optional[List[Dict[str, Any]]] = None,
+        stream: bool = False,
+        tools: Any = None,
+        tool_choice: Any = None,
+        temperature: Optional[float] = None,
+        max_tokens: Optional[int] = None,
+        top_p: Optional[float] = None,
+        stop: Any = None,
+        extra_body: Optional[Dict[str, Any]] = None,
+        timeout: Any = None,
+        **_: Any,
+    ) -> Any:
+        thinking_config = None
+        if isinstance(extra_body, dict):
+            thinking_config = extra_body.get("thinking_config") or extra_body.get("thinkingConfig")
+
+        request = build_gemini_request(
+            messages=messages or [],
+            tools=tools,
+            tool_choice=tool_choice,
+            temperature=temperature,
+            max_tokens=max_tokens,
+            top_p=top_p,
+            stop=stop,
+            thinking_config=thinking_config,
+        )
+
+        if stream:
+            return self._stream_completion(model=model, request=request, timeout=timeout)
+
+        url = f"{self.base_url}/models/{model}:generateContent"
+        response = self._http.post(url, json=request, headers=self._headers(), timeout=timeout)
+        if response.status_code != 200:
+            raise gemini_http_error(response)
+        try:
+            payload = response.json()
+        except ValueError as exc:
+            raise GeminiAPIError(
+                f"Invalid JSON from Gemini native API: {exc}",
+                code="gemini_invalid_json",
+                status_code=response.status_code,
+                response=response,
+            ) from exc
+        return translate_gemini_response(payload, model=model)
+
+    def _stream_completion(self, *, model: str, request: Dict[str, Any], timeout: Any = None) -> Iterator[_GeminiStreamChunk]:
+        url = f"{self.base_url}/models/{model}:streamGenerateContent?alt=sse"
+        stream_headers = dict(self._headers())
+        stream_headers["Accept"] = "text/event-stream"
+
+        def _generator() -> Iterator[_GeminiStreamChunk]:
+            try:
+                with self._http.stream("POST", url, json=request, headers=stream_headers, timeout=timeout) as response:
+                    if response.status_code != 200:
+                        response.read()
+                        raise gemini_http_error(response)
+                    tool_call_indices: Dict[str, Dict[str, Any]] = {}
+                    for event in _iter_sse_events(response):
+                        for chunk in translate_stream_event(event, model, tool_call_indices):
+                            yield chunk
+            except httpx.HTTPError as exc:
+                raise GeminiAPIError(
+                    f"Gemini streaming request failed: {exc}",
+                    code="gemini_stream_error",
+                ) from exc
+
+        return _generator()
+
+
+class AsyncGeminiNativeClient:
+    """Async wrapper used by auxiliary_client for native Gemini calls."""
+
+    def __init__(self, sync_client: GeminiNativeClient):
+        self._sync = sync_client
+        self.api_key = sync_client.api_key
+        self.base_url = sync_client.base_url
+        self.chat = _AsyncGeminiChatNamespace(self)
+
+    async def _create_chat_completion(self, **kwargs: Any) -> Any:
+        stream = bool(kwargs.get("stream"))
+        result = await asyncio.to_thread(self._sync.chat.completions.create, **kwargs)
+        if not stream:
+            return result
+
+        async def _async_stream() -> Any:
+            while True:
+                done, chunk = await asyncio.to_thread(self._sync._advance_stream_iterator, result)
+                if done:
+                    break
+                yield chunk
+
+        return _async_stream()
+
+    async def close(self) -> None:
+        await asyncio.to_thread(self._sync.close)

agent/gemini_schema.py

@@ -0,0 +1,85 @@
+"""Helpers for translating OpenAI-style tool schemas to Gemini's schema subset."""
+
+from __future__ import annotations
+
+from typing import Any, Dict, List
+
+# Gemini's ``FunctionDeclaration.parameters`` field accepts the ``Schema``
+# object, which is only a subset of OpenAPI 3.0 / JSON Schema.  Strip fields
+# outside that subset before sending Hermes tool schemas to Google.
+_GEMINI_SCHEMA_ALLOWED_KEYS = {
+    "type",
+    "format",
+    "title",
+    "description",
+    "nullable",
+    "enum",
+    "maxItems",
+    "minItems",
+    "properties",
+    "required",
+    "minProperties",
+    "maxProperties",
+    "minLength",
+    "maxLength",
+    "pattern",
+    "example",
+    "anyOf",
+    "propertyOrdering",
+    "default",
+    "items",
+    "minimum",
+    "maximum",
+}
+
+
+def sanitize_gemini_schema(schema: Any) -> Dict[str, Any]:
+    """Return a Gemini-compatible copy of a tool parameter schema.
+
+    Hermes tool schemas are OpenAI-flavored JSON Schema and may contain keys
+    such as ``$schema`` or ``additionalProperties`` that Google's Gemini
+    ``Schema`` object rejects.  This helper preserves the documented Gemini
+    subset and recursively sanitizes nested ``properties`` / ``items`` /
+    ``anyOf`` definitions.
+    """
+
+    if not isinstance(schema, dict):
+        return {}
+
+    cleaned: Dict[str, Any] = {}
+    for key, value in schema.items():
+        if key not in _GEMINI_SCHEMA_ALLOWED_KEYS:
+            continue
+        if key == "properties":
+            if not isinstance(value, dict):
+                continue
+            props: Dict[str, Any] = {}
+            for prop_name, prop_schema in value.items():
+                if not isinstance(prop_name, str):
+                    continue
+                props[prop_name] = sanitize_gemini_schema(prop_schema)
+            cleaned[key] = props
+            continue
+        if key == "items":
+            cleaned[key] = sanitize_gemini_schema(value)
+            continue
+        if key == "anyOf":
+            if not isinstance(value, list):
+                continue
+            cleaned[key] = [
+                sanitize_gemini_schema(item)
+                for item in value
+                if isinstance(item, dict)
+            ]
+            continue
+        cleaned[key] = value
+    return cleaned
+
+
+def sanitize_gemini_tool_parameters(parameters: Any) -> Dict[str, Any]:
+    """Normalize tool parameters to a valid Gemini object schema."""
+
+    cleaned = sanitize_gemini_schema(parameters)
+    if not cleaned:
+        return {"type": "object", "properties": {}}
+    return cleaned

agent/google_code_assist.py

@@ -0,0 +1,453 @@
+"""Google Code Assist API client — project discovery, onboarding, quota.
+
+The Code Assist API powers Google's official gemini-cli. It sits at
+``cloudcode-pa.googleapis.com`` and provides:
+
+- Free tier access (generous daily quota) for personal Google accounts
+- Paid tier access via GCP projects with billing / Workspace / Standard / Enterprise
+
+This module handles the control-plane dance needed before inference:
+
+1. ``load_code_assist()`` — probe the user's account to learn what tier they're on
+   and whether a ``cloudaicompanionProject`` is already assigned.
+2. ``onboard_user()`` — if the user hasn't been onboarded yet (new account, fresh
+   free tier, etc.), call this with the chosen tier + project id. Supports LRO
+   polling for slow provisioning.
+3. ``retrieve_user_quota()`` — fetch the ``buckets[]`` array showing remaining
+   quota per model, used by the ``/gquota`` slash command.
+
+VPC-SC handling: enterprise accounts under a VPC Service Controls perimeter
+will get ``SECURITY_POLICY_VIOLATED`` on ``load_code_assist``. We catch this
+and force the account to ``standard-tier`` so the call chain still succeeds.
+
+Derived from opencode-gemini-auth (MIT) and clawdbot/extensions/google. The
+request/response shapes are specific to Google's internal Code Assist API,
+documented nowhere public — we copy them from the reference implementations.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import time
+import urllib.error
+import urllib.parse
+import urllib.request
+import uuid
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional
+
+logger = logging.getLogger(__name__)
+
+
+# =============================================================================
+# Constants
+# =============================================================================
+
+CODE_ASSIST_ENDPOINT = "https://cloudcode-pa.googleapis.com"
+
+# Fallback endpoints tried when prod returns an error during project discovery
+FALLBACK_ENDPOINTS = [
+    "https://daily-cloudcode-pa.sandbox.googleapis.com",
+    "https://autopush-cloudcode-pa.sandbox.googleapis.com",
+]
+
+# Tier identifiers that Google's API uses
+FREE_TIER_ID = "free-tier"
+LEGACY_TIER_ID = "legacy-tier"
+STANDARD_TIER_ID = "standard-tier"
+
+# Default HTTP headers matching gemini-cli's fingerprint.
+# Google may reject unrecognized User-Agents on these internal endpoints.
+_GEMINI_CLI_USER_AGENT = "google-api-nodejs-client/9.15.1 (gzip)"
+_X_GOOG_API_CLIENT = "gl-node/24.0.0"
+_DEFAULT_REQUEST_TIMEOUT = 30.0
+_ONBOARDING_POLL_ATTEMPTS = 12
+_ONBOARDING_POLL_INTERVAL_SECONDS = 5.0
+
+
+class CodeAssistError(RuntimeError):
+    """Exception raised by the Code Assist (``cloudcode-pa``) integration.
+
+    Carries HTTP status / response / retry-after metadata so the agent's
+    ``error_classifier._extract_status_code`` and the main loop's Retry-After
+    handling (which walks ``error.response.headers``) pick up the right
+    signals.  Without these, 429s from the OAuth path look like opaque
+    ``RuntimeError`` and skip the rate-limit path.
+    """
+
+    def __init__(
+        self,
+        message: str,
+        *,
+        code: str = "code_assist_error",
+        status_code: Optional[int] = None,
+        response: Any = None,
+        retry_after: Optional[float] = None,
+        details: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        super().__init__(message)
+        self.code = code
+        # ``status_code`` is picked up by ``agent.error_classifier._extract_status_code``
+        # so a 429 from Code Assist classifies as FailoverReason.rate_limit and
+        # triggers the main loop's fallback_providers chain the same way SDK
+        # errors do.
+        self.status_code = status_code
+        # ``response`` is the underlying ``httpx.Response`` (or a shim with a
+        # ``.headers`` mapping and ``.json()`` method).  The main loop reads
+        # ``error.response.headers["Retry-After"]`` to honor Google's retry
+        # hints when the backend throttles us.
+        self.response = response
+        # Parsed ``Retry-After`` seconds (kept separately for convenience —
+        # Google returns retry hints in both the header and the error body's
+        # ``google.rpc.RetryInfo`` details, and we pick whichever we found).
+        self.retry_after = retry_after
+        # Parsed structured error details from the Google error envelope
+        # (e.g. ``{"reason": "MODEL_CAPACITY_EXHAUSTED", "status": "RESOURCE_EXHAUSTED"}``).
+        # Useful for logging and for tests that want to assert on specifics.
+        self.details = details or {}
+
+
+class ProjectIdRequiredError(CodeAssistError):
+    def __init__(self, message: str = "GCP project id required for this tier") -> None:
+        super().__init__(message, code="code_assist_project_id_required")
+
+
+# =============================================================================
+# HTTP primitive (auth via Bearer token passed per-call)
+# =============================================================================
+
+def _build_headers(access_token: str, *, user_agent_model: str = "") -> Dict[str, str]:
+    ua = _GEMINI_CLI_USER_AGENT
+    if user_agent_model:
+        ua = f"{ua} model/{user_agent_model}"
+    return {
+        "Content-Type": "application/json",
+        "Accept": "application/json",
+        "Authorization": f"Bearer {access_token}",
+        "User-Agent": ua,
+        "X-Goog-Api-Client": _X_GOOG_API_CLIENT,
+        "x-activity-request-id": str(uuid.uuid4()),
+    }
+
+
+def _client_metadata() -> Dict[str, str]:
+    """Match Google's gemini-cli exactly — unrecognized metadata may be rejected."""
+    return {
+        "ideType": "IDE_UNSPECIFIED",
+        "platform": "PLATFORM_UNSPECIFIED",
+        "pluginType": "GEMINI",
+    }
+
+
+def _post_json(
+    url: str,
+    body: Dict[str, Any],
+    access_token: str,
+    *,
+    timeout: float = _DEFAULT_REQUEST_TIMEOUT,
+    user_agent_model: str = "",
+) -> Dict[str, Any]:
+    data = json.dumps(body).encode("utf-8")
+    request = urllib.request.Request(
+        url, data=data, method="POST",
+        headers=_build_headers(access_token, user_agent_model=user_agent_model),
+    )
+    try:
+        with urllib.request.urlopen(request, timeout=timeout) as response:
+            raw = response.read().decode("utf-8", errors="replace")
+            return json.loads(raw) if raw else {}
+    except urllib.error.HTTPError as exc:
+        detail = ""
+        try:
+            detail = exc.read().decode("utf-8", errors="replace")
+        except Exception:
+            pass
+        # Special case: VPC-SC violation should be distinguishable
+        if _is_vpc_sc_violation(detail):
+            raise CodeAssistError(
+                f"VPC-SC policy violation: {detail}",
+                code="code_assist_vpc_sc",
+            ) from exc
+        raise CodeAssistError(
+            f"Code Assist HTTP {exc.code}: {detail or exc.reason}",
+            code=f"code_assist_http_{exc.code}",
+        ) from exc
+    except urllib.error.URLError as exc:
+        raise CodeAssistError(
+            f"Code Assist request failed: {exc}",
+            code="code_assist_network_error",
+        ) from exc
+
+
+def _is_vpc_sc_violation(body: str) -> bool:
+    """Detect a VPC Service Controls violation from a response body."""
+    if not body:
+        return False
+    try:
+        parsed = json.loads(body)
+    except (json.JSONDecodeError, ValueError):
+        return "SECURITY_POLICY_VIOLATED" in body
+    # Walk the nested error structure Google uses
+    error = parsed.get("error") if isinstance(parsed, dict) else None
+    if not isinstance(error, dict):
+        return False
+    details = error.get("details") or []
+    if isinstance(details, list):
+        for item in details:
+            if isinstance(item, dict):
+                reason = item.get("reason") or ""
+                if reason == "SECURITY_POLICY_VIOLATED":
+                    return True
+    msg = str(error.get("message", ""))
+    return "SECURITY_POLICY_VIOLATED" in msg
+
+
+# =============================================================================
+# load_code_assist — discovers current tier + assigned project
+# =============================================================================
+
+@dataclass
+class CodeAssistProjectInfo:
+    """Result from ``load_code_assist``."""
+    current_tier_id: str = ""
+    cloudaicompanion_project: str = ""   # Google-managed project (free tier)
+    allowed_tiers: List[str] = field(default_factory=list)
+    raw: Dict[str, Any] = field(default_factory=dict)
+
+
+def load_code_assist(
+    access_token: str,
+    *,
+    project_id: str = "",
+    user_agent_model: str = "",
+) -> CodeAssistProjectInfo:
+    """Call ``POST /v1internal:loadCodeAssist`` with prod → sandbox fallback.
+
+    Returns whatever tier + project info Google reports. On VPC-SC violations,
+    returns a synthetic ``standard-tier`` result so the chain can continue.
+    """
+    body: Dict[str, Any] = {
+        "metadata": {
+            "duetProject": project_id,
+            **_client_metadata(),
+        },
+    }
+    if project_id:
+        body["cloudaicompanionProject"] = project_id
+
+    endpoints = [CODE_ASSIST_ENDPOINT] + FALLBACK_ENDPOINTS
+    last_err: Optional[Exception] = None
+    for endpoint in endpoints:
+        url = f"{endpoint}/v1internal:loadCodeAssist"
+        try:
+            resp = _post_json(url, body, access_token, user_agent_model=user_agent_model)
+            return _parse_load_response(resp)
+        except CodeAssistError as exc:
+            if exc.code == "code_assist_vpc_sc":
+                logger.info("VPC-SC violation on %s — defaulting to standard-tier", endpoint)
+                return CodeAssistProjectInfo(
+                    current_tier_id=STANDARD_TIER_ID,
+                    cloudaicompanion_project=project_id,
+                )
+            last_err = exc
+            logger.warning("loadCodeAssist failed on %s: %s", endpoint, exc)
+            continue
+    if last_err:
+        raise last_err
+    return CodeAssistProjectInfo()
+
+
+def _parse_load_response(resp: Dict[str, Any]) -> CodeAssistProjectInfo:
+    current_tier = resp.get("currentTier") or {}
+    tier_id = str(current_tier.get("id") or "") if isinstance(current_tier, dict) else ""
+    project = str(resp.get("cloudaicompanionProject") or "")
+    allowed = resp.get("allowedTiers") or []
+    allowed_ids: List[str] = []
+    if isinstance(allowed, list):
+        for t in allowed:
+            if isinstance(t, dict):
+                tid = str(t.get("id") or "")
+                if tid:
+                    allowed_ids.append(tid)
+    return CodeAssistProjectInfo(
+        current_tier_id=tier_id,
+        cloudaicompanion_project=project,
+        allowed_tiers=allowed_ids,
+        raw=resp,
+    )
+
+
+# =============================================================================
+# onboard_user — provisions a new user on a tier (with LRO polling)
+# =============================================================================
+
+def onboard_user(
+    access_token: str,
+    *,
+    tier_id: str,
+    project_id: str = "",
+    user_agent_model: str = "",
+) -> Dict[str, Any]:
+    """Call ``POST /v1internal:onboardUser`` to provision the user.
+
+    For paid tiers, ``project_id`` is REQUIRED (raises ProjectIdRequiredError).
+    For free tiers, ``project_id`` is optional — Google will assign one.
+
+    Returns the final operation response. Polls ``/v1internal/<name>`` for up
+    to ``_ONBOARDING_POLL_ATTEMPTS`` × ``_ONBOARDING_POLL_INTERVAL_SECONDS``
+    (default: 12 × 5s = 1 min).
+    """
+    if tier_id != FREE_TIER_ID and tier_id != LEGACY_TIER_ID and not project_id:
+        raise ProjectIdRequiredError(
+            f"Tier {tier_id!r} requires a GCP project id. "
+            "Set HERMES_GEMINI_PROJECT_ID or GOOGLE_CLOUD_PROJECT."
+        )
+
+    body: Dict[str, Any] = {
+        "tierId": tier_id,
+        "metadata": _client_metadata(),
+    }
+    if project_id:
+        body["cloudaicompanionProject"] = project_id
+
+    endpoint = CODE_ASSIST_ENDPOINT
+    url = f"{endpoint}/v1internal:onboardUser"
+    resp = _post_json(url, body, access_token, user_agent_model=user_agent_model)
+
+    # Poll if LRO (long-running operation)
+    if not resp.get("done"):
+        op_name = resp.get("name", "")
+        if not op_name:
+            return resp
+        for attempt in range(_ONBOARDING_POLL_ATTEMPTS):
+            time.sleep(_ONBOARDING_POLL_INTERVAL_SECONDS)
+            poll_url = f"{endpoint}/v1internal/{op_name}"
+            try:
+                poll_resp = _post_json(poll_url, {}, access_token, user_agent_model=user_agent_model)
+            except CodeAssistError as exc:
+                logger.warning("Onboarding poll attempt %d failed: %s", attempt + 1, exc)
+                continue
+            if poll_resp.get("done"):
+                return poll_resp
+        logger.warning("Onboarding did not complete within %d attempts", _ONBOARDING_POLL_ATTEMPTS)
+    return resp
+
+
+# =============================================================================
+# retrieve_user_quota — for /gquota
+# =============================================================================
+
+@dataclass
+class QuotaBucket:
+    model_id: str
+    token_type: str = ""
+    remaining_fraction: float = 0.0
+    reset_time_iso: str = ""
+    raw: Dict[str, Any] = field(default_factory=dict)
+
+
+def retrieve_user_quota(
+    access_token: str,
+    *,
+    project_id: str = "",
+    user_agent_model: str = "",
+) -> List[QuotaBucket]:
+    """Call ``POST /v1internal:retrieveUserQuota`` and parse ``buckets[]``."""
+    body: Dict[str, Any] = {}
+    if project_id:
+        body["project"] = project_id
+    url = f"{CODE_ASSIST_ENDPOINT}/v1internal:retrieveUserQuota"
+    resp = _post_json(url, body, access_token, user_agent_model=user_agent_model)
+    raw_buckets = resp.get("buckets") or []
+    buckets: List[QuotaBucket] = []
+    if not isinstance(raw_buckets, list):
+        return buckets
+    for b in raw_buckets:
+        if not isinstance(b, dict):
+            continue
+        buckets.append(QuotaBucket(
+            model_id=str(b.get("modelId") or ""),
+            token_type=str(b.get("tokenType") or ""),
+            remaining_fraction=float(b.get("remainingFraction") or 0.0),
+            reset_time_iso=str(b.get("resetTime") or ""),
+            raw=b,
+        ))
+    return buckets
+
+
+# =============================================================================
+# Project context resolution
+# =============================================================================
+
+@dataclass
+class ProjectContext:
+    """Resolved state for a given OAuth session."""
+    project_id: str = ""           # effective project id sent on requests
+    managed_project_id: str = ""   # Google-assigned project (free tier)
+    tier_id: str = ""
+    source: str = ""               # "env", "config", "discovered", "onboarded"
+
+
+def resolve_project_context(
+    access_token: str,
+    *,
+    configured_project_id: str = "",
+    env_project_id: str = "",
+    user_agent_model: str = "",
+) -> ProjectContext:
+    """Figure out what project id + tier to use for requests.
+
+    Priority:
+      1. If configured_project_id or env_project_id is set, use that directly
+         and short-circuit (no discovery needed).
+      2. Otherwise call loadCodeAssist to see what Google says.
+      3. If no tier assigned yet, onboard the user (free tier default).
+    """
+    # Short-circuit: caller provided a project id
+    if configured_project_id:
+        return ProjectContext(
+            project_id=configured_project_id,
+            tier_id=STANDARD_TIER_ID,  # assume paid since they specified one
+            source="config",
+        )
+    if env_project_id:
+        return ProjectContext(
+            project_id=env_project_id,
+            tier_id=STANDARD_TIER_ID,
+            source="env",
+        )
+
+    # Discover via loadCodeAssist
+    info = load_code_assist(access_token, user_agent_model=user_agent_model)
+
+    effective_project = info.cloudaicompanion_project
+    tier = info.current_tier_id
+
+    if not tier:
+        # User hasn't been onboarded — provision them on free tier
+        onboard_resp = onboard_user(
+            access_token,
+            tier_id=FREE_TIER_ID,
+            project_id="",
+            user_agent_model=user_agent_model,
+        )
+        # Re-parse from the onboard response
+        response_body = onboard_resp.get("response") or {}
+        if isinstance(response_body, dict):
+            effective_project = (
+                effective_project
+                or str(response_body.get("cloudaicompanionProject") or "")
+            )
+        tier = FREE_TIER_ID
+        source = "onboarded"
+    else:
+        source = "discovered"
+
+    return ProjectContext(
+        project_id=effective_project,
+        managed_project_id=effective_project if tier == FREE_TIER_ID else "",
+        tier_id=tier,
+        source=source,
+    )

agent/insights.py

@@ -634,13 +634,7 @@ 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"${o['estimated_cost']:.2f}"
-        if o.get("models_without_pricing"):
-            cost_str += " *"
-        lines.append(f"  Total tokens:      {o['total_tokens']:<12,}  Est. cost:       {cost_str}")
+        lines.append(f"  Total tokens:      {o['total_tokens']:,}")
         if o["total_hours"] > 0:
             lines.append(f"  Active time:       ~{_format_duration(o['total_hours'] * 3600):<11}  Avg session:     ~{_format_duration(o['avg_session_duration'])}")
         lines.append(f"  Avg msgs/session:  {o['avg_messages_per_session']:.1f}")
@@ -650,16 +644,10 @@ class InsightsEngine:
         if report["models"]:
             lines.append("  🤖 Models Used")
             lines.append("  " + "─" * 56)
-            lines.append(f"  {'Model':<30} {'Sessions':>8} {'Tokens':>12} {'Cost':>8}")
+            lines.append(f"  {'Model':<30} {'Sessions':>8} {'Tokens':>12}")
             for m in report["models"]:
                 model_name = m["model"][:28]
-                if m.get("has_pricing"):
-                    cost_cell = f"${m['cost']:>6.2f}"
-                else:
-                    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("  * Cost N/A for custom/self-hosted models")
+                lines.append(f"  {model_name:<30} {m['sessions']:>8} {m['total_tokens']:>12,}")
             lines.append("")
 
         # Platform breakdown
@@ -739,15 +727,7 @@ class InsightsEngine:
 
         # Overview
         lines.append(f"**Sessions:** {o['total_sessions']} | **Messages:** {o['total_messages']:,} | **Tool calls:** {o['total_tool_calls']:,}")
-        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)_"
-        lines.append(f"**Est. cost:** ${o['estimated_cost']:.2f}{cost_note}")
+        lines.append(f"**Tokens:** {o['total_tokens']:,} (in: {o['total_input_tokens']:,} / out: {o['total_output_tokens']:,})")
         if o["total_hours"] > 0:
             lines.append(f"**Active time:** ~{_format_duration(o['total_hours'] * 3600)} | **Avg session:** ~{_format_duration(o['avg_session_duration'])}")
         lines.append("")
@@ -756,8 +736,7 @@ class InsightsEngine:
         if report["models"]:
             lines.append("**🤖 Models:**")
             for m in report["models"][:5]:
-                cost_str = f"${m['cost']:.2f}" if m.get("has_pricing") else "N/A"
-                lines.append(f"  {m['model'][:25]} — {m['sessions']} sessions, {m['total_tokens']:,} tokens, {cost_str}")
+                lines.append(f"  {m['model'][:25]} — {m['sessions']} sessions, {m['total_tokens']:,} tokens")
             lines.append("")
 
         # Platforms (if multi-platform)

agent/memory_manager.py

@@ -28,6 +28,7 @@ Usage in run_agent.py:
 
 from __future__ import annotations
 
+import json
 import logging
 import re
 from typing import Any, Dict, List, Optional
@@ -43,11 +44,22 @@ logger = logging.getLogger(__name__)
 # ---------------------------------------------------------------------------
 
 _FENCE_TAG_RE = re.compile(r'</?\s*memory-context\s*>', re.IGNORECASE)
+_INTERNAL_CONTEXT_RE = re.compile(
+    r'<\s*memory-context\s*>[\s\S]*?</\s*memory-context\s*>',
+    re.IGNORECASE,
+)
+_INTERNAL_NOTE_RE = re.compile(
+    r'\[System note:\s*The following is recalled memory context,\s*NOT new user input\.\s*Treat as informational background data\.\]\s*',
+    re.IGNORECASE,
+)
 
 
 def sanitize_context(text: str) -> str:
-    """Strip fence-escape sequences from provider output."""
-    return _FENCE_TAG_RE.sub('', text)
+    """Strip fence tags, injected context blocks, and system notes from provider output."""
+    text = _INTERNAL_CONTEXT_RE.sub('', text)
+    text = _INTERNAL_NOTE_RE.sub('', text)
+    text = _FENCE_TAG_RE.sub('', text)
+    return text
 
 
 def build_memory_context_block(raw_context: str) -> str:

agent/model_metadata.py

@@ -23,7 +23,7 @@ logger = logging.getLogger(__name__)
 # are preserved so the full model name reaches cache lookups and server queries.
 _PROVIDER_PREFIXES: frozenset[str] = frozenset({
     "openrouter", "nous", "openai-codex", "copilot", "copilot-acp",
-    "gemini", "zai", "kimi-coding", "kimi-coding-cn", "minimax", "minimax-cn", "anthropic", "deepseek",
+    "gemini", "ollama-cloud", "zai", "kimi-coding", "kimi-coding-cn", "minimax", "minimax-cn", "anthropic", "deepseek",
     "opencode-zen", "opencode-go", "ai-gateway", "kilocode", "alibaba",
     "qwen-oauth",
     "xiaomi",
@@ -33,10 +33,12 @@ _PROVIDER_PREFIXES: frozenset[str] = frozenset({
     "google", "google-gemini", "google-ai-studio",
     "glm", "z-ai", "z.ai", "zhipu", "github", "github-copilot",
     "github-models", "kimi", "moonshot", "kimi-cn", "moonshot-cn", "claude", "deep-seek",
+    "ollama",
     "opencode", "zen", "go", "vercel", "kilo", "dashscope", "aliyun", "qwen",
     "mimo", "xiaomi-mimo",
     "arcee-ai", "arceeai",
     "xai", "x-ai", "x.ai", "grok",
+    "nvidia", "nim", "nvidia-nim", "nemotron",
     "qwen-portal",
 })
 
@@ -101,6 +103,8 @@ DEFAULT_CONTEXT_LENGTHS = {
     # fuzzy-match collisions (e.g. "anthropic/claude-sonnet-4" is a
     # substring of "anthropic/claude-sonnet-4.6").
     # OpenRouter-prefixed models resolve via OpenRouter live API or models.dev.
+    "claude-opus-4-7": 1000000,
+    "claude-opus-4.7": 1000000,
     "claude-opus-4-6": 1000000,
     "claude-sonnet-4-6": 1000000,
     "claude-opus-4.6": 1000000,
@@ -121,7 +125,6 @@ DEFAULT_CONTEXT_LENGTHS = {
     "gemini": 1048576,
     # Gemma (open models served via AI Studio)
     "gemma-4-31b": 256000,
-    "gemma-4-26b": 256000,
     "gemma-3": 131072,
     "gemma": 8192,  # fallback for older gemma models
     # DeepSeek
@@ -155,6 +158,8 @@ DEFAULT_CONTEXT_LENGTHS = {
     "grok": 131072,             # catch-all (grok-beta, unknown grok-*)
     # Kimi
     "kimi": 262144,
+    # Nemotron — NVIDIA's open-weights series (128K context across all sizes)
+    "nemotron": 131072,
     # Arcee
     "trinity": 262144,
     # OpenRouter
@@ -237,8 +242,10 @@ _URL_TO_PROVIDER: Dict[str, str] = {
     "api.fireworks.ai": "fireworks",
     "opencode.ai": "opencode-go",
     "api.x.ai": "xai",
+    "integrate.api.nvidia.com": "nvidia",
     "api.xiaomimimo.com": "xiaomi",
     "xiaomimimo.com": "xiaomi",
+    "ollama.com": "ollama-cloud",
 }
 
 
@@ -1012,6 +1019,16 @@ def get_model_context_length(
         if ctx:
             return ctx
 
+    # 4b. AWS Bedrock — use static context length table.
+    # Bedrock's ListFoundationModels doesn't expose context window sizes,
+    # so we maintain a curated table in bedrock_adapter.py.
+    if provider == "bedrock" or (base_url and "bedrock-runtime" in base_url):
+        try:
+            from agent.bedrock_adapter import get_bedrock_context_length
+            return get_bedrock_context_length(model)
+        except ImportError:
+            pass  # boto3 not installed — fall through to generic resolution
+
     # 5. Provider-aware lookups (before generic OpenRouter cache)
     # These are provider-specific and take priority over the generic OR cache,
     # since the same model can have different context limits per provider

agent/models_dev.py

@@ -169,6 +169,7 @@ PROVIDER_TO_MODELS_DEV: Dict[str, str] = {
     "togetherai": "togetherai",
     "perplexity": "perplexity",
     "cohere": "cohere",
+    "ollama-cloud": "ollama-cloud",
 }
 
 # Reverse mapping: models.dev → Hermes (built lazily)
@@ -419,7 +420,10 @@ def list_provider_models(provider: str) -> List[str]:
     models = _get_provider_models(provider)
     if models is None:
         return []
-    return list(models.keys())
+    return [
+        mid for mid in models.keys()
+        if not _should_hide_from_provider_catalog(provider, mid)
+    ]
 
 
 # Patterns that indicate non-agentic or noise models (TTS, embedding,
@@ -431,6 +435,43 @@ _NOISE_PATTERNS: re.Pattern = re.compile(
     re.IGNORECASE,
 )
 
+# Google's live Gemini catalogs currently include a mix of stale slugs and
+# Gemma models whose TPM quotas are too small for normal Hermes agent traffic.
+# Keep capability metadata available for direct/manual use, but hide these from
+# the Gemini model catalogs we surface in setup and model selection.
+_GOOGLE_HIDDEN_MODELS = frozenset({
+    # Low-TPM Gemma models that trip Google input-token quota walls under
+    # agent-style traffic despite advertising large context windows.
+    "gemma-4-31b-it",
+    "gemma-4-26b-it",
+    "gemma-4-26b-a4b-it",
+    "gemma-3-1b",
+    "gemma-3-1b-it",
+    "gemma-3-2b",
+    "gemma-3-2b-it",
+    "gemma-3-4b",
+    "gemma-3-4b-it",
+    "gemma-3-12b",
+    "gemma-3-12b-it",
+    "gemma-3-27b",
+    "gemma-3-27b-it",
+    # Stale/retired Google slugs that still surface through models.dev-backed
+    # Gemini selection but 404 on the current Google endpoints.
+    "gemini-1.5-flash",
+    "gemini-1.5-pro",
+    "gemini-1.5-flash-8b",
+    "gemini-2.0-flash",
+    "gemini-2.0-flash-lite",
+})
+
+
+def _should_hide_from_provider_catalog(provider: str, model_id: str) -> bool:
+    provider_lower = (provider or "").strip().lower()
+    model_lower = (model_id or "").strip().lower()
+    if provider_lower in {"gemini", "google"} and model_lower in _GOOGLE_HIDDEN_MODELS:
+        return True
+    return False
+
 
 def list_agentic_models(provider: str) -> List[str]:
     """Return model IDs suitable for agentic use from models.dev.
@@ -447,6 +488,8 @@ def list_agentic_models(provider: str) -> List[str]:
     for mid, entry in models.items():
         if not isinstance(entry, dict):
             continue
+        if _should_hide_from_provider_catalog(provider, mid):
+            continue
         if not entry.get("tool_call", False):
             continue
         if _NOISE_PATTERNS.search(mid):
@@ -581,5 +624,3 @@ def get_model_info(
             return _parse_model_info(mid, mdata, mdev_id)
 
     return None
-
-

agent/nous_rate_guard.py

@@ -0,0 +1,182 @@
+"""Cross-session rate limit guard for Nous Portal.
+
+Writes rate limit state to a shared file so all sessions (CLI, gateway,
+cron, auxiliary) can check whether Nous Portal is currently rate-limited
+before making requests.  Prevents retry amplification when RPH is tapped.
+
+Each 429 from Nous triggers up to 9 API calls per conversation turn
+(3 SDK retries x 3 Hermes retries), and every one of those calls counts
+against RPH.  By recording the rate limit state on first 429 and checking
+it before subsequent attempts, we eliminate the amplification effect.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import tempfile
+import time
+from typing import Any, Mapping, Optional
+
+logger = logging.getLogger(__name__)
+
+_STATE_SUBDIR = "rate_limits"
+_STATE_FILENAME = "nous.json"
+
+
+def _state_path() -> str:
+    """Return the path to the Nous rate limit state file."""
+    try:
+        from hermes_constants import get_hermes_home
+        base = get_hermes_home()
+    except ImportError:
+        base = os.path.join(os.path.expanduser("~"), ".hermes")
+    return os.path.join(base, _STATE_SUBDIR, _STATE_FILENAME)
+
+
+def _parse_reset_seconds(headers: Optional[Mapping[str, str]]) -> Optional[float]:
+    """Extract the best available reset-time estimate from response headers.
+
+    Priority:
+      1. x-ratelimit-reset-requests-1h  (hourly RPH window — most useful)
+      2. x-ratelimit-reset-requests     (per-minute RPM window)
+      3. retry-after                     (generic HTTP header)
+
+    Returns seconds-from-now, or None if no usable header found.
+    """
+    if not headers:
+        return None
+
+    lowered = {k.lower(): v for k, v in headers.items()}
+
+    for key in (
+        "x-ratelimit-reset-requests-1h",
+        "x-ratelimit-reset-requests",
+        "retry-after",
+    ):
+        raw = lowered.get(key)
+        if raw is not None:
+            try:
+                val = float(raw)
+                if val > 0:
+                    return val
+            except (TypeError, ValueError):
+                pass
+
+    return None
+
+
+def record_nous_rate_limit(
+    *,
+    headers: Optional[Mapping[str, str]] = None,
+    error_context: Optional[dict[str, Any]] = None,
+    default_cooldown: float = 300.0,
+) -> None:
+    """Record that Nous Portal is rate-limited.
+
+    Parses the reset time from response headers or error context.
+    Falls back to ``default_cooldown`` (5 minutes) if no reset info
+    is available.  Writes to a shared file that all sessions can read.
+
+    Args:
+        headers: HTTP response headers from the 429 error.
+        error_context: Structured error context from _extract_api_error_context().
+        default_cooldown: Fallback cooldown in seconds when no header data.
+    """
+    now = time.time()
+    reset_at = None
+
+    # Try headers first (most accurate)
+    header_seconds = _parse_reset_seconds(headers)
+    if header_seconds is not None:
+        reset_at = now + header_seconds
+
+    # Try error_context reset_at (from body parsing)
+    if reset_at is None and isinstance(error_context, dict):
+        ctx_reset = error_context.get("reset_at")
+        if isinstance(ctx_reset, (int, float)) and ctx_reset > now:
+            reset_at = float(ctx_reset)
+
+    # Default cooldown
+    if reset_at is None:
+        reset_at = now + default_cooldown
+
+    path = _state_path()
+    try:
+        state_dir = os.path.dirname(path)
+        os.makedirs(state_dir, exist_ok=True)
+
+        state = {
+            "reset_at": reset_at,
+            "recorded_at": now,
+            "reset_seconds": reset_at - now,
+        }
+
+        # Atomic write: write to temp file + rename
+        fd, tmp_path = tempfile.mkstemp(dir=state_dir, suffix=".tmp")
+        try:
+            with os.fdopen(fd, "w") as f:
+                json.dump(state, f)
+            os.replace(tmp_path, path)
+        except Exception:
+            # Clean up temp file on failure
+            try:
+                os.unlink(tmp_path)
+            except OSError:
+                pass
+            raise
+
+        logger.info(
+            "Nous rate limit recorded: resets in %.0fs (at %.0f)",
+            reset_at - now, reset_at,
+        )
+    except Exception as exc:
+        logger.debug("Failed to write Nous rate limit state: %s", exc)
+
+
+def nous_rate_limit_remaining() -> Optional[float]:
+    """Check if Nous Portal is currently rate-limited.
+
+    Returns:
+        Seconds remaining until reset, or None if not rate-limited.
+    """
+    path = _state_path()
+    try:
+        with open(path) as f:
+            state = json.load(f)
+        reset_at = state.get("reset_at", 0)
+        remaining = reset_at - time.time()
+        if remaining > 0:
+            return remaining
+        # Expired — clean up
+        try:
+            os.unlink(path)
+        except OSError:
+            pass
+        return None
+    except (FileNotFoundError, json.JSONDecodeError, KeyError, TypeError):
+        return None
+
+
+def clear_nous_rate_limit() -> None:
+    """Clear the rate limit state (e.g., after a successful Nous request)."""
+    try:
+        os.unlink(_state_path())
+    except FileNotFoundError:
+        pass
+    except OSError as exc:
+        logger.debug("Failed to clear Nous rate limit state: %s", exc)
+
+
+def format_remaining(seconds: float) -> str:
+    """Format seconds remaining into human-readable duration."""
+    s = max(0, int(seconds))
+    if s < 60:
+        return f"{s}s"
+    if s < 3600:
+        m, sec = divmod(s, 60)
+        return f"{m}m {sec}s" if sec else f"{m}m"
+    h, remainder = divmod(s, 3600)
+    m = remainder // 60
+    return f"{h}h {m}m" if m else f"{h}h"

agent/prompt_builder.py

@@ -152,7 +152,13 @@ MEMORY_GUIDANCE = (
     "Do NOT save task progress, session outcomes, completed-work logs, or temporary TODO "
     "state to memory; use session_search to recall those from past transcripts. "
     "If you've discovered a new way to do something, solved a problem that could be "
-    "necessary later, save it as a skill with the skill tool."
+    "necessary later, save it as a skill with the skill tool.\n"
+    "Write memories as declarative facts, not instructions to yourself. "
+    "'User prefers concise responses' ✓ — 'Always respond concisely' ✗. "
+    "'Project uses pytest with xdist' ✓ — 'Run tests with pytest -n 4' ✗. "
+    "Imperative phrasing gets re-read as a directive in later sessions and can "
+    "cause repeated work or override the user's current request. Procedures and "
+    "workflows belong in skills, not memory."
 )
 
 SESSION_SEARCH_GUIDANCE = (
@@ -295,7 +301,9 @@ PLATFORM_HINTS = {
     ),
     "telegram": (
         "You are on a text messaging communication platform, Telegram. "
-        "Please do not use markdown as it does not render. "
+        "Standard markdown is automatically converted to Telegram format. "
+        "Supported: **bold**, *italic*, ~~strikethrough~~, ||spoiler||, "
+        "`inline code`, ```code blocks```, [links](url), and ## headers. "
         "You can send media files natively: to deliver a file to the user, "
         "include MEDIA:/absolute/path/to/file in your response. Images "
         "(.png, .jpg, .webp) appear as photos, audio (.ogg) sends as voice "
@@ -611,12 +619,14 @@ def build_skills_system_prompt(
         or get_session_env("HERMES_SESSION_PLATFORM")
         or ""
     )
+    disabled = get_disabled_skill_names()
     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,
+        tuple(sorted(disabled)),
     )
     with _SKILLS_PROMPT_CACHE_LOCK:
         cached = _SKILLS_PROMPT_CACHE.get(cache_key)
@@ -624,8 +634,6 @@ def build_skills_system_prompt(
             _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)
 
@@ -652,7 +660,7 @@ def build_skills_system_prompt(
             ):
                 continue
             skills_by_category.setdefault(category, []).append(
-                (skill_name, entry.get("description", ""))
+                (frontmatter_name, entry.get("description", ""))
             )
         category_descriptions = {
             str(k): str(v)
@@ -677,7 +685,7 @@ def build_skills_system_prompt(
             ):
                 continue
             skills_by_category.setdefault(entry["category"], []).append(
-                (skill_name, entry["description"])
+                (entry["frontmatter_name"], entry["description"])
             )
 
         # Read category-level DESCRIPTION.md files
@@ -720,9 +728,10 @@ def build_skills_system_prompt(
                     continue
                 entry = _build_snapshot_entry(skill_file, ext_dir, frontmatter, desc)
                 skill_name = entry["skill_name"]
-                if skill_name in seen_skill_names:
+                frontmatter_name = entry["frontmatter_name"]
+                if frontmatter_name in seen_skill_names:
                     continue
-                if entry["frontmatter_name"] in disabled or skill_name in disabled:
+                if frontmatter_name in disabled or skill_name in disabled:
                     continue
                 if not _skill_should_show(
                     extract_skill_conditions(frontmatter),
@@ -730,9 +739,9 @@ def build_skills_system_prompt(
                     available_toolsets,
                 ):
                     continue
-                seen_skill_names.add(skill_name)
+                seen_skill_names.add(frontmatter_name)
                 skills_by_category.setdefault(entry["category"], []).append(
-                    (skill_name, entry["description"])
+                    (frontmatter_name, entry["description"])
                 )
             except Exception as e:
                 logger.debug("Error reading external skill %s: %s", skill_file, e)

agent/redact.py

@@ -93,6 +93,17 @@ _DB_CONNSTR_RE = re.compile(
     re.IGNORECASE,
 )
 
+# JWT tokens: header.payload[.signature] — always start with "eyJ" (base64 for "{")
+# Matches 1-part (header only), 2-part (header.payload), and full 3-part JWTs.
+_JWT_RE = re.compile(
+    r"eyJ[A-Za-z0-9_-]{10,}"           # Header (always starts with eyJ)
+    r"(?:\.[A-Za-z0-9_=-]{4,}){0,2}"   # Optional payload and/or signature
+)
+
+# Discord user/role mentions: <@123456789012345678> or <@!123456789012345678>
+# Snowflake IDs are 17-20 digit integers that resolve to specific Discord accounts.
+_DISCORD_MENTION_RE = re.compile(r"<@!?(\d{17,20})>")
+
 # E.164 phone numbers: +<country><number>, 7-15 digits
 # Negative lookahead prevents matching hex strings or identifiers
 _SIGNAL_PHONE_RE = re.compile(r"(\+[1-9]\d{6,14})(?![A-Za-z0-9])")
@@ -159,6 +170,12 @@ def redact_sensitive_text(text: str) -> str:
     # Database connection string passwords
     text = _DB_CONNSTR_RE.sub(lambda m: f"{m.group(1)}***{m.group(3)}", text)
 
+    # JWT tokens (eyJ... — base64-encoded JSON headers)
+    text = _JWT_RE.sub(lambda m: _mask_token(m.group(0)), text)
+
+    # Discord user/role mentions (<@snowflake_id>)
+    text = _DISCORD_MENTION_RE.sub(lambda m: f"<@{'!' if '!' in m.group(0) else ''}***>", text)
+
     # E.164 phone numbers (Signal, WhatsApp)
     def _redact_phone(m):
         phone = m.group(1)

agent/skill_commands.py

@@ -12,6 +12,8 @@ from datetime import datetime
 from pathlib import Path
 from typing import Any, Dict, Optional
 
+from hermes_constants import display_hermes_home
+
 logger = logging.getLogger(__name__)
 
 _skill_commands: Dict[str, Dict[str, Any]] = {}
@@ -70,7 +72,14 @@ def _load_skill_payload(skill_identifier: str, task_id: str | None = None) -> tu
     skill_name = str(loaded_skill.get("name") or normalized)
     skill_path = str(loaded_skill.get("path") or "")
     skill_dir = None
-    if skill_path:
+    # Prefer the absolute skill_dir returned by skill_view() — this is
+    # correct for both local and external skills.  Fall back to the old
+    # SKILLS_DIR-relative reconstruction only when skill_dir is absent
+    # (e.g. legacy skill_view responses).
+    abs_skill_dir = loaded_skill.get("skill_dir")
+    if abs_skill_dir:
+        skill_dir = Path(abs_skill_dir)
+    elif skill_path:
         try:
             skill_dir = SKILLS_DIR / Path(skill_path).parent
         except Exception:
@@ -108,7 +117,7 @@ def _inject_skill_config(loaded_skill: dict[str, Any], parts: list[str]) -> None
         if not resolved:
             return
 
-        lines = ["", "[Skill config (from ~/.hermes/config.yaml):"]
+        lines = ["", f"[Skill config (from {display_hermes_home()}/config.yaml):"]
         for key, value in resolved.items():
             display_val = str(value) if value else "(not set)"
             lines.append(f"  {key} = {display_val}")

agent/smart_model_routing.py

@@ -1,195 +0,0 @@
-"""Helpers for optional cheap-vs-strong model routing."""
-
-from __future__ import annotations
-
-import os
-import re
-from typing import Any, Dict, Optional
-
-from utils import is_truthy_value
-
-_COMPLEX_KEYWORDS = {
-    "debug",
-    "debugging",
-    "implement",
-    "implementation",
-    "refactor",
-    "patch",
-    "traceback",
-    "stacktrace",
-    "exception",
-    "error",
-    "analyze",
-    "analysis",
-    "investigate",
-    "architecture",
-    "design",
-    "compare",
-    "benchmark",
-    "optimize",
-    "optimise",
-    "review",
-    "terminal",
-    "shell",
-    "tool",
-    "tools",
-    "pytest",
-    "test",
-    "tests",
-    "plan",
-    "planning",
-    "delegate",
-    "subagent",
-    "cron",
-    "docker",
-    "kubernetes",
-}
-
-_URL_RE = re.compile(r"https?://|www\.", re.IGNORECASE)
-
-
-def _coerce_bool(value: Any, default: bool = False) -> bool:
-    return is_truthy_value(value, default=default)
-
-
-def _coerce_int(value: Any, default: int) -> int:
-    try:
-        return int(value)
-    except (TypeError, ValueError):
-        return default
-
-
-def choose_cheap_model_route(user_message: str, routing_config: Optional[Dict[str, Any]]) -> Optional[Dict[str, Any]]:
-    """Return the configured cheap-model route when a message looks simple.
-
-    Conservative by design: if the message has signs of code/tool/debugging/
-    long-form work, keep the primary model.
-    """
-    cfg = routing_config or {}
-    if not _coerce_bool(cfg.get("enabled"), False):
-        return None
-
-    cheap_model = cfg.get("cheap_model") or {}
-    if not isinstance(cheap_model, dict):
-        return None
-    provider = str(cheap_model.get("provider") or "").strip().lower()
-    model = str(cheap_model.get("model") or "").strip()
-    if not provider or not model:
-        return None
-
-    text = (user_message or "").strip()
-    if not text:
-        return None
-
-    max_chars = _coerce_int(cfg.get("max_simple_chars"), 160)
-    max_words = _coerce_int(cfg.get("max_simple_words"), 28)
-
-    if len(text) > max_chars:
-        return None
-    if len(text.split()) > max_words:
-        return None
-    if text.count("\n") > 1:
-        return None
-    if "```" in text or "`" in text:
-        return None
-    if _URL_RE.search(text):
-        return None
-
-    lowered = text.lower()
-    words = {token.strip(".,:;!?()[]{}\"'`") for token in lowered.split()}
-    if words & _COMPLEX_KEYWORDS:
-        return None
-
-    route = dict(cheap_model)
-    route["provider"] = provider
-    route["model"] = model
-    route["routing_reason"] = "simple_turn"
-    return route
-
-
-def resolve_turn_route(user_message: str, routing_config: Optional[Dict[str, Any]], primary: Dict[str, Any]) -> Dict[str, Any]:
-    """Resolve the effective model/runtime for one turn.
-
-    Returns a dict with model/runtime/signature/label fields.
-    """
-    route = choose_cheap_model_route(user_message, routing_config)
-    if not route:
-        return {
-            "model": primary.get("model"),
-            "runtime": {
-                "api_key": primary.get("api_key"),
-                "base_url": primary.get("base_url"),
-                "provider": primary.get("provider"),
-                "api_mode": primary.get("api_mode"),
-                "command": primary.get("command"),
-                "args": list(primary.get("args") or []),
-                "credential_pool": primary.get("credential_pool"),
-            },
-            "label": None,
-            "signature": (
-                primary.get("model"),
-                primary.get("provider"),
-                primary.get("base_url"),
-                primary.get("api_mode"),
-                primary.get("command"),
-                tuple(primary.get("args") or ()),
-            ),
-        }
-
-    from hermes_cli.runtime_provider import resolve_runtime_provider
-
-    explicit_api_key = None
-    api_key_env = str(route.get("api_key_env") or "").strip()
-    if api_key_env:
-        explicit_api_key = os.getenv(api_key_env) or None
-
-    try:
-        runtime = resolve_runtime_provider(
-            requested=route.get("provider"),
-            explicit_api_key=explicit_api_key,
-            explicit_base_url=route.get("base_url"),
-        )
-    except Exception:
-        return {
-            "model": primary.get("model"),
-            "runtime": {
-                "api_key": primary.get("api_key"),
-                "base_url": primary.get("base_url"),
-                "provider": primary.get("provider"),
-                "api_mode": primary.get("api_mode"),
-                "command": primary.get("command"),
-                "args": list(primary.get("args") or []),
-                "credential_pool": primary.get("credential_pool"),
-            },
-            "label": None,
-            "signature": (
-                primary.get("model"),
-                primary.get("provider"),
-                primary.get("base_url"),
-                primary.get("api_mode"),
-                primary.get("command"),
-                tuple(primary.get("args") or ()),
-            ),
-        }
-
-    return {
-        "model": route.get("model"),
-        "runtime": {
-            "api_key": runtime.get("api_key"),
-            "base_url": runtime.get("base_url"),
-            "provider": runtime.get("provider"),
-            "api_mode": runtime.get("api_mode"),
-            "command": runtime.get("command"),
-            "args": list(runtime.get("args") or []),
-            "credential_pool": runtime.get("credential_pool"),
-        },
-        "label": f"smart route → {route.get('model')} ({runtime.get('provider')})",
-        "signature": (
-            route.get("model"),
-            runtime.get("provider"),
-            runtime.get("base_url"),
-            runtime.get("api_mode"),
-            runtime.get("command"),
-            tuple(runtime.get("args") or ()),
-        ),
-    }

agent/usage_pricing.py

@@ -284,6 +284,80 @@ _OFFICIAL_DOCS_PRICING: Dict[tuple[str, str], PricingEntry] = {
         source_url="https://ai.google.dev/pricing",
         pricing_version="google-pricing-2026-03-16",
     ),
+    # AWS Bedrock — pricing per the Bedrock pricing page.
+    # Bedrock charges the same per-token rates as the model provider but
+    # through AWS billing.  These are the on-demand prices (no commitment).
+    # Source: https://aws.amazon.com/bedrock/pricing/
+    (
+        "bedrock",
+        "anthropic.claude-opus-4-6",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("15.00"),
+        output_cost_per_million=Decimal("75.00"),
+        source="official_docs_snapshot",
+        source_url="https://aws.amazon.com/bedrock/pricing/",
+        pricing_version="bedrock-pricing-2026-04",
+    ),
+    (
+        "bedrock",
+        "anthropic.claude-sonnet-4-6",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("3.00"),
+        output_cost_per_million=Decimal("15.00"),
+        source="official_docs_snapshot",
+        source_url="https://aws.amazon.com/bedrock/pricing/",
+        pricing_version="bedrock-pricing-2026-04",
+    ),
+    (
+        "bedrock",
+        "anthropic.claude-sonnet-4-5",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("3.00"),
+        output_cost_per_million=Decimal("15.00"),
+        source="official_docs_snapshot",
+        source_url="https://aws.amazon.com/bedrock/pricing/",
+        pricing_version="bedrock-pricing-2026-04",
+    ),
+    (
+        "bedrock",
+        "anthropic.claude-haiku-4-5",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.80"),
+        output_cost_per_million=Decimal("4.00"),
+        source="official_docs_snapshot",
+        source_url="https://aws.amazon.com/bedrock/pricing/",
+        pricing_version="bedrock-pricing-2026-04",
+    ),
+    (
+        "bedrock",
+        "amazon.nova-pro",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.80"),
+        output_cost_per_million=Decimal("3.20"),
+        source="official_docs_snapshot",
+        source_url="https://aws.amazon.com/bedrock/pricing/",
+        pricing_version="bedrock-pricing-2026-04",
+    ),
+    (
+        "bedrock",
+        "amazon.nova-lite",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.06"),
+        output_cost_per_million=Decimal("0.24"),
+        source="official_docs_snapshot",
+        source_url="https://aws.amazon.com/bedrock/pricing/",
+        pricing_version="bedrock-pricing-2026-04",
+    ),
+    (
+        "bedrock",
+        "amazon.nova-micro",
+    ): PricingEntry(
+        input_cost_per_million=Decimal("0.035"),
+        output_cost_per_million=Decimal("0.14"),
+        source="official_docs_snapshot",
+        source_url="https://aws.amazon.com/bedrock/pricing/",
+        pricing_version="bedrock-pricing-2026-04",
+    ),
 }
 
 

batch_runner.py

@@ -561,7 +561,10 @@ class BatchRunner:
             provider_sort (str): Sort providers by price/throughput/latency (optional)
             max_tokens (int): Maximum tokens for model responses (optional, uses model default if not set)
             reasoning_config (Dict): OpenRouter reasoning config override (e.g. {"effort": "none"} to disable thinking)
-            prefill_messages (List[Dict]): Messages to prepend as prefilled conversation context (few-shot priming)
+            prefill_messages (List[Dict]): Messages to prepend as prefilled conversation context (few-shot priming).
+                NOTE: Anthropic Sonnet 4.6+ and Opus 4.6+ reject a trailing assistant-role prefill
+                (400 error).  For those models use output_config.format or structured-output
+                schemas instead.  Safe here for user-role priming and for older Claude / non-Claude models.
             max_samples (int): Only process the first N samples from the dataset (optional, processes all if not set)
         """
         self.dataset_file = Path(dataset_file)

cli-config.yaml.example

@@ -16,7 +16,7 @@ model:
   #   "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)
+  #   "openai-codex" - OpenAI Codex (requires: hermes auth)
   #   "copilot"      - GitHub Copilot / GitHub Models (requires: GITHUB_TOKEN)
   #   "gemini"      - Use Google AI Studio direct (requires: GOOGLE_API_KEY or GEMINI_API_KEY)
   #   "zai"         - Use z.ai / ZhipuAI GLM models (requires: GLM_API_KEY)
@@ -24,8 +24,10 @@ model:
   #   "minimax"      - MiniMax global (requires: MINIMAX_API_KEY)
   #   "minimax-cn"   - MiniMax China (requires: MINIMAX_CN_API_KEY)
   #   "huggingface"  - Hugging Face Inference (requires: HF_TOKEN)
+  #   "nvidia"       - NVIDIA NIM / build.nvidia.com (requires: NVIDIA_API_KEY)
   #   "xiaomi"       - Xiaomi MiMo (requires: XIAOMI_API_KEY)
   #   "arcee"        - Arcee AI Trinity models (requires: ARCEEAI_API_KEY)
+  #   "ollama-cloud" - Ollama Cloud (requires: OLLAMA_API_KEY — https://ollama.com/settings)
   #   "kilocode"     - KiloCode gateway (requires: KILOCODE_API_KEY)
   #   "ai-gateway"   - Vercel AI Gateway (requires: AI_GATEWAY_API_KEY)
   #
@@ -37,12 +39,6 @@ model:
   #     base_url: "http://localhost:1234/v1"
   #   No API key needed — local servers typically ignore auth.
   #
-  #   For Ollama Cloud (https://ollama.com/pricing):
-  #     provider: "custom"
-  #     base_url: "https://ollama.com/v1"
-  #   Set OLLAMA_API_KEY in .env — automatically picked up when base_url
-  #   points to ollama.com.
-  #
   # Can also be overridden with --provider flag or HERMES_INFERENCE_PROVIDER env var.
   provider: "auto"
   
@@ -67,7 +63,38 @@ model:
   #   Leave unset to use the model's native output ceiling (recommended).
   #   Set only if you want to deliberately limit individual response length.
   #
-  # max_tokens: 8192
+# max_tokens: 8192
+
+# Named provider overrides (optional)
+# Use this for per-provider request timeouts, non-stream stale timeouts,
+# and per-model exceptions.
+# Applies to the primary turn client on every api_mode (OpenAI-wire, native
+# Anthropic, and Anthropic-compatible providers), the fallback chain, and
+# client rebuilds during credential rotation.  For OpenAI-wire chat
+# completions (streaming and non-streaming) the configured value is also
+# used as the per-request ``timeout=`` kwarg so it wins over the legacy
+# HERMES_API_TIMEOUT env var (which still applies when no config is set).
+# ``stale_timeout_seconds`` controls the non-streaming stale-call detector and
+# wins over the legacy HERMES_API_CALL_STALE_TIMEOUT env var. Leaving these
+# unset keeps the legacy defaults (HERMES_API_TIMEOUT=1800s,
+# HERMES_API_CALL_STALE_TIMEOUT=300s, native Anthropic 900s).
+#
+# Not currently wired for AWS Bedrock (bedrock_converse + AnthropicBedrock
+# SDK paths) — those use boto3 with its own timeout configuration.
+#
+# providers:
+#   ollama-local:
+#     request_timeout_seconds: 300   # Longer timeout for local cold-starts
+#     stale_timeout_seconds: 900     # Explicitly re-enable stale detection on local endpoints
+#   anthropic:
+#     request_timeout_seconds: 30    # Fast-fail cloud requests
+#     models:
+#       claude-opus-4.6:
+#         timeout_seconds: 600       # Longer timeout for extended-thinking Opus calls
+#   openai-codex:
+#     models:
+#       gpt-5.4:
+#         stale_timeout_seconds: 1800  # Longer non-stream stale timeout for slow large-context turns
 
 # =============================================================================
 # OpenRouter Provider Routing (only applies when using OpenRouter)
@@ -95,20 +122,6 @@ model:
 #   # Data policy: "allow" (default) or "deny" to exclude providers that may store data
 #   # data_collection: "deny"
 
-# =============================================================================
-# Smart Model Routing (optional)
-# =============================================================================
-# Use a cheaper model for short/simple turns while keeping your main model for
-# more complex requests. Disabled by default.
-#
-# smart_model_routing:
-#   enabled: true
-#   max_simple_chars: 160
-#   max_simple_words: 28
-#   cheap_model:
-#     provider: openrouter
-#     model: google/gemini-2.5-flash
-
 # =============================================================================
 # Git Worktree Isolation
 # =============================================================================
@@ -337,6 +350,7 @@ compression:
 #   "openrouter" - Force OpenRouter (requires OPENROUTER_API_KEY)
 #   "nous"       - Force Nous Portal (requires: hermes login)
 #   "gemini"      - Force Google AI Studio direct (requires: GOOGLE_API_KEY or GEMINI_API_KEY)
+#   "ollama-cloud" - Ollama Cloud (requires: OLLAMA_API_KEY)
 #   "codex"       - Force Codex OAuth (requires: hermes model → Codex).
 #                  Uses gpt-5.3-codex which supports vision.
 #   "main"       - Use your custom endpoint (OPENAI_BASE_URL + OPENAI_API_KEY).
@@ -360,6 +374,18 @@ compression:
 #   web_extract:
 #     provider: "auto"
 #     model: ""
+#
+#   # Session search — summarizes matching past sessions
+#   session_search:
+#     provider: "auto"
+#     model: ""
+#     timeout: 30
+#     max_concurrency: 3    # Limit parallel summaries to reduce request-burst 429s
+#     extra_body: {}        # Provider-specific OpenAI-compatible request fields
+#                           # Example for providers that support request-body
+#                           # reasoning controls:
+#                           # extra_body:
+#                           #   enable_thinking: false
 
 # =============================================================================
 # Persistent Memory
@@ -564,6 +590,18 @@ platform_toolsets:
   homeassistant: [hermes-homeassistant]
   qqbot: [hermes-qqbot]
 
+# =============================================================================
+# Gateway Platform Settings
+# =============================================================================
+# Optional per-platform messaging settings.
+# Platform-specific knobs live under `extra`.
+#
+# platforms:
+#   telegram:
+#     reply_to_mode: "first"  # off | first | all
+#     extra:
+#       disable_link_previews: false  # Set true to suppress Telegram URL previews in bot messages
+
 # ─────────────────────────────────────────────────────────────────────────────
 # Available toolsets (use these names in platform_toolsets or the toolsets list)
 #

cron/jobs.py

@@ -501,6 +501,12 @@ def update_job(job_id: str, updates: Dict[str, Any]) -> Optional[Dict[str, Any]]
 
         if schedule_changed:
             updated_schedule = updated["schedule"]
+            # The API may pass schedule as a raw string (e.g. "every 10m")
+            # instead of a pre-parsed dict.  Normalize it the same way
+            # create_job() does so downstream code can call .get() safely.
+            if isinstance(updated_schedule, str):
+                updated_schedule = parse_schedule(updated_schedule)
+                updated["schedule"] = updated_schedule
             updated["schedule_display"] = updates.get(
                 "schedule_display",
                 updated_schedule.get("display", updated.get("schedule_display")),

cron/scheduler.py

@@ -10,6 +10,7 @@ runs at a time if multiple processes overlap.
 
 import asyncio
 import concurrent.futures
+import contextvars
 import json
 import logging
 import os
@@ -26,7 +27,7 @@ except ImportError:
     except ImportError:
         msvcrt = None
 from pathlib import Path
-from typing import Optional
+from typing import List, Optional
 
 # Add parent directory to path for imports BEFORE repo-level imports.
 # Without this, standalone invocations (e.g. after `hermes update` reloads
@@ -48,6 +49,33 @@ _KNOWN_DELIVERY_PLATFORMS = frozenset({
     "qqbot",
 })
 
+# Platforms that support a configured cron/notification home target, mapped to
+# the environment variable used by gateway setup/runtime config.
+_HOME_TARGET_ENV_VARS = {
+    "matrix": "MATRIX_HOME_ROOM",
+    "telegram": "TELEGRAM_HOME_CHANNEL",
+    "discord": "DISCORD_HOME_CHANNEL",
+    "slack": "SLACK_HOME_CHANNEL",
+    "signal": "SIGNAL_HOME_CHANNEL",
+    "mattermost": "MATTERMOST_HOME_CHANNEL",
+    "sms": "SMS_HOME_CHANNEL",
+    "email": "EMAIL_HOME_ADDRESS",
+    "dingtalk": "DINGTALK_HOME_CHANNEL",
+    "feishu": "FEISHU_HOME_CHANNEL",
+    "wecom": "WECOM_HOME_CHANNEL",
+    "weixin": "WEIXIN_HOME_CHANNEL",
+    "bluebubbles": "BLUEBUBBLES_HOME_CHANNEL",
+    "qqbot": "QQBOT_HOME_CHANNEL",
+}
+
+# Legacy env var names kept for back-compat.  Each entry is the current
+# primary env var → the previous name.  _get_home_target_chat_id falls
+# back to the legacy name if the primary is unset, so users who set the
+# old name before the rename keep working until they migrate.
+_LEGACY_HOME_TARGET_ENV_VARS = {
+    "QQBOT_HOME_CHANNEL": "QQ_HOME_CHANNEL",
+}
+
 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
@@ -75,15 +103,28 @@ def _resolve_origin(job: dict) -> Optional[dict]:
     return None
 
 
-def _resolve_delivery_target(job: dict) -> Optional[dict]:
-    """Resolve the concrete auto-delivery target for a cron job, if any."""
-    deliver = job.get("deliver", "local")
+def _get_home_target_chat_id(platform_name: str) -> str:
+    """Return the configured home target chat/room ID for a delivery platform."""
+    env_var = _HOME_TARGET_ENV_VARS.get(platform_name.lower())
+    if not env_var:
+        return ""
+    value = os.getenv(env_var, "")
+    if not value:
+        legacy = _LEGACY_HOME_TARGET_ENV_VARS.get(env_var)
+        if legacy:
+            value = os.getenv(legacy, "")
+    return value
+
+
+def _resolve_single_delivery_target(job: dict, deliver_value: str) -> Optional[dict]:
+    """Resolve one concrete auto-delivery target for a cron job."""
+
     origin = _resolve_origin(job)
 
-    if deliver == "local":
+    if deliver_value == "local":
         return None
 
-    if deliver == "origin":
+    if deliver_value == "origin":
         if origin:
             return {
                 "platform": origin["platform"],
@@ -92,8 +133,8 @@ def _resolve_delivery_target(job: dict) -> Optional[dict]:
             }
         # Origin missing (e.g. job created via API/script) — try each
         # platform's home channel as a fallback instead of silently dropping.
-        for platform_name in ("matrix", "telegram", "discord", "slack", "bluebubbles"):
-            chat_id = os.getenv(f"{platform_name.upper()}_HOME_CHANNEL", "")
+        for platform_name in _HOME_TARGET_ENV_VARS:
+            chat_id = _get_home_target_chat_id(platform_name)
             if chat_id:
                 logger.info(
                     "Job '%s' has deliver=origin but no origin; falling back to %s home channel",
@@ -107,8 +148,8 @@ def _resolve_delivery_target(job: dict) -> Optional[dict]:
                 }
         return None
 
-    if ":" in deliver:
-        platform_name, rest = deliver.split(":", 1)
+    if ":" in deliver_value:
+        platform_name, rest = deliver_value.split(":", 1)
         platform_key = platform_name.lower()
 
         from tools.send_message_tool import _parse_target_ref
@@ -138,7 +179,7 @@ def _resolve_delivery_target(job: dict) -> Optional[dict]:
             "thread_id": thread_id,
         }
 
-    platform_name = deliver
+    platform_name = deliver_value
     if origin and origin.get("platform") == platform_name:
         return {
             "platform": platform_name,
@@ -148,7 +189,7 @@ def _resolve_delivery_target(job: dict) -> Optional[dict]:
 
     if platform_name.lower() not in _KNOWN_DELIVERY_PLATFORMS:
         return None
-    chat_id = os.getenv(f"{platform_name.upper()}_HOME_CHANNEL", "")
+    chat_id = _get_home_target_chat_id(platform_name)
     if not chat_id:
         return None
 
@@ -159,6 +200,30 @@ def _resolve_delivery_target(job: dict) -> Optional[dict]:
     }
 
 
+def _resolve_delivery_targets(job: dict) -> List[dict]:
+    """Resolve all concrete auto-delivery targets for a cron job (supports comma-separated deliver)."""
+    deliver = job.get("deliver", "local")
+    if deliver == "local":
+        return []
+    parts = [p.strip() for p in str(deliver).split(",") if p.strip()]
+    seen = set()
+    targets = []
+    for part in parts:
+        target = _resolve_single_delivery_target(job, part)
+        if target:
+            key = (target["platform"].lower(), str(target["chat_id"]), target.get("thread_id"))
+            if key not in seen:
+                seen.add(key)
+                targets.append(target)
+    return targets
+
+
+def _resolve_delivery_target(job: dict) -> Optional[dict]:
+    """Resolve the concrete auto-delivery target for a cron job, if any."""
+    targets = _resolve_delivery_targets(job)
+    return targets[0] if targets else None
+
+
 # Media extension sets — keep in sync with gateway/platforms/base.py:_process_message_background
 _AUDIO_EXTS = frozenset({'.ogg', '.opus', '.mp3', '.wav', '.m4a'})
 _VIDEO_EXTS = frozenset({'.mp4', '.mov', '.avi', '.mkv', '.webm', '.3gp'})
@@ -199,7 +264,7 @@ def _send_media_via_adapter(adapter, chat_id: str, media_files: list, metadata:
 
 def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> Optional[str]:
     """
-    Deliver job output to the configured target (origin chat, specific platform, etc.).
+    Deliver job output to the configured target(s) (origin chat, specific platform, etc.).
 
     When ``adapters`` and ``loop`` are provided (gateway is running), tries to
     use the live adapter first — this supports E2EE rooms (e.g. Matrix) where
@@ -208,33 +273,14 @@ def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> Option
 
     Returns None on success, or an error string on failure.
     """
-    target = _resolve_delivery_target(job)
-    if not target:
+    targets = _resolve_delivery_targets(job)
+    if not targets:
         if job.get("deliver", "local") != "local":
             msg = f"no delivery target resolved for deliver={job.get('deliver', 'local')}"
             logger.warning("Job '%s': %s", job["id"], msg)
             return msg
         return None  # local-only jobs don't deliver — not a failure
 
-    platform_name = target["platform"]
-    chat_id = target["chat_id"]
-    thread_id = target.get("thread_id")
-
-    # Diagnostic: log thread_id for topic-aware delivery debugging
-    origin = job.get("origin") or {}
-    origin_thread = origin.get("thread_id")
-    if origin_thread and not thread_id:
-        logger.warning(
-            "Job '%s': origin has thread_id=%s but delivery target lost it "
-            "(deliver=%s, target=%s)",
-            job["id"], origin_thread, job.get("deliver", "local"), target,
-        )
-    elif thread_id:
-        logger.debug(
-            "Job '%s': delivering to %s:%s thread_id=%s",
-            job["id"], platform_name, chat_id, thread_id,
-        )
-
     from tools.send_message_tool import _send_to_platform
     from gateway.config import load_gateway_config, Platform
 
@@ -257,24 +303,6 @@ def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> Option
         "bluebubbles": Platform.BLUEBUBBLES,
         "qqbot": Platform.QQBOT,
     }
-    platform = platform_map.get(platform_name.lower())
-    if not platform:
-        msg = f"unknown platform '{platform_name}'"
-        logger.warning("Job '%s': %s", job["id"], msg)
-        return msg
-
-    try:
-        config = load_gateway_config()
-    except Exception as e:
-        msg = f"failed to load gateway config: {e}"
-        logger.error("Job '%s': %s", job["id"], msg)
-        return msg
-
-    pconfig = config.platforms.get(platform)
-    if not pconfig or not pconfig.enabled:
-        msg = f"platform '{platform_name}' not configured/enabled"
-        logger.warning("Job '%s': %s", job["id"], msg)
-        return msg
 
     # 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
@@ -288,11 +316,13 @@ def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> Option
 
     if wrap_response:
         task_name = job.get("name", job["id"])
+        job_id = job.get("id", "")
         delivery_content = (
             f"Cronjob Response: {task_name}\n"
+            f"(job_id: {job_id})\n"
             f"-------------\n\n"
             f"{content}\n\n"
-            f"Note: The agent cannot see this message, and therefore cannot respond to it."
+            f"To stop or manage this job, send me a new message (e.g. \"stop reminder {task_name}\")."
         )
     else:
         delivery_content = content
@@ -301,67 +331,117 @@ def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> Option
     from gateway.platforms.base import BasePlatformAdapter
     media_files, cleaned_delivery_content = BasePlatformAdapter.extract_media(delivery_content)
 
-    # Prefer the live adapter when the gateway is running — this supports E2EE
-    # rooms (e.g. Matrix) where the standalone HTTP path cannot encrypt.
-    runtime_adapter = (adapters or {}).get(platform)
-    if runtime_adapter is not None and loop is not None and getattr(loop, "is_running", lambda: False)():
-        send_metadata = {"thread_id": thread_id} if thread_id else None
-        try:
-            # Send cleaned text (MEDIA tags stripped) — not the raw content
-            text_to_send = cleaned_delivery_content.strip()
-            adapter_ok = True
-            if text_to_send:
-                future = asyncio.run_coroutine_threadsafe(
-                    runtime_adapter.send(chat_id, text_to_send, metadata=send_metadata),
-                    loop,
-                )
-                send_result = future.result(timeout=60)
-                if send_result and not getattr(send_result, "success", True):
-                    err = getattr(send_result, "error", "unknown")
-                    logger.warning(
-                        "Job '%s': live adapter send to %s:%s failed (%s), falling back to standalone",
-                        job["id"], platform_name, chat_id, err,
-                    )
-                    adapter_ok = False  # fall through to standalone path
+    try:
+        config = load_gateway_config()
+    except Exception as e:
+        msg = f"failed to load gateway config: {e}"
+        logger.error("Job '%s': %s", job["id"], msg)
+        return msg
 
-            # Send extracted media files as native attachments via the live adapter
-            if adapter_ok and media_files:
-                _send_media_via_adapter(runtime_adapter, chat_id, media_files, send_metadata, loop, job)
+    delivery_errors = []
 
-            if adapter_ok:
-                logger.info("Job '%s': delivered to %s:%s via live adapter", job["id"], platform_name, chat_id)
-                return None
-        except Exception as e:
+    for target in targets:
+        platform_name = target["platform"]
+        chat_id = target["chat_id"]
+        thread_id = target.get("thread_id")
+
+        # Diagnostic: log thread_id for topic-aware delivery debugging
+        origin = job.get("origin") or {}
+        origin_thread = origin.get("thread_id")
+        if origin_thread and not thread_id:
             logger.warning(
-                "Job '%s': live adapter delivery to %s:%s failed (%s), falling back to standalone",
-                job["id"], platform_name, chat_id, e,
+                "Job '%s': origin has thread_id=%s but delivery target lost it "
+                "(deliver=%s, target=%s)",
+                job["id"], origin_thread, job.get("deliver", "local"), target,
+            )
+        elif thread_id:
+            logger.debug(
+                "Job '%s': delivering to %s:%s thread_id=%s",
+                job["id"], platform_name, chat_id, thread_id,
             )
 
-    # Standalone path: run the async send in a fresh event loop (safe from any thread)
-    coro = _send_to_platform(platform, pconfig, chat_id, cleaned_delivery_content, thread_id=thread_id, media_files=media_files)
-    try:
-        result = asyncio.run(coro)
-    except RuntimeError:
-        # 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, cleaned_delivery_content, thread_id=thread_id, media_files=media_files))
-            result = future.result(timeout=30)
-    except Exception as e:
-        msg = f"delivery to {platform_name}:{chat_id} failed: {e}"
-        logger.error("Job '%s': %s", job["id"], msg)
-        return msg
+        platform = platform_map.get(platform_name.lower())
+        if not platform:
+            msg = f"unknown platform '{platform_name}'"
+            logger.warning("Job '%s': %s", job["id"], msg)
+            delivery_errors.append(msg)
+            continue
 
-    if result and result.get("error"):
-        msg = f"delivery error: {result['error']}"
-        logger.error("Job '%s': %s", job["id"], msg)
-        return msg
+        # Prefer the live adapter when the gateway is running — this supports E2EE
+        # rooms (e.g. Matrix) where the standalone HTTP path cannot encrypt.
+        runtime_adapter = (adapters or {}).get(platform)
+        delivered = False
+        if runtime_adapter is not None and loop is not None and getattr(loop, "is_running", lambda: False)():
+            send_metadata = {"thread_id": thread_id} if thread_id else None
+            try:
+                # Send cleaned text (MEDIA tags stripped) — not the raw content
+                text_to_send = cleaned_delivery_content.strip()
+                adapter_ok = True
+                if text_to_send:
+                    future = asyncio.run_coroutine_threadsafe(
+                        runtime_adapter.send(chat_id, text_to_send, metadata=send_metadata),
+                        loop,
+                    )
+                    send_result = future.result(timeout=60)
+                    if send_result and not getattr(send_result, "success", True):
+                        err = getattr(send_result, "error", "unknown")
+                        logger.warning(
+                            "Job '%s': live adapter send to %s:%s failed (%s), falling back to standalone",
+                            job["id"], platform_name, chat_id, err,
+                        )
+                        adapter_ok = False  # fall through to standalone path
 
-    logger.info("Job '%s': delivered to %s:%s", job["id"], platform_name, chat_id)
+                # Send extracted media files as native attachments via the live adapter
+                if adapter_ok and media_files:
+                    _send_media_via_adapter(runtime_adapter, chat_id, media_files, send_metadata, loop, job)
+
+                if adapter_ok:
+                    logger.info("Job '%s': delivered to %s:%s via live adapter", job["id"], platform_name, chat_id)
+                    delivered = True
+            except Exception as e:
+                logger.warning(
+                    "Job '%s': live adapter delivery to %s:%s failed (%s), falling back to standalone",
+                    job["id"], platform_name, chat_id, e,
+                )
+
+        if not delivered:
+            pconfig = config.platforms.get(platform)
+            if not pconfig or not pconfig.enabled:
+                msg = f"platform '{platform_name}' not configured/enabled"
+                logger.warning("Job '%s': %s", job["id"], msg)
+                delivery_errors.append(msg)
+                continue
+
+            # Standalone path: run the async send in a fresh event loop (safe from any thread)
+            coro = _send_to_platform(platform, pconfig, chat_id, cleaned_delivery_content, thread_id=thread_id, media_files=media_files)
+            try:
+                result = asyncio.run(coro)
+            except RuntimeError:
+                # 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, cleaned_delivery_content, thread_id=thread_id, media_files=media_files))
+                    result = future.result(timeout=30)
+            except Exception as e:
+                msg = f"delivery to {platform_name}:{chat_id} failed: {e}"
+                logger.error("Job '%s': %s", job["id"], msg)
+                delivery_errors.append(msg)
+                continue
+
+            if result and result.get("error"):
+                msg = f"delivery error: {result['error']}"
+                logger.error("Job '%s': %s", job["id"], msg)
+                delivery_errors.append(msg)
+                continue
+
+            logger.info("Job '%s': delivered to %s:%s", job["id"], platform_name, chat_id)
+
+    if delivery_errors:
+        return "; ".join(delivery_errors)
     return None
 
 
@@ -484,15 +564,53 @@ def _run_job_script(script_path: str) -> tuple[bool, str]:
         return False, f"Script execution failed: {exc}"
 
 
-def _build_job_prompt(job: dict) -> str:
-    """Build the effective prompt for a cron job, optionally loading one or more skills first."""
+def _parse_wake_gate(script_output: str) -> bool:
+    """Parse the last non-empty stdout line of a cron job's pre-check script
+    as a wake gate.
+
+    The convention (ported from nanoclaw #1232): if the last stdout line is
+    JSON like ``{"wakeAgent": false}``, the agent is skipped entirely — no
+    LLM run, no delivery. Any other output (non-JSON, missing flag, gate
+    absent, or ``wakeAgent: true``) means wake the agent normally.
+
+    Returns True if the agent should wake, False to skip.
+    """
+    if not script_output:
+        return True
+    stripped_lines = [line for line in script_output.splitlines() if line.strip()]
+    if not stripped_lines:
+        return True
+    last_line = stripped_lines[-1].strip()
+    try:
+        gate = json.loads(last_line)
+    except (json.JSONDecodeError, ValueError):
+        return True
+    if not isinstance(gate, dict):
+        return True
+    return gate.get("wakeAgent", True) is not False
+
+
+def _build_job_prompt(job: dict, prerun_script: Optional[tuple] = None) -> str:
+    """Build the effective prompt for a cron job, optionally loading one or more skills first.
+
+    Args:
+        job: The cron job dict.
+        prerun_script: Optional ``(success, stdout)`` from a script that has
+            already been executed by the caller (e.g. for a wake-gate check).
+            When provided, the script is not re-executed and the cached
+            result is used for prompt injection. When omitted, the script
+            (if any) runs inline as before.
+    """
     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 prerun_script is not None:
+            success, script_output = prerun_script
+        else:
+            success, script_output = _run_job_script(script_path)
         if success:
             if script_output:
                 prompt = (
@@ -594,13 +712,41 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
     
     job_id = job["id"]
     job_name = job["name"]
-    prompt = _build_job_prompt(job)
+
+    # Wake-gate: if this job has a pre-check script, run it BEFORE building
+    # the prompt so a ``{"wakeAgent": false}`` response can short-circuit
+    # the whole agent run. We pass the result into _build_job_prompt so
+    # the script is only executed once.
+    prerun_script = None
+    script_path = job.get("script")
+    if script_path:
+        prerun_script = _run_job_script(script_path)
+        _ran_ok, _script_output = prerun_script
+        if _ran_ok and not _parse_wake_gate(_script_output):
+            logger.info(
+                "Job '%s' (ID: %s): wakeAgent=false, skipping agent run",
+                job_name, job_id,
+            )
+            silent_doc = (
+                f"# Cron Job: {job_name}\n\n"
+                f"**Job ID:** {job_id}\n"
+                f"**Run Time:** {_hermes_now().strftime('%Y-%m-%d %H:%M:%S')}\n\n"
+                "Script gate returned `wakeAgent=false` — agent skipped.\n"
+            )
+            return True, silent_doc, SILENT_MARKER, None
+
+    prompt = _build_job_prompt(job, prerun_script=prerun_script)
     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])
 
+    # Mark this as a cron session so the approval system can apply cron_mode.
+    # This env var is process-wide and persists for the lifetime of the
+    # scheduler process — every job this process runs is a cron job.
+    os.environ["HERMES_CRON_SESSION"] = "1"
+
     try:
         # Inject origin context so the agent's send_message tool knows the chat.
         # Must be INSIDE the try block so the finally cleanup always runs.
@@ -680,7 +826,6 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
 
         # Provider routing
         pr = _cfg.get("provider_routing", {})
-        smart_routing = _cfg.get("smart_model_routing", {}) or {}
 
         from hermes_cli.runtime_provider import (
             resolve_runtime_provider,
@@ -697,24 +842,9 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
             message = format_runtime_provider_error(exc)
             raise RuntimeError(message) from exc
 
-        from agent.smart_model_routing import resolve_turn_route
-        turn_route = resolve_turn_route(
-            prompt,
-            smart_routing,
-            {
-                "model": model,
-                "api_key": runtime.get("api_key"),
-                "base_url": runtime.get("base_url"),
-                "provider": runtime.get("provider"),
-                "api_mode": runtime.get("api_mode"),
-                "command": runtime.get("command"),
-                "args": list(runtime.get("args") or []),
-            },
-        )
-
         fallback_model = _cfg.get("fallback_providers") or _cfg.get("fallback_model") or None
         credential_pool = None
-        runtime_provider = str(turn_route["runtime"].get("provider") or "").strip().lower()
+        runtime_provider = str(runtime.get("provider") or "").strip().lower()
         if runtime_provider:
             try:
                 from agent.credential_pool import load_pool
@@ -731,13 +861,13 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
                 logger.debug("Job '%s': failed to load credential pool for %s: %s", job_id, runtime_provider, e)
 
         agent = AIAgent(
-            model=turn_route["model"],
-            api_key=turn_route["runtime"].get("api_key"),
-            base_url=turn_route["runtime"].get("base_url"),
-            provider=turn_route["runtime"].get("provider"),
-            api_mode=turn_route["runtime"].get("api_mode"),
-            acp_command=turn_route["runtime"].get("command"),
-            acp_args=turn_route["runtime"].get("args"),
+            model=model,
+            api_key=runtime.get("api_key"),
+            base_url=runtime.get("base_url"),
+            provider=runtime.get("provider"),
+            api_mode=runtime.get("api_mode"),
+            acp_command=runtime.get("command"),
+            acp_args=runtime.get("args"),
             max_iterations=max_iterations,
             reasoning_config=reasoning_config,
             prefill_messages=prefill_messages,
@@ -768,7 +898,11 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
         _cron_inactivity_limit = _cron_timeout if _cron_timeout > 0 else None
         _POLL_INTERVAL = 5.0
         _cron_pool = concurrent.futures.ThreadPoolExecutor(max_workers=1)
-        _cron_future = _cron_pool.submit(agent.run_conversation, prompt)
+        # Preserve scheduler-scoped ContextVar state (for example skill-declared
+        # env passthrough registrations) when the cron run hops into the worker
+        # thread used for inactivity timeout monitoring.
+        _cron_context = contextvars.copy_context()
+        _cron_future = _cron_pool.submit(_cron_context.run, agent.run_conversation, prompt)
         _inactivity_timeout = False
         try:
             if _cron_inactivity_limit is None:
@@ -830,6 +964,9 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
             )
 
         final_response = result.get("final_response", "") or ""
+        # Strip leaked placeholder text that upstream may inject on empty completions.
+        if final_response.strip() == "(No response generated)":
+            final_response = ""
         # 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)"
@@ -969,6 +1106,13 @@ def tick(verbose: bool = True, adapters=None, loop=None) -> int:
                         delivery_error = str(de)
                         logger.error("Delivery failed for job %s: %s", job["id"], de)
 
+                # Treat empty final_response as a soft failure so last_status
+                # is not "ok" — the agent ran but produced nothing useful.
+                # (issue #8585)
+                if success and not final_response:
+                    success = False
+                    error = "Agent completed but produced empty response (model error, timeout, or misconfiguration)"
+
                 mark_job_run(job["id"], success, error, delivery_error=delivery_error)
                 executed += 1
 

docs/acp-setup.md

@@ -1,228 +0,0 @@
-# Hermes Agent — ACP (Agent Client Protocol) Setup Guide
-
-Hermes Agent supports the **Agent Client Protocol (ACP)**, allowing it to run as
-a coding agent inside your editor. ACP lets your IDE send tasks to Hermes, and
-Hermes responds with file edits, terminal commands, and explanations — all shown
-natively in the editor UI.
-
----
-
-## Prerequisites
-
-- Hermes Agent installed and configured (`hermes setup` completed)
-- An API key / provider set up in `~/.hermes/.env` or via `hermes login`
-- Python 3.11+
-
-Install the ACP extra:
-
-```bash
-pip install -e ".[acp]"
-```
-
----
-
-## VS Code Setup
-
-### 1. Install the ACP Client extension
-
-Open VS Code and install **ACP Client** from the marketplace:
-
-- Press `Ctrl+Shift+X` (or `Cmd+Shift+X` on macOS)
-- Search for **"ACP Client"**
-- Click **Install**
-
-Or install from the command line:
-
-```bash
-code --install-extension anysphere.acp-client
-```
-
-### 2. Configure settings.json
-
-Open your VS Code settings (`Ctrl+,` → click the `{}` icon for JSON) and add:
-
-```json
-{
-  "acpClient.agents": [
-    {
-      "name": "hermes-agent",
-      "registryDir": "/path/to/hermes-agent/acp_registry"
-    }
-  ]
-}
-```
-
-Replace `/path/to/hermes-agent` with the actual path to your Hermes Agent
-installation (e.g. `~/.hermes/hermes-agent`).
-
-Alternatively, if `hermes` is on your PATH, the ACP Client can discover it
-automatically via the registry directory.
-
-### 3. Restart VS Code
-
-After configuring, restart VS Code. You should see **Hermes Agent** appear in
-the ACP agent picker in the chat/agent panel.
-
----
-
-## Zed Setup
-
-Zed has built-in ACP support.
-
-### 1. Configure Zed settings
-
-Open Zed settings (`Cmd+,` on macOS or `Ctrl+,` on Linux) and add to your
-`settings.json`:
-
-```json
-{
-  "agent_servers": {
-    "hermes-agent": {
-      "type": "custom",
-      "command": "hermes",
-      "args": ["acp"],
-    },
-  },
-}
-```
-
-### 2. Restart Zed
-
-Hermes Agent will appear in the agent panel. Select it and start a conversation.
-
----
-
-## JetBrains Setup (IntelliJ, PyCharm, WebStorm, etc.)
-
-### 1. Install the ACP plugin
-
-- Open **Settings** → **Plugins** → **Marketplace**
-- Search for **"ACP"** or **"Agent Client Protocol"**
-- Install and restart the IDE
-
-### 2. Configure the agent
-
-- Open **Settings** → **Tools** → **ACP Agents**
-- Click **+** to add a new agent
-- Set the registry directory to your `acp_registry/` folder:
-  `/path/to/hermes-agent/acp_registry`
-- Click **OK**
-
-### 3. Use the agent
-
-Open the ACP panel (usually in the right sidebar) and select **Hermes Agent**.
-
----
-
-## What You Will See
-
-Once connected, your editor provides a native interface to Hermes Agent:
-
-### Chat Panel
-A conversational interface where you can describe tasks, ask questions, and
-give instructions. Hermes responds with explanations and actions.
-
-### File Diffs
-When Hermes edits files, you see standard diffs in the editor. You can:
-- **Accept** individual changes
-- **Reject** changes you don't want
-- **Review** the full diff before applying
-
-### Terminal Commands
-When Hermes needs to run shell commands (builds, tests, installs), the editor
-shows them in an integrated terminal. Depending on your settings:
-- Commands may run automatically
-- Or you may be prompted to **approve** each command
-
-### Approval Flow
-For potentially destructive operations, the editor will prompt you for
-approval before Hermes proceeds. This includes:
-- File deletions
-- Shell commands
-- Git operations
-
----
-
-## Configuration
-
-Hermes Agent under ACP uses the **same configuration** as the CLI:
-
-- **API keys / providers**: `~/.hermes/.env`
-- **Agent config**: `~/.hermes/config.yaml`
-- **Skills**: `~/.hermes/skills/`
-- **Sessions**: `~/.hermes/state.db`
-
-You can run `hermes setup` to configure providers, or edit `~/.hermes/.env`
-directly.
-
-### Changing the model
-
-Edit `~/.hermes/config.yaml`:
-
-```yaml
-model: openrouter/nous/hermes-3-llama-3.1-70b
-```
-
-Or set the `HERMES_MODEL` environment variable.
-
-### Toolsets
-
-ACP sessions use the curated `hermes-acp` toolset by default. It is designed for editor workflows and intentionally excludes things like messaging delivery, cronjob management, and audio-first UX features.
-
----
-
-## Troubleshooting
-
-### Agent doesn't appear in the editor
-
-1. **Check the registry path** — make sure the `acp_registry/` directory path
-   in your editor settings is correct and contains `agent.json`.
-2. **Check `hermes` is on PATH** — run `which hermes` in a terminal. If not
-   found, you may need to activate your virtualenv or add it to PATH.
-3. **Restart the editor** after changing settings.
-
-### Agent starts but errors immediately
-
-1. Run `hermes doctor` to check your configuration.
-2. Check that you have a valid API key: `hermes status`
-3. Try running `hermes acp` directly in a terminal to see error output.
-
-### "Module not found" errors
-
-Make sure you installed the ACP extra:
-
-```bash
-pip install -e ".[acp]"
-```
-
-### Slow responses
-
-- ACP streams responses, so you should see incremental output. If the agent
-  appears stuck, check your network connection and API provider status.
-- Some providers have rate limits. Try switching to a different model/provider.
-
-### Permission denied for terminal commands
-
-If the editor blocks terminal commands, check your ACP Client extension
-settings for auto-approval or manual-approval preferences.
-
-### Logs
-
-Hermes logs are written to stderr when running in ACP mode. Check:
-- VS Code: **Output** panel → select **ACP Client** or **Hermes Agent**
-- Zed: **View** → **Toggle Terminal** and check the process output
-- JetBrains: **Event Log** or the ACP tool window
-
-You can also enable verbose logging:
-
-```bash
-HERMES_LOG_LEVEL=DEBUG hermes acp
-```
-
----
-
-## Further Reading
-
-- [ACP Specification](https://github.com/anysphere/acp)
-- [Hermes Agent Documentation](https://github.com/NousResearch/hermes-agent)
-- Run `hermes --help` for all CLI options

docs/honcho-integration-spec.html

@@ -1,698 +0,0 @@
-<!DOCTYPE html>
-<html lang="en">
-<head>
-<meta charset="UTF-8">
-<meta name="viewport" content="width=device-width, initial-scale=1.0">
-<title>honcho-integration-spec</title>
-<style>
-  :root {
-    --bg:             #0b0e14;
-    --bg-surface:     #11151c;
-    --bg-elevated:    #181d27;
-    --bg-code:        #0d1018;
-    --fg:             #c9d1d9;
-    --fg-bright:      #e6edf3;
-    --fg-muted:       #6e7681;
-    --fg-subtle:      #484f58;
-    --accent:         #7eb8f6;
-    --accent-dim:     #3d6ea5;
-    --accent-glow:    rgba(126, 184, 246, 0.08);
-    --green:          #7ee6a8;
-    --green-dim:      #2ea04f;
-    --orange:         #e6a855;
-    --red:            #f47067;
-    --purple:         #bc8cff;
-    --cyan:           #56d4dd;
-    --border:         #21262d;
-    --border-subtle:  #161b22;
-    --radius:         6px;
-    --font-sans:      'New York', ui-serif, 'Iowan Old Style', 'Apple Garamond', Baskerville, 'Times New Roman', 'Noto Emoji', serif;
-    --font-mono:      'Departure Mono', 'Noto Emoji', monospace;
-  }
-
-  *, *::before, *::after { box-sizing: border-box; margin: 0; padding: 0; }
-  html { scroll-behavior: smooth; scroll-padding-top: 2rem; }
-  body {
-    font-family: var(--font-sans);
-    background: var(--bg);
-    color: var(--fg);
-    line-height: 1.7;
-    font-size: 15px;
-    -webkit-font-smoothing: antialiased;
-  }
-
-  .container { max-width: 860px; margin: 0 auto; padding: 3rem 2rem 6rem; }
-
-  .hero {
-    text-align: center;
-    padding: 4rem 0 3rem;
-    border-bottom: 1px solid var(--border);
-    margin-bottom: 3rem;
-  }
-  .hero h1 { font-family: var(--font-mono); font-size: 2.2rem; font-weight: 700; color: var(--fg-bright); letter-spacing: -0.03em; margin-bottom: 0.5rem; }
-  .hero h1 span { color: var(--accent); }
-  .hero .subtitle { font-family: var(--font-sans); color: var(--fg-muted); font-size: 0.92rem; max-width: 560px; margin: 0 auto; line-height: 1.6; }
-  .hero .meta { margin-top: 1.5rem; display: flex; justify-content: center; gap: 1.5rem; flex-wrap: wrap; }
-  .hero .meta span { font-size: 0.8rem; color: var(--fg-subtle); font-family: var(--font-mono); }
-
-  .toc { background: var(--bg-surface); border: 1px solid var(--border); border-radius: var(--radius); padding: 1.5rem 2rem; margin-bottom: 3rem; }
-  .toc h2 { font-size: 0.75rem; text-transform: uppercase; letter-spacing: 0.1em; color: var(--fg-muted); margin-bottom: 1rem; }
-  .toc ol { list-style: none; counter-reset: toc; columns: 2; column-gap: 2rem; }
-  .toc li { counter-increment: toc; break-inside: avoid; margin-bottom: 0.35rem; }
-  .toc li::before { content: counter(toc, decimal-leading-zero) " "; color: var(--fg-subtle); font-family: var(--font-mono); font-size: 0.75rem; margin-right: 0.25rem; }
-  .toc a { font-family: var(--font-mono); color: var(--fg); text-decoration: none; font-size: 0.82rem; transition: color 0.15s; }
-  .toc a:hover { color: var(--accent); }
-
-  section { margin-bottom: 4rem; }
-  section + section { padding-top: 1rem; }
-
-  h2 { font-family: var(--font-mono); font-size: 1.3rem; font-weight: 700; color: var(--fg-bright); letter-spacing: -0.01em; margin-bottom: 1.25rem; padding-bottom: 0.5rem; border-bottom: 1px solid var(--border); }
-  h3 { font-family: var(--font-mono); font-size: 1rem; font-weight: 600; color: var(--fg-bright); margin-top: 2rem; margin-bottom: 0.75rem; }
-  h4 { font-family: var(--font-mono); font-size: 0.9rem; font-weight: 600; color: var(--accent); margin-top: 1.5rem; margin-bottom: 0.5rem; }
-
-  p { margin-bottom: 1rem; font-size: 0.95rem; line-height: 1.75; }
-  strong { color: var(--fg-bright); font-weight: 600; }
-  a { color: var(--accent); text-decoration: none; }
-  a:hover { text-decoration: underline; }
-
-  ul, ol { margin-bottom: 1rem; padding-left: 1.5rem; font-size: 0.93rem; line-height: 1.7; }
-  li { margin-bottom: 0.35rem; }
-  li::marker { color: var(--fg-subtle); }
-
-  .table-wrap { overflow-x: auto; margin-bottom: 1.5rem; }
-  table { width: 100%; border-collapse: collapse; font-size: 0.88rem; }
-  th, td { text-align: left; padding: 0.6rem 1rem; border-bottom: 1px solid var(--border-subtle); }
-  th { font-family: var(--font-mono); font-size: 0.72rem; text-transform: uppercase; letter-spacing: 0.06em; color: var(--fg-muted); background: var(--bg-surface); border-bottom-color: var(--border); white-space: nowrap; }
-  td { font-family: var(--font-sans); font-size: 0.88rem; color: var(--fg); }
-  tr:hover td { background: var(--accent-glow); }
-  td code { background: var(--bg-elevated); padding: 0.15em 0.4em; border-radius: 3px; font-family: var(--font-mono); font-size: 0.82em; color: var(--cyan); }
-
-  pre { background: var(--bg-code); border: 1px solid var(--border); border-radius: var(--radius); padding: 1.25rem 1.5rem; overflow-x: auto; margin-bottom: 1.5rem; font-family: var(--font-mono); font-size: 0.82rem; line-height: 1.65; color: var(--fg); }
-  pre code { background: none; padding: 0; color: inherit; font-size: inherit; }
-  code { font-family: var(--font-mono); font-size: 0.85em; }
-  p code, li code { background: var(--bg-elevated); padding: 0.15em 0.4em; border-radius: 3px; color: var(--cyan); font-size: 0.85em; }
-
-  .kw { color: var(--purple); }
-  .str { color: var(--green); }
-  .cm { color: var(--fg-subtle); font-style: italic; }
-  .num { color: var(--orange); }
-  .key { color: var(--accent); }
-
-  .mermaid { margin: 1.5rem 0 2rem; text-align: center; }
-  .mermaid svg { max-width: 100%; height: auto; }
-
-  .callout { font-family: var(--font-sans); background: var(--bg-surface); border-left: 3px solid var(--accent-dim); border-radius: 0 var(--radius) var(--radius) 0; padding: 1rem 1.25rem; margin-bottom: 1.5rem; font-size: 0.88rem; color: var(--fg-muted); line-height: 1.6; }
-  .callout strong { font-family: var(--font-mono); color: var(--fg-bright); }
-  .callout.success { border-left-color: var(--green-dim); }
-  .callout.warn { border-left-color: var(--orange); }
-
-  .badge { display: inline-block; font-family: var(--font-mono); font-size: 0.65rem; font-weight: 600; text-transform: uppercase; letter-spacing: 0.05em; padding: 0.2em 0.6em; border-radius: 3px; vertical-align: middle; margin-left: 0.4rem; }
-  .badge-done { background: var(--green-dim); color: #fff; }
-  .badge-wip { background: var(--orange); color: #0b0e14; }
-  .badge-todo { background: var(--fg-subtle); color: var(--fg); }
-
-  .checklist { list-style: none; padding-left: 0; }
-  .checklist li { padding-left: 1.5rem; position: relative; margin-bottom: 0.5rem; }
-  .checklist li::before { position: absolute; left: 0; font-family: var(--font-mono); font-size: 0.85rem; }
-  .checklist li.done { color: var(--fg-muted); }
-  .checklist li.done::before { content: "\2713"; color: var(--green); }
-  .checklist li.todo::before { content: "\25CB"; color: var(--fg-subtle); }
-  .checklist li.wip::before { content: "\25D4"; color: var(--orange); }
-
-  .compare { display: grid; grid-template-columns: 1fr 1fr; gap: 1rem; margin-bottom: 2rem; }
-  .compare-card { background: var(--bg-surface); border: 1px solid var(--border); border-radius: var(--radius); padding: 1.25rem; }
-  .compare-card h4 { margin-top: 0; font-size: 0.82rem; }
-  .compare-card.after { border-color: var(--accent-dim); }
-  .compare-card ul { font-family: var(--font-mono); padding-left: 1.25rem; font-size: 0.8rem; }
-
-  hr { border: none; border-top: 1px solid var(--border); margin: 3rem 0; }
-
-  .progress-bar { position: fixed; top: 0; left: 0; height: 2px; background: var(--accent); z-index: 999; transition: width 0.1s linear; }
-
-  @media (max-width: 640px) {
-    .container { padding: 2rem 1rem 4rem; }
-    .hero h1 { font-size: 1.6rem; }
-    .toc ol { columns: 1; }
-    .compare { grid-template-columns: 1fr; }
-    table { font-size: 0.8rem; }
-    th, td { padding: 0.4rem 0.6rem; }
-  }
-</style>
-<link rel="preconnect" href="https://fonts.googleapis.com">
-<link href="https://fonts.googleapis.com/css2?family=Noto+Emoji&display=swap" rel="stylesheet">
-<style>
-  @font-face {
-    font-family: 'Departure Mono';
-    src: url('https://cdn.jsdelivr.net/gh/rektdeckard/departure-mono@latest/fonts/DepartureMono-Regular.woff2') format('woff2');
-    font-weight: normal;
-    font-style: normal;
-    font-display: swap;
-  }
-</style>
-</head>
-<body>
-
-<div class="progress-bar" id="progress"></div>
-
-<div class="container">
-
-<header class="hero">
-  <h1>honcho<span>-integration-spec</span></h1>
-  <p class="subtitle">Comparison of Hermes Agent vs. openclaw-honcho — and a porting spec for bringing Hermes patterns into other Honcho integrations.</p>
-  <div class="meta">
-    <span>hermes-agent / openclaw-honcho</span>
-    <span>Python + TypeScript</span>
-    <span>2026-03-09</span>
-  </div>
-</header>
-
-<nav class="toc">
-  <h2>Contents</h2>
-  <ol>
-    <li><a href="#overview">Overview</a></li>
-    <li><a href="#architecture">Architecture comparison</a></li>
-    <li><a href="#diff-table">Diff table</a></li>
-    <li><a href="#patterns">Hermes patterns to port</a></li>
-    <li><a href="#spec-async">Spec: async prefetch</a></li>
-    <li><a href="#spec-reasoning">Spec: dynamic reasoning level</a></li>
-    <li><a href="#spec-modes">Spec: per-peer memory modes</a></li>
-    <li><a href="#spec-identity">Spec: AI peer identity formation</a></li>
-    <li><a href="#spec-sessions">Spec: session naming strategies</a></li>
-    <li><a href="#spec-cli">Spec: CLI surface injection</a></li>
-    <li><a href="#openclaw-checklist">openclaw-honcho checklist</a></li>
-    <li><a href="#nanobot-checklist">nanobot-honcho checklist</a></li>
-  </ol>
-</nav>
-
-<!-- OVERVIEW -->
-<section id="overview">
-  <h2>Overview</h2>
-
-  <p>Two independent Honcho integrations have been built for two different agent runtimes: <strong>Hermes Agent</strong> (Python, baked into the runner) and <strong>openclaw-honcho</strong> (TypeScript plugin via hook/tool API). Both use the same Honcho peer paradigm — dual peer model, <code>session.context()</code>, <code>peer.chat()</code> — but they made different tradeoffs at every layer.</p>
-
-  <p>This document maps those tradeoffs and defines a porting spec: a set of Hermes-originated patterns, each stated as an integration-agnostic interface, that any Honcho integration can adopt regardless of runtime or language.</p>
-
-  <div class="callout">
-    <strong>Scope</strong> Both integrations work correctly today. This spec is about the delta — patterns in Hermes that are worth propagating and patterns in openclaw-honcho that Hermes should eventually adopt. The spec is additive, not prescriptive.
-  </div>
-</section>
-
-<!-- ARCHITECTURE -->
-<section id="architecture">
-  <h2>Architecture comparison</h2>
-
-  <h3>Hermes: baked-in runner</h3>
-  <p>Honcho is initialised directly inside <code>AIAgent.__init__</code>. There is no plugin boundary. Session management, context injection, async prefetch, and CLI surface are all first-class concerns of the runner. Context is injected once per session (baked into <code>_cached_system_prompt</code>) and never re-fetched mid-session — this maximises prefix cache hits at the LLM provider.</p>
-
-  <div class="mermaid">
-%%{init: {'theme': 'dark', 'themeVariables': { 'primaryColor': '#1f3150', 'primaryTextColor': '#c9d1d9', 'primaryBorderColor': '#3d6ea5', 'lineColor': '#3d6ea5', 'secondaryColor': '#162030', 'tertiaryColor': '#11151c' }}}%%
-flowchart TD
-    U["user message"] --> P["_honcho_prefetch()<br/>(reads cache — no HTTP)"]
-    P --> SP["_build_system_prompt()<br/>(first turn only, cached)"]
-    SP --> LLM["LLM call"]
-    LLM --> R["response"]
-    R --> FP["_honcho_fire_prefetch()<br/>(daemon threads, turn end)"]
-    FP --> C1["prefetch_context() thread"]
-    FP --> C2["prefetch_dialectic() thread"]
-    C1 --> CACHE["_context_cache / _dialectic_cache"]
-    C2 --> CACHE
-
-    style U fill:#162030,stroke:#3d6ea5,color:#c9d1d9
-    style P fill:#1f3150,stroke:#3d6ea5,color:#c9d1d9
-    style SP fill:#1f3150,stroke:#3d6ea5,color:#c9d1d9
-    style LLM fill:#162030,stroke:#3d6ea5,color:#c9d1d9
-    style R fill:#162030,stroke:#3d6ea5,color:#c9d1d9
-    style FP fill:#2a1a40,stroke:#bc8cff,color:#c9d1d9
-    style C1 fill:#2a1a40,stroke:#bc8cff,color:#c9d1d9
-    style C2 fill:#2a1a40,stroke:#bc8cff,color:#c9d1d9
-    style CACHE fill:#11151c,stroke:#484f58,color:#6e7681
-  </div>
-
-  <h3>openclaw-honcho: hook-based plugin</h3>
-  <p>The plugin registers hooks against OpenClaw's event bus. Context is fetched synchronously inside <code>before_prompt_build</code> on every turn. Message capture happens in <code>agent_end</code>. The multi-agent hierarchy is tracked via <code>subagent_spawned</code>. This model is correct but every turn pays a blocking Honcho round-trip before the LLM call can begin.</p>
-
-  <div class="mermaid">
-%%{init: {'theme': 'dark', 'themeVariables': { 'primaryColor': '#1f3150', 'primaryTextColor': '#c9d1d9', 'primaryBorderColor': '#3d6ea5', 'lineColor': '#3d6ea5', 'secondaryColor': '#162030', 'tertiaryColor': '#11151c' }}}%%
-flowchart TD
-    U2["user message"] --> BPB["before_prompt_build<br/>(BLOCKING HTTP — every turn)"]
-    BPB --> CTX["session.context()"]
-    CTX --> SP2["system prompt assembled"]
-    SP2 --> LLM2["LLM call"]
-    LLM2 --> R2["response"]
-    R2 --> AE["agent_end hook"]
-    AE --> SAVE["session.addMessages()<br/>session.setMetadata()"]
-
-    style U2 fill:#162030,stroke:#3d6ea5,color:#c9d1d9
-    style BPB fill:#3a1515,stroke:#f47067,color:#c9d1d9
-    style CTX fill:#3a1515,stroke:#f47067,color:#c9d1d9
-    style SP2 fill:#1f3150,stroke:#3d6ea5,color:#c9d1d9
-    style LLM2 fill:#162030,stroke:#3d6ea5,color:#c9d1d9
-    style R2 fill:#162030,stroke:#3d6ea5,color:#c9d1d9
-    style AE fill:#162030,stroke:#3d6ea5,color:#c9d1d9
-    style SAVE fill:#11151c,stroke:#484f58,color:#6e7681
-  </div>
-</section>
-
-<!-- DIFF TABLE -->
-<section id="diff-table">
-  <h2>Diff table</h2>
-
-  <div class="table-wrap">
-    <table>
-      <thead>
-        <tr>
-          <th>Dimension</th>
-          <th>Hermes Agent</th>
-          <th>openclaw-honcho</th>
-        </tr>
-      </thead>
-      <tbody>
-        <tr>
-          <td><strong>Context injection timing</strong></td>
-          <td>Once per session (cached). Zero HTTP on response path after turn 1.</td>
-          <td>Every turn, blocking. Fresh context per turn but adds latency.</td>
-        </tr>
-        <tr>
-          <td><strong>Prefetch strategy</strong></td>
-          <td>Daemon threads fire at turn end; consumed next turn from cache.</td>
-          <td>None. Blocking call at prompt-build time.</td>
-        </tr>
-        <tr>
-          <td><strong>Dialectic (peer.chat)</strong></td>
-          <td>Prefetched async; result injected into system prompt next turn.</td>
-          <td>On-demand via <code>honcho_recall</code> / <code>honcho_analyze</code> tools.</td>
-        </tr>
-        <tr>
-          <td><strong>Reasoning level</strong></td>
-          <td>Dynamic: scales with message length. Floor = config default. Cap = "high".</td>
-          <td>Fixed per tool: recall=minimal, analyze=medium.</td>
-        </tr>
-        <tr>
-          <td><strong>Memory modes</strong></td>
-          <td><code>user_memory_mode</code> / <code>agent_memory_mode</code>: hybrid / honcho / local.</td>
-          <td>None. Always writes to Honcho.</td>
-        </tr>
-        <tr>
-          <td><strong>Write frequency</strong></td>
-          <td>async (background queue), turn, session, N turns.</td>
-          <td>After every agent_end (no control).</td>
-        </tr>
-        <tr>
-          <td><strong>AI peer identity</strong></td>
-          <td><code>observe_me=True</code>, <code>seed_ai_identity()</code>, <code>get_ai_representation()</code>, SOUL.md → AI peer.</td>
-          <td>Agent files uploaded to agent peer at setup. No ongoing self-observation seeding.</td>
-        </tr>
-        <tr>
-          <td><strong>Context scope</strong></td>
-          <td>User peer + AI peer representation, both injected.</td>
-          <td>User peer (owner) representation + conversation summary. <code>peerPerspective</code> on context call.</td>
-        </tr>
-        <tr>
-          <td><strong>Session naming</strong></td>
-          <td>per-directory / global / manual map / title-based.</td>
-          <td>Derived from platform session key.</td>
-        </tr>
-        <tr>
-          <td><strong>Multi-agent</strong></td>
-          <td>Single-agent only.</td>
-          <td>Parent observer hierarchy via <code>subagent_spawned</code>.</td>
-        </tr>
-        <tr>
-          <td><strong>Tool surface</strong></td>
-          <td>Single <code>query_user_context</code> tool (on-demand dialectic).</td>
-          <td>6 tools: session, profile, search, context (fast) + recall, analyze (LLM).</td>
-        </tr>
-        <tr>
-          <td><strong>Platform metadata</strong></td>
-          <td>Not stripped.</td>
-          <td>Explicitly stripped before Honcho storage.</td>
-        </tr>
-        <tr>
-          <td><strong>Message dedup</strong></td>
-          <td>None (sends on every save cycle).</td>
-          <td><code>lastSavedIndex</code> in session metadata prevents re-sending.</td>
-        </tr>
-        <tr>
-          <td><strong>CLI surface in prompt</strong></td>
-          <td>Management commands injected into system prompt. Agent knows its own CLI.</td>
-          <td>Not injected.</td>
-        </tr>
-        <tr>
-          <td><strong>AI peer name in identity</strong></td>
-          <td>Replaces "Hermes Agent" in DEFAULT_AGENT_IDENTITY when configured.</td>
-          <td>Not implemented.</td>
-        </tr>
-        <tr>
-          <td><strong>QMD / local file search</strong></td>
-          <td>Not implemented.</td>
-          <td>Passthrough tools when QMD backend configured.</td>
-        </tr>
-        <tr>
-          <td><strong>Workspace metadata</strong></td>
-          <td>Not implemented.</td>
-          <td><code>agentPeerMap</code> in workspace metadata tracks agent&#8594;peer ID.</td>
-        </tr>
-      </tbody>
-    </table>
-  </div>
-</section>
-
-<!-- PATTERNS -->
-<section id="patterns">
-  <h2>Hermes patterns to port</h2>
-
-  <p>Six patterns from Hermes are worth adopting in any Honcho integration. They are described below as integration-agnostic interfaces — the implementation will differ per runtime, but the contract is the same.</p>
-
-  <div class="compare">
-    <div class="compare-card">
-      <h4>Patterns Hermes contributes</h4>
-      <ul>
-        <li>Async prefetch (zero-latency)</li>
-        <li>Dynamic reasoning level</li>
-        <li>Per-peer memory modes</li>
-        <li>AI peer identity formation</li>
-        <li>Session naming strategies</li>
-        <li>CLI surface injection</li>
-      </ul>
-    </div>
-    <div class="compare-card after">
-      <h4>Patterns openclaw contributes back</h4>
-      <ul>
-        <li>lastSavedIndex dedup</li>
-        <li>Platform metadata stripping</li>
-        <li>Multi-agent observer hierarchy</li>
-        <li>peerPerspective on context()</li>
-        <li>Tiered tool surface (fast/LLM)</li>
-        <li>Workspace agentPeerMap</li>
-      </ul>
-    </div>
-  </div>
-</section>
-
-<!-- SPEC: ASYNC PREFETCH -->
-<section id="spec-async">
-  <h2>Spec: async prefetch</h2>
-
-  <h3>Problem</h3>
-  <p>Calling <code>session.context()</code> and <code>peer.chat()</code> synchronously before each LLM call adds 200–800ms of Honcho round-trip latency to every turn. Users experience this as the agent "thinking slowly."</p>
-
-  <h3>Pattern</h3>
-  <p>Fire both calls as non-blocking background work at the <strong>end</strong> of each turn. Store results in a per-session cache keyed by session ID. At the <strong>start</strong> of the next turn, pop from cache — the HTTP is already done. First turn is cold (empty cache); all subsequent turns are zero-latency on the response path.</p>
-
-  <h3>Interface contract</h3>
-  <pre><code><span class="cm">// TypeScript (openclaw / nanobot plugin shape)</span>
-
-<span class="kw">interface</span> <span class="key">AsyncPrefetch</span> {
-  <span class="cm">// Fire context + dialectic fetches at turn end. Non-blocking.</span>
-  firePrefetch(sessionId: <span class="str">string</span>, userMessage: <span class="str">string</span>): <span class="kw">void</span>;
-
-  <span class="cm">// Pop cached results at turn start. Returns empty if cache is cold.</span>
-  popContextResult(sessionId: <span class="str">string</span>): ContextResult | <span class="kw">null</span>;
-  popDialecticResult(sessionId: <span class="str">string</span>): <span class="str">string</span> | <span class="kw">null</span>;
-}
-
-<span class="kw">type</span> <span class="key">ContextResult</span> = {
-  representation: <span class="str">string</span>;
-  card: <span class="str">string</span>[];
-  aiRepresentation?: <span class="str">string</span>;  <span class="cm">// AI peer context if enabled</span>
-  summary?: <span class="str">string</span>;            <span class="cm">// conversation summary if fetched</span>
-};</code></pre>
-
-  <h3>Implementation notes</h3>
-  <ul>
-    <li>Python: <code>threading.Thread(daemon=True)</code>. Write to <code>dict[session_id, result]</code> — GIL makes this safe for simple writes.</li>
-    <li>TypeScript: <code>Promise</code> stored in <code>Map&lt;string, Promise&lt;ContextResult&gt;&gt;</code>. Await at pop time. If not resolved yet, skip (return null) — do not block.</li>
-    <li>The pop is destructive: clears the cache entry after reading so stale data never accumulates.</li>
-    <li>Prefetch should also fire on first turn (even though it won't be consumed until turn 2) — this ensures turn 2 is never cold.</li>
-  </ul>
-
-  <h3>openclaw-honcho adoption</h3>
-  <p>Move <code>session.context()</code> from <code>before_prompt_build</code> to a post-<code>agent_end</code> background task. Store result in <code>state.contextCache</code>. In <code>before_prompt_build</code>, read from cache instead of calling Honcho. If cache is empty (turn 1), inject nothing — the prompt is still valid without Honcho context on the first turn.</p>
-</section>
-
-<!-- SPEC: DYNAMIC REASONING LEVEL -->
-<section id="spec-reasoning">
-  <h2>Spec: dynamic reasoning level</h2>
-
-  <h3>Problem</h3>
-  <p>Honcho's dialectic endpoint supports reasoning levels from <code>minimal</code> to <code>max</code>. A fixed level per tool wastes budget on simple queries and under-serves complex ones.</p>
-
-  <h3>Pattern</h3>
-  <p>Select the reasoning level dynamically based on the user's message. Use the configured default as a floor. Bump by message length. Cap auto-selection at <code>high</code> — never select <code>max</code> automatically.</p>
-
-  <h3>Interface contract</h3>
-  <pre><code><span class="cm">// Shared helper — identical logic in any language</span>
-
-<span class="kw">const</span> LEVELS = [<span class="str">"minimal"</span>, <span class="str">"low"</span>, <span class="str">"medium"</span>, <span class="str">"high"</span>, <span class="str">"max"</span>];
-
-<span class="kw">function</span> <span class="key">dynamicReasoningLevel</span>(
-  query: <span class="str">string</span>,
-  configDefault: <span class="str">string</span> = <span class="str">"low"</span>
-): <span class="str">string</span> {
-  <span class="kw">const</span> baseIdx = Math.max(<span class="num">0</span>, LEVELS.indexOf(configDefault));
-  <span class="kw">const</span> n = query.length;
-  <span class="kw">const</span> bump = n &lt; <span class="num">120</span> ? <span class="num">0</span> : n &lt; <span class="num">400</span> ? <span class="num">1</span> : <span class="num">2</span>;
-  <span class="kw">return</span> LEVELS[Math.min(baseIdx + bump, <span class="num">3</span>)]; <span class="cm">// cap at "high" (idx 3)</span>
-}</code></pre>
-
-  <h3>Config key</h3>
-  <p>Add a <code>dialecticReasoningLevel</code> config field (string, default <code>"low"</code>). This sets the floor. Users can raise or lower it. The dynamic bump always applies on top.</p>
-
-  <h3>openclaw-honcho adoption</h3>
-  <p>Apply in <code>honcho_recall</code> and <code>honcho_analyze</code>: replace the fixed <code>reasoningLevel</code> with the dynamic selector. <code>honcho_recall</code> should use floor <code>"minimal"</code> and <code>honcho_analyze</code> floor <code>"medium"</code> — both still bump with message length.</p>
-</section>
-
-<!-- SPEC: PER-PEER MEMORY MODES -->
-<section id="spec-modes">
-  <h2>Spec: per-peer memory modes</h2>
-
-  <h3>Problem</h3>
-  <p>Users want independent control over whether user context and agent context are written locally, to Honcho, or both. A single <code>memoryMode</code> shorthand is not granular enough.</p>
-
-  <h3>Pattern</h3>
-  <p>Three modes per peer: <code>hybrid</code> (write both local + Honcho), <code>honcho</code> (Honcho only, disable local files), <code>local</code> (local files only, skip Honcho sync for this peer). Two orthogonal axes: user peer and agent peer.</p>
-
-  <h3>Config schema</h3>
-  <pre><code><span class="cm">// ~/.openclaw/openclaw.json  (or ~/.nanobot/config.json)</span>
-{
-  <span class="str">"plugins"</span>: {
-    <span class="str">"openclaw-honcho"</span>: {
-      <span class="str">"config"</span>: {
-        <span class="str">"apiKey"</span>: <span class="str">"..."</span>,
-        <span class="str">"memoryMode"</span>: <span class="str">"hybrid"</span>,          <span class="cm">// shorthand: both peers</span>
-        <span class="str">"userMemoryMode"</span>: <span class="str">"honcho"</span>,       <span class="cm">// override for user peer</span>
-        <span class="str">"agentMemoryMode"</span>: <span class="str">"hybrid"</span>       <span class="cm">// override for agent peer</span>
-      }
-    }
-  }
-}</code></pre>
-
-  <h3>Resolution order</h3>
-  <ol>
-    <li>Per-peer field (<code>userMemoryMode</code> / <code>agentMemoryMode</code>) — wins if present.</li>
-    <li>Shorthand <code>memoryMode</code> — applies to both peers as default.</li>
-    <li>Hardcoded default: <code>"hybrid"</code>.</li>
-  </ol>
-
-  <h3>Effect on Honcho sync</h3>
-  <ul>
-    <li><code>userMemoryMode=local</code>: skip adding user peer messages to Honcho.</li>
-    <li><code>agentMemoryMode=local</code>: skip adding assistant peer messages to Honcho.</li>
-    <li>Both local: skip <code>session.addMessages()</code> entirely.</li>
-    <li><code>userMemoryMode=honcho</code>: disable local USER.md writes.</li>
-    <li><code>agentMemoryMode=honcho</code>: disable local MEMORY.md / SOUL.md writes.</li>
-  </ul>
-</section>
-
-<!-- SPEC: AI PEER IDENTITY -->
-<section id="spec-identity">
-  <h2>Spec: AI peer identity formation</h2>
-
-  <h3>Problem</h3>
-  <p>Honcho builds the user's representation organically by observing what the user says. The same mechanism exists for the AI peer — but only if <code>observe_me=True</code> is set for the agent peer. Without it, the agent peer accumulates nothing and Honcho's AI-side model never forms.</p>
-
-  <p>Additionally, existing persona files (SOUL.md, IDENTITY.md) should seed the AI peer's Honcho representation at first activation, rather than waiting for it to emerge from scratch.</p>
-
-  <h3>Part A: observe_me=True for agent peer</h3>
-  <pre><code><span class="cm">// TypeScript — in session.addPeers() call</span>
-<span class="kw">await</span> session.addPeers([
-  [ownerPeer.id, { observeMe: <span class="kw">true</span>,  observeOthers: <span class="kw">false</span> }],
-  [agentPeer.id, { observeMe: <span class="kw">true</span>,  observeOthers: <span class="kw">true</span>  }], <span class="cm">// was false</span>
-]);</code></pre>
-
-  <p>This is a one-line change but foundational. Without it, Honcho's AI peer representation stays empty regardless of what the agent says.</p>
-
-  <h3>Part B: seedAiIdentity()</h3>
-  <pre><code><span class="kw">async function</span> <span class="key">seedAiIdentity</span>(
-  session: HonchoSession,
-  agentPeer: Peer,
-  content: <span class="str">string</span>,
-  source: <span class="str">string</span>
-): Promise&lt;<span class="kw">boolean</span>&gt; {
-  <span class="kw">const</span> wrapped = [
-    <span class="str">`&lt;ai_identity_seed&gt;`</span>,
-    <span class="str">`&lt;source&gt;${source}&lt;/source&gt;`</span>,
-    <span class="str">``</span>,
-    content.trim(),
-    <span class="str">`&lt;/ai_identity_seed&gt;`</span>,
-  ].join(<span class="str">"\n"</span>);
-
-  <span class="kw">await</span> agentPeer.addMessage(<span class="str">"assistant"</span>, wrapped);
-  <span class="kw">return true</span>;
-}</code></pre>
-
-  <h3>Part C: migrate agent files at setup</h3>
-  <p>During <code>openclaw honcho setup</code>, upload agent-self files (SOUL.md, IDENTITY.md, AGENTS.md, BOOTSTRAP.md) to the agent peer using <code>seedAiIdentity()</code> instead of <code>session.uploadFile()</code>. This routes the content through Honcho's observation pipeline rather than the file store.</p>
-
-  <h3>Part D: AI peer name in identity</h3>
-  <p>When the agent has a configured name (non-default), inject it into the agent's self-identity prefix. In OpenClaw this means adding to the injected system prompt section:</p>
-  <pre><code><span class="cm">// In context hook return value</span>
-<span class="kw">return</span> {
-  systemPrompt: [
-    agentName ? <span class="str">`You are ${agentName}.`</span> : <span class="str">""</span>,
-    <span class="str">"## User Memory Context"</span>,
-    ...sections,
-  ].filter(Boolean).join(<span class="str">"\n\n"</span>)
-};</code></pre>
-
-  <h3>CLI surface: honcho identity subcommand</h3>
-  <pre><code>openclaw honcho identity &lt;file&gt;    <span class="cm"># seed from file</span>
-openclaw honcho identity --show    <span class="cm"># show current AI peer representation</span></code></pre>
-</section>
-
-<!-- SPEC: SESSION NAMING -->
-<section id="spec-sessions">
-  <h2>Spec: session naming strategies</h2>
-
-  <h3>Problem</h3>
-  <p>When Honcho is used across multiple projects or directories, a single global session means every project shares the same context. Per-directory sessions provide isolation without requiring users to name sessions manually.</p>
-
-  <h3>Strategies</h3>
-  <div class="table-wrap">
-    <table>
-      <thead><tr><th>Strategy</th><th>Session key</th><th>When to use</th></tr></thead>
-      <tbody>
-        <tr><td><code>per-directory</code></td><td>basename of CWD</td><td>Default. Each project gets its own session.</td></tr>
-        <tr><td><code>global</code></td><td>fixed string <code>"global"</code></td><td>Single cross-project session.</td></tr>
-        <tr><td>manual map</td><td>user-configured per path</td><td><code>sessions</code> config map overrides directory basename.</td></tr>
-        <tr><td>title-based</td><td>sanitized session title</td><td>When agent supports named sessions; title set mid-conversation.</td></tr>
-      </tbody>
-    </table>
-  </div>
-
-  <h3>Config schema</h3>
-  <pre><code>{
-  <span class="str">"sessionStrategy"</span>: <span class="str">"per-directory"</span>,   <span class="cm">// "per-directory" | "global"</span>
-  <span class="str">"sessionPeerPrefix"</span>: <span class="kw">false</span>,            <span class="cm">// prepend peer name to session key</span>
-  <span class="str">"sessions"</span>: {                            <span class="cm">// manual overrides</span>
-    <span class="str">"/home/user/projects/foo"</span>: <span class="str">"foo-project"</span>
-  }
-}</code></pre>
-
-  <h3>CLI surface</h3>
-  <pre><code>openclaw honcho sessions              <span class="cm"># list all mappings</span>
-openclaw honcho map &lt;name&gt;           <span class="cm"># map cwd to session name</span>
-openclaw honcho map                   <span class="cm"># no-arg = list mappings</span></code></pre>
-
-  <p>Resolution order: manual map wins &rarr; session title &rarr; directory basename &rarr; platform key.</p>
-</section>
-
-<!-- SPEC: CLI SURFACE INJECTION -->
-<section id="spec-cli">
-  <h2>Spec: CLI surface injection</h2>
-
-  <h3>Problem</h3>
-  <p>When a user asks "how do I change my memory settings?" or "what Honcho commands are available?" the agent either hallucinates or says it doesn't know. The agent should know its own management interface.</p>
-
-  <h3>Pattern</h3>
-  <p>When Honcho is active, append a compact command reference to the system prompt. The agent can cite these commands directly instead of guessing.</p>
-
-  <pre><code><span class="cm">// In context hook, append to systemPrompt</span>
-<span class="kw">const</span> honchoSection = [
-  <span class="str">"# Honcho memory integration"</span>,
-  <span class="str">`Active. Session: ${sessionKey}. Mode: ${mode}.`</span>,
-  <span class="str">"Management commands:"</span>,
-  <span class="str">"  openclaw honcho status                    — show config + connection"</span>,
-  <span class="str">"  openclaw honcho mode [hybrid|honcho|local] — show or set memory mode"</span>,
-  <span class="str">"  openclaw honcho sessions                  — list session mappings"</span>,
-  <span class="str">"  openclaw honcho map &lt;name&gt;                — map directory to session"</span>,
-  <span class="str">"  openclaw honcho identity [file] [--show]  — seed or show AI identity"</span>,
-  <span class="str">"  openclaw honcho setup                     — full interactive wizard"</span>,
-].join(<span class="str">"\n"</span>);</code></pre>
-
-  <div class="callout warn">
-    <strong>Keep it compact.</strong> This section is injected every turn. Keep it under 300 chars of context. List commands, not explanations — the agent can explain them on request.
-  </div>
-</section>
-
-<!-- OPENCLAW CHECKLIST -->
-<section id="openclaw-checklist">
-  <h2>openclaw-honcho checklist</h2>
-
-  <p>Ordered by impact. Each item maps to a spec section above.</p>
-
-  <ul class="checklist">
-    <li class="todo"><strong>Async prefetch</strong> — move <code>session.context()</code> out of <code>before_prompt_build</code> into post-<code>agent_end</code> background Promise. Pop from cache at prompt build. (<a href="#spec-async">spec</a>)</li>
-    <li class="todo"><strong>observe_me=True for agent peer</strong> — one-line change in <code>session.addPeers()</code> config for agent peer. (<a href="#spec-identity">spec</a>)</li>
-    <li class="todo"><strong>Dynamic reasoning level</strong> — add <code>dynamicReasoningLevel()</code> helper; apply in <code>honcho_recall</code> and <code>honcho_analyze</code>. Add <code>dialecticReasoningLevel</code> to config schema. (<a href="#spec-reasoning">spec</a>)</li>
-    <li class="todo"><strong>Per-peer memory modes</strong> — add <code>userMemoryMode</code> / <code>agentMemoryMode</code> to config; gate Honcho sync and local writes accordingly. (<a href="#spec-modes">spec</a>)</li>
-    <li class="todo"><strong>seedAiIdentity()</strong> — add helper; apply during setup migration for SOUL.md / IDENTITY.md instead of <code>session.uploadFile()</code>. (<a href="#spec-identity">spec</a>)</li>
-    <li class="todo"><strong>Session naming strategies</strong> — add <code>sessionStrategy</code>, <code>sessions</code> map, <code>sessionPeerPrefix</code> to config; implement resolution function. (<a href="#spec-sessions">spec</a>)</li>
-    <li class="todo"><strong>CLI surface injection</strong> — append command reference to <code>before_prompt_build</code> return value when Honcho is active. (<a href="#spec-cli">spec</a>)</li>
-    <li class="todo"><strong>honcho identity subcommand</strong> — add <code>openclaw honcho identity</code> CLI command. (<a href="#spec-identity">spec</a>)</li>
-    <li class="todo"><strong>AI peer name injection</strong> — if <code>aiPeer</code> name configured, prepend to injected system prompt. (<a href="#spec-identity">spec</a>)</li>
-    <li class="todo"><strong>honcho mode / honcho sessions / honcho map</strong> — CLI parity with Hermes. (<a href="#spec-sessions">spec</a>)</li>
-  </ul>
-
-  <div class="callout success">
-    <strong>Already done in openclaw-honcho (do not re-implement):</strong> lastSavedIndex dedup, platform metadata stripping, multi-agent parent observer hierarchy, peerPerspective on context(), tiered tool surface (fast/LLM), workspace agentPeerMap, QMD passthrough, self-hosted Honcho support.
-  </div>
-</section>
-
-<!-- NANOBOT CHECKLIST -->
-<section id="nanobot-checklist">
-  <h2>nanobot-honcho checklist</h2>
-
-  <p>nanobot-honcho is a greenfield integration. Start from openclaw-honcho's architecture (hook-based, dual peer) and apply all Hermes patterns from day one rather than retrofitting. Priority order:</p>
-
-  <h3>Phase 1 — core correctness</h3>
-  <ul class="checklist">
-    <li class="todo">Dual peer model (owner + agent peer), both with <code>observe_me=True</code></li>
-    <li class="todo">Message capture at turn end with <code>lastSavedIndex</code> dedup</li>
-    <li class="todo">Platform metadata stripping before Honcho storage</li>
-    <li class="todo">Async prefetch from day one — do not implement blocking context injection</li>
-    <li class="todo">Legacy file migration at first activation (USER.md → owner peer, SOUL.md → <code>seedAiIdentity()</code>)</li>
-  </ul>
-
-  <h3>Phase 2 — configuration</h3>
-  <ul class="checklist">
-    <li class="todo">Config schema: <code>apiKey</code>, <code>workspaceId</code>, <code>baseUrl</code>, <code>memoryMode</code>, <code>userMemoryMode</code>, <code>agentMemoryMode</code>, <code>dialecticReasoningLevel</code>, <code>sessionStrategy</code>, <code>sessions</code></li>
-    <li class="todo">Per-peer memory mode gating</li>
-    <li class="todo">Dynamic reasoning level</li>
-    <li class="todo">Session naming strategies</li>
-  </ul>
-
-  <h3>Phase 3 — tools and CLI</h3>
-  <ul class="checklist">
-    <li class="todo">Tool surface: <code>honcho_profile</code>, <code>honcho_recall</code>, <code>honcho_analyze</code>, <code>honcho_search</code>, <code>honcho_context</code></li>
-    <li class="todo">CLI: <code>setup</code>, <code>status</code>, <code>sessions</code>, <code>map</code>, <code>mode</code>, <code>identity</code></li>
-    <li class="todo">CLI surface injection into system prompt</li>
-    <li class="todo">AI peer name wired into agent identity</li>
-  </ul>
-</section>
-
-</div>
-
-<script type="module">
-  import mermaid from 'https://cdn.jsdelivr.net/npm/mermaid@11/dist/mermaid.esm.min.mjs';
-  mermaid.initialize({ startOnLoad: true, securityLevel: 'loose', fontFamily: 'Departure Mono, Noto Emoji, monospace' });
-</script>
-<script>
-  window.addEventListener('scroll', () => {
-    const bar = document.getElementById('progress');
-    const max = document.documentElement.scrollHeight - window.innerHeight;
-    bar.style.width = (max > 0 ? (window.scrollY / max) * 100 : 0) + '%';
-  });
-</script>
-</body>
-</html>

docs/honcho-integration-spec.md

@@ -1,377 +0,0 @@
-# honcho-integration-spec
-
-Comparison of Hermes Agent vs. openclaw-honcho — and a porting spec for bringing Hermes patterns into other Honcho integrations.
-
----
-
-## Overview
-
-Two independent Honcho integrations have been built for two different agent runtimes: **Hermes Agent** (Python, baked into the runner) and **openclaw-honcho** (TypeScript plugin via hook/tool API). Both use the same Honcho peer paradigm — dual peer model, `session.context()`, `peer.chat()` — but they made different tradeoffs at every layer.
-
-This document maps those tradeoffs and defines a porting spec: a set of Hermes-originated patterns, each stated as an integration-agnostic interface, that any Honcho integration can adopt regardless of runtime or language.
-
-> **Scope** Both integrations work correctly today. This spec is about the delta — patterns in Hermes that are worth propagating and patterns in openclaw-honcho that Hermes should eventually adopt. The spec is additive, not prescriptive.
-
----
-
-## Architecture comparison
-
-### Hermes: baked-in runner
-
-Honcho is initialised directly inside `AIAgent.__init__`. There is no plugin boundary. Session management, context injection, async prefetch, and CLI surface are all first-class concerns of the runner. Context is injected once per session (baked into `_cached_system_prompt`) and never re-fetched mid-session — this maximises prefix cache hits at the LLM provider.
-
-Turn flow:
-
-```
-user message
-  → _honcho_prefetch()       (reads cache — no HTTP)
-  → _build_system_prompt()   (first turn only, cached)
-  → LLM call
-  → response
-  → _honcho_fire_prefetch()  (daemon threads, turn end)
-       → prefetch_context() thread  ──┐
-       → prefetch_dialectic() thread ─┴→ _context_cache / _dialectic_cache
-```
-
-### openclaw-honcho: hook-based plugin
-
-The plugin registers hooks against OpenClaw's event bus. Context is fetched synchronously inside `before_prompt_build` on every turn. Message capture happens in `agent_end`. The multi-agent hierarchy is tracked via `subagent_spawned`. This model is correct but every turn pays a blocking Honcho round-trip before the LLM call can begin.
-
-Turn flow:
-
-```
-user message
-  → before_prompt_build (BLOCKING HTTP — every turn)
-       → session.context()
-  → system prompt assembled
-  → LLM call
-  → response
-  → agent_end hook
-       → session.addMessages()
-       → session.setMetadata()
-```
-
----
-
-## Diff table
-
-| Dimension | Hermes Agent | openclaw-honcho |
-|---|---|---|
-| **Context injection timing** | Once per session (cached). Zero HTTP on response path after turn 1. | Every turn, blocking. Fresh context per turn but adds latency. |
-| **Prefetch strategy** | Daemon threads fire at turn end; consumed next turn from cache. | None. Blocking call at prompt-build time. |
-| **Dialectic (peer.chat)** | Prefetched async; result injected into system prompt next turn. | On-demand via `honcho_recall` / `honcho_analyze` tools. |
-| **Reasoning level** | Dynamic: scales with message length. Floor = config default. Cap = "high". | Fixed per tool: recall=minimal, analyze=medium. |
-| **Memory modes** | `user_memory_mode` / `agent_memory_mode`: hybrid / honcho / local. | None. Always writes to Honcho. |
-| **Write frequency** | async (background queue), turn, session, N turns. | After every agent_end (no control). |
-| **AI peer identity** | `observe_me=True`, `seed_ai_identity()`, `get_ai_representation()`, SOUL.md → AI peer. | Agent files uploaded to agent peer at setup. No ongoing self-observation. |
-| **Context scope** | User peer + AI peer representation, both injected. | User peer (owner) representation + conversation summary. `peerPerspective` on context call. |
-| **Session naming** | per-directory / global / manual map / title-based. | Derived from platform session key. |
-| **Multi-agent** | Single-agent only. | Parent observer hierarchy via `subagent_spawned`. |
-| **Tool surface** | Single `query_user_context` tool (on-demand dialectic). | 6 tools: session, profile, search, context (fast) + recall, analyze (LLM). |
-| **Platform metadata** | Not stripped. | Explicitly stripped before Honcho storage. |
-| **Message dedup** | None. | `lastSavedIndex` in session metadata prevents re-sending. |
-| **CLI surface in prompt** | Management commands injected into system prompt. Agent knows its own CLI. | Not injected. |
-| **AI peer name in identity** | Replaces "Hermes Agent" in DEFAULT_AGENT_IDENTITY when configured. | Not implemented. |
-| **QMD / local file search** | Not implemented. | Passthrough tools when QMD backend configured. |
-| **Workspace metadata** | Not implemented. | `agentPeerMap` in workspace metadata tracks agent→peer ID. |
-
----
-
-## Patterns
-
-Six patterns from Hermes are worth adopting in any Honcho integration. Each is described as an integration-agnostic interface.
-
-**Hermes contributes:**
-- Async prefetch (zero-latency)
-- Dynamic reasoning level
-- Per-peer memory modes
-- AI peer identity formation
-- Session naming strategies
-- CLI surface injection
-
-**openclaw-honcho contributes back (Hermes should adopt):**
-- `lastSavedIndex` dedup
-- Platform metadata stripping
-- Multi-agent observer hierarchy
-- `peerPerspective` on `context()`
-- Tiered tool surface (fast/LLM)
-- Workspace `agentPeerMap`
-
----
-
-## Spec: async prefetch
-
-### Problem
-
-Calling `session.context()` and `peer.chat()` synchronously before each LLM call adds 200–800ms of Honcho round-trip latency to every turn.
-
-### Pattern
-
-Fire both calls as non-blocking background work at the **end** of each turn. Store results in a per-session cache keyed by session ID. At the **start** of the next turn, pop from cache — the HTTP is already done. First turn is cold (empty cache); all subsequent turns are zero-latency on the response path.
-
-### Interface contract
-
-```typescript
-interface AsyncPrefetch {
-  // Fire context + dialectic fetches at turn end. Non-blocking.
-  firePrefetch(sessionId: string, userMessage: string): void;
-
-  // Pop cached results at turn start. Returns empty if cache is cold.
-  popContextResult(sessionId: string): ContextResult | null;
-  popDialecticResult(sessionId: string): string | null;
-}
-
-type ContextResult = {
-  representation: string;
-  card: string[];
-  aiRepresentation?: string;  // AI peer context if enabled
-  summary?: string;           // conversation summary if fetched
-};
-```
-
-### Implementation notes
-
-- **Python:** `threading.Thread(daemon=True)`. Write to `dict[session_id, result]` — GIL makes this safe for simple writes.
-- **TypeScript:** `Promise` stored in `Map<string, Promise<ContextResult>>`. Await at pop time. If not resolved yet, return null — do not block.
-- The pop is destructive: clears the cache entry after reading so stale data never accumulates.
-- Prefetch should also fire on first turn (even though it won't be consumed until turn 2).
-
-### openclaw-honcho adoption
-
-Move `session.context()` from `before_prompt_build` to a post-`agent_end` background task. Store result in `state.contextCache`. In `before_prompt_build`, read from cache instead of calling Honcho. If cache is empty (turn 1), inject nothing — the prompt is still valid without Honcho context on the first turn.
-
----
-
-## Spec: dynamic reasoning level
-
-### Problem
-
-Honcho's dialectic endpoint supports reasoning levels from `minimal` to `max`. A fixed level per tool wastes budget on simple queries and under-serves complex ones.
-
-### Pattern
-
-Select the reasoning level dynamically based on the user's message. Use the configured default as a floor. Bump by message length. Cap auto-selection at `high` — never select `max` automatically.
-
-### Logic
-
-```
-< 120 chars  → default (typically "low")
-120–400 chars → one level above default (cap at "high")
-> 400 chars  → two levels above default (cap at "high")
-```
-
-### Config key
-
-Add `dialecticReasoningLevel` (string, default `"low"`). This sets the floor. The dynamic bump always applies on top.
-
-### openclaw-honcho adoption
-
-Apply in `honcho_recall` and `honcho_analyze`: replace fixed `reasoningLevel` with the dynamic selector. `honcho_recall` uses floor `"minimal"`, `honcho_analyze` uses floor `"medium"` — both still bump with message length.
-
----
-
-## Spec: per-peer memory modes
-
-### Problem
-
-Users want independent control over whether user context and agent context are written locally, to Honcho, or both.
-
-### Modes
-
-| Mode | Effect |
-|---|---|
-| `hybrid` | Write to both local files and Honcho (default) |
-| `honcho` | Honcho only — disable corresponding local file writes |
-| `local` | Local files only — skip Honcho sync for this peer |
-
-### Config schema
-
-```json
-{
-  "memoryMode": "hybrid",
-  "userMemoryMode": "honcho",
-  "agentMemoryMode": "hybrid"
-}
-```
-
-Resolution order: per-peer field wins → shorthand `memoryMode` → default `"hybrid"`.
-
-### Effect on Honcho sync
-
-- `userMemoryMode=local`: skip adding user peer messages to Honcho
-- `agentMemoryMode=local`: skip adding assistant peer messages to Honcho
-- Both local: skip `session.addMessages()` entirely
-- `userMemoryMode=honcho`: disable local USER.md writes
-- `agentMemoryMode=honcho`: disable local MEMORY.md / SOUL.md writes
-
----
-
-## Spec: AI peer identity formation
-
-### Problem
-
-Honcho builds the user's representation organically by observing what the user says. The same mechanism exists for the AI peer — but only if `observe_me=True` is set for the agent peer. Without it, the agent peer accumulates nothing.
-
-Additionally, existing persona files (SOUL.md, IDENTITY.md) should seed the AI peer's Honcho representation at first activation.
-
-### Part A: observe_me=True for agent peer
-
-```typescript
-await session.addPeers([
-  [ownerPeer.id, { observeMe: true,  observeOthers: false }],
-  [agentPeer.id, { observeMe: true,  observeOthers: true  }], // was false
-]);
-```
-
-One-line change. Foundational. Without it, the AI peer representation stays empty regardless of what the agent says.
-
-### Part B: seedAiIdentity()
-
-```typescript
-async function seedAiIdentity(
-  agentPeer: Peer,
-  content: string,
-  source: string
-): Promise<boolean> {
-  const wrapped = [
-    `<ai_identity_seed>`,
-    `<source>${source}</source>`,
-    ``,
-    content.trim(),
-    `</ai_identity_seed>`,
-  ].join("\n");
-
-  await agentPeer.addMessage("assistant", wrapped);
-  return true;
-}
-```
-
-### Part C: migrate agent files at setup
-
-During `honcho setup`, upload agent-self files (SOUL.md, IDENTITY.md, AGENTS.md) to the agent peer via `seedAiIdentity()` instead of `session.uploadFile()`. This routes content through Honcho's observation pipeline.
-
-### Part D: AI peer name in identity
-
-When the agent has a configured name, prepend it to the injected system prompt:
-
-```typescript
-const namePrefix = agentName ? `You are ${agentName}.\n\n` : "";
-return { systemPrompt: namePrefix + "## User Memory Context\n\n" + sections };
-```
-
-### CLI surface
-
-```
-honcho identity <file>    # seed from file
-honcho identity --show    # show current AI peer representation
-```
-
----
-
-## Spec: session naming strategies
-
-### Problem
-
-A single global session means every project shares the same Honcho context. Per-directory sessions provide isolation without requiring users to name sessions manually.
-
-### Strategies
-
-| Strategy | Session key | When to use |
-|---|---|---|
-| `per-directory` | basename of CWD | Default. Each project gets its own session. |
-| `global` | fixed string `"global"` | Single cross-project session. |
-| manual map | user-configured per path | `sessions` config map overrides directory basename. |
-| title-based | sanitized session title | When agent supports named sessions set mid-conversation. |
-
-### Config schema
-
-```json
-{
-  "sessionStrategy": "per-directory",
-  "sessionPeerPrefix": false,
-  "sessions": {
-    "/home/user/projects/foo": "foo-project"
-  }
-}
-```
-
-### CLI surface
-
-```
-honcho sessions              # list all mappings
-honcho map <name>            # map cwd to session name
-honcho map                   # no-arg = list mappings
-```
-
-Resolution order: manual map → session title → directory basename → platform key.
-
----
-
-## Spec: CLI surface injection
-
-### Problem
-
-When a user asks "how do I change my memory settings?" the agent either hallucinates or says it doesn't know. The agent should know its own management interface.
-
-### Pattern
-
-When Honcho is active, append a compact command reference to the system prompt. Keep it under 300 chars.
-
-```
-# Honcho memory integration
-Active. Session: {sessionKey}. Mode: {mode}.
-Management commands:
-  honcho status                    — show config + connection
-  honcho mode [hybrid|honcho|local] — show or set memory mode
-  honcho sessions                  — list session mappings
-  honcho map <name>                — map directory to session
-  honcho identity [file] [--show]  — seed or show AI identity
-  honcho setup                     — full interactive wizard
-```
-
----
-
-## openclaw-honcho checklist
-
-Ordered by impact:
-
-- [ ] **Async prefetch** — move `session.context()` out of `before_prompt_build` into post-`agent_end` background Promise
-- [ ] **observe_me=True for agent peer** — one-line change in `session.addPeers()`
-- [ ] **Dynamic reasoning level** — add helper; apply in `honcho_recall` and `honcho_analyze`; add `dialecticReasoningLevel` to config
-- [ ] **Per-peer memory modes** — add `userMemoryMode` / `agentMemoryMode` to config; gate Honcho sync and local writes
-- [ ] **seedAiIdentity()** — add helper; use during setup migration for SOUL.md / IDENTITY.md
-- [ ] **Session naming strategies** — add `sessionStrategy`, `sessions` map, `sessionPeerPrefix`
-- [ ] **CLI surface injection** — append command reference to `before_prompt_build` return value
-- [ ] **honcho identity subcommand** — seed from file or `--show` current representation
-- [ ] **AI peer name injection** — if `aiPeer` name configured, prepend to injected system prompt
-- [ ] **honcho mode / sessions / map** — CLI parity with Hermes
-
-Already done in openclaw-honcho (do not re-implement): `lastSavedIndex` dedup, platform metadata stripping, multi-agent parent observer, `peerPerspective` on `context()`, tiered tool surface, workspace `agentPeerMap`, QMD passthrough, self-hosted Honcho.
-
----
-
-## nanobot-honcho checklist
-
-Greenfield integration. Start from openclaw-honcho's architecture and apply all Hermes patterns from day one.
-
-### Phase 1 — core correctness
-
-- [ ] Dual peer model (owner + agent peer), both with `observe_me=True`
-- [ ] Message capture at turn end with `lastSavedIndex` dedup
-- [ ] Platform metadata stripping before Honcho storage
-- [ ] Async prefetch from day one — do not implement blocking context injection
-- [ ] Legacy file migration at first activation (USER.md → owner peer, SOUL.md → `seedAiIdentity()`)
-
-### Phase 2 — configuration
-
-- [ ] Config schema: `apiKey`, `workspaceId`, `baseUrl`, `memoryMode`, `userMemoryMode`, `agentMemoryMode`, `dialecticReasoningLevel`, `sessionStrategy`, `sessions`
-- [ ] Per-peer memory mode gating
-- [ ] Dynamic reasoning level
-- [ ] Session naming strategies
-
-### Phase 3 — tools and CLI
-
-- [ ] Tool surface: `honcho_profile`, `honcho_recall`, `honcho_analyze`, `honcho_search`, `honcho_context`
-- [ ] CLI: `setup`, `status`, `sessions`, `map`, `mode`, `identity`
-- [ ] CLI surface injection into system prompt
-- [ ] AI peer name wired into agent identity

docs/migration/openclaw.md

@@ -1,142 +0,0 @@
-# Migrating from OpenClaw to Hermes Agent
-
-This guide covers how to import your OpenClaw settings, memories, skills, and API keys into Hermes Agent.
-
-## Three Ways to Migrate
-
-### 1. Automatic (during first-time setup)
-
-When you run `hermes setup` for the first time and Hermes detects `~/.openclaw`, it automatically offers to import your OpenClaw data before configuration begins. Just accept the prompt and everything is handled for you.
-
-### 2. CLI Command (quick, scriptable)
-
-```bash
-hermes claw migrate                      # Preview then migrate (always shows preview first)
-hermes claw migrate --dry-run            # Preview only, no changes
-hermes claw migrate --preset user-data   # Migrate without API keys/secrets
-hermes claw migrate --yes                # Skip confirmation prompt
-```
-
-The migration always shows a full preview of what will be imported before making any changes. You review the preview and confirm before anything is written.
-
-**All options:**
-
-| Flag | Description |
-|------|-------------|
-| `--source PATH` | Path to OpenClaw directory (default: `~/.openclaw`) |
-| `--dry-run` | Preview only — no files are modified |
-| `--preset {user-data,full}` | Migration preset (default: `full`). `user-data` excludes secrets |
-| `--overwrite` | Overwrite existing files (default: skip conflicts) |
-| `--migrate-secrets` | Include allowlisted secrets (auto-enabled with `full` preset) |
-| `--workspace-target PATH` | Copy workspace instructions (AGENTS.md) to this absolute path |
-| `--skill-conflict {skip,overwrite,rename}` | How to handle skill name conflicts (default: `skip`) |
-| `--yes`, `-y` | Skip confirmation prompts |
-
-### 3. Agent-Guided (interactive, with previews)
-
-Ask the agent to run the migration for you:
-
-```
-> Migrate my OpenClaw setup to Hermes
-```
-
-The agent will use the `openclaw-migration` skill to:
-1. Run a preview first to show what would change
-2. Ask about conflict resolution (SOUL.md, skills, etc.)
-3. Let you choose between `user-data` and `full` presets
-4. Execute the migration with your choices
-5. Print a detailed summary of what was migrated
-
-## What Gets Migrated
-
-### `user-data` preset
-| Item | Source | Destination |
-|------|--------|-------------|
-| SOUL.md | `~/.openclaw/workspace/SOUL.md` | `~/.hermes/SOUL.md` |
-| Memory entries | `~/.openclaw/workspace/MEMORY.md` | `~/.hermes/memories/MEMORY.md` |
-| User profile | `~/.openclaw/workspace/USER.md` | `~/.hermes/memories/USER.md` |
-| Skills | `~/.openclaw/workspace/skills/` | `~/.hermes/skills/openclaw-imports/` |
-| Command allowlist | `~/.openclaw/workspace/exec_approval_patterns.yaml` | Merged into `~/.hermes/config.yaml` |
-| Messaging settings | `~/.openclaw/config.yaml` (TELEGRAM_ALLOWED_USERS, MESSAGING_CWD) | `~/.hermes/.env` |
-| TTS assets | `~/.openclaw/workspace/tts/` | `~/.hermes/tts/` |
-
-Workspace files are also checked at `workspace.default/` and `workspace-main/` as fallback paths (OpenClaw renamed `workspace/` to `workspace-main/` in recent versions).
-
-### `full` preset (adds to `user-data`)
-| Item | Source | Destination |
-|------|--------|-------------|
-| Telegram bot token | `openclaw.json` channels config | `~/.hermes/.env` |
-| OpenRouter API key | `.env`, `openclaw.json`, or `openclaw.json["env"]` | `~/.hermes/.env` |
-| OpenAI API key | `.env`, `openclaw.json`, or `openclaw.json["env"]` | `~/.hermes/.env` |
-| Anthropic API key | `.env`, `openclaw.json`, or `openclaw.json["env"]` | `~/.hermes/.env` |
-| ElevenLabs API key | `.env`, `openclaw.json`, or `openclaw.json["env"]` | `~/.hermes/.env` |
-
-API keys are searched across four sources: inline config values, `~/.openclaw/.env`, the `openclaw.json` `"env"` sub-object, and per-agent auth profiles.
-
-Only allowlisted secrets are ever imported. Other credentials are skipped and reported.
-
-## OpenClaw Schema Compatibility
-
-The migration handles both old and current OpenClaw config layouts:
-
-- **Channel tokens**: Reads from flat paths (`channels.telegram.botToken`) and the newer `accounts.default` layout (`channels.telegram.accounts.default.botToken`)
-- **TTS provider**: OpenClaw renamed "edge" to "microsoft" — both are recognized and mapped to Hermes' "edge"
-- **Provider API types**: Both short (`openai`, `anthropic`) and hyphenated (`openai-completions`, `anthropic-messages`, `google-generative-ai`) values are mapped correctly
-- **thinkingDefault**: All enum values are handled including newer ones (`minimal`, `xhigh`, `adaptive`)
-- **Matrix**: Uses `accessToken` field (not `botToken`)
-- **SecretRef formats**: Plain strings, env templates (`${VAR}`), and `source: "env"` SecretRefs are resolved. `source: "file"` and `source: "exec"` SecretRefs produce a warning — add those keys manually after migration.
-
-## Conflict Handling
-
-By default, the migration **will not overwrite** existing Hermes data:
-
-- **SOUL.md** — skipped if one already exists in `~/.hermes/`
-- **Memory entries** — skipped if memories already exist (to avoid duplicates)
-- **Skills** — skipped if a skill with the same name already exists
-- **API keys** — skipped if the key is already set in `~/.hermes/.env`
-
-To overwrite conflicts, use `--overwrite`. The migration creates backups before overwriting.
-
-For skills, you can also use `--skill-conflict rename` to import conflicting skills under a new name (e.g., `skill-name-imported`).
-
-## Migration Report
-
-Every migration produces a report showing:
-- **Migrated items** — what was successfully imported
-- **Conflicts** — items skipped because they already exist
-- **Skipped items** — items not found in the source
-- **Errors** — items that failed to import
-
-For executed migrations, the full report is saved to `~/.hermes/migration/openclaw/<timestamp>/`.
-
-## Post-Migration Notes
-
-- **Skills require a new session** — imported skills take effect after restarting your agent or starting a new chat.
-- **WhatsApp requires re-pairing** — WhatsApp uses QR-code pairing, not token-based auth. Run `hermes whatsapp` to pair.
-- **Archive cleanup** — after migration, you'll be offered to rename `~/.openclaw/` to `.openclaw.pre-migration/` to prevent state confusion. You can also run `hermes claw cleanup` later.
-
-## Troubleshooting
-
-### "OpenClaw directory not found"
-The migration looks for `~/.openclaw` by default, then tries `~/.clawdbot` and `~/.moltbot`. If your OpenClaw is installed elsewhere, use `--source`:
-```bash
-hermes claw migrate --source /path/to/.openclaw
-```
-
-### "Migration script not found"
-The migration script ships with Hermes Agent. If you installed via pip (not git clone), the `optional-skills/` directory may not be present. Install the skill from the Skills Hub:
-```bash
-hermes skills install openclaw-migration
-```
-
-### Memory overflow
-If your OpenClaw MEMORY.md or USER.md exceeds Hermes' character limits, excess entries are exported to an overflow file in the migration report directory. You can manually review and add the most important ones.
-
-### API keys not found
-Keys might be stored in different places depending on your OpenClaw setup:
-- `~/.openclaw/.env` file
-- Inline in `openclaw.json` under `models.providers.*.apiKey`
-- In `openclaw.json` under the `"env"` or `"env.vars"` sub-objects
-- In `~/.openclaw/agents/main/agent/auth-profiles.json`
-
-The migration checks all four. If keys use `source: "file"` or `source: "exec"` SecretRefs, they can't be resolved automatically — add them via `hermes config set`.

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

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

docs/skins/example-skin.yaml

@@ -1,97 +0,0 @@
-# ============================================================================
-# Hermes Agent — Example Skin Template
-# ============================================================================
-#
-# Copy this file to ~/.hermes/skins/<name>.yaml to create a custom skin.
-# All fields are optional — missing values inherit from the default skin.
-# Activate with: /skin <name>  or  display.skin: <name> in config.yaml
-#
-# See hermes_cli/skin_engine.py for the full schema reference.
-# ============================================================================
-
-# Required: unique skin name (used in /skin command and config)
-name: example
-description: An example custom skin — copy and modify this template
-
-# ── Colors ──────────────────────────────────────────────────────────────────
-# Hex color values for Rich markup. These control the CLI's visual palette.
-colors:
-  # Banner panel (the startup welcome box)
-  banner_border: "#CD7F32"        # Panel border
-  banner_title: "#FFD700"         # Panel title text
-  banner_accent: "#FFBF00"        # Section headers (Available Tools, Skills, etc.)
-  banner_dim: "#B8860B"           # Dim/muted text (separators, model info)
-  banner_text: "#FFF8DC"          # Body text (tool names, skill names)
-
-  # UI elements
-  ui_accent: "#FFBF00"            # General accent color
-  ui_label: "#4dd0e1"             # Labels
-  ui_ok: "#4caf50"                # Success indicators
-  ui_error: "#ef5350"             # Error indicators
-  ui_warn: "#ffa726"              # Warning indicators
-
-  # Input area
-  prompt: "#FFF8DC"               # Prompt text color
-  input_rule: "#CD7F32"           # Horizontal rule around input
-
-  # Response box
-  response_border: "#FFD700"      # Response box border (ANSI color)
-
-  # Session display
-  session_label: "#DAA520"        # Session label
-  session_border: "#8B8682"       # Session ID dim color
-
-  # TUI surfaces
-  status_bar_bg: "#1a1a2e"              # Status / usage bar background
-  voice_status_bg: "#1a1a2e"            # Voice-mode badge background
-  completion_menu_bg: "#1a1a2e"         # Completion list background
-  completion_menu_current_bg: "#333355" # Active completion row background
-  completion_menu_meta_bg: "#1a1a2e"    # Completion meta column background
-  completion_menu_meta_current_bg: "#333355"  # Active completion meta background
-
-# ── Spinner ─────────────────────────────────────────────────────────────────
-# Customize the animated spinner shown during API calls and tool execution.
-spinner:
-  # Faces shown while waiting for the API response
-  waiting_faces:
-    - "(｡◕‿◕｡)"
-    - "(◕‿◕✿)"
-    - "٩(◕‿◕｡)۶"
-
-  # Faces shown during extended thinking/reasoning
-  thinking_faces:
-    - "(｡•́︿•̀｡)"
-    - "(◔_◔)"
-    - "(¬‿¬)"
-
-  # Verbs used in spinner messages (e.g., "pondering your request...")
-  thinking_verbs:
-    - "pondering"
-    - "contemplating"
-    - "musing"
-    - "ruminating"
-
-  # Optional: left/right decorations around the spinner
-  # Each entry is a [left, right] pair. Omit entirely for no wings.
-  # wings:
-  #   - ["⟪⚔", "⚔⟫"]
-  #   - ["⟪▲", "▲⟫"]
-
-# ── Branding ────────────────────────────────────────────────────────────────
-# Text strings used throughout the CLI interface.
-branding:
-  agent_name: "Hermes Agent"          # Banner title, about display
-  welcome: "Welcome! Type your message or /help for commands."
-  goodbye: "Goodbye! ⚕"              # Exit message
-  response_label: " ⚕ Hermes "       # Response box header label
-  prompt_symbol: "❯ "                 # Input prompt symbol
-  help_header: "(^_^)? Available Commands"  # /help header text
-
-# ── Tool Output ─────────────────────────────────────────────────────────────
-# Character used as the prefix for tool output lines.
-# Default is "┊" (thin dotted vertical line). Some alternatives:
-#   "╎" (light triple dash vertical)
-#   "▏" (left one-eighth block)
-#   "│" (box drawing light vertical)
-#   "┃" (box drawing heavy vertical)
-tool_prefix: "┊"

docs/specs/container-cli-review-fixes.md

@@ -1,329 +0,0 @@
-# Container-Aware CLI Review Fixes Spec
-
-**PR:** NousResearch/hermes-agent#7543
-**Review:** cursor[bot] bugbot review (4094049442) + two prior rounds
-**Date:** 2026-04-12
-**Branch:** `feat/container-aware-cli-clean`
-
-## Review Issues Summary
-
-Six issues were raised across three bugbot review rounds. Three were fixed in intermediate commits (38277a6a, 726cf90f). This spec addresses remaining design concerns surfaced by those reviews and simplifies the implementation based on interview decisions.
-
-| # | Issue | Severity | Status |
-|---|-------|----------|--------|
-| 1 | `os.execvp` retry loop unreachable | Medium | Fixed in 79e8cd12 (switched to subprocess.run) |
-| 2 | Redundant `shutil.which("sudo")` | Medium | Fixed in 38277a6a (reuses `sudo` var) |
-| 3 | Missing `chown -h` on symlink update | Low | Fixed in 38277a6a |
-| 4 | Container routing after `parse_args()` | High | Fixed in 726cf90f |
-| 5 | Hardcoded `/home/${user}` | Medium | Fixed in 726cf90f |
-| 6 | Group membership not gated on `container.enable` | Low | Fixed in 726cf90f |
-
-The mechanical fixes are in place but the overall design needs revision. The retry loop, error swallowing, and process model have deeper issues than what the bugbot flagged.
-
----
-
-## Spec: Revised `_exec_in_container`
-
-### Design Principles
-
-1. **Let it crash.** No silent fallbacks. If `.container-mode` exists but something goes wrong, the error propagates naturally (Python traceback). The only case where container routing is skipped is when `.container-mode` doesn't exist or `HERMES_DEV=1`.
-2. **No retries.** Probe once for sudo, exec once. If it fails, docker/podman's stderr reaches the user verbatim.
-3. **Completely transparent.** No error wrapping, no prefixes, no spinners. Docker's output goes straight through.
-4. **`os.execvp` on the happy path.** Replace the Python process entirely so there's no idle parent during interactive sessions. Note: `execvp` never returns on success (process is replaced) and raises `OSError` on failure (it does not return a value). The container process's exit code becomes the process exit code by definition — no explicit propagation needed.
-5. **One human-readable exception to "let it crash".** `subprocess.TimeoutExpired` from the sudo probe gets a specific catch with a readable message, since a raw traceback for "your Docker daemon is slow" is confusing. All other exceptions propagate naturally.
-
-### Execution Flow
-
-```
-1. get_container_exec_info()
-   - HERMES_DEV=1 → return None (skip routing)
-   - Inside container → return None (skip routing)
-   - .container-mode doesn't exist → return None (skip routing)
-   - .container-mode exists → parse and return dict
-   - .container-mode exists but malformed/unreadable → LET IT CRASH (no try/except)
-
-2. _exec_in_container(container_info, sys.argv[1:])
-   a. shutil.which(backend) → if None, print "{backend} not found on PATH" and sys.exit(1)
-   b. Sudo probe: subprocess.run([runtime, "inspect", "--format", "ok", container_name], timeout=15)
-      - If succeeds → needs_sudo = False
-      - If fails → try subprocess.run([sudo, "-n", runtime, "inspect", ...], timeout=15)
-        - If succeeds → needs_sudo = True
-        - If fails → print error with sudoers hint (including why -n is required) and sys.exit(1)
-      - If TimeoutExpired → catch specifically, print human-readable message about slow daemon
-   c. Build exec_cmd: [sudo? + runtime, "exec", tty_flags, "-u", exec_user, env_flags, container, hermes_bin, *cli_args]
-   d. os.execvp(exec_cmd[0], exec_cmd)
-      - On success: process is replaced — Python is gone, container exit code IS the process exit code
-      - On OSError: let it crash (natural traceback)
-```
-
-### Changes to `hermes_cli/main.py`
-
-#### `_exec_in_container` — rewrite
-
-Remove:
-- The entire retry loop (`max_retries`, `for attempt in range(...)`)
-- Spinner logic (`"Waiting for container..."`, dots)
-- Exit code classification (125/126/127 handling)
-- `subprocess.run` for the exec call (keep it only for the sudo probe)
-- Special TTY vs non-TTY retry counts
-- The `time` import (no longer needed)
-
-Change:
-- Use `os.execvp(exec_cmd[0], exec_cmd)` as the final call
-- Keep the `subprocess` import only for the sudo probe
-- Keep TTY detection for the `-it` vs `-i` flag
-- Keep env var forwarding (TERM, COLORTERM, LANG, LC_ALL)
-- Keep the sudo probe as-is (it's the one "smart" part)
-- Bump probe `timeout` from 5s to 15s — cold podman on a loaded machine needs headroom
-- Catch `subprocess.TimeoutExpired` specifically on both probe calls — print a readable message about the daemon being unresponsive instead of a raw traceback
-- Expand the sudoers hint error message to explain *why* `-n` (non-interactive) is required: a password prompt would hang the CLI or break piped commands
-
-The function becomes roughly:
-
-```python
-def _exec_in_container(container_info: dict, cli_args: list):
-    """Replace the current process with a command inside the managed container.
-
-    Probes whether sudo is needed (rootful containers), then os.execvp
-    into the container. If exec fails, the OS error propagates naturally.
-    """
-    import shutil
-    import subprocess
-
-    backend = container_info["backend"]
-    container_name = container_info["container_name"]
-    exec_user = container_info["exec_user"]
-    hermes_bin = container_info["hermes_bin"]
-
-    runtime = shutil.which(backend)
-    if not runtime:
-        print(f"Error: {backend} not found on PATH. Cannot route to container.",
-              file=sys.stderr)
-        sys.exit(1)
-
-    # Probe whether we need sudo to see the rootful container.
-    # Timeout is 15s — cold podman on a loaded machine can take a while.
-    # TimeoutExpired is caught specifically for a human-readable message;
-    # all other exceptions propagate naturally.
-    needs_sudo = False
-    sudo = None
-    try:
-        probe = subprocess.run(
-            [runtime, "inspect", "--format", "ok", container_name],
-            capture_output=True, text=True, timeout=15,
-        )
-    except subprocess.TimeoutExpired:
-        print(
-            f"Error: timed out waiting for {backend} to respond.\n"
-            f"The {backend} daemon may be unresponsive or starting up.",
-            file=sys.stderr,
-        )
-        sys.exit(1)
-
-    if probe.returncode != 0:
-        sudo = shutil.which("sudo")
-        if sudo:
-            try:
-                probe2 = subprocess.run(
-                    [sudo, "-n", runtime, "inspect", "--format", "ok", container_name],
-                    capture_output=True, text=True, timeout=15,
-                )
-            except subprocess.TimeoutExpired:
-                print(
-                    f"Error: timed out waiting for sudo {backend} to respond.",
-                    file=sys.stderr,
-                )
-                sys.exit(1)
-
-            if probe2.returncode == 0:
-                needs_sudo = True
-            else:
-                print(
-                    f"Error: container '{container_name}' not found via {backend}.\n"
-                    f"\n"
-                    f"The NixOS service runs the container as root. Your user cannot\n"
-                    f"see it because {backend} uses per-user namespaces.\n"
-                    f"\n"
-                    f"Fix: grant passwordless sudo for {backend}. The -n (non-interactive)\n"
-                    f"flag is required because the CLI calls sudo non-interactively —\n"
-                    f"a password prompt would hang or break piped commands:\n"
-                    f"\n"
-                    f'  security.sudo.extraRules = [{{\n'
-                    f'    users = [ "{os.getenv("USER", "your-user")}" ];\n'
-                    f'    commands = [{{ command = "{runtime}"; options = [ "NOPASSWD" ]; }}];\n'
-                    f'  }}];\n'
-                    f"\n"
-                    f"Or run: sudo hermes {' '.join(cli_args)}",
-                    file=sys.stderr,
-                )
-                sys.exit(1)
-        else:
-            print(
-                f"Error: container '{container_name}' not found via {backend}.\n"
-                f"The container may be running under root. Try: sudo hermes {' '.join(cli_args)}",
-                file=sys.stderr,
-            )
-            sys.exit(1)
-
-    is_tty = sys.stdin.isatty()
-    tty_flags = ["-it"] if is_tty else ["-i"]
-
-    env_flags = []
-    for var in ("TERM", "COLORTERM", "LANG", "LC_ALL"):
-        val = os.environ.get(var)
-        if val:
-            env_flags.extend(["-e", f"{var}={val}"])
-
-    cmd_prefix = [sudo, "-n", runtime] if needs_sudo else [runtime]
-    exec_cmd = (
-        cmd_prefix + ["exec"]
-        + tty_flags
-        + ["-u", exec_user]
-        + env_flags
-        + [container_name, hermes_bin]
-        + cli_args
-    )
-
-    # execvp replaces this process entirely — it never returns on success.
-    # On failure it raises OSError, which propagates naturally.
-    os.execvp(exec_cmd[0], exec_cmd)
-```
-
-#### Container routing call site in `main()` — remove try/except
-
-Current:
-```python
-try:
-    from hermes_cli.config import get_container_exec_info
-    container_info = get_container_exec_info()
-    if container_info:
-        _exec_in_container(container_info, sys.argv[1:])
-        sys.exit(1)  # exec failed if we reach here
-except SystemExit:
-    raise
-except Exception:
-    pass  # Container routing unavailable, proceed locally
-```
-
-Revised:
-```python
-from hermes_cli.config import get_container_exec_info
-container_info = get_container_exec_info()
-if container_info:
-    _exec_in_container(container_info, sys.argv[1:])
-    # Unreachable: os.execvp never returns on success (process is replaced)
-    # and raises OSError on failure (which propagates as a traceback).
-    # This line exists only as a defensive assertion.
-    sys.exit(1)
-```
-
-No try/except. If `.container-mode` doesn't exist, `get_container_exec_info()` returns `None` and we skip routing. If it exists but is broken, the exception propagates with a natural traceback.
-
-Note: `sys.exit(1)` after `_exec_in_container` is dead code in all paths — `os.execvp` either replaces the process or raises. It's kept as a belt-and-suspenders assertion with a comment marking it unreachable, not as actual error handling.
-
-### Changes to `hermes_cli/config.py`
-
-#### `get_container_exec_info` — remove inner try/except
-
-Current code catches `(OSError, IOError)` and returns `None`. This silently hides permission errors, corrupt files, etc.
-
-Change: Remove the try/except around file reading. Keep the early returns for `HERMES_DEV=1` and `_is_inside_container()`. The `FileNotFoundError` from `open()` when `.container-mode` doesn't exist should still return `None` (this is the "container mode not enabled" case). All other exceptions propagate.
-
-```python
-def get_container_exec_info() -> Optional[dict]:
-    if os.environ.get("HERMES_DEV") == "1":
-        return None
-    if _is_inside_container():
-        return None
-
-    container_mode_file = get_hermes_home() / ".container-mode"
-
-    try:
-        with open(container_mode_file, "r") as f:
-            # ... parse key=value lines ...
-    except FileNotFoundError:
-        return None
-    # All other exceptions (PermissionError, malformed data, etc.) propagate
-
-    return { ... }
-```
-
----
-
-## Spec: NixOS Module Changes
-
-### Symlink creation — simplify to two branches
-
-Current: 4 branches (symlink exists, directory exists, other file, doesn't exist).
-
-Revised: 2 branches.
-
-```bash
-if [ -d "${symlinkPath}" ] && [ ! -L "${symlinkPath}" ]; then
-  # Real directory — back it up, then create symlink
-  _backup="${symlinkPath}.bak.$(date +%s)"
-  echo "hermes-agent: backing up existing ${symlinkPath} to $_backup"
-  mv "${symlinkPath}" "$_backup"
-fi
-# For everything else (symlink, doesn't exist, etc.) — just force-create
-ln -sfn "${target}" "${symlinkPath}"
-chown -h ${user}:${cfg.group} "${symlinkPath}"
-```
-
-`ln -sfn` handles: existing symlink (replaces), doesn't exist (creates), and after the `mv` above (creates). The only case that needs special handling is a real directory, because `ln -sfn` cannot atomically replace a directory.
-
-Note: there is a theoretical race between the `[ -d ... ]` check and the `mv` (something could create/remove the directory in between). In practice this is a NixOS activation script running as root during `nixos-rebuild switch` — no other process should be touching `~/.hermes` at that moment. Not worth adding locking for.
-
-### Sudoers — document, don't auto-configure
-
-Do NOT add `security.sudo.extraRules` to the module. Document the sudoers requirement in the module's description/comments and in the error message the CLI prints when sudo probe fails.
-
-### Group membership gating — keep as-is
-
-The fix in 726cf90f (`cfg.container.enable && cfg.container.hostUsers != []`) is correct. Leftover group membership when container mode is disabled is harmless. No cleanup needed.
-
----
-
-## Spec: Test Rewrite
-
-The existing test file (`tests/hermes_cli/test_container_aware_cli.py`) has 16 tests. With the simplified exec model, several are obsolete.
-
-### Tests to keep (update as needed)
-
-- `test_is_inside_container_dockerenv` — unchanged
-- `test_is_inside_container_containerenv` — unchanged
-- `test_is_inside_container_cgroup_docker` — unchanged
-- `test_is_inside_container_false_on_host` — unchanged
-- `test_get_container_exec_info_returns_metadata` — unchanged
-- `test_get_container_exec_info_none_inside_container` — unchanged
-- `test_get_container_exec_info_none_without_file` — unchanged
-- `test_get_container_exec_info_skipped_when_hermes_dev` — unchanged
-- `test_get_container_exec_info_not_skipped_when_hermes_dev_zero` — unchanged
-- `test_get_container_exec_info_defaults` — unchanged
-- `test_get_container_exec_info_docker_backend` — unchanged
-
-### Tests to add
-
-- `test_get_container_exec_info_crashes_on_permission_error` — verify that `PermissionError` propagates (no silent `None` return)
-- `test_exec_in_container_calls_execvp` — verify `os.execvp` is called with correct args (runtime, tty flags, user, env, container, binary, cli args)
-- `test_exec_in_container_sudo_probe_sets_prefix` — verify that when first probe fails and sudo probe succeeds, `os.execvp` is called with `sudo -n` prefix
-- `test_exec_in_container_no_runtime_hard_fails` — keep existing, verify `sys.exit(1)` when `shutil.which` returns None
-- `test_exec_in_container_non_tty_uses_i_only` — update to check `os.execvp` args instead of `subprocess.run` args
-- `test_exec_in_container_probe_timeout_prints_message` — verify that `subprocess.TimeoutExpired` from the probe produces a human-readable error and `sys.exit(1)`, not a raw traceback
-- `test_exec_in_container_container_not_running_no_sudo` — verify the path where runtime exists (`shutil.which` returns a path) but probe returns non-zero and no sudo is available. Should print the "container may be running under root" error. This is distinct from `no_runtime_hard_fails` which covers `shutil.which` returning None.
-
-### Tests to delete
-
-- `test_exec_in_container_tty_retries_on_container_failure` — retry loop removed
-- `test_exec_in_container_non_tty_retries_silently_exits_126` — retry loop removed
-- `test_exec_in_container_propagates_hermes_exit_code` — no subprocess.run to check exit codes; execvp replaces the process. Note: exit code propagation still works correctly — when `os.execvp` succeeds, the container's process *becomes* this process, so its exit code is the process exit code by OS semantics. No application code needed, no test needed. A comment in the function docstring documents this intent for future readers.
-
----
-
-## Out of Scope
-
-- Auto-configuring sudoers rules in the NixOS module
-- Any changes to `get_container_exec_info` parsing logic beyond the try/except narrowing
-- Changes to `.container-mode` file format
-- Changes to the `HERMES_DEV=1` bypass
-- Changes to container detection logic (`_is_inside_container`)

flake.lock

@@ -36,6 +36,26 @@
         "type": "github"
       }
     },
+    "npm-lockfile-fix": {
+      "inputs": {
+        "nixpkgs": [
+          "nixpkgs"
+        ]
+      },
+      "locked": {
+        "lastModified": 1775903712,
+        "narHash": "sha256-2GV79U6iVH4gKAPWYrxUReB0S41ty/Y3dBLquU8AlaA=",
+        "owner": "jeslie0",
+        "repo": "npm-lockfile-fix",
+        "rev": "c6093acb0c0548e0f9b8b3d82918823721930fe8",
+        "type": "github"
+      },
+      "original": {
+        "owner": "jeslie0",
+        "repo": "npm-lockfile-fix",
+        "type": "github"
+      }
+    },
     "pyproject-build-systems": {
       "inputs": {
         "nixpkgs": [
@@ -124,6 +144,7 @@
       "inputs": {
         "flake-parts": "flake-parts",
         "nixpkgs": "nixpkgs",
+        "npm-lockfile-fix": "npm-lockfile-fix",
         "pyproject-build-systems": "pyproject-build-systems",
         "pyproject-nix": "pyproject-nix_2",
         "uv2nix": "uv2nix_2"

flake.nix

@@ -19,11 +19,20 @@
       url = "github:pyproject-nix/build-system-pkgs";
       inputs.nixpkgs.follows = "nixpkgs";
     };
+    npm-lockfile-fix = {
+      url = "github:jeslie0/npm-lockfile-fix";
+      inputs.nixpkgs.follows = "nixpkgs";
+    };
   };
 
-  outputs = inputs:
+  outputs =
+    inputs:
     inputs.flake-parts.lib.mkFlake { inherit inputs; } {
-      systems = [ "x86_64-linux" "aarch64-linux" "aarch64-darwin" ];
+      systems = [
+        "x86_64-linux"
+        "aarch64-linux"
+        "aarch64-darwin"
+      ];
 
       imports = [
         ./nix/packages.nix

gateway/channel_directory.py

@@ -100,7 +100,7 @@ def build_channel_directory(adapters: Dict[Any, Any]) -> Dict[str, Any]:
 
 
 def _build_discord(adapter) -> List[Dict[str, str]]:
-    """Enumerate all text channels the Discord bot can see."""
+    """Enumerate all text channels and forum channels the Discord bot can see."""
     channels = []
     client = getattr(adapter, "_client", None)
     if not client:
@@ -119,6 +119,15 @@ def _build_discord(adapter) -> List[Dict[str, str]]:
                 "guild": guild.name,
                 "type": "channel",
             })
+        # Forum channels (type 15) — creating a message auto-spawns a thread post.
+        forums = getattr(guild, "forum_channels", None) or []
+        for ch in forums:
+            channels.append({
+                "id": str(ch.id),
+                "name": ch.name,
+                "guild": guild.name,
+                "type": "forum",
+            })
         # Also include DM-capable users we've interacted with is not
         # feasible via guild enumeration; those come from sessions.
 
@@ -191,6 +200,15 @@ def load_directory() -> Dict[str, Any]:
         return {"updated_at": None, "platforms": {}}
 
 
+def lookup_channel_type(platform_name: str, chat_id: str) -> Optional[str]:
+    """Return the channel ``type`` string (e.g. ``"channel"``, ``"forum"``) for *chat_id*, or *None* if unknown."""
+    directory = load_directory()
+    for ch in directory.get("platforms", {}).get(platform_name, []):
+        if ch.get("id") == chat_id:
+            return ch.get("type")
+    return None
+
+
 def resolve_channel_name(platform_name: str, name: str) -> Optional[str]:
     """
     Resolve a human-friendly channel name to a numeric ID.

gateway/config.py

@@ -258,6 +258,13 @@ class GatewayConfig:
     # Streaming configuration
     streaming: StreamingConfig = field(default_factory=StreamingConfig)
 
+    # Session store pruning: drop SessionEntry records older than this many
+    # days from the in-memory dict and sessions.json.  Keeps the store from
+    # growing unbounded in gateways serving many chats/threads/users over
+    # months.  Pruning is invisible to users — if they resume, they get a
+    # fresh session exactly as if the reset policy had fired.  0 = disabled.
+    session_store_max_age_days: int = 90
+
     def get_connected_platforms(self) -> List[Platform]:
         """Return list of platforms that are enabled and configured."""
         connected = []
@@ -307,6 +314,14 @@ class GatewayConfig:
             # QQBot uses extra dict for app credentials
             elif platform == Platform.QQBOT and config.extra.get("app_id") and config.extra.get("client_secret"):
                 connected.append(platform)
+            # DingTalk uses client_id/client_secret from config.extra or env vars
+            elif platform == Platform.DINGTALK and (
+                config.extra.get("client_id") or os.getenv("DINGTALK_CLIENT_ID")
+            ) and (
+                config.extra.get("client_secret") or os.getenv("DINGTALK_CLIENT_SECRET")
+            ):
+                connected.append(platform)
+        
         return connected
     
     def get_home_channel(self, platform: Platform) -> Optional[HomeChannel]:
@@ -357,6 +372,7 @@ class GatewayConfig:
             "thread_sessions_per_user": self.thread_sessions_per_user,
             "unauthorized_dm_behavior": self.unauthorized_dm_behavior,
             "streaming": self.streaming.to_dict(),
+            "session_store_max_age_days": self.session_store_max_age_days,
         }
     
     @classmethod
@@ -404,6 +420,13 @@ class GatewayConfig:
             "pair",
         )
 
+        try:
+            session_store_max_age_days = int(data.get("session_store_max_age_days", 90))
+            if session_store_max_age_days < 0:
+                session_store_max_age_days = 0
+        except (TypeError, ValueError):
+            session_store_max_age_days = 90
+
         return cls(
             platforms=platforms,
             default_reset_policy=default_policy,
@@ -418,6 +441,7 @@ class GatewayConfig:
             thread_sessions_per_user=_coerce_bool(thread_sessions_per_user, False),
             unauthorized_dm_behavior=unauthorized_dm_behavior,
             streaming=StreamingConfig.from_dict(data.get("streaming", {})),
+            session_store_max_age_days=session_store_max_age_days,
         )
 
     def get_unauthorized_dm_behavior(self, platform: Optional[Platform] = None) -> str:
@@ -554,6 +578,12 @@ def load_gateway_config() -> GatewayConfig:
                     bridged["mention_patterns"] = platform_cfg["mention_patterns"]
                 if plat == Platform.DISCORD and "channel_skill_bindings" in platform_cfg:
                     bridged["channel_skill_bindings"] = platform_cfg["channel_skill_bindings"]
+                if "channel_prompts" in platform_cfg:
+                    channel_prompts = platform_cfg["channel_prompts"]
+                    if isinstance(channel_prompts, dict):
+                        bridged["channel_prompts"] = {str(k): v for k, v in channel_prompts.items()}
+                    else:
+                        bridged["channel_prompts"] = channel_prompts
                 if not bridged:
                     continue
                 plat_data = platforms_data.setdefault(plat.value, {})
@@ -611,6 +641,20 @@ def load_gateway_config() -> GatewayConfig:
                     if isinstance(ntc, list):
                         ntc = ",".join(str(v) for v in ntc)
                     os.environ["DISCORD_NO_THREAD_CHANNELS"] = str(ntc)
+                # allow_mentions: granular control over what the bot can ping.
+                # Safe defaults (no @everyone/roles) are applied in the adapter;
+                # these YAML keys only override when set and let users opt back
+                # into unsafe modes (e.g. roles=true) if they actually want it.
+                allow_mentions_cfg = discord_cfg.get("allow_mentions")
+                if isinstance(allow_mentions_cfg, dict):
+                    for yaml_key, env_key in (
+                        ("everyone", "DISCORD_ALLOW_MENTION_EVERYONE"),
+                        ("roles", "DISCORD_ALLOW_MENTION_ROLES"),
+                        ("users", "DISCORD_ALLOW_MENTION_USERS"),
+                        ("replied_user", "DISCORD_ALLOW_MENTION_REPLIED_USER"),
+                    ):
+                        if yaml_key in allow_mentions_cfg and not os.getenv(env_key):
+                            os.environ[env_key] = str(allow_mentions_cfg[yaml_key]).lower()
 
             # Telegram settings → env vars (env vars take precedence)
             telegram_cfg = yaml_cfg.get("telegram", {})
@@ -632,6 +676,18 @@ def load_gateway_config() -> GatewayConfig:
                     os.environ["TELEGRAM_IGNORED_THREADS"] = str(ignored_threads)
                 if "reactions" in telegram_cfg and not os.getenv("TELEGRAM_REACTIONS"):
                     os.environ["TELEGRAM_REACTIONS"] = str(telegram_cfg["reactions"]).lower()
+                if "proxy_url" in telegram_cfg and not os.getenv("TELEGRAM_PROXY"):
+                    os.environ["TELEGRAM_PROXY"] = str(telegram_cfg["proxy_url"]).strip()
+                if "disable_link_previews" in telegram_cfg:
+                    plat_data = platforms_data.setdefault(Platform.TELEGRAM.value, {})
+                    if not isinstance(plat_data, dict):
+                        plat_data = {}
+                        platforms_data[Platform.TELEGRAM.value] = plat_data
+                    extra = plat_data.setdefault("extra", {})
+                    if not isinstance(extra, dict):
+                        extra = {}
+                        plat_data["extra"] = extra
+                    extra["disable_link_previews"] = telegram_cfg["disable_link_previews"]
 
             whatsapp_cfg = yaml_cfg.get("whatsapp", {})
             if isinstance(whatsapp_cfg, dict):
@@ -645,6 +701,24 @@ def load_gateway_config() -> GatewayConfig:
                         frc = ",".join(str(v) for v in frc)
                     os.environ["WHATSAPP_FREE_RESPONSE_CHATS"] = str(frc)
 
+            # DingTalk settings → env vars (env vars take precedence)
+            dingtalk_cfg = yaml_cfg.get("dingtalk", {})
+            if isinstance(dingtalk_cfg, dict):
+                if "require_mention" in dingtalk_cfg and not os.getenv("DINGTALK_REQUIRE_MENTION"):
+                    os.environ["DINGTALK_REQUIRE_MENTION"] = str(dingtalk_cfg["require_mention"]).lower()
+                if "mention_patterns" in dingtalk_cfg and not os.getenv("DINGTALK_MENTION_PATTERNS"):
+                    os.environ["DINGTALK_MENTION_PATTERNS"] = json.dumps(dingtalk_cfg["mention_patterns"])
+                frc = dingtalk_cfg.get("free_response_chats")
+                if frc is not None and not os.getenv("DINGTALK_FREE_RESPONSE_CHATS"):
+                    if isinstance(frc, list):
+                        frc = ",".join(str(v) for v in frc)
+                    os.environ["DINGTALK_FREE_RESPONSE_CHATS"] = str(frc)
+                allowed = dingtalk_cfg.get("allowed_users")
+                if allowed is not None and not os.getenv("DINGTALK_ALLOWED_USERS"):
+                    if isinstance(allowed, list):
+                        allowed = ",".join(str(v) for v in allowed)
+                    os.environ["DINGTALK_ALLOWED_USERS"] = str(allowed)
+
             # Matrix settings → env vars (env vars take precedence)
             matrix_cfg = yaml_cfg.get("matrix", {})
             if isinstance(matrix_cfg, dict):
@@ -988,6 +1062,25 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
         if webhook_secret:
             config.platforms[Platform.WEBHOOK].extra["secret"] = webhook_secret
 
+    # DingTalk
+    dingtalk_client_id = os.getenv("DINGTALK_CLIENT_ID")
+    dingtalk_client_secret = os.getenv("DINGTALK_CLIENT_SECRET")
+    if dingtalk_client_id and dingtalk_client_secret:
+        if Platform.DINGTALK not in config.platforms:
+            config.platforms[Platform.DINGTALK] = PlatformConfig()
+        config.platforms[Platform.DINGTALK].enabled = True
+        config.platforms[Platform.DINGTALK].extra.update({
+            "client_id": dingtalk_client_id,
+            "client_secret": dingtalk_client_secret,
+        })
+        dingtalk_home = os.getenv("DINGTALK_HOME_CHANNEL")
+        if dingtalk_home:
+            config.platforms[Platform.DINGTALK].home_channel = HomeChannel(
+                platform=Platform.DINGTALK,
+                chat_id=dingtalk_home,
+                name=os.getenv("DINGTALK_HOME_CHANNEL_NAME", "Home"),
+            )
+
     # Feishu / Lark
     feishu_app_id = os.getenv("FEISHU_APP_ID")
     feishu_app_secret = os.getenv("FEISHU_APP_SECRET")
@@ -1136,12 +1229,24 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
         qq_group_allowed = os.getenv("QQ_GROUP_ALLOWED_USERS", "").strip()
         if qq_group_allowed:
             extra["group_allow_from"] = qq_group_allowed
-        qq_home = os.getenv("QQ_HOME_CHANNEL", "").strip()
+        qq_home = os.getenv("QQBOT_HOME_CHANNEL", "").strip()
+        qq_home_name_env = "QQBOT_HOME_CHANNEL_NAME"
+        if not qq_home:
+            # Back-compat: accept the pre-rename name and log a one-time warning.
+            legacy_home = os.getenv("QQ_HOME_CHANNEL", "").strip()
+            if legacy_home:
+                qq_home = legacy_home
+                qq_home_name_env = "QQ_HOME_CHANNEL_NAME"
+                import logging
+                logging.getLogger(__name__).warning(
+                    "QQ_HOME_CHANNEL is deprecated; rename to QQBOT_HOME_CHANNEL "
+                    "in your .env for consistency with the platform key."
+                )
         if qq_home:
             config.platforms[Platform.QQBOT].home_channel = HomeChannel(
                 platform=Platform.QQBOT,
                 chat_id=qq_home,
-                name=os.getenv("QQ_HOME_CHANNEL_NAME", "Home"),
+                name=os.getenv("QQBOT_HOME_CHANNEL_NAME") or os.getenv(qq_home_name_env, "Home"),
             )
 
     # Session settings

gateway/platforms/api_server.py

@@ -902,7 +902,7 @@ class APIServerAdapter(BasePlatformAdapter):
                 return time.monotonic()
 
             # Stream content chunks as they arrive from the agent
-            loop = asyncio.get_event_loop()
+            loop = asyncio.get_running_loop()
             while True:
                 try:
                     delta = await loop.run_in_executor(None, lambda: stream_q.get(timeout=0.5))
@@ -1241,7 +1241,7 @@ class APIServerAdapter(BasePlatformAdapter):
                     await _emit_text_delta(it)
                 # Other types (non-string, non-tuple) are silently dropped.
 
-            loop = asyncio.get_event_loop()
+            loop = asyncio.get_running_loop()
             while True:
                 try:
                     item = await loop.run_in_executor(None, lambda: stream_q.get(timeout=0.5))
@@ -2004,7 +2004,7 @@ class APIServerAdapter(BasePlatformAdapter):
         callers (e.g. the SSE writer) to call ``agent.interrupt()`` from
         another thread to stop in-progress LLM calls.
         """
-        loop = asyncio.get_event_loop()
+        loop = asyncio.get_running_loop()
 
         def _run():
             agent = self._create_agent(

gateway/platforms/base.py

@@ -6,6 +6,7 @@ and implement the required methods.
 """
 
 import asyncio
+import inspect
 import ipaddress
 import logging
 import os
@@ -669,6 +670,15 @@ class MessageEvent:
     # Original platform data
     raw_message: Any = None
     message_id: Optional[str] = None
+
+    # Platform-specific update identifier.  For Telegram this is the
+    # ``update_id`` from the PTB Update wrapper; other platforms currently
+    # ignore it.  Used by ``/restart`` to record the triggering update so the
+    # new gateway can advance the Telegram offset past it and avoid processing
+    # the same ``/restart`` twice if PTB's graceful-shutdown ACK times out
+    # ("Error while calling `get_updates` one more time to mark all fetched
+    # updates" in gateway.log).
+    platform_update_id: Optional[int] = None
     
     # Media attachments
     # media_urls: local file paths (for vision tool access)
@@ -682,6 +692,10 @@ class MessageEvent:
     # Auto-loaded skill(s) for topic/channel bindings (e.g., Telegram DM Topics,
     # Discord channel_skill_bindings).  A single name or ordered list.
     auto_skill: Optional[str | list[str]] = None
+
+    # Per-channel ephemeral system prompt (e.g. Discord channel_prompts).
+    # Applied at API call time and never persisted to transcript history.
+    channel_prompt: Optional[str] = None
     
     # Internal flag — set for synthetic events (e.g. background process
     # completion notifications) that must bypass user authorization checks.
@@ -730,25 +744,56 @@ def merge_pending_message_event(
     pending_messages: Dict[str, MessageEvent],
     session_key: str,
     event: MessageEvent,
+    *,
+    merge_text: bool = False,
 ) -> None:
     """Store or merge a pending event for a session.
 
     Photo bursts/albums often arrive as multiple near-simultaneous PHOTO
     events. Merge those into the existing queued event so the next turn sees
-    the whole burst, while non-photo follow-ups still replace the pending
-    event normally.
+    the whole burst.
+
+    When ``merge_text`` is enabled, rapid follow-up TEXT events are appended
+    instead of replacing the pending turn. This is used for Telegram bursty
+    follow-ups so a multi-part user thought is not silently truncated to only
+    the last queued fragment.
     """
     existing = pending_messages.get(session_key)
-    if (
-        existing
-        and getattr(existing, "message_type", None) == MessageType.PHOTO
-        and event.message_type == MessageType.PHOTO
-    ):
-        existing.media_urls.extend(event.media_urls)
-        existing.media_types.extend(event.media_types)
-        if event.text:
-            existing.text = BasePlatformAdapter._merge_caption(existing.text, event.text)
-        return
+    if existing:
+        existing_is_photo = getattr(existing, "message_type", None) == MessageType.PHOTO
+        incoming_is_photo = event.message_type == MessageType.PHOTO
+        existing_has_media = bool(existing.media_urls)
+        incoming_has_media = bool(event.media_urls)
+
+        if existing_is_photo and incoming_is_photo:
+            existing.media_urls.extend(event.media_urls)
+            existing.media_types.extend(event.media_types)
+            if event.text:
+                existing.text = BasePlatformAdapter._merge_caption(existing.text, event.text)
+            return
+
+        if existing_has_media or incoming_has_media:
+            if incoming_has_media:
+                existing.media_urls.extend(event.media_urls)
+                existing.media_types.extend(event.media_types)
+            if event.text:
+                if existing.text:
+                    existing.text = BasePlatformAdapter._merge_caption(existing.text, event.text)
+                else:
+                    existing.text = event.text
+            if existing_is_photo or incoming_is_photo:
+                existing.message_type = MessageType.PHOTO
+            return
+
+        if (
+            merge_text
+            and getattr(existing, "message_type", None) == MessageType.TEXT
+            and event.message_type == MessageType.TEXT
+        ):
+            if event.text:
+                existing.text = f"{existing.text}\n{event.text}" if existing.text else event.text
+            return
+
     pending_messages[session_key] = event
 
 
@@ -776,6 +821,36 @@ _RETRYABLE_ERROR_PATTERNS = (
 MessageHandler = Callable[[MessageEvent], Awaitable[Optional[str]]]
 
 
+def resolve_channel_prompt(
+    config_extra: dict,
+    channel_id: str,
+    parent_id: str | None = None,
+) -> str | None:
+    """Resolve a per-channel ephemeral prompt from platform config.
+
+    Looks up ``channel_prompts`` in the adapter's ``config.extra`` dict.
+    Prefers an exact match on *channel_id*; falls back to *parent_id*
+    (useful for forum threads / child channels inheriting a parent prompt).
+
+    Returns the prompt string, or None if no match is found.  Blank/whitespace-
+    only prompts are treated as absent.
+    """
+    prompts = config_extra.get("channel_prompts") or {}
+    if not isinstance(prompts, dict):
+        return None
+
+    for key in (channel_id, parent_id):
+        if not key:
+            continue
+        prompt = prompts.get(key)
+        if prompt is None:
+            continue
+        prompt = str(prompt).strip()
+        if prompt:
+            return prompt
+    return None
+
+
 class BasePlatformAdapter(ABC):
     """
     Base class for platform adapters.
@@ -805,6 +880,12 @@ class BasePlatformAdapter(ABC):
         # Gateway shutdown cancels these so an old gateway instance doesn't keep
         # working on a task after --replace or manual restarts.
         self._background_tasks: set[asyncio.Task] = set()
+        # One-shot callbacks to fire after the main response is delivered.
+        # Keyed by session_key. Values are either a bare callback (legacy) or
+        # a ``(generation, callback)`` tuple so GatewayRunner can make deferred
+        # deliveries generation-aware and avoid stale runs clearing callbacks
+        # registered by a fresher run for the same session.
+        self._post_delivery_callbacks: Dict[str, Any] = {}
         self._expected_cancelled_tasks: set[asyncio.Task] = set()
         self._busy_session_handler: Optional[Callable[[MessageEvent, str], Awaitable[bool]]] = None
         # Chats where auto-TTS on voice input is disabled (set by /voice off)
@@ -975,16 +1056,40 @@ class BasePlatformAdapter(ABC):
         """
         pass
 
+    # Default: the adapter treats ``finalize=True`` on edit_message as a
+    # no-op and is happy to have the stream consumer skip redundant final
+    # edits.  Subclasses that *require* an explicit finalize call to close
+    # out the message lifecycle (e.g. rich card / AI assistant surfaces
+    # such as DingTalk AI Cards) override this to True (class attribute or
+    # property) so the stream consumer knows not to short-circuit.
+    REQUIRES_EDIT_FINALIZE: bool = False
+
     async def edit_message(
         self,
         chat_id: str,
         message_id: str,
         content: str,
+        *,
+        finalize: bool = False,
     ) -> SendResult:
         """
         Edit a previously sent message. Optional — platforms that don't
         support editing return success=False and callers fall back to
         sending a new message.
+
+        ``finalize`` signals that this is the last edit in a streaming
+        sequence.  Most platforms (Telegram, Slack, Discord, Matrix,
+        etc.) treat it as a no-op because their edit APIs have no notion
+        of message lifecycle state — an edit is an edit.  Platforms that
+        render streaming updates with a distinct "in progress" state and
+        require explicit closure (e.g. rich card / AI assistant surfaces
+        such as DingTalk AI Cards) use it to finalize the message and
+        transition the UI out of the streaming indicator — those should
+        also set ``REQUIRES_EDIT_FINALIZE = True`` so callers route a
+        final edit through even when content is unchanged.  Callers
+        should set ``finalize=True`` on the final edit of a streamed
+        response (typically when ``got_done`` fires in the stream
+        consumer) and leave it ``False`` on intermediate edits.
         """
         return SendResult(success=False, error="Not supported")
 
@@ -1221,7 +1326,7 @@ class BasePlatformAdapter(ABC):
                 path = path[1:-1].strip()
             path = path.lstrip("`\"'").rstrip("`\"',.;:)}]")
             if path:
-                media.append((path, has_voice_tag))
+                media.append((os.path.expanduser(path), has_voice_tag))
 
         # Remove MEDIA tags from content (including surrounding quote/backtick wrappers)
         if media:
@@ -1298,7 +1403,13 @@ class BasePlatformAdapter(ABC):
 
         return paths, cleaned
 
-    async def _keep_typing(self, chat_id: str, interval: float = 2.0, metadata=None) -> None:
+    async def _keep_typing(
+        self,
+        chat_id: str,
+        interval: float = 2.0,
+        metadata=None,
+        stop_event: asyncio.Event | None = None,
+    ) -> None:
         """
         Continuously send typing indicator until cancelled.
         
@@ -1312,9 +1423,18 @@ class BasePlatformAdapter(ABC):
         """
         try:
             while True:
+                if stop_event is not None and stop_event.is_set():
+                    return
                 if chat_id not in self._typing_paused:
                     await self.send_typing(chat_id, metadata=metadata)
-                await asyncio.sleep(interval)
+                if stop_event is None:
+                    await asyncio.sleep(interval)
+                    continue
+                try:
+                    await asyncio.wait_for(stop_event.wait(), timeout=interval)
+                except asyncio.TimeoutError:
+                    continue
+                return
         except asyncio.CancelledError:
             pass  # Normal cancellation when handler completes
         finally:
@@ -1341,6 +1461,59 @@ class BasePlatformAdapter(ABC):
         """Resume typing indicator for a chat after approval resolves."""
         self._typing_paused.discard(chat_id)
 
+    async def interrupt_session_activity(self, session_key: str, chat_id: str) -> None:
+        """Signal the active session loop to stop and clear typing immediately."""
+        if session_key:
+            interrupt_event = self._active_sessions.get(session_key)
+            if interrupt_event is not None:
+                interrupt_event.set()
+        try:
+            await self.stop_typing(chat_id)
+        except Exception:
+            pass
+
+    def register_post_delivery_callback(
+        self,
+        session_key: str,
+        callback: Callable,
+        *,
+        generation: int | None = None,
+    ) -> None:
+        """Register a deferred callback to fire after the main response.
+
+        ``generation`` lets callers tie the callback to a specific gateway run
+        generation so stale runs cannot clear callbacks owned by a fresher run.
+        """
+        if not session_key or not callable(callback):
+            return
+        if generation is None:
+            self._post_delivery_callbacks[session_key] = callback
+        else:
+            self._post_delivery_callbacks[session_key] = (int(generation), callback)
+
+    def pop_post_delivery_callback(
+        self,
+        session_key: str,
+        *,
+        generation: int | None = None,
+    ) -> Callable | None:
+        """Pop a deferred callback, optionally requiring generation ownership."""
+        if not session_key:
+            return None
+        entry = self._post_delivery_callbacks.get(session_key)
+        if entry is None:
+            return None
+        if isinstance(entry, tuple) and len(entry) == 2:
+            entry_generation, callback = entry
+            if generation is not None and int(entry_generation) != int(generation):
+                return None
+            self._post_delivery_callbacks.pop(session_key, None)
+            return callback if callable(callback) else None
+        if generation is not None:
+            return None
+        self._post_delivery_callbacks.pop(session_key, None)
+        return entry if callable(entry) else None
+
     # ── Processing lifecycle hooks ──────────────────────────────────────────
     # Subclasses override these to react to message processing events
     # (e.g. Discord adds 👀/✅/❌ reactions).
@@ -1509,7 +1682,9 @@ class BasePlatformAdapter(ABC):
             # session lifecycle and its cleanup races with the running task
             # (see PR #4926).
             cmd = event.get_command()
-            if cmd in ("approve", "deny", "status", "stop", "new", "reset", "background", "restart"):
+            from hermes_cli.commands import should_bypass_active_session
+
+            if should_bypass_active_session(cmd):
                 logger.debug(
                     "[%s] Command '/%s' bypassing active-session guard for %s",
                     self.name, cmd, session_key,
@@ -1609,10 +1784,23 @@ class BasePlatformAdapter(ABC):
         # 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
+        callback_generation = getattr(interrupt_event, "_hermes_run_generation", None)
         
         # Start continuous typing indicator (refreshes every 2 seconds)
         _thread_metadata = {"thread_id": event.source.thread_id} if event.source.thread_id else None
-        typing_task = asyncio.create_task(self._keep_typing(event.source.chat_id, metadata=_thread_metadata))
+        _keep_typing_kwargs = {"metadata": _thread_metadata}
+        try:
+            _keep_typing_sig = inspect.signature(self._keep_typing)
+        except (TypeError, ValueError):
+            _keep_typing_sig = None
+        if _keep_typing_sig is None or "stop_event" in _keep_typing_sig.parameters:
+            _keep_typing_kwargs["stop_event"] = interrupt_event
+        typing_task = asyncio.create_task(
+            self._keep_typing(
+                event.source.chat_id,
+                **_keep_typing_kwargs,
+            )
+        )
         
         try:
             await self._run_processing_hook("on_processing_start", event)
@@ -1624,6 +1812,21 @@ class BasePlatformAdapter(ABC):
             # 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.
+            #
+            # Suppress stale response when the session was interrupted by a
+            # new message that hasn't been consumed yet.  The pending message
+            # is processed by the pending-message handler below (#8221/#2483).
+            if (
+                response
+                and interrupt_event.is_set()
+                and session_key in self._pending_messages
+            ):
+                logger.info(
+                    "[%s] Suppressing stale response for interrupted session %s",
+                    self.name,
+                    session_key,
+                )
+                response = None
             if not response:
                 logger.debug("[%s] Handler returned empty/None response for %s", self.name, event.source.chat_id)
             if response:
@@ -1806,9 +2009,18 @@ class BasePlatformAdapter(ABC):
             if session_key in self._pending_messages:
                 pending_event = self._pending_messages.pop(session_key)
                 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]
+                # Keep the _active_sessions entry live across the turn chain
+                # and only CLEAR the interrupt Event — do NOT delete the entry.
+                # If we deleted here, a concurrent inbound message arriving
+                # during the awaits below would pass the Level-1 guard, spawn
+                # its own _process_message_background, and run simultaneously
+                # with the recursive drain below.  Two agents on one
+                # session_key = duplicate responses, duplicate tool calls.
+                # Clearing the Event keeps the guard live so follow-ups take
+                # the busy-handler path (queue + interrupt) as intended.
+                _active = self._active_sessions.get(session_key)
+                if _active is not None:
+                    _active.clear()
                 typing_task.cancel()
                 try:
                     await typing_task
@@ -1845,6 +2057,21 @@ class BasePlatformAdapter(ABC):
             except Exception:
                 pass  # Last resort — don't let error reporting crash the handler
         finally:
+            # Fire any one-shot post-delivery callback registered for this
+            # session (e.g. deferred background-review notifications).
+            _callback_generation = callback_generation
+            if hasattr(self, "pop_post_delivery_callback"):
+                _post_cb = self.pop_post_delivery_callback(
+                    session_key,
+                    generation=_callback_generation,
+                )
+            else:
+                _post_cb = getattr(self, "_post_delivery_callbacks", {}).pop(session_key, None)
+            if callable(_post_cb):
+                try:
+                    _post_cb()
+                except Exception:
+                    pass
             # Stop typing indicator
             typing_task.cancel()
             try:
@@ -1858,9 +2085,37 @@ class BasePlatformAdapter(ABC):
                     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]
+            # Late-arrival drain: a message may have arrived during the
+            # cleanup awaits above (typing_task cancel, stop_typing).  Such
+            # messages passed the Level-1 guard (entry still live, Event
+            # possibly set) and landed in _pending_messages via the
+            # busy-handler path.  Without this block, we would delete the
+            # active-session entry and the queued message would be silently
+            # dropped (user never gets a reply).
+            late_pending = self._pending_messages.pop(session_key, None)
+            if late_pending is not None:
+                logger.debug(
+                    "[%s] Late-arrival pending message during cleanup — spawning drain task",
+                    self.name,
+                )
+                _active = self._active_sessions.get(session_key)
+                if _active is not None:
+                    _active.clear()
+                drain_task = asyncio.create_task(
+                    self._process_message_background(late_pending, session_key)
+                )
+                try:
+                    self._background_tasks.add(drain_task)
+                    drain_task.add_done_callback(self._background_tasks.discard)
+                except TypeError:
+                    # Tests stub create_task() with non-hashable sentinels; tolerate.
+                    pass
+                # Leave _active_sessions[session_key] populated — the drain
+                # task's own lifecycle will clean it up.
+            else:
+                # Clean up session tracking
+                if session_key in self._active_sessions:
+                    del self._active_sessions[session_key]
     
     async def cancel_background_tasks(self) -> None:
         """Cancel any in-flight background message-processing tasks.
@@ -1868,12 +2123,26 @@ class BasePlatformAdapter(ABC):
         Used during gateway shutdown/replacement so active sessions from the old
         process do not keep running after adapters are being torn down.
         """
-        tasks = [task for task in self._background_tasks if not task.done()]
-        for task in tasks:
-            self._expected_cancelled_tasks.add(task)
-            task.cancel()
-        if tasks:
+        # Loop until no new tasks appear.  Without this, a message
+        # arriving during the `await asyncio.gather` below would spawn
+        # a fresh _process_message_background task (added to
+        # self._background_tasks at line ~1668 via handle_message),
+        # and the _background_tasks.clear() at the end of this method
+        # would drop the reference — the task runs untracked against a
+        # disconnecting adapter, logs send-failures, and may linger
+        # until it completes on its own.  Retrying the drain until the
+        # task set stabilizes closes the window.
+        MAX_DRAIN_ROUNDS = 5
+        for _ in range(MAX_DRAIN_ROUNDS):
+            tasks = [task for task in self._background_tasks if not task.done()]
+            if not tasks:
+                break
+            for task in tasks:
+                self._expected_cancelled_tasks.add(task)
+                task.cancel()
             await asyncio.gather(*tasks, return_exceptions=True)
+            # Loop: late-arrival tasks spawned during the gather above
+            # will be in self._background_tasks now.  Re-check.
         self._background_tasks.clear()
         self._expected_cancelled_tasks.clear()
         self._pending_messages.clear()
@@ -1898,6 +2167,7 @@ class BasePlatformAdapter(ABC):
         chat_topic: Optional[str] = None,
         user_id_alt: Optional[str] = None,
         chat_id_alt: Optional[str] = None,
+        is_bot: bool = False,
     ) -> SessionSource:
         """Helper to build a SessionSource for this platform."""
         # Normalize empty topic to None
@@ -1914,6 +2184,7 @@ class BasePlatformAdapter(ABC):
             chat_topic=chat_topic.strip() if chat_topic else None,
             user_id_alt=user_id_alt,
             chat_id_alt=chat_id_alt,
+            is_bot=is_bot,
         )
     
     @abstractmethod

gateway/platforms/feishu.py

@@ -8,7 +8,8 @@ Supports:
 - Gateway allowlist integration via FEISHU_ALLOWED_USERS
 - Persistent dedup state across restarts
 - Per-chat serial message processing (matches openclaw createChatQueue)
-- Persistent ACK emoji reaction on inbound messages
+- Processing status reactions: Typing while working, removed on success,
+  swapped for CrossMark on failure
 - Reaction events routed as synthetic text events (matches openclaw)
 - Interactive card button-click events routed as synthetic COMMAND events
 - Webhook anomaly tracking (matches openclaw createWebhookAnomalyTracker)
@@ -29,6 +30,7 @@ import re
 import threading
 import time
 import uuid
+from collections import OrderedDict
 from dataclasses import dataclass, field
 from datetime import datetime
 from pathlib import Path
@@ -98,6 +100,7 @@ from gateway.platforms.base import (
     BasePlatformAdapter,
     MessageEvent,
     MessageType,
+    ProcessingOutcome,
     SendResult,
     SUPPORTED_DOCUMENT_TYPES,
     cache_document_from_bytes,
@@ -119,6 +122,8 @@ _MARKDOWN_HINT_RE = re.compile(
     re.MULTILINE,
 )
 _MARKDOWN_LINK_RE = re.compile(r"\[([^\]]+)\]\(([^)]+)\)")
+_MARKDOWN_FENCE_OPEN_RE = re.compile(r"^```([^\n`]*)\s*$")
+_MARKDOWN_FENCE_CLOSE_RE = re.compile(r"^```\s*$")
 _MENTION_RE = re.compile(r"@_user_\d+")
 _MULTISPACE_RE = re.compile(r"[ \t]{2,}")
 _POST_CONTENT_INVALID_RE = re.compile(r"content format of the post type is incorrect", re.IGNORECASE)
@@ -188,7 +193,17 @@ _APPROVAL_LABEL_MAP: Dict[str, str] = {
 }
 _FEISHU_BOT_MSG_TRACK_SIZE = 512                   # LRU size for tracking sent message IDs
 _FEISHU_REPLY_FALLBACK_CODES = frozenset({230011, 231003})  # reply target withdrawn/missing → create fallback
-_FEISHU_ACK_EMOJI = "OK"
+
+# Feishu reactions render as prominent badges, unlike Discord/Telegram's
+# small footer emoji — a success badge on every message would add noise, so
+# we only mark start (Typing) and failure (CrossMark); the reply itself is
+# the success signal.
+_FEISHU_REACTION_IN_PROGRESS = "Typing"
+_FEISHU_REACTION_FAILURE = "CrossMark"
+# Bound on the (message_id → reaction_id) handle cache. Happy-path entries
+# drain on completion; the cap is a safeguard against unbounded growth from
+# delete-failures, not a capacity plan.
+_FEISHU_PROCESSING_REACTION_CACHE_SIZE = 1024
 
 # QR onboarding constants
 _ONBOARD_ACCOUNTS_URLS = {
@@ -430,23 +445,66 @@ def _coerce_required_int(value: Any, default: int, min_value: int = 0) -> int:
 
 
 def _build_markdown_post_payload(content: str) -> str:
+    rows = _build_markdown_post_rows(content)
     return json.dumps(
         {
             "zh_cn": {
-                "content": [
-                    [
-                        {
-                            "tag": "md",
-                            "text": content,
-                        }
-                    ]
-                ],
+                "content": rows,
             }
         },
         ensure_ascii=False,
     )
 
 
+def _build_markdown_post_rows(content: str) -> List[List[Dict[str, str]]]:
+    """Build Feishu post rows while isolating fenced code blocks.
+
+    Feishu's `md` renderer can swallow trailing content when a fenced code block
+    appears inside one large markdown element. Split the reply at real fence
+    lines so prose before/after the code block remains visible while code stays
+    in a dedicated row.
+    """
+    if not content:
+        return [[{"tag": "md", "text": ""}]]
+    if "```" not in content:
+        return [[{"tag": "md", "text": content}]]
+
+    rows: List[List[Dict[str, str]]] = []
+    current: List[str] = []
+    in_code_block = False
+
+    def _flush_current() -> None:
+        nonlocal current
+        if not current:
+            return
+        segment = "\n".join(current)
+        if segment.strip():
+            rows.append([{"tag": "md", "text": segment}])
+        current = []
+
+    for raw_line in content.splitlines():
+        stripped_line = raw_line.strip()
+        is_fence = bool(
+            _MARKDOWN_FENCE_CLOSE_RE.match(stripped_line)
+            if in_code_block
+            else _MARKDOWN_FENCE_OPEN_RE.match(stripped_line)
+        )
+
+        if is_fence:
+            if not in_code_block:
+                _flush_current()
+            current.append(raw_line)
+            in_code_block = not in_code_block
+            if not in_code_block:
+                _flush_current()
+            continue
+
+        current.append(raw_line)
+
+    _flush_current()
+    return rows or [[{"tag": "md", "text": content}]]
+
+
 def parse_feishu_post_payload(payload: Any) -> FeishuPostParseResult:
     resolved = _resolve_post_payload(payload)
     if not resolved:
@@ -1073,6 +1131,13 @@ class FeishuAdapter(BasePlatformAdapter):
         self._webhook_rate_counts: Dict[str, tuple[int, float]] = {}  # rate_key → (count, window_start)
         self._webhook_anomaly_counts: Dict[str, tuple[int, str, float]] = {}  # ip → (count, last_status, first_seen)
         self._card_action_tokens: Dict[str, float] = {}  # token → first_seen_time
+        # Inbound events that arrived before the adapter loop was ready
+        # (e.g. during startup/restart or network-flap reconnect). A single
+        # drainer thread replays them as soon as the loop becomes available.
+        self._pending_inbound_events: List[Any] = []
+        self._pending_inbound_lock = threading.Lock()
+        self._pending_drain_scheduled = False
+        self._pending_inbound_max_depth = 1000  # cap queue; drop oldest beyond
         self._chat_locks: Dict[str, asyncio.Lock] = {}  # chat_id → lock (per-chat serial processing)
         self._sent_message_ids_to_chat: Dict[str, str] = {}  # message_id → chat_id (for reaction routing)
         self._sent_message_id_order: List[str] = []  # LRU order for _sent_message_ids_to_chat
@@ -1089,6 +1154,9 @@ class FeishuAdapter(BasePlatformAdapter):
         # Exec approval button state (approval_id → {session_key, message_id, chat_id})
         self._approval_state: Dict[int, Dict[str, str]] = {}
         self._approval_counter = itertools.count(1)
+        # Feishu reaction deletion requires the opaque reaction_id returned
+        # by create, so we cache it per message_id.
+        self._pending_processing_reactions: "OrderedDict[str, str]" = OrderedDict()
         self._load_seen_message_ids()
 
     @staticmethod
@@ -1219,6 +1287,12 @@ class FeishuAdapter(BasePlatformAdapter):
             .register_p2_card_action_trigger(self._on_card_action_trigger)
             .register_p2_im_chat_member_bot_added_v1(self._on_bot_added_to_chat)
             .register_p2_im_chat_member_bot_deleted_v1(self._on_bot_removed_from_chat)
+            .register_p2_im_chat_access_event_bot_p2p_chat_entered_v1(self._on_p2p_chat_entered)
+            .register_p2_im_message_recalled_v1(self._on_message_recalled)
+            .register_p2_customized_event(
+                "drive.notice.comment_add_v1",
+                self._on_drive_comment_event,
+            )
             .build()
         )
 
@@ -1410,6 +1484,8 @@ class FeishuAdapter(BasePlatformAdapter):
         chat_id: str,
         message_id: str,
         content: str,
+        *,
+        finalize: bool = False,
     ) -> SendResult:
         """Edit a previously sent Feishu text/post message."""
         if not self._client:
@@ -1757,10 +1833,22 @@ class FeishuAdapter(BasePlatformAdapter):
     # =========================================================================
 
     def _on_message_event(self, data: Any) -> None:
-        """Normalize Feishu inbound events into MessageEvent."""
+        """Normalize Feishu inbound events into MessageEvent.
+
+        Called by the lark_oapi SDK's event dispatcher on a background thread.
+        If the adapter loop is not currently accepting callbacks (brief window
+        during startup/restart or network-flap reconnect), the event is queued
+        for replay instead of dropped.
+        """
         loop = self._loop
-        if loop is None or bool(getattr(loop, "is_closed", lambda: False)()):
-            logger.warning("[Feishu] Dropping inbound message before adapter loop is ready")
+        if not self._loop_accepts_callbacks(loop):
+            start_drainer = self._enqueue_pending_inbound_event(data)
+            if start_drainer:
+                threading.Thread(
+                    target=self._drain_pending_inbound_events,
+                    name="feishu-pending-inbound-drainer",
+                    daemon=True,
+                ).start()
             return
         future = asyncio.run_coroutine_threadsafe(
             self._handle_message_event_data(data),
@@ -1768,6 +1856,124 @@ class FeishuAdapter(BasePlatformAdapter):
         )
         future.add_done_callback(self._log_background_failure)
 
+    def _enqueue_pending_inbound_event(self, data: Any) -> bool:
+        """Append an event to the pending-inbound queue.
+
+        Returns True if the caller should spawn a drainer thread (no drainer
+        currently scheduled), False if a drainer is already running and will
+        pick up the new event on its next pass.
+        """
+        with self._pending_inbound_lock:
+            if len(self._pending_inbound_events) >= self._pending_inbound_max_depth:
+                # Queue full — drop the oldest to make room. This happens only
+                # if the loop stays unavailable for an extended period AND the
+                # WS keeps firing callbacks. Still better than silent drops.
+                dropped = self._pending_inbound_events.pop(0)
+                try:
+                    event = getattr(dropped, "event", None)
+                    message = getattr(event, "message", None)
+                    message_id = str(getattr(message, "message_id", "") or "unknown")
+                except Exception:
+                    message_id = "unknown"
+                logger.error(
+                    "[Feishu] Pending-inbound queue full (%d); dropped oldest event %s",
+                    self._pending_inbound_max_depth,
+                    message_id,
+                )
+            self._pending_inbound_events.append(data)
+            depth = len(self._pending_inbound_events)
+            should_start = not self._pending_drain_scheduled
+            if should_start:
+                self._pending_drain_scheduled = True
+        logger.warning(
+            "[Feishu] Queued inbound event for replay (loop not ready, queue depth=%d)",
+            depth,
+        )
+        return should_start
+
+    def _drain_pending_inbound_events(self) -> None:
+        """Replay queued inbound events once the adapter loop is ready.
+
+        Runs in a dedicated daemon thread. Polls ``_running`` and
+        ``_loop_accepts_callbacks`` until events can be dispatched or the
+        adapter shuts down. A single drainer handles the entire queue;
+        concurrent ``_on_message_event`` calls just append.
+        """
+        poll_interval = 0.25
+        max_wait_seconds = 120.0  # safety cap: drop queue after 2 minutes
+        waited = 0.0
+        try:
+            while True:
+                if not getattr(self, "_running", True):
+                    # Adapter shutting down — drop queued events rather than
+                    # holding them against a closed loop.
+                    with self._pending_inbound_lock:
+                        dropped = len(self._pending_inbound_events)
+                        self._pending_inbound_events.clear()
+                    if dropped:
+                        logger.warning(
+                            "[Feishu] Dropped %d queued inbound event(s) during shutdown",
+                            dropped,
+                        )
+                    return
+                loop = self._loop
+                if self._loop_accepts_callbacks(loop):
+                    with self._pending_inbound_lock:
+                        batch = self._pending_inbound_events[:]
+                        self._pending_inbound_events.clear()
+                    if not batch:
+                        # Queue emptied between check and grab; done.
+                        with self._pending_inbound_lock:
+                            if not self._pending_inbound_events:
+                                return
+                        continue
+                    dispatched = 0
+                    requeue: List[Any] = []
+                    for event in batch:
+                        try:
+                            fut = asyncio.run_coroutine_threadsafe(
+                                self._handle_message_event_data(event),
+                                loop,
+                            )
+                            fut.add_done_callback(self._log_background_failure)
+                            dispatched += 1
+                        except RuntimeError:
+                            # Loop closed between check and submit — requeue
+                            # and poll again.
+                            requeue.append(event)
+                    if requeue:
+                        with self._pending_inbound_lock:
+                            self._pending_inbound_events[:0] = requeue
+                    if dispatched:
+                        logger.info(
+                            "[Feishu] Replayed %d queued inbound event(s)",
+                            dispatched,
+                        )
+                    if not requeue:
+                        # Successfully drained; check if more arrived while
+                        # we were dispatching and exit if not.
+                        with self._pending_inbound_lock:
+                            if not self._pending_inbound_events:
+                                return
+                    # More events queued or requeue pending — loop again.
+                    continue
+                if waited >= max_wait_seconds:
+                    with self._pending_inbound_lock:
+                        dropped = len(self._pending_inbound_events)
+                        self._pending_inbound_events.clear()
+                    logger.error(
+                        "[Feishu] Adapter loop unavailable for %.0fs; "
+                        "dropped %d queued inbound event(s)",
+                        max_wait_seconds,
+                        dropped,
+                    )
+                    return
+                time.sleep(poll_interval)
+                waited += poll_interval
+        finally:
+            with self._pending_inbound_lock:
+                self._pending_drain_scheduled = False
+
     async def _handle_message_event_data(self, data: Any) -> None:
         """Shared inbound message handling for websocket and webhook transports."""
         event = getattr(data, "event", None)
@@ -1782,8 +1988,8 @@ class FeishuAdapter(BasePlatformAdapter):
         if not message_id or self._is_duplicate(message_id):
             logger.debug("[Feishu] Dropping duplicate/missing message_id: %s", message_id)
             return
-        if getattr(sender, "sender_type", "") == "bot":
-            logger.debug("[Feishu] Dropping bot-originated event: %s", message_id)
+        if self._is_self_sent_bot_message(event):
+            logger.debug("[Feishu] Dropping self-sent bot event: %s", message_id)
             return
 
         chat_type = getattr(message, "chat_type", "p2p")
@@ -1820,6 +2026,31 @@ class FeishuAdapter(BasePlatformAdapter):
         logger.info("[Feishu] Bot removed from chat: %s", chat_id)
         self._chat_info_cache.pop(chat_id, None)
 
+    def _on_p2p_chat_entered(self, data: Any) -> None:
+        logger.debug("[Feishu] User entered P2P chat with bot")
+
+    def _on_message_recalled(self, data: Any) -> None:
+        logger.debug("[Feishu] Message recalled by user")
+
+    def _on_drive_comment_event(self, data: Any) -> None:
+        """Handle drive document comment notification (drive.notice.comment_add_v1).
+
+        Delegates to :mod:`gateway.platforms.feishu_comment` for parsing,
+        logging, and reaction.  Scheduling follows the same
+        ``run_coroutine_threadsafe`` pattern used by ``_on_message_event``.
+        """
+        from gateway.platforms.feishu_comment import handle_drive_comment_event
+
+        loop = self._loop
+        if not self._loop_accepts_callbacks(loop):
+            logger.warning("[Feishu] Dropping drive comment event before adapter loop is ready")
+            return
+        future = asyncio.run_coroutine_threadsafe(
+            handle_drive_comment_event(self._client, data, self_open_id=self._bot_open_id),
+            loop,
+        )
+        future.add_done_callback(self._log_background_failure)
+
     def _on_reaction_event(self, event_type: str, data: Any) -> None:
         """Route user reactions on bot messages as synthetic text events."""
         event = getattr(data, "event", None)
@@ -1835,12 +2066,12 @@ class FeishuAdapter(BasePlatformAdapter):
             operator_type,
             emoji_type,
         )
-        # Only process reactions from real users. Ignore app/bot-generated reactions
-        # and Hermes' own ACK emoji to avoid feedback loops.
+        # Drop bot/app-origin reactions to break the feedback loop from our
+        # own lifecycle reactions. A human reacting with the same emoji (e.g.
+        # clicking Typing on a bot message) is still routed through.
         loop = self._loop
         if (
             operator_type in {"bot", "app"}
-            or emoji_type == _FEISHU_ACK_EMOJI
             or not message_id
             or loop is None
             or bool(getattr(loop, "is_closed", lambda: False)())
@@ -2064,33 +2295,35 @@ class FeishuAdapter(BasePlatformAdapter):
 
     async def _handle_message_with_guards(self, event: MessageEvent) -> None:
         """Dispatch a single event through the agent pipeline with per-chat serialization
-        and a persistent ACK emoji reaction before processing starts.
+        before handing the event off to the agent.
 
-        - Per-chat lock: ensures messages in the same chat are processed one at a time
-          (matches openclaw's createChatQueue serial queue behaviour).
-        - ACK indicator: adds a CHECK reaction to the triggering message before handing
-          off to the agent and leaves it in place as a receipt marker.
+        Per-chat lock ensures messages in the same chat are processed one at a
+        time (matches openclaw's createChatQueue serial queue behaviour).
         """
         chat_id = getattr(event.source, "chat_id", "") or "" if event.source else ""
         chat_lock = self._get_chat_lock(chat_id)
         async with chat_lock:
-            message_id = event.message_id
-            if message_id:
-                await self._add_ack_reaction(message_id)
             await self.handle_message(event)
 
-    async def _add_ack_reaction(self, message_id: str) -> Optional[str]:
-        """Add a persistent ACK emoji reaction to signal the message was received."""
-        if not self._client or not message_id:
+    # =========================================================================
+    # Processing status reactions
+    # =========================================================================
+
+    def _reactions_enabled(self) -> bool:
+        return os.getenv("FEISHU_REACTIONS", "true").strip().lower() not in ("false", "0", "no")
+
+    async def _add_reaction(self, message_id: str, emoji_type: str) -> Optional[str]:
+        """Return the reaction_id on success, else None. The id is needed later for deletion."""
+        if not self._client or not message_id or not emoji_type:
             return None
         try:
-            from lark_oapi.api.im.v1 import (  # lazy import — keeps optional dep optional
+            from lark_oapi.api.im.v1 import (
                 CreateMessageReactionRequest,
                 CreateMessageReactionRequestBody,
             )
             body = (
                 CreateMessageReactionRequestBody.builder()
-                .reaction_type({"emoji_type": _FEISHU_ACK_EMOJI})
+                .reaction_type({"emoji_type": emoji_type})
                 .build()
             )
             request = (
@@ -2103,16 +2336,93 @@ class FeishuAdapter(BasePlatformAdapter):
             if response and getattr(response, "success", lambda: False)():
                 data = getattr(response, "data", None)
                 return getattr(data, "reaction_id", None)
-            logger.warning(
-                "[Feishu] Failed to add ack reaction to %s: code=%s msg=%s",
+            logger.debug(
+                "[Feishu] Add reaction %s on %s rejected: code=%s msg=%s",
+                emoji_type,
                 message_id,
                 getattr(response, "code", None),
                 getattr(response, "msg", None),
             )
         except Exception:
-            logger.warning("[Feishu] Failed to add ack reaction to %s", message_id, exc_info=True)
+            logger.warning(
+                "[Feishu] Add reaction %s on %s raised",
+                emoji_type,
+                message_id,
+                exc_info=True,
+            )
         return None
 
+    async def _remove_reaction(self, message_id: str, reaction_id: str) -> bool:
+        if not self._client or not message_id or not reaction_id:
+            return False
+        try:
+            from lark_oapi.api.im.v1 import DeleteMessageReactionRequest
+            request = (
+                DeleteMessageReactionRequest.builder()
+                .message_id(message_id)
+                .reaction_id(reaction_id)
+                .build()
+            )
+            response = await asyncio.to_thread(self._client.im.v1.message_reaction.delete, request)
+            if response and getattr(response, "success", lambda: False)():
+                return True
+            logger.debug(
+                "[Feishu] Remove reaction %s on %s rejected: code=%s msg=%s",
+                reaction_id,
+                message_id,
+                getattr(response, "code", None),
+                getattr(response, "msg", None),
+            )
+        except Exception:
+            logger.warning(
+                "[Feishu] Remove reaction %s on %s raised",
+                reaction_id,
+                message_id,
+                exc_info=True,
+            )
+        return False
+
+    def _remember_processing_reaction(self, message_id: str, reaction_id: str) -> None:
+        cache = self._pending_processing_reactions
+        cache[message_id] = reaction_id
+        cache.move_to_end(message_id)
+        while len(cache) > _FEISHU_PROCESSING_REACTION_CACHE_SIZE:
+            cache.popitem(last=False)
+
+    def _pop_processing_reaction(self, message_id: str) -> Optional[str]:
+        return self._pending_processing_reactions.pop(message_id, None)
+
+    async def on_processing_start(self, event: MessageEvent) -> None:
+        if not self._reactions_enabled():
+            return
+        message_id = event.message_id
+        if not message_id or message_id in self._pending_processing_reactions:
+            return
+        reaction_id = await self._add_reaction(message_id, _FEISHU_REACTION_IN_PROGRESS)
+        if reaction_id:
+            self._remember_processing_reaction(message_id, reaction_id)
+
+    async def on_processing_complete(
+        self, event: MessageEvent, outcome: ProcessingOutcome
+    ) -> None:
+        if not self._reactions_enabled():
+            return
+        message_id = event.message_id
+        if not message_id:
+            return
+
+        start_reaction_id = self._pending_processing_reactions.get(message_id)
+        if start_reaction_id:
+            if not await self._remove_reaction(message_id, start_reaction_id):
+                # Don't stack a second badge on top of a Typing we couldn't
+                # remove — UI would read as both "working" and "done/failed"
+                # simultaneously. Keep the handle so LRU eventually evicts it.
+                return
+            self._pop_processing_reaction(message_id)
+
+        if outcome is ProcessingOutcome.FAILURE:
+            await self._add_reaction(message_id, _FEISHU_REACTION_FAILURE)
+
     # =========================================================================
     # Webhook server and security
     # =========================================================================
@@ -2445,6 +2755,8 @@ class FeishuAdapter(BasePlatformAdapter):
             self._on_reaction_event(event_type, data)
         elif event_type == "card.action.trigger":
             self._on_card_action_trigger(data)
+        elif event_type == "drive.notice.comment_add_v1":
+            self._on_drive_comment_event(data)
         else:
             logger.debug("[Feishu] Ignoring webhook event type: %s", event_type or "unknown")
         return web.json_response({"code": 0, "msg": "ok"})
@@ -3079,6 +3391,23 @@ class FeishuAdapter(BasePlatformAdapter):
             return self._post_mentions_bot(normalized.mentioned_ids)
         return False
 
+    def _is_self_sent_bot_message(self, event: Any) -> bool:
+        """Return True only for Feishu events emitted by this Hermes bot."""
+        sender = getattr(event, "sender", None)
+        sender_type = str(getattr(sender, "sender_type", "") or "").strip().lower()
+        if sender_type not in {"bot", "app"}:
+            return False
+
+        sender_id = getattr(sender, "sender_id", None)
+        sender_open_id = str(getattr(sender_id, "open_id", "") or "").strip()
+        sender_user_id = str(getattr(sender_id, "user_id", "") or "").strip()
+
+        if self._bot_open_id and sender_open_id == self._bot_open_id:
+            return True
+        if self._bot_user_id and sender_user_id == self._bot_user_id:
+            return True
+        return False
+
     def _message_mentions_bot(self, mentions: List[Any]) -> bool:
         """Check whether any mention targets the configured or inferred bot identity."""
         for mention in mentions:
@@ -3106,10 +3435,55 @@ class FeishuAdapter(BasePlatformAdapter):
         return False
 
     async def _hydrate_bot_identity(self) -> None:
-        """Best-effort discovery of bot identity for precise group mention gating."""
+        """Best-effort discovery of bot identity for precise group mention gating
+        and self-sent bot event filtering.
+
+        Populates ``_bot_open_id`` and ``_bot_name`` from /open-apis/bot/v3/info
+        (no extra scopes required beyond the tenant access token). Falls back to
+        the application info endpoint for ``_bot_name`` only when the first probe
+        doesn't return it. Each field is hydrated independently — a value already
+        supplied via env vars (FEISHU_BOT_OPEN_ID / FEISHU_BOT_USER_ID /
+        FEISHU_BOT_NAME) is preserved and skips its probe.
+        """
         if not self._client:
             return
-        if any((self._bot_open_id, self._bot_user_id, self._bot_name)):
+        if self._bot_open_id and self._bot_name:
+            # Everything the self-send filter and precise mention gate need is
+            # already in place; nothing to probe.
+            return
+
+        # Primary probe: /open-apis/bot/v3/info — returns bot_name + open_id, no
+        # extra scopes required. This is the same endpoint the onboarding wizard
+        # uses via probe_bot().
+        if not self._bot_open_id or not self._bot_name:
+            try:
+                resp = await asyncio.to_thread(
+                    self._client.request,
+                    method="GET",
+                    url="/open-apis/bot/v3/info",
+                    body=None,
+                    raw_response=True,
+                )
+                content = getattr(resp, "content", None)
+                if content:
+                    payload = json.loads(content)
+                    parsed = _parse_bot_response(payload) or {}
+                    open_id = (parsed.get("bot_open_id") or "").strip()
+                    bot_name = (parsed.get("bot_name") or "").strip()
+                    if open_id and not self._bot_open_id:
+                        self._bot_open_id = open_id
+                    if bot_name and not self._bot_name:
+                        self._bot_name = bot_name
+            except Exception:
+                logger.debug(
+                    "[Feishu] /bot/v3/info probe failed during hydration",
+                    exc_info=True,
+                )
+
+        # Fallback probe for _bot_name only: application info endpoint. Needs
+        # admin:app.info:readonly or application:application:self_manage scope,
+        # so it's best-effort.
+        if self._bot_name:
             return
         try:
             request = self._build_get_application_request(app_id=self._app_id, lang="en_us")
@@ -3118,17 +3492,17 @@ class FeishuAdapter(BasePlatformAdapter):
                 code = getattr(response, "code", None)
                 if code == 99991672:
                     logger.warning(
-                        "[Feishu] Unable to hydrate bot identity from application info. "
+                        "[Feishu] Unable to hydrate bot name from application info. "
                         "Grant admin:app.info:readonly or application:application:self_manage "
                         "so group @mention gating can resolve the bot name precisely."
                     )
                 return
             app = getattr(getattr(response, "data", None), "app", None)
             app_name = (getattr(app, "app_name", None) or "").strip()
-            if app_name:
+            if app_name and not self._bot_name:
                 self._bot_name = app_name
         except Exception:
-            logger.debug("[Feishu] Failed to hydrate bot identity", exc_info=True)
+            logger.debug("[Feishu] Failed to hydrate bot name from application info", exc_info=True)
 
     # =========================================================================
     # Deduplication — seen message ID cache (persistent)

gateway/platforms/feishu_comment_rules.py

@@ -0,0 +1,429 @@
+"""
+Feishu document comment access-control rules.
+
+3-tier rule resolution: exact doc > wildcard "*" > top-level > code defaults.
+Each field (enabled/policy/allow_from) falls back independently.
+Config: ~/.hermes/feishu_comment_rules.json (mtime-cached, hot-reload).
+Pairing store: ~/.hermes/feishu_comment_pairing.json.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from hermes_constants import get_hermes_home
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Paths
+# ---------------------------------------------------------------------------
+#
+# Uses the canonical ``get_hermes_home()`` helper (HERMES_HOME-aware and
+# profile-safe). Resolved at import time; this module is lazy-imported by
+# the Feishu comment event handler, which runs long after profile overrides
+# have been applied, so freezing paths here is safe.
+
+RULES_FILE = get_hermes_home() / "feishu_comment_rules.json"
+PAIRING_FILE = get_hermes_home() / "feishu_comment_pairing.json"
+
+# ---------------------------------------------------------------------------
+# Data models
+# ---------------------------------------------------------------------------
+
+_VALID_POLICIES = ("allowlist", "pairing")
+
+
+@dataclass(frozen=True)
+class CommentDocumentRule:
+    """Per-document rule.  ``None`` means 'inherit from lower tier'."""
+    enabled: Optional[bool] = None
+    policy: Optional[str] = None
+    allow_from: Optional[frozenset] = None
+
+
+@dataclass(frozen=True)
+class CommentsConfig:
+    """Top-level comment access config."""
+    enabled: bool = True
+    policy: str = "pairing"
+    allow_from: frozenset = field(default_factory=frozenset)
+    documents: Dict[str, CommentDocumentRule] = field(default_factory=dict)
+
+
+@dataclass(frozen=True)
+class ResolvedCommentRule:
+    """Fully resolved rule after field-by-field fallback."""
+    enabled: bool
+    policy: str
+    allow_from: frozenset
+    match_source: str  # e.g. "exact:docx:xxx" | "wildcard" | "top" | "default"
+
+
+# ---------------------------------------------------------------------------
+# Mtime-cached file loading
+# ---------------------------------------------------------------------------
+
+class _MtimeCache:
+    """Generic mtime-based file cache.  ``stat()`` per access, re-read only on change."""
+
+    def __init__(self, path: Path):
+        self._path = path
+        self._mtime: float = 0.0
+        self._data: Optional[dict] = None
+
+    def load(self) -> dict:
+        try:
+            st = self._path.stat()
+            mtime = st.st_mtime
+        except FileNotFoundError:
+            self._mtime = 0.0
+            self._data = {}
+            return {}
+
+        if mtime == self._mtime and self._data is not None:
+            return self._data
+
+        try:
+            with open(self._path, "r", encoding="utf-8") as f:
+                data = json.load(f)
+            if not isinstance(data, dict):
+                data = {}
+        except (json.JSONDecodeError, OSError):
+            logger.warning("[Feishu-Rules] Failed to read %s, using empty config", self._path)
+            data = {}
+
+        self._mtime = mtime
+        self._data = data
+        return data
+
+
+_rules_cache = _MtimeCache(RULES_FILE)
+_pairing_cache = _MtimeCache(PAIRING_FILE)
+
+
+# ---------------------------------------------------------------------------
+# Config parsing
+# ---------------------------------------------------------------------------
+
+def _parse_frozenset(raw: Any) -> Optional[frozenset]:
+    """Parse a list of strings into a frozenset; return None if key absent."""
+    if raw is None:
+        return None
+    if isinstance(raw, (list, tuple)):
+        return frozenset(str(u).strip() for u in raw if str(u).strip())
+    return None
+
+
+def _parse_document_rule(raw: dict) -> CommentDocumentRule:
+    enabled = raw.get("enabled")
+    if enabled is not None:
+        enabled = bool(enabled)
+    policy = raw.get("policy")
+    if policy is not None:
+        policy = str(policy).strip().lower()
+        if policy not in _VALID_POLICIES:
+            policy = None
+    allow_from = _parse_frozenset(raw.get("allow_from"))
+    return CommentDocumentRule(enabled=enabled, policy=policy, allow_from=allow_from)
+
+
+def load_config() -> CommentsConfig:
+    """Load comment rules from disk (mtime-cached)."""
+    raw = _rules_cache.load()
+    if not raw:
+        return CommentsConfig()
+
+    documents: Dict[str, CommentDocumentRule] = {}
+    raw_docs = raw.get("documents", {})
+    if isinstance(raw_docs, dict):
+        for key, rule_raw in raw_docs.items():
+            if isinstance(rule_raw, dict):
+                documents[str(key)] = _parse_document_rule(rule_raw)
+
+    policy = str(raw.get("policy", "pairing")).strip().lower()
+    if policy not in _VALID_POLICIES:
+        policy = "pairing"
+
+    return CommentsConfig(
+        enabled=raw.get("enabled", True),
+        policy=policy,
+        allow_from=_parse_frozenset(raw.get("allow_from")) or frozenset(),
+        documents=documents,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Rule resolution  (§8.4 field-by-field fallback)
+# ---------------------------------------------------------------------------
+
+def has_wiki_keys(cfg: CommentsConfig) -> bool:
+    """Check if any document rule key starts with 'wiki:'."""
+    return any(k.startswith("wiki:") for k in cfg.documents)
+
+
+def resolve_rule(
+    cfg: CommentsConfig,
+    file_type: str,
+    file_token: str,
+    wiki_token: str = "",
+) -> ResolvedCommentRule:
+    """Resolve effective rule: exact doc → wiki key → wildcard → top-level → defaults."""
+    exact_key = f"{file_type}:{file_token}"
+
+    exact = cfg.documents.get(exact_key)
+    exact_src = f"exact:{exact_key}"
+    if exact is None and wiki_token:
+        wiki_key = f"wiki:{wiki_token}"
+        exact = cfg.documents.get(wiki_key)
+        exact_src = f"exact:{wiki_key}"
+
+    wildcard = cfg.documents.get("*")
+
+    layers = []
+    if exact is not None:
+        layers.append((exact, exact_src))
+    if wildcard is not None:
+        layers.append((wildcard, "wildcard"))
+
+    def _pick(field_name: str):
+        for layer, source in layers:
+            val = getattr(layer, field_name)
+            if val is not None:
+                return val, source
+        return getattr(cfg, field_name), "top"
+
+    enabled, en_src = _pick("enabled")
+    policy, pol_src = _pick("policy")
+    allow_from, _ = _pick("allow_from")
+
+    # match_source = highest-priority tier that contributed any field
+    priority_order = {"exact": 0, "wildcard": 1, "top": 2}
+    best_src = min(
+        [en_src, pol_src],
+        key=lambda s: priority_order.get(s.split(":")[0], 3),
+    )
+
+    return ResolvedCommentRule(
+        enabled=enabled,
+        policy=policy,
+        allow_from=allow_from,
+        match_source=best_src,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Pairing store
+# ---------------------------------------------------------------------------
+
+def _load_pairing_approved() -> set:
+    """Return set of approved user open_ids (mtime-cached)."""
+    data = _pairing_cache.load()
+    approved = data.get("approved", {})
+    if isinstance(approved, dict):
+        return set(approved.keys())
+    if isinstance(approved, list):
+        return set(str(u) for u in approved if u)
+    return set()
+
+
+def _save_pairing(data: dict) -> None:
+    PAIRING_FILE.parent.mkdir(parents=True, exist_ok=True)
+    tmp = PAIRING_FILE.with_suffix(".tmp")
+    with open(tmp, "w", encoding="utf-8") as f:
+        json.dump(data, f, indent=2, ensure_ascii=False)
+    tmp.replace(PAIRING_FILE)
+    # Invalidate cache so next load picks up change
+    _pairing_cache._mtime = 0.0
+    _pairing_cache._data = None
+
+
+def pairing_add(user_open_id: str) -> bool:
+    """Add a user to the pairing-approved list. Returns True if newly added."""
+    data = _pairing_cache.load()
+    approved = data.get("approved", {})
+    if not isinstance(approved, dict):
+        approved = {}
+    if user_open_id in approved:
+        return False
+    approved[user_open_id] = {"approved_at": time.time()}
+    data["approved"] = approved
+    _save_pairing(data)
+    return True
+
+
+def pairing_remove(user_open_id: str) -> bool:
+    """Remove a user from the pairing-approved list. Returns True if removed."""
+    data = _pairing_cache.load()
+    approved = data.get("approved", {})
+    if not isinstance(approved, dict):
+        return False
+    if user_open_id not in approved:
+        return False
+    del approved[user_open_id]
+    data["approved"] = approved
+    _save_pairing(data)
+    return True
+
+
+def pairing_list() -> Dict[str, Any]:
+    """Return the approved dict  {user_open_id: {approved_at: ...}}."""
+    data = _pairing_cache.load()
+    approved = data.get("approved", {})
+    return dict(approved) if isinstance(approved, dict) else {}
+
+
+# ---------------------------------------------------------------------------
+# Access check  (public API for feishu_comment.py)
+# ---------------------------------------------------------------------------
+
+def is_user_allowed(rule: ResolvedCommentRule, user_open_id: str) -> bool:
+    """Check if user passes the resolved rule's policy gate."""
+    if user_open_id in rule.allow_from:
+        return True
+    if rule.policy == "pairing":
+        return user_open_id in _load_pairing_approved()
+    return False
+
+
+# ---------------------------------------------------------------------------
+# CLI
+# ---------------------------------------------------------------------------
+
+def _print_status() -> None:
+    cfg = load_config()
+    print(f"Rules file: {RULES_FILE}")
+    print(f"  exists: {RULES_FILE.exists()}")
+    print(f"Pairing file: {PAIRING_FILE}")
+    print(f"  exists: {PAIRING_FILE.exists()}")
+    print()
+    print(f"Top-level:")
+    print(f"  enabled:    {cfg.enabled}")
+    print(f"  policy:     {cfg.policy}")
+    print(f"  allow_from: {sorted(cfg.allow_from) if cfg.allow_from else '[]'}")
+    print()
+    if cfg.documents:
+        print(f"Document rules ({len(cfg.documents)}):")
+        for key, rule in sorted(cfg.documents.items()):
+            parts = []
+            if rule.enabled is not None:
+                parts.append(f"enabled={rule.enabled}")
+            if rule.policy is not None:
+                parts.append(f"policy={rule.policy}")
+            if rule.allow_from is not None:
+                parts.append(f"allow_from={sorted(rule.allow_from)}")
+            print(f"  [{key}] {', '.join(parts) if parts else '(empty — inherits all)'}")
+    else:
+        print("Document rules: (none)")
+    print()
+    approved = pairing_list()
+    print(f"Pairing approved ({len(approved)}):")
+    for uid, meta in sorted(approved.items()):
+        ts = meta.get("approved_at", 0)
+        print(f"  {uid}  (approved_at={ts})")
+
+
+def _do_check(doc_key: str, user_open_id: str) -> None:
+    cfg = load_config()
+    parts = doc_key.split(":", 1)
+    if len(parts) != 2:
+        print(f"Error: doc_key must be 'fileType:fileToken', got '{doc_key}'")
+        return
+    file_type, file_token = parts
+    rule = resolve_rule(cfg, file_type, file_token)
+    allowed = is_user_allowed(rule, user_open_id)
+    print(f"Document:     {doc_key}")
+    print(f"User:         {user_open_id}")
+    print(f"Resolved rule:")
+    print(f"  enabled:      {rule.enabled}")
+    print(f"  policy:       {rule.policy}")
+    print(f"  allow_from:   {sorted(rule.allow_from) if rule.allow_from else '[]'}")
+    print(f"  match_source: {rule.match_source}")
+    print(f"Result:       {'ALLOWED' if allowed else 'DENIED'}")
+
+
+def _main() -> int:
+    import sys
+
+    try:
+        from hermes_cli.env_loader import load_hermes_dotenv
+        load_hermes_dotenv()
+    except Exception:
+        pass
+
+    usage = (
+        "Usage: python -m gateway.platforms.feishu_comment_rules <command> [args]\n"
+        "\n"
+        "Commands:\n"
+        "  status                              Show rules config and pairing state\n"
+        "  check <fileType:token> <user>        Simulate access check\n"
+        "  pairing add <user_open_id>           Add user to pairing-approved list\n"
+        "  pairing remove <user_open_id>        Remove user from pairing-approved list\n"
+        "  pairing list                         List pairing-approved users\n"
+        "\n"
+        f"Rules config file: {RULES_FILE}\n"
+        "  Edit this JSON file directly to configure policies and document rules.\n"
+        "  Changes take effect on the next comment event (no restart needed).\n"
+    )
+
+    args = sys.argv[1:]
+    if not args:
+        print(usage)
+        return 1
+
+    cmd = args[0]
+
+    if cmd == "status":
+        _print_status()
+
+    elif cmd == "check":
+        if len(args) < 3:
+            print("Usage: check <fileType:fileToken> <user_open_id>")
+            return 1
+        _do_check(args[1], args[2])
+
+    elif cmd == "pairing":
+        if len(args) < 2:
+            print("Usage: pairing <add|remove|list> [args]")
+            return 1
+        sub = args[1]
+        if sub == "add":
+            if len(args) < 3:
+                print("Usage: pairing add <user_open_id>")
+                return 1
+            if pairing_add(args[2]):
+                print(f"Added: {args[2]}")
+            else:
+                print(f"Already approved: {args[2]}")
+        elif sub == "remove":
+            if len(args) < 3:
+                print("Usage: pairing remove <user_open_id>")
+                return 1
+            if pairing_remove(args[2]):
+                print(f"Removed: {args[2]}")
+            else:
+                print(f"Not in approved list: {args[2]}")
+        elif sub == "list":
+            approved = pairing_list()
+            if not approved:
+                print("(no approved users)")
+            for uid, meta in sorted(approved.items()):
+                print(f"  {uid}  approved_at={meta.get('approved_at', '?')}")
+        else:
+            print(f"Unknown pairing subcommand: {sub}")
+            return 1
+    else:
+        print(f"Unknown command: {cmd}\n")
+        print(usage)
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    import sys
+    sys.exit(_main())

gateway/platforms/helpers.py

@@ -49,7 +49,10 @@ class MessageDeduplicator:
             return False
         now = time.time()
         if msg_id in self._seen:
-            return True
+            if now - self._seen[msg_id] < self._ttl:
+                return True
+            # Entry has expired — remove it and treat as new
+            del self._seen[msg_id]
         self._seen[msg_id] = now
         if len(self._seen) > self._max_size:
             cutoff = now - self._ttl

gateway/platforms/mattermost.py

@@ -304,7 +304,7 @@ class MattermostAdapter(BasePlatformAdapter):
         )
 
     async def edit_message(
-        self, chat_id: str, message_id: str, content: str
+        self, chat_id: str, message_id: str, content: str, *, finalize: bool = False
     ) -> SendResult:
         """Edit an existing post."""
         formatted = self.format_message(content)
@@ -718,6 +718,12 @@ class MattermostAdapter(BasePlatformAdapter):
             thread_id=thread_id,
         )
 
+        # Per-channel ephemeral prompt
+        from gateway.platforms.base import resolve_channel_prompt
+        _channel_prompt = resolve_channel_prompt(
+            self.config.extra, channel_id, None,
+        )
+
         msg_event = MessageEvent(
             text=message_text,
             message_type=msg_type,
@@ -726,6 +732,7 @@ class MattermostAdapter(BasePlatformAdapter):
             message_id=post_id,
             media_urls=media_urls if media_urls else None,
             media_types=media_types if media_types else None,
+            channel_prompt=_channel_prompt,
         )
 
         await self.handle_message(msg_event)

gateway/platforms/qqbot/__init__.py

@@ -0,0 +1,57 @@
+"""
+QQBot platform package.
+
+Re-exports the main adapter symbols from ``adapter.py`` (the original
+``qqbot.py``) so that **all existing import paths remain unchanged**::
+
+    from gateway.platforms.qqbot import QQAdapter          # works
+    from gateway.platforms.qqbot import check_qq_requirements  # works
+
+New modules:
+    - ``constants`` — shared constants (API URLs, timeouts, message types)
+    - ``utils`` — User-Agent builder, config helpers
+    - ``crypto`` — AES-256-GCM key generation and decryption
+    - ``onboard`` — QR-code scan-to-configure flow
+"""
+
+# -- Adapter (original qqbot.py) ------------------------------------------
+from .adapter import (  # noqa: F401
+    QQAdapter,
+    QQCloseError,
+    check_qq_requirements,
+    _coerce_list,
+    _ssrf_redirect_guard,
+)
+
+# -- Onboard (QR-code scan-to-configure) -----------------------------------
+from .onboard import (  # noqa: F401
+    BindStatus,
+    create_bind_task,
+    poll_bind_result,
+    build_connect_url,
+)
+from .crypto import decrypt_secret, generate_bind_key  # noqa: F401
+
+# -- Utils -----------------------------------------------------------------
+from .utils import build_user_agent, get_api_headers, coerce_list  # noqa: F401
+
+__all__ = [
+    # adapter
+    "QQAdapter",
+    "QQCloseError",
+    "check_qq_requirements",
+    "_coerce_list",
+    "_ssrf_redirect_guard",
+    # onboard
+    "BindStatus",
+    "create_bind_task",
+    "poll_bind_result",
+    "build_connect_url",
+    # crypto
+    "decrypt_secret",
+    "generate_bind_key",
+    # utils
+    "build_user_agent",
+    "get_api_headers",
+    "coerce_list",
+]

gateway/platforms/qqbot/constants.py

@@ -0,0 +1,74 @@
+"""QQBot package-level constants shared across adapter, onboard, and other modules."""
+
+from __future__ import annotations
+
+import os
+
+# ---------------------------------------------------------------------------
+# QQBot adapter version — bump on functional changes to the adapter package.
+# ---------------------------------------------------------------------------
+
+QQBOT_VERSION = "1.1.0"
+
+# ---------------------------------------------------------------------------
+# API endpoints
+# ---------------------------------------------------------------------------
+
+# The portal domain is configurable via QQ_API_HOST for corporate proxies
+# or test environments.  Default: q.qq.com (production).
+PORTAL_HOST = os.getenv("QQ_PORTAL_HOST", "q.qq.com")
+
+API_BASE = "https://api.sgroup.qq.com"
+TOKEN_URL = "https://bots.qq.com/app/getAppAccessToken"
+GATEWAY_URL_PATH = "/gateway"
+
+# QR-code onboard endpoints (on the portal host)
+ONBOARD_CREATE_PATH = "/lite/create_bind_task"
+ONBOARD_POLL_PATH = "/lite/poll_bind_result"
+QR_URL_TEMPLATE = (
+    "https://q.qq.com/qqbot/openclaw/connect.html"
+    "?task_id={task_id}&_wv=2&source=hermes"
+)
+
+# ---------------------------------------------------------------------------
+# Timeouts & retry
+# ---------------------------------------------------------------------------
+
+DEFAULT_API_TIMEOUT = 30.0
+FILE_UPLOAD_TIMEOUT = 120.0
+CONNECT_TIMEOUT_SECONDS = 20.0
+
+RECONNECT_BACKOFF = [2, 5, 10, 30, 60]
+MAX_RECONNECT_ATTEMPTS = 100
+RATE_LIMIT_DELAY = 60  # seconds
+QUICK_DISCONNECT_THRESHOLD = 5.0  # seconds
+MAX_QUICK_DISCONNECT_COUNT = 3
+
+ONBOARD_POLL_INTERVAL = 2.0  # seconds between poll_bind_result calls
+ONBOARD_API_TIMEOUT = 10.0
+
+# ---------------------------------------------------------------------------
+# Message limits
+# ---------------------------------------------------------------------------
+
+MAX_MESSAGE_LENGTH = 4000
+DEDUP_WINDOW_SECONDS = 300
+DEDUP_MAX_SIZE = 1000
+
+# ---------------------------------------------------------------------------
+# QQ Bot message types
+# ---------------------------------------------------------------------------
+
+MSG_TYPE_TEXT = 0
+MSG_TYPE_MARKDOWN = 2
+MSG_TYPE_MEDIA = 7
+MSG_TYPE_INPUT_NOTIFY = 6
+
+# ---------------------------------------------------------------------------
+# QQ Bot file media types
+# ---------------------------------------------------------------------------
+
+MEDIA_TYPE_IMAGE = 1
+MEDIA_TYPE_VIDEO = 2
+MEDIA_TYPE_VOICE = 3
+MEDIA_TYPE_FILE = 4

gateway/platforms/qqbot/crypto.py

@@ -0,0 +1,45 @@
+"""AES-256-GCM utilities for QQBot scan-to-configure credential decryption."""
+
+from __future__ import annotations
+
+import base64
+import os
+
+
+def generate_bind_key() -> str:
+    """Generate a 256-bit random AES key and return it as base64.
+
+    The key is passed to ``create_bind_task`` so the server can encrypt
+    the bot's *client_secret* before returning it.  Only this CLI holds
+    the key, ensuring the secret never travels in plaintext.
+    """
+    return base64.b64encode(os.urandom(32)).decode()
+
+
+def decrypt_secret(encrypted_base64: str, key_base64: str) -> str:
+    """Decrypt a base64-encoded AES-256-GCM ciphertext.
+
+    Ciphertext layout (after base64-decoding)::
+
+        IV (12 bytes) ‖ ciphertext (N bytes) ‖ AuthTag (16 bytes)
+
+    Args:
+        encrypted_base64: The ``bot_encrypt_secret`` value from
+            ``poll_bind_result``.
+        key_base64: The base64 AES key generated by
+            :func:`generate_bind_key`.
+
+    Returns:
+        The decrypted *client_secret* as a UTF-8 string.
+    """
+    from cryptography.hazmat.primitives.ciphers.aead import AESGCM
+
+    key = base64.b64decode(key_base64)
+    raw = base64.b64decode(encrypted_base64)
+
+    iv = raw[:12]
+    ciphertext_with_tag = raw[12:]  # AESGCM expects ciphertext + tag concatenated
+
+    aesgcm = AESGCM(key)
+    plaintext = aesgcm.decrypt(iv, ciphertext_with_tag, None)
+    return plaintext.decode("utf-8")

gateway/platforms/qqbot/onboard.py

@@ -0,0 +1,124 @@
+"""
+QQBot scan-to-configure (QR code onboard) module.
+
+Calls the ``q.qq.com`` ``create_bind_task`` / ``poll_bind_result`` APIs to
+generate a QR-code URL and poll for scan completion.  On success the caller
+receives the bot's *app_id*, *client_secret* (decrypted locally), and the
+scanner's *user_openid* — enough to fully configure the QQBot gateway.
+
+Reference: https://bot.q.qq.com/wiki/develop/api-v2/
+"""
+
+from __future__ import annotations
+
+import logging
+from enum import IntEnum
+from typing import Tuple
+from urllib.parse import quote
+
+from .constants import (
+    ONBOARD_API_TIMEOUT,
+    ONBOARD_CREATE_PATH,
+    ONBOARD_POLL_PATH,
+    PORTAL_HOST,
+    QR_URL_TEMPLATE,
+)
+from .crypto import generate_bind_key
+from .utils import get_api_headers
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Bind status
+# ---------------------------------------------------------------------------
+
+
+class BindStatus(IntEnum):
+    """Status codes returned by ``poll_bind_result``."""
+
+    NONE = 0
+    PENDING = 1
+    COMPLETED = 2
+    EXPIRED = 3
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+
+async def create_bind_task(
+    timeout: float = ONBOARD_API_TIMEOUT,
+) -> Tuple[str, str]:
+    """Create a bind task and return *(task_id, aes_key_base64)*.
+
+    The AES key is generated locally and sent to the server so it can
+    encrypt the bot credentials before returning them.
+
+    Raises:
+        RuntimeError: If the API returns a non-zero ``retcode``.
+    """
+    import httpx
+
+    url = f"https://{PORTAL_HOST}{ONBOARD_CREATE_PATH}"
+    key = generate_bind_key()
+
+    async with httpx.AsyncClient(timeout=timeout, follow_redirects=True) as client:
+        resp = await client.post(url, json={"key": key}, headers=get_api_headers())
+        resp.raise_for_status()
+        data = resp.json()
+
+    if data.get("retcode") != 0:
+        raise RuntimeError(data.get("msg", "create_bind_task failed"))
+
+    task_id = data.get("data", {}).get("task_id")
+    if not task_id:
+        raise RuntimeError("create_bind_task: missing task_id in response")
+
+    logger.debug("create_bind_task ok: task_id=%s", task_id)
+    return task_id, key
+
+
+async def poll_bind_result(
+    task_id: str,
+    timeout: float = ONBOARD_API_TIMEOUT,
+) -> Tuple[BindStatus, str, str, str]:
+    """Poll the bind result for *task_id*.
+
+    Returns:
+        A 4-tuple of ``(status, bot_appid, bot_encrypt_secret, user_openid)``.
+
+        * ``bot_encrypt_secret`` is AES-256-GCM encrypted — decrypt it with
+          :func:`~gateway.platforms.qqbot.crypto.decrypt_secret` using the
+          key from :func:`create_bind_task`.
+        * ``user_openid`` is the OpenID of the person who scanned the code
+          (available when ``status == COMPLETED``).
+
+    Raises:
+        RuntimeError: If the API returns a non-zero ``retcode``.
+    """
+    import httpx
+
+    url = f"https://{PORTAL_HOST}{ONBOARD_POLL_PATH}"
+
+    async with httpx.AsyncClient(timeout=timeout, follow_redirects=True) as client:
+        resp = await client.post(url, json={"task_id": task_id}, headers=get_api_headers())
+        resp.raise_for_status()
+        data = resp.json()
+
+    if data.get("retcode") != 0:
+        raise RuntimeError(data.get("msg", "poll_bind_result failed"))
+
+    d = data.get("data", {})
+    return (
+        BindStatus(d.get("status", 0)),
+        str(d.get("bot_appid", "")),
+        d.get("bot_encrypt_secret", ""),
+        d.get("user_openid", ""),
+    )
+
+
+def build_connect_url(task_id: str) -> str:
+    """Build the QR-code target URL for a given *task_id*."""
+    return QR_URL_TEMPLATE.format(task_id=quote(task_id))

gateway/platforms/qqbot/utils.py

@@ -0,0 +1,71 @@
+"""QQBot shared utilities — User-Agent, HTTP helpers, config coercion."""
+
+from __future__ import annotations
+
+import platform
+import sys
+from typing import Any, Dict, List
+
+from .constants import QQBOT_VERSION
+
+
+# ---------------------------------------------------------------------------
+# User-Agent
+# ---------------------------------------------------------------------------
+
+def _get_hermes_version() -> str:
+    """Return the hermes-agent package version, or 'dev' if unavailable."""
+    try:
+        from importlib.metadata import version
+        return version("hermes-agent")
+    except Exception:
+        return "dev"
+
+
+def build_user_agent() -> str:
+    """Build a descriptive User-Agent string.
+
+    Format::
+
+        QQBotAdapter/<qqbot_version> (Python/<py_version>; <os>; Hermes/<hermes_version>)
+
+    Example::
+
+        QQBotAdapter/1.0.0 (Python/3.11.15; darwin; Hermes/0.9.0)
+    """
+    py_version = f"{sys.version_info.major}.{sys.version_info.minor}.{sys.version_info.micro}"
+    os_name = platform.system().lower()
+    hermes_version = _get_hermes_version()
+    return f"QQBotAdapter/{QQBOT_VERSION} (Python/{py_version}; {os_name}; Hermes/{hermes_version})"
+
+
+def get_api_headers() -> Dict[str, str]:
+    """Return standard HTTP headers for QQBot API requests.
+
+    Includes ``Content-Type``, ``Accept``, and a dynamic ``User-Agent``.
+    ``q.qq.com`` requires ``Accept: application/json`` — without it,
+    the server returns a JavaScript anti-bot challenge page.
+    """
+    return {
+        "Content-Type": "application/json",
+        "Accept": "application/json",
+        "User-Agent": build_user_agent(),
+    }
+
+
+# ---------------------------------------------------------------------------
+# Config helpers
+# ---------------------------------------------------------------------------
+
+def coerce_list(value: Any) -> List[str]:
+    """Coerce config values into a trimmed string list.
+
+    Accepts comma-separated strings, lists, tuples, sets, or single values.
+    """
+    if value is None:
+        return []
+    if isinstance(value, str):
+        return [item.strip() for item in value.split(",") if item.strip()]
+    if isinstance(value, (list, tuple, set)):
+        return [str(item).strip() for item in value if str(item).strip()]
+    return [str(value).strip()] if str(value).strip() else []

gateway/platforms/signal.py

@@ -18,6 +18,7 @@ import logging
 import os
 import random
 import time
+import uuid
 from datetime import datetime, timezone
 from pathlib import Path
 from typing import Dict, List, Optional, Any
@@ -127,6 +128,27 @@ def _render_mentions(text: str, mentions: list) -> str:
     return text
 
 
+def _is_signal_service_id(value: str) -> bool:
+    """Return True if *value* already looks like a Signal service identifier."""
+    if not value:
+        return False
+    if value.startswith("PNI:") or value.startswith("u:"):
+        return True
+    try:
+        uuid.UUID(value)
+        return True
+    except (ValueError, AttributeError, TypeError):
+        return False
+
+
+def _looks_like_e164_number(value: str) -> bool:
+    """Return True for a plausible E.164 phone number."""
+    if not value or not value.startswith("+"):
+        return False
+    digits = value[1:]
+    return digits.isdigit() and 7 <= len(digits) <= 15
+
+
 def check_signal_requirements() -> bool:
     """Check if Signal is configured (has URL and account)."""
     return bool(os.getenv("SIGNAL_HTTP_URL") and os.getenv("SIGNAL_ACCOUNT"))
@@ -160,6 +182,14 @@ class SignalAdapter(BasePlatformAdapter):
         self._sse_task: Optional[asyncio.Task] = None
         self._health_monitor_task: Optional[asyncio.Task] = None
         self._typing_tasks: Dict[str, asyncio.Task] = {}
+        # Per-chat typing-indicator backoff. When signal-cli reports
+        # NETWORK_FAILURE (recipient offline / unroutable), base.py's
+        # _keep_typing refresh loop would otherwise hammer sendTyping every
+        # ~2s indefinitely, producing WARNING-level log spam and pointless
+        # RPC traffic. We track consecutive failures per chat and skip the
+        # RPC during a cooldown window instead.
+        self._typing_failures: Dict[str, int] = {}
+        self._typing_skip_until: Dict[str, float] = {}
         self._running = False
         self._last_sse_activity = 0.0
         self._sse_response: Optional[httpx.Response] = None
@@ -171,6 +201,12 @@ class SignalAdapter(BasePlatformAdapter):
         # in Note to Self / self-chat mode (mirrors WhatsApp recentlySentIds)
         self._recent_sent_timestamps: set = set()
         self._max_recent_timestamps = 50
+        # Signal increasingly exposes ACI/PNI UUIDs as stable recipient IDs.
+        # Keep a best-effort mapping so outbound sends can upgrade from a
+        # phone number to the corresponding UUID when signal-cli prefers it.
+        self._recipient_uuid_by_number: Dict[str, str] = {}
+        self._recipient_number_by_uuid: Dict[str, str] = {}
+        self._recipient_cache_lock = asyncio.Lock()
 
         logger.info("Signal adapter initialized: url=%s account=%s groups=%s",
                      self.http_url, redact_phone(self.account),
@@ -187,31 +223,40 @@ class SignalAdapter(BasePlatformAdapter):
             return False
 
         # Acquire scoped lock to prevent duplicate Signal listeners for the same phone
+        lock_acquired = False
         try:
             if not self._acquire_platform_lock('signal-phone', self.account, 'Signal account'):
                 return False
+            lock_acquired = True
         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
         try:
-            resp = await self.client.get(f"{self.http_url}/api/v1/check", timeout=10.0)
-            if resp.status_code != 200:
-                logger.error("Signal: health check failed (status %d)", resp.status_code)
+            # Health check — verify signal-cli daemon is reachable
+            try:
+                resp = await self.client.get(f"{self.http_url}/api/v1/check", timeout=10.0)
+                if resp.status_code != 200:
+                    logger.error("Signal: health check failed (status %d)", resp.status_code)
+                    return False
+            except Exception as e:
+                logger.error("Signal: cannot reach signal-cli at %s: %s", self.http_url, e)
                 return False
-        except Exception as e:
-            logger.error("Signal: cannot reach signal-cli at %s: %s", self.http_url, e)
-            return False
 
-        self._running = True
-        self._last_sse_activity = time.time()
-        self._sse_task = asyncio.create_task(self._sse_listener())
-        self._health_monitor_task = asyncio.create_task(self._health_monitor())
+            self._running = True
+            self._last_sse_activity = time.time()
+            self._sse_task = asyncio.create_task(self._sse_listener())
+            self._health_monitor_task = asyncio.create_task(self._health_monitor())
 
-        logger.info("Signal: connected to %s", self.http_url)
-        return True
+            logger.info("Signal: connected to %s", self.http_url)
+            return True
+        finally:
+            if not self._running:
+                if self.client:
+                    await self.client.aclose()
+                    self.client = None
+                if lock_acquired:
+                    self._release_platform_lock()
 
     async def disconnect(self) -> None:
         """Stop SSE listener and clean up."""
@@ -392,6 +437,7 @@ class SignalAdapter(BasePlatformAdapter):
         )
         sender_name = envelope_data.get("sourceName", "")
         sender_uuid = envelope_data.get("sourceUuid", "")
+        self._remember_recipient_identifiers(sender, sender_uuid)
 
         if not sender:
             logger.debug("Signal: ignoring envelope with no sender")
@@ -510,6 +556,64 @@ class SignalAdapter(BasePlatformAdapter):
 
         await self.handle_message(event)
 
+    def _remember_recipient_identifiers(self, number: Optional[str], service_id: Optional[str]) -> None:
+        """Cache any number↔UUID mapping observed from Signal envelopes."""
+        if not number or not service_id or not _is_signal_service_id(service_id):
+            return
+        self._recipient_uuid_by_number[number] = service_id
+        self._recipient_number_by_uuid[service_id] = number
+
+    def _extract_contact_uuid(self, contact: Any, phone_number: str) -> Optional[str]:
+        """Best-effort extraction of a Signal service ID from listContacts output."""
+        if not isinstance(contact, dict):
+            return None
+
+        number = contact.get("number")
+        recipient = contact.get("recipient")
+        service_id = contact.get("uuid") or contact.get("serviceId")
+        if not service_id:
+            profile = contact.get("profile")
+            if isinstance(profile, dict):
+                service_id = profile.get("serviceId") or profile.get("uuid")
+
+        if service_id and _is_signal_service_id(service_id):
+            matches_number = number == phone_number or recipient == phone_number
+            if matches_number:
+                return service_id
+        return None
+
+    async def _resolve_recipient(self, chat_id: str) -> str:
+        """Return the preferred Signal recipient identifier for a direct chat."""
+        if (
+            not chat_id
+            or chat_id.startswith("group:")
+            or _is_signal_service_id(chat_id)
+            or not _looks_like_e164_number(chat_id)
+        ):
+            return chat_id
+
+        cached = self._recipient_uuid_by_number.get(chat_id)
+        if cached:
+            return cached
+
+        async with self._recipient_cache_lock:
+            cached = self._recipient_uuid_by_number.get(chat_id)
+            if cached:
+                return cached
+
+            contacts = await self._rpc("listContacts", {
+                "account": self.account,
+                "allRecipients": True,
+            })
+            if isinstance(contacts, list):
+                for contact in contacts:
+                    number = contact.get("number") if isinstance(contact, dict) else None
+                    service_id = self._extract_contact_uuid(contact, chat_id)
+                    if number and service_id:
+                        self._remember_recipient_identifiers(number, service_id)
+
+            return self._recipient_uuid_by_number.get(chat_id, chat_id)
+
     # ------------------------------------------------------------------
     # Attachment Handling
     # ------------------------------------------------------------------
@@ -548,8 +652,22 @@ class SignalAdapter(BasePlatformAdapter):
     # JSON-RPC Communication
     # ------------------------------------------------------------------
 
-    async def _rpc(self, method: str, params: dict, rpc_id: str = None) -> Any:
-        """Send a JSON-RPC 2.0 request to signal-cli daemon."""
+    async def _rpc(
+        self,
+        method: str,
+        params: dict,
+        rpc_id: str = None,
+        *,
+        log_failures: bool = True,
+    ) -> Any:
+        """Send a JSON-RPC 2.0 request to signal-cli daemon.
+
+        When ``log_failures=False``, error and exception paths log at DEBUG
+        instead of WARNING — used by the typing-indicator path to silence
+        repeated NETWORK_FAILURE spam for unreachable recipients while
+        still preserving visibility for the first occurrence and for
+        unrelated RPCs.
+        """
         if not self.client:
             logger.warning("Signal: RPC called but client not connected")
             return None
@@ -574,13 +692,19 @@ class SignalAdapter(BasePlatformAdapter):
             data = resp.json()
 
             if "error" in data:
-                logger.warning("Signal RPC error (%s): %s", method, data["error"])
+                if log_failures:
+                    logger.warning("Signal RPC error (%s): %s", method, data["error"])
+                else:
+                    logger.debug("Signal RPC error (%s): %s", method, data["error"])
                 return None
 
             return data.get("result")
 
         except Exception as e:
-            logger.warning("Signal RPC %s failed: %s", method, e)
+            if log_failures:
+                logger.warning("Signal RPC %s failed: %s", method, e)
+            else:
+                logger.debug("Signal RPC %s failed: %s", method, e)
             return None
 
     # ------------------------------------------------------------------
@@ -605,7 +729,7 @@ class SignalAdapter(BasePlatformAdapter):
         if chat_id.startswith("group:"):
             params["groupId"] = chat_id[6:]
         else:
-            params["recipient"] = [chat_id]
+            params["recipient"] = [await self._resolve_recipient(chat_id)]
 
         result = await self._rpc("send", params)
 
@@ -627,7 +751,28 @@ class SignalAdapter(BasePlatformAdapter):
                 self._recent_sent_timestamps.pop()
 
     async def send_typing(self, chat_id: str, metadata=None) -> None:
-        """Send a typing indicator."""
+        """Send a typing indicator.
+
+        base.py's ``_keep_typing`` refresh loop calls this every ~2s while
+        the agent is processing. If signal-cli returns NETWORK_FAILURE for
+        this recipient (offline, unroutable, group membership lost, etc.)
+        the unmitigated behaviour is: a WARNING log every 2 seconds for as
+        long as the agent keeps running. Instead we:
+
+        - silence the WARNING after the first consecutive failure (subsequent
+          attempts log at DEBUG) so transport issues are still visible once
+          but don't flood the log,
+        - skip the RPC entirely during an exponential cooldown window once
+          three consecutive failures have happened, so we stop hammering
+          signal-cli with requests it can't deliver.
+
+        A successful sendTyping clears the counters.
+        """
+        now = time.monotonic()
+        skip_until = self._typing_skip_until.get(chat_id, 0.0)
+        if now < skip_until:
+            return
+
         params: Dict[str, Any] = {
             "account": self.account,
         }
@@ -635,9 +780,28 @@ class SignalAdapter(BasePlatformAdapter):
         if chat_id.startswith("group:"):
             params["groupId"] = chat_id[6:]
         else:
-            params["recipient"] = [chat_id]
+            params["recipient"] = [await self._resolve_recipient(chat_id)]
 
-        await self._rpc("sendTyping", params, rpc_id="typing")
+        fails = self._typing_failures.get(chat_id, 0)
+        result = await self._rpc(
+            "sendTyping",
+            params,
+            rpc_id="typing",
+            log_failures=(fails == 0),
+        )
+
+        if result is None:
+            fails += 1
+            self._typing_failures[chat_id] = fails
+            # After 3 consecutive failures, back off exponentially (16s,
+            # 32s, 60s cap) to stop spamming signal-cli for a recipient
+            # that clearly isn't reachable right now.
+            if fails >= 3:
+                backoff = min(60.0, 16.0 * (2 ** (fails - 3)))
+                self._typing_skip_until[chat_id] = now + backoff
+        else:
+            self._typing_failures.pop(chat_id, None)
+            self._typing_skip_until.pop(chat_id, None)
 
     async def send_image(
         self,
@@ -677,7 +841,7 @@ class SignalAdapter(BasePlatformAdapter):
         if chat_id.startswith("group:"):
             params["groupId"] = chat_id[6:]
         else:
-            params["recipient"] = [chat_id]
+            params["recipient"] = [await self._resolve_recipient(chat_id)]
 
         result = await self._rpc("send", params)
         if result is not None:
@@ -716,7 +880,7 @@ class SignalAdapter(BasePlatformAdapter):
         if chat_id.startswith("group:"):
             params["groupId"] = chat_id[6:]
         else:
-            params["recipient"] = [chat_id]
+            params["recipient"] = [await self._resolve_recipient(chat_id)]
 
         result = await self._rpc("send", params)
         if result is not None:
@@ -789,6 +953,10 @@ class SignalAdapter(BasePlatformAdapter):
                 await task
             except asyncio.CancelledError:
                 pass
+        # Reset per-chat typing backoff state so the next agent turn starts
+        # fresh rather than inheriting a cooldown from a prior conversation.
+        self._typing_failures.pop(chat_id, None)
+        self._typing_skip_until.pop(chat_id, None)
 
     async def stop_typing(self, chat_id: str) -> None:
         """Public interface for stopping typing — called by base adapter's

gateway/platforms/slack.py

@@ -150,9 +150,11 @@ class SlackAdapter(BasePlatformAdapter):
             except Exception as e:
                 logger.warning("[Slack] Failed to read %s: %s", tokens_file, e)
 
+        lock_acquired = False
         try:
             if not self._acquire_platform_lock('slack-app-token', app_token, 'Slack app token'):
                 return False
+            lock_acquired = True
 
             # First token is the primary — used for AsyncApp / Socket Mode
             primary_token = bot_tokens[0]
@@ -228,6 +230,9 @@ class SlackAdapter(BasePlatformAdapter):
         except Exception as e:  # pragma: no cover - defensive logging
             logger.error("[Slack] Connection failed: %s", e, exc_info=True)
             return False
+        finally:
+            if lock_acquired and not self._running:
+                self._release_platform_lock()
 
     async def disconnect(self) -> None:
         """Disconnect from Slack."""
@@ -316,6 +321,8 @@ class SlackAdapter(BasePlatformAdapter):
         chat_id: str,
         message_id: str,
         content: str,
+        *,
+        finalize: bool = False,
     ) -> SendResult:
         """Edit a previously sent Slack message."""
         if not self._app:
@@ -366,6 +373,20 @@ class SlackAdapter(BasePlatformAdapter):
             # in an assistant-enabled context. Falls back to reactions.
             logger.debug("[Slack] assistant.threads.setStatus failed: %s", e)
 
+    def _dm_top_level_threads_as_sessions(self) -> bool:
+        """Whether top-level Slack DMs get per-message session threads.
+
+        Defaults to ``True`` so each visible DM reply thread is isolated as its
+        own Hermes session — matching the per-thread behavior channels already
+        have.  Set ``platforms.slack.extra.dm_top_level_threads_as_sessions``
+        to ``false`` in config.yaml to revert to the legacy behavior where all
+        top-level DMs share one continuous session.
+        """
+        raw = self.config.extra.get("dm_top_level_threads_as_sessions")
+        if raw is None:
+            return True  # default: each DM thread is its own session
+        return str(raw).strip().lower() in ("1", "true", "yes", "on")
+
     def _resolve_thread_ts(
         self,
         reply_to: Optional[str] = None,
@@ -996,10 +1017,14 @@ class SlackAdapter(BasePlatformAdapter):
         # Build thread_ts for session keying.
         # In channels: fall back to ts so each top-level @mention starts a
         #   new thread/session (the bot always replies in a thread).
-        # In DMs: only use the real thread_ts — top-level DMs should share
-        #   one continuous session, threaded DMs get their own session.
+        # In DMs: fall back to ts so each top-level DM reply thread gets
+        #   its own session key (matching channel behavior). Set
+        #   dm_top_level_threads_as_sessions: false in config to revert to
+        #   legacy single-session-per-DM-channel behavior.
         if is_dm:
-            thread_ts = event.get("thread_ts") or assistant_meta.get("thread_ts")  # None for top-level DMs
+            thread_ts = event.get("thread_ts") or assistant_meta.get("thread_ts")
+            if not thread_ts and self._dm_top_level_threads_as_sessions():
+                thread_ts = ts
         else:
             thread_ts = event.get("thread_ts") or ts  # ts fallback for channels
 
@@ -1167,6 +1192,12 @@ class SlackAdapter(BasePlatformAdapter):
             thread_id=thread_ts,
         )
 
+        # Per-channel ephemeral prompt
+        from gateway.platforms.base import resolve_channel_prompt
+        _channel_prompt = resolve_channel_prompt(
+            self.config.extra, channel_id, None,
+        )
+
         msg_event = MessageEvent(
             text=text,
             message_type=msg_type,
@@ -1176,6 +1207,7 @@ class SlackAdapter(BasePlatformAdapter):
             media_urls=media_urls,
             media_types=media_types,
             reply_to_message_id=thread_ts if thread_ts != ts else None,
+            channel_prompt=_channel_prompt,
         )
 
         # Only react when bot is directly addressed (DM or @mention).

gateway/platforms/telegram.py

@@ -11,6 +11,8 @@ import asyncio
 import json
 import logging
 import os
+import tempfile
+import html as _html
 import re
 from typing import Dict, List, Optional, Any
 
@@ -18,6 +20,10 @@ logger = logging.getLogger(__name__)
 
 try:
     from telegram import Update, Bot, Message, InlineKeyboardButton, InlineKeyboardMarkup
+    try:
+        from telegram import LinkPreviewOptions
+    except ImportError:
+        LinkPreviewOptions = None
     from telegram.ext import (
         Application,
         CommandHandler,
@@ -36,6 +42,7 @@ except ImportError:
     Message = Any
     InlineKeyboardButton = Any
     InlineKeyboardMarkup = Any
+    LinkPreviewOptions = None
     Application = Any
     CommandHandler = Any
     CallbackQueryHandler = Any
@@ -112,6 +119,84 @@ def _strip_mdv2(text: str) -> str:
     return cleaned
 
 
+# ---------------------------------------------------------------------------
+# Markdown table → code block conversion
+# ---------------------------------------------------------------------------
+# Telegram's MarkdownV2 has no table syntax — '|' is just an escaped literal,
+# so pipe tables render as noisy backslash-pipe text with no alignment.
+# Wrapping the table in a fenced code block makes Telegram render it as
+# monospace preformatted text with columns intact.
+
+# Matches a GFM table delimiter row: optional outer pipes, cells containing
+# only dashes (with optional leading/trailing colons for alignment) separated
+# by '|'.  Requires at least one internal '|' so lone '---' horizontal rules
+# are NOT matched.
+_TABLE_SEPARATOR_RE = re.compile(
+    r'^\s*\|?\s*:?-+:?\s*(?:\|\s*:?-+:?\s*){1,}\|?\s*$'
+)
+
+
+def _is_table_row(line: str) -> bool:
+    """Return True if *line* could plausibly be a table data row."""
+    stripped = line.strip()
+    return bool(stripped) and '|' in stripped
+
+
+def _wrap_markdown_tables(text: str) -> str:
+    """Wrap GFM-style pipe tables in ``` fences so Telegram renders them.
+
+    Detected by a row containing '|' immediately followed by a delimiter
+    row matching :data:`_TABLE_SEPARATOR_RE`.  Subsequent pipe-containing
+    non-blank lines are consumed as the table body and included in the
+    wrapped block.  Tables inside existing fenced code blocks are left
+    alone.
+    """
+    if '|' not in text or '-' not in text:
+        return text
+
+    lines = text.split('\n')
+    out: list[str] = []
+    in_fence = False
+    i = 0
+    while i < len(lines):
+        line = lines[i]
+        stripped = line.lstrip()
+
+        # Track existing fenced code blocks — never touch content inside.
+        if stripped.startswith('```'):
+            in_fence = not in_fence
+            out.append(line)
+            i += 1
+            continue
+        if in_fence:
+            out.append(line)
+            i += 1
+            continue
+
+        # Look for a header row (contains '|') immediately followed by a
+        # delimiter row.
+        if (
+            '|' in line
+            and i + 1 < len(lines)
+            and _TABLE_SEPARATOR_RE.match(lines[i + 1])
+        ):
+            table_block = [line, lines[i + 1]]
+            j = i + 2
+            while j < len(lines) and _is_table_row(lines[j]):
+                table_block.append(lines[j])
+                j += 1
+            out.append('```')
+            out.extend(table_block)
+            out.append('```')
+            i = j
+            continue
+
+        out.append(line)
+        i += 1
+
+    return '\n'.join(out)
+
+
 class TelegramAdapter(BasePlatformAdapter):
     """
     Telegram bot adapter.
@@ -129,6 +214,7 @@ class TelegramAdapter(BasePlatformAdapter):
     # When a chunk is near this limit, a continuation is almost certain.
     _SPLIT_THRESHOLD = 4000
     MEDIA_GROUP_WAIT_SECONDS = 0.8
+    _GENERAL_TOPIC_THREAD_ID = "1"
     
     def __init__(self, config: PlatformConfig):
         super().__init__(config, Platform.TELEGRAM)
@@ -137,6 +223,7 @@ class TelegramAdapter(BasePlatformAdapter):
         self._webhook_mode: bool = False
         self._mention_patterns = self._compile_mention_patterns()
         self._reply_to_mode: str = getattr(config, 'reply_to_mode', 'first') or 'first'
+        self._disable_link_previews: bool = self._coerce_bool_extra("disable_link_previews", False)
         # Buffer rapid/album photo updates so Telegram image bursts are handled
         # as a single MessageEvent instead of self-interrupting multiple turns.
         self._media_batch_delay_seconds = float(os.getenv("HERMES_TELEGRAM_MEDIA_BATCH_DELAY_SECONDS", "0.8"))
@@ -163,6 +250,38 @@ class TelegramAdapter(BasePlatformAdapter):
         # Approval button state: message_id → session_key
         self._approval_state: Dict[int, str] = {}
 
+    @staticmethod
+    def _is_callback_user_authorized(user_id: str) -> bool:
+        """Return whether a Telegram inline-button caller may perform gated actions."""
+        allowed_csv = os.getenv("TELEGRAM_ALLOWED_USERS", "").strip()
+        if not allowed_csv:
+            return True
+        allowed_ids = {uid.strip() for uid in allowed_csv.split(",") if uid.strip()}
+        return "*" in allowed_ids or user_id in allowed_ids
+
+    @classmethod
+    def _metadata_thread_id(cls, metadata: Optional[Dict[str, Any]]) -> Optional[str]:
+        if not metadata:
+            return None
+        thread_id = metadata.get("thread_id") or metadata.get("message_thread_id")
+        return str(thread_id) if thread_id is not None else None
+
+    @classmethod
+    def _message_thread_id_for_send(cls, thread_id: Optional[str]) -> Optional[int]:
+        if not thread_id or str(thread_id) == cls._GENERAL_TOPIC_THREAD_ID:
+            return None
+        return int(thread_id)
+
+    @classmethod
+    def _message_thread_id_for_typing(cls, thread_id: Optional[str]) -> Optional[int]:
+        if not thread_id:
+            return None
+        return int(thread_id)
+
+    @staticmethod
+    def _is_thread_not_found_error(error: Exception) -> bool:
+        return "thread not found" in str(error).lower()
+
     def _fallback_ips(self) -> list[str]:
         """Return validated fallback IPs from config (populated by _apply_env_overrides)."""
         configured = self.config.extra.get("fallback_ips", []) if getattr(self.config, "extra", None) else []
@@ -193,6 +312,26 @@ class TelegramAdapter(BasePlatformAdapter):
             pass
         return isinstance(error, OSError)
 
+    def _coerce_bool_extra(self, key: str, default: bool = False) -> bool:
+        value = self.config.extra.get(key) if getattr(self.config, "extra", None) else None
+        if value is None:
+            return default
+        if isinstance(value, str):
+            lowered = value.strip().lower()
+            if lowered in ("true", "1", "yes", "on"):
+                return True
+            if lowered in ("false", "0", "no", "off"):
+                return False
+            return default
+        return bool(value)
+
+    def _link_preview_kwargs(self) -> Dict[str, Any]:
+        if not getattr(self, "_disable_link_previews", False):
+            return {}
+        if LinkPreviewOptions is not None:
+            return {"link_preview_options": LinkPreviewOptions(is_disabled=True)}
+        return {"disable_web_page_preview": True}
+
     async def _handle_polling_network_error(self, error: Exception) -> None:
         """Reconnect polling after a transient network interruption.
 
@@ -396,8 +535,23 @@ class TelegramAdapter(BasePlatformAdapter):
                         break
 
             if changed:
-                with open(config_path, "w") as f:
-                    _yaml.dump(config, f, default_flow_style=False, sort_keys=False)
+                fd, tmp_path = tempfile.mkstemp(
+                    dir=str(config_path.parent),
+                    suffix=".tmp",
+                    prefix=".config_",
+                )
+                try:
+                    with os.fdopen(fd, "w", encoding="utf-8") as f:
+                        _yaml.dump(config, f, default_flow_style=False, sort_keys=False)
+                        f.flush()
+                        os.fsync(f.fileno())
+                    os.replace(tmp_path, config_path)
+                except BaseException:
+                    try:
+                        os.unlink(tmp_path)
+                    except OSError:
+                        pass
+                    raise
                 logger.info(
                     "[%s] Persisted thread_id=%s for topic '%s' in config.yaml",
                     self.name, thread_id, topic_name,
@@ -540,7 +694,7 @@ class TelegramAdapter(BasePlatformAdapter):
                 "write_timeout": _env_float("HERMES_TELEGRAM_HTTP_WRITE_TIMEOUT", 20.0),
             }
 
-            proxy_url = resolve_proxy_url()
+            proxy_url = resolve_proxy_url("TELEGRAM_PROXY")
             disable_fallback = (os.getenv("HERMES_TELEGRAM_DISABLE_FALLBACK_IPS", "").strip().lower() in ("1", "true", "yes", "on"))
             fallback_ips = self._fallback_ips()
             if not fallback_ips:
@@ -606,14 +760,14 @@ class TelegramAdapter(BasePlatformAdapter):
                 from telegram.error import NetworkError, TimedOut
             except ImportError:
                 NetworkError = TimedOut = OSError  # type: ignore[misc,assignment]
-            _max_connect = 3
+            _max_connect = 8
             for _attempt in range(_max_connect):
                 try:
                     await self._app.initialize()
                     break
                 except (NetworkError, TimedOut, OSError) as init_err:
                     if _attempt < _max_connect - 1:
-                        wait = 2 ** _attempt
+                        wait = min(2 ** _attempt, 15)
                         logger.warning(
                             "[%s] Connect attempt %d/%d failed: %s — retrying in %ds",
                             self.name, _attempt + 1, _max_connect, init_err, wait,
@@ -814,7 +968,7 @@ class TelegramAdapter(BasePlatformAdapter):
                 ]
             
             message_ids = []
-            thread_id = metadata.get("thread_id") if metadata else None
+            thread_id = self._metadata_thread_id(metadata)
             
             try:
                 from telegram.error import NetworkError as _NetErr
@@ -834,7 +988,7 @@ class TelegramAdapter(BasePlatformAdapter):
             for i, chunk in enumerate(chunks):
                 should_thread = self._should_thread_reply(reply_to, i)
                 reply_to_id = int(reply_to) if should_thread else None
-                effective_thread_id = int(thread_id) if thread_id else None
+                effective_thread_id = self._message_thread_id_for_send(thread_id)
 
                 msg = None
                 for _send_attempt in range(3):
@@ -847,6 +1001,7 @@ class TelegramAdapter(BasePlatformAdapter):
                                 parse_mode=ParseMode.MARKDOWN_V2,
                                 reply_to_message_id=reply_to_id,
                                 message_thread_id=effective_thread_id,
+                                **self._link_preview_kwargs(),
                             )
                         except Exception as md_error:
                             # Markdown parsing failed, try plain text
@@ -859,6 +1014,7 @@ class TelegramAdapter(BasePlatformAdapter):
                                     parse_mode=None,
                                     reply_to_message_id=reply_to_id,
                                     message_thread_id=effective_thread_id,
+                                    **self._link_preview_kwargs(),
                                 )
                             else:
                                 raise
@@ -869,8 +1025,7 @@ class TelegramAdapter(BasePlatformAdapter):
                         # (not transient network issues). Detect and handle
                         # specific cases instead of blindly retrying.
                         if _BadReq and isinstance(send_err, _BadReq):
-                            err_lower = str(send_err).lower()
-                            if "thread not found" in err_lower and effective_thread_id is not None:
+                            if self._is_thread_not_found_error(send_err) and effective_thread_id is not None:
                                 # Thread doesn't exist — retry without
                                 # message_thread_id so the message still
                                 # reaches the chat.
@@ -880,6 +1035,7 @@ class TelegramAdapter(BasePlatformAdapter):
                                 )
                                 effective_thread_id = None
                                 continue
+                            err_lower = str(send_err).lower()
                             if "message to be replied not found" in err_lower and reply_to_id is not None:
                                 # Original message was deleted before we
                                 # could reply — clear reply target and retry
@@ -941,6 +1097,8 @@ class TelegramAdapter(BasePlatformAdapter):
         chat_id: str,
         message_id: str,
         content: str,
+        *,
+        finalize: bool = False,
     ) -> SendResult:
         """Edit a previously sent Telegram message."""
         if not self._bot:
@@ -1046,6 +1204,7 @@ class TelegramAdapter(BasePlatformAdapter):
                 text=text,
                 parse_mode=ParseMode.MARKDOWN,
                 reply_markup=keyboard,
+                **self._link_preview_kwargs(),
             )
             return SendResult(success=True, message_id=str(msg.message_id))
         except Exception as e:
@@ -1068,15 +1227,13 @@ class TelegramAdapter(BasePlatformAdapter):
         try:
             cmd_preview = command[:3800] + "..." if len(command) > 3800 else command
             text = (
-                f"⚠️ *Command Approval Required*\n\n"
-                f"`{cmd_preview}`\n\n"
-                f"Reason: {description}"
+                f"⚠️ <b>Command Approval Required</b>\n\n"
+                f"<pre>{_html.escape(cmd_preview)}</pre>\n\n"
+                f"Reason: {_html.escape(description)}"
             )
 
             # Resolve thread context for thread replies
-            thread_id = None
-            if metadata:
-                thread_id = metadata.get("thread_id") or metadata.get("message_thread_id")
+            thread_id = self._metadata_thread_id(metadata)
 
             # We'll use the message_id as part of callback_data to look up session_key
             # Send a placeholder first, then update — or use a counter.
@@ -1100,11 +1257,13 @@ class TelegramAdapter(BasePlatformAdapter):
             kwargs: Dict[str, Any] = {
                 "chat_id": int(chat_id),
                 "text": text,
-                "parse_mode": ParseMode.MARKDOWN,
+                "parse_mode": ParseMode.HTML,
                 "reply_markup": keyboard,
+                **self._link_preview_kwargs(),
             }
-            if thread_id:
-                kwargs["message_thread_id"] = int(thread_id)
+            message_thread_id = self._message_thread_id_for_send(thread_id)
+            if message_thread_id is not None:
+                kwargs["message_thread_id"] = message_thread_id
 
             msg = await self._bot.send_message(**kwargs)
 
@@ -1172,6 +1331,7 @@ class TelegramAdapter(BasePlatformAdapter):
                 parse_mode=ParseMode.MARKDOWN,
                 reply_markup=keyboard,
                 message_thread_id=int(thread_id) if thread_id else None,
+                **self._link_preview_kwargs(),
             )
 
             # Store picker state keyed by chat_id
@@ -1440,12 +1600,9 @@ class TelegramAdapter(BasePlatformAdapter):
 
                 # Only authorized users may click approval buttons.
                 caller_id = str(getattr(query.from_user, "id", ""))
-                allowed_csv = os.getenv("TELEGRAM_ALLOWED_USERS", "").strip()
-                if allowed_csv:
-                    allowed_ids = {uid.strip() for uid in allowed_csv.split(",") if uid.strip()}
-                    if "*" not in allowed_ids and caller_id not in allowed_ids:
-                        await query.answer(text="⛔ You are not authorized to approve commands.")
-                        return
+                if not self._is_callback_user_authorized(caller_id):
+                    await query.answer(text="⛔ You are not authorized to approve commands.")
+                    return
 
                 session_key = self._approval_state.pop(approval_id, None)
                 if not session_key:
@@ -1490,6 +1647,10 @@ class TelegramAdapter(BasePlatformAdapter):
         if not data.startswith("update_prompt:"):
             return
         answer = data.split(":", 1)[1]  # "y" or "n"
+        caller_id = str(getattr(query.from_user, "id", ""))
+        if not self._is_callback_user_authorized(caller_id):
+            await query.answer(text="⛔ You are not authorized to answer update prompts.")
+            return
         await query.answer(text=f"Sent '{answer}' to the update process.")
         # Edit the message to show the choice and remove buttons
         label = "Yes" if answer == "y" else "No"
@@ -1514,6 +1675,21 @@ class TelegramAdapter(BasePlatformAdapter):
         except Exception as exc:
             logger.error("Failed to write update response from callback: %s", exc)
 
+    def _missing_media_path_error(self, label: str, path: str) -> str:
+        """Build an actionable file-not-found error for gateway MEDIA delivery.
+
+        Paths like /workspace/... or /output/... often only exist inside the
+        Docker sandbox, while the gateway process runs on the host.
+        """
+        error = f"{label} file not found: {path}"
+        if path.startswith(("/workspace/", "/output/", "/outputs/")):
+            error += (
+                " (path may only exist inside the Docker sandbox. "
+                "Bind-mount a host directory and emit the host-visible "
+                "path in MEDIA: for gateway file delivery.)"
+            )
+        return error
+
     async def send_voice(
         self,
         chat_id: str,
@@ -1530,28 +1706,28 @@ class TelegramAdapter(BasePlatformAdapter):
         try:
             import os
             if not os.path.exists(audio_path):
-                return SendResult(success=False, error=f"Audio file not found: {audio_path}")
+                return SendResult(success=False, error=self._missing_media_path_error("Audio", audio_path))
             
             with open(audio_path, "rb") as audio_file:
                 # .ogg files -> send as voice (round playable bubble)
                 if audio_path.endswith((".ogg", ".opus")):
-                    _voice_thread = metadata.get("thread_id") if metadata else None
+                    _voice_thread = self._metadata_thread_id(metadata)
                     msg = await self._bot.send_voice(
                         chat_id=int(chat_id),
                         voice=audio_file,
                         caption=caption[:1024] if caption else None,
                         reply_to_message_id=int(reply_to) if reply_to else None,
-                        message_thread_id=int(_voice_thread) if _voice_thread else None,
+                        message_thread_id=self._message_thread_id_for_send(_voice_thread),
                     )
                 else:
                     # .mp3 and others -> send as audio file
-                    _audio_thread = metadata.get("thread_id") if metadata else None
+                    _audio_thread = self._metadata_thread_id(metadata)
                     msg = await self._bot.send_audio(
                         chat_id=int(chat_id),
                         audio=audio_file,
                         caption=caption[:1024] if caption else None,
                         reply_to_message_id=int(reply_to) if reply_to else None,
-                        message_thread_id=int(_audio_thread) if _audio_thread else None,
+                        message_thread_id=self._message_thread_id_for_send(_audio_thread),
                     )
             return SendResult(success=True, message_id=str(msg.message_id))
         except Exception as e:
@@ -1579,16 +1755,16 @@ class TelegramAdapter(BasePlatformAdapter):
         try:
             import os
             if not os.path.exists(image_path):
-                return SendResult(success=False, error=f"Image file not found: {image_path}")
+                return SendResult(success=False, error=self._missing_media_path_error("Image", image_path))
 
-            _thread = metadata.get("thread_id") if metadata else None
+            _thread = self._metadata_thread_id(metadata)
             with open(image_path, "rb") as image_file:
                 msg = await self._bot.send_photo(
                     chat_id=int(chat_id),
                     photo=image_file,
                     caption=caption[:1024] if caption else None,
                     reply_to_message_id=int(reply_to) if reply_to else None,
-                    message_thread_id=int(_thread) if _thread else None,
+                    message_thread_id=self._message_thread_id_for_send(_thread),
                 )
             return SendResult(success=True, message_id=str(msg.message_id))
         except Exception as e:
@@ -1616,10 +1792,10 @@ class TelegramAdapter(BasePlatformAdapter):
 
         try:
             if not os.path.exists(file_path):
-                return SendResult(success=False, error=f"File not found: {file_path}")
+                return SendResult(success=False, error=self._missing_media_path_error("File", file_path))
 
             display_name = file_name or os.path.basename(file_path)
-            _thread = metadata.get("thread_id") if metadata else None
+            _thread = self._metadata_thread_id(metadata)
 
             with open(file_path, "rb") as f:
                 msg = await self._bot.send_document(
@@ -1628,7 +1804,7 @@ class TelegramAdapter(BasePlatformAdapter):
                     filename=display_name,
                     caption=caption[:1024] if caption else None,
                     reply_to_message_id=int(reply_to) if reply_to else None,
-                    message_thread_id=int(_thread) if _thread else None,
+                    message_thread_id=self._message_thread_id_for_send(_thread),
                 )
             return SendResult(success=True, message_id=str(msg.message_id))
         except Exception as e:
@@ -1650,16 +1826,16 @@ class TelegramAdapter(BasePlatformAdapter):
 
         try:
             if not os.path.exists(video_path):
-                return SendResult(success=False, error=f"Video file not found: {video_path}")
+                return SendResult(success=False, error=self._missing_media_path_error("Video", video_path))
 
-            _thread = metadata.get("thread_id") if metadata else None
+            _thread = self._metadata_thread_id(metadata)
             with open(video_path, "rb") as f:
                 msg = await self._bot.send_video(
                     chat_id=int(chat_id),
                     video=f,
                     caption=caption[:1024] if caption else None,
                     reply_to_message_id=int(reply_to) if reply_to else None,
-                    message_thread_id=int(_thread) if _thread else None,
+                    message_thread_id=self._message_thread_id_for_send(_thread),
                 )
             return SendResult(success=True, message_id=str(msg.message_id))
         except Exception as e:
@@ -1689,13 +1865,13 @@ class TelegramAdapter(BasePlatformAdapter):
 
         try:
             # Telegram can send photos directly from URLs (up to ~5MB)
-            _photo_thread = metadata.get("thread_id") if metadata else None
+            _photo_thread = self._metadata_thread_id(metadata)
             msg = await self._bot.send_photo(
                 chat_id=int(chat_id),
                 photo=image_url,
                 caption=caption[:1024] if caption else None,  # Telegram caption limit
                 reply_to_message_id=int(reply_to) if reply_to else None,
-                message_thread_id=int(_photo_thread) if _photo_thread else None,
+                message_thread_id=self._message_thread_id_for_send(_photo_thread),
             )
             return SendResult(success=True, message_id=str(msg.message_id))
         except Exception as e:
@@ -1718,6 +1894,7 @@ class TelegramAdapter(BasePlatformAdapter):
                     photo=image_data,
                     caption=caption[:1024] if caption else None,
                     reply_to_message_id=int(reply_to) if reply_to else None,
+                    message_thread_id=self._message_thread_id_for_send(_photo_thread),
                 )
                 return SendResult(success=True, message_id=str(msg.message_id))
             except Exception as e2:
@@ -1743,13 +1920,13 @@ class TelegramAdapter(BasePlatformAdapter):
             return SendResult(success=False, error="Not connected")
         
         try:
-            _anim_thread = metadata.get("thread_id") if metadata else None
+            _anim_thread = self._metadata_thread_id(metadata)
             msg = await self._bot.send_animation(
                 chat_id=int(chat_id),
                 animation=animation_url,
                 caption=caption[:1024] if caption else None,
                 reply_to_message_id=int(reply_to) if reply_to else None,
-                message_thread_id=int(_anim_thread) if _anim_thread else None,
+                message_thread_id=self._message_thread_id_for_send(_anim_thread),
             )
             return SendResult(success=True, message_id=str(msg.message_id))
         except Exception as e:
@@ -1766,12 +1943,23 @@ class TelegramAdapter(BasePlatformAdapter):
         """Send typing indicator."""
         if self._bot:
             try:
-                _typing_thread = metadata.get("thread_id") if metadata else None
-                await self._bot.send_chat_action(
-                    chat_id=int(chat_id),
-                    action="typing",
-                    message_thread_id=int(_typing_thread) if _typing_thread else None,
-                )
+                _typing_thread = self._metadata_thread_id(metadata)
+                message_thread_id = self._message_thread_id_for_typing(_typing_thread)
+                try:
+                    await self._bot.send_chat_action(
+                        chat_id=int(chat_id),
+                        action="typing",
+                        message_thread_id=message_thread_id,
+                    )
+                except Exception as e:
+                    if message_thread_id is not None and self._is_thread_not_found_error(e):
+                        await self._bot.send_chat_action(
+                            chat_id=int(chat_id),
+                            action="typing",
+                            message_thread_id=None,
+                        )
+                    else:
+                        raise
             except Exception as e:
                 # Typing failures are non-fatal; log at debug level only.
                 logger.debug(
@@ -1839,6 +2027,12 @@ class TelegramAdapter(BasePlatformAdapter):
 
         text = content
 
+        # 0) Pre-wrap GFM-style pipe tables in ``` fences.  Telegram can't
+        #    render tables natively, but fenced code blocks render as
+        #    monospace preformatted text with columns intact.  The wrapped
+        #    tables then flow through step (1) below as protected regions.
+        text = _wrap_markdown_tables(text)
+
         # 1) Protect fenced code blocks (``` ... ```)
         #    Per MarkdownV2 spec, \ and ` inside pre/code must be escaped.
         def _protect_fenced(m):
@@ -2080,22 +2274,27 @@ class TelegramAdapter(BasePlatformAdapter):
 
         bot_username = (getattr(self._bot, "username", None) or "").lstrip("@").lower()
         bot_id = getattr(self._bot, "id", None)
+        expected = f"@{bot_username}" if bot_username else None
 
         def _iter_sources():
             yield getattr(message, "text", None) or "", getattr(message, "entities", None) or []
             yield getattr(message, "caption", None) or "", getattr(message, "caption_entities", None) or []
 
+        # Telegram parses mentions server-side and emits MessageEntity objects
+        # (type=mention for @username, type=text_mention for @FirstName targeting
+        # a user without a public username). Only those entities are authoritative —
+        # raw substring matches like "foo@hermes_bot.example" are not mentions
+        # (bug #12545). Entities also correctly handle @handles inside URLs, code
+        # blocks, and quoted text, where a regex scan would over-match.
         for source_text, entities in _iter_sources():
-            if bot_username and f"@{bot_username}" in source_text.lower():
-                return True
             for entity in entities:
                 entity_type = str(getattr(entity, "type", "")).split(".")[-1].lower()
-                if entity_type == "mention" and bot_username:
+                if entity_type == "mention" and expected:
                     offset = int(getattr(entity, "offset", -1))
                     length = int(getattr(entity, "length", 0))
                     if offset < 0 or length <= 0:
                         continue
-                    if source_text[offset:offset + length].strip().lower() == f"@{bot_username}":
+                    if source_text[offset:offset + length].strip().lower() == expected:
                         return True
                 elif entity_type == "text_mention":
                     user = getattr(entity, "user", None)
@@ -2165,7 +2364,7 @@ class TelegramAdapter(BasePlatformAdapter):
         if not self._should_process_message(update.message):
             return
 
-        event = self._build_message_event(update.message, MessageType.TEXT)
+        event = self._build_message_event(update.message, MessageType.TEXT, update_id=update.update_id)
         event.text = self._clean_bot_trigger_text(event.text)
         self._enqueue_text_event(event)
     
@@ -2176,7 +2375,7 @@ class TelegramAdapter(BasePlatformAdapter):
         if not self._should_process_message(update.message, is_command=True):
             return
         
-        event = self._build_message_event(update.message, MessageType.COMMAND)
+        event = self._build_message_event(update.message, MessageType.COMMAND, update_id=update.update_id)
         await self.handle_message(event)
     
     async def _handle_location_message(self, update: Update, context: ContextTypes.DEFAULT_TYPE) -> None:
@@ -2212,7 +2411,7 @@ class TelegramAdapter(BasePlatformAdapter):
         parts.append(f"Map: https://www.google.com/maps/search/?api=1&query={lat},{lon}")
         parts.append("Ask what they'd like to find nearby (restaurants, cafes, etc.) and any preferences.")
 
-        event = self._build_message_event(msg, MessageType.LOCATION)
+        event = self._build_message_event(msg, MessageType.LOCATION, update_id=update.update_id)
         event.text = "\n".join(parts)
         await self.handle_message(event)
 
@@ -2363,7 +2562,7 @@ class TelegramAdapter(BasePlatformAdapter):
         else:
             msg_type = MessageType.DOCUMENT
         
-        event = self._build_message_event(msg, msg_type)
+        event = self._build_message_event(msg, msg_type, update_id=update.update_id)
         
         # Add caption as text
         if msg.caption:
@@ -2702,8 +2901,19 @@ class TelegramAdapter(BasePlatformAdapter):
                 self.name, cache_key, thread_id,
             )
 
-    def _build_message_event(self, message: Message, msg_type: MessageType) -> MessageEvent:
-        """Build a MessageEvent from a Telegram message."""
+    def _build_message_event(
+        self,
+        message: Message,
+        msg_type: MessageType,
+        update_id: Optional[int] = None,
+    ) -> MessageEvent:
+        """Build a MessageEvent from a Telegram message.
+
+        ``update_id`` is the ``Update.update_id`` from PTB; passing it through
+        lets ``/restart`` record the triggering offset so the new gateway
+        process can advance past it (prevents ``/restart`` being re-delivered
+        when PTB's graceful-shutdown ACK fails).
+        """
         chat = message.chat
         user = message.from_user
         
@@ -2716,7 +2926,9 @@ class TelegramAdapter(BasePlatformAdapter):
 
         # Resolve DM topic name and skill binding
         thread_id_raw = message.message_thread_id
-        thread_id_str = str(thread_id_raw) if thread_id_raw else None
+        thread_id_str = str(thread_id_raw) if thread_id_raw is not None else None
+        if chat_type == "group" and thread_id_str is None and getattr(chat, "is_forum", False):
+            thread_id_str = self._GENERAL_TOPIC_THREAD_ID
         chat_topic = None
         topic_skill = None
 
@@ -2752,8 +2964,8 @@ class TelegramAdapter(BasePlatformAdapter):
             chat_id=str(chat.id),
             chat_name=chat.title or (chat.full_name if hasattr(chat, "full_name") else None),
             chat_type=chat_type,
-            user_id=str(user.id) if user else None,
-            user_name=user.full_name if user else None,
+            user_id=str(user.id) if user else (str(chat.id) if chat_type == "dm" else None),
+            user_name=user.full_name if user else (chat.full_name if hasattr(chat, "full_name") and chat_type == "dm" else None),
             thread_id=thread_id_str,
             chat_topic=chat_topic,
         )
@@ -2765,15 +2977,26 @@ class TelegramAdapter(BasePlatformAdapter):
             reply_to_id = str(message.reply_to_message.message_id)
             reply_to_text = message.reply_to_message.text or message.reply_to_message.caption or None
 
+        # Per-channel/topic ephemeral prompt
+        from gateway.platforms.base import resolve_channel_prompt
+        _chat_id_str = str(chat.id)
+        _channel_prompt = resolve_channel_prompt(
+            self.config.extra,
+            thread_id_str or _chat_id_str,
+            _chat_id_str if thread_id_str else None,
+        )
+
         return MessageEvent(
             text=message.text or "",
             message_type=msg_type,
             source=source,
             raw_message=message,
             message_id=str(message.message_id),
+            platform_update_id=update_id,
             reply_to_message_id=reply_to_id,
             reply_to_text=reply_to_text,
             auto_skill=topic_skill,
+            channel_prompt=_channel_prompt,
             timestamp=message.date,
         )
 

gateway/platforms/telegram_network.py

@@ -46,7 +46,7 @@ _SEED_FALLBACK_IPS: list[str] = ["149.154.167.220"]
 def _resolve_proxy_url() -> str | None:
     # Delegate to shared implementation (env vars + macOS system proxy detection)
     from gateway.platforms.base import resolve_proxy_url
-    return resolve_proxy_url()
+    return resolve_proxy_url("TELEGRAM_PROXY")
 
 
 class TelegramFallbackTransport(httpx.AsyncBaseTransport):

gateway/platforms/webhook.py

@@ -13,6 +13,10 @@ Each route defines:
   - skills: optional list of skills to load for the agent
   - deliver: where to send the response (github_comment, telegram, etc.)
   - deliver_extra: additional delivery config (repo, pr_number, chat_id)
+  - deliver_only: if true, skip the agent — the rendered prompt IS the
+    message that gets delivered.  Use for external push notifications
+    (Supabase, monitoring alerts, inter-agent pings) where zero LLM cost
+    and sub-second delivery matter more than agent reasoning.
 
 Security:
   - HMAC secret is required per route (validated at startup)
@@ -122,6 +126,19 @@ class WebhookAdapter(BasePlatformAdapter):
                     f"For testing without auth, set secret to '{_INSECURE_NO_AUTH}'."
                 )
 
+            # deliver_only routes bypass the agent — the POST body becomes a
+            # direct push notification via the configured delivery target.
+            # Validate up-front so misconfiguration surfaces at startup rather
+            # than on the first webhook POST.
+            if route.get("deliver_only"):
+                deliver = route.get("deliver", "log")
+                if not deliver or deliver == "log":
+                    raise ValueError(
+                        f"[webhook] Route '{name}' has deliver_only=true but "
+                        f"deliver is '{deliver}'. Direct delivery requires a "
+                        f"real target (telegram, discord, slack, github_comment, etc.)."
+                    )
+
         app = web.Application()
         app.router.add_get("/health", self._handle_health)
         app.router.add_post("/webhooks/{route_name}", self._handle_webhook)
@@ -296,24 +313,14 @@ class WebhookAdapter(BasePlatformAdapter):
                 {"error": "Payload too large"}, status=413
             )
 
-        # ── Rate limiting ────────────────────────────────────────
-        now = time.time()
-        window = self._rate_counts.setdefault(route_name, [])
-        window[:] = [t for t in window if now - t < 60]
-        if len(window) >= self._rate_limit:
-            return web.json_response(
-                {"error": "Rate limit exceeded"}, status=429
-            )
-        window.append(now)
-
-        # Read body
+        # Read body (must be done before any validation)
         try:
             raw_body = await request.read()
         except Exception as e:
             logger.error("[webhook] Failed to read body: %s", e)
             return web.json_response({"error": "Bad request"}, status=400)
 
-        # Validate HMAC signature (skip for INSECURE_NO_AUTH testing mode)
+        # Validate HMAC signature FIRST (skip for INSECURE_NO_AUTH testing mode)
         secret = route_config.get("secret", self._global_secret)
         if secret and secret != _INSECURE_NO_AUTH:
             if not self._validate_signature(request, raw_body, secret):
@@ -324,6 +331,16 @@ class WebhookAdapter(BasePlatformAdapter):
                     {"error": "Invalid signature"}, status=401
                 )
 
+        # ── Rate limiting (after auth) ───────────────────────────
+        now = time.time()
+        window = self._rate_counts.setdefault(route_name, [])
+        window[:] = [t for t in window if now - t < 60]
+        if len(window) >= self._rate_limit:
+            return web.json_response(
+                {"error": "Rate limit exceeded"}, status=429
+            )
+        window.append(now)
+
         # Parse payload
         try:
             payload = json.loads(raw_body)
@@ -419,6 +436,64 @@ class WebhookAdapter(BasePlatformAdapter):
             )
         self._seen_deliveries[delivery_id] = now
 
+        # ── Direct delivery mode (deliver_only) ─────────────────
+        # Skip the agent entirely — the rendered prompt IS the message we
+        # deliver.  Use case: external services (Supabase, monitoring,
+        # cron jobs, other agents) that need to push a plain notification
+        # to a user's chat with zero LLM cost.  Reuses the same HMAC auth,
+        # rate limiting, idempotency, and template rendering as agent mode.
+        if route_config.get("deliver_only"):
+            delivery = {
+                "deliver": route_config.get("deliver", "log"),
+                "deliver_extra": self._render_delivery_extra(
+                    route_config.get("deliver_extra", {}), payload
+                ),
+                "payload": payload,
+            }
+            logger.info(
+                "[webhook] direct-deliver event=%s route=%s target=%s msg_len=%d delivery=%s",
+                event_type,
+                route_name,
+                delivery["deliver"],
+                len(prompt),
+                delivery_id,
+            )
+            try:
+                result = await self._direct_deliver(prompt, delivery)
+            except Exception:
+                logger.exception(
+                    "[webhook] direct-deliver failed route=%s delivery=%s",
+                    route_name,
+                    delivery_id,
+                )
+                return web.json_response(
+                    {"status": "error", "error": "Delivery failed", "delivery_id": delivery_id},
+                    status=502,
+                )
+
+            if result.success:
+                return web.json_response(
+                    {
+                        "status": "delivered",
+                        "route": route_name,
+                        "target": delivery["deliver"],
+                        "delivery_id": delivery_id,
+                    },
+                    status=200,
+                )
+            # Delivery attempted but target rejected it — surface as 502
+            # with a generic error (don't leak adapter-level detail).
+            logger.warning(
+                "[webhook] direct-deliver target rejected route=%s target=%s error=%s",
+                route_name,
+                delivery["deliver"],
+                result.error,
+            )
+            return web.json_response(
+                {"status": "error", "error": "Delivery failed", "delivery_id": delivery_id},
+                status=502,
+            )
+
         # Use delivery_id in session key so concurrent webhooks on the
         # same route get independent agent runs (not queued/interrupted).
         session_chat_id = f"webhook:{route_name}:{delivery_id}"
@@ -572,6 +647,34 @@ class WebhookAdapter(BasePlatformAdapter):
     # Response delivery
     # ------------------------------------------------------------------
 
+    async def _direct_deliver(
+        self, content: str, delivery: dict
+    ) -> SendResult:
+        """Deliver *content* directly without invoking the agent.
+
+        Used by ``deliver_only`` routes: the rendered template becomes the
+        literal message body, and we dispatch to the same delivery helpers
+        that the agent-mode ``send()`` flow uses.  All target types that
+        work in agent mode work here — Telegram, Discord, Slack, GitHub
+        PR comments, etc.
+        """
+        deliver_type = delivery.get("deliver", "log")
+
+        if deliver_type == "log":
+            # Shouldn't reach here — startup validation rejects deliver_only
+            # with deliver=log — but guard defensively.
+            logger.info("[webhook] direct-deliver log-only: %s", content[:200])
+            return SendResult(success=True)
+
+        if deliver_type == "github_comment":
+            return await self._deliver_github_comment(content, delivery)
+
+        # Fall through to the cross-platform dispatcher, which validates the
+        # target name and routes via the gateway runner.
+        return await self._deliver_cross_platform(
+            deliver_type, content, delivery
+        )
+
     async def _deliver_github_comment(
         self, content: str, delivery: dict
     ) -> SendResult:

gateway/platforms/wecom.py

@@ -180,6 +180,8 @@ class WeComAdapter(BasePlatformAdapter):
         self._text_batch_split_delay_seconds = float(os.getenv("HERMES_WECOM_TEXT_BATCH_SPLIT_DELAY_SECONDS", "2.0"))
         self._pending_text_batches: Dict[str, MessageEvent] = {}
         self._pending_text_batch_tasks: Dict[str, asyncio.Task] = {}
+        self._device_id = uuid.uuid4().hex
+        self._last_chat_req_ids: Dict[str, str] = {}
 
     # ------------------------------------------------------------------
     # Connection lifecycle
@@ -277,7 +279,11 @@ class WeComAdapter(BasePlatformAdapter):
             {
                 "cmd": APP_CMD_SUBSCRIBE,
                 "headers": {"req_id": req_id},
-                "body": {"bot_id": self._bot_id, "secret": self._secret},
+                "body": {
+                    "bot_id": self._bot_id,
+                    "secret": self._secret,
+                    "device_id": self._device_id,
+                },
             }
         )
 
@@ -496,6 +502,11 @@ class WeComAdapter(BasePlatformAdapter):
             logger.debug("[%s] DM sender %s blocked by policy", self.name, sender_id)
             return
 
+        # Cache the inbound req_id after policy checks so proactive sends to
+        # this chat can fall back to APP_CMD_RESPONSE (required for groups —
+        # WeCom AI Bots cannot initiate APP_CMD_SEND in group chats).
+        self._remember_chat_req_id(chat_id, self._payload_req_id(payload))
+
         text, reply_text = self._extract_text(body)
         media_urls, media_types = await self._extract_media(body)
         message_type = self._derive_message_type(body, text, media_types)
@@ -847,6 +858,23 @@ class WeComAdapter(BasePlatformAdapter):
         while len(self._reply_req_ids) > DEDUP_MAX_SIZE:
             self._reply_req_ids.pop(next(iter(self._reply_req_ids)))
 
+    def _remember_chat_req_id(self, chat_id: str, req_id: str) -> None:
+        """Cache the most recent inbound req_id per chat.
+
+        Used as a fallback reply target when we need to send into a group
+        without an explicit ``reply_to`` — WeCom AI Bots are blocked from
+        APP_CMD_SEND in groups and must use APP_CMD_RESPONSE bound to some
+        prior req_id. Bounded like _reply_req_ids so long-running gateways
+        don't leak memory across many chats.
+        """
+        normalized_chat_id = str(chat_id or "").strip()
+        normalized_req_id = str(req_id or "").strip()
+        if not normalized_chat_id or not normalized_req_id:
+            return
+        self._last_chat_req_ids[normalized_chat_id] = normalized_req_id
+        while len(self._last_chat_req_ids) > DEDUP_MAX_SIZE:
+            self._last_chat_req_ids.pop(next(iter(self._last_chat_req_ids)))
+
     def _reply_req_id_for_message(self, reply_to: Optional[str]) -> Optional[str]:
         normalized = str(reply_to or "").strip()
         if not normalized or normalized.startswith("quote:"):
@@ -1163,19 +1191,15 @@ class WeComAdapter(BasePlatformAdapter):
         self._raise_for_wecom_error(response, "send media message")
         return response
 
-    async def _send_reply_stream(self, reply_req_id: str, content: str) -> Dict[str, Any]:
+    async def _send_reply_markdown(self, reply_req_id: str, content: str) -> Dict[str, Any]:
         response = await self._send_reply_request(
             reply_req_id,
             {
-                "msgtype": "stream",
-                "stream": {
-                    "id": self._new_req_id("stream"),
-                    "finish": True,
-                    "content": content[:self.MAX_MESSAGE_LENGTH],
-                },
+                "msgtype": "markdown",
+                "markdown": {"content": content[:self.MAX_MESSAGE_LENGTH]},
             },
         )
-        self._raise_for_wecom_error(response, "send reply stream")
+        self._raise_for_wecom_error(response, "send reply markdown")
         return response
 
     async def _send_reply_media_message(
@@ -1235,6 +1259,9 @@ class WeComAdapter(BasePlatformAdapter):
             return SendResult(success=False, error=prepared["reject_reason"])
 
         reply_req_id = self._reply_req_id_for_message(reply_to)
+        if not reply_req_id and chat_id in self._last_chat_req_ids:
+            reply_req_id = self._last_chat_req_ids[chat_id]
+
         try:
             upload_result = await self._upload_media_bytes(
                 prepared["data"],
@@ -1302,8 +1329,12 @@ class WeComAdapter(BasePlatformAdapter):
 
         try:
             reply_req_id = self._reply_req_id_for_message(reply_to)
+
+            if not reply_req_id and chat_id in self._last_chat_req_ids:
+                reply_req_id = self._last_chat_req_ids[chat_id]
+
             if reply_req_id:
-                response = await self._send_reply_stream(reply_req_id, content)
+                response = await self._send_reply_markdown(reply_req_id, content)
             else:
                 response = await self._send_request(
                     APP_CMD_SEND,

gateway/platforms/wecom_callback.py

@@ -258,6 +258,20 @@ class WecomCallbackAdapter(BasePlatformAdapter):
                 )
                 event = self._build_event(app, decrypted)
                 if event is not None:
+                    # Deduplicate: WeCom retries callbacks on timeout,
+                    # producing duplicate inbound messages (#10305).
+                    if event.message_id:
+                        now = time.time()
+                        if event.message_id in self._seen_messages:
+                            if now - self._seen_messages[event.message_id] < MESSAGE_DEDUP_TTL_SECONDS:
+                                logger.debug("[WecomCallback] Duplicate MsgId %s, skipping", event.message_id)
+                                return web.Response(text="success", content_type="text/plain")
+                            del self._seen_messages[event.message_id]
+                        self._seen_messages[event.message_id] = now
+                        # Prune expired entries when cache grows large
+                        if len(self._seen_messages) > 2000:
+                            cutoff = now - MESSAGE_DEDUP_TTL_SECONDS
+                            self._seen_messages = {k: v for k, v in self._seen_messages.items() if v > cutoff}
                     # Record which app this user belongs to.
                     if event.source and event.source.user_id:
                         map_key = self._user_app_key(

gateway/platforms/weixin.py

@@ -28,7 +28,7 @@ import uuid
 from datetime import datetime
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Tuple
-from urllib.parse import quote
+from urllib.parse import quote, urlparse
 
 logger = logging.getLogger(__name__)
 
@@ -96,6 +96,28 @@ MEDIA_VIDEO = 2
 MEDIA_FILE = 3
 MEDIA_VOICE = 4
 
+_LIVE_ADAPTERS: Dict[str, Any] = {}
+
+
+def _make_ssl_connector() -> Optional["aiohttp.TCPConnector"]:
+    """Return a TCPConnector with a certifi CA bundle, or None if certifi is unavailable.
+
+    Tencent's iLink server (``ilinkai.weixin.qq.com``) is not verifiable against
+    some system CA stores (notably Homebrew's OpenSSL on macOS Apple Silicon).
+    When ``certifi`` is installed, use its Mozilla CA bundle to guarantee
+    verification. Otherwise fall back to aiohttp's default (which honors
+    ``SSL_CERT_FILE`` env var via ``trust_env=True``).
+    """
+    try:
+        import ssl
+        import certifi
+    except ImportError:
+        return None
+    if not AIOHTTP_AVAILABLE:
+        return None
+    ssl_ctx = ssl.create_default_context(cafile=certifi.where())
+    return aiohttp.TCPConnector(ssl=ssl_ctx)
+
 ITEM_TEXT = 1
 ITEM_IMAGE = 2
 ITEM_VOICE = 3
@@ -398,7 +420,12 @@ async def _send_message(
     text: str,
     context_token: Optional[str],
     client_id: str,
-) -> None:
+) -> Dict[str, Any]:
+    """Send a text message via iLink sendmessage API.
+
+    Returns the raw API response dict (may contain error codes like
+    ``errcode: -14`` for session expiry that the caller can inspect).
+    """
     if not text or not text.strip():
         raise ValueError("_send_message: text must not be empty")
     message: Dict[str, Any] = {
@@ -411,7 +438,7 @@ async def _send_message(
     }
     if context_token:
         message["context_token"] = context_token
-    await _api_post(
+    return await _api_post(
         session,
         base_url=base_url,
         endpoint=EP_SEND_MESSAGE,
@@ -533,6 +560,39 @@ async def _download_bytes(
         return await response.read()
 
 
+_WEIXIN_CDN_ALLOWLIST: frozenset[str] = frozenset(
+    {
+        "novac2c.cdn.weixin.qq.com",
+        "ilinkai.weixin.qq.com",
+        "wx.qlogo.cn",
+        "thirdwx.qlogo.cn",
+        "res.wx.qq.com",
+        "mmbiz.qpic.cn",
+        "mmbiz.qlogo.cn",
+    }
+)
+
+
+def _assert_weixin_cdn_url(url: str) -> None:
+    """Raise ValueError if *url* does not point at a known WeChat CDN host."""
+    try:
+        parsed = urlparse(url)
+        scheme = parsed.scheme.lower()
+        host = parsed.hostname or ""
+    except Exception as exc:  # noqa: BLE001
+        raise ValueError(f"Unparseable media URL: {url!r}") from exc
+
+    if scheme not in ("http", "https"):
+        raise ValueError(
+            f"Media URL has disallowed scheme {scheme!r}; only http/https are permitted."
+        )
+    if host not in _WEIXIN_CDN_ALLOWLIST:
+        raise ValueError(
+            f"Media URL host {host!r} is not in the WeChat CDN allowlist. "
+            "Refusing to fetch to prevent SSRF."
+        )
+
+
 def _media_reference(item: Dict[str, Any], key: str) -> Dict[str, Any]:
     return (item.get(key) or {}).get("media") or {}
 
@@ -553,6 +613,7 @@ async def _download_and_decrypt_media(
             timeout_seconds=timeout_seconds,
         )
     elif full_url:
+        _assert_weixin_cdn_url(full_url)
         raw = await _download_bytes(session, url=full_url, timeout_seconds=timeout_seconds)
     else:
         raise RuntimeError("media item had neither encrypt_query_param nor full_url")
@@ -623,42 +684,31 @@ def _rewrite_table_block_for_weixin(lines: List[str]) -> str:
 def _normalize_markdown_blocks(content: str) -> str:
     lines = content.splitlines()
     result: List[str] = []
-    i = 0
     in_code_block = False
+    blank_run = 0
 
-    while i < len(lines):
-        line = lines[i].rstrip()
-        fence_match = _FENCE_RE.match(line.strip())
-        if fence_match:
+    for raw_line in lines:
+        line = raw_line.rstrip()
+        if _FENCE_RE.match(line.strip()):
             in_code_block = not in_code_block
             result.append(line)
-            i += 1
+            blank_run = 0
             continue
 
         if in_code_block:
             result.append(line)
-            i += 1
             continue
 
-        if (
-            i + 1 < len(lines)
-            and "|" in lines[i]
-            and _TABLE_RULE_RE.match(lines[i + 1].rstrip())
-        ):
-            table_lines = [lines[i].rstrip(), lines[i + 1].rstrip()]
-            i += 2
-            while i < len(lines) and "|" in lines[i]:
-                table_lines.append(lines[i].rstrip())
-                i += 1
-            result.append(_rewrite_table_block_for_weixin(table_lines))
+        if not line.strip():
+            blank_run += 1
+            if blank_run <= 1:
+                result.append("")
             continue
 
-        result.append(_MARKDOWN_LINK_RE.sub(r"\1 (\2)", _rewrite_headers_for_weixin(line)))
-        i += 1
+        blank_run = 0
+        result.append(line)
 
-    normalized = "\n".join(item.rstrip() for item in result)
-    normalized = re.sub(r"\n{3,}", "\n\n", normalized)
-    return normalized.strip()
+    return "\n".join(result).strip()
 
 
 def _split_markdown_blocks(content: str) -> List[str]:
@@ -704,8 +754,8 @@ def _split_delivery_units_for_weixin(content: str) -> List[str]:
 
     Weixin can render Markdown, but chat readability is better when top-level
     line breaks become separate messages. Keep fenced code blocks intact and
-    attach indented continuation lines to the previous top-level line so
-    transformed tables/lists do not get torn apart.
+    attach indented continuation lines to the previous top-level line so nested
+    list items do not get torn apart.
     """
     units: List[str] = []
 
@@ -747,7 +797,9 @@ def _looks_like_chatty_line_for_weixin(line: str) -> bool:
         return False
     if line.startswith((" ", "\t")):
         return False
-    if stripped.startswith((">", "-", "*", "【")):
+    if stripped.startswith((">", "-", "*", "【", "#", "|")):
+        return False
+    if _TABLE_RULE_RE.match(stripped):
         return False
     if re.match(r"^\*\*[^*]+\*\*$", stripped):
         return False
@@ -757,10 +809,12 @@ def _looks_like_chatty_line_for_weixin(line: str) -> bool:
 
 
 def _looks_like_heading_line_for_weixin(line: str) -> bool:
-    """Return True when a short line behaves like a plain-text heading."""
+    """Return True when a short line behaves like a heading."""
     stripped = line.strip()
     if not stripped:
         return False
+    if _HEADER_RE.match(stripped):
+        return True
     return len(stripped) <= 24 and stripped.endswith((":", "："))
 
 
@@ -935,7 +989,7 @@ async def qr_login(
     if not AIOHTTP_AVAILABLE:
         raise RuntimeError("aiohttp is required for Weixin QR login")
 
-    async with aiohttp.ClientSession(trust_env=True) as session:
+    async with aiohttp.ClientSession(trust_env=True, connector=_make_ssl_connector()) as session:
         try:
             qr_resp = await _api_get(
                 session,
@@ -953,6 +1007,10 @@ async def qr_login(
             logger.error("weixin: QR response missing qrcode")
             return None
 
+        # qrcode_url is the full scannable liteapp URL; qrcode_value is just the hex token
+        # WeChat needs to scan the full URL, not the raw hex string
+        qr_scan_data = qrcode_url if qrcode_url else qrcode_value
+
         print("\n请使用微信扫描以下二维码：")
         if qrcode_url:
             print(qrcode_url)
@@ -960,11 +1018,11 @@ async def qr_login(
             import qrcode
 
             qr = qrcode.QRCode()
-            qr.add_data(qrcode_url or qrcode_value)
+            qr.add_data(qr_scan_data)
             qr.make(fit=True)
             qr.print_ascii(invert=True)
-        except Exception:
-            print("（终端二维码渲染失败，请直接打开上面的二维码链接）")
+        except Exception as _qr_exc:
+            print(f"（终端二维码渲染失败: {_qr_exc}，请直接打开上面的二维码链接）")
 
         deadline = time.time() + timeout_seconds
         current_base_url = ILINK_BASE_URL
@@ -1010,8 +1068,17 @@ async def qr_login(
                     )
                     qrcode_value = str(qr_resp.get("qrcode") or "")
                     qrcode_url = str(qr_resp.get("qrcode_img_content") or "")
+                    qr_scan_data = qrcode_url if qrcode_url else qrcode_value
                     if qrcode_url:
                         print(qrcode_url)
+                    try:
+                        import qrcode as _qrcode
+                        qr = _qrcode.QRCode()
+                        qr.add_data(qr_scan_data)
+                        qr.make(fit=True)
+                        qr.print_ascii(invert=True)
+                    except Exception:
+                        pass
                 except Exception as exc:
                     logger.error("weixin: QR refresh failed: %s", exc)
                     return None
@@ -1059,7 +1126,8 @@ class WeixinAdapter(BasePlatformAdapter):
         self._hermes_home = hermes_home
         self._token_store = ContextTokenStore(hermes_home)
         self._typing_cache = TypingTicketCache()
-        self._session: Optional[aiohttp.ClientSession] = None
+        self._poll_session: Optional[aiohttp.ClientSession] = None
+        self._send_session: Optional[aiohttp.ClientSession] = None
         self._poll_task: Optional[asyncio.Task] = None
         self._dedup = MessageDeduplicator(ttl_seconds=MESSAGE_DEDUP_TTL_SECONDS)
 
@@ -1134,14 +1202,17 @@ class WeixinAdapter(BasePlatformAdapter):
         except Exception as exc:
             logger.debug("[%s] Token lock unavailable (non-fatal): %s", self.name, exc)
 
-        self._session = aiohttp.ClientSession(trust_env=True)
+        self._poll_session = aiohttp.ClientSession(trust_env=True, connector=_make_ssl_connector())
+        self._send_session = aiohttp.ClientSession(trust_env=True, connector=_make_ssl_connector())
         self._token_store.restore(self._account_id)
         self._poll_task = asyncio.create_task(self._poll_loop(), name="weixin-poll")
         self._mark_connected()
+        _LIVE_ADAPTERS[self._token] = self
         logger.info("[%s] Connected account=%s base=%s", self.name, _safe_id(self._account_id), self._base_url)
         return True
 
     async def disconnect(self) -> None:
+        _LIVE_ADAPTERS.pop(self._token, None)
         self._running = False
         if self._poll_task and not self._poll_task.done():
             self._poll_task.cancel()
@@ -1150,15 +1221,18 @@ class WeixinAdapter(BasePlatformAdapter):
             except asyncio.CancelledError:
                 pass
         self._poll_task = None
-        if self._session and not self._session.closed:
-            await self._session.close()
-        self._session = None
+        if self._poll_session and not self._poll_session.closed:
+            await self._poll_session.close()
+        self._poll_session = None
+        if self._send_session and not self._send_session.closed:
+            await self._send_session.close()
+        self._send_session = None
         self._release_platform_lock()
         self._mark_disconnected()
         logger.info("[%s] Disconnected", self.name)
 
     async def _poll_loop(self) -> None:
-        assert self._session is not None
+        assert self._poll_session is not None
         sync_buf = _load_sync_buf(self._hermes_home, self._account_id)
         timeout_ms = LONG_POLL_TIMEOUT_MS
         consecutive_failures = 0
@@ -1166,7 +1240,7 @@ class WeixinAdapter(BasePlatformAdapter):
         while self._running:
             try:
                 response = await _get_updates(
-                    self._session,
+                    self._poll_session,
                     base_url=self._base_url,
                     token=self._token,
                     sync_buf=sync_buf,
@@ -1223,7 +1297,7 @@ class WeixinAdapter(BasePlatformAdapter):
             logger.error("[%s] unhandled inbound error from=%s: %s", self.name, _safe_id(message.get("from_user_id")), exc, exc_info=True)
 
     async def _process_message(self, message: Dict[str, Any]) -> None:
-        assert self._session is not None
+        assert self._poll_session is not None
         sender_id = str(message.get("from_user_id") or "").strip()
         if not sender_id:
             return
@@ -1316,7 +1390,7 @@ class WeixinAdapter(BasePlatformAdapter):
         media = _media_reference(item, "image_item")
         try:
             data = await _download_and_decrypt_media(
-                self._session,
+                self._poll_session,
                 cdn_base_url=self._cdn_base_url,
                 encrypted_query_param=media.get("encrypt_query_param"),
                 aes_key_b64=(item.get("image_item") or {}).get("aeskey")
@@ -1334,7 +1408,7 @@ class WeixinAdapter(BasePlatformAdapter):
         media = _media_reference(item, "video_item")
         try:
             data = await _download_and_decrypt_media(
-                self._session,
+                self._poll_session,
                 cdn_base_url=self._cdn_base_url,
                 encrypted_query_param=media.get("encrypt_query_param"),
                 aes_key_b64=media.get("aes_key"),
@@ -1353,7 +1427,7 @@ class WeixinAdapter(BasePlatformAdapter):
         mime = _mime_from_filename(filename)
         try:
             data = await _download_and_decrypt_media(
-                self._session,
+                self._poll_session,
                 cdn_base_url=self._cdn_base_url,
                 encrypted_query_param=media.get("encrypt_query_param"),
                 aes_key_b64=media.get("aes_key"),
@@ -1372,7 +1446,7 @@ class WeixinAdapter(BasePlatformAdapter):
             return None
         try:
             data = await _download_and_decrypt_media(
-                self._session,
+                self._poll_session,
                 cdn_base_url=self._cdn_base_url,
                 encrypted_query_param=media.get("encrypt_query_param"),
                 aes_key_b64=media.get("aes_key"),
@@ -1385,13 +1459,13 @@ class WeixinAdapter(BasePlatformAdapter):
             return None
 
     async def _maybe_fetch_typing_ticket(self, user_id: str, context_token: Optional[str]) -> None:
-        if not self._session or not self._token:
+        if not self._poll_session or not self._token:
             return
         if self._typing_cache.get(user_id):
             return
         try:
             response = await _get_config(
-                self._session,
+                self._poll_session,
                 base_url=self._base_url,
                 token=self._token,
                 user_id=user_id,
@@ -1416,12 +1490,19 @@ class WeixinAdapter(BasePlatformAdapter):
         context_token: Optional[str],
         client_id: str,
     ) -> None:
-        """Send a single text chunk with per-chunk retry and backoff."""
+        """Send a single text chunk with per-chunk retry and backoff.
+
+        On session-expired errors (errcode -14), automatically retries
+        *without* ``context_token`` — iLink accepts tokenless sends as a
+        degraded fallback, which keeps cron-initiated push messages working
+        even when no user message has refreshed the session recently.
+        """
         last_error: Optional[Exception] = None
+        retried_without_token = False
         for attempt in range(self._send_chunk_retries + 1):
             try:
-                await _send_message(
-                    self._session,
+                resp = await _send_message(
+                    self._send_session,
                     base_url=self._base_url,
                     token=self._token,
                     to=chat_id,
@@ -1429,6 +1510,31 @@ class WeixinAdapter(BasePlatformAdapter):
                     context_token=context_token,
                     client_id=client_id,
                 )
+                # Check iLink response for session-expired error
+                if resp and isinstance(resp, dict):
+                    ret = resp.get("ret")
+                    errcode = resp.get("errcode")
+                    if (ret is not None and ret not in (0,)) or (errcode is not None and errcode not in (0,)):
+                        is_session_expired = (
+                            ret == SESSION_EXPIRED_ERRCODE
+                            or errcode == SESSION_EXPIRED_ERRCODE
+                        )
+                        # Session expired — strip token and retry once
+                        if is_session_expired and not retried_without_token and context_token:
+                            retried_without_token = True
+                            context_token = None
+                            self._token_store._cache.pop(
+                                self._token_store._key(self._account_id, chat_id), None
+                            )
+                            logger.warning(
+                                "[%s] session expired for %s; retrying without context_token",
+                                self.name, _safe_id(chat_id),
+                            )
+                            continue
+                        errmsg = resp.get("errmsg") or resp.get("msg") or "unknown error"
+                        raise RuntimeError(
+                            f"iLink sendmessage error: ret={ret} errcode={errcode} errmsg={errmsg}"
+                        )
                 return
             except Exception as exc:
                 last_error = exc
@@ -1456,12 +1562,48 @@ class WeixinAdapter(BasePlatformAdapter):
         reply_to: Optional[str] = None,
         metadata: Optional[Dict[str, Any]] = None,
     ) -> SendResult:
-        if not self._session or not self._token:
+        if not self._send_session or not self._token:
             return SendResult(success=False, error="Not connected")
         context_token = self._token_store.get(self._account_id, chat_id)
         last_message_id: Optional[str] = None
+
+        # Extract MEDIA: tags and bare local file paths before text delivery.
+        media_files, cleaned_content = self.extract_media(content)
+        _, image_cleaned = self.extract_images(cleaned_content)
+        local_files, final_content = self.extract_local_files(image_cleaned)
+
+        _AUDIO_EXTS = {".ogg", ".opus", ".mp3", ".wav", ".m4a"}
+        _VIDEO_EXTS = {".mp4", ".mov", ".avi", ".mkv", ".webm", ".3gp"}
+        _IMAGE_EXTS = {".jpg", ".jpeg", ".png", ".webp", ".gif"}
+
+        async def _deliver_media(path: str, is_voice: bool = False) -> None:
+            ext = Path(path).suffix.lower()
+            if is_voice or ext in _AUDIO_EXTS:
+                await self.send_voice(chat_id=chat_id, audio_path=path, metadata=metadata)
+            elif ext in _VIDEO_EXTS:
+                await self.send_video(chat_id=chat_id, video_path=path, metadata=metadata)
+            elif ext in _IMAGE_EXTS:
+                await self.send_image_file(chat_id=chat_id, image_path=path, metadata=metadata)
+            else:
+                await self.send_document(chat_id=chat_id, file_path=path, metadata=metadata)
+
         try:
-            chunks = [c for c in self._split_text(self.format_message(content)) if c and c.strip()]
+            # Deliver extracted MEDIA: attachments first.
+            for media_path, is_voice in media_files:
+                try:
+                    await _deliver_media(media_path, is_voice)
+                except Exception as exc:
+                    logger.warning("[%s] media delivery failed for %s: %s", self.name, media_path, exc)
+
+            # Deliver bare local file paths.
+            for file_path in local_files:
+                try:
+                    await _deliver_media(file_path, is_voice=False)
+                except Exception as exc:
+                    logger.warning("[%s] local file delivery failed for %s: %s", self.name, file_path, exc)
+
+            # Deliver text content.
+            chunks = [c for c in self._split_text(self.format_message(final_content)) if c and c.strip()]
             for idx, chunk in enumerate(chunks):
                 client_id = f"hermes-weixin-{uuid.uuid4().hex}"
                 await self._send_text_chunk(
@@ -1479,14 +1621,14 @@ class WeixinAdapter(BasePlatformAdapter):
             return SendResult(success=False, error=str(exc))
 
     async def send_typing(self, chat_id: str, metadata: Optional[Dict[str, Any]] = None) -> None:
-        if not self._session or not self._token:
+        if not self._send_session or not self._token:
             return
         typing_ticket = self._typing_cache.get(chat_id)
         if not typing_ticket:
             return
         try:
             await _send_typing(
-                self._session,
+                self._send_session,
                 base_url=self._base_url,
                 token=self._token,
                 to_user_id=chat_id,
@@ -1497,14 +1639,14 @@ class WeixinAdapter(BasePlatformAdapter):
             logger.debug("[%s] typing start failed for %s: %s", self.name, _safe_id(chat_id), exc)
 
     async def stop_typing(self, chat_id: str) -> None:
-        if not self._session or not self._token:
+        if not self._send_session or not self._token:
             return
         typing_ticket = self._typing_cache.get(chat_id)
         if not typing_ticket:
             return
         try:
             await _send_typing(
-                self._session,
+                self._send_session,
                 base_url=self._base_url,
                 token=self._token,
                 to_user_id=chat_id,
@@ -1542,24 +1684,35 @@ class WeixinAdapter(BasePlatformAdapter):
     async def send_image_file(
         self,
         chat_id: str,
-        path: str,
-        caption: str = "",
+        image_path: str,
+        caption: Optional[str] = None,
         reply_to: Optional[str] = None,
         metadata: Optional[Dict[str, Any]] = None,
+        **kwargs,
     ) -> SendResult:
-        return await self.send_document(chat_id, file_path=path, caption=caption, metadata=metadata)
+        del reply_to, kwargs
+        return await self.send_document(
+            chat_id=chat_id,
+            file_path=image_path,
+            caption=caption,
+            metadata=metadata,
+        )
 
     async def send_document(
         self,
         chat_id: str,
         file_path: str,
-        caption: str = "",
+        caption: Optional[str] = None,
+        file_name: Optional[str] = None,
+        reply_to: Optional[str] = None,
         metadata: Optional[Dict[str, Any]] = None,
+        **kwargs,
     ) -> SendResult:
-        if not self._session or not self._token:
+        del file_name, reply_to, metadata, kwargs
+        if not self._send_session or not self._token:
             return SendResult(success=False, error="Not connected")
         try:
-            message_id = await self._send_file(chat_id, file_path, caption)
+            message_id = await self._send_file(chat_id, file_path, caption or "")
             return SendResult(success=True, message_id=message_id)
         except Exception as exc:
             logger.error("[%s] send_document failed to=%s: %s", self.name, _safe_id(chat_id), exc)
@@ -1573,7 +1726,7 @@ class WeixinAdapter(BasePlatformAdapter):
         reply_to: Optional[str] = None,
         metadata: Optional[Dict[str, Any]] = None,
     ) -> SendResult:
-        if not self._session or not self._token:
+        if not self._send_session or not self._token:
             return SendResult(success=False, error="Not connected")
         try:
             message_id = await self._send_file(chat_id, video_path, caption or "")
@@ -1590,7 +1743,24 @@ class WeixinAdapter(BasePlatformAdapter):
         reply_to: Optional[str] = None,
         metadata: Optional[Dict[str, Any]] = None,
     ) -> SendResult:
-        return await self.send_document(chat_id, audio_path, caption=caption or "", metadata=metadata)
+        if not self._send_session or not self._token:
+            return SendResult(success=False, error="Not connected")
+
+        # Native outbound Weixin voice bubbles are not proven-working in the
+        # upstream reference implementation. Prefer a reliable file attachment
+        # fallback so users at least receive playable audio, even for .silk.
+        fallback_caption = caption or "[voice message as attachment]"
+        try:
+            message_id = await self._send_file(
+                chat_id,
+                audio_path,
+                fallback_caption,
+                force_file_attachment=True,
+            )
+            return SendResult(success=True, message_id=message_id)
+        except Exception as exc:
+            logger.error("[%s] send_voice failed to=%s: %s", self.name, _safe_id(chat_id), exc)
+            return SendResult(success=False, error=str(exc))
 
     async def _download_remote_media(self, url: str) -> str:
         from tools.url_safety import is_safe_url
@@ -1598,8 +1768,8 @@ class WeixinAdapter(BasePlatformAdapter):
         if not is_safe_url(url):
             raise ValueError(f"Blocked unsafe URL (SSRF protection): {url}")
 
-        assert self._session is not None
-        async with self._session.get(url, timeout=aiohttp.ClientTimeout(total=30)) as response:
+        assert self._send_session is not None
+        async with self._send_session.get(url, timeout=aiohttp.ClientTimeout(total=30)) as response:
             response.raise_for_status()
             data = await response.read()
             suffix = Path(url.split("?", 1)[0]).suffix or ".bin"
@@ -1607,16 +1777,22 @@ class WeixinAdapter(BasePlatformAdapter):
             handle.write(data)
             return handle.name
 
-    async def _send_file(self, chat_id: str, path: str, caption: str) -> str:
-        assert self._session is not None and self._token is not None
+    async def _send_file(
+        self,
+        chat_id: str,
+        path: str,
+        caption: str,
+        force_file_attachment: bool = False,
+    ) -> str:
+        assert self._send_session is not None and self._token is not None
         plaintext = Path(path).read_bytes()
-        media_type, item_builder = self._outbound_media_builder(path)
+        media_type, item_builder = self._outbound_media_builder(path, force_file_attachment=force_file_attachment)
         filekey = secrets.token_hex(16)
         aes_key = secrets.token_bytes(16)
         rawsize = len(plaintext)
         rawfilemd5 = hashlib.md5(plaintext).hexdigest()
         upload_response = await _get_upload_url(
-            self._session,
+            self._send_session,
             base_url=self._base_url,
             token=self._token,
             to_user_id=chat_id,
@@ -1642,30 +1818,34 @@ class WeixinAdapter(BasePlatformAdapter):
             raise RuntimeError(f"getUploadUrl returned neither upload_param nor upload_full_url: {upload_response}")
 
         encrypted_query_param = await _upload_ciphertext(
-            self._session,
+            self._send_session,
             ciphertext=ciphertext,
             upload_url=upload_url,
         )
-
         context_token = self._token_store.get(self._account_id, chat_id)
         # The iLink API expects aes_key as base64(hex_string), not base64(raw_bytes).
         # Sending base64(raw_bytes) causes images to show as grey boxes on the
         # receiver side because the decryption key doesn't match.
         aes_key_for_api = base64.b64encode(aes_key.hex().encode("ascii")).decode("ascii")
-        media_item = item_builder(
-            encrypt_query_param=encrypted_query_param,
-            aes_key_for_api=aes_key_for_api,
-            ciphertext_size=len(ciphertext),
-            plaintext_size=rawsize,
-            filename=Path(path).name,
-            rawfilemd5=rawfilemd5,
-        )
+        item_kwargs = {
+            "encrypt_query_param": encrypted_query_param,
+            "aes_key_for_api": aes_key_for_api,
+            "ciphertext_size": len(ciphertext),
+            "plaintext_size": rawsize,
+            "filename": Path(path).name,
+            "rawfilemd5": rawfilemd5,
+        }
+        if media_type == MEDIA_VOICE and path.endswith(".silk"):
+            item_kwargs["encode_type"] = 6
+            item_kwargs["sample_rate"] = 24000
+            item_kwargs["bits_per_sample"] = 16
+        media_item = item_builder(**item_kwargs)
 
         last_message_id = None
         if caption:
             last_message_id = f"hermes-weixin-{uuid.uuid4().hex}"
             await _send_message(
-                self._session,
+                self._send_session,
                 base_url=self._base_url,
                 token=self._token,
                 to=chat_id,
@@ -1676,7 +1856,7 @@ class WeixinAdapter(BasePlatformAdapter):
 
         last_message_id = f"hermes-weixin-{uuid.uuid4().hex}"
         await _api_post(
-            self._session,
+            self._send_session,
             base_url=self._base_url,
             endpoint=EP_SEND_MESSAGE,
             payload={
@@ -1695,7 +1875,7 @@ class WeixinAdapter(BasePlatformAdapter):
         )
         return last_message_id
 
-    def _outbound_media_builder(self, path: str):
+    def _outbound_media_builder(self, path: str, force_file_attachment: bool = False):
         mime = mimetypes.guess_type(path)[0] or "application/octet-stream"
         if mime.startswith("image/"):
             return MEDIA_IMAGE, lambda **kw: {
@@ -1723,7 +1903,7 @@ class WeixinAdapter(BasePlatformAdapter):
                     "video_md5": kw.get("rawfilemd5", ""),
                 },
             }
-        if mime.startswith("audio/") or path.endswith(".silk"):
+        if path.endswith(".silk") and not force_file_attachment:
             return MEDIA_VOICE, lambda **kw: {
                 "type": ITEM_VOICE,
                 "voice_item": {
@@ -1732,9 +1912,25 @@ class WeixinAdapter(BasePlatformAdapter):
                         "aes_key": kw["aes_key_for_api"],
                         "encrypt_type": 1,
                     },
+                    "encode_type": kw.get("encode_type"),
+                    "bits_per_sample": kw.get("bits_per_sample"),
+                    "sample_rate": kw.get("sample_rate"),
                     "playtime": kw.get("playtime", 0),
                 },
             }
+        if mime.startswith("audio/"):
+            return MEDIA_FILE, lambda **kw: {
+                "type": ITEM_FILE,
+                "file_item": {
+                    "media": {
+                        "encrypt_query_param": kw["encrypt_query_param"],
+                        "aes_key": kw["aes_key_for_api"],
+                        "encrypt_type": 1,
+                    },
+                    "file_name": kw["filename"],
+                    "len": str(kw["plaintext_size"]),
+                },
+            }
         return MEDIA_FILE, lambda **kw: {
             "type": ITEM_FILE,
             "file_item": {
@@ -1784,7 +1980,34 @@ async def send_weixin_direct(
     token_store.restore(account_id)
     context_token = token_store.get(account_id, chat_id)
 
-    async with aiohttp.ClientSession(trust_env=True) as session:
+    live_adapter = _LIVE_ADAPTERS.get(resolved_token)
+    send_session = getattr(live_adapter, '_send_session', None)
+    if live_adapter is not None and send_session is not None and not send_session.closed:
+        last_result: Optional[SendResult] = None
+        cleaned = live_adapter.format_message(message)
+        if cleaned:
+            last_result = await live_adapter.send(chat_id, cleaned)
+            if not last_result.success:
+                return {"error": f"Weixin send failed: {last_result.error}"}
+
+        for media_path, _is_voice in media_files or []:
+            ext = Path(media_path).suffix.lower()
+            if ext in {".jpg", ".jpeg", ".png", ".gif", ".webp", ".bmp"}:
+                last_result = await live_adapter.send_image_file(chat_id, media_path)
+            else:
+                last_result = await live_adapter.send_document(chat_id, media_path)
+            if not last_result.success:
+                return {"error": f"Weixin media send failed: {last_result.error}"}
+
+        return {
+            "success": True,
+            "platform": "weixin",
+            "chat_id": chat_id,
+            "message_id": last_result.message_id if last_result else None,
+            "context_token_used": bool(context_token),
+        }
+
+    async with aiohttp.ClientSession(trust_env=True, connector=_make_ssl_connector()) as session:
         adapter = WeixinAdapter(
             PlatformConfig(
                 enabled=True,
@@ -1797,6 +2020,7 @@ async def send_weixin_direct(
                 },
             )
         )
+        adapter._send_session = session
         adapter._session = session
         adapter._token = resolved_token
         adapter._account_id = account_id

gateway/platforms/whatsapp.py

@@ -289,33 +289,35 @@ class WhatsAppAdapter(BasePlatformAdapter):
         logger.info("[%s] Bridge found at %s", self.name, bridge_path)
         
         # Acquire scoped lock to prevent duplicate sessions
+        lock_acquired = False
         try:
             if not self._acquire_platform_lock('whatsapp-session', str(self._session_path), 'WhatsApp session'):
                 return False
+            lock_acquired = True
         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():
-            print(f"[{self.name}] Installing WhatsApp bridge dependencies...")
-            try:
-                install_result = subprocess.run(
-                    ["npm", "install", "--silent"],
-                    cwd=str(bridge_dir),
-                    capture_output=True,
-                    text=True,
-                    timeout=60,
-                )
-                if install_result.returncode != 0:
-                    print(f"[{self.name}] npm install failed: {install_result.stderr}")
-                    return False
-                print(f"[{self.name}] Dependencies installed")
-            except Exception as e:
-                print(f"[{self.name}] Failed to install dependencies: {e}")
-                return False
-        
         try:
+            # Auto-install npm dependencies if node_modules doesn't exist
+            bridge_dir = bridge_path.parent
+            if not (bridge_dir / "node_modules").exists():
+                print(f"[{self.name}] Installing WhatsApp bridge dependencies...")
+                try:
+                    install_result = subprocess.run(
+                        ["npm", "install", "--silent"],
+                        cwd=str(bridge_dir),
+                        capture_output=True,
+                        text=True,
+                        timeout=60,
+                    )
+                    if install_result.returncode != 0:
+                        print(f"[{self.name}] npm install failed: {install_result.stderr}")
+                        return False
+                    print(f"[{self.name}] Dependencies installed")
+                except Exception as e:
+                    print(f"[{self.name}] Failed to install dependencies: {e}")
+                    return False
+
             # Ensure session directory exists
             self._session_path.mkdir(parents=True, exist_ok=True)
             
@@ -452,10 +454,13 @@ class WhatsAppAdapter(BasePlatformAdapter):
             return True
             
         except Exception as e:
-            self._release_platform_lock()
             logger.error("[%s] Failed to start bridge: %s", self.name, e, exc_info=True)
-            self._close_bridge_log()
             return False
+        finally:
+            if not self._running:
+                if lock_acquired:
+                    self._release_platform_lock()
+                self._close_bridge_log()
     
     def _close_bridge_log(self) -> None:
         """Close the bridge log file handle if open."""
@@ -655,6 +660,8 @@ class WhatsAppAdapter(BasePlatformAdapter):
         chat_id: str,
         message_id: str,
         content: str,
+        *,
+        finalize: bool = False,
     ) -> SendResult:
         """Edit a previously sent message via the WhatsApp bridge."""
         if not self._running or not self._http_session:

gateway/session.py

@@ -82,6 +82,7 @@ class SessionSource:
     chat_topic: Optional[str] = None  # Channel topic/description (Discord, Slack)
     user_id_alt: Optional[str] = None  # Signal UUID (alternative to phone number)
     chat_id_alt: Optional[str] = None  # Signal group internal ID
+    is_bot: bool = False  # True when the message author is a bot/webhook (Discord)
     
     @property
     def description(self) -> str:
@@ -301,6 +302,8 @@ def build_session_context_prompt(
     lines.append("")
     lines.append("**Delivery options for scheduled tasks:**")
     
+    from hermes_constants import display_hermes_home
+
     # Origin delivery
     if context.source.platform == Platform.LOCAL:
         lines.append("- `\"origin\"` → Local output (saved to files)")
@@ -309,9 +312,11 @@ def build_session_context_prompt(
             _hash_chat_id(context.source.chat_id) if redact_pii else context.source.chat_id
         )
         lines.append(f"- `\"origin\"` → Back to this chat ({_origin_label})")
-    
+
     # Local always available
-    lines.append("- `\"local\"` → Save to local files only (~/.hermes/cron/output/)")
+    lines.append(
+        f"- `\"local\"` → Save to local files only ({display_hermes_home()}/cron/output/)"
+    )
     
     # Platform home channels
     for platform, home in context.home_channels.items():
@@ -372,7 +377,19 @@ class SessionEntry:
     # this session (create a new session_id) so the user starts fresh.
     # Set by /stop to break stuck-resume loops (#7536).
     suspended: bool = False
-    
+
+    # When True the session was interrupted by a gateway restart/shutdown
+    # drain timeout, but recovery is still expected.  Unlike ``suspended``,
+    # ``resume_pending`` preserves the existing session_id on next access —
+    # the user stays on the same transcript and the agent auto-continues
+    # from where it left off.  Cleared after the next successful turn.
+    # Escalation to ``suspended`` is handled by the existing
+    # ``.restart_failure_counts`` stuck-loop counter (#7536), not by a
+    # parallel counter on this entry.
+    resume_pending: bool = False
+    resume_reason: Optional[str] = None  # e.g. "restart_timeout"
+    last_resume_marked_at: Optional[datetime] = None
+
     def to_dict(self) -> Dict[str, Any]:
         result = {
             "session_key": self.session_key,
@@ -392,6 +409,13 @@ class SessionEntry:
             "cost_status": self.cost_status,
             "memory_flushed": self.memory_flushed,
             "suspended": self.suspended,
+            "resume_pending": self.resume_pending,
+            "resume_reason": self.resume_reason,
+            "last_resume_marked_at": (
+                self.last_resume_marked_at.isoformat()
+                if self.last_resume_marked_at
+                else None
+            ),
         }
         if self.origin:
             result["origin"] = self.origin.to_dict()
@@ -409,7 +433,15 @@ class SessionEntry:
                 platform = Platform(data["platform"])
             except ValueError as e:
                 logger.debug("Unknown platform value %r: %s", data["platform"], e)
-        
+
+        last_resume_marked_at = None
+        _lrma = data.get("last_resume_marked_at")
+        if _lrma:
+            try:
+                last_resume_marked_at = datetime.fromisoformat(_lrma)
+            except (TypeError, ValueError):
+                last_resume_marked_at = None
+
         return cls(
             session_key=data["session_key"],
             session_id=data["session_id"],
@@ -429,6 +461,9 @@ class SessionEntry:
             cost_status=data.get("cost_status", "unknown"),
             memory_flushed=data.get("memory_flushed", False),
             suspended=data.get("suspended", False),
+            resume_pending=data.get("resume_pending", False),
+            resume_reason=data.get("resume_reason"),
+            last_resume_marked_at=last_resume_marked_at,
         )
 
 
@@ -705,9 +740,23 @@ class SessionStore:
                 entry = self._entries[session_key]
 
                 # Auto-reset sessions marked as suspended (e.g. after /stop
-                # broke a stuck loop — #7536).
+                # broke a stuck loop — #7536).  ``suspended`` is the hard
+                # forced-wipe signal and always wins over ``resume_pending``,
+                # so repeated interrupted restarts that escalate via the
+                # existing ``.restart_failure_counts`` stuck-loop counter
+                # still converge to a clean slate.
                 if entry.suspended:
                     reset_reason = "suspended"
+                elif entry.resume_pending:
+                    # Restart-interrupted session: preserve the session_id
+                    # and return the existing entry so the transcript
+                    # reloads intact.  ``resume_pending`` is cleared after
+                    # the NEXT successful turn completes (not here), which
+                    # means a re-interrupted retry keeps trying — the
+                    # stuck-loop counter handles terminal escalation.
+                    entry.updated_at = now
+                    self._save()
+                    return entry
                 else:
                     reset_reason = self._should_reset(entry, source)
                 if not reset_reason:
@@ -797,6 +846,112 @@ class SessionStore:
                 return True
         return False
 
+    def mark_resume_pending(
+        self,
+        session_key: str,
+        reason: str = "restart_timeout",
+    ) -> bool:
+        """Mark a session as resumable after a restart interruption.
+
+        Unlike ``suspend_session()``, this preserves the existing
+        ``session_id`` and the transcript.  The next call to
+        ``get_or_create_session()`` for this key returns the same entry
+        so the user auto-resumes on the same conversation lane.
+
+        Returns True if the session existed and was marked.
+        """
+        with self._lock:
+            self._ensure_loaded_locked()
+            if session_key in self._entries:
+                entry = self._entries[session_key]
+                # Never override an explicit ``suspended`` — that is a hard
+                # forced-wipe signal (from /stop or stuck-loop escalation).
+                if entry.suspended:
+                    return False
+                entry.resume_pending = True
+                entry.resume_reason = reason
+                entry.last_resume_marked_at = _now()
+                self._save()
+                return True
+        return False
+
+    def clear_resume_pending(self, session_key: str) -> bool:
+        """Clear the resume-pending flag after a successful resumed turn.
+
+        Called from the gateway after ``run_conversation()`` returns a
+        final response for a session that had ``resume_pending=True``,
+        signalling that recovery succeeded.
+
+        Returns True if a flag was cleared.
+        """
+        with self._lock:
+            self._ensure_loaded_locked()
+            entry = self._entries.get(session_key)
+            if entry is None or not entry.resume_pending:
+                return False
+            entry.resume_pending = False
+            entry.resume_reason = None
+            entry.last_resume_marked_at = None
+            self._save()
+            return True
+
+    def prune_old_entries(self, max_age_days: int) -> int:
+        """Drop SessionEntry records older than max_age_days.
+
+        Pruning is based on ``updated_at`` (last activity), not ``created_at``.
+        A session that's been active within the window is kept regardless of
+        how old it is.  Entries marked ``suspended`` are kept — the user
+        explicitly paused them for later resume.  Entries held by an active
+        process (via has_active_processes_fn) are also kept so long-running
+        background work isn't orphaned.
+
+        Pruning is functionally identical to a natural reset-policy expiry:
+        the transcript in SQLite stays, but the session_key → session_id
+        mapping is dropped and the user starts a fresh session on return.
+
+        ``max_age_days <= 0`` disables pruning; returns 0 immediately.
+        Returns the number of entries removed.
+        """
+        if max_age_days is None or max_age_days <= 0:
+            return 0
+        from datetime import timedelta
+
+        cutoff = _now() - timedelta(days=max_age_days)
+        removed_keys: list[str] = []
+
+        with self._lock:
+            self._ensure_loaded_locked()
+            for key, entry in list(self._entries.items()):
+                if entry.suspended:
+                    continue
+                # Never prune sessions with an active background process
+                # attached — the user may still be waiting on output.
+                # The callback is keyed by session_key (see process_registry.
+                # has_active_for_session); passing session_id here used to
+                # never match, so active sessions got pruned anyway.
+                if self._has_active_processes_fn is not None:
+                    try:
+                        if self._has_active_processes_fn(entry.session_key):
+                            continue
+                    except Exception as exc:
+                        logger.debug(
+                            "has_active_processes_fn raised during prune for %s: %s",
+                            entry.session_key, exc,
+                        )
+                if entry.updated_at < cutoff:
+                    removed_keys.append(key)
+            for key in removed_keys:
+                self._entries.pop(key, None)
+            if removed_keys:
+                self._save()
+
+        if removed_keys:
+            logger.info(
+                "SessionStore pruned %d entries older than %d days",
+                len(removed_keys), max_age_days,
+            )
+        return len(removed_keys)
+
     def suspend_recently_active(self, max_age_seconds: int = 120) -> int:
         """Mark recently-active sessions as suspended.
 
@@ -805,6 +960,12 @@ class SessionStore:
         (#7536).  Only suspends sessions updated within *max_age_seconds*
         to avoid resetting long-idle sessions that are harmless to resume.
         Returns the number of sessions that were suspended.
+
+        Entries flagged ``resume_pending=True`` are skipped — those were
+        marked intentionally by the drain-timeout path as recoverable.
+        Terminal escalation for genuinely stuck ``resume_pending`` sessions
+        is handled by the existing ``.restart_failure_counts`` stuck-loop
+        counter, which runs after this method on startup.
         """
         from datetime import timedelta
 
@@ -813,6 +974,8 @@ class SessionStore:
         with self._lock:
             self._ensure_loaded_locked()
             for entry in self._entries.values():
+                if entry.resume_pending:
+                    continue
                 if not entry.suspended and entry.updated_at >= cutoff:
                     entry.suspended = True
                     count += 1

gateway/session_context.py

@@ -37,18 +37,24 @@ needs to replace the import + call site:
 """
 
 from contextvars import ContextVar
+from typing import Any
+
+# Sentinel to distinguish "never set in this context" from "explicitly set to empty".
+# When a contextvar holds _UNSET, we fall back to os.environ (CLI/cron compat).
+# When it holds "" (after clear_session_vars resets it), we return "" — no fallback.
+_UNSET: Any = object()
 
 # ---------------------------------------------------------------------------
 # Per-task session variables
 # ---------------------------------------------------------------------------
 
-_SESSION_PLATFORM: ContextVar[str] = ContextVar("HERMES_SESSION_PLATFORM", default="")
-_SESSION_CHAT_ID: ContextVar[str] = ContextVar("HERMES_SESSION_CHAT_ID", default="")
-_SESSION_CHAT_NAME: ContextVar[str] = ContextVar("HERMES_SESSION_CHAT_NAME", default="")
-_SESSION_THREAD_ID: ContextVar[str] = ContextVar("HERMES_SESSION_THREAD_ID", default="")
-_SESSION_USER_ID: ContextVar[str] = ContextVar("HERMES_SESSION_USER_ID", default="")
-_SESSION_USER_NAME: ContextVar[str] = ContextVar("HERMES_SESSION_USER_NAME", default="")
-_SESSION_KEY: ContextVar[str] = ContextVar("HERMES_SESSION_KEY", default="")
+_SESSION_PLATFORM: ContextVar = ContextVar("HERMES_SESSION_PLATFORM", default=_UNSET)
+_SESSION_CHAT_ID: ContextVar = ContextVar("HERMES_SESSION_CHAT_ID", default=_UNSET)
+_SESSION_CHAT_NAME: ContextVar = ContextVar("HERMES_SESSION_CHAT_NAME", default=_UNSET)
+_SESSION_THREAD_ID: ContextVar = ContextVar("HERMES_SESSION_THREAD_ID", default=_UNSET)
+_SESSION_USER_ID: ContextVar = ContextVar("HERMES_SESSION_USER_ID", default=_UNSET)
+_SESSION_USER_NAME: ContextVar = ContextVar("HERMES_SESSION_USER_NAME", default=_UNSET)
+_SESSION_KEY: ContextVar = ContextVar("HERMES_SESSION_KEY", default=_UNSET)
 
 _VAR_MAP = {
     "HERMES_SESSION_PLATFORM": _SESSION_PLATFORM,
@@ -91,10 +97,17 @@ def set_session_vars(
 
 
 def clear_session_vars(tokens: list) -> None:
-    """Restore session context variables to their pre-handler values."""
-    if not tokens:
-        return
-    vars_in_order = [
+    """Mark session context variables as explicitly cleared.
+
+    Sets all variables to ``""`` so that ``get_session_env`` returns an empty
+    string instead of falling back to (potentially stale) ``os.environ``
+    values.  The *tokens* argument is accepted for API compatibility with
+    callers that saved the return value of ``set_session_vars``, but the
+    actual clearing uses ``var.set("")`` rather than ``var.reset(token)``
+    to ensure the "explicitly cleared" state is distinguishable from
+    "never set" (which holds the ``_UNSET`` sentinel).
+    """
+    for var in (
         _SESSION_PLATFORM,
         _SESSION_CHAT_ID,
         _SESSION_CHAT_NAME,
@@ -102,9 +115,8 @@ def clear_session_vars(tokens: list) -> None:
         _SESSION_USER_ID,
         _SESSION_USER_NAME,
         _SESSION_KEY,
-    ]
-    for var, token in zip(vars_in_order, tokens):
-        var.reset(token)
+    ):
+        var.set("")
 
 
 def get_session_env(name: str, default: str = "") -> str:
@@ -113,8 +125,13 @@ def get_session_env(name: str, default: str = "") -> str:
     Drop-in replacement for ``os.getenv("HERMES_SESSION_*", default)``.
 
     Resolution order:
-    1. Context variable (set by the gateway for concurrency-safe access)
-    2. ``os.environ`` (used by CLI, cron scheduler, and tests)
+    1. Context variable (set by the gateway for concurrency-safe access).
+       If the variable was explicitly set (even to ``""``) via
+       ``set_session_vars`` or ``clear_session_vars``, that value is
+       returned — **no fallback to os.environ**.
+    2. ``os.environ`` (only when the context variable was never set in
+       this context — i.e. CLI, cron scheduler, and test processes that
+       don't use ``set_session_vars`` at all).
     3. *default*
     """
     import os
@@ -122,7 +139,7 @@ def get_session_env(name: str, default: str = "") -> str:
     var = _VAR_MAP.get(name)
     if var is not None:
         value = var.get()
-        if value:
+        if value is not _UNSET:
             return value
     # Fall back to os.environ for CLI, cron, and test compatibility
     return os.getenv(name, default)

gateway/status.py

@@ -188,8 +188,8 @@ def _write_json_file(path: Path, payload: dict[str, Any]) -> None:
     path.write_text(json.dumps(payload))
 
 
-def _read_pid_record() -> Optional[dict]:
-    pid_path = _get_pid_path()
+def _read_pid_record(pid_path: Optional[Path] = None) -> Optional[dict]:
+    pid_path = pid_path or _get_pid_path()
     if not pid_path.exists():
         return None
 
@@ -212,6 +212,18 @@ def _read_pid_record() -> Optional[dict]:
     return None
 
 
+def _cleanup_invalid_pid_path(pid_path: Path, *, cleanup_stale: bool) -> None:
+    if not cleanup_stale:
+        return
+    try:
+        if pid_path == _get_pid_path():
+            remove_pid_file()
+        else:
+            pid_path.unlink(missing_ok=True)
+    except Exception:
+        pass
+
+
 def write_pid_file() -> None:
     """Write the current process PID and metadata to the gateway PID file."""
     _write_json_file(_get_pid_path(), _build_pid_record())
@@ -413,43 +425,179 @@ def release_all_scoped_locks() -> int:
     return removed
 
 
-def get_running_pid() -> Optional[int]:
+# ── --replace takeover marker ─────────────────────────────────────────
+#
+# When a new gateway starts with ``--replace``, it SIGTERMs the existing
+# gateway so it can take over the bot token. PR #5646 made SIGTERM exit
+# the gateway with code 1 so ``Restart=on-failure`` can revive it after
+# unexpected kills — but that also means a --replace takeover target
+# exits 1, which tricks systemd into reviving it 30 seconds later,
+# starting a flap loop against the replacer when both services are
+# enabled in the user's systemd (e.g. ``hermes.service`` + ``hermes-
+# gateway.service``).
+#
+# The takeover marker breaks the loop: the replacer writes a short-lived
+# file naming the target PID + start_time BEFORE sending SIGTERM.
+# The target's shutdown handler reads the marker and, if it names
+# this process, treats the SIGTERM as a planned takeover and exits 0.
+# The marker is unlinked after the target has consumed it, so a stale
+# marker left by a crashed replacer can grief at most one future
+# shutdown on the same PID — and only within _TAKEOVER_MARKER_TTL_S.
+
+_TAKEOVER_MARKER_FILENAME = ".gateway-takeover.json"
+_TAKEOVER_MARKER_TTL_S = 60  # Marker older than this is treated as stale
+
+
+def _get_takeover_marker_path() -> Path:
+    """Return the path to the --replace takeover marker file."""
+    home = get_hermes_home()
+    return home / _TAKEOVER_MARKER_FILENAME
+
+
+def write_takeover_marker(target_pid: int) -> bool:
+    """Record that ``target_pid`` is being replaced by the current process.
+
+    Captures the target's ``start_time`` so that PID reuse after the
+    target exits cannot later match the marker. Also records the
+    replacer's PID and a UTC timestamp for TTL-based staleness checks.
+
+    Returns True on successful write, False on any failure. The caller
+    should proceed with the SIGTERM even if the write fails (the marker
+    is a best-effort signal, not a correctness requirement).
+    """
+    try:
+        target_start_time = _get_process_start_time(target_pid)
+        record = {
+            "target_pid": target_pid,
+            "target_start_time": target_start_time,
+            "replacer_pid": os.getpid(),
+            "written_at": _utc_now_iso(),
+        }
+        _write_json_file(_get_takeover_marker_path(), record)
+        return True
+    except (OSError, PermissionError):
+        return False
+
+
+def consume_takeover_marker_for_self() -> bool:
+    """Check & unlink the takeover marker if it names the current process.
+
+    Returns True only when a valid (non-stale) marker names this PID +
+    start_time. A returning True indicates the current SIGTERM is a
+    planned --replace takeover; the caller should exit 0 instead of
+    signalling ``_signal_initiated_shutdown``.
+
+    Always unlinks the marker on match (and on detected staleness) so
+    subsequent unrelated signals don't re-trigger.
+    """
+    path = _get_takeover_marker_path()
+    record = _read_json_file(path)
+    if not record:
+        return False
+
+    # Any malformed or stale marker → drop it and return False
+    try:
+        target_pid = int(record["target_pid"])
+        target_start_time = record.get("target_start_time")
+        written_at = record.get("written_at") or ""
+    except (KeyError, TypeError, ValueError):
+        try:
+            path.unlink(missing_ok=True)
+        except OSError:
+            pass
+        return False
+
+    # TTL guard: a stale marker older than _TAKEOVER_MARKER_TTL_S is ignored.
+    stale = False
+    try:
+        written_dt = datetime.fromisoformat(written_at)
+        age = (datetime.now(timezone.utc) - written_dt).total_seconds()
+        if age > _TAKEOVER_MARKER_TTL_S:
+            stale = True
+    except (TypeError, ValueError):
+        stale = True  # Unparseable timestamp — treat as stale
+
+    if stale:
+        try:
+            path.unlink(missing_ok=True)
+        except OSError:
+            pass
+        return False
+
+    # Does the marker name THIS process?
+    our_pid = os.getpid()
+    our_start_time = _get_process_start_time(our_pid)
+    matches = (
+        target_pid == our_pid
+        and target_start_time is not None
+        and our_start_time is not None
+        and target_start_time == our_start_time
+    )
+
+    # Consume the marker whether it matched or not — a marker that doesn't
+    # match our identity is stale-for-us anyway.
+    try:
+        path.unlink(missing_ok=True)
+    except OSError:
+        pass
+
+    return matches
+
+
+def clear_takeover_marker() -> None:
+    """Remove the takeover marker unconditionally. Safe to call repeatedly."""
+    try:
+        _get_takeover_marker_path().unlink(missing_ok=True)
+    except OSError:
+        pass
+
+
+def get_running_pid(
+    pid_path: Optional[Path] = None,
+    *,
+    cleanup_stale: bool = True,
+) -> Optional[int]:
     """Return the PID of a running gateway instance, or ``None``.
 
     Checks the PID file and verifies the process is actually alive.
     Cleans up stale PID files automatically.
     """
-    record = _read_pid_record()
+    resolved_pid_path = pid_path or _get_pid_path()
+    record = _read_pid_record(resolved_pid_path)
     if not record:
-        remove_pid_file()
+        _cleanup_invalid_pid_path(resolved_pid_path, cleanup_stale=cleanup_stale)
         return None
 
     try:
         pid = int(record["pid"])
     except (KeyError, TypeError, ValueError):
-        remove_pid_file()
+        _cleanup_invalid_pid_path(resolved_pid_path, cleanup_stale=cleanup_stale)
         return None
 
     try:
         os.kill(pid, 0)  # signal 0 = existence check, no actual signal sent
     except (ProcessLookupError, PermissionError):
-        remove_pid_file()
+        _cleanup_invalid_pid_path(resolved_pid_path, cleanup_stale=cleanup_stale)
         return None
 
     recorded_start = record.get("start_time")
     current_start = _get_process_start_time(pid)
     if recorded_start is not None and current_start is not None and current_start != recorded_start:
-        remove_pid_file()
+        _cleanup_invalid_pid_path(resolved_pid_path, cleanup_stale=cleanup_stale)
         return None
 
     if not _looks_like_gateway_process(pid):
         if not _record_looks_like_gateway(record):
-            remove_pid_file()
+            _cleanup_invalid_pid_path(resolved_pid_path, cleanup_stale=cleanup_stale)
             return None
 
     return pid
 
 
-def is_gateway_running() -> bool:
+def is_gateway_running(
+    pid_path: Optional[Path] = None,
+    *,
+    cleanup_stale: bool = True,
+) -> bool:
     """Check if the gateway daemon is currently running."""
-    return get_running_pid() is not None
+    return get_running_pid(pid_path, cleanup_stale=cleanup_stale) is not None

gateway/stream_consumer.py

@@ -43,6 +43,7 @@ class StreamConsumerConfig:
     edit_interval: float = 1.0
     buffer_threshold: int = 40
     cursor: str = " ▉"
+    buffer_only: bool = False
 
 
 class GatewayStreamConsumer:
@@ -99,6 +100,14 @@ class GatewayStreamConsumer:
         self._flood_strikes = 0         # Consecutive flood-control edit failures
         self._current_edit_interval = self.cfg.edit_interval  # Adaptive backoff
         self._final_response_sent = False
+        # Cache adapter lifecycle capability: only platforms that need an
+        # explicit finalize call (e.g. DingTalk AI Cards) force us to make
+        # a redundant final edit.  Everyone else keeps the fast path.
+        # Use ``is True`` (not ``bool(...)``) so MagicMock attribute access
+        # in tests doesn't incorrectly enable this path.
+        self._adapter_requires_finalize: bool = (
+            getattr(adapter, "REQUIRES_EDIT_FINALIZE", False) is True
+        )
 
         # Think-block filter state (mirrors CLI's _stream_delta tag suppression)
         self._in_think_block = False
@@ -295,10 +304,13 @@ class GatewayStreamConsumer:
                     got_done
                     or got_segment_break
                     or commentary_text is not None
-                    or (elapsed >= self._current_edit_interval
-                        and self._accumulated)
-                    or len(self._accumulated) >= self.cfg.buffer_threshold
                 )
+                if not self.cfg.buffer_only:
+                    should_edit = should_edit or (
+                        (elapsed >= self._current_edit_interval
+                            and self._accumulated)
+                        or len(self._accumulated) >= self.cfg.buffer_threshold
+                    )
 
                 current_update_visible = False
                 if should_edit and self._accumulated:
@@ -357,7 +369,16 @@ class GatewayStreamConsumer:
                     if not got_done and not got_segment_break and commentary_text is None:
                         display_text += self.cfg.cursor
 
-                    current_update_visible = await self._send_or_edit(display_text)
+                    # Segment break: finalize the current message so platforms
+                    # that need explicit closure (e.g. DingTalk AI Cards) don't
+                    # leave the previous segment stuck in a loading state when
+                    # the next segment (tool progress, next chunk) creates a
+                    # new message below it.  got_done has its own finalize
+                    # path below so we don't finalize here for it.
+                    current_update_visible = await self._send_or_edit(
+                        display_text,
+                        finalize=got_segment_break,
+                    )
                     self._last_edit_time = time.monotonic()
 
                 if got_done:
@@ -368,10 +389,22 @@ class GatewayStreamConsumer:
                     if self._accumulated:
                         if self._fallback_final_send:
                             await self._send_fallback_final(self._accumulated)
-                        elif current_update_visible:
+                        elif (
+                            current_update_visible
+                            and not self._adapter_requires_finalize
+                        ):
+                            # Mid-stream edit above already delivered the
+                            # final accumulated content.  Skip the redundant
+                            # final edit — but only for adapters that don't
+                            # need an explicit finalize signal.
                             self._final_response_sent = True
                         elif self._message_id:
-                            self._final_response_sent = await self._send_or_edit(self._accumulated)
+                            # Either the mid-stream edit didn't run (no
+                            # visible update this tick) OR the adapter needs
+                            # explicit finalize=True to close the stream.
+                            self._final_response_sent = await self._send_or_edit(
+                                self._accumulated, finalize=True,
+                            )
                         elif not self._already_sent:
                             self._final_response_sent = await self._send_or_edit(self._accumulated)
                     return
@@ -397,24 +430,41 @@ class GatewayStreamConsumer:
                 # a real string like "msg_1", not "__no_edit__", so that case
                 # still resets and creates a fresh segment as intended.)
                 if got_segment_break:
+                    # If the segment-break edit failed to deliver the
+                    # accumulated content (flood control that has not yet
+                    # promoted to fallback mode, or fallback mode itself),
+                    # _accumulated still holds pre-boundary text the user
+                    # never saw. Flush that tail as a continuation message
+                    # before the reset below wipes _accumulated — otherwise
+                    # text generated before the tool boundary is silently
+                    # dropped (issue #8124).
+                    if (
+                        self._accumulated
+                        and not current_update_visible
+                        and self._message_id
+                        and self._message_id != "__no_edit__"
+                    ):
+                        await self._flush_segment_tail_on_edit_failure()
                     self._reset_segment_state(preserve_no_edit=True)
 
                 await asyncio.sleep(0.05)  # Small yield to not busy-loop
 
         except asyncio.CancelledError:
             # Best-effort final edit on cancellation
+            _best_effort_ok = False
             if self._accumulated and self._message_id:
                 try:
-                    await self._send_or_edit(self._accumulated)
+                    _best_effort_ok = bool(await self._send_or_edit(self._accumulated))
                 except Exception:
                     pass
-            # If we delivered any content before being cancelled, mark the
-            # final response as sent so the gateway's already_sent check
-            # doesn't trigger a duplicate message.  The 5-second
-            # stream_task timeout (gateway/run.py) can cancel us while
-            # waiting on a slow Telegram API call — without this flag the
-            # gateway falls through to the normal send path.
-            if self._already_sent:
+            # Only confirm final delivery if the best-effort send above
+            # actually succeeded OR if the final response was already
+            # confirmed before we were cancelled.  Previously this
+            # promoted any partial send (already_sent=True) to
+            # final_response_sent — which suppressed the gateway's
+            # fallback send even when only intermediate text (e.g.
+            # "Let me search…") had been delivered, not the real answer.
+            if _best_effort_ok and not self._final_response_sent:
                 self._final_response_sent = True
         except Exception as e:
             logger.error("Stream consumer error: %s", e)
@@ -513,9 +563,41 @@ class GatewayStreamConsumer:
         self._fallback_final_send = False
         if not continuation.strip():
             # Nothing new to send — the visible partial already matches final text.
-            self._already_sent = True
-            self._final_response_sent = True
-            return
+            # BUT: if final_text itself has meaningful content (e.g. a timeout
+            # message after a long tool call), the prefix-based continuation
+            # calculation may wrongly conclude "already shown" because the
+            # streamed prefix was from a *previous* segment (before the tool
+            # boundary).  In that case, send the full final_text as-is (#10807).
+            if final_text.strip() and final_text != self._visible_prefix():
+                continuation = final_text
+            else:
+                # Defence-in-depth for #7183: the last edit may still show the
+                # cursor character because fallback mode was entered after an
+                # edit failure left it stuck.  Try one final edit to strip it
+                # so the message doesn't freeze with a visible ▉.  Best-effort
+                # — if this edit also fails (flood control still active),
+                # _try_strip_cursor has already been called on fallback entry
+                # and the adaptive-backoff retries will have had their shot.
+                if (
+                    self._message_id
+                    and self._last_sent_text
+                    and self.cfg.cursor
+                    and self._last_sent_text.endswith(self.cfg.cursor)
+                ):
+                    clean_text = self._last_sent_text[:-len(self.cfg.cursor)]
+                    try:
+                        result = await self.adapter.edit_message(
+                            chat_id=self.chat_id,
+                            message_id=self._message_id,
+                            content=clean_text,
+                        )
+                        if result.success:
+                            self._last_sent_text = clean_text
+                    except Exception:
+                        pass
+                self._already_sent = True
+                self._final_response_sent = True
+                return
 
         raw_limit = getattr(self.adapter, "MAX_MESSAGE_LENGTH", 4096)
         safe_limit = max(500, raw_limit - 100)
@@ -577,6 +659,39 @@ class GatewayStreamConsumer:
         err_lower = err.lower()
         return "flood" in err_lower or "retry after" in err_lower or "rate" in err_lower
 
+    async def _flush_segment_tail_on_edit_failure(self) -> None:
+        """Deliver un-sent tail content before a segment-break reset.
+
+        When an edit fails (flood control, transport error) and a tool
+        boundary arrives before the next retry, ``_accumulated`` holds text
+        that was generated but never shown to the user. Without this flush,
+        the segment reset would discard that tail and leave a frozen cursor
+        in the partial message.
+
+        Sends the tail that sits after the last successfully-delivered
+        prefix as a new message, and best-effort strips the stuck cursor
+        from the previous partial message.
+        """
+        if not self._fallback_final_send:
+            await self._try_strip_cursor()
+        visible = self._fallback_prefix or self._visible_prefix()
+        tail = self._accumulated
+        if visible and tail.startswith(visible):
+            tail = tail[len(visible):].lstrip()
+        tail = self._clean_for_display(tail)
+        if not tail.strip():
+            return
+        try:
+            result = await self.adapter.send(
+                chat_id=self.chat_id,
+                content=tail,
+                metadata=self.metadata,
+            )
+            if result.success:
+                self._already_sent = True
+        except Exception as e:
+            logger.error("Segment-break tail flush error: %s", e)
+
     async def _try_strip_cursor(self) -> None:
         """Best-effort edit to remove the cursor from the last visible message.
 
@@ -609,19 +724,25 @@ class GatewayStreamConsumer:
                 content=text,
                 metadata=self.metadata,
             )
-            if result.success:
-                self._already_sent = True
-                return True
+            # Note: do NOT set _already_sent = True here.
+            # Commentary messages are interim status updates (e.g. "Using browser
+            # tool..."), not the final response. Setting already_sent would cause
+            # the final response to be incorrectly suppressed when there are
+            # multiple tool calls. See: https://github.com/NousResearch/hermes-agent/issues/10454
+            return result.success
         except Exception as e:
             logger.error("Commentary send error: %s", e)
-        return False
+            return False
 
-    async def _send_or_edit(self, text: str) -> bool:
+    async def _send_or_edit(self, text: str, *, finalize: bool = False) -> bool:
         """Send or edit the streaming message.
 
         Returns True if the text was successfully delivered (sent or edited),
         False otherwise.  Callers like the overflow split loop use this to
         decide whether to advance past the delivered chunk.
+
+        ``finalize`` is True when this is the last edit in a streaming
+        sequence.
         """
         # Strip MEDIA: directives so they don't appear as visible text.
         # Media files are delivered as native attachments after the stream
@@ -655,14 +776,22 @@ class GatewayStreamConsumer:
         try:
             if self._message_id is not None:
                 if self._edit_supported:
-                    # Skip if text is identical to what we last sent
-                    if text == self._last_sent_text:
+                    # Skip if text is identical to what we last sent.
+                    # Exception: adapters that require an explicit finalize
+                    # call (REQUIRES_EDIT_FINALIZE) must still receive the
+                    # finalize=True edit even when content is unchanged, so
+                    # their streaming UI can transition out of the in-
+                    # progress state.  Everyone else short-circuits.
+                    if text == self._last_sent_text and not (
+                        finalize and self._adapter_requires_finalize
+                    ):
                         return True
                     # Edit existing message
                     result = await self.adapter.edit_message(
                         chat_id=self.chat_id,
                         message_id=self._message_id,
                         content=text,
+                        finalize=finalize,
                     )
                     if result.success:
                         self._already_sent = True

hermes_cli/__init__.py

@@ -11,5 +11,5 @@ Provides subcommands for:
 - hermes cron          - Manage cron jobs
 """
 
-__version__ = "0.9.0"
-__release_date__ = "2026.4.13"
+__version__ = "0.10.0"
+__release_date__ = "2026.4.16"

hermes_cli/auth.py

@@ -20,6 +20,7 @@ import logging
 import os
 import shutil
 import shlex
+import ssl
 import stat
 import base64
 import hashlib
@@ -70,6 +71,7 @@ DEFAULT_CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex"
 DEFAULT_QWEN_BASE_URL = "https://portal.qwen.ai/v1"
 DEFAULT_GITHUB_MODELS_BASE_URL = "https://api.githubcopilot.com"
 DEFAULT_COPILOT_ACP_BASE_URL = "acp://copilot"
+DEFAULT_OLLAMA_CLOUD_BASE_URL = "https://ollama.com/v1"
 CODEX_OAUTH_CLIENT_ID = "app_EMoamEEZ73f0CkXaXp7hrann"
 CODEX_OAUTH_TOKEN_URL = "https://auth.openai.com/oauth/token"
 CODEX_ACCESS_TOKEN_REFRESH_SKEW_SECONDS = 120
@@ -77,6 +79,10 @@ QWEN_OAUTH_CLIENT_ID = "f0304373b74a44d2b584a3fb70ca9e56"
 QWEN_OAUTH_TOKEN_URL = "https://chat.qwen.ai/api/v1/oauth2/token"
 QWEN_ACCESS_TOKEN_REFRESH_SKEW_SECONDS = 120
 
+# Google Gemini OAuth (google-gemini-cli provider, Cloud Code Assist backend)
+DEFAULT_GEMINI_CLOUDCODE_BASE_URL = "cloudcode-pa://google"
+GEMINI_OAUTH_ACCESS_TOKEN_REFRESH_SKEW_SECONDS = 60  # refresh 60s before expiry
+
 
 # =============================================================================
 # Provider Registry
@@ -121,6 +127,12 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
         auth_type="oauth_external",
         inference_base_url=DEFAULT_QWEN_BASE_URL,
     ),
+    "google-gemini-cli": ProviderConfig(
+        id="google-gemini-cli",
+        name="Google Gemini (OAuth)",
+        auth_type="oauth_external",
+        inference_base_url=DEFAULT_GEMINI_CLOUDCODE_BASE_URL,
+    ),
     "copilot": ProviderConfig(
         id="copilot",
         name="GitHub Copilot",
@@ -140,7 +152,7 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
         id="gemini",
         name="Google AI Studio",
         auth_type="api_key",
-        inference_base_url="https://generativelanguage.googleapis.com/v1beta/openai",
+        inference_base_url="https://generativelanguage.googleapis.com/v1beta",
         api_key_env_vars=("GOOGLE_API_KEY", "GEMINI_API_KEY"),
         base_url_env_var="GEMINI_BASE_URL",
     ),
@@ -222,6 +234,14 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
         api_key_env_vars=("XAI_API_KEY",),
         base_url_env_var="XAI_BASE_URL",
     ),
+    "nvidia": ProviderConfig(
+        id="nvidia",
+        name="NVIDIA NIM",
+        auth_type="api_key",
+        inference_base_url="https://integrate.api.nvidia.com/v1",
+        api_key_env_vars=("NVIDIA_API_KEY",),
+        base_url_env_var="NVIDIA_BASE_URL",
+    ),
     "ai-gateway": ProviderConfig(
         id="ai-gateway",
         name="Vercel AI Gateway",
@@ -274,6 +294,22 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
         api_key_env_vars=("XIAOMI_API_KEY",),
         base_url_env_var="XIAOMI_BASE_URL",
     ),
+    "ollama-cloud": ProviderConfig(
+        id="ollama-cloud",
+        name="Ollama Cloud",
+        auth_type="api_key",
+        inference_base_url=DEFAULT_OLLAMA_CLOUD_BASE_URL,
+        api_key_env_vars=("OLLAMA_API_KEY",),
+        base_url_env_var="OLLAMA_BASE_URL",
+    ),
+    "bedrock": ProviderConfig(
+        id="bedrock",
+        name="AWS Bedrock",
+        auth_type="aws_sdk",
+        inference_base_url="https://bedrock-runtime.us-east-1.amazonaws.com",
+        api_key_env_vars=(),
+        base_url_env_var="BEDROCK_BASE_URL",
+    ),
 }
 
 
@@ -318,6 +354,9 @@ def _resolve_kimi_base_url(api_key: str, default_url: str, env_override: str) ->
     """
     if env_override:
         return env_override
+    # No key → nothing to infer from.  Return default without inspecting.
+    if not api_key:
+        return default_url
     if api_key.startswith("sk-kimi-"):
         return KIMI_CODE_BASE_URL
     return default_url
@@ -445,6 +484,14 @@ def _resolve_zai_base_url(api_key: str, default_url: str, env_override: str) ->
     if env_override:
         return env_override
 
+    # No API key set → don't probe (would fire N×M HTTPS requests with an
+    # empty Bearer token, all returning 401).  This path is hit during
+    # auxiliary-client auto-detection when the user has no Z.AI credentials
+    # at all — the caller discards the result immediately, so the probe is
+    # pure latency for every AIAgent construction.
+    if not api_key:
+        return default_url
+
     # Check provider-state cache for a previously-detected endpoint.
     auth_store = _load_auth_store()
     state = _load_provider_state(auth_store, "zai") or {}
@@ -746,6 +793,28 @@ def is_source_suppressed(provider_id: str, source: str) -> bool:
         return False
 
 
+def unsuppress_credential_source(provider_id: str, source: str) -> bool:
+    """Clear a suppression marker so the source will be re-seeded on the next load.
+
+    Returns True if a marker was cleared, False if no marker existed.
+    """
+    with _auth_store_lock():
+        auth_store = _load_auth_store()
+        suppressed = auth_store.get("suppressed_sources")
+        if not isinstance(suppressed, dict):
+            return False
+        provider_list = suppressed.get(provider_id)
+        if not isinstance(provider_list, list) or source not in provider_list:
+            return False
+        provider_list.remove(source)
+        if not provider_list:
+            suppressed.pop(provider_id, None)
+        if not suppressed:
+            auth_store.pop("suppressed_sources", None)
+        _save_auth_store(auth_store)
+        return True
+
+
 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()
@@ -911,6 +980,7 @@ def resolve_provider(
     _PROVIDER_ALIASES = {
         "glm": "zai", "z-ai": "zai", "z.ai": "zai", "zhipu": "zai",
         "google": "gemini", "google-gemini": "gemini", "google-ai-studio": "gemini",
+        "x-ai": "xai", "x.ai": "xai", "grok": "xai",
         "kimi": "kimi-coding", "kimi-for-coding": "kimi-coding", "moonshot": "kimi-coding",
         "kimi-cn": "kimi-coding-cn", "moonshot-cn": "kimi-coding-cn",
         "arcee-ai": "arcee", "arceeai": "arcee",
@@ -921,14 +991,16 @@ 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",
-        "qwen-portal": "qwen-oauth", "qwen-cli": "qwen-oauth", "qwen-oauth": "qwen-oauth",
+        "qwen-portal": "qwen-oauth", "qwen-cli": "qwen-oauth", "qwen-oauth": "qwen-oauth", "google-gemini-cli": "google-gemini-cli", "gemini-cli": "google-gemini-cli", "gemini-oauth": "google-gemini-cli",
         "hf": "huggingface", "hugging-face": "huggingface", "huggingface-hub": "huggingface",
         "mimo": "xiaomi", "xiaomi-mimo": "xiaomi",
+        "aws": "bedrock", "aws-bedrock": "bedrock", "amazon-bedrock": "bedrock", "amazon": "bedrock",
         "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",
+        "ollama": "custom", "ollama_cloud": "ollama-cloud",
+        "vllm": "custom", "llamacpp": "custom",
         "llama.cpp": "custom", "llama-cpp": "custom",
     }
     normalized = _PROVIDER_ALIASES.get(normalized, normalized)
@@ -980,6 +1052,15 @@ def resolve_provider(
             if has_usable_secret(os.getenv(env_var, "")):
                 return pid
 
+    # AWS Bedrock — detect via boto3 credential chain (IAM roles, SSO, env vars).
+    # This runs after API-key providers so explicit keys always win.
+    try:
+        from agent.bedrock_adapter import has_aws_credentials
+        if has_aws_credentials():
+            return "bedrock"
+    except ImportError:
+        pass  # boto3 not installed — skip Bedrock auto-detection
+
     raise AuthError(
         "No inference provider configured. Run 'hermes model' to choose a "
         "provider and model, or set an API key (OPENROUTER_API_KEY, "
@@ -1222,6 +1303,83 @@ def get_qwen_auth_status() -> Dict[str, Any]:
         }
 
 
+# =============================================================================
+# Google Gemini OAuth (google-gemini-cli) — PKCE flow + Cloud Code Assist.
+#
+# Tokens live in ~/.hermes/auth/google_oauth.json (managed by agent.google_oauth).
+# The `base_url` here is the marker "cloudcode-pa://google" that run_agent.py
+# uses to construct a GeminiCloudCodeClient instead of the default OpenAI SDK.
+# Actual HTTP traffic goes to https://cloudcode-pa.googleapis.com/v1internal:*.
+# =============================================================================
+
+def resolve_gemini_oauth_runtime_credentials(
+    *,
+    force_refresh: bool = False,
+) -> Dict[str, Any]:
+    """Resolve runtime OAuth creds for google-gemini-cli."""
+    try:
+        from agent.google_oauth import (
+            GoogleOAuthError,
+            _credentials_path,
+            get_valid_access_token,
+            load_credentials,
+        )
+    except ImportError as exc:
+        raise AuthError(
+            f"agent.google_oauth is not importable: {exc}",
+            provider="google-gemini-cli",
+            code="google_oauth_module_missing",
+        ) from exc
+
+    try:
+        access_token = get_valid_access_token(force_refresh=force_refresh)
+    except GoogleOAuthError as exc:
+        raise AuthError(
+            str(exc),
+            provider="google-gemini-cli",
+            code=exc.code,
+        ) from exc
+
+    creds = load_credentials()
+    base_url = DEFAULT_GEMINI_CLOUDCODE_BASE_URL
+    return {
+        "provider": "google-gemini-cli",
+        "base_url": base_url,
+        "api_key": access_token,
+        "source": "google-oauth",
+        "expires_at_ms": (creds.expires_ms if creds else None),
+        "auth_file": str(_credentials_path()),
+        "email": (creds.email if creds else "") or "",
+        "project_id": (creds.project_id if creds else "") or "",
+    }
+
+
+def get_gemini_oauth_auth_status() -> Dict[str, Any]:
+    """Return a status dict for `hermes auth list` / `hermes status`."""
+    try:
+        from agent.google_oauth import _credentials_path, load_credentials
+    except ImportError:
+        return {"logged_in": False, "error": "agent.google_oauth unavailable"}
+    auth_path = _credentials_path()
+    creds = load_credentials()
+    if creds is None or not creds.access_token:
+        return {
+            "logged_in": False,
+            "auth_file": str(auth_path),
+            "error": "not logged in",
+        }
+    return {
+        "logged_in": True,
+        "auth_file": str(auth_path),
+        "source": "google-oauth",
+        "api_key": creds.access_token,
+        "expires_at_ms": creds.expires_ms,
+        "email": creds.email,
+        "project_id": creds.project_id,
+    }
+
+
+
 # =============================================================================
 # SSH / remote session detection
 # =============================================================================
@@ -1288,49 +1446,6 @@ def _read_codex_tokens(*, _lock: bool = True) -> Dict[str, Any]:
     }
 
 
-def _write_codex_cli_tokens(
-    access_token: str,
-    refresh_token: str,
-    *,
-    last_refresh: Optional[str] = None,
-) -> None:
-    """Write refreshed tokens back to ~/.codex/auth.json.
-
-    OpenAI OAuth refresh tokens are single-use and rotate on every refresh.
-    When Hermes refreshes a token it consumes the old refresh_token; if we
-    don't write the new pair back, the Codex CLI (or VS Code extension) will
-    fail with ``refresh_token_reused`` on its next refresh attempt.
-
-    This mirrors the Anthropic write-back to ~/.claude/.credentials.json
-    via ``_write_claude_code_credentials()``.
-    """
-    codex_home = os.getenv("CODEX_HOME", "").strip()
-    if not codex_home:
-        codex_home = str(Path.home() / ".codex")
-    auth_path = Path(codex_home).expanduser() / "auth.json"
-    try:
-        existing: Dict[str, Any] = {}
-        if auth_path.is_file():
-            existing = json.loads(auth_path.read_text(encoding="utf-8"))
-        if not isinstance(existing, dict):
-            existing = {}
-
-        tokens_dict = existing.get("tokens")
-        if not isinstance(tokens_dict, dict):
-            tokens_dict = {}
-        tokens_dict["access_token"] = access_token
-        tokens_dict["refresh_token"] = refresh_token
-        existing["tokens"] = tokens_dict
-        if last_refresh is not None:
-            existing["last_refresh"] = last_refresh
-
-        auth_path.parent.mkdir(parents=True, exist_ok=True)
-        auth_path.write_text(json.dumps(existing, indent=2), encoding="utf-8")
-        auth_path.chmod(0o600)
-    except (OSError, IOError) as exc:
-        logger.debug("Failed to write refreshed tokens to %s: %s", auth_path, exc)
-
-
 def _save_codex_tokens(tokens: Dict[str, str], last_refresh: str = None) -> None:
     """Save Codex OAuth tokens to Hermes auth store (~/.hermes/auth.json)."""
     if last_refresh is None:
@@ -1398,6 +1513,11 @@ def refresh_codex_oauth_pure(
                 "then run `hermes auth` to re-authenticate."
             )
             relogin_required = True
+        # A 401/403 from the token endpoint always means the refresh token
+        # is invalid/expired — force relogin even if the body error code
+        # wasn't one of the known strings above.
+        if response.status_code in (401, 403) and not relogin_required:
+            relogin_required = True
         raise AuthError(
             message,
             provider="openai-codex",
@@ -1453,12 +1573,6 @@ def _refresh_codex_auth_tokens(
     updated_tokens["refresh_token"] = refreshed["refresh_token"]
 
     _save_codex_tokens(updated_tokens)
-    # Write back to ~/.codex/auth.json so Codex CLI / VS Code stay in sync.
-    _write_codex_cli_tokens(
-        refreshed["access_token"],
-        refreshed["refresh_token"],
-        last_refresh=refreshed.get("last_refresh"),
-    )
     return updated_tokens
 
 
@@ -1503,25 +1617,7 @@ def resolve_codex_runtime_credentials(
     refresh_skew_seconds: int = CODEX_ACCESS_TOKEN_REFRESH_SKEW_SECONDS,
 ) -> Dict[str, Any]:
     """Resolve runtime credentials from Hermes's own Codex token store."""
-    try:
-        data = _read_codex_tokens()
-    except AuthError as orig_err:
-        # Only attempt migration when there are NO tokens stored at all
-        # (code == "codex_auth_missing"), not when tokens exist but are invalid.
-        if orig_err.code != "codex_auth_missing":
-            raise
-
-        # Migration: user had Codex as active provider with old storage (~/.codex/).
-        cli_tokens = _import_codex_cli_tokens()
-        if cli_tokens:
-            logger.info("Migrating Codex credentials from ~/.codex/ to Hermes auth store")
-            print("⚠️  Migrating Codex credentials to Hermes's own auth store.")
-            print("   This avoids conflicts with Codex CLI and VS Code.")
-            print("   Run `hermes auth` to create a fully independent session.\n")
-            _save_codex_tokens(cli_tokens)
-            data = _read_codex_tokens()
-        else:
-            raise
+    data = _read_codex_tokens()
     tokens = dict(data["tokens"])
     access_token = str(tokens.get("access_token", "") or "").strip()
     refresh_timeout_seconds = float(os.getenv("HERMES_CODEX_REFRESH_TIMEOUT_SECONDS", "20"))
@@ -1568,7 +1664,7 @@ def _resolve_verify(
     insecure: Optional[bool] = None,
     ca_bundle: Optional[str] = None,
     auth_state: Optional[Dict[str, Any]] = None,
-) -> bool | str:
+) -> bool | ssl.SSLContext:
     tls_state = auth_state.get("tls") if isinstance(auth_state, dict) else {}
     tls_state = tls_state if isinstance(tls_state, dict) else {}
 
@@ -1588,13 +1684,12 @@ def _resolve_verify(
     if effective_ca:
         ca_path = str(effective_ca)
         if not os.path.isfile(ca_path):
-            import logging
-            logging.getLogger("hermes.auth").warning(
+            logger.warning(
                 "CA bundle path does not exist: %s — falling back to default certificates",
                 ca_path,
             )
             return True
-        return ca_path
+        return ssl.create_default_context(cafile=ca_path)
     return True
 
 
@@ -2013,6 +2108,62 @@ def refresh_nous_oauth_from_state(
     )
 
 
+NOUS_DEVICE_CODE_SOURCE = "device_code"
+
+
+def persist_nous_credentials(
+    creds: Dict[str, Any],
+    *,
+    label: Optional[str] = None,
+):
+    """Persist minted Nous OAuth credentials as the singleton provider state
+    and ensure the credential pool is in sync.
+
+    Nous credentials are read at runtime from two independent locations:
+
+    - ``providers.nous``: singleton state read by
+      ``resolve_nous_runtime_credentials()`` during 401 recovery and by
+      ``_seed_from_singletons()`` during pool load.
+    - ``credential_pool.nous``: used by the runtime ``pool.select()`` path.
+
+    Historically ``hermes auth add nous`` wrote a ``manual:device_code`` pool
+    entry only, skipping ``providers.nous``.  When the 24h agent_key TTL
+    expired, the recovery path read the empty singleton state and raised
+    ``AuthError`` silently (``logger.debug`` at INFO level).
+
+    This helper writes ``providers.nous`` then calls ``load_pool("nous")`` so
+    ``_seed_from_singletons`` materialises the canonical ``device_code`` pool
+    entry from the singleton.  Re-running login upserts the same entry in
+    place; the pool never accumulates duplicate device_code rows.
+
+    ``label`` is an optional user-chosen display name (from
+    ``hermes auth add nous --label <name>``).  It gets embedded in the
+    singleton state so that ``_seed_from_singletons`` uses it as the pool
+    entry's label on every subsequent ``load_pool("nous")`` instead of the
+    auto-derived token fingerprint.  When ``None``, the auto-derived label
+    via ``label_from_token`` is used (unchanged default behaviour).
+
+    Returns the upserted :class:`PooledCredential` entry (or ``None`` if
+    seeding somehow produced no match — shouldn't happen).
+    """
+    from agent.credential_pool import load_pool
+
+    state = dict(creds)
+    if label and str(label).strip():
+        state["label"] = str(label).strip()
+
+    with _auth_store_lock():
+        auth_store = _load_auth_store()
+        _save_provider_state(auth_store, "nous", state)
+        _save_auth_store(auth_store)
+
+    pool = load_pool("nous")
+    return next(
+        (e for e in pool.entries() if e.source == NOUS_DEVICE_CODE_SOURCE),
+        None,
+    )
+
+
 def resolve_nous_runtime_credentials(
     *,
     min_key_ttl_seconds: int = DEFAULT_AGENT_KEY_MIN_TTL_SECONDS,
@@ -2384,7 +2535,7 @@ def get_api_key_provider_status(provider_id: str) -> Dict[str, Any]:
     if pconfig.base_url_env_var:
         env_url = os.getenv(pconfig.base_url_env_var, "").strip()
 
-    if provider_id == "kimi-coding":
+    if provider_id in ("kimi-coding", "kimi-coding-cn"):
         base_url = _resolve_kimi_base_url(api_key, pconfig.inference_base_url, env_url)
     elif env_url:
         base_url = env_url
@@ -2440,12 +2591,21 @@ def get_auth_status(provider_id: Optional[str] = None) -> Dict[str, Any]:
         return get_codex_auth_status()
     if target == "qwen-oauth":
         return get_qwen_auth_status()
+    if target == "google-gemini-cli":
+        return get_gemini_oauth_auth_status()
     if target == "copilot-acp":
         return get_external_process_provider_status(target)
     # API-key providers
     pconfig = PROVIDER_REGISTRY.get(target)
     if pconfig and pconfig.auth_type == "api_key":
         return get_api_key_provider_status(target)
+    # AWS SDK providers (Bedrock) — check via boto3 credential chain
+    if pconfig and pconfig.auth_type == "aws_sdk":
+        try:
+            from agent.bedrock_adapter import has_aws_credentials
+            return {"logged_in": has_aws_credentials(), "provider": target}
+        except ImportError:
+            return {"logged_in": False, "provider": target, "error": "boto3 not installed"}
     return {"logged_in": False}
 
 
@@ -2470,7 +2630,7 @@ def resolve_api_key_provider_credentials(provider_id: str) -> Dict[str, Any]:
     if pconfig.base_url_env_var:
         env_url = os.getenv(pconfig.base_url_env_var, "").strip()
 
-    if provider_id == "kimi-coding":
+    if provider_id in ("kimi-coding", "kimi-coding-cn"):
         base_url = _resolve_kimi_base_url(api_key, pconfig.inference_base_url, env_url)
     elif provider_id == "zai":
         base_url = _resolve_zai_base_url(api_key, pconfig.inference_base_url, env_url)
@@ -2572,6 +2732,17 @@ def _update_config_for_provider(
         # Clear stale base_url to prevent contamination when switching providers
         model_cfg.pop("base_url", None)
 
+    # Clear stale api_key/api_mode left over from a previous custom provider.
+    # When the user switches from e.g. a MiniMax custom endpoint
+    # (api_mode=anthropic_messages, api_key=mxp-...) to a built-in provider
+    # (e.g. OpenRouter), the stale api_key/api_mode would override the new
+    # provider's credentials and transport choice.  Built-in providers that
+    # need a specific api_mode (copilot, xai) set it at request-resolution
+    # time via `_copilot_runtime_api_mode` / `_detect_api_mode_for_url`, so
+    # removing the persisted value here is safe.
+    model_cfg.pop("api_key", None)
+    model_cfg.pop("api_mode", None)
+
     # When switching to a non-OpenRouter provider, ensure model.default is
     # valid for the new provider.  An OpenRouter-formatted name like
     # "anthropic/claude-opus-4.6" will fail on direct-API providers.
@@ -3172,6 +3343,14 @@ def _login_nous(args, pconfig: ProviderConfig) -> None:
 
         inference_base_url = auth_state["inference_base_url"]
 
+        # Snapshot the prior active_provider BEFORE _save_provider_state
+        # overwrites it to "nous".  If the user picks "Skip (keep current)"
+        # during model selection below, we restore this so the user's previous
+        # provider (e.g. openrouter) is preserved.
+        with _auth_store_lock():
+            _prior_store = _load_auth_store()
+            prior_active_provider = _prior_store.get("active_provider")
+
         with _auth_store_lock():
             auth_store = _load_auth_store()
             _save_provider_state(auth_store, "nous", auth_state)
@@ -3231,6 +3410,27 @@ def _login_nous(args, pconfig: ProviderConfig) -> None:
             print(f"Login succeeded, but could not fetch available models. Reason: {message}")
 
         # Write provider + model atomically so config is never mismatched.
+        # If no model was selected (user picked "Skip (keep current)",
+        # model list fetch failed, or no curated models were available),
+        # preserve the user's previous provider — don't silently switch
+        # them to Nous with a mismatched model.  The Nous OAuth tokens
+        # stay saved for future use.
+        if not selected_model:
+            # Restore the prior active_provider that _save_provider_state
+            # overwrote to "nous".  config.yaml model.provider is left
+            # untouched, so the user's previous provider is fully preserved.
+            with _auth_store_lock():
+                auth_store = _load_auth_store()
+                if prior_active_provider:
+                    auth_store["active_provider"] = prior_active_provider
+                else:
+                    auth_store.pop("active_provider", None)
+                _save_auth_store(auth_store)
+            print()
+            print("No provider change. Nous credentials saved for future use.")
+            print("  Run `hermes model` again to switch to Nous Portal.")
+            return
+
         config_path = _update_config_for_provider(
             "nous", inference_base_url, default_model=selected_model,
         )

hermes_cli/auth_commands.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 from getpass import getpass
 import math
+import sys
 import time
 from types import SimpleNamespace
 import uuid
@@ -32,7 +33,7 @@ from hermes_constants import OPENROUTER_BASE_URL
 
 
 # Providers that support OAuth login in addition to API keys.
-_OAUTH_CAPABLE_PROVIDERS = {"anthropic", "nous", "openai-codex", "qwen-oauth"}
+_OAUTH_CAPABLE_PROVIDERS = {"anthropic", "nous", "openai-codex", "qwen-oauth", "google-gemini-cli"}
 
 
 def _get_custom_provider_names() -> list:
@@ -147,7 +148,7 @@ def auth_add_command(args) -> None:
         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", "qwen-oauth"} else AUTH_TYPE_API_KEY
+            requested_type = AUTH_TYPE_OAUTH if provider in {"anthropic", "nous", "openai-codex", "qwen-oauth", "google-gemini-cli"} else AUTH_TYPE_API_KEY
 
     pool = load_pool(provider)
 
@@ -160,7 +161,10 @@ def auth_add_command(args) -> None:
         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
+            if sys.stdin.isatty():
+                label = input(f"Label (optional, default: {default_label}): ").strip() or default_label
+            else:
+                label = default_label
         entry = PooledCredential(
             provider=provider,
             id=uuid.uuid4().hex[:6],
@@ -213,22 +217,21 @@ def auth_add_command(args) -> None:
             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),
+        # Honor `--label <name>` so nous matches other providers' UX.  The
+        # helper embeds this into providers.nous so that label_from_token
+        # doesn't overwrite it on every subsequent load_pool("nous").
+        custom_label = (getattr(args, "label", None) or "").strip() or None
+        entry = auth_mod.persist_nous_credentials(creds, label=custom_label)
+        shown_label = entry.label if entry is not None else label_from_token(
+            creds.get("access_token", ""), _oauth_default_label(provider, 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}"')
+        print(f'Saved {provider} OAuth device-code credentials: "{shown_label}"')
         return
 
     if provider == "openai-codex":
+        # Clear any existing suppression marker so a re-link after `hermes auth
+        # remove openai-codex` works without the new tokens being skipped.
+        auth_mod.unsuppress_credential_source(provider, "device_code")
         creds = auth_mod._codex_device_code_login()
         label = (getattr(args, "label", None) or "").strip() or label_from_token(
             creds["tokens"]["access_token"],
@@ -250,6 +253,27 @@ def auth_add_command(args) -> None:
         print(f'Added {provider} OAuth credential #{len(pool.entries())}: "{entry.label}"')
         return
 
+    if provider == "google-gemini-cli":
+        from agent.google_oauth import run_gemini_oauth_login_pure
+
+        creds = run_gemini_oauth_login_pure()
+        label = (getattr(args, "label", None) or "").strip() or (
+            creds.get("email") or _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}:google_pkce",
+            access_token=creds["access_token"],
+            refresh_token=creds.get("refresh_token"),
+        )
+        pool.add_entry(entry)
+        print(f'Added {provider} OAuth credential #{len(pool.entries())}: "{entry.label}"')
+        return
+
     if provider == "qwen-oauth":
         creds = auth_mod.resolve_qwen_runtime_credentials(refresh_if_expiring=False)
         label = (getattr(args, "label", None) or "").strip() or label_from_token(
@@ -327,7 +351,34 @@ def auth_remove_command(args) -> None:
     # If this was a singleton-seeded credential (OAuth device_code, hermes_pkce),
     # clear the underlying auth store / credential file so it doesn't get
     # re-seeded on the next load_pool() call.
-    elif removed.source == "device_code" and provider in ("openai-codex", "nous"):
+    elif provider == "openai-codex" and (
+        removed.source == "device_code" or removed.source.endswith(":device_code")
+    ):
+        # Codex tokens live in TWO places: the Hermes auth store and
+        # ~/.codex/auth.json (the Codex CLI shared file).  On every refresh,
+        # refresh_codex_oauth_pure() writes to both.  So clearing only the
+        # Hermes auth store is not enough — _seed_from_singletons() will
+        # auto-import from ~/.codex/auth.json on the next load_pool() and
+        # the removal is instantly undone.  Mark the source as suppressed
+        # so auto-import is skipped; leave ~/.codex/auth.json untouched so
+        # the Codex CLI itself keeps working.
+        from hermes_cli.auth import (
+            _load_auth_store, _save_auth_store, _auth_store_lock,
+            suppress_credential_source,
+        )
+        with _auth_store_lock():
+            auth_store = _load_auth_store()
+            providers_dict = auth_store.get("providers")
+            if isinstance(providers_dict, dict) and provider in providers_dict:
+                del providers_dict[provider]
+                _save_auth_store(auth_store)
+                print(f"Cleared {provider} OAuth tokens from auth store")
+        suppress_credential_source(provider, "device_code")
+        print("Suppressed openai-codex device_code source — it will not be re-seeded.")
+        print("Note: Codex CLI credentials still live in ~/.codex/auth.json")
+        print("Run `hermes auth add openai-codex` to re-enable if needed.")
+
+    elif removed.source == "device_code" and provider == "nous":
         from hermes_cli.auth import (
             _load_auth_store, _save_auth_store, _auth_store_lock,
         )
@@ -368,6 +419,27 @@ def _interactive_auth() -> None:
     print("=" * 50)
 
     auth_list_command(SimpleNamespace(provider=None))
+
+    # Show AWS Bedrock credential status (not in the pool — uses boto3 chain)
+    try:
+        from agent.bedrock_adapter import has_aws_credentials, resolve_aws_auth_env_var, resolve_bedrock_region
+        if has_aws_credentials():
+            auth_source = resolve_aws_auth_env_var() or "unknown"
+            region = resolve_bedrock_region()
+            print(f"bedrock (AWS SDK credential chain):")
+            print(f"  Auth: {auth_source}")
+            print(f"  Region: {region}")
+            try:
+                import boto3
+                sts = boto3.client("sts", region_name=region)
+                identity = sts.get_caller_identity()
+                arn = identity.get("Arn", "unknown")
+                print(f"  Identity: {arn}")
+            except Exception:
+                print(f"  Identity: (could not resolve — boto3 STS call failed)")
+            print()
+    except ImportError:
+        pass  # boto3 or bedrock_adapter not available
     print()
 
     # Main menu

hermes_cli/backup.py

@@ -201,7 +201,7 @@ def run_backup(args) -> None:
                 else:
                     zf.write(abs_path, arcname=str(rel_path))
                     total_bytes += abs_path.stat().st_size
-            except (PermissionError, OSError) as exc:
+            except (PermissionError, OSError, ValueError) as exc:
                 errors.append(f"  {rel_path}: {exc}")
                 continue
 

hermes_cli/clipboard.py

@@ -7,8 +7,8 @@ CLI tools that ship with the platform (or are commonly installed).
 
 Platform support:
   macOS   — osascript (always available), pngpaste (if installed)
-  Windows — PowerShell via .NET System.Windows.Forms.Clipboard
-  WSL2    — powershell.exe via .NET System.Windows.Forms.Clipboard
+  Windows — PowerShell via WinForms, Get-Clipboard, file-drop fallback
+  WSL2    — powershell.exe via WinForms, Get-Clipboard, file-drop fallback
   Linux   — wl-paste (Wayland), xclip (X11)
 """
 
@@ -46,10 +46,11 @@ def has_clipboard_image() -> bool:
         return _macos_has_image()
     if sys.platform == "win32":
         return _windows_has_image()
-    if _is_wsl():
-        return _wsl_has_image()
-    if os.environ.get("WAYLAND_DISPLAY"):
-        return _wayland_has_image()
+    # Match _linux_save fallthrough order: WSL → Wayland → X11
+    if _is_wsl() and _wsl_has_image():
+        return True
+    if os.environ.get("WAYLAND_DISPLAY") and _wayland_has_image():
+        return True
     return _xclip_has_image()
 
 
@@ -135,6 +136,114 @@ _PS_EXTRACT_IMAGE = (
     "[System.Convert]::ToBase64String($ms.ToArray())"
 )
 
+_PS_CHECK_IMAGE_GET_CLIPBOARD = (
+    "try { "
+    "$img = Get-Clipboard -Format Image -ErrorAction Stop;"
+    "if ($null -ne $img) { 'True' } else { 'False' }"
+    "} catch { 'False' }"
+)
+
+_PS_EXTRACT_IMAGE_GET_CLIPBOARD = (
+    "try { "
+    "Add-Type -AssemblyName System.Drawing;"
+    "Add-Type -AssemblyName PresentationCore;"
+    "Add-Type -AssemblyName WindowsBase;"
+    "$img = Get-Clipboard -Format Image -ErrorAction Stop;"
+    "if ($null -eq $img) { exit 1 }"
+    "$ms = New-Object System.IO.MemoryStream;"
+    "if ($img -is [System.Drawing.Image]) {"
+    "$img.Save($ms, [System.Drawing.Imaging.ImageFormat]::Png)"
+    "} elseif ($img -is [System.Windows.Media.Imaging.BitmapSource]) {"
+    "$enc = New-Object System.Windows.Media.Imaging.PngBitmapEncoder;"
+    "$enc.Frames.Add([System.Windows.Media.Imaging.BitmapFrame]::Create($img));"
+    "$enc.Save($ms)"
+    "} else { exit 2 }"
+    "[System.Convert]::ToBase64String($ms.ToArray())"
+    "} catch { exit 1 }"
+)
+
+_FILEDROP_IMAGE_EXTS = "'.png','.jpg','.jpeg','.gif','.webp','.bmp','.tiff','.tif'"
+
+_PS_CHECK_FILEDROP_IMAGE = (
+    "try { "
+    "$files = Get-Clipboard -Format FileDropList -ErrorAction Stop;"
+    f"$exts = @({_FILEDROP_IMAGE_EXTS});"
+    "$hit = $files | Where-Object { $exts -contains ([System.IO.Path]::GetExtension($_).ToLowerInvariant()) } | Select-Object -First 1;"
+    "if ($null -ne $hit) { 'True' } else { 'False' }"
+    "} catch { 'False' }"
+)
+
+_PS_EXTRACT_FILEDROP_IMAGE = (
+    "try { "
+    "$files = Get-Clipboard -Format FileDropList -ErrorAction Stop;"
+    f"$exts = @({_FILEDROP_IMAGE_EXTS});"
+    "$hit = $files | Where-Object { $exts -contains ([System.IO.Path]::GetExtension($_).ToLowerInvariant()) } | Select-Object -First 1;"
+    "if ($null -eq $hit) { exit 1 }"
+    "[System.Convert]::ToBase64String([System.IO.File]::ReadAllBytes($hit))"
+    "} catch { exit 1 }"
+)
+
+_POWERSHELL_HAS_IMAGE_SCRIPTS = (
+    _PS_CHECK_IMAGE,
+    _PS_CHECK_IMAGE_GET_CLIPBOARD,
+    _PS_CHECK_FILEDROP_IMAGE,
+)
+
+_POWERSHELL_EXTRACT_IMAGE_SCRIPTS = (
+    _PS_EXTRACT_IMAGE,
+    _PS_EXTRACT_IMAGE_GET_CLIPBOARD,
+    _PS_EXTRACT_FILEDROP_IMAGE,
+)
+
+
+def _run_powershell(exe: str, script: str, timeout: int) -> subprocess.CompletedProcess:
+    return subprocess.run(
+        [exe, "-NoProfile", "-NonInteractive", "-Command", script],
+        capture_output=True, text=True, timeout=timeout,
+    )
+
+
+def _write_base64_image(dest: Path, b64_data: str) -> bool:
+    image_bytes = base64.b64decode(b64_data, validate=True)
+    dest.write_bytes(image_bytes)
+    return dest.exists() and dest.stat().st_size > 0
+
+
+def _powershell_has_image(exe: str, *, timeout: int, label: str) -> bool:
+    for script in _POWERSHELL_HAS_IMAGE_SCRIPTS:
+        try:
+            r = _run_powershell(exe, script, timeout=timeout)
+            if r.returncode == 0 and "True" in r.stdout:
+                return True
+        except FileNotFoundError:
+            logger.debug("%s not found — clipboard unavailable", exe)
+            return False
+        except Exception as e:
+            logger.debug("%s clipboard image check failed: %s", label, e)
+    return False
+
+
+def _powershell_save_image(exe: str, dest: Path, *, timeout: int, label: str) -> bool:
+    for script in _POWERSHELL_EXTRACT_IMAGE_SCRIPTS:
+        try:
+            r = _run_powershell(exe, script, timeout=timeout)
+            if r.returncode != 0:
+                continue
+
+            b64_data = r.stdout.strip()
+            if not b64_data:
+                continue
+
+            if _write_base64_image(dest, b64_data):
+                return True
+        except FileNotFoundError:
+            logger.debug("%s not found — clipboard unavailable", exe)
+            return False
+        except Exception as e:
+            logger.debug("%s clipboard image extraction failed: %s", label, e)
+            dest.unlink(missing_ok=True)
+    return False
+
 
 # ── Native Windows ────────────────────────────────────────────────────────
 
@@ -175,15 +284,7 @@ def _windows_has_image() -> bool:
     ps = _get_ps_exe()
     if ps is None:
         return False
-    try:
-        r = subprocess.run(
-            [ps, "-NoProfile", "-NonInteractive", "-Command", _PS_CHECK_IMAGE],
-            capture_output=True, text=True, timeout=5,
-        )
-        return r.returncode == 0 and "True" in r.stdout
-    except Exception as e:
-        logger.debug("Windows clipboard image check failed: %s", e)
-    return False
+    return _powershell_has_image(ps, timeout=5, label="Windows")
 
 
 def _windows_save(dest: Path) -> bool:
@@ -192,26 +293,7 @@ def _windows_save(dest: Path) -> bool:
     if ps is None:
         logger.debug("No PowerShell found — Windows clipboard image paste unavailable")
         return False
-    try:
-        r = subprocess.run(
-            [ps, "-NoProfile", "-NonInteractive", "-Command", _PS_EXTRACT_IMAGE],
-            capture_output=True, text=True, timeout=15,
-        )
-        if r.returncode != 0:
-            return False
-
-        b64_data = r.stdout.strip()
-        if not b64_data:
-            return False
-
-        png_bytes = base64.b64decode(b64_data)
-        dest.write_bytes(png_bytes)
-        return dest.exists() and dest.stat().st_size > 0
-
-    except Exception as e:
-        logger.debug("Windows clipboard image extraction failed: %s", e)
-        dest.unlink(missing_ok=True)
-    return False
+    return _powershell_save_image(ps, dest, timeout=15, label="Windows")
 
 
 # ── Linux ────────────────────────────────────────────────────────────────
@@ -235,45 +317,12 @@ def _linux_save(dest: Path) -> bool:
 
 def _wsl_has_image() -> bool:
     """Check if Windows clipboard has an image (via powershell.exe)."""
-    try:
-        r = subprocess.run(
-            ["powershell.exe", "-NoProfile", "-NonInteractive", "-Command",
-             _PS_CHECK_IMAGE],
-            capture_output=True, text=True, timeout=8,
-        )
-        return r.returncode == 0 and "True" in r.stdout
-    except FileNotFoundError:
-        logger.debug("powershell.exe not found — WSL clipboard unavailable")
-    except Exception as e:
-        logger.debug("WSL clipboard check failed: %s", e)
-    return False
+    return _powershell_has_image("powershell.exe", timeout=8, label="WSL")
 
 
 def _wsl_save(dest: Path) -> bool:
     """Extract clipboard image via powershell.exe → base64 → decode to PNG."""
-    try:
-        r = subprocess.run(
-            ["powershell.exe", "-NoProfile", "-NonInteractive", "-Command",
-             _PS_EXTRACT_IMAGE],
-            capture_output=True, text=True, timeout=15,
-        )
-        if r.returncode != 0:
-            return False
-
-        b64_data = r.stdout.strip()
-        if not b64_data:
-            return False
-
-        png_bytes = base64.b64decode(b64_data)
-        dest.write_bytes(png_bytes)
-        return dest.exists() and dest.stat().st_size > 0
-
-    except FileNotFoundError:
-        logger.debug("powershell.exe not found — WSL clipboard unavailable")
-    except Exception as e:
-        logger.debug("WSL clipboard extraction failed: %s", e)
-        dest.unlink(missing_ok=True)
-    return False
+    return _powershell_save_image("powershell.exe", dest, timeout=15, label="WSL")
 
 
 # ── Wayland (wl-paste) ──────────────────────────────────────────────────

hermes_cli/commands.py

@@ -87,8 +87,12 @@ COMMAND_REGISTRY: list[CommandDef] = [
                aliases=("bg",), args_hint="<prompt>"),
     CommandDef("btw", "Ephemeral side question using session context (no tools, not persisted)", "Session",
                args_hint="<question>"),
+    CommandDef("agents", "Show active agents and running tasks", "Session",
+               aliases=("tasks",)),
     CommandDef("queue", "Queue a prompt for the next turn (doesn't interrupt)", "Session",
                aliases=("q",), args_hint="<prompt>"),
+    CommandDef("steer", "Inject a message after the next tool call without interrupting", "Session",
+               args_hint="<prompt>"),
     CommandDef("status", "Show session info", "Session"),
     CommandDef("profile", "Show active profile name and home directory", "Info"),
     CommandDef("sethome", "Set this chat as the home channel", "Session",
@@ -99,9 +103,10 @@ COMMAND_REGISTRY: list[CommandDef] = [
     # Configuration
     CommandDef("config", "Show current configuration", "Configuration",
                cli_only=True),
-    CommandDef("model", "Switch model for this session", "Configuration", args_hint="[model] [--global]"),
+    CommandDef("model", "Switch model for this session", "Configuration", args_hint="[model] [--provider name] [--global]"),
     CommandDef("provider", "Show available providers and current provider",
                "Configuration"),
+    CommandDef("gquota", "Show Google Gemini Code Assist quota usage", "Info"),
 
     CommandDef("personality", "Set a predefined personality", "Configuration",
                args_hint="[name]"),
@@ -119,7 +124,7 @@ COMMAND_REGISTRY: list[CommandDef] = [
                args_hint="[normal|fast|status]",
                subcommands=("normal", "fast", "status", "on", "off")),
     CommandDef("skin", "Show or change the display skin/theme", "Configuration",
-               cli_only=True, args_hint="[name]"),
+               args_hint="[name]"),
     CommandDef("voice", "Toggle voice mode", "Configuration",
                args_hint="[on|off|tts|status]", subcommands=("on", "off", "tts", "status")),
 
@@ -154,7 +159,9 @@ COMMAND_REGISTRY: list[CommandDef] = [
                args_hint="[days]"),
     CommandDef("platforms", "Show gateway/messaging platform status", "Info",
                cli_only=True, aliases=("gateway",)),
-    CommandDef("paste", "Check clipboard for an image and attach it", "Info",
+    CommandDef("copy", "Copy the last assistant response to clipboard", "Info",
+               cli_only=True, args_hint="[number]"),
+    CommandDef("paste", "Attach clipboard image from your clipboard", "Info",
                cli_only=True),
     CommandDef("image", "Attach a local image file for your next prompt", "Info",
                cli_only=True, args_hint="<path>"),
@@ -164,7 +171,7 @@ COMMAND_REGISTRY: list[CommandDef] = [
 
     # Exit
     CommandDef("quit", "Exit the CLI", "Exit",
-               cli_only=True, aliases=("exit", "q")),
+               cli_only=True, aliases=("exit",)),
 ]
 
 
@@ -253,6 +260,53 @@ GATEWAY_KNOWN_COMMANDS: frozenset[str] = frozenset(
 )
 
 
+# Commands with explicit Level-2 running-agent handlers in gateway/run.py.
+# Listed here for introspection / tests; semantically a subset of
+# "all resolvable commands" — which is the real bypass set (see
+# should_bypass_active_session below).
+ACTIVE_SESSION_BYPASS_COMMANDS: frozenset[str] = frozenset(
+    {
+        "agents",
+        "approve",
+        "background",
+        "commands",
+        "deny",
+        "help",
+        "new",
+        "profile",
+        "queue",
+        "restart",
+        "status",
+        "steer",
+        "stop",
+        "update",
+    }
+)
+
+
+def should_bypass_active_session(command_name: str | None) -> bool:
+    """Return True for any resolvable slash command.
+
+    Rationale: every gateway-registered slash command either has a
+    specific Level-2 handler in gateway/run.py (/stop, /new, /model,
+    /approve, etc.) or reaches the running-agent catch-all that returns
+    a "busy — wait or /stop first" response. In both paths the command
+    is dispatched, not queued.
+
+    Queueing is always wrong for a recognized slash command because the
+    safety net in gateway.run discards any command text that reaches
+    the pending queue — which meant a mid-run /model (or /reasoning,
+    /voice, /insights, /title, /resume, /retry, /undo, /compress,
+    /usage, /provider, /reload-mcp, /sethome, /reset) would silently
+    interrupt the agent AND get discarded, producing a zero-char
+    response. See issue #5057 / PRs #6252, #10370, #4665.
+
+    ACTIVE_SESSION_BYPASS_COMMANDS remains the subset of commands with
+    explicit Level-2 handlers; the rest fall through to the catch-all.
+    """
+    return resolve_command(command_name) is not None if command_name else False
+
+
 def _resolve_config_gates() -> set[str]:
     """Return canonical names of commands whose ``gateway_config_gate`` is truthy.
 
@@ -450,7 +504,7 @@ def _collect_gateway_skill_entries(
             name = sanitize_name(cmd_name) if sanitize_name else cmd_name
             if not name:
                 continue
-            desc = "Plugin command"
+            desc = plugin_cmds[cmd_name].get("description", "Plugin command")
             if len(desc) > desc_limit:
                 desc = desc[:desc_limit - 3] + "..."
             plugin_pairs.append((name, desc))
@@ -1043,6 +1097,51 @@ class SlashCommandCompleter(Completer):
                 display_meta=f"{fp}  {meta}" if meta else fp,
             )
 
+    @staticmethod
+    def _skin_completions(sub_text: str, sub_lower: str):
+        """Yield completions for /skin from available skins."""
+        try:
+            from hermes_cli.skin_engine import list_skins
+            for s in list_skins():
+                name = s["name"]
+                if name.startswith(sub_lower) and name != sub_lower:
+                    yield Completion(
+                        name,
+                        start_position=-len(sub_text),
+                        display=name,
+                        display_meta=s.get("description", "") or s.get("source", ""),
+                    )
+        except Exception:
+            pass
+
+    @staticmethod
+    def _personality_completions(sub_text: str, sub_lower: str):
+        """Yield completions for /personality from configured personalities."""
+        try:
+            from hermes_cli.config import load_config
+            personalities = load_config().get("agent", {}).get("personalities", {})
+            if "none".startswith(sub_lower) and "none" != sub_lower:
+                yield Completion(
+                    "none",
+                    start_position=-len(sub_text),
+                    display="none",
+                    display_meta="clear personality overlay",
+                )
+            for name, prompt in personalities.items():
+                if name.startswith(sub_lower) and name != sub_lower:
+                    if isinstance(prompt, dict):
+                        meta = prompt.get("description") or prompt.get("system_prompt", "")[:50]
+                    else:
+                        meta = str(prompt)[:50]
+                    yield Completion(
+                        name,
+                        start_position=-len(sub_text),
+                        display=name,
+                        display_meta=meta,
+                    )
+        except Exception:
+            pass
+
     def _model_completions(self, sub_text: str, sub_lower: str):
         """Yield completions for /model from config aliases + built-in aliases."""
         seen = set()
@@ -1097,10 +1196,17 @@ class SlashCommandCompleter(Completer):
             sub_text = parts[1] if len(parts) > 1 else ""
             sub_lower = sub_text.lower()
 
-            # Dynamic model alias completions for /model
-            if " " not in sub_text and base_cmd == "/model":
-                yield from self._model_completions(sub_text, sub_lower)
-                return
+            # Dynamic completions for commands with runtime lists
+            if " " not in sub_text:
+                if base_cmd == "/model":
+                    yield from self._model_completions(sub_text, sub_lower)
+                    return
+                if base_cmd == "/skin":
+                    yield from self._skin_completions(sub_text, sub_lower)
+                    return
+                if base_cmd == "/personality":
+                    yield from self._personality_completions(sub_text, sub_lower)
+                    return
 
             # Static subcommand completions
             if " " not in sub_text and base_cmd in SUBCOMMANDS and self._command_allowed(base_cmd):
@@ -1139,6 +1245,22 @@ class SlashCommandCompleter(Completer):
                     display_meta=f"⚡ {short_desc}",
                 )
 
+        # Plugin-registered slash commands
+        try:
+            from hermes_cli.plugins import get_plugin_commands
+            for cmd_name, cmd_info in get_plugin_commands().items():
+                if cmd_name.startswith(word):
+                    desc = str(cmd_info.get("description", "Plugin command"))
+                    short_desc = desc[:50] + ("..." if len(desc) > 50 else "")
+                    yield Completion(
+                        self._completion_text(cmd_name, word),
+                        start_position=-len(word),
+                        display=f"/{cmd_name}",
+                        display_meta=f"🔌 {short_desc}",
+                    )
+        except Exception:
+            pass
+
 
 # ---------------------------------------------------------------------------
 # Inline auto-suggest (ghost text) for slash commands

hermes_cli/config.py

@@ -12,6 +12,7 @@ This module provides:
 - hermes config wizard   - Re-run setup wizard
 """
 
+import copy
 import os
 import platform
 import re
@@ -23,10 +24,10 @@ from dataclasses import dataclass
 from pathlib import Path
 from typing import Dict, Any, Optional, List, Tuple
 
-from tools.tool_backend_helpers import managed_nous_tools_enabled as _managed_nous_tools_enabled
 
 _IS_WINDOWS = platform.system() == "Windows"
 _ENV_VAR_NAME_RE = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*$")
+_LAST_EXPANDED_CONFIG_BY_PATH: Dict[str, Any] = {}
 # Env var names written to .env that aren't in OPTIONAL_ENV_VARS
 # (managed by setup/provider flows directly).
 _EXTRA_ENV_KEYS = frozenset({
@@ -45,7 +46,8 @@ _EXTRA_ENV_KEYS = frozenset({
     "WEIXIN_HOME_CHANNEL", "WEIXIN_HOME_CHANNEL_NAME", "WEIXIN_DM_POLICY", "WEIXIN_GROUP_POLICY",
     "WEIXIN_ALLOWED_USERS", "WEIXIN_GROUP_ALLOWED_USERS", "WEIXIN_ALLOW_ALL_USERS",
     "BLUEBUBBLES_SERVER_URL", "BLUEBUBBLES_PASSWORD",
-    "QQ_APP_ID", "QQ_CLIENT_SECRET", "QQ_HOME_CHANNEL", "QQ_HOME_CHANNEL_NAME",
+    "QQ_APP_ID", "QQ_CLIENT_SECRET", "QQBOT_HOME_CHANNEL", "QQBOT_HOME_CHANNEL_NAME",
+    "QQ_HOME_CHANNEL", "QQ_HOME_CHANNEL_NAME",  # legacy aliases (pre-rename, still read for back-compat)
     "QQ_ALLOWED_USERS", "QQ_GROUP_ALLOWED_USERS", "QQ_ALLOW_ALL_USERS", "QQ_MARKDOWN_SUPPORT",
     "QQ_STT_API_KEY", "QQ_STT_BASE_URL", "QQ_STT_MODEL",
     "TERMINAL_ENV", "TERMINAL_SSH_KEY", "TERMINAL_SSH_PORT",
@@ -241,13 +243,41 @@ def _secure_dir(path):
         pass
 
 
+def _is_container() -> bool:
+    """Detect if we're running inside a Docker/Podman/LXC container.
+
+    When Hermes runs in a container with volume-mounted config files, forcing
+    0o600 permissions breaks multi-process setups where the gateway and
+    dashboard run as different UIDs or the volume mount requires broader
+    permissions.
+    """
+    # Explicit opt-out
+    if os.environ.get("HERMES_CONTAINER") or os.environ.get("HERMES_SKIP_CHMOD"):
+        return True
+    # Docker / Podman marker file
+    if os.path.exists("/.dockerenv"):
+        return True
+    # LXC / cgroup-based detection
+    try:
+        with open("/proc/1/cgroup", "r") as f:
+            cgroup_content = f.read()
+        if "docker" in cgroup_content or "lxc" in cgroup_content or "kubepods" in cgroup_content:
+            return True
+    except (OSError, IOError):
+        pass
+    return False
+
+
 def _secure_file(path):
     """Set file to owner-only read/write (0600). No-op on Windows.
 
     Skipped in managed mode — the NixOS activation script sets
     group-readable permissions (0640) on config files.
+
+    Skipped in containers — Docker/Podman volume mounts often need broader
+    permissions.  Set HERMES_SKIP_CHMOD=1 to force-skip on other systems.
     """
-    if is_managed():
+    if is_managed() or _is_container():
         return
     try:
         if os.path.exists(str(path)):
@@ -373,7 +403,11 @@ DEFAULT_CONFIG = {
         "container_persistent": True,   # Persist filesystem across sessions
         # Docker volume mounts — share host directories with the container.
         # Each entry is "host_path:container_path" (standard Docker -v syntax).
-        # Example: ["/home/user/projects:/workspace/projects", "/data:/data"]
+        # Example:
+        # ["/home/user/projects:/workspace/projects",
+        #  "/home/user/.hermes/cache/documents:/output"]
+        # For gateway MEDIA delivery, write inside Docker to /output/... and emit
+        # the host-visible path in MEDIA:, not the container path.
         "docker_volumes": [],
         # Explicit opt-in: mount the host cwd into /workspace for Docker sessions.
         # Default off because passing host directories into a sandbox weakens isolation.
@@ -390,10 +424,10 @@ DEFAULT_CONFIG = {
         "command_timeout": 30,  # Timeout for browser commands in seconds (screenshot, navigate, etc.)
         "record_sessions": False,  # Auto-record browser sessions as WebM videos
         "allow_private_urls": False,  # Allow navigating to private/internal IPs (localhost, 192.168.x.x, etc.)
+        "cdp_url": "",  # Optional persistent CDP endpoint for attaching to an existing Chromium/Chrome
         "camofox": {
             # When true, Hermes sends a stable profile-scoped userId to Camofox
-            # so the server can map it to a persistent browser profile directory.
-            # Requires Camofox server to be configured with CAMOFOX_PROFILE_DIR.
+            # so the server maps it to a persistent Firefox profile automatically.
             # When false (default), each session gets a random userId (ephemeral).
             "managed_persistence": False,
         },
@@ -419,13 +453,27 @@ DEFAULT_CONFIG = {
         "protect_last_n": 20,         # minimum recent messages to keep uncompressed
 
     },
-    "smart_model_routing": {
-        "enabled": False,
-        "max_simple_chars": 160,
-        "max_simple_words": 28,
-        "cheap_model": {},
+
+    # AWS Bedrock provider configuration.
+    # Only used when model.provider is "bedrock".
+    "bedrock": {
+        "region": "",  # AWS region for Bedrock API calls (empty = AWS_REGION env var → us-east-1)
+        "discovery": {
+            "enabled": True,           # Auto-discover models via ListFoundationModels
+            "provider_filter": [],     # Only show models from these providers (e.g. ["anthropic", "amazon"])
+            "refresh_interval": 3600,  # Cache discovery results for this many seconds
+        },
+        "guardrail": {
+            # Amazon Bedrock Guardrails — content filtering and safety policies.
+            # Create a guardrail in the Bedrock console, then set the ID and version here.
+            # See: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html
+            "guardrail_identifier": "",  # e.g. "abc123def456"
+            "guardrail_version": "",     # e.g. "1" or "DRAFT"
+            "stream_processing_mode": "async",  # "sync" or "async"
+            "trace": "disabled",         # "enabled", "disabled", or "enabled_full"
+        },
     },
-    
+
     # Auxiliary model config — provider:model for each side task.
     # Format: provider is the provider name, model is the model slug.
     # "auto" for provider = auto-detect best available provider.
@@ -439,6 +487,7 @@ DEFAULT_CONFIG = {
             "base_url": "",        # direct OpenAI-compatible endpoint (takes precedence over provider)
             "api_key": "",         # API key for base_url (falls back to OPENAI_API_KEY)
             "timeout": 120,        # seconds — LLM API call timeout; vision payloads need generous timeout
+            "extra_body": {},      # OpenAI-compatible provider-specific request fields
             "download_timeout": 30,  # seconds — image HTTP download timeout; increase for slow connections
         },
         "web_extract": {
@@ -447,6 +496,7 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 360,        # seconds (6min) — per-attempt LLM summarization timeout; increase for slow local models
+            "extra_body": {},
         },
         "compression": {
             "provider": "auto",
@@ -454,6 +504,7 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 120,        # seconds — compression summarises large contexts; increase for local models
+            "extra_body": {},
         },
         "session_search": {
             "provider": "auto",
@@ -461,6 +512,8 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 30,
+            "extra_body": {},
+            "max_concurrency": 3,  # Clamp parallel summaries to avoid request-burst 429s on small providers
         },
         "skills_hub": {
             "provider": "auto",
@@ -468,6 +521,7 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 30,
+            "extra_body": {},
         },
         "approval": {
             "provider": "auto",
@@ -475,6 +529,7 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 30,
+            "extra_body": {},
         },
         "mcp": {
             "provider": "auto",
@@ -482,6 +537,7 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 30,
+            "extra_body": {},
         },
         "flush_memories": {
             "provider": "auto",
@@ -489,6 +545,15 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 30,
+            "extra_body": {},
+        },
+        "title_generation": {
+            "provider": "auto",
+            "model": "",
+            "base_url": "",
+            "api_key": "",
+            "timeout": 30,
+            "extra_body": {},
         },
     },
     
@@ -500,9 +565,14 @@ DEFAULT_CONFIG = {
         "bell_on_complete": False,
         "show_reasoning": False,
         "streaming": False,
+        "final_response_markdown": "strip",  # render | strip | raw
         "inline_diffs": True,     # Show inline diff previews for write actions (write_file, patch, skill_manage)
         "show_cost": False,       # Show $ cost in the status bar (off by default)
         "skin": "default",
+        "user_message_preview": {  # CLI: how many submitted user-message lines to echo back in scrollback
+            "first_lines": 2,
+            "last_lines": 2,
+        },
         "interim_assistant_messages": True,  # Gateway: show natural mid-turn assistant status messages
         "tool_progress_command": False,  # Enable /verbose command in messaging gateway
         "tool_progress_overrides": {},  # DEPRECATED — use display.platforms instead
@@ -510,6 +580,11 @@ DEFAULT_CONFIG = {
         "platforms": {},  # Per-platform display overrides: {"telegram": {"tool_progress": "all"}, "slack": {"tool_progress": "off"}}
     },
 
+    # Web dashboard settings
+    "dashboard": {
+        "theme": "default",  # Dashboard visual theme: "default", "midnight", "ember", "mono", "cyberpunk", "rose"
+    },
+
     # Privacy settings
     "privacy": {
         "redact_pii": False,  # When True, hash user IDs and strip phone numbers from LLM context
@@ -517,7 +592,7 @@ DEFAULT_CONFIG = {
     
     # Text-to-speech configuration
     "tts": {
-        "provider": "edge",  # "edge" (free) | "elevenlabs" (premium) | "openai" | "minimax" | "mistral" | "neutts" (local)
+        "provider": "edge",  # "edge" (free) | "elevenlabs" (premium) | "openai" | "xai" | "minimax" | "mistral" | "neutts" (local)
         "edge": {
             "voice": "en-US-AriaNeural",
             # Popular: AriaNeural, JennyNeural, AndrewNeural, BrianNeural, SoniaNeural
@@ -531,6 +606,12 @@ DEFAULT_CONFIG = {
             "voice": "alloy",
             # Voices: alloy, echo, fable, onyx, nova, shimmer
         },
+        "xai": {
+            "voice_id": "eve",
+            "language": "en",
+            "sample_rate": 24000,
+            "bit_rate": 128000,
+        },
         "mistral": {
             "model": "voxtral-mini-tts-2603",
             "voice_id": "c69964a6-ab8b-4f8a-9465-ec0925096ec8",  # Paul - Neutral
@@ -638,6 +719,15 @@ DEFAULT_CONFIG = {
         "allowed_channels": "",        # If set, bot ONLY responds in these channel IDs (whitelist)
         "auto_thread": True,           # Auto-create threads on @mention in channels (like Slack)
         "reactions": True,             # Add 👀/✅/❌ reactions to messages during processing
+        "channel_prompts": {},         # Per-channel ephemeral system prompts (forum parents apply to child threads)
+        # discord_server tool: restrict which actions the agent may call.
+        # Default (empty) = all actions allowed (subject to bot privileged intents).
+        # Accepts comma-separated string ("list_guilds,list_channels,fetch_messages")
+        # or YAML list. Unknown names are dropped with a warning at load time.
+        # Actions: list_guilds, server_info, list_channels, channel_info,
+        # list_roles, member_info, search_members, fetch_messages, list_pins,
+        # pin_message, unpin_message, create_thread, add_role, remove_role.
+        "server_actions": "",
     },
 
     # WhatsApp platform settings (gateway mode)
@@ -648,13 +738,33 @@ DEFAULT_CONFIG = {
         # Supports \n for newlines, e.g. "🤖 *My Bot*\n──────\n"
     },
 
+    # Telegram platform settings (gateway mode)
+    "telegram": {
+        "channel_prompts": {},         # Per-chat/topic ephemeral system prompts (topics inherit from parent group)
+    },
+
+    # Slack platform settings (gateway mode)
+    "slack": {
+        "channel_prompts": {},         # Per-channel ephemeral system prompts
+    },
+
+    # Mattermost platform settings (gateway mode)
+    "mattermost": {
+        "channel_prompts": {},         # Per-channel ephemeral system prompts
+    },
+
     # Approval mode for dangerous commands:
     #   manual — always prompt the user (default)
     #   smart  — use auxiliary LLM to auto-approve low-risk commands, prompt for high-risk
     #   off    — skip all approval prompts (equivalent to --yolo)
+    #
+    # cron_mode — what to do when a cron job hits a dangerous command:
+    #   deny    — block the command and let the agent find another way (default, safe)
+    #   approve — auto-approve all dangerous commands in cron jobs
     "approvals": {
         "mode": "manual",
         "timeout": 60,
+        "cron_mode": "deny",
     },
 
     # Permanently allowed dangerous command patterns (added via "always" approval)
@@ -686,6 +796,20 @@ DEFAULT_CONFIG = {
         "wrap_response": True,
     },
 
+    # execute_code settings — controls the tool used for programmatic tool calls.
+    "code_execution": {
+        # Execution mode:
+        #   project (default) — scripts run in the session's working directory
+        #     with the active virtualenv/conda env's python, so project deps
+        #     (pandas, torch, project packages) and relative paths resolve.
+        #   strict            — scripts run in an isolated temp directory with
+        #     hermes-agent's own python (sys.executable). Maximum isolation
+        #     and reproducibility; project deps and relative paths won't work.
+        # Env scrubbing (strips *_API_KEY, *_TOKEN, *_SECRET, ...) and the
+        # tool whitelist apply identically in both modes.
+        "mode": "project",
+    },
+
     # Logging — controls file logging to ~/.hermes/logs/.
     # agent.log captures INFO+ (all agent activity); errors.log captures WARNING+.
     "logging": {
@@ -703,7 +827,7 @@ DEFAULT_CONFIG = {
     },
 
     # Config schema version - bump this when adding new required fields
-    "_config_version": 17,
+    "_config_version": 20,
 }
 
 # =============================================================================
@@ -771,6 +895,38 @@ OPTIONAL_ENV_VARS = {
         "category": "provider",
         "advanced": True,
     },
+    "XAI_API_KEY": {
+        "description": "xAI API key",
+        "prompt": "xAI API key",
+        "url": "https://console.x.ai/",
+        "password": True,
+        "category": "provider",
+        "advanced": True,
+    },
+    "XAI_BASE_URL": {
+        "description": "xAI base URL override",
+        "prompt": "xAI base URL (leave empty for default)",
+        "url": None,
+        "password": False,
+        "category": "provider",
+        "advanced": True,
+    },
+    "NVIDIA_API_KEY": {
+        "description": "NVIDIA NIM API key (build.nvidia.com or local NIM endpoint)",
+        "prompt": "NVIDIA NIM API key",
+        "url": "https://build.nvidia.com/",
+        "password": True,
+        "category": "provider",
+        "advanced": True,
+    },
+    "NVIDIA_BASE_URL": {
+        "description": "NVIDIA NIM base URL override (e.g. http://localhost:8000/v1 for local NIM)",
+        "prompt": "NVIDIA NIM base URL (leave empty for default)",
+        "url": None,
+        "password": False,
+        "category": "provider",
+        "advanced": True,
+    },
     "GLM_API_KEY": {
         "description": "Z.AI / GLM API key (also recognized as ZAI_API_KEY / Z_AI_API_KEY)",
         "prompt": "Z.AI / GLM API key",
@@ -912,6 +1068,30 @@ OPTIONAL_ENV_VARS = {
         "category": "provider",
         "advanced": True,
     },
+    "HERMES_GEMINI_CLIENT_ID": {
+        "description": "Google OAuth client ID for google-gemini-cli (optional; defaults to Google's public gemini-cli client)",
+        "prompt": "Google OAuth client ID (optional — leave empty to use the public default)",
+        "url": "https://console.cloud.google.com/apis/credentials",
+        "password": False,
+        "category": "provider",
+        "advanced": True,
+    },
+    "HERMES_GEMINI_CLIENT_SECRET": {
+        "description": "Google OAuth client secret for google-gemini-cli (optional)",
+        "prompt": "Google OAuth client secret (optional)",
+        "url": "https://console.cloud.google.com/apis/credentials",
+        "password": True,
+        "category": "provider",
+        "advanced": True,
+    },
+    "HERMES_GEMINI_PROJECT_ID": {
+        "description": "GCP project ID for paid Gemini tiers (free tier auto-provisions)",
+        "prompt": "GCP project ID for Gemini OAuth (leave empty for free tier)",
+        "url": None,
+        "password": False,
+        "category": "provider",
+        "advanced": True,
+    },
     "OPENCODE_ZEN_API_KEY": {
         "description": "OpenCode Zen API key (pay-as-you-go access to curated models)",
         "prompt": "OpenCode Zen API key",
@@ -959,6 +1139,22 @@ OPTIONAL_ENV_VARS = {
         "category": "provider",
         "advanced": True,
     },
+    "OLLAMA_API_KEY": {
+        "description": "Ollama Cloud API key (ollama.com — cloud-hosted open models)",
+        "prompt": "Ollama Cloud API key",
+        "url": "https://ollama.com/settings",
+        "password": True,
+        "category": "provider",
+        "advanced": True,
+    },
+    "OLLAMA_BASE_URL": {
+        "description": "Ollama Cloud base URL override (default: https://ollama.com/v1)",
+        "prompt": "Ollama base URL (leave empty for default)",
+        "url": None,
+        "password": False,
+        "category": "provider",
+        "advanced": True,
+    },
     "XIAOMI_API_KEY": {
         "description": "Xiaomi MiMo API key for MiMo models (mimo-v2-pro, mimo-v2-omni, mimo-v2-flash)",
         "prompt": "Xiaomi MiMo API Key",
@@ -974,6 +1170,22 @@ OPTIONAL_ENV_VARS = {
         "category": "provider",
         "advanced": True,
     },
+    "AWS_REGION": {
+        "description": "AWS region for Bedrock API calls (e.g. us-east-1, eu-central-1)",
+        "prompt": "AWS Region",
+        "url": "https://docs.aws.amazon.com/bedrock/latest/userguide/bedrock-regions.html",
+        "password": False,
+        "category": "provider",
+        "advanced": True,
+    },
+    "AWS_PROFILE": {
+        "description": "AWS named profile for Bedrock authentication (from ~/.aws/credentials)",
+        "prompt": "AWS Profile",
+        "url": None,
+        "password": False,
+        "category": "provider",
+        "advanced": True,
+    },
 
     # ── Tool API keys ──
     "EXA_API_KEY": {
@@ -1171,6 +1383,12 @@ OPTIONAL_ENV_VARS = {
         "password": False,
         "category": "messaging",
     },
+    "TELEGRAM_PROXY": {
+        "description": "Proxy URL for Telegram connections (overrides HTTPS_PROXY). Supports http://, https://, socks5://",
+        "prompt": "Telegram proxy URL (optional)",
+        "password": False,
+        "category": "messaging",
+    },
     "DISCORD_BOT_TOKEN": {
         "description": "Discord bot token from Developer Portal",
         "prompt": "Discord bot token",
@@ -1366,12 +1584,12 @@ OPTIONAL_ENV_VARS = {
         "prompt": "Allow All QQ Users",
         "category": "messaging",
     },
-    "QQ_HOME_CHANNEL": {
+    "QQBOT_HOME_CHANNEL": {
         "description": "Default QQ channel/group for cron delivery and notifications",
         "prompt": "QQ Home Channel",
         "category": "messaging",
     },
-    "QQ_HOME_CHANNEL_NAME": {
+    "QQBOT_HOME_CHANNEL_NAME": {
         "description": "Display name for the QQ home channel",
         "prompt": "QQ Home Channel Name",
         "category": "messaging",
@@ -1468,13 +1686,8 @@ OPTIONAL_ENV_VARS = {
     },
 
     # ── Agent settings ──
-    "MESSAGING_CWD": {
-        "description": "Working directory for terminal commands via messaging",
-        "prompt": "Messaging working directory (default: home)",
-        "url": None,
-        "password": False,
-        "category": "setting",
-    },
+    # NOTE: MESSAGING_CWD was removed here — use terminal.cwd in config.yaml
+    # instead.  The gateway reads TERMINAL_CWD (bridged from terminal.cwd).
     "SUDO_PASSWORD": {
         "description": "Sudo password for terminal commands requiring root access; set to an explicit empty string to try empty without prompting",
         "prompt": "Sudo password",
@@ -1522,14 +1735,8 @@ OPTIONAL_ENV_VARS = {
     },
 }
 
-if not _managed_nous_tools_enabled():
-    for _hidden_var in (
-        "FIRECRAWL_GATEWAY_URL",
-        "TOOL_GATEWAY_DOMAIN",
-        "TOOL_GATEWAY_SCHEME",
-        "TOOL_GATEWAY_USER_TOKEN",
-    ):
-        OPTIONAL_ENV_VARS.pop(_hidden_var, None)
+# Tool Gateway env vars are always visible — they're useful for
+# self-hosted / custom gateway setups regardless of subscription state.
 
 
 def get_missing_env_vars(required_only: bool = False) -> List[Dict[str, Any]]:
@@ -1953,6 +2160,52 @@ def print_config_warnings(config: Optional[Dict[str, Any]] = None) -> None:
     sys.stderr.write("\n".join(lines) + "\n\n")
 
 
+def warn_deprecated_cwd_env_vars(config: Optional[Dict[str, Any]] = None) -> None:
+    """Warn if MESSAGING_CWD or TERMINAL_CWD is set in .env instead of config.yaml.
+
+    These env vars are deprecated — the canonical setting is terminal.cwd
+    in config.yaml.  Prints a migration hint to stderr.
+    """
+    import os, sys
+    messaging_cwd = os.environ.get("MESSAGING_CWD")
+    terminal_cwd_env = os.environ.get("TERMINAL_CWD")
+
+    if config is None:
+        try:
+            config = load_config()
+        except Exception:
+            return
+
+    terminal_cfg = config.get("terminal", {})
+    config_cwd = terminal_cfg.get("cwd", ".") if isinstance(terminal_cfg, dict) else "."
+    # Only warn if config.yaml doesn't have an explicit path
+    config_has_explicit_cwd = config_cwd not in (".", "auto", "cwd", "")
+
+    lines: list[str] = []
+    if messaging_cwd:
+        lines.append(
+            f"  \033[33m⚠\033[0m MESSAGING_CWD={messaging_cwd} found in .env — "
+            f"this is deprecated."
+        )
+    if terminal_cwd_env and not config_has_explicit_cwd:
+        # TERMINAL_CWD in env but not from config bridge — likely from .env
+        lines.append(
+            f"  \033[33m⚠\033[0m TERMINAL_CWD={terminal_cwd_env} found in .env — "
+            f"this is deprecated."
+        )
+    if lines:
+        hint_path = os.environ.get("HERMES_HOME", "~/.hermes")
+        lines.insert(0, "\033[33m⚠ Deprecated .env settings detected:\033[0m")
+        lines.append(
+            f"  \033[2mMove to config.yaml instead:  "
+            f"terminal:\\n    cwd: /your/project/path\033[0m"
+        )
+        lines.append(
+            f"  \033[2mThen remove the old entries from {hint_path}/.env\033[0m"
+        )
+        sys.stderr.write("\n".join(lines) + "\n\n")
+
+
 def migrate_config(interactive: bool = True, quiet: bool = False) -> Dict[str, Any]:
     """
     Migrate config to latest version, prompting for new required fields.
@@ -2423,6 +2676,85 @@ def _expand_env_vars(obj):
     return obj
 
 
+def _items_by_unique_name(items):
+    """Return a name-indexed dict only when all items have unique string names."""
+    if not isinstance(items, list):
+        return None
+    indexed = {}
+    for item in items:
+        if not isinstance(item, dict) or not isinstance(item.get("name"), str):
+            return None
+        name = item["name"]
+        if name in indexed:
+            return None
+        indexed[name] = item
+    return indexed
+
+
+def _preserve_env_ref_templates(current, raw, loaded_expanded=None):
+    """Restore raw ``${VAR}`` templates when a value is otherwise unchanged.
+
+    ``load_config()`` expands env refs for runtime use. When a caller later
+    persists that config after modifying some unrelated setting, keep the
+    original on-disk template instead of writing the expanded plaintext
+    secret back to ``config.yaml``.
+
+    Prefer preserving the raw template when ``current`` still matches either
+    the value previously returned by ``load_config()`` for this config path or
+    the current environment expansion of ``raw``. This handles env-var
+    rotation between load and save while still treating mixed literal/template
+    string edits as caller-owned once their rendered value diverges.
+    """
+    if isinstance(current, str) and isinstance(raw, str) and re.search(r"\${[^}]+}", raw):
+        if current == raw:
+            return raw
+        if isinstance(loaded_expanded, str) and current == loaded_expanded:
+            return raw
+        if _expand_env_vars(raw) == current:
+            return raw
+        return current
+
+    if isinstance(current, dict) and isinstance(raw, dict):
+        return {
+            key: _preserve_env_ref_templates(
+                value,
+                raw.get(key),
+                loaded_expanded.get(key) if isinstance(loaded_expanded, dict) else None,
+            )
+            for key, value in current.items()
+        }
+
+    if isinstance(current, list) and isinstance(raw, list):
+        # Prefer matching named config objects (e.g. custom_providers) by name
+        # so harmless reordering doesn't drop the original template. If names
+        # are duplicated, fall back to positional matching instead of silently
+        # shadowing one entry.
+        current_by_name = _items_by_unique_name(current)
+        raw_by_name = _items_by_unique_name(raw)
+        loaded_by_name = _items_by_unique_name(loaded_expanded)
+        if current_by_name is not None and raw_by_name is not None:
+            return [
+                _preserve_env_ref_templates(
+                    item,
+                    raw_by_name.get(item.get("name")),
+                    loaded_by_name.get(item.get("name")) if loaded_by_name is not None else None,
+                )
+                for item in current
+            ]
+        return [
+            _preserve_env_ref_templates(
+                item,
+                raw[index] if index < len(raw) else None,
+                loaded_expanded[index]
+                if isinstance(loaded_expanded, list) and index < len(loaded_expanded)
+                else None,
+            )
+            for index, item in enumerate(current)
+        ]
+
+    return current
+
+
 def _normalize_root_model_keys(config: Dict[str, Any]) -> Dict[str, Any]:
     """Move stale root-level provider/base_url into model section.
 
@@ -2490,7 +2822,6 @@ def read_raw_config() -> Dict[str, Any]:
 
 def load_config() -> Dict[str, Any]:
     """Load configuration from ~/.hermes/config.yaml."""
-    import copy
     ensure_hermes_home()
     config_path = get_config_path()
     
@@ -2511,8 +2842,11 @@ def load_config() -> Dict[str, Any]:
             config = _deep_merge(config, user_config)
         except Exception as e:
             print(f"Warning: Failed to load config: {e}")
-    
-    return _expand_env_vars(_normalize_root_model_keys(_normalize_max_turns_config(config)))
+
+    normalized = _normalize_root_model_keys(_normalize_max_turns_config(config))
+    expanded = _expand_env_vars(normalized)
+    _LAST_EXPANDED_CONFIG_BY_PATH[str(config_path)] = copy.deepcopy(expanded)
+    return expanded
 
 
 _SECURITY_COMMENT = """
@@ -2547,24 +2881,11 @@ _FALLBACK_COMMENT = """
 #   minimax      (MINIMAX_API_KEY)     — MiniMax
 #   minimax-cn   (MINIMAX_CN_API_KEY)  — MiniMax (China)
 #
-# For custom OpenAI-compatible endpoints, add base_url and api_key_env.
+# For custom OpenAI-compatible endpoints, add base_url and key_env.
 #
 # fallback_model:
 #   provider: openrouter
 #   model: anthropic/claude-sonnet-4
-#
-# ── Smart Model Routing ────────────────────────────────────────────────
-# Optional cheap-vs-strong routing for simple turns.
-# Keeps the primary model for complex work, but can route short/simple
-# messages to a cheaper model across providers.
-#
-# smart_model_routing:
-#   enabled: true
-#   max_simple_chars: 160
-#   max_simple_words: 28
-#   cheap_model:
-#     provider: openrouter
-#     model: google/gemini-2.5-flash
 """
 
 
@@ -2591,24 +2912,11 @@ _COMMENTED_SECTIONS = """
 #   minimax      (MINIMAX_API_KEY)     — MiniMax
 #   minimax-cn   (MINIMAX_CN_API_KEY)  — MiniMax (China)
 #
-# For custom OpenAI-compatible endpoints, add base_url and api_key_env.
+# For custom OpenAI-compatible endpoints, add base_url and key_env.
 #
 # fallback_model:
 #   provider: openrouter
 #   model: anthropic/claude-sonnet-4
-#
-# ── Smart Model Routing ────────────────────────────────────────────────
-# Optional cheap-vs-strong routing for simple turns.
-# Keeps the primary model for complex work, but can route short/simple
-# messages to a cheaper model across providers.
-#
-# smart_model_routing:
-#   enabled: true
-#   max_simple_chars: 160
-#   max_simple_words: 28
-#   cheap_model:
-#     provider: openrouter
-#     model: google/gemini-2.5-flash
 """
 
 
@@ -2621,7 +2929,15 @@ def save_config(config: Dict[str, Any]):
 
     ensure_hermes_home()
     config_path = get_config_path()
-    normalized = _normalize_root_model_keys(_normalize_max_turns_config(config))
+    current_normalized = _normalize_root_model_keys(_normalize_max_turns_config(config))
+    normalized = current_normalized
+    raw_existing = _normalize_root_model_keys(_normalize_max_turns_config(read_raw_config()))
+    if raw_existing:
+        normalized = _preserve_env_ref_templates(
+            normalized,
+            raw_existing,
+            _LAST_EXPANDED_CONFIG_BY_PATH.get(str(config_path)),
+        )
 
     # Build optional commented-out sections for features that are off by
     # default or only relevant when explicitly configured.
@@ -2639,6 +2955,7 @@ def save_config(config: Dict[str, Any]):
         extra_content="".join(parts) if parts else None,
     )
     _secure_file(config_path)
+    _LAST_EXPANDED_CONFIG_BY_PATH[str(config_path)] = copy.deepcopy(current_normalized)
 
 
 def load_env() -> Dict[str, str]:
@@ -2847,12 +3164,25 @@ def save_env_value(key: str, value: str):
         lines.append(f"{key}={value}\n")
     
     fd, tmp_path = tempfile.mkstemp(dir=str(env_path.parent), suffix='.tmp', prefix='.env_')
+    # Preserve original permissions so Docker volume mounts aren't clobbered.
+    original_mode = None
+    if env_path.exists():
+        try:
+            original_mode = stat.S_IMODE(env_path.stat().st_mode)
+        except OSError:
+            pass
     try:
         with os.fdopen(fd, 'w', **write_kw) as f:
             f.writelines(lines)
             f.flush()
             os.fsync(f.fileno())
         os.replace(tmp_path, env_path)
+        # Restore original permissions before _secure_file may tighten them.
+        if original_mode is not None:
+            try:
+                os.chmod(env_path, original_mode)
+            except OSError:
+                pass
     except BaseException:
         try:
             os.unlink(tmp_path)
@@ -2863,13 +3193,6 @@ def save_env_value(key: str, value: str):
 
     os.environ[key] = value
 
-    # Restrict .env permissions to owner-only (contains API keys)
-    if not _IS_WINDOWS:
-        try:
-            os.chmod(env_path, stat.S_IRUSR | stat.S_IWUSR)
-        except OSError:
-            pass
-
 
 def remove_env_value(key: str) -> bool:
     """Remove a key from ~/.hermes/.env and os.environ.
@@ -2898,12 +3221,23 @@ def remove_env_value(key: str) -> bool:
 
     if found:
         fd, tmp_path = tempfile.mkstemp(dir=str(env_path.parent), suffix='.tmp', prefix='.env_')
+        # Preserve original permissions so Docker volume mounts aren't clobbered.
+        original_mode = None
+        try:
+            original_mode = stat.S_IMODE(env_path.stat().st_mode)
+        except OSError:
+            pass
         try:
             with os.fdopen(fd, 'w', **write_kw) as f:
                 f.writelines(new_lines)
                 f.flush()
                 os.fsync(f.fileno())
             os.replace(tmp_path, env_path)
+            if original_mode is not None:
+                try:
+                    os.chmod(env_path, original_mode)
+                except OSError:
+                    pass
         except BaseException:
             try:
                 os.unlink(tmp_path)
@@ -3045,6 +3379,10 @@ def show_config():
     print(f"  Personality:  {display.get('personality', 'kawaii')}")
     print(f"  Reasoning:    {'on' if display.get('show_reasoning', False) else 'off'}")
     print(f"  Bell:         {'on' if display.get('bell_on_complete', False) else 'off'}")
+    ump = display.get('user_message_preview', {}) if isinstance(display.get('user_message_preview', {}), dict) else {}
+    ump_first = ump.get('first_lines', 2)
+    ump_last = ump.get('last_lines', 2)
+    print(f"  User preview: first {ump_first} line(s), last {ump_last} line(s)")
 
     # Terminal
     print()

hermes_cli/curses_ui.py

@@ -166,6 +166,7 @@ def curses_radiolist(
     selected: int = 0,
     *,
     cancel_returns: int | None = None,
+    description: str | None = None,
 ) -> int:
     """Curses single-select radio list. Returns the selected index.
 
@@ -174,6 +175,9 @@ def curses_radiolist(
         items: Display labels for each row.
         selected: Index that starts selected (pre-selected).
         cancel_returns: Returned on ESC/q. Defaults to the original *selected*.
+        description: Optional multi-line text shown between the title and
+            the item list.  Useful for context that should survive the
+            curses screen clear.
     """
     if cancel_returns is None:
         cancel_returns = selected
@@ -181,6 +185,10 @@ def curses_radiolist(
     if not sys.stdin.isatty():
         return cancel_returns
 
+    desc_lines: list[str] = []
+    if description:
+        desc_lines = description.splitlines()
+
     try:
         import curses
         result_holder: list = [None]
@@ -199,22 +207,35 @@ def curses_radiolist(
                 stdscr.clear()
                 max_y, max_x = stdscr.getmaxyx()
 
+                row = 0
+
                 # Header
                 try:
                     hattr = curses.A_BOLD
                     if curses.has_colors():
                         hattr |= curses.color_pair(2)
-                    stdscr.addnstr(0, 0, title, max_x - 1, hattr)
+                    stdscr.addnstr(row, 0, title, max_x - 1, hattr)
+                    row += 1
+
+                    # Description lines
+                    for dline in desc_lines:
+                        if row >= max_y - 1:
+                            break
+                        stdscr.addnstr(row, 0, dline, max_x - 1, curses.A_NORMAL)
+                        row += 1
+
                     stdscr.addnstr(
-                        1, 0,
+                        row, 0,
                         "  \u2191\u2193 navigate  ENTER/SPACE select  ESC cancel",
                         max_x - 1, curses.A_DIM,
                     )
+                    row += 1
                 except curses.error:
                     pass
 
                 # Scrollable item list
-                visible_rows = max_y - 4
+                items_start = row + 1
+                visible_rows = max_y - items_start - 1
                 if cursor < scroll_offset:
                     scroll_offset = cursor
                 elif cursor >= scroll_offset + visible_rows:
@@ -223,7 +244,7 @@ def curses_radiolist(
                 for draw_i, i in enumerate(
                     range(scroll_offset, min(len(items), scroll_offset + visible_rows))
                 ):
-                    y = draw_i + 3
+                    y = draw_i + items_start
                     if y >= max_y - 1:
                         break
                     radio = "\u25cf" if i == selected else "\u25cb"

hermes_cli/debug.py

@@ -6,7 +6,10 @@ Currently supports:
 """
 
 import io
+import json
+import os
 import sys
+import time
 import urllib.error
 import urllib.parse
 import urllib.request
@@ -27,6 +30,205 @@ _DPASTE_COM_URL = "https://dpaste.com/api/"
 # paste.rs caps at ~1 MB; we stay under that with headroom.
 _MAX_LOG_BYTES = 512_000
 
+# Auto-delete pastes after this many seconds (6 hours).
+_AUTO_DELETE_SECONDS = 21600
+
+
+# ---------------------------------------------------------------------------
+# Pending-deletion tracking (replaces the old fork-and-sleep subprocess).
+# ---------------------------------------------------------------------------
+
+def _pending_file() -> Path:
+    """Path to ``~/.hermes/pastes/pending.json``.
+
+    Each entry: ``{"url": "...", "expire_at": <unix_ts>}``.  Scheduled
+    DELETEs used to be handled by spawning a detached Python process per
+    paste that slept for 6 hours; those accumulated forever if the user
+    ran ``hermes debug share`` repeatedly.  We now persist the schedule
+    to disk and sweep expired entries on the next debug invocation.
+    """
+    return get_hermes_home() / "pastes" / "pending.json"
+
+
+def _load_pending() -> list[dict]:
+    path = _pending_file()
+    if not path.exists():
+        return []
+    try:
+        data = json.loads(path.read_text(encoding="utf-8"))
+        if isinstance(data, list):
+            # Filter to well-formed entries only
+            return [
+                e for e in data
+                if isinstance(e, dict) and "url" in e and "expire_at" in e
+            ]
+    except (OSError, ValueError, json.JSONDecodeError):
+        pass
+    return []
+
+
+def _save_pending(entries: list[dict]) -> None:
+    path = _pending_file()
+    try:
+        path.parent.mkdir(parents=True, exist_ok=True)
+        tmp = path.with_suffix(".json.tmp")
+        tmp.write_text(json.dumps(entries, indent=2), encoding="utf-8")
+        os.replace(tmp, path)
+    except OSError:
+        # Non-fatal — worst case the user has to run ``hermes debug delete``
+        # manually.
+        pass
+
+
+def _record_pending(urls: list[str], delay_seconds: int = _AUTO_DELETE_SECONDS) -> None:
+    """Record *urls* for deletion at ``now + delay_seconds``.
+
+    Only paste.rs URLs are recorded (dpaste.com auto-expires).  Entries
+    are merged into any existing pending.json.
+    """
+    paste_rs_urls = [u for u in urls if _extract_paste_id(u)]
+    if not paste_rs_urls:
+        return
+
+    entries = _load_pending()
+    # Dedupe by URL: keep the later expire_at if same URL appears twice
+    by_url: dict[str, float] = {e["url"]: float(e["expire_at"]) for e in entries}
+    expire_at = time.time() + delay_seconds
+    for u in paste_rs_urls:
+        by_url[u] = max(expire_at, by_url.get(u, 0.0))
+    merged = [{"url": u, "expire_at": ts} for u, ts in by_url.items()]
+    _save_pending(merged)
+
+
+def _sweep_expired_pastes(now: Optional[float] = None) -> tuple[int, int]:
+    """Synchronously DELETE any pending pastes whose ``expire_at`` has passed.
+
+    Returns ``(deleted, remaining)``.  Best-effort: failed deletes stay in
+    the pending file and will be retried on the next sweep.  Silent —
+    intended to be called from every ``hermes debug`` invocation with
+    minimal noise.
+    """
+    entries = _load_pending()
+    if not entries:
+        return (0, 0)
+
+    current = time.time() if now is None else now
+    deleted = 0
+    remaining: list[dict] = []
+
+    for entry in entries:
+        try:
+            expire_at = float(entry.get("expire_at", 0))
+        except (TypeError, ValueError):
+            continue  # drop malformed entries
+        if expire_at > current:
+            remaining.append(entry)
+            continue
+
+        url = entry.get("url", "")
+        try:
+            if delete_paste(url):
+                deleted += 1
+                continue
+        except Exception:
+            # Network hiccup, 404 (already gone), etc. — drop the entry
+            # after a grace period; don't retry forever.
+            pass
+
+        # Retain failed deletes for up to 24h past expiration, then give up.
+        if expire_at + 86400 > current:
+            remaining.append(entry)
+        else:
+            deleted += 1  # count as reaped (paste.rs will GC eventually)
+
+    if deleted:
+        _save_pending(remaining)
+
+    return (deleted, len(remaining))
+
+
+# ---------------------------------------------------------------------------
+# Privacy / delete helpers
+# ---------------------------------------------------------------------------
+
+_PRIVACY_NOTICE = """\
+⚠️  This will upload the following to a public paste service:
+  • System info (OS, Python version, Hermes version, provider, which API keys
+    are configured — NOT the actual keys)
+  • Recent log lines (agent.log, errors.log, gateway.log — may contain
+    conversation fragments and file paths)
+  • Full agent.log and gateway.log (up to 512 KB each — likely contains
+    conversation content, tool outputs, and file paths)
+
+Pastes auto-delete after 6 hours.
+"""
+
+_GATEWAY_PRIVACY_NOTICE = (
+    "⚠️ **Privacy notice:** This uploads system info + recent log tails "
+    "(may contain conversation fragments) to a public paste service. "
+    "Full logs are NOT included from the gateway — use `hermes debug share` "
+    "from the CLI for full log uploads.\n"
+    "Pastes auto-delete after 6 hours."
+)
+
+
+def _extract_paste_id(url: str) -> Optional[str]:
+    """Extract the paste ID from a paste.rs or dpaste.com URL.
+
+    Returns the ID string, or None if the URL doesn't match a known service.
+    """
+    url = url.strip().rstrip("/")
+    for prefix in ("https://paste.rs/", "http://paste.rs/"):
+        if url.startswith(prefix):
+            return url[len(prefix):]
+    return None
+
+
+def delete_paste(url: str) -> bool:
+    """Delete a paste from paste.rs.  Returns True on success.
+
+    Only paste.rs supports unauthenticated DELETE.  dpaste.com pastes
+    expire automatically but cannot be deleted via API.
+    """
+    paste_id = _extract_paste_id(url)
+    if not paste_id:
+        raise ValueError(
+            f"Cannot delete: only paste.rs URLs are supported.  Got: {url}"
+        )
+
+    target = f"{_PASTE_RS_URL}{paste_id}"
+    req = urllib.request.Request(
+        target, method="DELETE",
+        headers={"User-Agent": "hermes-agent/debug-share"},
+    )
+    with urllib.request.urlopen(req, timeout=30) as resp:
+        return 200 <= resp.status < 300
+
+
+def _schedule_auto_delete(urls: list[str], delay_seconds: int = _AUTO_DELETE_SECONDS):
+    """Record *urls* for deletion ``delay_seconds`` from now.
+
+    Previously this spawned a detached Python subprocess per call that slept
+    for 6 hours and then issued DELETE requests.  Those subprocesses leaked —
+    every ``hermes debug share`` invocation added ~20 MB of resident Python
+    interpreters that never exited until the sleep completed.
+
+    The replacement is stateless: we append to ``~/.hermes/pastes/pending.json``
+    and rely on opportunistic sweeps (``_sweep_expired_pastes``) called from
+    every ``hermes debug`` invocation.  If the user never runs ``hermes debug``
+    again, paste.rs's own retention policy handles cleanup.
+    """
+    _record_pending(urls, delay_seconds=delay_seconds)
+
+
+def _delete_hint(url: str) -> str:
+    """Return a one-liner delete command for the given paste URL."""
+    paste_id = _extract_paste_id(url)
+    if paste_id:
+        return f"hermes debug delete {url}"
+    # dpaste.com — no API delete, expires on its own.
+    return "(auto-expires per dpaste.com policy)"
+
 
 def _upload_paste_rs(content: str) -> str:
     """Upload to paste.rs.  Returns the paste URL.
@@ -250,6 +452,9 @@ def run_debug_share(args):
     expiry = getattr(args, "expire", 7)
     local_only = getattr(args, "local", False)
 
+    if not local_only:
+        print(_PRIVACY_NOTICE)
+
     print("Collecting debug report...")
 
     # Capture dump once — prepended to every paste for context.
@@ -315,22 +520,66 @@ def run_debug_share(args):
     if failures:
         print(f"\n  (failed to upload: {', '.join(failures)})")
 
+    # Schedule auto-deletion after 6 hours
+    _schedule_auto_delete(list(urls.values()))
+    print(f"\n⏱  Pastes will auto-delete in 6 hours.")
+
+    # Manual delete fallback
+    print(f"To delete now:  hermes debug delete <url>")
+
     print(f"\nShare these links with the Hermes team for support.")
 
 
+def run_debug_delete(args):
+    """Delete one or more paste URLs uploaded by /debug."""
+    urls = getattr(args, "urls", [])
+    if not urls:
+        print("Usage: hermes debug delete <url> [<url> ...]")
+        print("  Deletes paste.rs pastes uploaded by 'hermes debug share'.")
+        return
+
+    for url in urls:
+        try:
+            ok = delete_paste(url)
+            if ok:
+                print(f"  ✓ Deleted: {url}")
+            else:
+                print(f"  ✗ Failed to delete: {url} (unexpected response)")
+        except ValueError as exc:
+            print(f"  ✗ {exc}")
+        except Exception as exc:
+            print(f"  ✗ Could not delete {url}: {exc}")
+
+
 def run_debug(args):
     """Route debug subcommands."""
+    # Opportunistic sweep of expired pastes on every ``hermes debug`` call.
+    # Replaces the old per-paste sleeping subprocess that used to leak as
+    # one orphaned Python interpreter per scheduled deletion.  Silent and
+    # best-effort — any failure is swallowed so ``hermes debug`` stays
+    # reliable even when offline.
+    try:
+        _sweep_expired_pastes()
+    except Exception:
+        pass
+
     subcmd = getattr(args, "debug_command", None)
     if subcmd == "share":
         run_debug_share(args)
+    elif subcmd == "delete":
+        run_debug_delete(args)
     else:
         # Default: show help
-        print("Usage: hermes debug share [--lines N] [--expire N] [--local]")
+        print("Usage: hermes debug <command>")
         print()
         print("Commands:")
         print("  share    Upload debug report to a paste service and print URL")
+        print("  delete   Delete a previously uploaded paste")
         print()
-        print("Options:")
+        print("Options (share):")
         print("  --lines N    Number of log lines to include (default: 200)")
         print("  --expire N   Paste expiry in days (default: 7)")
         print("  --local      Print report locally instead of uploading")
+        print()
+        print("Options (delete):")
+        print("  <url> ...    One or more paste URLs to delete")

hermes_cli/dingtalk_auth.py

@@ -0,0 +1,294 @@
+"""
+DingTalk Device Flow authorization.
+
+Implements the same 3-step registration flow as dingtalk-openclaw-connector:
+  1. POST /app/registration/init   → get nonce
+  2. POST /app/registration/begin  → get device_code + verification_uri_complete
+  3. POST /app/registration/poll   → poll until SUCCESS → get client_id + client_secret
+
+The verification_uri_complete is rendered as a QR code in the terminal so the
+user can scan it with DingTalk to authorize, yielding AppKey + AppSecret
+automatically.
+"""
+
+from __future__ import annotations
+
+import io
+import os
+import sys
+import time
+import logging
+from typing import Optional, Tuple
+
+import requests
+
+logger = logging.getLogger(__name__)
+
+# ── Configuration ──────────────────────────────────────────────────────────
+
+REGISTRATION_BASE_URL = os.environ.get(
+    "DINGTALK_REGISTRATION_BASE_URL", "https://oapi.dingtalk.com"
+).rstrip("/")
+
+REGISTRATION_SOURCE = os.environ.get("DINGTALK_REGISTRATION_SOURCE", "openClaw")
+
+
+# ── API helpers ────────────────────────────────────────────────────────────
+
+class RegistrationError(Exception):
+    """Raised when a DingTalk registration API call fails."""
+
+
+def _api_post(path: str, payload: dict) -> dict:
+    """POST to the registration API and return the parsed JSON body."""
+    url = f"{REGISTRATION_BASE_URL}{path}"
+    try:
+        resp = requests.post(url, json=payload, timeout=15)
+        resp.raise_for_status()
+        data = resp.json()
+    except requests.RequestException as exc:
+        raise RegistrationError(f"Network error calling {url}: {exc}") from exc
+
+    errcode = data.get("errcode", -1)
+    if errcode != 0:
+        errmsg = data.get("errmsg", "unknown error")
+        raise RegistrationError(f"API error [{path}]: {errmsg} (errcode={errcode})")
+    return data
+
+
+# ── Core flow ──────────────────────────────────────────────────────────────
+
+def begin_registration() -> dict:
+    """Start a device-flow registration.
+
+    Returns a dict with keys:
+        device_code, verification_uri_complete, expires_in, interval
+    """
+    # Step 1: init → nonce
+    init_data = _api_post("/app/registration/init", {"source": REGISTRATION_SOURCE})
+    nonce = str(init_data.get("nonce", "")).strip()
+    if not nonce:
+        raise RegistrationError("init response missing nonce")
+
+    # Step 2: begin → device_code, verification_uri_complete
+    begin_data = _api_post("/app/registration/begin", {"nonce": nonce})
+    device_code = str(begin_data.get("device_code", "")).strip()
+    verification_uri_complete = str(begin_data.get("verification_uri_complete", "")).strip()
+    if not device_code:
+        raise RegistrationError("begin response missing device_code")
+    if not verification_uri_complete:
+        raise RegistrationError("begin response missing verification_uri_complete")
+
+    return {
+        "device_code": device_code,
+        "verification_uri_complete": verification_uri_complete,
+        "expires_in": int(begin_data.get("expires_in", 7200)),
+        "interval": max(int(begin_data.get("interval", 3)), 2),
+    }
+
+
+def poll_registration(device_code: str) -> dict:
+    """Poll the registration status once.
+
+    Returns a dict with keys:  status, client_id?, client_secret?, fail_reason?
+    """
+    data = _api_post("/app/registration/poll", {"device_code": device_code})
+    status_raw = str(data.get("status", "")).strip().upper()
+    if status_raw not in ("WAITING", "SUCCESS", "FAIL", "EXPIRED"):
+        status_raw = "UNKNOWN"
+    return {
+        "status": status_raw,
+        "client_id": str(data.get("client_id", "")).strip() or None,
+        "client_secret": str(data.get("client_secret", "")).strip() or None,
+        "fail_reason": str(data.get("fail_reason", "")).strip() or None,
+    }
+
+
+def wait_for_registration_success(
+    device_code: str,
+    interval: int = 3,
+    expires_in: int = 7200,
+    on_waiting: Optional[callable] = None,
+) -> Tuple[str, str]:
+    """Block until the registration succeeds or times out.
+
+    Returns (client_id, client_secret).
+    """
+    deadline = time.monotonic() + expires_in
+    retry_window = 120  # 2 minutes for transient errors
+    retry_start = 0.0
+
+    while time.monotonic() < deadline:
+        time.sleep(interval)
+        try:
+            result = poll_registration(device_code)
+        except RegistrationError:
+            if retry_start == 0:
+                retry_start = time.monotonic()
+            if time.monotonic() - retry_start < retry_window:
+                continue
+            raise
+
+        status = result["status"]
+        if status == "WAITING":
+            retry_start = 0
+            if on_waiting:
+                on_waiting()
+            continue
+        if status == "SUCCESS":
+            cid = result["client_id"]
+            csecret = result["client_secret"]
+            if not cid or not csecret:
+                raise RegistrationError("authorization succeeded but credentials are missing")
+            return cid, csecret
+        # FAIL / EXPIRED / UNKNOWN
+        if retry_start == 0:
+            retry_start = time.monotonic()
+        if time.monotonic() - retry_start < retry_window:
+            continue
+        reason = result.get("fail_reason") or status
+        raise RegistrationError(f"authorization failed: {reason}")
+
+    raise RegistrationError("authorization timed out, please retry")
+
+
+# ── QR code rendering ─────────────────────────────────────────────────────
+
+def _ensure_qrcode_installed() -> bool:
+    """Try to import qrcode; if missing, auto-install it via pip/uv."""
+    try:
+        import qrcode  # noqa: F401
+        return True
+    except ImportError:
+        pass
+
+    import subprocess
+
+    # Try uv first (Hermes convention), then pip
+    for cmd in (
+        [sys.executable, "-m", "uv", "pip", "install", "qrcode"],
+        [sys.executable, "-m", "pip", "install", "-q", "qrcode"],
+    ):
+        try:
+            subprocess.check_call(cmd, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+            import qrcode  # noqa: F401,F811
+            return True
+        except (subprocess.CalledProcessError, ImportError, FileNotFoundError):
+            continue
+    return False
+
+
+def render_qr_to_terminal(url: str) -> bool:
+    """Render *url* as a compact QR code in the terminal.
+
+    Returns True if the QR code was printed, False if the library is missing.
+    """
+    try:
+        import qrcode
+    except ImportError:
+        return False
+
+    qr = qrcode.QRCode(
+        version=1,
+        error_correction=qrcode.constants.ERROR_CORRECT_L,
+        box_size=1,
+        border=1,
+    )
+    qr.add_data(url)
+    qr.make(fit=True)
+
+    # Use half-block characters for compact rendering (2 rows per character)
+    matrix = qr.get_matrix()
+    rows = len(matrix)
+    lines: list[str] = []
+
+    TOP_HALF = "\u2580"      # ▀
+    BOTTOM_HALF = "\u2584"   # ▄
+    FULL_BLOCK = "\u2588"    # █
+    EMPTY = " "
+
+    for r in range(0, rows, 2):
+        line_chars: list[str] = []
+        for c in range(len(matrix[r])):
+            top = matrix[r][c]
+            bottom = matrix[r + 1][c] if r + 1 < rows else False
+            if top and bottom:
+                line_chars.append(FULL_BLOCK)
+            elif top:
+                line_chars.append(TOP_HALF)
+            elif bottom:
+                line_chars.append(BOTTOM_HALF)
+            else:
+                line_chars.append(EMPTY)
+        lines.append("    " + "".join(line_chars))
+
+    print("\n".join(lines))
+    return True
+
+
+# ── High-level entry point for the setup wizard ───────────────────────────
+
+def dingtalk_qr_auth() -> Optional[Tuple[str, str]]:
+    """Run the interactive QR-code device-flow authorization.
+
+    Returns (client_id, client_secret) on success, or None if the user
+    cancelled or the flow failed.
+    """
+    from hermes_cli.setup import print_info, print_success, print_warning, print_error
+
+    print()
+    print_info("  Initializing DingTalk device authorization...")
+    print_info("  Note: the scan page is branded 'OpenClaw' — DingTalk's")
+    print_info("        ecosystem onboarding bridge. Safe to use.")
+
+    try:
+        reg = begin_registration()
+    except RegistrationError as exc:
+        print_error(f"  Authorization init failed: {exc}")
+        return None
+
+    url = reg["verification_uri_complete"]
+
+    # Ensure qrcode library is available (auto-install if missing)
+    if not _ensure_qrcode_installed():
+        print_warning("  qrcode library install failed, will show link only.")
+
+    print()
+    print_info("  Please scan the QR code below with DingTalk to authorize:")
+    print()
+
+    if not render_qr_to_terminal(url):
+        print_warning(f"  QR code render failed, please open the link below to authorize:")
+
+    print()
+    print_info(f"  Or open this link manually: {url}")
+    print()
+    print_info("  Waiting for QR scan authorization... (timeout: 2 hours)")
+
+    dot_count = 0
+
+    def _on_waiting():
+        nonlocal dot_count
+        dot_count += 1
+        if dot_count % 10 == 0:
+            sys.stdout.write(".")
+            sys.stdout.flush()
+
+    try:
+        client_id, client_secret = wait_for_registration_success(
+            device_code=reg["device_code"],
+            interval=reg["interval"],
+            expires_in=reg["expires_in"],
+            on_waiting=_on_waiting,
+        )
+    except RegistrationError as exc:
+        print()
+        print_error(f"  Authorization failed: {exc}")
+        return None
+
+    print()
+    print_success("  QR scan authorization successful!")
+    print_success(f"  Client ID:     {client_id}")
+    print_success(f"  Client Secret: {client_secret[:8]}{'*' * (len(client_secret) - 8)}")
+
+    return client_id, client_secret

hermes_cli/doctor.py

@@ -277,6 +277,86 @@ def run_doctor(args):
     config_path = HERMES_HOME / 'config.yaml'
     if config_path.exists():
         check_ok(f"{_DHH}/config.yaml exists")
+
+        # Validate model.provider and model.default values
+        try:
+            import yaml as _yaml
+            cfg = _yaml.safe_load(config_path.read_text(encoding="utf-8")) or {}
+            model_section = cfg.get("model") or {}
+            provider_raw = (model_section.get("provider") or "").strip()
+            provider = provider_raw.lower()
+            default_model = (model_section.get("default") or model_section.get("model") or "").strip()
+
+            known_providers: set = set()
+            try:
+                from hermes_cli.auth import PROVIDER_REGISTRY
+                known_providers = set(PROVIDER_REGISTRY.keys()) | {"openrouter", "custom", "auto"}
+            except Exception:
+                pass
+            try:
+                from hermes_cli.auth import resolve_provider as _resolve_provider
+            except Exception:
+                _resolve_provider = None
+
+            canonical_provider = provider
+            if provider and _resolve_provider is not None and provider != "auto":
+                try:
+                    canonical_provider = _resolve_provider(provider)
+                except Exception:
+                    canonical_provider = None
+
+            if provider and provider != "auto":
+                if canonical_provider is None or (known_providers and canonical_provider not in known_providers):
+                    known_list = ", ".join(sorted(known_providers)) if known_providers else "(unavailable)"
+                    check_fail(
+                        f"model.provider '{provider_raw}' is not a recognised provider",
+                        f"(known: {known_list})",
+                    )
+                    issues.append(
+                        f"model.provider '{provider_raw}' is unknown. "
+                        f"Valid providers: {known_list}. "
+                        f"Fix: run 'hermes config set model.provider <valid_provider>'"
+                    )
+
+            # Warn if model is set to a provider-prefixed name on a provider that doesn't use them
+            if default_model and "/" in default_model and canonical_provider and canonical_provider not in ("openrouter", "custom", "auto", "ai-gateway", "kilocode", "opencode-zen", "huggingface", "nous"):
+                check_warn(
+                    f"model.default '{default_model}' uses a vendor/model slug but provider is '{provider_raw}'",
+                    "(vendor-prefixed slugs belong to aggregators like openrouter)",
+                )
+                issues.append(
+                    f"model.default '{default_model}' is vendor-prefixed but model.provider is '{provider_raw}'. "
+                    "Either set model.provider to 'openrouter', or drop the vendor prefix."
+                )
+
+            # Check credentials for the configured provider.
+            # Limit to API-key providers in PROVIDER_REGISTRY — other provider
+            # types (OAuth, SDK, openrouter/anthropic/custom/auto) have their
+            # own env-var checks elsewhere in doctor, and get_auth_status()
+            # returns a bare {logged_in: False} for anything it doesn't
+            # explicitly dispatch, which would produce false positives.
+            if canonical_provider and canonical_provider not in ("auto", "custom", "openrouter"):
+                try:
+                    from hermes_cli.auth import PROVIDER_REGISTRY, get_auth_status
+                    pconfig = PROVIDER_REGISTRY.get(canonical_provider)
+                    if pconfig and getattr(pconfig, "auth_type", "") == "api_key":
+                        status = get_auth_status(canonical_provider) or {}
+                        configured = bool(status.get("configured") or status.get("logged_in") or status.get("api_key"))
+                        if not configured:
+                            check_fail(
+                                f"model.provider '{canonical_provider}' is set but no API key is configured",
+                                "(check ~/.hermes/.env or run 'hermes setup')",
+                            )
+                            issues.append(
+                                f"No credentials found for provider '{canonical_provider}'. "
+                                f"Run 'hermes setup' or set the provider's API key in {_DHH}/.env, "
+                                f"or switch providers with 'hermes config set model.provider <name>'"
+                            )
+                except Exception:
+                    pass
+
+        except Exception as e:
+            check_warn("Could not validate model/provider config", f"({e})")
     else:
         fallback_config = PROJECT_ROOT / 'cli-config.yaml'
         if fallback_config.exists():
@@ -373,7 +453,11 @@ def run_doctor(args):
     print(color("◆ Auth Providers", Colors.CYAN, Colors.BOLD))
 
     try:
-        from hermes_cli.auth import get_nous_auth_status, get_codex_auth_status
+        from hermes_cli.auth import (
+            get_nous_auth_status,
+            get_codex_auth_status,
+            get_gemini_oauth_auth_status,
+        )
 
         nous_status = get_nous_auth_status()
         if nous_status.get("logged_in"):
@@ -388,6 +472,20 @@ def run_doctor(args):
             check_warn("OpenAI Codex auth", "(not logged in)")
             if codex_status.get("error"):
                 check_info(codex_status["error"])
+
+        gemini_status = get_gemini_oauth_auth_status()
+        if gemini_status.get("logged_in"):
+            email = gemini_status.get("email") or ""
+            project = gemini_status.get("project_id") or ""
+            pieces = []
+            if email:
+                pieces.append(email)
+            if project:
+                pieces.append(f"project={project}")
+            suffix = f" ({', '.join(pieces)})" if pieces else ""
+            check_ok("Google Gemini OAuth", f"(logged in{suffix})")
+        else:
+            check_warn("Google Gemini OAuth", "(not logged in)")
     except Exception as e:
         check_warn("Auth provider status", f"(could not check: {e})")
 
@@ -760,6 +858,16 @@ def run_doctor(args):
             elif response.status_code == 401:
                 print(f"\r  {color('✗', Colors.RED)} OpenRouter API {color('(invalid API key)', Colors.DIM)}                ")
                 issues.append("Check OPENROUTER_API_KEY in .env")
+            elif response.status_code == 402:
+                print(f"\r  {color('✗', Colors.RED)} OpenRouter API {color('(out of credits — payment required)', Colors.DIM)}")
+                issues.append(
+                    "OpenRouter account has insufficient credits. "
+                    "Fix: run 'hermes config set model.provider <provider>' to switch providers, "
+                    "or fund your OpenRouter account at https://openrouter.ai/settings/credits"
+                )
+            elif response.status_code == 429:
+                print(f"\r  {color('✗', Colors.RED)} OpenRouter API {color('(rate limited)', Colors.DIM)}                ")
+                issues.append("OpenRouter rate limit hit — consider switching to a different provider or waiting")
             else:
                 print(f"\r  {color('✗', Colors.RED)} OpenRouter API {color(f'(HTTP {response.status_code})', Colors.DIM)}                ")
         except Exception as e:
@@ -807,6 +915,7 @@ def run_doctor(args):
         ("Arcee AI",         ("ARCEEAI_API_KEY",),                            "https://api.arcee.ai/api/v1/models",  "ARCEE_BASE_URL", True),
         ("DeepSeek",         ("DEEPSEEK_API_KEY",),                           "https://api.deepseek.com/v1/models",  "DEEPSEEK_BASE_URL", True),
         ("Hugging Face",     ("HF_TOKEN",),                                   "https://router.huggingface.co/v1/models", "HF_BASE_URL", True),
+        ("NVIDIA NIM",       ("NVIDIA_API_KEY",),                             "https://integrate.api.nvidia.com/v1/models", "NVIDIA_BASE_URL", True),
         ("Alibaba/DashScope", ("DASHSCOPE_API_KEY",),                         "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/models", "DASHSCOPE_BASE_URL", True),
         # MiniMax: the /anthropic endpoint doesn't support /models, but the /v1 endpoint does.
         ("MiniMax",          ("MINIMAX_API_KEY",),                            "https://api.minimax.io/v1/models",    "MINIMAX_BASE_URL", True),
@@ -814,7 +923,8 @@ def run_doctor(args):
         ("Vercel AI Gateway",       ("AI_GATEWAY_API_KEY",),                          "https://ai-gateway.vercel.sh/v1/models", "AI_GATEWAY_BASE_URL", True),
         ("Kilo Code",        ("KILOCODE_API_KEY",),                            "https://api.kilo.ai/api/gateway/models",  "KILOCODE_BASE_URL", True),
         ("OpenCode Zen",     ("OPENCODE_ZEN_API_KEY",),                        "https://opencode.ai/zen/v1/models",  "OPENCODE_ZEN_BASE_URL", True),
-        ("OpenCode Go",      ("OPENCODE_GO_API_KEY",),                         "https://opencode.ai/zen/go/v1/models", "OPENCODE_GO_BASE_URL", True),
+        # OpenCode Go has no shared /models endpoint; skip the health check.
+        ("OpenCode Go",      ("OPENCODE_GO_API_KEY",),                         None,                                  "OPENCODE_GO_BASE_URL", False),
     ]
     for _pname, _env_vars, _default_url, _base_env, _supports_health_check in _apikey_providers:
         _key = ""
@@ -859,6 +969,31 @@ def run_doctor(args):
             except Exception as _e:
                 print(f"\r  {color('⚠', Colors.YELLOW)} {_label} {color(f'({_e})', Colors.DIM)}           ")
 
+    # -- AWS Bedrock --
+    # Bedrock uses the AWS SDK credential chain, not API keys.
+    try:
+        from agent.bedrock_adapter import has_aws_credentials, resolve_aws_auth_env_var, resolve_bedrock_region
+        if has_aws_credentials():
+            _auth_var = resolve_aws_auth_env_var()
+            _region = resolve_bedrock_region()
+            _label = "AWS Bedrock".ljust(20)
+            print(f"  Checking AWS Bedrock...", end="", flush=True)
+            try:
+                import boto3
+                _br_client = boto3.client("bedrock", region_name=_region)
+                _br_resp = _br_client.list_foundation_models()
+                _model_count = len(_br_resp.get("modelSummaries", []))
+                print(f"\r  {color('✓', Colors.GREEN)} {_label} {color(f'({_auth_var}, {_region}, {_model_count} models)', Colors.DIM)}           ")
+            except ImportError:
+                print(f"\r  {color('⚠', Colors.YELLOW)} {_label} {color(f'(boto3 not installed — {sys.executable} -m pip install boto3)', Colors.DIM)}           ")
+                issues.append(f"Install boto3 for Bedrock: {sys.executable} -m pip install boto3")
+            except Exception as _e:
+                _err_name = type(_e).__name__
+                print(f"\r  {color('⚠', Colors.YELLOW)} {_label} {color(f'({_err_name}: {_e})', Colors.DIM)}           ")
+                issues.append(f"AWS Bedrock: {_err_name} — check IAM permissions for bedrock:ListFoundationModels")
+    except ImportError:
+        pass  # bedrock_adapter not available — skip silently
+
     # =========================================================================
     # Check: Submodules
     # =========================================================================