feat(tui): add /mouse [on|off|toggle] runtime slash command

Toggle SGR mouse tracking (DEC 1000/1002/1003/1006) at runtime without
restart or env-var spelunking. Fix path when a terminal doesn't honor raw
mode / no-echo and echoes mouse events as visible escape sequences (e.g.
`<35;111;133M` scrolling up the transcript on every mouse move).

- New `/mouse [on|off|toggle]` slash command (persists via config.set key=mouse
  → display.tui_mouse in ~/.hermes/config.yaml).
- New hermes-ink export `setAltScreenMouseTracking(enabled)` that writes
  ENABLE/DISABLE bytes and updates the instance flag without re-entering
  the alt-screen — so live toggles are flicker-free.
- `<AlternateScreen>` mouseTracking prop is frozen at initial value (from
  `HERMES_TUI_DISABLE_MOUSE` env); runtime state lives in `$uiState` and is
  applied via useEffect. Env-var opt-out wins over config so explicit
  HERMES_TUI_DISABLE_MOUSE=1 stays off regardless of persisted state.
- Server: folds `mouse` into the existing compact/statusbar branch in
  config.set/get, defaulting to on.

.github/actions/nix-setup/action.yml

@@ -0,0 +1,8 @@
+name: 'Setup Nix'
+description: 'Install Nix with DeterminateSystems and enable magic-nix-cache'
+
+runs:
+  using: composite
+  steps:
+    - uses: DeterminateSystems/nix-installer-action@ef8a148080ab6020fd15196c2084a2eea5ff2d25 # v22
+    - uses: DeterminateSystems/magic-nix-cache-action@565684385bcd71bad329742eefe8d12f2e765b39 # v13

.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/nix-lockfile-check.yml

@@ -0,0 +1,68 @@
+name: Nix Lockfile Check
+
+on:
+  pull_request:
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  pull-requests: write
+
+concurrency:
+  group: nix-lockfile-check-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  check:
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5  # v4
+
+      - uses: ./.github/actions/nix-setup
+
+      - name: Resolve head SHA
+        id: sha
+        shell: bash
+        run: |
+          FULL="${{ github.event.pull_request.head.sha || github.sha }}"
+          echo "full=$FULL" >> "$GITHUB_OUTPUT"
+          echo "short=${FULL:0:7}" >> "$GITHUB_OUTPUT"
+
+      - name: Check lockfile hashes
+        id: check
+        continue-on-error: true
+        env:
+          LINK_SHA: ${{ steps.sha.outputs.full }}
+        run: nix run .#fix-lockfiles -- --check
+
+      - name: Post sticky PR comment (stale)
+        if: steps.check.outputs.stale == 'true' && github.event_name == 'pull_request'
+        uses: marocchino/sticky-pull-request-comment@52423e01640425a022ef5fd42c6fb5f633a02728  # v2.9.1
+        with:
+          header: nix-lockfile-check
+          message: |
+            ### ⚠️ npm lockfile hash out of date
+
+            Checked against commit [`${{ steps.sha.outputs.short }}`](${{ github.server_url }}/${{ github.repository }}/commit/${{ steps.sha.outputs.full }}) (PR head at check time).
+
+            The `hash = "sha256-..."` line in these nix files no longer matches the committed `package-lock.json`:
+
+            ${{ steps.check.outputs.report }}
+
+            #### Apply the fix
+
+            - [ ] **Apply lockfile fix** — tick to push a commit with the correct hashes to this PR branch
+            - Or [run the Nix Lockfile Fix workflow](${{ github.server_url }}/${{ github.repository }}/actions/workflows/nix-lockfile-fix.yml) manually (pass PR `#${{ github.event.pull_request.number }}`)
+            - Or locally: `nix run .#fix-lockfiles -- --apply` and commit the diff
+
+      - name: Clear sticky PR comment (resolved)
+        if: steps.check.outputs.stale == 'false' && github.event_name == 'pull_request'
+        uses: marocchino/sticky-pull-request-comment@52423e01640425a022ef5fd42c6fb5f633a02728  # v2.9.1
+        with:
+          header: nix-lockfile-check
+          delete: true
+
+      - name: Fail if stale
+        if: steps.check.outputs.stale == 'true'
+        run: exit 1

.github/workflows/nix-lockfile-fix.yml

@@ -0,0 +1,149 @@
+name: Nix Lockfile Fix
+
+on:
+  workflow_dispatch:
+    inputs:
+      pr_number:
+        description: 'PR number to fix (leave empty to run on the selected branch)'
+        required: false
+        type: string
+  issue_comment:
+    types: [edited]
+
+permissions:
+  contents: write
+  pull-requests: write
+
+concurrency:
+  group: nix-lockfile-fix-${{ github.event.issue.number || github.event.inputs.pr_number || github.ref }}
+  cancel-in-progress: false
+
+jobs:
+  fix:
+    # Run on manual dispatch OR when a task-list checkbox in the sticky
+    # lockfile-check comment flips from `[ ]` to `[x]`.
+    if: |
+      github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'issue_comment'
+       && github.event.issue.pull_request != null
+       && contains(github.event.comment.body, '[x] **Apply lockfile fix**')
+       && !contains(github.event.changes.body.from, '[x] **Apply lockfile fix**'))
+    runs-on: ubuntu-latest
+    timeout-minutes: 25
+    steps:
+      - name: Authorize & resolve PR
+        id: resolve
+        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea  # v7.0.1
+        with:
+          script: |
+            // 1. Verify the actor has write access — applies to both checkbox
+            //    clicks and manual dispatch.
+            const { data: perm } =
+              await github.rest.repos.getCollaboratorPermissionLevel({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                username: context.actor,
+              });
+            if (!['admin', 'write', 'maintain'].includes(perm.permission)) {
+              core.setFailed(
+                `${context.actor} lacks write access (has: ${perm.permission})`
+              );
+              return;
+            }
+
+            // 2. Resolve which ref to check out.
+            let prNumber = '';
+            if (context.eventName === 'issue_comment') {
+              prNumber = String(context.payload.issue.number);
+            } else if (context.eventName === 'workflow_dispatch') {
+              prNumber = context.payload.inputs.pr_number || '';
+            }
+
+            if (!prNumber) {
+              core.setOutput('ref', context.ref.replace(/^refs\/heads\//, ''));
+              core.setOutput('repo', context.repo.repo);
+              core.setOutput('owner', context.repo.owner);
+              core.setOutput('pr', '');
+              return;
+            }
+
+            const { data: pr } = await github.rest.pulls.get({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              pull_number: Number(prNumber),
+            });
+            core.setOutput('ref', pr.head.ref);
+            core.setOutput('repo', pr.head.repo.name);
+            core.setOutput('owner', pr.head.repo.owner.login);
+            core.setOutput('pr', String(pr.number));
+
+      # Wipe the sticky lockfile-check comment to a "running" state as soon
+      # as the job is authorized, so the user sees their click was picked up
+      # before the ~minute of nix build work.
+      - name: Mark sticky as running
+        if: steps.resolve.outputs.pr != ''
+        uses: marocchino/sticky-pull-request-comment@52423e01640425a022ef5fd42c6fb5f633a02728  # v2.9.1
+        with:
+          header: nix-lockfile-check
+          number: ${{ steps.resolve.outputs.pr }}
+          message: |
+            ### 🔄 Applying lockfile fix…
+
+            Triggered by @${{ github.actor }} — [workflow run](${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}).
+
+      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5  # v4
+        with:
+          repository: ${{ steps.resolve.outputs.owner }}/${{ steps.resolve.outputs.repo }}
+          ref: ${{ steps.resolve.outputs.ref }}
+          token: ${{ secrets.GITHUB_TOKEN }}
+          fetch-depth: 0
+
+      - uses: ./.github/actions/nix-setup
+
+      - name: Apply lockfile hashes
+        id: apply
+        run: nix run .#fix-lockfiles -- --apply
+
+      - name: Commit & push
+        if: steps.apply.outputs.changed == 'true'
+        shell: bash
+        run: |
+          set -euo pipefail
+          git config user.name 'github-actions[bot]'
+          git config user.email '41898282+github-actions[bot]@users.noreply.github.com'
+          git add nix/tui.nix nix/web.nix
+          git commit -m "fix(nix): refresh npm lockfile hashes"
+          git push
+
+      - name: Update sticky (applied)
+        if: steps.apply.outputs.changed == 'true' && steps.resolve.outputs.pr != ''
+        uses: marocchino/sticky-pull-request-comment@52423e01640425a022ef5fd42c6fb5f633a02728  # v2.9.1
+        with:
+          header: nix-lockfile-check
+          number: ${{ steps.resolve.outputs.pr }}
+          message: |
+            ### ✅ Lockfile fix applied
+
+            Pushed a commit refreshing the npm lockfile hashes — [workflow run](${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}).
+
+      - name: Update sticky (already current)
+        if: steps.apply.outputs.changed == 'false' && steps.resolve.outputs.pr != ''
+        uses: marocchino/sticky-pull-request-comment@52423e01640425a022ef5fd42c6fb5f633a02728  # v2.9.1
+        with:
+          header: nix-lockfile-check
+          number: ${{ steps.resolve.outputs.pr }}
+          message: |
+            ### ✅ Lockfile hashes already current
+
+            Nothing to commit — [workflow run](${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}).
+
+      - name: Update sticky (failed)
+        if: failure() && steps.resolve.outputs.pr != ''
+        uses: marocchino/sticky-pull-request-comment@52423e01640425a022ef5fd42c6fb5f633a02728  # v2.9.1
+        with:
+          header: nix-lockfile-check
+          number: ${{ steps.resolve.outputs.pr }}
+          message: |
+            ### ❌ Lockfile fix failed
+
+            See the [workflow run](${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}) for logs.

.github/workflows/nix.yml

@@ -4,15 +4,6 @@ on:
   push:
     branches: [main]
   pull_request:
-    paths:
-      - 'flake.nix'
-      - 'flake.lock'
-      - 'nix/**'
-      - 'pyproject.toml'
-      - 'uv.lock'
-      - 'hermes_cli/**'
-      - 'run_agent.py'
-      - 'acp_adapter/**'
 
 permissions:
   contents: read
@@ -29,9 +20,8 @@ jobs:
     runs-on: ${{ matrix.os }}
     timeout-minutes: 30
     steps:
-      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5  # v4
-      - uses: DeterminateSystems/nix-installer-action@ef8a148080ab6020fd15196c2084a2eea5ff2d25  # v22
-      - uses: DeterminateSystems/magic-nix-cache-action@565684385bcd71bad329742eefe8d12f2e765b39  # v13
+      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
+      - uses: ./.github/actions/nix-setup
       - name: Check flake
         if: runner.os == 'Linux'
         run: nix flake check --print-build-logs

.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

@@ -54,6 +54,11 @@ 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/

AGENTS.md

@@ -566,3 +566,52 @@ 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.
+
+### Don't write change-detector tests
+
+A test is a **change-detector** if it fails whenever data that is **expected
+to change** gets updated — model catalogs, config version numbers,
+enumeration counts, hardcoded lists of provider models. These tests add no
+behavioral coverage; they just guarantee that routine source updates break
+CI and cost engineering time to "fix."
+
+**Do not write:**
+
+```python
+# catalog snapshot — breaks every model release
+assert "gemini-2.5-pro" in _PROVIDER_MODELS["gemini"]
+assert "MiniMax-M2.7" in models
+
+# config version literal — breaks every schema bump
+assert DEFAULT_CONFIG["_config_version"] == 21
+
+# enumeration count — breaks every time a skill/provider is added
+assert len(_PROVIDER_MODELS["huggingface"]) == 8
+```
+
+**Do write:**
+
+```python
+# behavior: does the catalog plumbing work at all?
+assert "gemini" in _PROVIDER_MODELS
+assert len(_PROVIDER_MODELS["gemini"]) >= 1
+
+# behavior: does migration bump the user's version to current latest?
+assert raw["_config_version"] == DEFAULT_CONFIG["_config_version"]
+
+# invariant: no plan-only model leaks into the legacy list
+assert not (set(moonshot_models) & coding_plan_only_models)
+
+# invariant: every model in the catalog has a context-length entry
+for m in _PROVIDER_MODELS["huggingface"]:
+    assert m.lower() in DEFAULT_CONTEXT_LENGTHS_LOWER
+```
+
+The rule: if the test reads like a snapshot of current data, delete it. If
+it reads like a contract about how two pieces of data must relate, keep it.
+When a PR adds a new provider/model and you want a test, make the test
+assert the relationship (e.g. "catalog entries all have context lengths"),
+not the specific names.
+
+Reviewers should reject new change-detector tests; authors should convert
+them into invariants before re-requesting review.

Dockerfile

@@ -27,12 +27,10 @@ WORKDIR /opt/hermes
 # 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 scripts/whatsapp-bridge && npm install --prefer-offline --no-audit) && \
     (cd web && npm install --prefer-offline --no-audit) && \
     npm cache clean --force
 

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/permissions.py

@@ -63,6 +63,9 @@ def make_approval_callback(
             logger.warning("Permission request timed out or failed: %s", exc)
             return "deny"
 
+        if response is None:
+            return "deny"
+
         outcome = response.outcome
         if isinstance(outcome, AllowedOutcome):
             option_id = outcome.option_id

acp_adapter/server.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import asyncio
 import logging
+import os
 from collections import defaultdict, deque
 from concurrent.futures import ThreadPoolExecutor
 from typing import Any, Deque, Optional
@@ -51,7 +52,7 @@ try:
 except ImportError:
     from acp.schema import AuthMethod as AuthMethodAgent  # type: ignore[attr-defined]
 
-from acp_adapter.auth import detect_provider, has_provider
+from acp_adapter.auth import detect_provider
 from acp_adapter.events import (
     make_message_cb,
     make_step_cb,
@@ -71,6 +72,11 @@ except Exception:
 # Thread pool for running AIAgent (synchronous) in parallel.
 _executor = ThreadPoolExecutor(max_workers=4, thread_name_prefix="acp-agent")
 
+# Server-side page size for list_sessions. The ACP ListSessionsRequest schema
+# does not expose a client-side limit, so this is a fixed cap that clients
+# paginate against using `cursor` / `next_cursor`.
+_LIST_SESSIONS_PAGE_SIZE = 50
+
 
 def _extract_text(
     prompt: list[
@@ -351,9 +357,18 @@ class HermesACPAgent(acp.Agent):
         )
 
     async def authenticate(self, method_id: str, **kwargs: Any) -> AuthenticateResponse | None:
-        if has_provider():
-            return AuthenticateResponse()
-        return None
+        # Only accept authenticate() calls whose method_id matches the
+        # provider we advertised in initialize(). Without this check,
+        # authenticate() would acknowledge any method_id as long as the
+        # server has provider credentials configured — harmless under
+        # Hermes' threat model (ACP is stdio-only, local-trust), but poor
+        # API hygiene and confusing if ACP ever grows multi-method auth.
+        provider = detect_provider()
+        if not provider:
+            return None
+        if not isinstance(method_id, str) or method_id.strip().lower() != provider:
+            return None
+        return AuthenticateResponse()
 
     # ---- Session management -------------------------------------------------
 
@@ -437,7 +452,28 @@ class HermesACPAgent(acp.Agent):
         cwd: str | None = None,
         **kwargs: Any,
     ) -> ListSessionsResponse:
+        """List ACP sessions with optional ``cwd`` filtering and cursor pagination.
+
+        ``cwd`` is passed through to ``SessionManager.list_sessions`` which already
+        normalizes and filters by working directory. ``cursor`` is a ``session_id``
+        previously returned as ``next_cursor``; results resume after that entry.
+        Server-side page size is capped at ``_LIST_SESSIONS_PAGE_SIZE``; when more
+        results remain, ``next_cursor`` is set to the last returned ``session_id``.
+        """
         infos = self.session_manager.list_sessions(cwd=cwd)
+
+        if cursor:
+            for idx, s in enumerate(infos):
+                if s["session_id"] == cursor:
+                    infos = infos[idx + 1:]
+                    break
+            else:
+                # Unknown cursor -> empty page (do not fall back to full list).
+                infos = []
+
+        has_more = len(infos) > _LIST_SESSIONS_PAGE_SIZE
+        infos = infos[:_LIST_SESSIONS_PAGE_SIZE]
+
         sessions = []
         for s in infos:
             updated_at = s.get("updated_at")
@@ -451,7 +487,9 @@ class HermesACPAgent(acp.Agent):
                     updated_at=updated_at,
                 )
             )
-        return ListSessionsResponse(sessions=sessions)
+
+        next_cursor = sessions[-1].session_id if has_more and sessions else None
+        return ListSessionsResponse(sessions=sessions, next_cursor=next_cursor)
 
     # ---- Prompt (core) ------------------------------------------------------
 
@@ -517,15 +555,32 @@ class HermesACPAgent(acp.Agent):
         agent.step_callback = step_cb
         agent.message_callback = message_cb
 
-        if approval_cb:
-            try:
-                from tools import terminal_tool as _terminal_tool
-                previous_approval_cb = getattr(_terminal_tool, "_approval_callback", None)
-                _terminal_tool.set_approval_callback(approval_cb)
-            except Exception:
-                logger.debug("Could not set ACP approval callback", exc_info=True)
+        # Approval callback is per-thread (thread-local, GHSA-qg5c-hvr5-hjgr).
+        # Set it INSIDE _run_agent so the TLS write happens in the executor
+        # thread — setting it here would write to the event-loop thread's TLS,
+        # not the executor's. Also set HERMES_INTERACTIVE so approval.py
+        # takes the CLI-interactive path (which calls the registered
+        # callback via prompt_dangerous_approval) instead of the
+        # non-interactive auto-approve branch (GHSA-96vc-wcxf-jjff).
+        # ACP's conn.request_permission maps cleanly to the interactive
+        # callback shape — not the gateway-queue HERMES_EXEC_ASK path,
+        # which requires a notify_cb registered in _gateway_notify_cbs.
+        previous_approval_cb = None
+        previous_interactive = None
 
         def _run_agent() -> dict:
+            nonlocal previous_approval_cb, previous_interactive
+            if approval_cb:
+                try:
+                    from tools import terminal_tool as _terminal_tool
+                    previous_approval_cb = _terminal_tool._get_approval_callback()
+                    _terminal_tool.set_approval_callback(approval_cb)
+                except Exception:
+                    logger.debug("Could not set ACP approval callback", exc_info=True)
+            # Signal to tools.approval that we have an interactive callback
+            # and the non-interactive auto-approve path must not fire.
+            previous_interactive = os.environ.get("HERMES_INTERACTIVE")
+            os.environ["HERMES_INTERACTIVE"] = "1"
             try:
                 result = agent.run_conversation(
                     user_message=user_text,
@@ -537,6 +592,11 @@ class HermesACPAgent(acp.Agent):
                 logger.exception("Agent error in session %s", session_id)
                 return {"final_response": f"Error: {e}", "messages": state.history}
             finally:
+                # Restore HERMES_INTERACTIVE.
+                if previous_interactive is None:
+                    os.environ.pop("HERMES_INTERACTIVE", None)
+                else:
+                    os.environ["HERMES_INTERACTIVE"] = previous_interactive
                 if approval_cb:
                     try:
                         from tools import terminal_tool as _terminal_tool
@@ -613,8 +673,8 @@ class HermesACPAgent(acp.Agent):
             await self._conn.session_update(
                 session_id=session_id,
                 update=AvailableCommandsUpdate(
-                    sessionUpdate="available_commands_update",
-                    availableCommands=self._available_commands(),
+                    session_update="available_commands_update",
+                    available_commands=self._available_commands(),
                 ),
             )
         except Exception:

agent/account_usage.py

@@ -0,0 +1,326 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from typing import Any, Optional
+
+import httpx
+
+from agent.anthropic_adapter import _is_oauth_token, resolve_anthropic_token
+from hermes_cli.auth import _read_codex_tokens, resolve_codex_runtime_credentials
+from hermes_cli.runtime_provider import resolve_runtime_provider
+
+
+def _utc_now() -> datetime:
+    return datetime.now(timezone.utc)
+
+
+@dataclass(frozen=True)
+class AccountUsageWindow:
+    label: str
+    used_percent: Optional[float] = None
+    reset_at: Optional[datetime] = None
+    detail: Optional[str] = None
+
+
+@dataclass(frozen=True)
+class AccountUsageSnapshot:
+    provider: str
+    source: str
+    fetched_at: datetime
+    title: str = "Account limits"
+    plan: Optional[str] = None
+    windows: tuple[AccountUsageWindow, ...] = ()
+    details: tuple[str, ...] = ()
+    unavailable_reason: Optional[str] = None
+
+    @property
+    def available(self) -> bool:
+        return bool(self.windows or self.details) and not self.unavailable_reason
+
+
+def _title_case_slug(value: Optional[str]) -> Optional[str]:
+    cleaned = str(value or "").strip()
+    if not cleaned:
+        return None
+    return cleaned.replace("_", " ").replace("-", " ").title()
+
+
+def _parse_dt(value: Any) -> Optional[datetime]:
+    if value in (None, ""):
+        return None
+    if isinstance(value, (int, float)):
+        return datetime.fromtimestamp(float(value), tz=timezone.utc)
+    if isinstance(value, str):
+        text = value.strip()
+        if not text:
+            return None
+        if text.endswith("Z"):
+            text = text[:-1] + "+00:00"
+        try:
+            dt = datetime.fromisoformat(text)
+            return dt if dt.tzinfo else dt.replace(tzinfo=timezone.utc)
+        except ValueError:
+            return None
+    return None
+
+
+def _format_reset(dt: Optional[datetime]) -> str:
+    if not dt:
+        return "unknown"
+    local_dt = dt.astimezone()
+    delta = dt - _utc_now()
+    total_seconds = int(delta.total_seconds())
+    if total_seconds <= 0:
+        return f"now ({local_dt.strftime('%Y-%m-%d %H:%M %Z')})"
+    hours, rem = divmod(total_seconds, 3600)
+    minutes = rem // 60
+    if hours >= 24:
+        days, hours = divmod(hours, 24)
+        rel = f"in {days}d {hours}h"
+    elif hours > 0:
+        rel = f"in {hours}h {minutes}m"
+    else:
+        rel = f"in {minutes}m"
+    return f"{rel} ({local_dt.strftime('%Y-%m-%d %H:%M %Z')})"
+
+
+def render_account_usage_lines(snapshot: Optional[AccountUsageSnapshot], *, markdown: bool = False) -> list[str]:
+    if not snapshot:
+        return []
+    header = f"📈 {'**' if markdown else ''}{snapshot.title}{'**' if markdown else ''}"
+    lines = [header]
+    if snapshot.plan:
+        lines.append(f"Provider: {snapshot.provider} ({snapshot.plan})")
+    else:
+        lines.append(f"Provider: {snapshot.provider}")
+    for window in snapshot.windows:
+        if window.used_percent is None:
+            base = f"{window.label}: unavailable"
+        else:
+            remaining = max(0, round(100 - float(window.used_percent)))
+            used = max(0, round(float(window.used_percent)))
+            base = f"{window.label}: {remaining}% remaining ({used}% used)"
+        if window.reset_at:
+            base += f" • resets {_format_reset(window.reset_at)}"
+        elif window.detail:
+            base += f" • {window.detail}"
+        lines.append(base)
+    for detail in snapshot.details:
+        lines.append(detail)
+    if snapshot.unavailable_reason:
+        lines.append(f"Unavailable: {snapshot.unavailable_reason}")
+    return lines
+
+
+def _resolve_codex_usage_url(base_url: str) -> str:
+    normalized = (base_url or "").strip().rstrip("/")
+    if not normalized:
+        normalized = "https://chatgpt.com/backend-api/codex"
+    if normalized.endswith("/codex"):
+        normalized = normalized[: -len("/codex")]
+    if "/backend-api" in normalized:
+        return normalized + "/wham/usage"
+    return normalized + "/api/codex/usage"
+
+
+def _fetch_codex_account_usage() -> Optional[AccountUsageSnapshot]:
+    creds = resolve_codex_runtime_credentials(refresh_if_expiring=True)
+    token_data = _read_codex_tokens()
+    tokens = token_data.get("tokens") or {}
+    account_id = str(tokens.get("account_id", "") or "").strip() or None
+    headers = {
+        "Authorization": f"Bearer {creds['api_key']}",
+        "Accept": "application/json",
+        "User-Agent": "codex-cli",
+    }
+    if account_id:
+        headers["ChatGPT-Account-Id"] = account_id
+    with httpx.Client(timeout=15.0) as client:
+        response = client.get(_resolve_codex_usage_url(creds.get("base_url", "")), headers=headers)
+        response.raise_for_status()
+    payload = response.json() or {}
+    rate_limit = payload.get("rate_limit") or {}
+    windows: list[AccountUsageWindow] = []
+    for key, label in (("primary_window", "Session"), ("secondary_window", "Weekly")):
+        window = rate_limit.get(key) or {}
+        used = window.get("used_percent")
+        if used is None:
+            continue
+        windows.append(
+            AccountUsageWindow(
+                label=label,
+                used_percent=float(used),
+                reset_at=_parse_dt(window.get("reset_at")),
+            )
+        )
+    details: list[str] = []
+    credits = payload.get("credits") or {}
+    if credits.get("has_credits"):
+        balance = credits.get("balance")
+        if isinstance(balance, (int, float)):
+            details.append(f"Credits balance: ${float(balance):.2f}")
+        elif credits.get("unlimited"):
+            details.append("Credits balance: unlimited")
+    return AccountUsageSnapshot(
+        provider="openai-codex",
+        source="usage_api",
+        fetched_at=_utc_now(),
+        plan=_title_case_slug(payload.get("plan_type")),
+        windows=tuple(windows),
+        details=tuple(details),
+    )
+
+
+def _fetch_anthropic_account_usage() -> Optional[AccountUsageSnapshot]:
+    token = (resolve_anthropic_token() or "").strip()
+    if not token:
+        return None
+    if not _is_oauth_token(token):
+        return AccountUsageSnapshot(
+            provider="anthropic",
+            source="oauth_usage_api",
+            fetched_at=_utc_now(),
+            unavailable_reason="Anthropic account limits are only available for OAuth-backed Claude accounts.",
+        )
+    headers = {
+        "Authorization": f"Bearer {token}",
+        "Accept": "application/json",
+        "Content-Type": "application/json",
+        "anthropic-beta": "oauth-2025-04-20",
+        "User-Agent": "claude-code/2.1.0",
+    }
+    with httpx.Client(timeout=15.0) as client:
+        response = client.get("https://api.anthropic.com/api/oauth/usage", headers=headers)
+        response.raise_for_status()
+    payload = response.json() or {}
+    windows: list[AccountUsageWindow] = []
+    mapping = (
+        ("five_hour", "Current session"),
+        ("seven_day", "Current week"),
+        ("seven_day_opus", "Opus week"),
+        ("seven_day_sonnet", "Sonnet week"),
+    )
+    for key, label in mapping:
+        window = payload.get(key) or {}
+        util = window.get("utilization")
+        if util is None:
+            continue
+        used = float(util) * 100 if float(util) <= 1 else float(util)
+        windows.append(
+            AccountUsageWindow(
+                label=label,
+                used_percent=used,
+                reset_at=_parse_dt(window.get("resets_at")),
+            )
+        )
+    details: list[str] = []
+    extra = payload.get("extra_usage") or {}
+    if extra.get("is_enabled"):
+        used_credits = extra.get("used_credits")
+        monthly_limit = extra.get("monthly_limit")
+        currency = extra.get("currency") or "USD"
+        if isinstance(used_credits, (int, float)) and isinstance(monthly_limit, (int, float)):
+            details.append(
+                f"Extra usage: {used_credits:.2f} / {monthly_limit:.2f} {currency}"
+            )
+    return AccountUsageSnapshot(
+        provider="anthropic",
+        source="oauth_usage_api",
+        fetched_at=_utc_now(),
+        windows=tuple(windows),
+        details=tuple(details),
+    )
+
+
+def _fetch_openrouter_account_usage(base_url: Optional[str], api_key: Optional[str]) -> Optional[AccountUsageSnapshot]:
+    runtime = resolve_runtime_provider(
+        requested="openrouter",
+        explicit_base_url=base_url,
+        explicit_api_key=api_key,
+    )
+    token = str(runtime.get("api_key", "") or "").strip()
+    if not token:
+        return None
+    normalized = str(runtime.get("base_url", "") or "").rstrip("/")
+    credits_url = f"{normalized}/credits"
+    key_url = f"{normalized}/key"
+    headers = {
+        "Authorization": f"Bearer {token}",
+        "Accept": "application/json",
+    }
+    with httpx.Client(timeout=10.0) as client:
+        credits_resp = client.get(credits_url, headers=headers)
+        credits_resp.raise_for_status()
+        credits = (credits_resp.json() or {}).get("data") or {}
+        try:
+            key_resp = client.get(key_url, headers=headers)
+            key_resp.raise_for_status()
+            key_data = (key_resp.json() or {}).get("data") or {}
+        except Exception:
+            key_data = {}
+    total_credits = float(credits.get("total_credits") or 0.0)
+    total_usage = float(credits.get("total_usage") or 0.0)
+    details = [f"Credits balance: ${max(0.0, total_credits - total_usage):.2f}"]
+    windows: list[AccountUsageWindow] = []
+    limit = key_data.get("limit")
+    limit_remaining = key_data.get("limit_remaining")
+    limit_reset = str(key_data.get("limit_reset") or "").strip()
+    usage = key_data.get("usage")
+    if (
+        isinstance(limit, (int, float))
+        and float(limit) > 0
+        and isinstance(limit_remaining, (int, float))
+        and 0 <= float(limit_remaining) <= float(limit)
+    ):
+        limit_value = float(limit)
+        remaining_value = float(limit_remaining)
+        used_percent = ((limit_value - remaining_value) / limit_value) * 100
+        detail_parts = [f"${remaining_value:.2f} of ${limit_value:.2f} remaining"]
+        if limit_reset:
+            detail_parts.append(f"resets {limit_reset}")
+        windows.append(
+            AccountUsageWindow(
+                label="API key quota",
+                used_percent=used_percent,
+                detail=" • ".join(detail_parts),
+            )
+        )
+    if isinstance(usage, (int, float)):
+        usage_parts = [f"API key usage: ${float(usage):.2f} total"]
+        for value, label in (
+            (key_data.get("usage_daily"), "today"),
+            (key_data.get("usage_weekly"), "this week"),
+            (key_data.get("usage_monthly"), "this month"),
+        ):
+            if isinstance(value, (int, float)) and float(value) > 0:
+                usage_parts.append(f"${float(value):.2f} {label}")
+        details.append(" • ".join(usage_parts))
+    return AccountUsageSnapshot(
+        provider="openrouter",
+        source="credits_api",
+        fetched_at=_utc_now(),
+        windows=tuple(windows),
+        details=tuple(details),
+    )
+
+
+def fetch_account_usage(
+    provider: Optional[str],
+    *,
+    base_url: Optional[str] = None,
+    api_key: Optional[str] = None,
+) -> Optional[AccountUsageSnapshot]:
+    normalized = str(provider or "").strip().lower()
+    if normalized in {"", "auto", "custom"}:
+        return None
+    try:
+        if normalized == "openai-codex":
+            return _fetch_codex_account_usage()
+        if normalized == "anthropic":
+            return _fetch_anthropic_account_usage()
+        if normalized == "openrouter":
+            return _fetch_openrouter_account_usage(base_url, api_key)
+    except Exception:
+        return None
+    return None

agent/anthropic_adapter.py

@@ -19,6 +19,7 @@ from pathlib import Path
 from hermes_constants import get_hermes_home
 from types import SimpleNamespace
 from typing import Any, Dict, List, Optional, Tuple
+from utils import normalize_proxy_env_vars
 
 try:
     import anthropic as _anthropic_sdk
@@ -292,9 +293,15 @@ 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:
@@ -302,11 +309,15 @@ def build_anthropic_client(api_key: str, base_url: str = None):
             "The 'anthropic' package is required for the Anthropic provider. "
             "Install it with: pip install 'anthropic>=0.39.0'"
         )
+
+    normalize_proxy_env_vars()
+
     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
@@ -1518,3 +1529,42 @@ def normalize_anthropic_response(
         ),
         finish_reason,
     )
+
+
+def normalize_anthropic_response_v2(
+    response,
+    strip_tool_prefix: bool = False,
+) -> "NormalizedResponse":
+    """Normalize Anthropic response to NormalizedResponse.
+
+    Wraps the existing normalize_anthropic_response() and maps its output
+    to the shared transport types.  This allows incremental migration —
+    one call site at a time — without changing the original function.
+    """
+    from agent.transports.types import NormalizedResponse, build_tool_call
+
+    assistant_msg, finish_reason = normalize_anthropic_response(response, strip_tool_prefix)
+
+    tool_calls = None
+    if assistant_msg.tool_calls:
+        tool_calls = [
+            build_tool_call(
+                id=tc.id,
+                name=tc.function.name,
+                arguments=tc.function.arguments,
+            )
+            for tc in assistant_msg.tool_calls
+        ]
+
+    provider_data = {}
+    if getattr(assistant_msg, "reasoning_details", None):
+        provider_data["reasoning_details"] = assistant_msg.reasoning_details
+
+    return NormalizedResponse(
+        content=assistant_msg.content,
+        tool_calls=tool_calls,
+        finish_reason=finish_reason,
+        reasoning=getattr(assistant_msg, "reasoning", None),
+        usage=None,  # Anthropic usage is on the raw response, not the normaliser
+        provider_data=provider_data or None,
+    )

agent/auxiliary_client.py

@@ -48,6 +48,7 @@ from openai import OpenAI
 from agent.credential_pool import load_pool
 from hermes_cli.config import get_hermes_home
 from hermes_constants import OPENROUTER_BASE_URL
+from utils import base_url_host_matches, base_url_hostname, normalize_proxy_env_vars
 
 logger = logging.getLogger(__name__)
 
@@ -95,51 +96,37 @@ def _normalize_aux_provider(provider: Optional[str]) -> str:
     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",
-})
+# Sentinel: when returned by _fixed_temperature_for_model(), callers must
+# strip the ``temperature`` key from API kwargs entirely so the provider's
+# server-side default applies.  Kimi/Moonshot models manage temperature
+# internally — sending *any* value (even the "correct" one) can conflict
+# with gateway-side mode selection (thinking → 1.0, non-thinking → 0.6).
+OMIT_TEMPERATURE: object = object()
 
 
-def _fixed_temperature_for_model(model: Optional[str]) -> Optional[float]:
-    """Return a required temperature override for models with strict contracts.
+def _is_kimi_model(model: Optional[str]) -> bool:
+    """True for any Kimi / Moonshot model that manages temperature server-side."""
+    bare = (model or "").strip().lower().rsplit("/", 1)[-1]
+    return bare.startswith("kimi-") or bare == "kimi"
 
-    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.
 
-    Returns ``None`` for every other model, including ``kimi-k2-instruct*``
-    which is the separate non-coding K2 family with variable temperature.
+def _fixed_temperature_for_model(
+    model: Optional[str],
+    base_url: Optional[str] = None,
+) -> "Optional[float] | object":
+    """Return a temperature directive for models with strict contracts.
+
+    Returns:
+        ``OMIT_TEMPERATURE`` — caller must remove the ``temperature`` key so the
+            provider chooses its own default.  Used for all Kimi / Moonshot
+            models whose gateway selects temperature server-side.
+        ``float`` — a specific value the caller must use (reserved for future
+            models with fixed-temperature contracts).
+        ``None`` — no override; caller should use its own default.
     """
-    normalized = (model or "").strip().lower()
-    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
-    bare = normalized.rsplit("/", 1)[-1]
-    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
+    if _is_kimi_model(model):
+        logger.debug("Omitting temperature for Kimi model %r (server-managed)", model)
+        return OMIT_TEMPERATURE
     return None
 
 # Default auxiliary models for direct API-key providers (cheap/fast for side tasks)
@@ -174,6 +161,16 @@ _OR_HEADERS = {
     "X-OpenRouter-Categories": "productivity,cli-agent",
 }
 
+# Vercel AI Gateway app attribution headers. HTTP-Referer maps to
+# referrerUrl and X-Title maps to appName in the gateway's analytics.
+from hermes_cli import __version__ as _HERMES_VERSION
+
+_AI_GATEWAY_HEADERS = {
+    "HTTP-Referer": "https://hermes-agent.nousresearch.com",
+    "X-Title": "Hermes Agent",
+    "User-Agent": f"HermesAgent/{_HERMES_VERSION}",
+}
+
 # Nous Portal extra_body for product attribution.
 # Callers should pass this as extra_body in chat.completions.create()
 # when the auxiliary client is backed by Nous Portal.
@@ -200,6 +197,45 @@ _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.
 
@@ -692,6 +728,33 @@ def _nous_base_url() -> str:
     return os.getenv("NOUS_INFERENCE_BASE_URL", _NOUS_DEFAULT_BASE_URL)
 
 
+def _resolve_nous_runtime_api(*, force_refresh: bool = False) -> Optional[tuple[str, str]]:
+    """Return fresh Nous runtime credentials when available.
+
+    This mirrors the main agent's 401 recovery path and keeps auxiliary
+    clients aligned with the singleton auth store + mint flow instead of
+    relying only on whatever raw tokens happen to be sitting in auth.json
+    or the credential pool.
+    """
+    try:
+        from hermes_cli.auth import resolve_nous_runtime_credentials
+
+        creds = resolve_nous_runtime_credentials(
+            min_key_ttl_seconds=max(60, int(os.getenv("HERMES_NOUS_MIN_KEY_TTL_SECONDS", "1800"))),
+            timeout_seconds=float(os.getenv("HERMES_NOUS_TIMEOUT_SECONDS", "15")),
+            force_mint=force_refresh,
+        )
+    except Exception as exc:
+        logger.debug("Auxiliary Nous runtime credential resolution failed: %s", exc)
+        return None
+
+    api_key = str(creds.get("api_key") or "").strip()
+    base_url = str(creds.get("base_url") or "").strip().rstrip("/")
+    if not api_key or not base_url:
+        return None
+    return api_key, base_url
+
+
 def _read_codex_access_token() -> Optional[str]:
     """Read a valid, non-expired Codex OAuth access token from Hermes auth store.
 
@@ -775,10 +838,15 @@ 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():
+            if base_url_host_matches(base_url, "api.kimi.com"):
                 extra["default_headers"] = {"User-Agent": "KimiCLI/1.30.0"}
-            elif "api.githubcopilot.com" in base_url.lower():
+            elif base_url_host_matches(base_url, "api.githubcopilot.com"):
                 from hermes_cli.models import copilot_default_headers
 
                 extra["default_headers"] = copilot_default_headers()
@@ -796,10 +864,15 @@ 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():
+        if base_url_host_matches(base_url, "api.kimi.com"):
             extra["default_headers"] = {"User-Agent": "KimiCLI/1.30.0"}
-        elif "api.githubcopilot.com" in base_url.lower():
+        elif base_url_host_matches(base_url, "api.githubcopilot.com"):
             from hermes_cli.models import copilot_default_headers
 
             extra["default_headers"] = copilot_default_headers()
@@ -848,7 +921,8 @@ def _try_nous(vision: bool = False) -> Tuple[Optional[OpenAI], Optional[str]]:
         pass
 
     nous = _read_nous_auth()
-    if not nous:
+    runtime = _resolve_nous_runtime_api(force_refresh=False)
+    if runtime is None and not nous:
         return None, None
     global auxiliary_is_nous
     auxiliary_is_nous = True
@@ -859,6 +933,8 @@ def _try_nous(vision: bool = False) -> Tuple[Optional[OpenAI], Optional[str]]:
         model = _NOUS_MODEL
     # Free-tier users can't use paid auxiliary models — use the free
     # models instead: mimo-v2-omni for vision, mimo-v2-pro for text tasks.
+    # Paid accounts keep their tier-appropriate models: gemini-3-flash-preview
+    # for both text and vision tasks.
     try:
         from hermes_cli.models import check_nous_free_tier
         if check_nous_free_tier():
@@ -867,10 +943,15 @@ def _try_nous(vision: bool = False) -> Tuple[Optional[OpenAI], Optional[str]]:
                          model, "vision" if vision else "text")
     except Exception:
         pass
+    if runtime is not None:
+        api_key, base_url = runtime
+    else:
+        api_key = _nous_api_key(nous or {})
+        base_url = str((nous or {}).get("inference_base_url") or _nous_base_url()).rstrip("/")
     return (
         OpenAI(
-            api_key=_nous_api_key(nous),
-            base_url=str(nous.get("inference_base_url") or _nous_base_url()).rstrip("/"),
+            api_key=api_key,
+            base_url=base_url,
         ),
         model,
     )
@@ -948,7 +1029,7 @@ def _resolve_custom_runtime() -> Tuple[Optional[str], Optional[str], Optional[st
         return None, None, None
 
     custom_base = custom_base.strip().rstrip("/")
-    if "openrouter.ai" in custom_base.lower():
+    if base_url_host_matches(custom_base, "openrouter.ai"):
         # requested='custom' falls back to OpenRouter when no custom endpoint is
         # configured. Treat that as "no custom endpoint" for auxiliary routing.
         return None, None, None
@@ -982,6 +1063,8 @@ def _validate_proxy_env_urls() -> None:
     """
     from urllib.parse import urlparse
 
+    normalize_proxy_env_vars()
+
     for key in ("HTTPS_PROXY", "HTTP_PROXY", "ALL_PROXY",
                 "https_proxy", "http_proxy", "all_proxy"):
         value = str(os.environ.get(key) or "").strip()
@@ -1016,7 +1099,7 @@ def _validate_base_url(base_url: str) -> None:
         ) from exc
 
 
-def _try_custom_endpoint() -> Tuple[Optional[OpenAI], Optional[str]]:
+def _try_custom_endpoint() -> Tuple[Optional[Any], Optional[str]]:
     runtime = _resolve_custom_runtime()
     if len(runtime) == 2:
         custom_base, custom_key = runtime
@@ -1032,6 +1115,23 @@ 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
 
 
@@ -1052,7 +1152,11 @@ 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
 
 
@@ -1191,6 +1295,15 @@ def _is_connection_error(exc: Exception) -> bool:
     return False
 
 
+def _is_auth_error(exc: Exception) -> bool:
+    """Detect auth failures that should trigger provider-specific refresh."""
+    status = getattr(exc, "status_code", None)
+    if status == 401:
+        return True
+    err_lower = str(exc).lower()
+    return "error code: 401" in err_lower or "authenticationerror" in type(exc).__name__.lower()
+
+
 def _try_payment_fallback(
     failed_provider: str,
     task: str = None,
@@ -1348,6 +1461,13 @@ 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):
@@ -1359,14 +1479,14 @@ def _to_async_client(sync_client, model: str):
         "api_key": sync_client.api_key,
         "base_url": str(sync_client.base_url),
     }
-    base_lower = str(sync_client.base_url).lower()
-    if "openrouter" in base_lower:
+    sync_base_url = str(sync_client.base_url)
+    if base_url_host_matches(sync_base_url, "openrouter.ai"):
         async_kwargs["default_headers"] = dict(_OR_HEADERS)
-    elif "api.githubcopilot.com" in base_lower:
+    elif base_url_host_matches(sync_base_url, "api.githubcopilot.com"):
         from hermes_cli.models import copilot_default_headers
 
         async_kwargs["default_headers"] = copilot_default_headers()
-    elif "api.kimi.com" in base_lower:
+    elif base_url_host_matches(sync_base_url, "api.kimi.com"):
         async_kwargs["default_headers"] = {"User-Agent": "KimiCLI/1.30.0"}
     return AsyncOpenAI(**async_kwargs), model
 
@@ -1443,8 +1563,7 @@ def resolve_provider_client(
         # Auto-detect: api.openai.com + codex model name pattern
         if api_mode and api_mode != "codex_responses":
             return False  # explicit non-codex mode
-        normalized_base = (base_url_str or "").strip().lower()
-        if "api.openai.com" in normalized_base and "openrouter" not in normalized_base:
+        if base_url_hostname(base_url_str) == "api.openai.com":
             model_lower = (model_str or "").lower()
             if "codex" in model_lower:
                 return True
@@ -1492,7 +1611,13 @@ def resolve_provider_client(
 
     # ── Nous Portal (OAuth) ──────────────────────────────────────────
     if provider == "nous":
-        client, default = _try_nous()
+        # Detect vision tasks: either explicit model override from
+        # _PROVIDER_VISION_MODELS, or caller passed a known vision model.
+        _is_vision = (
+            model in _PROVIDER_VISION_MODELS.values()
+            or (model or "").strip().lower() == "mimo-v2-omni"
+        )
+        client, default = _try_nous(vision=_is_vision)
         if client is None:
             logger.warning("resolve_provider_client: nous requested "
                            "but Nous Portal not configured (run: hermes auth)")
@@ -1512,7 +1637,11 @@ 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()
@@ -1544,9 +1673,9 @@ def resolve_provider_client(
                 provider,
             )
             extra = {}
-            if "api.kimi.com" in custom_base.lower():
+            if base_url_host_matches(custom_base, "api.kimi.com"):
                 extra["default_headers"] = {"User-Agent": "KimiCLI/1.30.0"}
-            elif "api.githubcopilot.com" in custom_base.lower():
+            elif base_url_host_matches(custom_base, "api.githubcopilot.com"):
                 from hermes_cli.models import copilot_default_headers
                 extra["default_headers"] = copilot_default_headers()
             client = OpenAI(api_key=custom_key, base_url=custom_base, **extra)
@@ -1640,11 +1769,20 @@ 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():
+        if base_url_host_matches(base_url, "api.kimi.com"):
             headers["User-Agent"] = "KimiCLI/1.30.0"
-        elif "api.githubcopilot.com" in base_url.lower():
+        elif base_url_host_matches(base_url, "api.githubcopilot.com"):
             from hermes_cli.models import copilot_default_headers
 
             headers.update(copilot_default_headers())
@@ -1875,24 +2013,35 @@ def resolve_vision_provider_client(
         #      _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).
+        #      zai → glm-5v-turbo). Nous is the exception: it has a dedicated
+        #      strict vision backend with tier-aware defaults, so it must not
+        #      fall through to the user's text chat model here.
         #   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", ""):
-            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)
+            if main_provider == "nous":
+                sync_client, default_model = _resolve_strict_vision_backend(main_provider)
+                if sync_client is not None:
+                    logger.info(
+                        "Vision auto-detect: using main provider %s (%s)",
+                        main_provider, default_model or resolved_model or main_model,
+                    )
+                    return _finalize(main_provider, sync_client, default_model)
+            else:
+                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 (uses their dedicated vision model,
         # not the user's main model) when main provider has no client.
@@ -1939,7 +2088,7 @@ def auxiliary_max_tokens_param(value: int) -> dict:
     # Only use max_completion_tokens for direct OpenAI custom endpoints
     if (not or_key
             and _read_nous_auth() is None
-            and "api.openai.com" in custom_base.lower()):
+            and base_url_hostname(custom_base) == "api.openai.com"):
         return {"max_completion_tokens": value}
     return {"max_tokens": value}
 
@@ -1967,6 +2116,76 @@ _client_cache_lock = threading.Lock()
 _CLIENT_CACHE_MAX_SIZE = 64  # safety belt — evict oldest when exceeded
 
 
+def _client_cache_key(
+    provider: str,
+    *,
+    async_mode: bool,
+    base_url: Optional[str] = None,
+    api_key: Optional[str] = None,
+    api_mode: Optional[str] = None,
+    main_runtime: Optional[Dict[str, Any]] = None,
+) -> tuple:
+    runtime = _normalize_main_runtime(main_runtime)
+    runtime_key = tuple(runtime.get(field, "") for field in _MAIN_RUNTIME_FIELDS) if provider == "auto" else ()
+    return (provider, async_mode, base_url or "", api_key or "", api_mode or "", runtime_key)
+
+
+def _store_cached_client(cache_key: tuple, client: Any, default_model: Optional[str], *, bound_loop: Any = None) -> None:
+    with _client_cache_lock:
+        old_entry = _client_cache.get(cache_key)
+        if old_entry is not None and old_entry[0] is not client:
+            _force_close_async_httpx(old_entry[0])
+            try:
+                close_fn = getattr(old_entry[0], "close", None)
+                if callable(close_fn):
+                    close_fn()
+            except Exception:
+                pass
+        _client_cache[cache_key] = (client, default_model, bound_loop)
+
+
+def _refresh_nous_auxiliary_client(
+    *,
+    cache_provider: str,
+    model: Optional[str],
+    async_mode: bool,
+    base_url: Optional[str] = None,
+    api_key: Optional[str] = None,
+    api_mode: Optional[str] = None,
+    main_runtime: Optional[Dict[str, Any]] = None,
+) -> Tuple[Optional[Any], Optional[str]]:
+    """Refresh Nous runtime creds, rebuild the client, and replace the cache entry."""
+    runtime = _resolve_nous_runtime_api(force_refresh=True)
+    if runtime is None:
+        return None, model
+
+    fresh_key, fresh_base_url = runtime
+    sync_client = OpenAI(api_key=fresh_key, base_url=fresh_base_url)
+    final_model = model
+
+    current_loop = None
+    if async_mode:
+        try:
+            import asyncio as _aio
+            current_loop = _aio.get_event_loop()
+        except RuntimeError:
+            pass
+        client, final_model = _to_async_client(sync_client, final_model or "")
+    else:
+        client = sync_client
+
+    cache_key = _client_cache_key(
+        cache_provider,
+        async_mode=async_mode,
+        base_url=base_url,
+        api_key=api_key,
+        api_mode=api_mode,
+        main_runtime=main_runtime,
+    )
+    _store_cached_client(cache_key, client, final_model, bound_loop=current_loop)
+    return client, final_model
+
+
 def neuter_async_httpx_del() -> None:
     """Monkey-patch ``AsyncHttpxClientWrapper.__del__`` to be a no-op.
 
@@ -2068,7 +2287,7 @@ def cleanup_stale_async_clients() -> None:
 
 def _is_openrouter_client(client: Any) -> bool:
     for obj in (client, getattr(client, "_client", None), getattr(client, "client", None)):
-        if obj and "openrouter" in str(getattr(obj, "base_url", "") or "").lower():
+        if obj and base_url_host_matches(str(getattr(obj, "base_url", "") or ""), "openrouter.ai"):
             return True
     return False
 
@@ -2120,8 +2339,14 @@ def _get_cached_client(
         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 "", runtime_key)
+    cache_key = _client_cache_key(
+        provider,
+        async_mode=async_mode,
+        base_url=base_url,
+        api_key=api_key,
+        api_mode=api_mode,
+        main_runtime=main_runtime,
+    )
     with _client_cache_lock:
         if cache_key in _client_cache:
             cached_client, cached_default, cached_loop = _client_cache[cache_key]
@@ -2190,7 +2415,6 @@ 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
@@ -2198,16 +2422,7 @@ 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
@@ -2237,17 +2452,25 @@ 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:
@@ -2257,6 +2480,15 @@ 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
 # ---------------------------------------------------------------------------
@@ -2344,8 +2576,10 @@ def _build_call_kwargs(
         "timeout": timeout,
     }
 
-    fixed_temperature = _fixed_temperature_for_model(model)
-    if fixed_temperature is not None:
+    fixed_temperature = _fixed_temperature_for_model(model, base_url)
+    if fixed_temperature is OMIT_TEMPERATURE:
+        temperature = None  # strip — let server choose
+    elif fixed_temperature is not None:
         temperature = fixed_temperature
 
     # Opus 4.7+ rejects any non-default temperature/top_p/top_k — silently
@@ -2365,7 +2599,7 @@ def _build_call_kwargs(
         # Direct OpenAI api.openai.com with newer models needs max_completion_tokens.
         if provider == "custom":
             custom_base = base_url or _current_custom_base_url()
-            if "api.openai.com" in custom_base.lower():
+            if base_url_hostname(custom_base) == "api.openai.com":
                 kwargs["max_completion_tokens"] = max_tokens
             else:
                 kwargs["max_tokens"] = max_tokens
@@ -2457,6 +2691,8 @@ 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(
@@ -2525,11 +2761,14 @@ 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 "")
@@ -2555,6 +2794,29 @@ def call_llm(
                     raise
                 first_err = retry_err
 
+        # ── Nous auth refresh parity with main agent ──────────────────
+        client_is_nous = (
+            resolved_provider == "nous"
+            or base_url_host_matches(_base_info, "inference-api.nousresearch.com")
+        )
+        if _is_auth_error(first_err) and client_is_nous:
+            refreshed_client, refreshed_model = _refresh_nous_auxiliary_client(
+                cache_provider=resolved_provider or "nous",
+                model=final_model,
+                async_mode=False,
+                base_url=resolved_base_url,
+                api_key=resolved_api_key,
+                api_mode=resolved_api_mode,
+                main_runtime=main_runtime,
+            )
+            if refreshed_client is not None:
+                logger.info("Auxiliary %s: refreshed Nous runtime credentials after 401, retrying",
+                            task or "call")
+                if refreshed_model and refreshed_model != kwargs.get("model"):
+                    kwargs["model"] = refreshed_model
+                return _validate_llm_response(
+                    refreshed_client.chat.completions.create(**kwargs), task)
+
         # ── Payment / credit exhaustion fallback ──────────────────────
         # When the resolved provider returns 402 or a credit-related error,
         # try alternative providers instead of giving up.  This handles the
@@ -2583,7 +2845,8 @@ 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
@@ -2665,6 +2928,8 @@ 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(
@@ -2718,14 +2983,17 @@ 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"])
 
@@ -2747,6 +3015,28 @@ async def async_call_llm(
                     raise
                 first_err = retry_err
 
+        # ── Nous auth refresh parity with main agent ──────────────────
+        client_is_nous = (
+            resolved_provider == "nous"
+            or base_url_host_matches(_client_base, "inference-api.nousresearch.com")
+        )
+        if _is_auth_error(first_err) and client_is_nous:
+            refreshed_client, refreshed_model = _refresh_nous_auxiliary_client(
+                cache_provider=resolved_provider or "nous",
+                model=final_model,
+                async_mode=True,
+                base_url=resolved_base_url,
+                api_key=resolved_api_key,
+                api_mode=resolved_api_mode,
+            )
+            if refreshed_client is not None:
+                logger.info("Auxiliary %s (async): refreshed Nous runtime credentials after 401, retrying",
+                            task or "call")
+                if refreshed_model and refreshed_model != kwargs.get("model"):
+                    kwargs["model"] = refreshed_model
+                return _validate_llm_response(
+                    await refreshed_client.chat.completions.create(**kwargs), task)
+
         # ── Payment / connection fallback (mirrors sync call_llm) ─────
         should_fallback = _is_payment_error(first_err) or _is_connection_error(first_err)
         is_auto = resolved_provider in ("auto", "", None)
@@ -2761,7 +3051,8 @@ 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 +1,813 @@
+"""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__)
+
+
+# ---------------------------------------------------------------------------
+# Multimodal content helpers
+# ---------------------------------------------------------------------------
+
+def _chat_content_to_responses_parts(content: Any) -> List[Dict[str, Any]]:
+    """Convert chat-style multimodal content to Responses API input parts.
+
+    Input:  ``[{"type":"text"|"image_url", ...}]`` (native OpenAI Chat format)
+    Output: ``[{"type":"input_text"|"input_image", ...}]`` (Responses format)
+
+    Returns an empty list when ``content`` is not a list or contains no
+    recognized parts — callers fall back to the string path.
+    """
+    if not isinstance(content, list):
+        return []
+    converted: List[Dict[str, Any]] = []
+    for part in content:
+        if isinstance(part, str):
+            if part:
+                converted.append({"type": "input_text", "text": part})
+            continue
+        if not isinstance(part, dict):
+            continue
+        ptype = str(part.get("type") or "").strip().lower()
+        if ptype in {"text", "input_text", "output_text"}:
+            text = part.get("text")
+            if isinstance(text, str) and text:
+                converted.append({"type": "input_text", "text": text})
+            continue
+        if ptype in {"image_url", "input_image"}:
+            image_ref = part.get("image_url")
+            detail = part.get("detail")
+            if isinstance(image_ref, dict):
+                url = image_ref.get("url")
+                detail = image_ref.get("detail", detail)
+            else:
+                url = image_ref
+            if not isinstance(url, str) or not url:
+                continue
+            image_part: Dict[str, Any] = {"type": "input_image", "image_url": url}
+            if isinstance(detail, str) and detail.strip():
+                image_part["detail"] = detail.strip()
+            converted.append(image_part)
+    return converted
+
+
+def _summarize_user_message_for_log(content: Any) -> str:
+    """Return a short text summary of a user message for logging/trajectory.
+
+    Multimodal messages arrive as a list of ``{type:"text"|"image_url", ...}``
+    parts from the API server.  Logging, spinner previews, and trajectory
+    files all want a plain string — this helper extracts the first chunk of
+    text and notes any attached images.  Returns an empty string for empty
+    lists and ``str(content)`` for unexpected scalar types.
+    """
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        text_bits: List[str] = []
+        image_count = 0
+        for part in content:
+            if isinstance(part, str):
+                if part:
+                    text_bits.append(part)
+                continue
+            if not isinstance(part, dict):
+                continue
+            ptype = str(part.get("type") or "").strip().lower()
+            if ptype in {"text", "input_text", "output_text"}:
+                text = part.get("text")
+                if isinstance(text, str) and text:
+                    text_bits.append(text)
+            elif ptype in {"image_url", "input_image"}:
+                image_count += 1
+        summary = " ".join(text_bits).strip()
+        if image_count:
+            note = f"[{image_count} image{'s' if image_count != 1 else ''}]"
+            summary = f"{note} {summary}" if summary else note
+        return summary
+    try:
+        return str(content)
+    except Exception:
+        return ""
+
+
+# ---------------------------------------------------------------------------
+# ID helpers
+# ---------------------------------------------------------------------------
+
+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}"
+
+
+# ---------------------------------------------------------------------------
+# Schema conversion
+# ---------------------------------------------------------------------------
+
+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."""
+    if not tools:
+        return None
+
+    converted: List[Dict[str, Any]] = []
+    for item in 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
+
+
+# ---------------------------------------------------------------------------
+# Message format conversion
+# ---------------------------------------------------------------------------
+
+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", "")
+            if isinstance(content, list):
+                content_parts = _chat_content_to_responses_parts(content)
+                content_text = "".join(
+                    p.get("text", "") for p in content_parts if p.get("type") == "input_text"
+                )
+            else:
+                content_parts = []
+                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_parts:
+                    items.append({"role": "assistant", "content": content_parts})
+                elif 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
+
+            # Non-assistant (user) role: emit multimodal parts when present,
+            # otherwise fall back to the text payload.
+            if content_parts:
+                items.append({"role": role, "content": content_parts})
+            else:
+                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
+
+
+# ---------------------------------------------------------------------------
+# Input preflight / validation
+# ---------------------------------------------------------------------------
+
+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 isinstance(content, list):
+                # Multimodal content from ``_chat_messages_to_responses_input``
+                # is already in Responses format (``input_text`` / ``input_image``).
+                # Validate each part and pass through.
+                validated: List[Dict[str, Any]] = []
+                for part_idx, part in enumerate(content):
+                    if isinstance(part, str):
+                        if part:
+                            validated.append({"type": "input_text", "text": part})
+                        continue
+                    if not isinstance(part, dict):
+                        raise ValueError(
+                            f"Codex Responses input[{idx}].content[{part_idx}] must be an object or string."
+                        )
+                    ptype = str(part.get("type") or "").strip().lower()
+                    if ptype in {"input_text", "text", "output_text"}:
+                        text = part.get("text", "")
+                        if not isinstance(text, str):
+                            text = str(text or "")
+                        validated.append({"type": "input_text", "text": text})
+                    elif ptype in {"input_image", "image_url"}:
+                        image_ref = part.get("image_url", "")
+                        detail = part.get("detail")
+                        if isinstance(image_ref, dict):
+                            url = image_ref.get("url", "")
+                            detail = image_ref.get("detail", detail)
+                        else:
+                            url = image_ref
+                        if not isinstance(url, str):
+                            url = str(url or "")
+                        image_part: Dict[str, Any] = {"type": "input_image", "image_url": url}
+                        if isinstance(detail, str) and detail.strip():
+                            image_part["detail"] = detail.strip()
+                        validated.append(image_part)
+                    else:
+                        raise ValueError(
+                            f"Codex Responses input[{idx}].content[{part_idx}] has unsupported type {part.get('type')!r}."
+                        )
+                normalized.append({"role": role, "content": validated})
+                continue
+            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
+
+
+# ---------------------------------------------------------------------------
+# Response extraction helpers
+# ---------------------------------------------------------------------------
+
+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 ""
+
+
+# ---------------------------------------------------------------------------
+# Full response normalization
+# ---------------------------------------------------------------------------
+
+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

@@ -31,6 +31,7 @@ from agent.model_metadata import (
     get_model_context_length,
     estimate_messages_tokens_rough,
 )
+from agent.redact import redact_sensitive_text
 
 logger = logging.getLogger(__name__)
 
@@ -550,11 +551,15 @@ class ContextCompressor(ContextEngine):
         Includes tool call arguments and result content (up to
         ``_CONTENT_MAX`` chars per message) so the summarizer can preserve
         specific details like file paths, commands, and outputs.
+
+        All content is redacted before serialization to prevent secrets
+        (API keys, tokens, passwords) from leaking into the summary that
+        gets sent to the auxiliary model and persisted across compactions.
         """
         parts = []
         for msg in turns:
             role = msg.get("role", "unknown")
-            content = msg.get("content") or ""
+            content = redact_sensitive_text(msg.get("content") or "")
 
             # Tool results: keep enough content for the summarizer
             if role == "tool":
@@ -575,7 +580,7 @@ class ContextCompressor(ContextEngine):
                         if isinstance(tc, dict):
                             fn = tc.get("function", {})
                             name = fn.get("name", "?")
-                            args = fn.get("arguments", "")
+                            args = redact_sensitive_text(fn.get("arguments", ""))
                             # Truncate long arguments but keep enough for context
                             if len(args) > self._TOOL_ARGS_MAX:
                                 args = args[:self._TOOL_ARGS_HEAD] + "..."
@@ -635,7 +640,11 @@ class ContextCompressor(ContextEngine):
             "only output the structured summary. "
             "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."
+            "conversation — do not translate or switch to English. "
+            "NEVER include API keys, tokens, passwords, secrets, credentials, "
+            "or connection strings in the summary — replace any that appear "
+            "with [REDACTED]. Note that the user had credentials present, but "
+            "do not preserve their values."
         )
 
         # Shared structured template (used by both paths).
@@ -692,7 +701,7 @@ Be specific with file paths, commands, line numbers, and results.]
 [What remains to be done — framed as context, not instructions]
 
 ## Critical Context
-[Any specific values, error messages, configuration details, or data that would be lost without explicit preservation]
+[Any specific values, error messages, configuration details, or data that would be lost without explicit preservation. NEVER include API keys, tokens, passwords, or credentials — write [REDACTED] instead.]
 
 Target ~{summary_budget} tokens. Be CONCRETE — include file paths, command outputs, error messages, line numbers, and specific values. Avoid vague descriptions like "made some changes" — say exactly what changed.
 
@@ -732,7 +741,7 @@ Use this exact structure:
             prompt += f"""
 
 FOCUS TOPIC: "{focus_topic}"
-The user has requested that this compaction PRIORITISE preserving all information related to the focus topic above. For content related to "{focus_topic}", include full detail — exact values, file paths, command outputs, error messages, and decisions. For content NOT related to the focus topic, summarise more aggressively (brief one-liners or omit if truly irrelevant). The focus topic sections should receive roughly 60-70% of the summary token budget."""
+The user has requested that this compaction PRIORITISE preserving all information related to the focus topic above. For content related to "{focus_topic}", include full detail — exact values, file paths, command outputs, error messages, and decisions. For content NOT related to the focus topic, summarise more aggressively (brief one-liners or omit if truly irrelevant). The focus topic sections should receive roughly 60-70% of the summary token budget. Even for the focus topic, NEVER preserve API keys, tokens, passwords, or credentials — use [REDACTED]."""
 
         try:
             call_kwargs = {
@@ -755,7 +764,9 @@ The user has requested that this compaction PRIORITISE preserving all informatio
             # Handle cases where content is not a string (e.g., dict from llama.cpp)
             if not isinstance(content, str):
                 content = str(content) if content else ""
-            summary = content.strip()
+            # Redact the summary output as well — the summarizer LLM may
+            # ignore prompt instructions and echo back secrets verbatim.
+            summary = redact_sensitive_text(content.strip())
             # Store for iterative updates on next compaction
             self._previous_summary = summary
             self._summary_failure_cooldown_until = 0.0
@@ -796,7 +807,7 @@ The user has requested that this compaction PRIORITISE preserving all informatio
                 )
                 self.summary_model = ""  # empty = use main model
                 self._summary_failure_cooldown_until = 0.0  # no cooldown
-                return self._generate_summary(messages, summary_budget)  # retry immediately
+                return self._generate_summary(turns_to_summarize)  # retry immediately
 
             # Transient errors (timeout, rate limit, network) — shorter cooldown
             _transient_cooldown = 60

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

@@ -21,6 +21,9 @@ from pathlib import Path
 from types import SimpleNamespace
 from typing import Any
 
+from agent.file_safety import get_read_block_error, is_write_denied
+from agent.redact import redact_sensitive_text
+
 ACP_MARKER_BASE_URL = "acp://copilot"
 _DEFAULT_TIMEOUT_SECONDS = 900.0
 
@@ -54,6 +57,18 @@ def _jsonrpc_error(message_id: Any, code: int, message: str) -> dict[str, Any]:
     }
 
 
+def _permission_denied(message_id: Any) -> dict[str, Any]:
+    return {
+        "jsonrpc": "2.0",
+        "id": message_id,
+        "result": {
+            "outcome": {
+                "outcome": "cancelled",
+            }
+        },
+    }
+
+
 def _format_messages_as_prompt(
     messages: list[dict[str, Any]],
     model: str | None = None,
@@ -386,6 +401,8 @@ class CopilotACPClient:
         stderr_tail: deque[str] = deque(maxlen=40)
 
         def _stdout_reader() -> None:
+            if proc.stdout is None:
+                return
             for line in proc.stdout:
                 try:
                     inbox.put(json.loads(line))
@@ -533,18 +550,13 @@ class CopilotACPClient:
         params = msg.get("params") or {}
 
         if method == "session/request_permission":
-            response = {
-                "jsonrpc": "2.0",
-                "id": message_id,
-                "result": {
-                    "outcome": {
-                        "outcome": "allow_once",
-                    }
-                },
-            }
+            response = _permission_denied(message_id)
         elif method == "fs/read_text_file":
             try:
                 path = _ensure_path_within_cwd(str(params.get("path") or ""), cwd)
+                block_error = get_read_block_error(str(path))
+                if block_error:
+                    raise PermissionError(block_error)
                 content = path.read_text() if path.exists() else ""
                 line = params.get("line")
                 limit = params.get("limit")
@@ -553,6 +565,8 @@ class CopilotACPClient:
                     start = line - 1
                     end = start + limit if isinstance(limit, int) and limit > 0 else None
                     content = "".join(lines[start:end])
+                if content:
+                    content = redact_sensitive_text(content)
                 response = {
                     "jsonrpc": "2.0",
                     "id": message_id,
@@ -565,6 +579,10 @@ class CopilotACPClient:
         elif method == "fs/write_text_file":
             try:
                 path = _ensure_path_within_cwd(str(params.get("path") or ""), cwd)
+                if is_write_denied(str(path)):
+                    raise PermissionError(
+                        f"Write denied: '{path}' is a protected system/credential file."
+                    )
                 path.parent.mkdir(parents=True, exist_ok=True)
                 path.write_text(str(params.get("content") or ""))
                 response = {

agent/credential_pool.py

@@ -983,6 +983,14 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
     active_sources: Set[str] = set()
     auth_store = _load_auth_store()
 
+    # Shared suppression gate — used at every upsert site so
+    # `hermes auth remove <provider> <N>` is stable across all source types.
+    try:
+        from hermes_cli.auth import is_source_suppressed as _is_suppressed
+    except ImportError:
+        def _is_suppressed(_p, _s):  # type: ignore[misc]
+            return False
+
     if provider == "anthropic":
         # Only auto-discover external credentials (Claude Code, Hermes PKCE)
         # when the user has explicitly configured anthropic as their provider.
@@ -1002,13 +1010,8 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
             ("claude_code", read_claude_code_credentials()),
         ):
             if creds and creds.get("accessToken"):
-                # Check if user explicitly removed this source
-                try:
-                    from hermes_cli.auth import is_source_suppressed
-                    if is_source_suppressed(provider, source_name):
-                        continue
-                except ImportError:
-                    pass
+                if _is_suppressed(provider, source_name):
+                    continue
                 active_sources.add(source_name)
                 changed |= _upsert_entry(
                     entries,
@@ -1026,7 +1029,7 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
 
     elif provider == "nous":
         state = _load_provider_state(auth_store, "nous")
-        if state:
+        if state and not _is_suppressed(provider, "device_code"):
             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
@@ -1067,20 +1070,21 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
             token, source = resolve_copilot_token()
             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,
-                    source_name,
-                    {
-                        "source": source_name,
-                        "auth_type": AUTH_TYPE_API_KEY,
-                        "access_token": token,
-                        "base_url": pconfig.inference_base_url if pconfig else "",
-                        "label": source,
-                    },
-                )
+                if not _is_suppressed(provider, source_name):
+                    active_sources.add(source_name)
+                    pconfig = PROVIDER_REGISTRY.get(provider)
+                    changed |= _upsert_entry(
+                        entries,
+                        provider,
+                        source_name,
+                        {
+                            "source": source_name,
+                            "auth_type": AUTH_TYPE_API_KEY,
+                            "access_token": token,
+                            "base_url": pconfig.inference_base_url if pconfig else "",
+                            "label": source,
+                        },
+                    )
         except Exception as exc:
             logger.debug("Copilot token seed failed: %s", exc)
 
@@ -1096,20 +1100,21 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
             token = creds.get("api_key", "")
             if token:
                 source_name = creds.get("source", "qwen-cli")
-                active_sources.add(source_name)
-                changed |= _upsert_entry(
-                    entries,
-                    provider,
-                    source_name,
-                    {
-                        "source": source_name,
-                        "auth_type": AUTH_TYPE_OAUTH,
-                        "access_token": token,
-                        "expires_at_ms": creds.get("expires_at_ms"),
-                        "base_url": creds.get("base_url", ""),
-                        "label": creds.get("auth_file", source_name),
-                    },
-                )
+                if not _is_suppressed(provider, source_name):
+                    active_sources.add(source_name)
+                    changed |= _upsert_entry(
+                        entries,
+                        provider,
+                        source_name,
+                        {
+                            "source": source_name,
+                            "auth_type": AUTH_TYPE_OAUTH,
+                            "access_token": token,
+                            "expires_at_ms": creds.get("expires_at_ms"),
+                            "base_url": creds.get("base_url", ""),
+                            "label": creds.get("auth_file", source_name),
+                        },
+                    )
         except Exception as exc:
             logger.debug("Qwen OAuth token seed failed: %s", exc)
 
@@ -1118,13 +1123,7 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
         # 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:
+        if _is_suppressed(provider, "device_code"):
             return changed, active_sources
 
         state = _load_provider_state(auth_store, "openai-codex")
@@ -1158,10 +1157,22 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
 def _seed_from_env(provider: str, entries: List[PooledCredential]) -> Tuple[bool, Set[str]]:
     changed = False
     active_sources: Set[str] = set()
+    # Honour user suppression — `hermes auth remove <provider> <N>` for an
+    # env-seeded credential marks the env:<VAR> source as suppressed so it
+    # won't be re-seeded from the user's shell environment or ~/.hermes/.env.
+    # Without this gate the removal is silently undone on the next
+    # load_pool() call whenever the var is still exported by the shell.
+    try:
+        from hermes_cli.auth import is_source_suppressed as _is_source_suppressed
+    except ImportError:
+        def _is_source_suppressed(_p, _s):  # type: ignore[misc]
+            return False
     if provider == "openrouter":
         token = os.getenv("OPENROUTER_API_KEY", "").strip()
         if token:
             source = "env:OPENROUTER_API_KEY"
+            if _is_source_suppressed(provider, source):
+                return changed, active_sources
             active_sources.add(source)
             changed |= _upsert_entry(
                 entries,
@@ -1198,6 +1209,8 @@ def _seed_from_env(provider: str, entries: List[PooledCredential]) -> Tuple[bool
         if not token:
             continue
         source = f"env:{env_var}"
+        if _is_source_suppressed(provider, source):
+            continue
         active_sources.add(source)
         auth_type = AUTH_TYPE_OAUTH if provider == "anthropic" and not token.startswith("sk-ant-api") else AUTH_TYPE_API_KEY
         base_url = env_url or pconfig.inference_base_url
@@ -1242,6 +1255,13 @@ def _seed_custom_pool(pool_key: str, entries: List[PooledCredential]) -> Tuple[b
     changed = False
     active_sources: Set[str] = set()
 
+    # Shared suppression gate — same pattern as _seed_from_env/_seed_from_singletons.
+    try:
+        from hermes_cli.auth import is_source_suppressed as _is_suppressed
+    except ImportError:
+        def _is_suppressed(_p, _s):  # type: ignore[misc]
+            return False
+
     # Seed from the custom_providers config entry's api_key field
     cp_config = _get_custom_provider_config(pool_key)
     if cp_config:
@@ -1250,19 +1270,20 @@ def _seed_custom_pool(pool_key: str, entries: List[PooledCredential]) -> Tuple[b
         name = str(cp_config.get("name") or "").strip()
         if api_key:
             source = f"config:{name}"
-            active_sources.add(source)
-            changed |= _upsert_entry(
-                entries,
-                pool_key,
-                source,
-                {
-                    "source": source,
-                    "auth_type": AUTH_TYPE_API_KEY,
-                    "access_token": api_key,
-                    "base_url": base_url,
-                    "label": name or source,
-                },
-            )
+            if not _is_suppressed(pool_key, source):
+                active_sources.add(source)
+                changed |= _upsert_entry(
+                    entries,
+                    pool_key,
+                    source,
+                    {
+                        "source": source,
+                        "auth_type": AUTH_TYPE_API_KEY,
+                        "access_token": api_key,
+                        "base_url": base_url,
+                        "label": name or source,
+                    },
+                )
 
     # Seed from model.api_key if model.provider=='custom' and model.base_url matches
     try:
@@ -1282,19 +1303,20 @@ def _seed_custom_pool(pool_key: str, entries: List[PooledCredential]) -> Tuple[b
                 matched_key = get_custom_provider_pool_key(model_base_url)
                 if matched_key == pool_key:
                     source = "model_config"
-                    active_sources.add(source)
-                    changed |= _upsert_entry(
-                        entries,
-                        pool_key,
-                        source,
-                        {
-                            "source": source,
-                            "auth_type": AUTH_TYPE_API_KEY,
-                            "access_token": model_api_key,
-                            "base_url": model_base_url,
-                            "label": "model_config",
-                        },
-                    )
+                    if not _is_suppressed(pool_key, source):
+                        active_sources.add(source)
+                        changed |= _upsert_entry(
+                            entries,
+                            pool_key,
+                            source,
+                            {
+                                "source": source,
+                                "auth_type": AUTH_TYPE_API_KEY,
+                                "access_token": model_api_key,
+                                "base_url": model_base_url,
+                                "label": "model_config",
+                            },
+                        )
     except Exception:
         pass
 

agent/credential_sources.py

@@ -0,0 +1,401 @@
+"""Unified removal contract for every credential source Hermes reads from.
+
+Hermes seeds its credential pool from many places:
+
+    env:<VAR>     — os.environ / ~/.hermes/.env
+    claude_code   — ~/.claude/.credentials.json
+    hermes_pkce   — ~/.hermes/.anthropic_oauth.json
+    device_code   — auth.json providers.<provider> (nous, openai-codex, ...)
+    qwen-cli      — ~/.qwen/oauth_creds.json
+    gh_cli        — gh auth token
+    config:<name> — custom_providers config entry
+    model_config  — model.api_key when model.provider == "custom"
+    manual        — user ran `hermes auth add`
+
+Each source has its own reader inside ``agent.credential_pool._seed_from_*``
+(which keep their existing shape — we haven't restructured them).  What we
+unify here is **removal**:
+
+    ``hermes auth remove <provider> <N>`` must make the pool entry stay gone.
+
+Before this module, every source had an ad-hoc removal branch in
+``auth_remove_command``, and several sources had no branch at all — so
+``auth remove`` silently reverted on the next ``load_pool()`` call for
+qwen-cli, nous device_code (partial), hermes_pkce, copilot gh_cli, and
+custom-config sources.
+
+Now every source registers a ``RemovalStep`` that does exactly three things
+in the same shape:
+
+    1. Clean up whatever externally-readable state the source reads from
+       (.env line, auth.json block, OAuth file, etc.)
+    2. Suppress the ``(provider, source_id)`` in auth.json so the
+       corresponding ``_seed_from_*`` branch skips the upsert on re-load
+    3. Return ``RemovalResult`` describing what was cleaned and any
+       diagnostic hints the user should see (shell-exported env vars,
+       external credential files we deliberately don't delete, etc.)
+
+Adding a new credential source is:
+    - wire up a reader branch in ``_seed_from_*`` (existing pattern)
+    - gate that reader behind ``is_source_suppressed(provider, source_id)``
+    - register a ``RemovalStep`` here
+
+No more per-source if/elif chain in ``auth_remove_command``.
+"""
+
+from __future__ import annotations
+
+import os
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Callable, List, Optional
+
+
+@dataclass
+class RemovalResult:
+    """Outcome of removing a credential source.
+
+    Attributes:
+        cleaned: Short strings describing external state that was actually
+            mutated (``"Cleared XAI_API_KEY from .env"``,
+            ``"Cleared openai-codex OAuth tokens from auth store"``).
+            Printed as plain lines to the user.
+        hints: Diagnostic lines ABOUT state the user may need to clean up
+            themselves or is deliberately left intact (shell-exported env
+            var, Claude Code credential file we don't delete, etc.).
+            Printed as plain lines to the user.  Always non-destructive.
+        suppress: Whether to call ``suppress_credential_source`` after
+            cleanup so future ``load_pool`` calls skip this source.
+            Default True — almost every source needs this to stay sticky.
+            The only legitimate False is ``manual`` entries, which aren't
+            seeded from anywhere external.
+    """
+
+    cleaned: List[str] = field(default_factory=list)
+    hints: List[str] = field(default_factory=list)
+    suppress: bool = True
+
+
+@dataclass
+class RemovalStep:
+    """How to remove one specific credential source cleanly.
+
+    Attributes:
+        provider: Provider pool key (``"xai"``, ``"anthropic"``, ``"nous"``, ...).
+            Special value ``"*"`` means "matches any provider" — used for
+            sources like ``manual`` that aren't provider-specific.
+        source_id: Source identifier as it appears in
+            ``PooledCredential.source``.  May be a literal (``"claude_code"``)
+            or a prefix pattern matched via ``match_fn``.
+        match_fn: Optional predicate overriding literal ``source_id``
+            matching.  Gets the removed entry's source string.  Used for
+            ``env:*`` (any env-seeded key), ``config:*`` (any custom
+            pool), and ``manual:*`` (any manual-source variant).
+        remove_fn: ``(provider, removed_entry) -> RemovalResult``.  Does the
+            actual cleanup and returns what happened for the user.
+        description: One-line human-readable description for docs / tests.
+    """
+
+    provider: str
+    source_id: str
+    remove_fn: Callable[..., RemovalResult]
+    match_fn: Optional[Callable[[str], bool]] = None
+    description: str = ""
+
+    def matches(self, provider: str, source: str) -> bool:
+        if self.provider != "*" and self.provider != provider:
+            return False
+        if self.match_fn is not None:
+            return self.match_fn(source)
+        return source == self.source_id
+
+
+_REGISTRY: List[RemovalStep] = []
+
+
+def register(step: RemovalStep) -> RemovalStep:
+    _REGISTRY.append(step)
+    return step
+
+
+def find_removal_step(provider: str, source: str) -> Optional[RemovalStep]:
+    """Return the first matching RemovalStep, or None if unregistered.
+
+    Unregistered sources fall through to the default remove path in
+    ``auth_remove_command``: the pool entry is already gone (that happens
+    before dispatch), no external cleanup, no suppression.  This is the
+    correct behaviour for ``manual`` entries — they were only ever stored
+    in the pool, nothing external to clean up.
+    """
+    for step in _REGISTRY:
+        if step.matches(provider, source):
+            return step
+    return None
+
+
+# ---------------------------------------------------------------------------
+# Individual RemovalStep implementations — one per source.
+# ---------------------------------------------------------------------------
+# Each remove_fn is intentionally small and single-purpose.  Adding a new
+# credential source means adding ONE entry here — no other changes to
+# auth_remove_command.
+
+
+def _remove_env_source(provider: str, removed) -> RemovalResult:
+    """env:<VAR> — the most common case.
+
+    Handles three user situations:
+      1. Var lives only in ~/.hermes/.env  → clear it
+      2. Var lives only in the user's shell (shell profile, systemd
+         EnvironmentFile, launchd plist) → hint them where to unset it
+      3. Var lives in both → clear from .env, hint about shell
+    """
+    from hermes_cli.config import get_env_path, remove_env_value
+
+    result = RemovalResult()
+    env_var = removed.source[len("env:"):]
+    if not env_var:
+        return result
+
+    # Detect shell vs .env BEFORE remove_env_value pops os.environ.
+    env_in_process = bool(os.getenv(env_var))
+    env_in_dotenv = False
+    try:
+        env_path = get_env_path()
+        if env_path.exists():
+            env_in_dotenv = any(
+                line.strip().startswith(f"{env_var}=")
+                for line in env_path.read_text(errors="replace").splitlines()
+            )
+    except OSError:
+        pass
+    shell_exported = env_in_process and not env_in_dotenv
+
+    cleared = remove_env_value(env_var)
+    if cleared:
+        result.cleaned.append(f"Cleared {env_var} from .env")
+
+    if shell_exported:
+        result.hints.extend([
+            f"Note: {env_var} is still set in your shell environment "
+            f"(not in ~/.hermes/.env).",
+            "  Unset it there (shell profile, systemd EnvironmentFile, "
+            "launchd plist, etc.) or it will keep being visible to Hermes.",
+            f"  The pool entry is now suppressed — Hermes will ignore "
+            f"{env_var} until you run `hermes auth add {provider}`.",
+        ])
+    else:
+        result.hints.append(
+            f"Suppressed env:{env_var} — it will not be re-seeded even "
+            f"if the variable is re-exported later."
+        )
+    return result
+
+
+def _remove_claude_code(provider: str, removed) -> RemovalResult:
+    """~/.claude/.credentials.json is owned by Claude Code itself.
+
+    We don't delete it — the user's Claude Code install still needs to
+    work.  We just suppress it so Hermes stops reading it.
+    """
+    return RemovalResult(hints=[
+        "Suppressed claude_code credential — it will not be re-seeded.",
+        "Note: Claude Code credentials still live in ~/.claude/.credentials.json",
+        "Run `hermes auth add anthropic` to re-enable if needed.",
+    ])
+
+
+def _remove_hermes_pkce(provider: str, removed) -> RemovalResult:
+    """~/.hermes/.anthropic_oauth.json is ours — delete it outright."""
+    from hermes_constants import get_hermes_home
+
+    result = RemovalResult()
+    oauth_file = get_hermes_home() / ".anthropic_oauth.json"
+    if oauth_file.exists():
+        try:
+            oauth_file.unlink()
+            result.cleaned.append("Cleared Hermes Anthropic OAuth credentials")
+        except OSError as exc:
+            result.hints.append(f"Could not delete {oauth_file}: {exc}")
+    return result
+
+
+def _clear_auth_store_provider(provider: str) -> bool:
+    """Delete auth_store.providers[provider].  Returns True if deleted."""
+    from hermes_cli.auth import (
+        _auth_store_lock,
+        _load_auth_store,
+        _save_auth_store,
+    )
+
+    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)
+            return True
+    return False
+
+
+def _remove_nous_device_code(provider: str, removed) -> RemovalResult:
+    """Nous OAuth lives in auth.json providers.nous — clear it and suppress.
+
+    We suppress in addition to clearing because nothing else stops the
+    user's next `hermes login` run from writing providers.nous again
+    before they decide to.  Suppression forces them to go through
+    `hermes auth add nous` to re-engage, which is the documented re-add
+    path and clears the suppression atomically.
+    """
+    result = RemovalResult()
+    if _clear_auth_store_provider(provider):
+        result.cleaned.append(f"Cleared {provider} OAuth tokens from auth store")
+    return result
+
+
+def _remove_codex_device_code(provider: str, removed) -> RemovalResult:
+    """Codex tokens live in TWO places: our auth store AND ~/.codex/auth.json.
+
+    refresh_codex_oauth_pure() writes both every time, so clearing only
+    the Hermes auth store is not enough — _seed_from_singletons() would
+    re-import from ~/.codex/auth.json on the next load_pool() call and
+    the removal would be instantly undone.  We suppress instead of
+    deleting Codex CLI's file, so the Codex CLI itself keeps working.
+
+    The canonical source name in ``_seed_from_singletons`` is
+    ``"device_code"`` (no prefix).  Entries may show up in the pool as
+    either ``"device_code"`` (seeded) or ``"manual:device_code"`` (added
+    via ``hermes auth add openai-codex``), but in both cases the re-seed
+    gate lives at the ``"device_code"`` suppression key.  We suppress
+    that canonical key here; the central dispatcher also suppresses
+    ``removed.source`` which is fine — belt-and-suspenders, idempotent.
+    """
+    from hermes_cli.auth import suppress_credential_source
+
+    result = RemovalResult()
+    if _clear_auth_store_provider(provider):
+        result.cleaned.append(f"Cleared {provider} OAuth tokens from auth store")
+    # Suppress the canonical re-seed source, not just whatever source the
+    # removed entry had.  Otherwise `manual:device_code` removals wouldn't
+    # block the `device_code` re-seed path.
+    suppress_credential_source(provider, "device_code")
+    result.hints.extend([
+        "Suppressed openai-codex device_code source — it will not be re-seeded.",
+        "Note: Codex CLI credentials still live in ~/.codex/auth.json",
+        "Run `hermes auth add openai-codex` to re-enable if needed.",
+    ])
+    return result
+
+
+def _remove_qwen_cli(provider: str, removed) -> RemovalResult:
+    """~/.qwen/oauth_creds.json is owned by the Qwen CLI.
+
+    Same pattern as claude_code — suppress, don't delete.  The user's
+    Qwen CLI install still reads from that file.
+    """
+    return RemovalResult(hints=[
+        "Suppressed qwen-cli credential — it will not be re-seeded.",
+        "Note: Qwen CLI credentials still live in ~/.qwen/oauth_creds.json",
+        "Run `hermes auth add qwen-oauth` to re-enable if needed.",
+    ])
+
+
+def _remove_copilot_gh(provider: str, removed) -> RemovalResult:
+    """Copilot token comes from `gh auth token` or COPILOT_GITHUB_TOKEN / GH_TOKEN / GITHUB_TOKEN.
+
+    Copilot is special: the same token can be seeded as multiple source
+    entries (gh_cli from ``_seed_from_singletons`` plus env:<VAR> from
+    ``_seed_from_env``), so removing one entry without suppressing the
+    others lets the duplicates resurrect.  We suppress ALL known copilot
+    sources here so removal is stable regardless of which entry the
+    user clicked.
+
+    We don't touch the user's gh CLI or shell state — just suppress so
+    Hermes stops picking the token up.
+    """
+    # Suppress ALL copilot source variants up-front so no path resurrects
+    # the pool entry.  The central dispatcher in auth_remove_command will
+    # ALSO suppress removed.source, but it's idempotent so double-calling
+    # is harmless.
+    from hermes_cli.auth import suppress_credential_source
+    suppress_credential_source(provider, "gh_cli")
+    for env_var in ("COPILOT_GITHUB_TOKEN", "GH_TOKEN", "GITHUB_TOKEN"):
+        suppress_credential_source(provider, f"env:{env_var}")
+
+    return RemovalResult(hints=[
+        "Suppressed all copilot token sources (gh_cli + env vars) — they will not be re-seeded.",
+        "Note: Your gh CLI / shell environment is unchanged.",
+        "Run `hermes auth add copilot` to re-enable if needed.",
+    ])
+
+
+def _remove_custom_config(provider: str, removed) -> RemovalResult:
+    """Custom provider pools are seeded from custom_providers config or
+    model.api_key.  Both are in config.yaml — modifying that from here
+    is more invasive than suppression.  We suppress; the user can edit
+    config.yaml if they want to remove the key from disk entirely.
+    """
+    source_label = removed.source
+    return RemovalResult(hints=[
+        f"Suppressed {source_label} — it will not be re-seeded.",
+        "Note: The underlying value in config.yaml is unchanged.  Edit it "
+        "directly if you want to remove the credential from disk.",
+    ])
+
+
+def _register_all_sources() -> None:
+    """Called once on module import.
+
+    ORDER MATTERS — ``find_removal_step`` returns the first match.  Put
+    provider-specific steps before the generic ``env:*`` step so that e.g.
+    copilot's ``env:GH_TOKEN`` goes through the copilot removal (which
+    doesn't touch the user's shell), not the generic env-var removal
+    (which would try to clear .env).
+    """
+    register(RemovalStep(
+        provider="copilot", source_id="gh_cli",
+        match_fn=lambda src: src == "gh_cli" or src.startswith("env:"),
+        remove_fn=_remove_copilot_gh,
+        description="gh auth token / COPILOT_GITHUB_TOKEN / GH_TOKEN",
+    ))
+    register(RemovalStep(
+        provider="*", source_id="env:",
+        match_fn=lambda src: src.startswith("env:"),
+        remove_fn=_remove_env_source,
+        description="Any env-seeded credential (XAI_API_KEY, DEEPSEEK_API_KEY, etc.)",
+    ))
+    register(RemovalStep(
+        provider="anthropic", source_id="claude_code",
+        remove_fn=_remove_claude_code,
+        description="~/.claude/.credentials.json",
+    ))
+    register(RemovalStep(
+        provider="anthropic", source_id="hermes_pkce",
+        remove_fn=_remove_hermes_pkce,
+        description="~/.hermes/.anthropic_oauth.json",
+    ))
+    register(RemovalStep(
+        provider="nous", source_id="device_code",
+        remove_fn=_remove_nous_device_code,
+        description="auth.json providers.nous",
+    ))
+    register(RemovalStep(
+        provider="openai-codex", source_id="device_code",
+        match_fn=lambda src: src == "device_code" or src.endswith(":device_code"),
+        remove_fn=_remove_codex_device_code,
+        description="auth.json providers.openai-codex + ~/.codex/auth.json",
+    ))
+    register(RemovalStep(
+        provider="qwen-oauth", source_id="qwen-cli",
+        remove_fn=_remove_qwen_cli,
+        description="~/.qwen/oauth_creds.json",
+    ))
+    register(RemovalStep(
+        provider="*", source_id="config:",
+        match_fn=lambda src: src.startswith("config:") or src == "model_config",
+        remove_fn=_remove_custom_config,
+        description="Custom provider config.yaml api_key field",
+    ))
+
+
+_register_all_sources()

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":
@@ -939,9 +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}")

agent/error_classifier.py

@@ -290,7 +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):
@@ -302,11 +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:
@@ -606,10 +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/file_safety.py

@@ -0,0 +1,111 @@
+"""Shared file safety rules used by both tools and ACP shims."""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+from typing import Optional
+
+
+def _hermes_home_path() -> Path:
+    """Resolve the active HERMES_HOME (profile-aware) without circular imports."""
+    try:
+        from hermes_constants import get_hermes_home  # local import to avoid cycles
+        return get_hermes_home()
+    except Exception:
+        return Path(os.path.expanduser("~/.hermes"))
+
+
+def build_write_denied_paths(home: str) -> set[str]:
+    """Return exact sensitive paths that must never be written."""
+    hermes_home = _hermes_home_path()
+    return {
+        os.path.realpath(p)
+        for p in [
+            os.path.join(home, ".ssh", "authorized_keys"),
+            os.path.join(home, ".ssh", "id_rsa"),
+            os.path.join(home, ".ssh", "id_ed25519"),
+            os.path.join(home, ".ssh", "config"),
+            str(hermes_home / ".env"),
+            os.path.join(home, ".bashrc"),
+            os.path.join(home, ".zshrc"),
+            os.path.join(home, ".profile"),
+            os.path.join(home, ".bash_profile"),
+            os.path.join(home, ".zprofile"),
+            os.path.join(home, ".netrc"),
+            os.path.join(home, ".pgpass"),
+            os.path.join(home, ".npmrc"),
+            os.path.join(home, ".pypirc"),
+            "/etc/sudoers",
+            "/etc/passwd",
+            "/etc/shadow",
+        ]
+    }
+
+
+def build_write_denied_prefixes(home: str) -> list[str]:
+    """Return sensitive directory prefixes that must never be written."""
+    return [
+        os.path.realpath(p) + os.sep
+        for p in [
+            os.path.join(home, ".ssh"),
+            os.path.join(home, ".aws"),
+            os.path.join(home, ".gnupg"),
+            os.path.join(home, ".kube"),
+            "/etc/sudoers.d",
+            "/etc/systemd",
+            os.path.join(home, ".docker"),
+            os.path.join(home, ".azure"),
+            os.path.join(home, ".config", "gh"),
+        ]
+    ]
+
+
+def get_safe_write_root() -> Optional[str]:
+    """Return the resolved HERMES_WRITE_SAFE_ROOT path, or None if unset."""
+    root = os.getenv("HERMES_WRITE_SAFE_ROOT", "")
+    if not root:
+        return None
+    try:
+        return os.path.realpath(os.path.expanduser(root))
+    except Exception:
+        return None
+
+
+def is_write_denied(path: str) -> bool:
+    """Return True if path is blocked by the write denylist or safe root."""
+    home = os.path.realpath(os.path.expanduser("~"))
+    resolved = os.path.realpath(os.path.expanduser(str(path)))
+
+    if resolved in build_write_denied_paths(home):
+        return True
+    for prefix in build_write_denied_prefixes(home):
+        if resolved.startswith(prefix):
+            return True
+
+    safe_root = get_safe_write_root()
+    if safe_root and not (resolved == safe_root or resolved.startswith(safe_root + os.sep)):
+        return True
+
+    return False
+
+
+def get_read_block_error(path: str) -> Optional[str]:
+    """Return an error message when a read targets internal Hermes cache files."""
+    resolved = Path(path).expanduser().resolve()
+    hermes_home = _hermes_home_path().resolve()
+    blocked_dirs = [
+        hermes_home / "skills" / ".hub" / "index-cache",
+        hermes_home / "skills" / ".hub",
+    ]
+    for blocked in blocked_dirs:
+        try:
+            resolved.relative_to(blocked)
+        except ValueError:
+            continue
+        return (
+            f"Access denied: {path} is an internal Hermes cache file "
+            "and cannot be read directly to prevent prompt injection. "
+            "Use the skills_list or skill_view tools instead."
+        )
+    return None

agent/gemini_cloudcode_adapter.py

@@ -39,6 +39,7 @@ 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,
@@ -205,7 +206,7 @@ def _translate_tools_to_gemini(tools: Any) -> List[Dict[str, Any]]:
             decl["description"] = str(fn["description"])
         params = fn.get("parameters")
         if isinstance(params, dict):
-            decl["parameters"] = params
+            decl["parameters"] = sanitize_gemini_tool_parameters(params)
         declarations.append(decl)
     if not declarations:
         return []
@@ -504,9 +505,16 @@ def _iter_sse_events(response: httpx.Response) -> Iterator[Dict[str, Any]]:
 def _translate_stream_event(
     event: Dict[str, Any],
     model: str,
-    tool_call_indices: Dict[str, int],
+    tool_call_counter: List[int],
 ) -> List[_GeminiStreamChunk]:
-    """Unwrap Code Assist envelope and emit OpenAI-shaped chunk(s)."""
+    """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:
@@ -532,7 +540,8 @@ def _translate_stream_event(
         fc = part.get("functionCall")
         if isinstance(fc, dict) and fc.get("name"):
             name = str(fc["name"])
-            idx = tool_call_indices.setdefault(name, len(tool_call_indices))
+            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):
@@ -549,7 +558,7 @@ def _translate_stream_event(
     finish_reason_raw = str(cand.get("finishReason") or "")
     if finish_reason_raw:
         mapped = _map_gemini_finish_reason(finish_reason_raw)
-        if tool_call_indices:
+        if tool_call_counter[0] > 0:
             mapped = "tool_calls"
         chunks.append(_make_stream_chunk(model=model, finish_reason=mapped))
     return chunks
@@ -733,9 +742,9 @@ class GeminiCloudCodeClient:
                         # Materialize error body for better diagnostics
                         response.read()
                         raise _gemini_http_error(response)
-                    tool_call_indices: Dict[str, int] = {}
+                    tool_call_counter: List[int] = [0]
                     for event in _iter_sse_events(response):
-                        for chunk in _translate_stream_event(event, model, tool_call_indices):
+                        for chunk in _translate_stream_event(event, model, tool_call_counter):
                             yield chunk
             except httpx.HTTPError as exc:
                 raise CodeAssistError(
@@ -790,7 +799,8 @@ def _gemini_http_error(response: httpx.Response) -> CodeAssistError:
         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 []
+    _raw_details = err_obj.get("details")
+    err_details_list = _raw_details if isinstance(_raw_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.

agent/gemini_native_adapter.py

@@ -0,0 +1,847 @@
+"""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()
+    _raw_details = err_obj.get("details")
+    details_list = _raw_details if isinstance(_raw_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/insights.py

@@ -124,6 +124,7 @@ class InsightsEngine:
         # Gather raw data
         sessions = self._get_sessions(cutoff, source)
         tool_usage = self._get_tool_usage(cutoff, source)
+        skill_usage = self._get_skill_usage(cutoff, source)
         message_stats = self._get_message_stats(cutoff, source)
 
         if not sessions:
@@ -135,6 +136,15 @@ class InsightsEngine:
                 "models": [],
                 "platforms": [],
                 "tools": [],
+                "skills": {
+                    "summary": {
+                        "total_skill_loads": 0,
+                        "total_skill_edits": 0,
+                        "total_skill_actions": 0,
+                        "distinct_skills_used": 0,
+                    },
+                    "top_skills": [],
+                },
                 "activity": {},
                 "top_sessions": [],
             }
@@ -144,6 +154,7 @@ class InsightsEngine:
         models = self._compute_model_breakdown(sessions)
         platforms = self._compute_platform_breakdown(sessions)
         tools = self._compute_tool_breakdown(tool_usage)
+        skills = self._compute_skill_breakdown(skill_usage)
         activity = self._compute_activity_patterns(sessions)
         top_sessions = self._compute_top_sessions(sessions)
 
@@ -156,6 +167,7 @@ class InsightsEngine:
             "models": models,
             "platforms": platforms,
             "tools": tools,
+            "skills": skills,
             "activity": activity,
             "top_sessions": top_sessions,
         }
@@ -284,6 +296,82 @@ class InsightsEngine:
             for name, count in tool_counts.most_common()
         ]
 
+    def _get_skill_usage(self, cutoff: float, source: str = None) -> List[Dict]:
+        """Extract per-skill usage from assistant tool calls."""
+        skill_counts: Dict[str, Dict[str, Any]] = {}
+
+        if source:
+            cursor = self._conn.execute(
+                """SELECT m.tool_calls, m.timestamp
+                   FROM messages m
+                   JOIN sessions s ON s.id = m.session_id
+                   WHERE s.started_at >= ? AND s.source = ?
+                     AND m.role = 'assistant' AND m.tool_calls IS NOT NULL""",
+                (cutoff, source),
+            )
+        else:
+            cursor = self._conn.execute(
+                """SELECT m.tool_calls, m.timestamp
+                   FROM messages m
+                   JOIN sessions s ON s.id = m.session_id
+                   WHERE s.started_at >= ?
+                     AND m.role = 'assistant' AND m.tool_calls IS NOT NULL""",
+                (cutoff,),
+            )
+
+        for row in cursor.fetchall():
+            try:
+                calls = row["tool_calls"]
+                if isinstance(calls, str):
+                    calls = json.loads(calls)
+                if not isinstance(calls, list):
+                    continue
+            except (json.JSONDecodeError, TypeError):
+                continue
+
+            timestamp = row["timestamp"]
+            for call in calls:
+                if not isinstance(call, dict):
+                    continue
+                func = call.get("function", {})
+                tool_name = func.get("name")
+                if tool_name not in {"skill_view", "skill_manage"}:
+                    continue
+
+                args = func.get("arguments")
+                if isinstance(args, str):
+                    try:
+                        args = json.loads(args)
+                    except (json.JSONDecodeError, TypeError):
+                        continue
+                if not isinstance(args, dict):
+                    continue
+
+                skill_name = args.get("name")
+                if not isinstance(skill_name, str) or not skill_name.strip():
+                    continue
+
+                entry = skill_counts.setdefault(
+                    skill_name,
+                    {
+                        "skill": skill_name,
+                        "view_count": 0,
+                        "manage_count": 0,
+                        "last_used_at": None,
+                    },
+                )
+                if tool_name == "skill_view":
+                    entry["view_count"] += 1
+                else:
+                    entry["manage_count"] += 1
+
+                if timestamp is not None and (
+                    entry["last_used_at"] is None or timestamp > entry["last_used_at"]
+                ):
+                    entry["last_used_at"] = timestamp
+
+        return list(skill_counts.values())
+
     def _get_message_stats(self, cutoff: float, source: str = None) -> Dict:
         """Get aggregate message statistics."""
         if source:
@@ -475,6 +563,46 @@ class InsightsEngine:
             })
         return result
 
+    def _compute_skill_breakdown(self, skill_usage: List[Dict]) -> Dict[str, Any]:
+        """Process per-skill usage into summary + ranked list."""
+        total_skill_loads = sum(s["view_count"] for s in skill_usage) if skill_usage else 0
+        total_skill_edits = sum(s["manage_count"] for s in skill_usage) if skill_usage else 0
+        total_skill_actions = total_skill_loads + total_skill_edits
+
+        top_skills = []
+        for skill in skill_usage:
+            total_count = skill["view_count"] + skill["manage_count"]
+            percentage = (total_count / total_skill_actions * 100) if total_skill_actions else 0
+            top_skills.append({
+                "skill": skill["skill"],
+                "view_count": skill["view_count"],
+                "manage_count": skill["manage_count"],
+                "total_count": total_count,
+                "percentage": percentage,
+                "last_used_at": skill.get("last_used_at"),
+            })
+
+        top_skills.sort(
+            key=lambda s: (
+                s["total_count"],
+                s["view_count"],
+                s["manage_count"],
+                s["last_used_at"] or 0,
+                s["skill"],
+            ),
+            reverse=True,
+        )
+
+        return {
+            "summary": {
+                "total_skill_loads": total_skill_loads,
+                "total_skill_edits": total_skill_edits,
+                "total_skill_actions": total_skill_actions,
+                "distinct_skills_used": len(skill_usage),
+            },
+            "top_skills": top_skills,
+        }
+
     def _compute_activity_patterns(self, sessions: List[Dict]) -> Dict:
         """Analyze activity patterns by day of week and hour."""
         day_counts = Counter()  # 0=Monday ... 6=Sunday
@@ -670,6 +798,28 @@ class InsightsEngine:
                 lines.append(f"  ... and {len(report['tools']) - 15} more tools")
             lines.append("")
 
+        # Skill usage
+        skills = report.get("skills", {})
+        top_skills = skills.get("top_skills", [])
+        if top_skills:
+            lines.append("  🧠 Top Skills")
+            lines.append("  " + "─" * 56)
+            lines.append(f"  {'Skill':<28} {'Loads':>7} {'Edits':>7} {'Last used':>11}")
+            for skill in top_skills[:10]:
+                last_used = "—"
+                if skill.get("last_used_at"):
+                    last_used = datetime.fromtimestamp(skill["last_used_at"]).strftime("%b %d")
+                lines.append(
+                    f"  {skill['skill'][:28]:<28} {skill['view_count']:>7,} {skill['manage_count']:>7,} {last_used:>11}"
+                )
+            summary = skills.get("summary", {})
+            lines.append(
+                f"  Distinct skills: {summary.get('distinct_skills_used', 0)}  "
+                f"Loads: {summary.get('total_skill_loads', 0):,}  "
+                f"Edits: {summary.get('total_skill_edits', 0):,}"
+            )
+            lines.append("")
+
         # Activity patterns
         act = report.get("activity", {})
         if act.get("by_day"):
@@ -753,6 +903,18 @@ class InsightsEngine:
                 lines.append(f"  {t['tool']} — {t['count']:,} calls ({t['percentage']:.1f}%)")
             lines.append("")
 
+        skills = report.get("skills", {})
+        if skills.get("top_skills"):
+            lines.append("**🧠 Top Skills:**")
+            for skill in skills["top_skills"][:5]:
+                suffix = ""
+                if skill.get("last_used_at"):
+                    suffix = f", last used {datetime.fromtimestamp(skill['last_used_at']).strftime('%b %d')}"
+                lines.append(
+                    f"  {skill['skill']} — {skill['view_count']:,} loads, {skill['manage_count']:,} edits{suffix}"
+                )
+            lines.append("")
+
         # Activity summary
         act = report.get("activity", {})
         if act.get("busiest_day") and act.get("busiest_hour"):

agent/model_metadata.py

@@ -14,6 +14,8 @@ from urllib.parse import urlparse
 import requests
 import yaml
 
+from utils import base_url_host_matches, base_url_hostname
+
 from hermes_constants import OPENROUTER_MODELS_URL
 
 logger = logging.getLogger(__name__)
@@ -116,7 +118,6 @@ DEFAULT_CONTEXT_LENGTHS = {
     "gpt-5.4-nano": 400000,           # 400k (not 1.05M like full 5.4)
     "gpt-5.4-mini": 400000,           # 400k (not 1.05M like full 5.4)
     "gpt-5.4": 1050000,               # GPT-5.4, GPT-5.4 Pro (1.05M context)
-    "gpt-5.3-codex-spark": 128000,    # Spark variant has reduced 128k context
     "gpt-5.1-chat": 128000,           # Chat variant has 128k context
     "gpt-5": 400000,                  # GPT-5.x base, mini, codex variants (400k)
     "gpt-4.1": 1047576,
@@ -169,6 +170,7 @@ DEFAULT_CONTEXT_LENGTHS = {
     "Qwen/Qwen3.5-35B-A3B": 131072,
     "deepseek-ai/DeepSeek-V3.2": 65536,
     "moonshotai/Kimi-K2.5": 262144,
+    "moonshotai/Kimi-K2.6": 262144,
     "moonshotai/Kimi-K2-Thinking": 262144,
     "MiniMaxAI/MiniMax-M2.5": 204800,
     "XiaomiMiMo/MiMo-V2-Flash": 256000,
@@ -211,8 +213,15 @@ def _normalize_base_url(base_url: str) -> str:
     return (base_url or "").strip().rstrip("/")
 
 
+def _auth_headers(api_key: str = "") -> Dict[str, str]:
+    token = str(api_key or "").strip()
+    if not token:
+        return {}
+    return {"Authorization": f"Bearer {token}"}
+
+
 def _is_openrouter_base_url(base_url: str) -> bool:
-    return "openrouter.ai" in _normalize_base_url(base_url).lower()
+    return base_url_host_matches(base_url, "openrouter.ai")
 
 
 def _is_custom_endpoint(base_url: str) -> bool:
@@ -310,7 +319,7 @@ def is_local_endpoint(base_url: str) -> bool:
     return False
 
 
-def detect_local_server_type(base_url: str) -> Optional[str]:
+def detect_local_server_type(base_url: str, api_key: str = "") -> Optional[str]:
     """Detect which local server is running at base_url by probing known endpoints.
 
     Returns one of: "ollama", "lm-studio", "vllm", "llamacpp", or None.
@@ -322,8 +331,10 @@ def detect_local_server_type(base_url: str) -> Optional[str]:
     if server_url.endswith("/v1"):
         server_url = server_url[:-3]
 
+    headers = _auth_headers(api_key)
+
     try:
-        with httpx.Client(timeout=2.0) as client:
+        with httpx.Client(timeout=2.0, headers=headers) as client:
             # LM Studio exposes /api/v1/models — check first (most specific)
             try:
                 r = client.get(f"{server_url}/api/v1/models")
@@ -510,6 +521,59 @@ def fetch_endpoint_model_metadata(
     headers = {"Authorization": f"Bearer {api_key}"} if api_key else {}
     last_error: Optional[Exception] = None
 
+    if is_local_endpoint(normalized):
+        try:
+            if detect_local_server_type(normalized, api_key=api_key) == "lm-studio":
+                server_url = normalized[:-3].rstrip("/") if normalized.endswith("/v1") else normalized
+                response = requests.get(
+                    server_url.rstrip("/") + "/api/v1/models",
+                    headers=headers,
+                    timeout=10,
+                )
+                response.raise_for_status()
+                payload = response.json()
+                cache: Dict[str, Dict[str, Any]] = {}
+                for model in payload.get("models", []):
+                    if not isinstance(model, dict):
+                        continue
+                    model_id = model.get("key") or model.get("id")
+                    if not model_id:
+                        continue
+                    entry: Dict[str, Any] = {"name": model.get("name", model_id)}
+
+                    context_length = None
+                    for inst in model.get("loaded_instances", []) or []:
+                        if not isinstance(inst, dict):
+                            continue
+                        cfg = inst.get("config", {})
+                        ctx = cfg.get("context_length") if isinstance(cfg, dict) else None
+                        if isinstance(ctx, int) and ctx > 0:
+                            context_length = ctx
+                            break
+                    if context_length is None:
+                        context_length = _extract_context_length(model)
+                    if context_length is not None:
+                        entry["context_length"] = context_length
+
+                    max_completion_tokens = _extract_max_completion_tokens(model)
+                    if max_completion_tokens is not None:
+                        entry["max_completion_tokens"] = max_completion_tokens
+
+                    pricing = _extract_pricing(model)
+                    if pricing:
+                        entry["pricing"] = pricing
+
+                    _add_model_aliases(cache, model_id, entry)
+                    alt_id = model.get("id")
+                    if isinstance(alt_id, str) and alt_id and alt_id != model_id:
+                        _add_model_aliases(cache, alt_id, entry)
+
+                _endpoint_model_metadata_cache[normalized] = cache
+                _endpoint_model_metadata_cache_time[normalized] = time.time()
+                return cache
+        except Exception as exc:
+            last_error = exc
+
     for candidate in candidates:
         url = candidate.rstrip("/") + "/models"
         try:
@@ -716,7 +780,7 @@ def _model_id_matches(candidate_id: str, lookup_model: str) -> bool:
     return False
 
 
-def query_ollama_num_ctx(model: str, base_url: str) -> Optional[int]:
+def query_ollama_num_ctx(model: str, base_url: str, api_key: str = "") -> Optional[int]:
     """Query an Ollama server for the model's context length.
 
     Returns the model's maximum context from GGUF metadata via ``/api/show``,
@@ -734,14 +798,16 @@ def query_ollama_num_ctx(model: str, base_url: str) -> Optional[int]:
         server_url = server_url[:-3]
 
     try:
-        server_type = detect_local_server_type(base_url)
+        server_type = detect_local_server_type(base_url, api_key=api_key)
     except Exception:
         return None
     if server_type != "ollama":
         return None
 
+    headers = _auth_headers(api_key)
+
     try:
-        with httpx.Client(timeout=3.0) as client:
+        with httpx.Client(timeout=3.0, headers=headers) as client:
             resp = client.post(f"{server_url}/api/show", json={"name": bare_model})
             if resp.status_code != 200:
                 return None
@@ -769,7 +835,7 @@ def query_ollama_num_ctx(model: str, base_url: str) -> Optional[int]:
     return None
 
 
-def _query_local_context_length(model: str, base_url: str) -> Optional[int]:
+def _query_local_context_length(model: str, base_url: str, api_key: str = "") -> Optional[int]:
     """Query a local server for the model's context length."""
     import httpx
 
@@ -782,13 +848,15 @@ def _query_local_context_length(model: str, base_url: str) -> Optional[int]:
     if server_url.endswith("/v1"):
         server_url = server_url[:-3]
 
+    headers = _auth_headers(api_key)
+
     try:
-        server_type = detect_local_server_type(base_url)
+        server_type = detect_local_server_type(base_url, api_key=api_key)
     except Exception:
         server_type = None
 
     try:
-        with httpx.Client(timeout=3.0) as client:
+        with httpx.Client(timeout=3.0, headers=headers) as client:
             # Ollama: /api/show returns model details with context info
             if server_type == "ollama":
                 resp = client.post(f"{server_url}/api/show", json={"name": model})
@@ -999,7 +1067,7 @@ def get_model_context_length(
         if not _is_known_provider_base_url(base_url):
             # 3. Try querying local server directly
             if is_local_endpoint(base_url):
-                local_ctx = _query_local_context_length(model, base_url)
+                local_ctx = _query_local_context_length(model, base_url, api_key=api_key)
                 if local_ctx and local_ctx > 0:
                     save_context_length(model, base_url, local_ctx)
                     return local_ctx
@@ -1013,7 +1081,7 @@ def get_model_context_length(
 
     # 4. Anthropic /v1/models API (only for regular API keys, not OAuth)
     if provider == "anthropic" or (
-        base_url and "api.anthropic.com" in base_url
+        base_url and base_url_hostname(base_url) == "api.anthropic.com"
     ):
         ctx = _query_anthropic_context_length(model, base_url or "https://api.anthropic.com", api_key)
         if ctx:
@@ -1022,7 +1090,11 @@ def get_model_context_length(
     # 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):
+    if provider == "bedrock" or (
+        base_url
+        and base_url_hostname(base_url).startswith("bedrock-runtime.")
+        and base_url_host_matches(base_url, "amazonaws.com")
+    ):
         try:
             from agent.bedrock_adapter import get_bedrock_context_length
             return get_bedrock_context_length(model)
@@ -1069,7 +1141,7 @@ def get_model_context_length(
 
     # 9. Query local server as last resort
     if base_url and is_local_endpoint(base_url):
-        local_ctx = _query_local_context_length(model, base_url)
+        local_ctx = _query_local_context_length(model, base_url, api_key=api_key)
         if local_ctx and local_ctx > 0:
             save_context_length(model, base_url, local_ctx)
             return local_ctx

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 = (
@@ -613,12 +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)
@@ -626,8 +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)
 

agent/redact.py

@@ -13,6 +13,48 @@ import re
 
 logger = logging.getLogger(__name__)
 
+# Sensitive query-string parameter names (case-insensitive exact match).
+# Ported from nearai/ironclaw#2529 — catches tokens whose values don't match
+# any known vendor prefix regex (e.g. opaque tokens, short OAuth codes).
+_SENSITIVE_QUERY_PARAMS = frozenset({
+    "access_token",
+    "refresh_token",
+    "id_token",
+    "token",
+    "api_key",
+    "apikey",
+    "client_secret",
+    "password",
+    "auth",
+    "jwt",
+    "session",
+    "secret",
+    "key",
+    "code",           # OAuth authorization codes
+    "signature",      # pre-signed URL signatures
+    "x-amz-signature",
+})
+
+# Sensitive form-urlencoded / JSON body key names (case-insensitive exact match).
+# Exact match, NOT substring — "token_count" and "session_id" must NOT match.
+# Ported from nearai/ironclaw#2529.
+_SENSITIVE_BODY_KEYS = frozenset({
+    "access_token",
+    "refresh_token",
+    "id_token",
+    "token",
+    "api_key",
+    "apikey",
+    "client_secret",
+    "password",
+    "auth",
+    "jwt",
+    "secret",
+    "private_key",
+    "authorization",
+    "key",
+})
+
 # Snapshot at import time so runtime env mutations (e.g. LLM-generated
 # `export HERMES_REDACT_SECRETS=false`) cannot disable redaction mid-session.
 _REDACT_ENABLED = os.getenv("HERMES_REDACT_SECRETS", "").lower() not in ("0", "false", "no", "off")
@@ -108,6 +150,30 @@ _DISCORD_MENTION_RE = re.compile(r"<@!?(\d{17,20})>")
 # Negative lookahead prevents matching hex strings or identifiers
 _SIGNAL_PHONE_RE = re.compile(r"(\+[1-9]\d{6,14})(?![A-Za-z0-9])")
 
+# URLs containing query strings — matches `scheme://...?...[# or end]`.
+# Used to scan text for URLs whose query params may contain secrets.
+# Ported from nearai/ironclaw#2529.
+_URL_WITH_QUERY_RE = re.compile(
+    r"(https?|wss?|ftp)://"          # scheme
+    r"([^\s/?#]+)"                    # authority (may include userinfo)
+    r"([^\s?#]*)"                     # path
+    r"\?([^\s#]+)"                    # query (required)
+    r"(#\S*)?",                       # optional fragment
+)
+
+# URLs containing userinfo — `scheme://user:password@host` for ANY scheme
+# (not just DB protocols already covered by _DB_CONNSTR_RE above).
+# Catches things like `https://user:token@api.example.com/v1/foo`.
+_URL_USERINFO_RE = re.compile(
+    r"(https?|wss?|ftp)://([^/\s:@]+):([^/\s@]+)@",
+)
+
+# Form-urlencoded body detection: conservative — only applies when the entire
+# text looks like a query string (k=v&k=v pattern with no newlines).
+_FORM_BODY_RE = re.compile(
+    r"^[A-Za-z_][A-Za-z0-9_.-]*=[^&\s]*(?:&[A-Za-z_][A-Za-z0-9_.-]*=[^&\s]*)+$"
+)
+
 # Compile known prefix patterns into one alternation
 _PREFIX_RE = re.compile(
     r"(?<![A-Za-z0-9_-])(" + "|".join(_PREFIX_PATTERNS) + r")(?![A-Za-z0-9_-])"
@@ -121,6 +187,72 @@ def _mask_token(token: str) -> str:
     return f"{token[:6]}...{token[-4:]}"
 
 
+def _redact_query_string(query: str) -> str:
+    """Redact sensitive parameter values in a URL query string.
+
+    Handles `k=v&k=v` format. Sensitive keys (case-insensitive) have values
+    replaced with `***`. Non-sensitive keys pass through unchanged.
+    Empty or malformed pairs are preserved as-is.
+    """
+    if not query:
+        return query
+    parts = []
+    for pair in query.split("&"):
+        if "=" not in pair:
+            parts.append(pair)
+            continue
+        key, _, value = pair.partition("=")
+        if key.lower() in _SENSITIVE_QUERY_PARAMS:
+            parts.append(f"{key}=***")
+        else:
+            parts.append(pair)
+    return "&".join(parts)
+
+
+def _redact_url_query_params(text: str) -> str:
+    """Scan text for URLs with query strings and redact sensitive params.
+
+    Catches opaque tokens that don't match vendor prefix regexes, e.g.
+    `https://example.com/cb?code=ABC123&state=xyz` → `...?code=***&state=xyz`.
+    """
+    def _sub(m: re.Match) -> str:
+        scheme = m.group(1)
+        authority = m.group(2)
+        path = m.group(3)
+        query = _redact_query_string(m.group(4))
+        fragment = m.group(5) or ""
+        return f"{scheme}://{authority}{path}?{query}{fragment}"
+    return _URL_WITH_QUERY_RE.sub(_sub, text)
+
+
+def _redact_url_userinfo(text: str) -> str:
+    """Strip `user:password@` from HTTP/WS/FTP URLs.
+
+    DB protocols (postgres, mysql, mongodb, redis, amqp) are handled
+    separately by `_DB_CONNSTR_RE`.
+    """
+    return _URL_USERINFO_RE.sub(
+        lambda m: f"{m.group(1)}://{m.group(2)}:***@",
+        text,
+    )
+
+
+def _redact_form_body(text: str) -> str:
+    """Redact sensitive values in a form-urlencoded body.
+
+    Only applies when the entire input looks like a pure form body
+    (k=v&k=v with no newlines, no other text). Single-line non-form
+    text passes through unchanged. This is a conservative pass — the
+    `_redact_url_query_params` function handles embedded query strings.
+    """
+    if not text or "\n" in text or "&" not in text:
+        return text
+    # The body-body form check is strict: only trigger on clean k=v&k=v.
+    if not _FORM_BODY_RE.match(text.strip()):
+        return text
+    return _redact_query_string(text.strip())
+
+
 def redact_sensitive_text(text: str) -> str:
     """Apply all redaction patterns to a block of text.
 
@@ -173,6 +305,16 @@ def redact_sensitive_text(text: str) -> str:
     # JWT tokens (eyJ... — base64-encoded JSON headers)
     text = _JWT_RE.sub(lambda m: _mask_token(m.group(0)), text)
 
+    # URL userinfo (http(s)://user:pass@host) — redact for non-DB schemes.
+    # DB schemes are handled above by _DB_CONNSTR_RE.
+    text = _redact_url_userinfo(text)
+
+    # URL query params containing opaque tokens (?access_token=…&code=…)
+    text = _redact_url_query_params(text)
+
+    # Form-urlencoded bodies (only triggers on clean k=v&k=v inputs).
+    text = _redact_form_body(text)
+
     # Discord user/role mentions (<@snowflake_id>)
     text = _DISCORD_MENTION_RE.sub(lambda m: f"<@{'!' if '!' in m.group(0) else ''}***>", text)
 

agent/shell_hooks.py

@@ -0,0 +1,831 @@
+"""
+Shell-script hooks bridge.
+
+Reads the ``hooks:`` block from ``cli-config.yaml``, prompts the user for
+consent on first use of each ``(event, command)`` pair, and registers
+callbacks on the existing plugin hook manager so every existing
+``invoke_hook()`` site dispatches to the configured shell scripts — with
+zero changes to call sites.
+
+Design notes
+------------
+* Python plugins and shell hooks compose naturally: both flow through
+  :func:`hermes_cli.plugins.invoke_hook` and its aggregators.  Python
+  plugins are registered first (via ``discover_and_load()``) so their
+  block decisions win ties over shell-hook blocks.
+* Subprocess execution uses ``shlex.split(os.path.expanduser(command))``
+  with ``shell=False`` — no shell injection footguns.  Users that need
+  pipes/redirection wrap their logic in a script.
+* First-use consent is gated by the allowlist under
+  ``~/.hermes/shell-hooks-allowlist.json``.  Non-TTY callers must pass
+  ``accept_hooks=True`` (resolved from ``--accept-hooks``,
+  ``HERMES_ACCEPT_HOOKS``, or ``hooks_auto_accept: true`` in config)
+  for registration to succeed without a prompt.
+* Registration is idempotent — safe to invoke from both the CLI entry
+  point (``hermes_cli/main.py``) and the gateway entry point
+  (``gateway/run.py``).
+
+Wire protocol
+-------------
+**stdin** (JSON, piped to the script)::
+
+    {
+        "hook_event_name": "pre_tool_call",
+        "tool_name":       "terminal",
+        "tool_input":      {"command": "rm -rf /"},
+        "session_id":      "sess_abc123",
+        "cwd":             "/home/user/project",
+        "extra":           {...}   # event-specific kwargs
+    }
+
+**stdout** (JSON, optional — anything else is ignored)::
+
+    # Block a pre_tool_call (either shape accepted; normalised internally):
+    {"decision": "block", "reason":  "Forbidden command"}   # Claude-Code-style
+    {"action":   "block", "message": "Forbidden command"}   # Hermes-canonical
+
+    # Inject context for pre_llm_call:
+    {"context": "Today is Friday"}
+
+    # Silent no-op:
+    <empty or any non-matching JSON object>
+"""
+
+from __future__ import annotations
+
+import difflib
+import json
+import logging
+import os
+import re
+import shlex
+import subprocess
+import sys
+import tempfile
+import threading
+import time
+from contextlib import contextmanager
+from dataclasses import dataclass, field
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Callable, Dict, Iterator, List, Optional, Set, Tuple
+
+try:
+    import fcntl  # POSIX only; Windows falls back to best-effort without flock.
+except ImportError:  # pragma: no cover
+    fcntl = None  # type: ignore[assignment]
+
+from hermes_constants import get_hermes_home
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_TIMEOUT_SECONDS = 60
+MAX_TIMEOUT_SECONDS = 300
+ALLOWLIST_FILENAME = "shell-hooks-allowlist.json"
+
+# (event, matcher, command) triples that have been wired to the plugin
+# manager in the current process.  Matcher is part of the key because
+# the same script can legitimately register for different matchers under
+# the same event (e.g. one entry per tool the user wants to gate).
+# Second registration attempts for the exact same triple become no-ops
+# so the CLI and gateway can both call register_from_config() safely.
+_registered: Set[Tuple[str, Optional[str], str]] = set()
+_registered_lock = threading.Lock()
+
+# Intra-process lock for allowlist read-modify-write on platforms that
+# lack ``fcntl`` (non-POSIX).  Kept separate from ``_registered_lock``
+# because ``register_from_config`` already holds ``_registered_lock`` when
+# it triggers ``_record_approval`` — reusing it here would self-deadlock
+# (``threading.Lock`` is non-reentrant).  POSIX callers use the sibling
+# ``.lock`` file via ``fcntl.flock`` and bypass this.
+_allowlist_write_lock = threading.Lock()
+
+
+@dataclass
+class ShellHookSpec:
+    """Parsed and validated representation of a single ``hooks:`` entry."""
+
+    event: str
+    command: str
+    matcher: Optional[str] = None
+    timeout: int = DEFAULT_TIMEOUT_SECONDS
+    compiled_matcher: Optional[re.Pattern] = field(default=None, repr=False)
+
+    def __post_init__(self) -> None:
+        # Strip whitespace introduced by YAML quirks (e.g. multi-line string
+        # folding) — a matcher of " terminal" would otherwise silently fail
+        # to match "terminal" without any diagnostic.
+        if isinstance(self.matcher, str):
+            stripped = self.matcher.strip()
+            self.matcher = stripped if stripped else None
+        if self.matcher:
+            try:
+                self.compiled_matcher = re.compile(self.matcher)
+            except re.error as exc:
+                logger.warning(
+                    "shell hook matcher %r is invalid (%s) — treating as "
+                    "literal equality", self.matcher, exc,
+                )
+                self.compiled_matcher = None
+
+    def matches_tool(self, tool_name: Optional[str]) -> bool:
+        if not self.matcher:
+            return True
+        if tool_name is None:
+            return False
+        if self.compiled_matcher is not None:
+            return self.compiled_matcher.fullmatch(tool_name) is not None
+        # compiled_matcher is None only when the regex failed to compile,
+        # in which case we already warned and fall back to literal equality.
+        return tool_name == self.matcher
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+def register_from_config(
+    cfg: Optional[Dict[str, Any]],
+    *,
+    accept_hooks: bool = False,
+) -> List[ShellHookSpec]:
+    """Register every configured shell hook on the plugin manager.
+
+    ``cfg`` is the full parsed config dict (``hermes_cli.config.load_config``
+    output).  The ``hooks:`` key is read out of it.  Missing, empty, or
+    non-dict ``hooks`` is treated as zero configured hooks.
+
+    ``accept_hooks=True`` skips the TTY consent prompt — the caller is
+    promising that the user has opted in via a flag, env var, or config
+    setting.  ``HERMES_ACCEPT_HOOKS=1`` and ``hooks_auto_accept: true`` are
+    also honored inside this function so either CLI or gateway call sites
+    pick them up.
+
+    Returns the list of :class:`ShellHookSpec` entries that ended up wired
+    up on the plugin manager.  Skipped entries (unknown events, malformed,
+    not allowlisted, already registered) are logged but not returned.
+    """
+    if not isinstance(cfg, dict):
+        return []
+
+    effective_accept = _resolve_effective_accept(cfg, accept_hooks)
+
+    specs = _parse_hooks_block(cfg.get("hooks"))
+    if not specs:
+        return []
+
+    registered: List[ShellHookSpec] = []
+
+    # Import lazily — avoids circular imports at module-load time.
+    from hermes_cli.plugins import get_plugin_manager
+
+    manager = get_plugin_manager()
+
+    # Idempotence + allowlist read happen under the lock; the TTY
+    # prompt runs outside so other threads aren't parked on a blocking
+    # input().  Mutation re-takes the lock with a defensive idempotence
+    # re-check in case two callers ever race through the prompt.
+    for spec in specs:
+        key = (spec.event, spec.matcher, spec.command)
+        with _registered_lock:
+            if key in _registered:
+                continue
+            already_allowlisted = _is_allowlisted(spec.event, spec.command)
+
+        if not already_allowlisted:
+            if not _prompt_and_record(
+                spec.event, spec.command, accept_hooks=effective_accept,
+            ):
+                logger.warning(
+                    "shell hook for %s (%s) not allowlisted — skipped. "
+                    "Use --accept-hooks / HERMES_ACCEPT_HOOKS=1 / "
+                    "hooks_auto_accept: true, or approve at the TTY "
+                    "prompt next run.",
+                    spec.event, spec.command,
+                )
+                continue
+
+        with _registered_lock:
+            if key in _registered:
+                continue
+            manager._hooks.setdefault(spec.event, []).append(_make_callback(spec))
+            _registered.add(key)
+            registered.append(spec)
+            logger.info(
+                "shell hook registered: %s -> %s (matcher=%s, timeout=%ds)",
+                spec.event, spec.command, spec.matcher, spec.timeout,
+            )
+
+    return registered
+
+
+def iter_configured_hooks(cfg: Optional[Dict[str, Any]]) -> List[ShellHookSpec]:
+    """Return the parsed ``ShellHookSpec`` entries from config without
+    registering anything.  Used by ``hermes hooks list`` and ``doctor``."""
+    if not isinstance(cfg, dict):
+        return []
+    return _parse_hooks_block(cfg.get("hooks"))
+
+
+def reset_for_tests() -> None:
+    """Clear the idempotence set.  Test-only helper."""
+    with _registered_lock:
+        _registered.clear()
+
+
+# ---------------------------------------------------------------------------
+# Config parsing
+# ---------------------------------------------------------------------------
+
+def _parse_hooks_block(hooks_cfg: Any) -> List[ShellHookSpec]:
+    """Normalise the ``hooks:`` dict into a flat list of ``ShellHookSpec``.
+
+    Malformed entries warn-and-skip — we never raise from config parsing
+    because a broken hook must not crash the agent.
+    """
+    from hermes_cli.plugins import VALID_HOOKS
+
+    if not isinstance(hooks_cfg, dict):
+        return []
+
+    specs: List[ShellHookSpec] = []
+
+    for event_name, entries in hooks_cfg.items():
+        if event_name not in VALID_HOOKS:
+            suggestion = difflib.get_close_matches(
+                str(event_name), VALID_HOOKS, n=1, cutoff=0.6,
+            )
+            if suggestion:
+                logger.warning(
+                    "unknown hook event %r in hooks: config — did you mean %r?",
+                    event_name, suggestion[0],
+                )
+            else:
+                logger.warning(
+                    "unknown hook event %r in hooks: config (valid: %s)",
+                    event_name, ", ".join(sorted(VALID_HOOKS)),
+                )
+            continue
+
+        if entries is None:
+            continue
+
+        if not isinstance(entries, list):
+            logger.warning(
+                "hooks.%s must be a list of hook definitions; got %s",
+                event_name, type(entries).__name__,
+            )
+            continue
+
+        for i, raw in enumerate(entries):
+            spec = _parse_single_entry(event_name, i, raw)
+            if spec is not None:
+                specs.append(spec)
+
+    return specs
+
+
+def _parse_single_entry(
+    event: str, index: int, raw: Any,
+) -> Optional[ShellHookSpec]:
+    if not isinstance(raw, dict):
+        logger.warning(
+            "hooks.%s[%d] must be a mapping with a 'command' key; got %s",
+            event, index, type(raw).__name__,
+        )
+        return None
+
+    command = raw.get("command")
+    if not isinstance(command, str) or not command.strip():
+        logger.warning(
+            "hooks.%s[%d] is missing a non-empty 'command' field",
+            event, index,
+        )
+        return None
+
+    matcher = raw.get("matcher")
+    if matcher is not None and not isinstance(matcher, str):
+        logger.warning(
+            "hooks.%s[%d].matcher must be a string regex; ignoring",
+            event, index,
+        )
+        matcher = None
+
+    if matcher is not None and event not in ("pre_tool_call", "post_tool_call"):
+        logger.warning(
+            "hooks.%s[%d].matcher=%r will be ignored at runtime — the "
+            "matcher field is only honored for pre_tool_call / "
+            "post_tool_call.  The hook will fire on every %s event.",
+            event, index, matcher, event,
+        )
+        matcher = None
+
+    timeout_raw = raw.get("timeout", DEFAULT_TIMEOUT_SECONDS)
+    try:
+        timeout = int(timeout_raw)
+    except (TypeError, ValueError):
+        logger.warning(
+            "hooks.%s[%d].timeout must be an int (got %r); using default %ds",
+            event, index, timeout_raw, DEFAULT_TIMEOUT_SECONDS,
+        )
+        timeout = DEFAULT_TIMEOUT_SECONDS
+
+    if timeout < 1:
+        logger.warning(
+            "hooks.%s[%d].timeout must be >=1; using default %ds",
+            event, index, DEFAULT_TIMEOUT_SECONDS,
+        )
+        timeout = DEFAULT_TIMEOUT_SECONDS
+
+    if timeout > MAX_TIMEOUT_SECONDS:
+        logger.warning(
+            "hooks.%s[%d].timeout=%ds exceeds max %ds; clamping",
+            event, index, timeout, MAX_TIMEOUT_SECONDS,
+        )
+        timeout = MAX_TIMEOUT_SECONDS
+
+    return ShellHookSpec(
+        event=event,
+        command=command.strip(),
+        matcher=matcher,
+        timeout=timeout,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Subprocess callback
+# ---------------------------------------------------------------------------
+
+_TOP_LEVEL_PAYLOAD_KEYS = {"tool_name", "args", "session_id", "parent_session_id"}
+
+
+def _spawn(spec: ShellHookSpec, stdin_json: str) -> Dict[str, Any]:
+    """Run ``spec.command`` as a subprocess with ``stdin_json`` on stdin.
+
+    Returns a diagnostic dict with the same keys for every outcome
+    (``returncode``, ``stdout``, ``stderr``, ``timed_out``,
+    ``elapsed_seconds``, ``error``).  This is the single place the
+    subprocess is actually invoked — both the live callback path
+    (:func:`_make_callback`) and the CLI test helper (:func:`run_once`)
+    go through it.
+    """
+    result: Dict[str, Any] = {
+        "returncode": None,
+        "stdout": "",
+        "stderr": "",
+        "timed_out": False,
+        "elapsed_seconds": 0.0,
+        "error": None,
+    }
+    try:
+        argv = shlex.split(os.path.expanduser(spec.command))
+    except ValueError as exc:
+        result["error"] = f"command {spec.command!r} cannot be parsed: {exc}"
+        return result
+    if not argv:
+        result["error"] = "empty command"
+        return result
+
+    t0 = time.monotonic()
+    try:
+        proc = subprocess.run(
+            argv,
+            input=stdin_json,
+            capture_output=True,
+            timeout=spec.timeout,
+            text=True,
+            shell=False,
+        )
+    except subprocess.TimeoutExpired:
+        result["timed_out"] = True
+        result["elapsed_seconds"] = round(time.monotonic() - t0, 3)
+        return result
+    except FileNotFoundError:
+        result["error"] = "command not found"
+        return result
+    except PermissionError:
+        result["error"] = "command not executable"
+        return result
+    except Exception as exc:  # pragma: no cover — defensive
+        result["error"] = str(exc)
+        return result
+
+    result["returncode"] = proc.returncode
+    result["stdout"] = proc.stdout or ""
+    result["stderr"] = proc.stderr or ""
+    result["elapsed_seconds"] = round(time.monotonic() - t0, 3)
+    return result
+
+
+def _make_callback(spec: ShellHookSpec) -> Callable[..., Optional[Dict[str, Any]]]:
+    """Build the closure that ``invoke_hook()`` will call per firing."""
+
+    def _callback(**kwargs: Any) -> Optional[Dict[str, Any]]:
+        # Matcher gate — only meaningful for tool-scoped events.
+        if spec.event in ("pre_tool_call", "post_tool_call"):
+            if not spec.matches_tool(kwargs.get("tool_name")):
+                return None
+
+        r = _spawn(spec, _serialize_payload(spec.event, kwargs))
+
+        if r["error"]:
+            logger.warning(
+                "shell hook failed (event=%s command=%s): %s",
+                spec.event, spec.command, r["error"],
+            )
+            return None
+        if r["timed_out"]:
+            logger.warning(
+                "shell hook timed out after %.2fs (event=%s command=%s)",
+                r["elapsed_seconds"], spec.event, spec.command,
+            )
+            return None
+
+        stderr = r["stderr"].strip()
+        if stderr:
+            logger.debug(
+                "shell hook stderr (event=%s command=%s): %s",
+                spec.event, spec.command, stderr[:400],
+            )
+        # Non-zero exits: log but still parse stdout so scripts that
+        # signal failure via exit code can also return a block directive.
+        if r["returncode"] != 0:
+            logger.warning(
+                "shell hook exited %d (event=%s command=%s); stderr=%s",
+                r["returncode"], spec.event, spec.command, stderr[:400],
+            )
+        return _parse_response(spec.event, r["stdout"])
+
+    _callback.__name__ = f"shell_hook[{spec.event}:{spec.command}]"
+    _callback.__qualname__ = _callback.__name__
+    return _callback
+
+
+def _serialize_payload(event: str, kwargs: Dict[str, Any]) -> str:
+    """Render the stdin JSON payload.  Unserialisable values are
+    stringified via ``default=str`` rather than dropped."""
+    extras = {k: v for k, v in kwargs.items() if k not in _TOP_LEVEL_PAYLOAD_KEYS}
+    try:
+        cwd = str(Path.cwd())
+    except OSError:
+        cwd = ""
+    payload = {
+        "hook_event_name": event,
+        "tool_name": kwargs.get("tool_name"),
+        "tool_input": kwargs.get("args") if isinstance(kwargs.get("args"), dict) else None,
+        "session_id": kwargs.get("session_id") or kwargs.get("parent_session_id") or "",
+        "cwd": cwd,
+        "extra": extras,
+    }
+    return json.dumps(payload, ensure_ascii=False, default=str)
+
+
+def _parse_response(event: str, stdout: str) -> Optional[Dict[str, Any]]:
+    """Translate stdout JSON into a Hermes wire-shape dict.
+
+    For ``pre_tool_call`` the Claude-Code-style ``{"decision": "block",
+    "reason": "..."}`` payload is translated into the canonical Hermes
+    ``{"action": "block", "message": "..."}`` shape expected by
+    :func:`hermes_cli.plugins.get_pre_tool_call_block_message`.  This is
+    the single most important correctness invariant in this module —
+    skipping the translation silently breaks every ``pre_tool_call``
+    block directive.
+
+    For ``pre_llm_call``, ``{"context": "..."}`` is passed through
+    unchanged to match the existing plugin-hook contract.
+
+    Anything else returns ``None``.
+    """
+    stdout = (stdout or "").strip()
+    if not stdout:
+        return None
+
+    try:
+        data = json.loads(stdout)
+    except json.JSONDecodeError:
+        logger.warning(
+            "shell hook stdout was not valid JSON (event=%s): %s",
+            event, stdout[:200],
+        )
+        return None
+
+    if not isinstance(data, dict):
+        return None
+
+    if event == "pre_tool_call":
+        if data.get("action") == "block":
+            message = data.get("message") or data.get("reason") or ""
+            if isinstance(message, str) and message:
+                return {"action": "block", "message": message}
+        if data.get("decision") == "block":
+            message = data.get("reason") or data.get("message") or ""
+            if isinstance(message, str) and message:
+                return {"action": "block", "message": message}
+        return None
+
+    context = data.get("context")
+    if isinstance(context, str) and context.strip():
+        return {"context": context}
+
+    return None
+
+
+# ---------------------------------------------------------------------------
+# Allowlist / consent
+# ---------------------------------------------------------------------------
+
+def allowlist_path() -> Path:
+    """Path to the per-user shell-hook allowlist file."""
+    return get_hermes_home() / ALLOWLIST_FILENAME
+
+
+def load_allowlist() -> Dict[str, Any]:
+    """Return the parsed allowlist, or an empty skeleton if absent."""
+    try:
+        raw = json.loads(allowlist_path().read_text())
+    except (FileNotFoundError, json.JSONDecodeError, OSError):
+        return {"approvals": []}
+    if not isinstance(raw, dict):
+        return {"approvals": []}
+    approvals = raw.get("approvals")
+    if not isinstance(approvals, list):
+        raw["approvals"] = []
+    return raw
+
+
+def save_allowlist(data: Dict[str, Any]) -> None:
+    """Atomically persist the allowlist via per-process ``mkstemp`` +
+    ``os.replace``.  Cross-process read-modify-write races are handled
+    by :func:`_locked_update_approvals` (``fcntl.flock``).  On OSError
+    the failure is logged; the in-process hook still registers but
+    the approval won't survive across runs."""
+    p = allowlist_path()
+    try:
+        p.parent.mkdir(parents=True, exist_ok=True)
+        fd, tmp_path = tempfile.mkstemp(
+            prefix=f"{p.name}.", suffix=".tmp", dir=str(p.parent),
+        )
+        try:
+            with os.fdopen(fd, "w") as fh:
+                fh.write(json.dumps(data, indent=2, sort_keys=True))
+            os.replace(tmp_path, p)
+        except Exception:
+            try:
+                os.unlink(tmp_path)
+            except OSError:
+                pass
+            raise
+    except OSError as exc:
+        logger.warning(
+            "Failed to persist shell hook allowlist to %s: %s. "
+            "The approval is in-memory for this run, but the next "
+            "startup will re-prompt (or skip registration on non-TTY "
+            "runs without --accept-hooks / HERMES_ACCEPT_HOOKS).",
+            p, exc,
+        )
+
+
+def _is_allowlisted(event: str, command: str) -> bool:
+    data = load_allowlist()
+    return any(
+        isinstance(e, dict)
+        and e.get("event") == event
+        and e.get("command") == command
+        for e in data.get("approvals", [])
+    )
+
+
+@contextmanager
+def _locked_update_approvals() -> Iterator[Dict[str, Any]]:
+    """Serialise read-modify-write on the allowlist across processes.
+
+    Holds an exclusive ``flock`` on a sibling lock file for the duration
+    of the update so concurrent ``_record_approval``/``revoke`` callers
+    cannot clobber each other's changes (the race Codex reproduced with
+    20–50 simultaneous writers).  Falls back to an in-process lock on
+    platforms without ``fcntl``.
+    """
+    p = allowlist_path()
+    p.parent.mkdir(parents=True, exist_ok=True)
+    lock_path = p.with_suffix(p.suffix + ".lock")
+
+    if fcntl is None:  # pragma: no cover — non-POSIX fallback
+        with _allowlist_write_lock:
+            data = load_allowlist()
+            yield data
+            save_allowlist(data)
+        return
+
+    with open(lock_path, "a+") as lock_fh:
+        fcntl.flock(lock_fh.fileno(), fcntl.LOCK_EX)
+        try:
+            data = load_allowlist()
+            yield data
+            save_allowlist(data)
+        finally:
+            fcntl.flock(lock_fh.fileno(), fcntl.LOCK_UN)
+
+
+def _prompt_and_record(
+    event: str, command: str, *, accept_hooks: bool,
+) -> bool:
+    """Decide whether to approve an unseen ``(event, command)`` pair.
+    Returns ``True`` iff the approval was granted and recorded.
+    """
+    if accept_hooks:
+        _record_approval(event, command)
+        logger.info(
+            "shell hook auto-approved via --accept-hooks / env / config: "
+            "%s -> %s", event, command,
+        )
+        return True
+
+    if not sys.stdin.isatty():
+        return False
+
+    print(
+        f"\n⚠ Hermes is about to register a shell hook that will run a\n"
+        f"  command on your behalf.\n\n"
+        f"    Event:   {event}\n"
+        f"    Command: {command}\n\n"
+        f"  Commands run with your full user credentials.  Only approve\n"
+        f"  commands you trust."
+    )
+    try:
+        answer = input("Allow this hook to run? [y/N]: ").strip().lower()
+    except (EOFError, KeyboardInterrupt):
+        print()  # keep the terminal tidy after ^C
+        return False
+
+    if answer in ("y", "yes"):
+        _record_approval(event, command)
+        return True
+
+    return False
+
+
+def _record_approval(event: str, command: str) -> None:
+    entry = {
+        "event": event,
+        "command": command,
+        "approved_at": _utc_now_iso(),
+        "script_mtime_at_approval": script_mtime_iso(command),
+    }
+    with _locked_update_approvals() as data:
+        data["approvals"] = [
+            e for e in data.get("approvals", [])
+            if not (
+                isinstance(e, dict)
+                and e.get("event") == event
+                and e.get("command") == command
+            )
+        ] + [entry]
+
+
+def _utc_now_iso() -> str:
+    return datetime.now(tz=timezone.utc).isoformat().replace("+00:00", "Z")
+
+
+def revoke(command: str) -> int:
+    """Remove every allowlist entry matching ``command``.
+
+    Returns the number of entries removed.  Does not unregister any
+    callbacks that are already live on the plugin manager in the current
+    process — restart the CLI / gateway to drop them.
+    """
+    with _locked_update_approvals() as data:
+        before = len(data.get("approvals", []))
+        data["approvals"] = [
+            e for e in data.get("approvals", [])
+            if not (isinstance(e, dict) and e.get("command") == command)
+        ]
+        after = len(data["approvals"])
+    return before - after
+
+
+_SCRIPT_EXTENSIONS: Tuple[str, ...] = (
+    ".sh", ".bash", ".zsh", ".fish",
+    ".py", ".pyw",
+    ".rb", ".pl", ".lua",
+    ".js", ".mjs", ".cjs", ".ts",
+)
+
+
+def _command_script_path(command: str) -> str:
+    """Return the script path from ``command`` for doctor / drift checks.
+
+    Prefers a token ending in a known script extension, then a token
+    containing ``/`` or leading ``~``, then the first token.  Handles
+    ``python3 /path/hook.py``, ``/usr/bin/env bash hook.sh``, and the
+    common bare-path form.
+    """
+    try:
+        parts = shlex.split(command)
+    except ValueError:
+        return command
+    if not parts:
+        return command
+    for part in parts:
+        if part.lower().endswith(_SCRIPT_EXTENSIONS):
+            return part
+    for part in parts:
+        if "/" in part or part.startswith("~"):
+            return part
+    return parts[0]
+
+
+# ---------------------------------------------------------------------------
+# Helpers for accept-hooks resolution
+# ---------------------------------------------------------------------------
+
+def _resolve_effective_accept(
+    cfg: Dict[str, Any], accept_hooks_arg: bool,
+) -> bool:
+    """Combine all three opt-in channels into a single boolean.
+
+    Precedence (any truthy source flips us on):
+      1. ``--accept-hooks`` flag (CLI) / explicit argument
+      2. ``HERMES_ACCEPT_HOOKS`` env var
+      3. ``hooks_auto_accept: true`` in ``cli-config.yaml``
+    """
+    if accept_hooks_arg:
+        return True
+    env = os.environ.get("HERMES_ACCEPT_HOOKS", "").strip().lower()
+    if env in ("1", "true", "yes", "on"):
+        return True
+    cfg_val = cfg.get("hooks_auto_accept", False)
+    return bool(cfg_val)
+
+
+# ---------------------------------------------------------------------------
+# Introspection (used by `hermes hooks` CLI)
+# ---------------------------------------------------------------------------
+
+def allowlist_entry_for(event: str, command: str) -> Optional[Dict[str, Any]]:
+    """Return the allowlist record for this pair, if any."""
+    for e in load_allowlist().get("approvals", []):
+        if (
+            isinstance(e, dict)
+            and e.get("event") == event
+            and e.get("command") == command
+        ):
+            return e
+    return None
+
+
+def script_mtime_iso(command: str) -> Optional[str]:
+    """ISO-8601 mtime of the resolved script path, or ``None`` if the
+    script is missing."""
+    path = _command_script_path(command)
+    if not path:
+        return None
+    try:
+        expanded = os.path.expanduser(path)
+        return datetime.fromtimestamp(
+            os.path.getmtime(expanded), tz=timezone.utc,
+        ).isoformat().replace("+00:00", "Z")
+    except OSError:
+        return None
+
+
+def script_is_executable(command: str) -> bool:
+    """Return ``True`` iff ``command`` is runnable as configured.
+
+    For a bare invocation (``/path/hook.sh``) the script itself must be
+    executable.  For interpreter-prefixed commands (``python3
+    /path/hook.py``, ``/usr/bin/env bash hook.sh``) the script just has
+    to be readable — the interpreter doesn't care about the ``X_OK``
+    bit.  Mirrors what ``_spawn`` would actually do at runtime."""
+    path = _command_script_path(command)
+    if not path:
+        return False
+    expanded = os.path.expanduser(path)
+    if not os.path.isfile(expanded):
+        return False
+    try:
+        argv = shlex.split(command)
+    except ValueError:
+        return False
+    is_bare_invocation = bool(argv) and argv[0] == path
+    required = os.X_OK if is_bare_invocation else os.R_OK
+    return os.access(expanded, required)
+
+
+def run_once(
+    spec: ShellHookSpec, kwargs: Dict[str, Any],
+) -> Dict[str, Any]:
+    """Fire a single shell-hook invocation with a synthetic payload.
+    Used by ``hermes hooks test`` and ``hermes hooks doctor``.
+
+    ``kwargs`` is the same dict that :func:`hermes_cli.plugins.invoke_hook`
+    would pass at runtime.  It is routed through :func:`_serialize_payload`
+    so the synthetic stdin exactly matches what a real hook firing would
+    produce — otherwise scripts tested via ``hermes hooks test`` could
+    diverge silently from production behaviour.
+
+    Returns the :func:`_spawn` diagnostic dict plus a ``parsed`` field
+    holding the canonical Hermes-wire-shape response."""
+    stdin_json = _serialize_payload(spec.event, kwargs)
+    result = _spawn(spec, stdin_json)
+    result["parsed"] = _parse_response(spec.event, result["stdout"])
+    return result

agent/skill_commands.py

@@ -8,6 +8,7 @@ can invoke skills via /skill-name commands and prompt-only built-ins like
 import json
 import logging
 import re
+import subprocess
 from datetime import datetime
 from pathlib import Path
 from typing import Any, Dict, Optional
@@ -22,6 +23,110 @@ _PLAN_SLUG_RE = re.compile(r"[^a-z0-9]+")
 _SKILL_INVALID_CHARS = re.compile(r"[^a-z0-9-]")
 _SKILL_MULTI_HYPHEN = re.compile(r"-{2,}")
 
+# Matches ${HERMES_SKILL_DIR} / ${HERMES_SESSION_ID} tokens in SKILL.md.
+# Tokens that don't resolve (e.g. ${HERMES_SESSION_ID} with no session) are
+# left as-is so the user can debug them.
+_SKILL_TEMPLATE_RE = re.compile(r"\$\{(HERMES_SKILL_DIR|HERMES_SESSION_ID)\}")
+
+# Matches inline shell snippets like:  !`date +%Y-%m-%d`
+# Non-greedy, single-line only — no newlines inside the backticks.
+_INLINE_SHELL_RE = re.compile(r"!`([^`\n]+)`")
+
+# Cap inline-shell output so a runaway command can't blow out the context.
+_INLINE_SHELL_MAX_OUTPUT = 4000
+
+
+def _load_skills_config() -> dict:
+    """Load the ``skills`` section of config.yaml (best-effort)."""
+    try:
+        from hermes_cli.config import load_config
+
+        cfg = load_config() or {}
+        skills_cfg = cfg.get("skills")
+        if isinstance(skills_cfg, dict):
+            return skills_cfg
+    except Exception:
+        logger.debug("Could not read skills config", exc_info=True)
+    return {}
+
+
+def _substitute_template_vars(
+    content: str,
+    skill_dir: Path | None,
+    session_id: str | None,
+) -> str:
+    """Replace ${HERMES_SKILL_DIR} / ${HERMES_SESSION_ID} in skill content.
+
+    Only substitutes tokens for which a concrete value is available —
+    unresolved tokens are left in place so the author can spot them.
+    """
+    if not content:
+        return content
+
+    skill_dir_str = str(skill_dir) if skill_dir else None
+
+    def _replace(match: re.Match) -> str:
+        token = match.group(1)
+        if token == "HERMES_SKILL_DIR" and skill_dir_str:
+            return skill_dir_str
+        if token == "HERMES_SESSION_ID" and session_id:
+            return str(session_id)
+        return match.group(0)
+
+    return _SKILL_TEMPLATE_RE.sub(_replace, content)
+
+
+def _run_inline_shell(command: str, cwd: Path | None, timeout: int) -> str:
+    """Execute a single inline-shell snippet and return its stdout (trimmed).
+
+    Failures return a short ``[inline-shell error: ...]`` marker instead of
+    raising, so one bad snippet can't wreck the whole skill message.
+    """
+    try:
+        completed = subprocess.run(
+            ["bash", "-c", command],
+            cwd=str(cwd) if cwd else None,
+            capture_output=True,
+            text=True,
+            timeout=max(1, int(timeout)),
+            check=False,
+        )
+    except subprocess.TimeoutExpired:
+        return f"[inline-shell timeout after {timeout}s: {command}]"
+    except FileNotFoundError:
+        return f"[inline-shell error: bash not found]"
+    except Exception as exc:
+        return f"[inline-shell error: {exc}]"
+
+    output = (completed.stdout or "").rstrip("\n")
+    if not output and completed.stderr:
+        output = completed.stderr.rstrip("\n")
+    if len(output) > _INLINE_SHELL_MAX_OUTPUT:
+        output = output[:_INLINE_SHELL_MAX_OUTPUT] + "…[truncated]"
+    return output
+
+
+def _expand_inline_shell(
+    content: str,
+    skill_dir: Path | None,
+    timeout: int,
+) -> str:
+    """Replace every !`cmd` snippet in ``content`` with its stdout.
+
+    Runs each snippet with the skill directory as CWD so relative paths in
+    the snippet work the way the author expects.
+    """
+    if "!`" not in content:
+        return content
+
+    def _replace(match: re.Match) -> str:
+        cmd = match.group(1).strip()
+        if not cmd:
+            return ""
+        return _run_inline_shell(cmd, skill_dir, timeout)
+
+    return _INLINE_SHELL_RE.sub(_replace, content)
+
 
 def build_plan_path(
     user_instruction: str = "",
@@ -133,14 +238,36 @@ def _build_skill_message(
     activation_note: str,
     user_instruction: str = "",
     runtime_note: str = "",
+    session_id: str | None = None,
 ) -> str:
     """Format a loaded skill into a user/system message payload."""
     from tools.skills_tool import SKILLS_DIR
 
     content = str(loaded_skill.get("content") or "")
 
+    # ── Template substitution and inline-shell expansion ──
+    # Done before anything else so downstream blocks (setup notes,
+    # supporting-file hints) see the expanded content.
+    skills_cfg = _load_skills_config()
+    if skills_cfg.get("template_vars", True):
+        content = _substitute_template_vars(content, skill_dir, session_id)
+    if skills_cfg.get("inline_shell", False):
+        timeout = int(skills_cfg.get("inline_shell_timeout", 10) or 10)
+        content = _expand_inline_shell(content, skill_dir, timeout)
+
     parts = [activation_note, "", content.strip()]
 
+    # ── Inject the absolute skill directory so the agent can reference
+    #    bundled scripts without an extra skill_view() round-trip. ──
+    if skill_dir:
+        parts.append("")
+        parts.append(f"[Skill directory: {skill_dir}]")
+        parts.append(
+            "Resolve any relative paths in this skill (e.g. `scripts/foo.js`, "
+            "`templates/config.yaml`) against that directory, then run them "
+            "with the terminal tool using the absolute path."
+        )
+
     # ── Inject resolved skill config values ──
     _inject_skill_config(loaded_skill, parts)
 
@@ -188,11 +315,13 @@ def _build_skill_message(
             # Skill is from an external dir — use the skill name instead
             skill_view_target = skill_dir.name
         parts.append("")
-        parts.append("[This skill has supporting files you can load with the skill_view tool:]")
+        parts.append("[This skill has supporting files:]")
         for sf in supporting:
-            parts.append(f"- {sf}")
+            parts.append(f"- {sf}  ->  {skill_dir / sf}")
         parts.append(
-            f'\nTo view any of these, use: skill_view(name="{skill_view_target}", file_path="<path>")'
+            f'\nLoad any of these with skill_view(name="{skill_view_target}", '
+            f'file_path="<path>"), or run scripts directly by absolute path '
+            f"(e.g. `node {skill_dir}/scripts/foo.js`)."
         )
 
     if user_instruction:
@@ -332,6 +461,7 @@ def build_skill_invocation_message(
         activation_note,
         user_instruction=user_instruction,
         runtime_note=runtime_note,
+        session_id=task_id,
     )
 
 
@@ -370,6 +500,7 @@ def build_preloaded_skills_prompt(
                 loaded_skill,
                 skill_dir,
                 activation_note,
+                session_id=task_id,
             )
         )
         loaded_names.append(skill_name)

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/transports/__init__.py

@@ -0,0 +1,39 @@
+"""Transport layer types and registry for provider response normalization.
+
+Usage:
+    from agent.transports import get_transport
+    transport = get_transport("anthropic_messages")
+    result = transport.normalize_response(raw_response)
+"""
+
+from agent.transports.types import NormalizedResponse, ToolCall, Usage, build_tool_call, map_finish_reason  # noqa: F401
+
+_REGISTRY: dict = {}
+
+
+def register_transport(api_mode: str, transport_cls: type) -> None:
+    """Register a transport class for an api_mode string."""
+    _REGISTRY[api_mode] = transport_cls
+
+
+def get_transport(api_mode: str):
+    """Get a transport instance for the given api_mode.
+
+    Returns None if no transport is registered for this api_mode.
+    This allows gradual migration — call sites can check for None
+    and fall back to the legacy code path.
+    """
+    if not _REGISTRY:
+        _discover_transports()
+    cls = _REGISTRY.get(api_mode)
+    if cls is None:
+        return None
+    return cls()
+
+
+def _discover_transports() -> None:
+    """Import all transport modules to trigger auto-registration."""
+    try:
+        import agent.transports.anthropic  # noqa: F401
+    except ImportError:
+        pass

agent/transports/anthropic.py

@@ -0,0 +1,129 @@
+"""Anthropic Messages API transport.
+
+Delegates to the existing adapter functions in agent/anthropic_adapter.py.
+This transport owns format conversion and normalization — NOT client lifecycle.
+"""
+
+from typing import Any, Dict, List, Optional
+
+from agent.transports.base import ProviderTransport
+from agent.transports.types import NormalizedResponse
+
+
+class AnthropicTransport(ProviderTransport):
+    """Transport for api_mode='anthropic_messages'.
+
+    Wraps the existing functions in anthropic_adapter.py behind the
+    ProviderTransport ABC.  Each method delegates — no logic is duplicated.
+    """
+
+    @property
+    def api_mode(self) -> str:
+        return "anthropic_messages"
+
+    def convert_messages(self, messages: List[Dict[str, Any]], **kwargs) -> Any:
+        """Convert OpenAI messages to Anthropic (system, messages) tuple.
+
+        kwargs:
+            base_url: Optional[str] — affects thinking signature handling.
+        """
+        from agent.anthropic_adapter import convert_messages_to_anthropic
+
+        base_url = kwargs.get("base_url")
+        return convert_messages_to_anthropic(messages, base_url=base_url)
+
+    def convert_tools(self, tools: List[Dict[str, Any]]) -> Any:
+        """Convert OpenAI tool schemas to Anthropic input_schema format."""
+        from agent.anthropic_adapter import convert_tools_to_anthropic
+
+        return convert_tools_to_anthropic(tools)
+
+    def build_kwargs(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        tools: Optional[List[Dict[str, Any]]] = None,
+        **params,
+    ) -> Dict[str, Any]:
+        """Build Anthropic messages.create() kwargs.
+
+        Calls convert_messages and convert_tools internally.
+
+        params (all optional):
+            max_tokens: int
+            reasoning_config: dict | None
+            tool_choice: str | None
+            is_oauth: bool
+            preserve_dots: bool
+            context_length: int | None
+            base_url: str | None
+            fast_mode: bool
+        """
+        from agent.anthropic_adapter import build_anthropic_kwargs
+
+        return build_anthropic_kwargs(
+            model=model,
+            messages=messages,
+            tools=tools,
+            max_tokens=params.get("max_tokens", 16384),
+            reasoning_config=params.get("reasoning_config"),
+            tool_choice=params.get("tool_choice"),
+            is_oauth=params.get("is_oauth", False),
+            preserve_dots=params.get("preserve_dots", False),
+            context_length=params.get("context_length"),
+            base_url=params.get("base_url"),
+            fast_mode=params.get("fast_mode", False),
+        )
+
+    def normalize_response(self, response: Any, **kwargs) -> NormalizedResponse:
+        """Normalize Anthropic response to NormalizedResponse.
+
+        kwargs:
+            strip_tool_prefix: bool — strip 'mcp_mcp_' prefixes from tool names.
+        """
+        from agent.anthropic_adapter import normalize_anthropic_response_v2
+
+        strip_tool_prefix = kwargs.get("strip_tool_prefix", False)
+        return normalize_anthropic_response_v2(response, strip_tool_prefix=strip_tool_prefix)
+
+    def validate_response(self, response: Any) -> bool:
+        """Check Anthropic response structure is valid."""
+        if response is None:
+            return False
+        content_blocks = getattr(response, "content", None)
+        if not isinstance(content_blocks, list):
+            return False
+        if not content_blocks:
+            return False
+        return True
+
+    def extract_cache_stats(self, response: Any) -> Optional[Dict[str, int]]:
+        """Extract Anthropic cache_read and cache_creation token counts."""
+        usage = getattr(response, "usage", None)
+        if usage is None:
+            return None
+        cached = getattr(usage, "cache_read_input_tokens", 0) or 0
+        written = getattr(usage, "cache_creation_input_tokens", 0) or 0
+        if cached or written:
+            return {"cached_tokens": cached, "creation_tokens": written}
+        return None
+
+    # Promote the adapter's canonical mapping to module level so it's shared
+    _STOP_REASON_MAP = {
+        "end_turn": "stop",
+        "tool_use": "tool_calls",
+        "max_tokens": "length",
+        "stop_sequence": "stop",
+        "refusal": "content_filter",
+        "model_context_window_exceeded": "length",
+    }
+
+    def map_finish_reason(self, raw_reason: str) -> str:
+        """Map Anthropic stop_reason to OpenAI finish_reason."""
+        return self._STOP_REASON_MAP.get(raw_reason, "stop")
+
+
+# Auto-register on import
+from agent.transports import register_transport  # noqa: E402
+
+register_transport("anthropic_messages", AnthropicTransport)

agent/transports/base.py

@@ -0,0 +1,89 @@
+"""Abstract base for provider transports.
+
+A transport owns the data path for one api_mode:
+  convert_messages → convert_tools → build_kwargs → normalize_response
+
+It does NOT own: client construction, streaming, credential refresh,
+prompt caching, interrupt handling, or retry logic.  Those stay on AIAgent.
+"""
+
+from abc import ABC, abstractmethod
+from typing import Any, Dict, List, Optional
+
+from agent.transports.types import NormalizedResponse
+
+
+class ProviderTransport(ABC):
+    """Base class for provider-specific format conversion and normalization."""
+
+    @property
+    @abstractmethod
+    def api_mode(self) -> str:
+        """The api_mode string this transport handles (e.g. 'anthropic_messages')."""
+        ...
+
+    @abstractmethod
+    def convert_messages(self, messages: List[Dict[str, Any]], **kwargs) -> Any:
+        """Convert OpenAI-format messages to provider-native format.
+
+        Returns provider-specific structure (e.g. (system, messages) for Anthropic,
+        or the messages list unchanged for chat_completions).
+        """
+        ...
+
+    @abstractmethod
+    def convert_tools(self, tools: List[Dict[str, Any]]) -> Any:
+        """Convert OpenAI-format tool definitions to provider-native format.
+
+        Returns provider-specific tool list (e.g. Anthropic input_schema format).
+        """
+        ...
+
+    @abstractmethod
+    def build_kwargs(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        tools: Optional[List[Dict[str, Any]]] = None,
+        **params,
+    ) -> Dict[str, Any]:
+        """Build the complete API call kwargs dict.
+
+        This is the primary entry point — it typically calls convert_messages()
+        and convert_tools() internally, then adds model-specific config.
+
+        Returns a dict ready to be passed to the provider's SDK client.
+        """
+        ...
+
+    @abstractmethod
+    def normalize_response(self, response: Any, **kwargs) -> NormalizedResponse:
+        """Normalize a raw provider response to the shared NormalizedResponse type.
+
+        This is the only method that returns a transport-layer type.
+        """
+        ...
+
+    def validate_response(self, response: Any) -> bool:
+        """Optional: check if the raw response is structurally valid.
+
+        Returns True if valid, False if the response should be treated as invalid.
+        Default implementation always returns True.
+        """
+        return True
+
+    def extract_cache_stats(self, response: Any) -> Optional[Dict[str, int]]:
+        """Optional: extract provider-specific cache hit/creation stats.
+
+        Returns dict with 'cached_tokens' and 'creation_tokens', or None.
+        Default returns None.
+        """
+        return None
+
+    def map_finish_reason(self, raw_reason: str) -> str:
+        """Optional: map provider-specific stop reason to OpenAI equivalent.
+
+        Default returns the raw reason unchanged.  Override for providers
+        with different stop reason vocabularies.
+        """
+        return raw_reason

agent/transports/types.py

@@ -0,0 +1,100 @@
+"""Shared types for normalized provider responses.
+
+These dataclasses define the canonical shape that all provider adapters
+normalize responses to.  The shared surface is intentionally minimal —
+only fields that every downstream consumer reads are top-level.
+Protocol-specific state goes in ``provider_data`` dicts (response-level
+and per-tool-call) so that protocol-aware code paths can access it
+without polluting the shared type.
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional
+
+
+@dataclass
+class ToolCall:
+    """A normalized tool call from any provider.
+
+    ``id`` is the protocol's canonical identifier — what gets used in
+    ``tool_call_id`` / ``tool_use_id`` when constructing tool result
+    messages.  May be ``None`` when the provider omits it; the agent
+    fills it via ``_deterministic_call_id()`` before storing in history.
+
+    ``provider_data`` carries per-tool-call protocol metadata that only
+    protocol-aware code reads:
+
+    * Codex: ``{"call_id": "call_XXX", "response_item_id": "fc_XXX"}``
+    * Gemini: ``{"extra_content": {"google": {"thought_signature": "..."}}}``
+    * Others: ``None``
+    """
+
+    id: Optional[str]
+    name: str
+    arguments: str  # JSON string
+    provider_data: Optional[Dict[str, Any]] = field(default=None, repr=False)
+
+
+@dataclass
+class Usage:
+    """Token usage from an API response."""
+
+    prompt_tokens: int = 0
+    completion_tokens: int = 0
+    total_tokens: int = 0
+    cached_tokens: int = 0
+
+
+@dataclass
+class NormalizedResponse:
+    """Normalized API response from any provider.
+
+    Shared fields are truly cross-provider — every caller can rely on
+    them without branching on api_mode.  Protocol-specific state goes in
+    ``provider_data`` so that only protocol-aware code paths read it.
+
+    Response-level ``provider_data`` examples:
+
+    * Anthropic: ``{"reasoning_details": [...]}``
+    * Codex: ``{"codex_reasoning_items": [...]}``
+    * Others: ``None``
+    """
+
+    content: Optional[str]
+    tool_calls: Optional[List[ToolCall]]
+    finish_reason: str  # "stop", "tool_calls", "length", "content_filter"
+    reasoning: Optional[str] = None
+    usage: Optional[Usage] = None
+    provider_data: Optional[Dict[str, Any]] = field(default=None, repr=False)
+
+
+# ---------------------------------------------------------------------------
+# Factory helpers
+# ---------------------------------------------------------------------------
+
+def build_tool_call(
+    id: Optional[str],
+    name: str,
+    arguments: Any,
+    **provider_fields: Any,
+) -> ToolCall:
+    """Build a ``ToolCall``, auto-serialising *arguments* if it's a dict.
+
+    Any extra keyword arguments are collected into ``provider_data``.
+    """
+    args_str = json.dumps(arguments) if isinstance(arguments, dict) else str(arguments)
+    pd = dict(provider_fields) if provider_fields else None
+    return ToolCall(id=id, name=name, arguments=args_str, provider_data=pd)
+
+
+def map_finish_reason(reason: Optional[str], mapping: Dict[str, str]) -> str:
+    """Translate a provider-specific stop reason to the normalised set.
+
+    Falls back to ``"stop"`` for unknown or ``None`` reasons.
+    """
+    if reason is None:
+        return "stop"
+    return mapping.get(reason, "stop")

agent/usage_pricing.py

@@ -6,6 +6,7 @@ from decimal import Decimal
 from typing import Any, Dict, Literal, Optional
 
 from agent.model_metadata import fetch_endpoint_model_metadata, fetch_model_metadata
+from utils import base_url_host_matches
 
 DEFAULT_PRICING = {"input": 0.0, "output": 0.0}
 
@@ -393,7 +394,7 @@ def resolve_billing_route(
 
     if provider_name == "openai-codex":
         return BillingRoute(provider="openai-codex", model=model, base_url=base_url or "", billing_mode="subscription_included")
-    if provider_name == "openrouter" or "openrouter.ai" in base:
+    if provider_name == "openrouter" or base_url_host_matches(base_url or "", "openrouter.ai"):
         return BillingRoute(provider="openrouter", model=model, base_url=base_url or "", billing_mode="official_models_api")
     if provider_name == "anthropic":
         return BillingRoute(provider="anthropic", model=model.split("/")[-1], base_url=base_url or "", billing_mode="official_docs_snapshot")

batch_runner.py

@@ -444,6 +444,7 @@ def _process_batch_worker(args: Tuple) -> Dict[str, Any]:
             if not reasoning.get("has_any_reasoning", True):
                 print(f"   🚫 Prompt {prompt_index} discarded (no reasoning in any turn)")
                 discarded_no_reasoning += 1
+                completed_in_batch.append(prompt_index)
                 continue
             
             # Get and normalize tool stats for consistent schema across all entries
@@ -1189,12 +1190,12 @@ def main(
     """
     # Handle list distributions
     if list_distributions:
-        from toolset_distributions import list_distributions as get_all_dists, print_distribution_info
-        
+        from toolset_distributions import print_distribution_info
+
         print("📊 Available Toolset Distributions")
         print("=" * 70)
-        
-        all_dists = get_all_dists()
+
+        all_dists = list_distributions()
         for dist_name in sorted(all_dists.keys()):
             print_distribution_info(dist_name)
         

cli-config.yaml.example

@@ -63,7 +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)
@@ -91,20 +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
 # =============================================================================
@@ -357,6 +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
@@ -741,10 +770,12 @@ code_execution:
 # Subagent Delegation
 # =============================================================================
 # The delegate_task tool spawns child agents with isolated context.
-# Supports single tasks and batch mode (up to 3 parallel).
+# Supports single tasks and batch mode (default 3 parallel, configurable).
 delegation:
   max_iterations: 50                          # Max tool-calling turns per child (default: 50)
-  default_toolsets: ["terminal", "file", "web"]  # Default toolsets for subagents
+  # max_concurrent_children: 3                # Max parallel child agents (default: 3)
+  # max_spawn_depth: 1                        # Tree depth cap (1-3, default: 1 = flat). Raise to 2 or 3 to allow orchestrator children to spawn their own workers.
+  # orchestrator_enabled: true                # Kill switch for role="orchestrator" children (default: true).
   # model: "google/gemini-3-flash-preview"    # Override model for subagents (empty = inherit parent)
   # provider: "openrouter"                    # Override provider for subagents (empty = inherit parent)
   #                                           # Resolves full credentials (base_url, api_key) automatically.
@@ -888,3 +919,39 @@ display:
 #   # Names and usernames are NOT affected (user-chosen, publicly visible).
 #   # Routing/delivery still uses the original values internally.
 #   redact_pii: false
+
+# =============================================================================
+# Shell-script hooks
+# =============================================================================
+# Register shell scripts as plugin-hook callbacks.  Each entry is executed as
+# a subprocess (shell=False, shlex.split) with a JSON payload on stdin.  On
+# stdout the script may return JSON that either blocks the tool call or
+# injects context into the next LLM call.
+#
+# Valid events (mirror hermes_cli.plugins.VALID_HOOKS):
+#   pre_tool_call, post_tool_call, pre_llm_call, post_llm_call,
+#   pre_api_request, post_api_request, on_session_start, on_session_end,
+#   on_session_finalize, on_session_reset, subagent_stop
+#
+# First-use consent: each (event, command) pair prompts once on a TTY, then
+# is persisted to ~/.hermes/shell-hooks-allowlist.json.  Non-interactive
+# runs (gateway, cron) need --accept-hooks, HERMES_ACCEPT_HOOKS=1, or the
+# hooks_auto_accept key below.
+#
+# See website/docs/user-guide/features/hooks.md for the full JSON wire
+# protocol and worked examples.
+#
+# hooks:
+#   pre_tool_call:
+#     - matcher: "terminal"
+#       command: "~/.hermes/agent-hooks/block-rm-rf.sh"
+#       timeout: 10
+#   post_tool_call:
+#     - matcher: "write_file|patch"
+#       command: "~/.hermes/agent-hooks/auto-format.sh"
+#   pre_llm_call:
+#     - command: "~/.hermes/agent-hooks/inject-cwd-context.sh"
+#   subagent_stop:
+#     - command: "~/.hermes/agent-hooks/log-orchestration.sh"
+#
+# hooks_auto_accept: false

cron/jobs.py

@@ -9,6 +9,7 @@ import copy
 import json
 import logging
 import tempfile
+import threading
 import os
 import re
 import uuid
@@ -34,6 +35,11 @@ except ImportError:
 HERMES_DIR = get_hermes_home().resolve()
 CRON_DIR = HERMES_DIR / "cron"
 JOBS_FILE = CRON_DIR / "jobs.json"
+
+# In-process lock protecting load_jobs→modify→save_jobs cycles.
+# Required when tick() runs jobs in parallel threads — without this,
+# concurrent mark_job_run / advance_next_run calls can clobber each other.
+_jobs_file_lock = threading.Lock()
 OUTPUT_DIR = CRON_DIR / "output"
 ONESHOT_GRACE_SECONDS = 120
 
@@ -594,43 +600,44 @@ def mark_job_run(job_id: str, success: bool, error: Optional[str] = None,
     ``delivery_error`` is tracked separately from the agent error — a job
     can succeed (agent produced output) but fail delivery (platform down).
     """
-    jobs = load_jobs()
-    for i, job in enumerate(jobs):
-        if job["id"] == job_id:
-            now = _hermes_now().isoformat()
-            job["last_run_at"] = now
-            job["last_status"] = "ok" if success else "error"
-            job["last_error"] = error if not success else None
-            # Track delivery failures separately — cleared on successful delivery
-            job["last_delivery_error"] = delivery_error
-            
-            # Increment completed count
-            if job.get("repeat"):
-                job["repeat"]["completed"] = job["repeat"].get("completed", 0) + 1
+    with _jobs_file_lock:
+        jobs = load_jobs()
+        for i, job in enumerate(jobs):
+            if job["id"] == job_id:
+                now = _hermes_now().isoformat()
+                job["last_run_at"] = now
+                job["last_status"] = "ok" if success else "error"
+                job["last_error"] = error if not success else None
+                # Track delivery failures separately — cleared on successful delivery
+                job["last_delivery_error"] = delivery_error
                 
-                # Check if we've hit the repeat limit
-                times = job["repeat"].get("times")
-                completed = job["repeat"]["completed"]
-                if times is not None and times > 0 and completed >= times:
-                    # Remove the job (limit reached)
-                    jobs.pop(i)
-                    save_jobs(jobs)
-                    return
-            
-            # Compute next run
-            job["next_run_at"] = compute_next_run(job["schedule"], now)
+                # Increment completed count
+                if job.get("repeat"):
+                    job["repeat"]["completed"] = job["repeat"].get("completed", 0) + 1
+                    
+                    # Check if we've hit the repeat limit
+                    times = job["repeat"].get("times")
+                    completed = job["repeat"]["completed"]
+                    if times is not None and times > 0 and completed >= times:
+                        # Remove the job (limit reached)
+                        jobs.pop(i)
+                        save_jobs(jobs)
+                        return
+                
+                # Compute next run
+                job["next_run_at"] = compute_next_run(job["schedule"], now)
 
-            # If no next run (one-shot completed), disable
-            if job["next_run_at"] is None:
-                job["enabled"] = False
-                job["state"] = "completed"
-            elif job.get("state") != "paused":
-                job["state"] = "scheduled"
+                # If no next run (one-shot completed), disable
+                if job["next_run_at"] is None:
+                    job["enabled"] = False
+                    job["state"] = "completed"
+                elif job.get("state") != "paused":
+                    job["state"] = "scheduled"
 
-            save_jobs(jobs)
-            return
+                save_jobs(jobs)
+                return
 
-    logger.warning("mark_job_run: job_id %s not found, skipping save", job_id)
+        logger.warning("mark_job_run: job_id %s not found, skipping save", job_id)
 
 
 def advance_next_run(job_id: str) -> bool:
@@ -645,20 +652,21 @@ def advance_next_run(job_id: str) -> bool:
 
     Returns True if next_run_at was advanced, False otherwise.
     """
-    jobs = load_jobs()
-    for job in jobs:
-        if job["id"] == job_id:
-            kind = job.get("schedule", {}).get("kind")
-            if kind not in ("cron", "interval"):
+    with _jobs_file_lock:
+        jobs = load_jobs()
+        for job in jobs:
+            if job["id"] == job_id:
+                kind = job.get("schedule", {}).get("kind")
+                if kind not in ("cron", "interval"):
+                    return False
+                now = _hermes_now().isoformat()
+                new_next = compute_next_run(job["schedule"], now)
+                if new_next and new_next != job.get("next_run_at"):
+                    job["next_run_at"] = new_next
+                    save_jobs(jobs)
+                    return True
                 return False
-            now = _hermes_now().isoformat()
-            new_next = compute_next_run(job["schedule"], now)
-            if new_next and new_next != job.get("next_run_at"):
-                job["next_run_at"] = new_next
-                save_jobs(jobs)
-                return True
-            return False
-    return False
+        return False
 
 
 def get_due_jobs() -> List[Dict[str, Any]]:

cron/scheduler.py

@@ -252,7 +252,11 @@ def _send_media_via_adapter(adapter, chat_id: str, media_files: list, metadata:
                 coro = adapter.send_document(chat_id=chat_id, file_path=media_path, metadata=metadata)
 
             future = asyncio.run_coroutine_threadsafe(coro, loop)
-            result = future.result(timeout=30)
+            try:
+                result = future.result(timeout=30)
+            except TimeoutError:
+                future.cancel()
+                raise
             if result and not getattr(result, "success", True):
                 logger.warning(
                     "Job '%s': media send failed for %s: %s",
@@ -382,7 +386,11 @@ def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> Option
                         runtime_adapter.send(chat_id, text_to_send, metadata=send_metadata),
                         loop,
                     )
-                    send_result = future.result(timeout=60)
+                    try:
+                        send_result = future.result(timeout=60)
+                    except TimeoutError:
+                        future.cancel()
+                        raise
                     if send_result and not getattr(send_result, "success", True):
                         err = getattr(send_result, "error", "unknown")
                         logger.warning(
@@ -422,7 +430,6 @@ def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> Option
                 # 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)
@@ -747,14 +754,17 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
     # scheduler process — every job this process runs is a cron job.
     os.environ["HERMES_CRON_SESSION"] = "1"
 
+    # Use ContextVars for per-job session/delivery state so parallel jobs
+    # don't clobber each other's targets (os.environ is process-global).
+    from gateway.session_context import set_session_vars, clear_session_vars, _VAR_MAP
+
+    _ctx_tokens = set_session_vars(
+        platform=origin["platform"] if origin else "",
+        chat_id=str(origin["chat_id"]) if origin else "",
+        chat_name=origin.get("chat_name", "") if origin else "",
+    )
+
     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.
-        if origin:
-            os.environ["HERMES_SESSION_PLATFORM"] = origin["platform"]
-            os.environ["HERMES_SESSION_CHAT_ID"] = str(origin["chat_id"])
-            if origin.get("chat_name"):
-                os.environ["HERMES_SESSION_CHAT_NAME"] = origin["chat_name"]
         # Re-read .env and config.yaml fresh every run so provider/key
         # changes take effect without a gateway restart.
         from dotenv import load_dotenv
@@ -765,10 +775,10 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
 
         delivery_target = _resolve_delivery_target(job)
         if delivery_target:
-            os.environ["HERMES_CRON_AUTO_DELIVER_PLATFORM"] = delivery_target["platform"]
-            os.environ["HERMES_CRON_AUTO_DELIVER_CHAT_ID"] = str(delivery_target["chat_id"])
+            _VAR_MAP["HERMES_CRON_AUTO_DELIVER_PLATFORM"].set(delivery_target["platform"])
+            _VAR_MAP["HERMES_CRON_AUTO_DELIVER_CHAT_ID"].set(str(delivery_target["chat_id"]))
             if delivery_target.get("thread_id") is not None:
-                os.environ["HERMES_CRON_AUTO_DELIVER_THREAD_ID"] = str(delivery_target["thread_id"])
+                _VAR_MAP["HERMES_CRON_AUTO_DELIVER_THREAD_ID"].set(str(delivery_target["thread_id"]))
 
         model = job.get("model") or os.getenv("HERMES_MODEL") or ""
 
@@ -807,14 +817,13 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
         prefill_messages = None
         prefill_file = os.getenv("HERMES_PREFILL_MESSAGES_FILE", "") or _cfg.get("prefill_messages_file", "")
         if prefill_file:
-            import json as _json
             pfpath = Path(prefill_file).expanduser()
             if not pfpath.is_absolute():
                 pfpath = _hermes_home / pfpath
             if pfpath.exists():
                 try:
                     with open(pfpath, "r", encoding="utf-8") as _pf:
-                        prefill_messages = _json.load(_pf)
+                        prefill_messages = json.load(_pf)
                     if not isinstance(prefill_messages, list):
                         prefill_messages = None
                 except Exception as e:
@@ -826,7 +835,6 @@ 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,
@@ -843,24 +851,9 @@ 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
@@ -877,13 +870,13 @@ 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,
@@ -1028,16 +1021,8 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
         return False, output, "", error_msg
 
     finally:
-        # Clean up injected env vars so they don't leak to other jobs
-        for key in (
-            "HERMES_SESSION_PLATFORM",
-            "HERMES_SESSION_CHAT_ID",
-            "HERMES_SESSION_CHAT_NAME",
-            "HERMES_CRON_AUTO_DELIVER_PLATFORM",
-            "HERMES_CRON_AUTO_DELIVER_CHAT_ID",
-            "HERMES_CRON_AUTO_DELIVER_THREAD_ID",
-        ):
-            os.environ.pop(key, None)
+        # Clean up ContextVar session/delivery state for this job.
+        clear_session_vars(_ctx_tokens)
         if _session_db:
             try:
                 _session_db.end_session(_cron_session_id, "cron_complete")
@@ -1090,15 +1075,41 @@ def tick(verbose: bool = True, adapters=None, loop=None) -> int:
         if verbose:
             logger.info("%s - %s job(s) due", _hermes_now().strftime('%H:%M:%S'), len(due_jobs))
 
-        executed = 0
+        # Advance next_run_at for all recurring jobs FIRST, under the file lock,
+        # before any execution begins.  This preserves at-most-once semantics.
         for job in due_jobs:
-            try:
-                # For recurring jobs (cron/interval), advance next_run_at to the
-                # next future occurrence BEFORE execution.  This way, if the
-                # process crashes mid-run, the job won't re-fire on restart.
-                # One-shot jobs are left alone so they can retry on restart.
-                advance_next_run(job["id"])
+            advance_next_run(job["id"])
 
+        # Resolve max parallel workers: env var > config.yaml > unbounded.
+        # Set HERMES_CRON_MAX_PARALLEL=1 to restore old serial behaviour.
+        _max_workers: Optional[int] = None
+        try:
+            _env_par = os.getenv("HERMES_CRON_MAX_PARALLEL", "").strip()
+            if _env_par:
+                _max_workers = int(_env_par) or None
+        except (ValueError, TypeError):
+            logger.warning("Invalid HERMES_CRON_MAX_PARALLEL value; defaulting to unbounded")
+        if _max_workers is None:
+            try:
+                _ucfg = load_config() or {}
+                _cfg_par = (
+                    _ucfg.get("cron", {}) if isinstance(_ucfg, dict) else {}
+                ).get("max_parallel_jobs")
+                if _cfg_par is not None:
+                    _max_workers = int(_cfg_par) or None
+            except Exception:
+                pass
+
+        if verbose:
+            logger.info(
+                "Running %d job(s) in parallel (max_workers=%s)",
+                len(due_jobs),
+                _max_workers if _max_workers else "unbounded",
+            )
+
+        def _process_job(job: dict) -> bool:
+            """Run one due job end-to-end: execute, save, deliver, mark."""
+            try:
                 success, output, final_response, error = run_job(job)
 
                 output_file = save_job_output(job["id"], output)
@@ -1130,13 +1141,23 @@ def tick(verbose: bool = True, adapters=None, loop=None) -> int:
                     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
+                return True
 
             except Exception as e:
                 logger.error("Error processing job %s: %s", job['id'], e)
                 mark_job_run(job["id"], False, str(e))
+                return False
 
-        return executed
+        # Run all due jobs concurrently, each in its own ContextVar copy
+        # so session/delivery state stays isolated per-thread.
+        with concurrent.futures.ThreadPoolExecutor(max_workers=_max_workers) as _tick_pool:
+            _futures = []
+            for job in due_jobs:
+                _ctx = contextvars.copy_context()
+                _futures.append(_tick_pool.submit(_ctx.run, _process_job, job))
+            _results = [f.result() for f in _futures]
+
+        return sum(_results)
     finally:
         if fcntl:
             fcntl.flock(lock_fd, fcntl.LOCK_UN)

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/plans/2026-04-01-ink-gateway-tui-migration-plan.md

@@ -1,108 +0,0 @@
-# Ink Gateway TUI Migration — Post-mortem
-
-Planned: 2026-04-01 · Delivered: 2026-04 · Status: shipped, classic (prompt_toolkit) CLI still present
-
-## What Shipped
-
-Three layers, same repo, Python runtime unchanged.
-
-```
-ui-tui (Node/TS)  ──stdio JSON-RPC──▶  tui_gateway (Py)  ──▶  AIAgent (run_agent.py)
-```
-
-### Backend — `tui_gateway/`
-
-```
-tui_gateway/
-├── entry.py          # subprocess entrypoint, stdio read/write loop
-├── server.py         # everything: sessions dict, @method handlers, _emit
-├── render.py         # stream renderer, diff rendering, message rendering
-├── slash_worker.py   # subprocess that runs hermes_cli slash commands
-└── __init__.py
-```
-
-`server.py` owns the full runtime-control surface: session store (`_sessions: dict[str, dict]`), method registry (`@method("…")` decorator), event emitter (`_emit`), agent lifecycle (`_make_agent`, `_init_session`, `_wire_callbacks`), approval/sudo/clarify round-trips, and JSON-RPC dispatch.
-
-Protocol methods (`@method(...)` in `server.py`):
-
-- session: `session.{create, resume, list, close, interrupt, usage, history, compress, branch, title, save, undo}`
-- prompt: `prompt.{submit, background, btw}`
-- tools: `tools.{list, show, configure}`
-- slash: `slash.exec`, `command.{dispatch, resolve}`, `commands.catalog`, `complete.{path, slash}`
-- approvals: `approval.respond`, `sudo.respond`, `clarify.respond`, `secret.respond`
-- config/state: `config.{get, set, show}`, `model.options`, `reload.mcp`
-- ops: `shell.exec`, `cli.exec`, `terminal.resize`, `input.detect_drop`, `clipboard.paste`, `paste.collapse`, `image.attach`, `process.stop`
-- misc: `agents.list`, `skills.manage`, `plugins.list`, `cron.manage`, `insights.get`, `rollback.{list, diff, restore}`, `browser.manage`
-
-Protocol events (`_emit(…)` → handled in `ui-tui/src/app/createGatewayEventHandler.ts`):
-
-- lifecycle: `gateway.{ready, stderr}`, `session.info`, `skin.changed`
-- stream: `message.{start, delta, complete}`, `thinking.delta`, `reasoning.{delta, available}`, `status.update`
-- tools: `tool.{start, progress, complete, generating}`, `subagent.{start, thinking, tool, progress, complete}`
-- interactive: `approval.request`, `sudo.request`, `clarify.request`, `secret.request`
-- async: `background.complete`, `btw.complete`, `error`
-
-### Frontend — `ui-tui/src/`
-
-```
-src/
-├── entry.tsx            # node bootstrap: bootBanner → spawn python → dynamic-import Ink → render(<App/>)
-├── app.tsx              # <GatewayProvider> wraps <AppLayout>
-├── bootBanner.ts        # raw-ANSI banner to stdout in ~2ms, pre-React
-├── gatewayClient.ts     # JSON-RPC client over child_process stdio
-├── gatewayTypes.ts      # typed RPC responses + GatewayEvent union
-├── theme.ts             # DEFAULT_THEME + fromSkin
-│
-├── app/                 # hooks + stores — the orchestration layer
-│   ├── uiStore.ts             # nanostore: sid, info, busy, usage, theme, status…
-│   ├── turnStore.ts           # nanostore: per-turn activity / reasoning / tools
-│   ├── turnController.ts      # imperative singleton for stream-time operations
-│   ├── overlayStore.ts        # nanostore: modal/overlay state
-│   ├── useMainApp.ts          # top-level composition hook
-│   ├── useSessionLifecycle.ts # session.create/resume/close/reset
-│   ├── useSubmission.ts       # shell/slash/prompt dispatch + interpolation
-│   ├── useConfigSync.ts       # config.get + mtime poll
-│   ├── useComposerState.ts    # input buffer, paste snippets, editor mode
-│   ├── useInputHandlers.ts    # key bindings
-│   ├── createGatewayEventHandler.ts  # event-stream dispatcher
-│   ├── createSlashHandler.ts         # slash command router (registry + python fallback)
-│   └── slash/commands/        # core.ts, ops.ts, session.ts — TS-owned slash commands
-│
-├── components/          # AppLayout, AppChrome, AppOverlays, MessageLine, Thinking, Markdown, pickers, prompts, Banner, SessionPanel
-├── config/              # env, limits, timing constants
-├── content/             # charms, faces, fortunes, hotkeys, placeholders, verbs
-├── domain/              # details, messages, paths, roles, slash, usage, viewport
-├── protocol/            # interpolation, paste regex
-├── hooks/               # useCompletion, useInputHistory, useQueue, useVirtualHistory
-└── lib/                 # history, messages, osc52, rpc, text
-```
-
-### CLI entry points — `hermes_cli/main.py`
-
-- `hermes --tui`      → `node dist/entry.js` (auto-builds when `.ts`/`.tsx` newer than `dist/entry.js`)
-- `hermes --tui --dev` → `tsx src/entry.tsx` (skip build)
-- `HERMES_TUI_DIR=…`  → external prebuilt dist (nix, distro packaging)
-
-## Diverged From Original Plan
-
-| Plan | Reality | Why |
-|---|---|---|
-| `tui_gateway/{controller,session_state,events,protocol}.py` | all collapsed into `server.py` | no second consumer ever emerged, keeping one file cheaper than four |
-| `ui-tui/src/main.tsx` | split into `entry.tsx` (bootstrap) + `app.tsx` (shell) | boot banner + early python spawn wanted a pre-React moment |
-| `ui-tui/src/state/store.ts` | three nanostores (`uiStore`, `turnStore`, `overlayStore`) | separate lifetimes: ui persists, turn resets per reply, overlay is modal |
-| `approval.requested` / `sudo.requested` / `clarify.requested` | `*.request` (no `-ed`) | cosmetic |
-| `session.cancel` | dropped | `session.interrupt` covers it |
-| `HERMES_EXPERIMENTAL_TUI=1`, `display.experimental_tui: true`, `/tui on/off/status` | none shipped | `--tui` went from opt-in to first-class without an experimental phase |
-
-## Post-migration Additions (not in original plan)
-
-- **Async `session.create`** — returns sid in ~1ms, agent builds on a background thread, `session.info` broadcasts when ready; `_wait_agent()` gates every agent-touching handler via `_sess`
-- **`bootBanner`** — raw-ANSI logo painted to stdout at T≈2ms, before Ink loads; `<AlternateScreen>` wipes it seamlessly when React mounts
-- **Selection uniform bg** — `theme.color.selectionBg` wired via `useSelection().setSelectionBgColor`; replaces SGR-inverse per-cell swap that fragmented over amber/gold fg
-- **Slash command registry** — TS-owned commands in `app/slash/commands/{core,ops,session}.ts`, everything else falls through to `slash.exec` (python worker)
-- **Turn store + controller split** — imperative singleton (`turnController`) holds refs/timers, nanostore (`turnStore`) holds render-visible state
-
-## What's Still Open
-
-- **Classic CLI not deleted.** `cli.py` still has ~80 `prompt_toolkit` references; classic REPL is still the default when `--tui` is absent. The original plan's "Cut 4 · prompt_toolkit removal later" hasn't happened.
-- **No config-file opt-in.** `HERMES_EXPERIMENTAL_TUI` and `display.experimental_tui` were never built; only the CLI flag exists. Fine for now — if we want "default to TUI", a single line in `main.py` flips it.

docs/skins/example-skin.yaml

@@ -1,106 +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
-#
-# Keys are marked:
-#   (both)    — applies to both the classic CLI and the TUI
-#   (classic) — classic CLI only (see hermes --tui in user-guide/tui.md)
-#   (tui)     — TUI only
-#
-# 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. These control the visual palette.
-colors:
-  # Banner panel (the startup welcome box) — (both)
-  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 — (both)
-  ui_accent: "#FFBF00"            # General accent (falls back to banner_accent)
-  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 / `❯` glyph color (both)
-  input_rule: "#CD7F32"           # Horizontal rule above input (classic)
-
-  # Response box — (classic)
-  response_border: "#FFD700"      # Response box border
-
-  # Session display — (both)
-  session_label: "#DAA520"        # "Session: " label
-  session_border: "#8B8682"       # Session ID text
-
-  # TUI / CLI surfaces — (classic: status bar, voice badge, completion meta)
-  status_bar_bg: "#1a1a2e"              # Status / usage bar background (classic)
-  voice_status_bg: "#1a1a2e"            # Voice-mode badge background (classic)
-  completion_menu_bg: "#1a1a2e"         # Completion list background (both)
-  completion_menu_current_bg: "#333355" # Active completion row background (both)
-  completion_menu_meta_bg: "#1a1a2e"    # Completion meta column bg (classic)
-  completion_menu_meta_current_bg: "#333355"  # Active meta bg (classic)
-
-  # Drag-to-select background — (tui)
-  selection_bg: "#3a3a55"               # Uniform selection highlight in the TUI
-
-# ── Spinner ─────────────────────────────────────────────────────────────────
-# (classic) — the TUI uses its own animated indicators; spinner config here
-# is only read by the classic prompt_toolkit CLI.
-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 interface.
-branding:
-  agent_name: "Hermes Agent"                  # (both) Banner title, about display
-  welcome: "Welcome! Type your message or /help for commands."  # (both)
-  goodbye: "Goodbye! ⚕"                       # (both) Exit message
-  response_label: " ⚕ Hermes "                # (classic) Response box header label
-  prompt_symbol: "❯ "                          # (both) Input prompt glyph
-  help_header: "(^_^)? Available Commands"     # (both) /help overlay title
-
-# ── Tool Output ─────────────────────────────────────────────────────────────
-# Character used as the prefix for tool output lines. (both)
-# 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`)

environments/tool_context.py

@@ -53,7 +53,6 @@ def _run_tool_in_thread(tool_name: str, arguments: Dict[str, Any], task_id: str)
     try:
         loop = asyncio.get_running_loop()
         # We're in an async context -- need to run in thread
-        import concurrent.futures
         with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
             future = pool.submit(
                 handle_function_call, tool_name, arguments, task_id

gateway/config.py

@@ -576,6 +576,14 @@ def load_gateway_config() -> GatewayConfig:
                     bridged["free_response_channels"] = platform_cfg["free_response_channels"]
                 if "mention_patterns" in platform_cfg:
                     bridged["mention_patterns"] = platform_cfg["mention_patterns"]
+                if "dm_policy" in platform_cfg:
+                    bridged["dm_policy"] = platform_cfg["dm_policy"]
+                if "allow_from" in platform_cfg:
+                    bridged["allow_from"] = platform_cfg["allow_from"]
+                if "group_policy" in platform_cfg:
+                    bridged["group_policy"] = platform_cfg["group_policy"]
+                if "group_allow_from" in platform_cfg:
+                    bridged["group_allow_from"] = platform_cfg["group_allow_from"]
                 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:
@@ -662,8 +670,7 @@ def load_gateway_config() -> GatewayConfig:
                 if "require_mention" in telegram_cfg and not os.getenv("TELEGRAM_REQUIRE_MENTION"):
                     os.environ["TELEGRAM_REQUIRE_MENTION"] = str(telegram_cfg["require_mention"]).lower()
                 if "mention_patterns" in telegram_cfg and not os.getenv("TELEGRAM_MENTION_PATTERNS"):
-                    import json as _json
-                    os.environ["TELEGRAM_MENTION_PATTERNS"] = _json.dumps(telegram_cfg["mention_patterns"])
+                    os.environ["TELEGRAM_MENTION_PATTERNS"] = json.dumps(telegram_cfg["mention_patterns"])
                 frc = telegram_cfg.get("free_response_chats")
                 if frc is not None and not os.getenv("TELEGRAM_FREE_RESPONSE_CHATS"):
                     if isinstance(frc, list):
@@ -700,6 +707,20 @@ def load_gateway_config() -> GatewayConfig:
                     if isinstance(frc, list):
                         frc = ",".join(str(v) for v in frc)
                     os.environ["WHATSAPP_FREE_RESPONSE_CHATS"] = str(frc)
+                if "dm_policy" in whatsapp_cfg and not os.getenv("WHATSAPP_DM_POLICY"):
+                    os.environ["WHATSAPP_DM_POLICY"] = str(whatsapp_cfg["dm_policy"]).lower()
+                af = whatsapp_cfg.get("allow_from")
+                if af is not None and not os.getenv("WHATSAPP_ALLOWED_USERS"):
+                    if isinstance(af, list):
+                        af = ",".join(str(v) for v in af)
+                    os.environ["WHATSAPP_ALLOWED_USERS"] = str(af)
+                if "group_policy" in whatsapp_cfg and not os.getenv("WHATSAPP_GROUP_POLICY"):
+                    os.environ["WHATSAPP_GROUP_POLICY"] = str(whatsapp_cfg["group_policy"]).lower()
+                gaf = whatsapp_cfg.get("group_allow_from")
+                if gaf is not None and not os.getenv("WHATSAPP_GROUP_ALLOWED_USERS"):
+                    if isinstance(gaf, list):
+                        gaf = ",".join(str(v) for v in gaf)
+                    os.environ["WHATSAPP_GROUP_ALLOWED_USERS"] = str(gaf)
 
             # DingTalk settings → env vars (env vars take precedence)
             dingtalk_cfg = yaml_cfg.get("dingtalk", {})
@@ -1237,7 +1258,6 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             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."

gateway/platforms/api_server.py

@@ -117,6 +117,160 @@ def _normalize_chat_content(
         return ""
 
 
+# Content part type aliases used by the OpenAI Chat Completions and Responses
+# APIs.  We accept both spellings on input and emit a single canonical internal
+# shape (``{"type": "text", ...}`` / ``{"type": "image_url", ...}``) that the
+# rest of the agent pipeline already understands.
+_TEXT_PART_TYPES = frozenset({"text", "input_text", "output_text"})
+_IMAGE_PART_TYPES = frozenset({"image_url", "input_image"})
+_FILE_PART_TYPES = frozenset({"file", "input_file"})
+
+
+def _normalize_multimodal_content(content: Any) -> Any:
+    """Validate and normalize multimodal content for the API server.
+
+    Returns a plain string when the content is text-only, or a list of
+    ``{"type": "text"|"image_url", ...}`` parts when images are present.
+    The output shape is the native OpenAI Chat Completions vision format,
+    which the agent pipeline accepts verbatim (OpenAI-wire providers) or
+    converts (``_preprocess_anthropic_content`` for Anthropic).
+
+    Raises ``ValueError`` with an OpenAI-style code on invalid input:
+      * ``unsupported_content_type`` — file/input_file/file_id parts, or
+        non-image ``data:`` URLs.
+      * ``invalid_image_url`` — missing URL or unsupported scheme.
+      * ``invalid_content_part`` — malformed text/image objects.
+
+    Callers translate the ValueError into a 400 response.
+    """
+    # Scalar passthrough mirrors ``_normalize_chat_content``.
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content[:MAX_NORMALIZED_TEXT_LENGTH] if len(content) > MAX_NORMALIZED_TEXT_LENGTH else content
+    if not isinstance(content, list):
+        # Mirror the legacy text-normalizer's fallback so callers that
+        # pre-existed image support still get a string back.
+        return _normalize_chat_content(content)
+
+    items = content[:MAX_CONTENT_LIST_SIZE] if len(content) > MAX_CONTENT_LIST_SIZE else content
+    normalized_parts: List[Dict[str, Any]] = []
+    text_accum_len = 0
+
+    for part in items:
+        if isinstance(part, str):
+            if part:
+                trimmed = part[:MAX_NORMALIZED_TEXT_LENGTH]
+                normalized_parts.append({"type": "text", "text": trimmed})
+                text_accum_len += len(trimmed)
+            continue
+
+        if not isinstance(part, dict):
+            # Ignore unknown scalars for forward compatibility with future
+            # Responses API additions (e.g. ``refusal``).  The same policy
+            # the text normalizer applies.
+            continue
+
+        raw_type = part.get("type")
+        part_type = str(raw_type or "").strip().lower()
+
+        if part_type in _TEXT_PART_TYPES:
+            text = part.get("text")
+            if text is None:
+                continue
+            if not isinstance(text, str):
+                text = str(text)
+            if text:
+                trimmed = text[:MAX_NORMALIZED_TEXT_LENGTH]
+                normalized_parts.append({"type": "text", "text": trimmed})
+                text_accum_len += len(trimmed)
+            continue
+
+        if part_type in _IMAGE_PART_TYPES:
+            detail = part.get("detail")
+            image_ref = part.get("image_url")
+            # OpenAI Responses sends ``input_image`` with a top-level
+            # ``image_url`` string; Chat Completions sends ``image_url`` as
+            # ``{"url": "...", "detail": "..."}``.  Support both.
+            if isinstance(image_ref, dict):
+                url_value = image_ref.get("url")
+                detail = image_ref.get("detail", detail)
+            else:
+                url_value = image_ref
+            if not isinstance(url_value, str) or not url_value.strip():
+                raise ValueError("invalid_image_url:Image parts must include a non-empty image URL.")
+            url_value = url_value.strip()
+            lowered = url_value.lower()
+            if lowered.startswith("data:"):
+                if not lowered.startswith("data:image/") or "," not in url_value:
+                    raise ValueError(
+                        "unsupported_content_type:Only image data URLs are supported. "
+                        "Non-image data payloads are not supported."
+                    )
+            elif not (lowered.startswith("http://") or lowered.startswith("https://")):
+                raise ValueError(
+                    "invalid_image_url:Image inputs must use http(s) URLs or data:image/... URLs."
+                )
+            image_part: Dict[str, Any] = {"type": "image_url", "image_url": {"url": url_value}}
+            if detail is not None:
+                if not isinstance(detail, str) or not detail.strip():
+                    raise ValueError("invalid_content_part:Image detail must be a non-empty string when provided.")
+                image_part["image_url"]["detail"] = detail.strip()
+            normalized_parts.append(image_part)
+            continue
+
+        if part_type in _FILE_PART_TYPES:
+            raise ValueError(
+                "unsupported_content_type:Inline image inputs are supported, "
+                "but uploaded files and document inputs are not supported on this endpoint."
+            )
+
+        # Unknown part type — reject explicitly so clients get a clear error
+        # instead of a silently dropped turn.
+        raise ValueError(
+            f"unsupported_content_type:Unsupported content part type {raw_type!r}. "
+            "Only text and image_url/input_image parts are supported."
+        )
+
+    if not normalized_parts:
+        return ""
+
+    # Text-only: collapse to a plain string so downstream logging/trajectory
+    # code sees the native shape and prompt caching on text-only turns is
+    # unaffected.
+    if all(p.get("type") == "text" for p in normalized_parts):
+        return "\n".join(p["text"] for p in normalized_parts if p.get("text"))
+
+    return normalized_parts
+
+
+def _content_has_visible_payload(content: Any) -> bool:
+    """True when content has any text or image attachment.  Used to reject empty turns."""
+    if isinstance(content, str):
+        return bool(content.strip())
+    if isinstance(content, list):
+        for part in content:
+            if isinstance(part, dict):
+                ptype = str(part.get("type") or "").strip().lower()
+                if ptype in _TEXT_PART_TYPES and str(part.get("text") or "").strip():
+                    return True
+                if ptype in _IMAGE_PART_TYPES:
+                    return True
+    return False
+
+
+def _multimodal_validation_error(exc: ValueError, *, param: str) -> "web.Response":
+    """Translate a ``_normalize_multimodal_content`` ValueError into a 400 response."""
+    raw = str(exc)
+    code, _, message = raw.partition(":")
+    if not message:
+        code, message = "invalid_content_part", raw
+    return web.json_response(
+        _openai_error(message, code=code, param=param),
+        status=400,
+    )
+
+
 def check_api_server_requirements() -> bool:
     """Check if API server dependencies are available."""
     return AIOHTTP_AVAILABLE
@@ -169,7 +323,6 @@ class ResponseStore:
         ).fetchone()
         if row is None:
             return None
-        import time
         self._conn.execute(
             "UPDATE responses SET accessed_at = ? WHERE response_id = ?",
             (time.time(), response_id),
@@ -179,7 +332,6 @@ class ResponseStore:
 
     def put(self, response_id: str, data: Dict[str, Any]) -> None:
         """Store a response, evicting the oldest if at capacity."""
-        import time
         self._conn.execute(
             "INSERT OR REPLACE INTO responses (response_id, data, accessed_at) VALUES (?, ?, ?)",
             (response_id, json.dumps(data, default=str), time.time()),
@@ -315,12 +467,12 @@ class _IdempotencyCache:
     def __init__(self, max_items: int = 1000, ttl_seconds: int = 300):
         from collections import OrderedDict
         self._store = OrderedDict()
+        self._inflight: Dict[tuple[str, str], "asyncio.Task[Any]"] = {}
         self._ttl = ttl_seconds
         self._max = max_items
 
     def _purge(self):
-        import time as _t
-        now = _t.time()
+        now = time.time()
         expired = [k for k, v in self._store.items() if now - v["ts"] > self._ttl]
         for k in expired:
             self._store.pop(k, None)
@@ -332,11 +484,27 @@ class _IdempotencyCache:
         item = self._store.get(key)
         if item and item["fp"] == fingerprint:
             return item["resp"]
-        resp = await compute_coro()
-        import time as _t
-        self._store[key] = {"resp": resp, "fp": fingerprint, "ts": _t.time()}
-        self._purge()
-        return resp
+
+        inflight_key = (key, fingerprint)
+        task = self._inflight.get(inflight_key)
+        if task is None:
+            async def _compute_and_store():
+                resp = await compute_coro()
+                import time as _t
+                self._store[key] = {"resp": resp, "fp": fingerprint, "ts": _t.time()}
+                self._purge()
+                return resp
+
+            task = asyncio.create_task(_compute_and_store())
+            self._inflight[inflight_key] = task
+
+            def _clear_inflight(done_task: "asyncio.Task[Any]") -> None:
+                if self._inflight.get(inflight_key) is done_task:
+                    self._inflight.pop(inflight_key, None)
+
+            task.add_done_callback(_clear_inflight)
+
+        return await asyncio.shield(task)
 
 
 _idem_cache = _IdempotencyCache()
@@ -366,6 +534,30 @@ def _derive_chat_session_id(
     return f"api-{digest}"
 
 
+_CRON_AVAILABLE = False
+try:
+    from cron.jobs import (
+        list_jobs as _cron_list,
+        get_job as _cron_get,
+        create_job as _cron_create,
+        update_job as _cron_update,
+        remove_job as _cron_remove,
+        pause_job as _cron_pause,
+        resume_job as _cron_resume,
+        trigger_job as _cron_trigger,
+    )
+    _CRON_AVAILABLE = True
+except ImportError:
+    _cron_list = None
+    _cron_get = None
+    _cron_create = None
+    _cron_update = None
+    _cron_remove = None
+    _cron_pause = None
+    _cron_resume = None
+    _cron_trigger = None
+
+
 class APIServerAdapter(BasePlatformAdapter):
     """
     OpenAI-compatible HTTP API server adapter.
@@ -637,26 +829,32 @@ class APIServerAdapter(BasePlatformAdapter):
         system_prompt = None
         conversation_messages: List[Dict[str, str]] = []
 
-        for msg in messages:
+        for idx, msg in enumerate(messages):
             role = msg.get("role", "")
-            content = _normalize_chat_content(msg.get("content", ""))
+            raw_content = msg.get("content", "")
             if role == "system":
-                # Accumulate system messages
+                # System messages don't support images (Anthropic rejects, OpenAI
+                # text-model systems don't render them).  Flatten to text.
+                content = _normalize_chat_content(raw_content)
                 if system_prompt is None:
                     system_prompt = content
                 else:
                     system_prompt = system_prompt + "\n" + content
             elif role in ("user", "assistant"):
+                try:
+                    content = _normalize_multimodal_content(raw_content)
+                except ValueError as exc:
+                    return _multimodal_validation_error(exc, param=f"messages[{idx}].content")
                 conversation_messages.append({"role": role, "content": content})
 
         # Extract the last user message as the primary input
-        user_message = ""
+        user_message: Any = ""
         history = []
         if conversation_messages:
             user_message = conversation_messages[-1].get("content", "")
             history = conversation_messages[:-1]
 
-        if not user_message:
+        if not _content_has_visible_payload(user_message):
             return web.json_response(
                 {"error": {"message": "No user message found in messages", "type": "invalid_request_error"}},
                 status=400,
@@ -1424,16 +1622,19 @@ class APIServerAdapter(BasePlatformAdapter):
             # No error if conversation doesn't exist yet — it's a new conversation
 
         # Normalize input to message list
-        input_messages: List[Dict[str, str]] = []
+        input_messages: List[Dict[str, Any]] = []
         if isinstance(raw_input, str):
             input_messages = [{"role": "user", "content": raw_input}]
         elif isinstance(raw_input, list):
-            for item in raw_input:
+            for idx, item in enumerate(raw_input):
                 if isinstance(item, str):
                     input_messages.append({"role": "user", "content": item})
                 elif isinstance(item, dict):
                     role = item.get("role", "user")
-                    content = _normalize_chat_content(item.get("content", ""))
+                    try:
+                        content = _normalize_multimodal_content(item.get("content", ""))
+                    except ValueError as exc:
+                        return _multimodal_validation_error(exc, param=f"input[{idx}].content")
                     input_messages.append({"role": role, "content": content})
         else:
             return web.json_response(_openai_error("'input' must be a string or array"), status=400)
@@ -1442,7 +1643,7 @@ class APIServerAdapter(BasePlatformAdapter):
         # This lets stateless clients supply their own history instead of
         # relying on server-side response chaining via previous_response_id.
         # Precedence: explicit conversation_history > previous_response_id.
-        conversation_history: List[Dict[str, str]] = []
+        conversation_history: List[Dict[str, Any]] = []
         raw_history = body.get("conversation_history")
         if raw_history:
             if not isinstance(raw_history, list):
@@ -1456,7 +1657,11 @@ class APIServerAdapter(BasePlatformAdapter):
                         _openai_error(f"conversation_history[{i}] must have 'role' and 'content' fields"),
                         status=400,
                     )
-                conversation_history.append({"role": str(entry["role"]), "content": str(entry["content"])})
+                try:
+                    entry_content = _normalize_multimodal_content(entry["content"])
+                except ValueError as exc:
+                    return _multimodal_validation_error(exc, param=f"conversation_history[{i}].content")
+                conversation_history.append({"role": str(entry["role"]), "content": entry_content})
             if previous_response_id:
                 logger.debug("Both conversation_history and previous_response_id provided; using conversation_history")
 
@@ -1476,8 +1681,8 @@ class APIServerAdapter(BasePlatformAdapter):
             conversation_history.append(msg)
 
         # Last input message is the user_message
-        user_message = input_messages[-1].get("content", "") if input_messages else ""
-        if not user_message:
+        user_message: Any = input_messages[-1].get("content", "") if input_messages else ""
+        if not _content_has_visible_payload(user_message):
             return web.json_response(_openai_error("No user message found in input"), status=400)
 
         # Truncation support
@@ -1682,44 +1887,16 @@ class APIServerAdapter(BasePlatformAdapter):
     # Cron jobs API
     # ------------------------------------------------------------------
 
-    # Check cron module availability once (not per-request)
-    _CRON_AVAILABLE = False
-    try:
-        from cron.jobs import (
-            list_jobs as _cron_list,
-            get_job as _cron_get,
-            create_job as _cron_create,
-            update_job as _cron_update,
-            remove_job as _cron_remove,
-            pause_job as _cron_pause,
-            resume_job as _cron_resume,
-            trigger_job as _cron_trigger,
-        )
-        # Wrap as staticmethod to prevent descriptor binding — these are plain
-        # module functions, not instance methods.  Without this, self._cron_*()
-        # injects ``self`` as the first positional argument and every call
-        # raises TypeError.
-        _cron_list = staticmethod(_cron_list)
-        _cron_get = staticmethod(_cron_get)
-        _cron_create = staticmethod(_cron_create)
-        _cron_update = staticmethod(_cron_update)
-        _cron_remove = staticmethod(_cron_remove)
-        _cron_pause = staticmethod(_cron_pause)
-        _cron_resume = staticmethod(_cron_resume)
-        _cron_trigger = staticmethod(_cron_trigger)
-        _CRON_AVAILABLE = True
-    except ImportError:
-        pass
-
     _JOB_ID_RE = __import__("re").compile(r"[a-f0-9]{12}")
     # Allowed fields for update — prevents clients injecting arbitrary keys
     _UPDATE_ALLOWED_FIELDS = {"name", "schedule", "prompt", "deliver", "skills", "skill", "repeat", "enabled"}
     _MAX_NAME_LENGTH = 200
     _MAX_PROMPT_LENGTH = 5000
 
-    def _check_jobs_available(self) -> Optional["web.Response"]:
+    @staticmethod
+    def _check_jobs_available() -> Optional["web.Response"]:
         """Return error response if cron module isn't available."""
-        if not self._CRON_AVAILABLE:
+        if not _CRON_AVAILABLE:
             return web.json_response(
                 {"error": "Cron module not available"}, status=501,
             )
@@ -1744,7 +1921,7 @@ class APIServerAdapter(BasePlatformAdapter):
             return cron_err
         try:
             include_disabled = request.query.get("include_disabled", "").lower() in ("true", "1")
-            jobs = self._cron_list(include_disabled=include_disabled)
+            jobs = _cron_list(include_disabled=include_disabled)
             return web.json_response({"jobs": jobs})
         except Exception as e:
             return web.json_response({"error": str(e)}, status=500)
@@ -1792,7 +1969,7 @@ class APIServerAdapter(BasePlatformAdapter):
             if repeat is not None:
                 kwargs["repeat"] = repeat
 
-            job = self._cron_create(**kwargs)
+            job = _cron_create(**kwargs)
             return web.json_response({"job": job})
         except Exception as e:
             return web.json_response({"error": str(e)}, status=500)
@@ -1809,7 +1986,7 @@ class APIServerAdapter(BasePlatformAdapter):
         if id_err:
             return id_err
         try:
-            job = self._cron_get(job_id)
+            job = _cron_get(job_id)
             if not job:
                 return web.json_response({"error": "Job not found"}, status=404)
             return web.json_response({"job": job})
@@ -1842,7 +2019,7 @@ class APIServerAdapter(BasePlatformAdapter):
                 return web.json_response(
                     {"error": f"Prompt must be ≤ {self._MAX_PROMPT_LENGTH} characters"}, status=400,
                 )
-            job = self._cron_update(job_id, sanitized)
+            job = _cron_update(job_id, sanitized)
             if not job:
                 return web.json_response({"error": "Job not found"}, status=404)
             return web.json_response({"job": job})
@@ -1861,7 +2038,7 @@ class APIServerAdapter(BasePlatformAdapter):
         if id_err:
             return id_err
         try:
-            success = self._cron_remove(job_id)
+            success = _cron_remove(job_id)
             if not success:
                 return web.json_response({"error": "Job not found"}, status=404)
             return web.json_response({"ok": True})
@@ -1880,7 +2057,7 @@ class APIServerAdapter(BasePlatformAdapter):
         if id_err:
             return id_err
         try:
-            job = self._cron_pause(job_id)
+            job = _cron_pause(job_id)
             if not job:
                 return web.json_response({"error": "Job not found"}, status=404)
             return web.json_response({"job": job})
@@ -1899,7 +2076,7 @@ class APIServerAdapter(BasePlatformAdapter):
         if id_err:
             return id_err
         try:
-            job = self._cron_resume(job_id)
+            job = _cron_resume(job_id)
             if not job:
                 return web.json_response({"error": "Job not found"}, status=404)
             return web.json_response({"job": job})
@@ -1918,7 +2095,7 @@ class APIServerAdapter(BasePlatformAdapter):
         if id_err:
             return id_err
         try:
-            job = self._cron_trigger(job_id)
+            job = _cron_trigger(job_id)
             if not job:
                 return web.json_response({"error": "Job not found"}, status=404)
             return web.json_response({"job": job})

gateway/platforms/base.py

@@ -19,6 +19,8 @@ import uuid
 from abc import ABC, abstractmethod
 from urllib.parse import urlsplit
 
+from utils import normalize_proxy_url
+
 logger = logging.getLogger(__name__)
 
 
@@ -159,13 +161,13 @@ def resolve_proxy_url(platform_env_var: str | None = None) -> str | None:
     if platform_env_var:
         value = (os.environ.get(platform_env_var) or "").strip()
         if value:
-            return value
+            return normalize_proxy_url(value)
     for key in ("HTTPS_PROXY", "HTTP_PROXY", "ALL_PROXY",
                 "https_proxy", "http_proxy", "all_proxy"):
         value = (os.environ.get(key) or "").strip()
         if value:
-            return value
-    return _detect_macos_system_proxy()
+            return normalize_proxy_url(value)
+    return normalize_proxy_url(_detect_macos_system_proxy())
 
 
 def proxy_kwargs_for_bot(proxy_url: str | None) -> dict:
@@ -391,12 +393,9 @@ async def cache_image_from_url(url: str, ext: str = ".jpg", retries: int = 2) ->
     if not is_safe_url(url):
         raise ValueError(f"Blocked unsafe URL (SSRF protection): {safe_url_for_log(url)}")
 
-    import asyncio
     import httpx
-    import logging as _logging
-    _log = _logging.getLogger(__name__)
+    _log = logging.getLogger(__name__)
 
-    last_exc = None
     async with httpx.AsyncClient(
         timeout=30.0,
         follow_redirects=True,
@@ -414,7 +413,6 @@ async def cache_image_from_url(url: str, ext: str = ".jpg", retries: int = 2) ->
                 response.raise_for_status()
                 return cache_image_from_bytes(response.content, ext)
             except (httpx.TimeoutException, httpx.HTTPStatusError) as exc:
-                last_exc = exc
                 if isinstance(exc, httpx.HTTPStatusError) and exc.response.status_code < 429:
                     raise
                 if attempt < retries:
@@ -430,7 +428,6 @@ async def cache_image_from_url(url: str, ext: str = ".jpg", retries: int = 2) ->
                     await asyncio.sleep(wait)
                     continue
                 raise
-    raise last_exc
 
 
 def cleanup_image_cache(max_age_hours: int = 24) -> int:
@@ -510,12 +507,9 @@ async def cache_audio_from_url(url: str, ext: str = ".ogg", retries: int = 2) ->
     if not is_safe_url(url):
         raise ValueError(f"Blocked unsafe URL (SSRF protection): {safe_url_for_log(url)}")
 
-    import asyncio
     import httpx
-    import logging as _logging
-    _log = _logging.getLogger(__name__)
+    _log = logging.getLogger(__name__)
 
-    last_exc = None
     async with httpx.AsyncClient(
         timeout=30.0,
         follow_redirects=True,
@@ -533,7 +527,6 @@ async def cache_audio_from_url(url: str, ext: str = ".ogg", retries: int = 2) ->
                 response.raise_for_status()
                 return cache_audio_from_bytes(response.content, ext)
             except (httpx.TimeoutException, httpx.HTTPStatusError) as exc:
-                last_exc = exc
                 if isinstance(exc, httpx.HTTPStatusError) and exc.response.status_code < 429:
                     raise
                 if attempt < retries:
@@ -549,7 +542,39 @@ async def cache_audio_from_url(url: str, ext: str = ".ogg", retries: int = 2) ->
                     await asyncio.sleep(wait)
                     continue
                 raise
-    raise last_exc
+
+
+# ---------------------------------------------------------------------------
+# Video cache utilities
+#
+# Same pattern as image/audio cache -- videos from platforms are downloaded
+# here so the agent can reference them by local file path.
+# ---------------------------------------------------------------------------
+
+VIDEO_CACHE_DIR = get_hermes_dir("cache/videos", "video_cache")
+
+SUPPORTED_VIDEO_TYPES = {
+    ".mp4": "video/mp4",
+    ".mov": "video/quicktime",
+    ".webm": "video/webm",
+    ".mkv": "video/x-matroska",
+    ".avi": "video/x-msvideo",
+}
+
+
+def get_video_cache_dir() -> Path:
+    """Return the video cache directory, creating it if it doesn't exist."""
+    VIDEO_CACHE_DIR.mkdir(parents=True, exist_ok=True)
+    return VIDEO_CACHE_DIR
+
+
+def cache_video_from_bytes(data: bytes, ext: str = ".mp4") -> str:
+    """Save raw video bytes to the cache and return the absolute file path."""
+    cache_dir = get_video_cache_dir()
+    filename = f"video_{uuid.uuid4().hex[:12]}{ext}"
+    filepath = cache_dir / filename
+    filepath.write_bytes(data)
+    return str(filepath)
 
 
 # ---------------------------------------------------------------------------
@@ -1318,7 +1343,7 @@ class BasePlatformAdapter(ABC):
         # Extract MEDIA:<path> tags, allowing optional whitespace after the colon
         # and quoted/backticked paths for LLM-formatted outputs.
         media_pattern = re.compile(
-            r'''[`"']?MEDIA:\s*(?P<path>`[^`\n]+`|"[^"\n]+"|'[^'\n]+'|(?:~/|/)\S+(?:[^\S\n]+\S+)*?\.(?:png|jpe?g|gif|webp|mp4|mov|avi|mkv|webm|ogg|opus|mp3|wav|m4a)(?=[\s`"',;:)\]}]|$)|\S+)[`"']?'''
+            r'''[`"']?MEDIA:\s*(?P<path>`[^`\n]+`|"[^"\n]+"|'[^'\n]+'|(?:~/|/)\S+(?:[^\S\n]+\S+)*?\.(?:png|jpe?g|gif|webp|mp4|mov|avi|mkv|webm|ogg|opus|mp3|wav|m4a|pdf)(?=[\s`"',;:)\]}]|$)|\S+)[`"']?'''
         )
         for match in media_pattern.finditer(content):
             path = match.group("path").strip()
@@ -1754,8 +1779,6 @@ class BasePlatformAdapter(ABC):
           HERMES_HUMAN_DELAY_MIN_MS: minimum delay in ms (default 800, custom mode)
           HERMES_HUMAN_DELAY_MAX_MS: maximum delay in ms (default 2500, custom mode)
         """
-        import random
-
         mode = os.getenv("HERMES_HUMAN_DELAY_MODE", "off").lower()
         if mode == "off":
             return 0.0

gateway/platforms/bluebubbles.py

@@ -75,7 +75,7 @@ def _redact(text: str) -> str:
 def check_bluebubbles_requirements() -> bool:
     try:
         import aiohttp  # noqa: F401
-        import httpx as _httpx  # noqa: F401
+        import httpx  # noqa: F401
     except ImportError:
         return False
     return True

gateway/platforms/discord.py

@@ -541,7 +541,6 @@ class DiscordAdapter(BasePlatformAdapter):
             # ctypes.util.find_library fails on macOS with Homebrew-installed libs,
             # so fall back to known Homebrew paths if needed.
             if not opus_path:
-                import sys
                 _homebrew_paths = (
                     "/opt/homebrew/lib/libopus.dylib",  # Apple Silicon
                     "/usr/local/lib/libopus.dylib",     # Intel Mac
@@ -637,29 +636,14 @@ class DiscordAdapter(BasePlatformAdapter):
 
             @self._client.event
             async def on_message(message: DiscordMessage):
-                # Wait for on_ready to finish resolving username-based
-                # allowlist entries.  Without this block, messages
-                # arriving between Discord's READY event and the end
-                # of _resolve_allowed_usernames compare author IDs
-                # (numeric) against a set that may still contain raw
-                # usernames (strings) from DISCORD_ALLOWED_USERS —
-                # legitimate users get silently rejected for the first
-                # few seconds after every reconnect.  The wait is a
-                # near-instant no-op in steady state (_ready_event is
-                # already set); only the startup / reconnect window
-                # ever blocks.
+                # Block until _resolve_allowed_usernames has swapped
+                # any raw usernames in DISCORD_ALLOWED_USERS for numeric
+                # IDs (otherwise on_message's author.id lookup can miss).
                 if not adapter_self._ready_event.is_set():
                     try:
-                        await asyncio.wait_for(
-                            adapter_self._ready_event.wait(),
-                            timeout=30.0,
-                        )
+                        await asyncio.wait_for(adapter_self._ready_event.wait(), timeout=30.0)
                     except asyncio.TimeoutError:
-                        logger.warning(
-                            "[%s] on_message timed out waiting for _ready_event; "
-                            "allowlist check may use pre-resolved entries",
-                            adapter_self.name,
-                        )
+                        pass
 
                 # Dedup: Discord RESUME replays events after reconnects (#4777)
                 if adapter_self._dedup.is_duplicate(str(message.id)):
@@ -1096,6 +1080,8 @@ class DiscordAdapter(BasePlatformAdapter):
         chat_id: str,
         message_id: str,
         content: str,
+        *,
+        finalize: bool = False,
     ) -> SendResult:
         """Edit a previously sent Discord message."""
         if not self._client:
@@ -1256,28 +1242,13 @@ class DiscordAdapter(BasePlatformAdapter):
     # Voice channel methods (join / leave / play)
     # ------------------------------------------------------------------
 
-    def _voice_lock_for(self, guild_id: int) -> "asyncio.Lock":
-        """Return the per-guild lock, creating it on first use.
-
-        Voice join/leave/move must be serialized per guild — without
-        this, two concurrent /voice channel invocations both see
-        _voice_clients.get(guild_id) return None, both call
-        channel.connect(), and discord.py raises ClientException
-        ('Already connected') on the loser.
-        """
-        lock = self._voice_locks.get(guild_id)
-        if lock is None:
-            lock = asyncio.Lock()
-            self._voice_locks[guild_id] = lock
-        return lock
-
     async def join_voice_channel(self, channel) -> bool:
         """Join a Discord voice channel. Returns True on success."""
         if not self._client or not DISCORD_AVAILABLE:
             return False
         guild_id = channel.guild.id
 
-        async with self._voice_lock_for(guild_id):
+        async with self._voice_locks.setdefault(guild_id, asyncio.Lock()):
             # Already connected in this guild?
             existing = self._voice_clients.get(guild_id)
             if existing and existing.is_connected():
@@ -1307,7 +1278,7 @@ class DiscordAdapter(BasePlatformAdapter):
 
     async def leave_voice_channel(self, guild_id: int) -> None:
         """Disconnect from the voice channel in a guild."""
-        async with self._voice_lock_for(guild_id):
+        async with self._voice_locks.setdefault(guild_id, asyncio.Lock()):
             # Stop voice receiver first
             receiver = self._voice_receivers.pop(guild_id, None)
             if receiver:
@@ -1450,8 +1421,7 @@ class DiscordAdapter(BasePlatformAdapter):
         speaking_user_ids: set = set()
         receiver = self._voice_receivers.get(guild_id)
         if receiver:
-            import time as _time
-            now = _time.monotonic()
+            now = time.monotonic()
             with receiver._lock:
                 for ssrc, last_t in receiver._last_packet_time.items():
                     # Consider "speaking" if audio received within last 2 seconds
@@ -2990,6 +2960,17 @@ class DiscordAdapter(BasePlatformAdapter):
             parent_channel_id = self._get_parent_channel_id(message.channel)
 
         is_voice_linked_channel = False
+
+        # Save mention-stripped text before auto-threading since create_thread()
+        # can clobber message.content, breaking /command detection in channels.
+        raw_content = message.content.strip()
+        normalized_content = raw_content
+        mention_prefix = False
+        if self._client.user and self._client.user in message.mentions:
+            mention_prefix = True
+            normalized_content = normalized_content.replace(f"<@{self._client.user.id}>", "").strip()
+            normalized_content = normalized_content.replace(f"<@!{self._client.user.id}>", "").strip()
+            message.content = normalized_content
         if not isinstance(message.channel, discord.DMChannel):
             channel_ids = {str(message.channel.id)}
             if parent_channel_id:
@@ -3027,13 +3008,8 @@ class DiscordAdapter(BasePlatformAdapter):
             in_bot_thread = is_thread and thread_id in self._threads
 
             if require_mention and not is_free_channel and not in_bot_thread:
-                if self._client.user not in message.mentions:
+                if self._client.user not in message.mentions and not mention_prefix:
                     return
-
-            if self._client.user and self._client.user in message.mentions:
-                message.content = message.content.replace(f"<@{self._client.user.id}>", "").strip()
-                message.content = message.content.replace(f"<@!{self._client.user.id}>", "").strip()
-
         # Auto-thread: when enabled, automatically create a thread for every
         # @mention in a text channel so each conversation is isolated (like Slack).
         # Messages already inside threads or DMs are unaffected.
@@ -3055,7 +3031,7 @@ class DiscordAdapter(BasePlatformAdapter):
 
         # Determine message type
         msg_type = MessageType.TEXT
-        if message.content.startswith("/"):
+        if normalized_content.startswith("/"):
             msg_type = MessageType.COMMAND
         elif message.attachments:
             # Check attachment types
@@ -3195,7 +3171,9 @@ class DiscordAdapter(BasePlatformAdapter):
                                 att.filename, e, exc_info=True,
                             )
 
-        event_text = message.content
+        # Use normalized_content (saved before auto-threading) instead of message.content,
+        # to detect /slash commands in channel messages.
+        event_text = normalized_content
         if pending_text_injection:
             event_text = f"{pending_text_injection}\n\n{event_text}" if event_text else pending_text_injection
 

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,
@@ -190,7 +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 = {
@@ -1141,6 +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
@@ -1468,6 +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:
@@ -1970,8 +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")
@@ -2048,12 +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)())
@@ -2277,33 +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 = (
@@ -2316,16 +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
     # =========================================================================
@@ -3294,6 +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:
@@ -3321,10 +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")
@@ -3333,17 +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/matrix.py

@@ -825,7 +825,7 @@ class MatrixAdapter(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 message (via m.replace)."""
 

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)
@@ -410,7 +410,6 @@ class MattermostAdapter(BasePlatformAdapter):
             logger.warning("Mattermost: blocked unsafe URL (SSRF protection)")
             return await self.send(chat_id, f"{caption or ''}\n{url}".strip(), reply_to)
 
-        import asyncio
         import aiohttp
 
         last_exc = None

gateway/platforms/qqbot/adapter.py

@@ -1086,11 +1086,8 @@ class QQAdapter(BasePlatformAdapter):
             return MessageType.VIDEO
         if "image" in first_type or "photo" in first_type:
             return MessageType.PHOTO
-        # Unknown content type with an attachment — don't assume PHOTO
-        # to prevent non-image files from being sent to vision analysis.
         logger.debug(
-            "[%s] Unknown media content_type '%s', defaulting to TEXT",
-            self._log_tag,
+            "Unknown media content_type '%s', defaulting to TEXT",
             first_type,
         )
         return MessageType.TEXT
@@ -1826,14 +1823,12 @@ class QQAdapter(BasePlatformAdapter):
             body["file_name"] = file_name
 
         # Retry transient upload failures
-        last_exc = None
         for attempt in range(3):
             try:
                 return await self._api_request(
                     "POST", path, body, timeout=FILE_UPLOAD_TIMEOUT
                 )
             except RuntimeError as exc:
-                last_exc = exc
                 err_msg = str(exc)
                 if any(
                         kw in err_msg
@@ -1842,8 +1837,8 @@ class QQAdapter(BasePlatformAdapter):
                     raise
                 if attempt < 2:
                     await asyncio.sleep(1.5 * (attempt + 1))
-
-        raise last_exc  # type: ignore[misc]
+                else:
+                    raise
 
     # Maximum time (seconds) to wait for reconnection before giving up on send.
     _RECONNECT_WAIT_SECONDS = 15.0

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"))
@@ -179,6 +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),
@@ -195,31 +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."""
@@ -400,6 +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")
@@ -518,6 +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
     # ------------------------------------------------------------------
@@ -633,7 +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)
 
@@ -684,7 +780,7 @@ 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)]
 
         fails = self._typing_failures.get(chat_id, 0)
         result = await self._rpc(
@@ -745,7 +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:
@@ -784,7 +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:

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:
@@ -1593,11 +1600,9 @@ class SlackAdapter(BasePlatformAdapter):
 
     async def _download_slack_file(self, url: str, ext: str, audio: bool = False, team_id: str = "") -> str:
         """Download a Slack file using the bot token for auth, with retry."""
-        import asyncio
         import httpx
 
         bot_token = self._team_clients[team_id].token if team_id and team_id in self._team_clients else self.config.token
-        last_exc = None
 
         async with httpx.AsyncClient(timeout=30.0, follow_redirects=True) as client:
             for attempt in range(3):
@@ -1627,7 +1632,6 @@ class SlackAdapter(BasePlatformAdapter):
                         from gateway.platforms.base import cache_image_from_bytes
                         return cache_image_from_bytes(response.content, ext)
                 except (httpx.TimeoutException, httpx.HTTPStatusError) as exc:
-                    last_exc = exc
                     if isinstance(exc, httpx.HTTPStatusError) and exc.response.status_code < 429:
                         raise
                     if attempt < 2:
@@ -1636,15 +1640,12 @@ class SlackAdapter(BasePlatformAdapter):
                         await asyncio.sleep(1.5 * (attempt + 1))
                         continue
                     raise
-        raise last_exc
 
     async def _download_slack_file_bytes(self, url: str, team_id: str = "") -> bytes:
         """Download a Slack file and return raw bytes, with retry."""
-        import asyncio
         import httpx
 
         bot_token = self._team_clients[team_id].token if team_id and team_id in self._team_clients else self.config.token
-        last_exc = None
 
         async with httpx.AsyncClient(timeout=30.0, follow_redirects=True) as client:
             for attempt in range(3):
@@ -1656,7 +1657,6 @@ class SlackAdapter(BasePlatformAdapter):
                     response.raise_for_status()
                     return response.content
                 except (httpx.TimeoutException, httpx.HTTPStatusError) as exc:
-                    last_exc = exc
                     if isinstance(exc, httpx.HTTPStatusError) and exc.response.status_code < 429:
                         raise
                     if attempt < 2:
@@ -1665,7 +1665,6 @@ class SlackAdapter(BasePlatformAdapter):
                         await asyncio.sleep(1.5 * (attempt + 1))
                         continue
                     raise
-        raise last_exc
 
     # ── Channel mention gating ─────────────────────────────────────────────
 

gateway/platforms/telegram.py

@@ -11,6 +11,7 @@ import asyncio
 import json
 import logging
 import os
+import tempfile
 import html as _html
 import re
 from typing import Dict, List, Optional, Any
@@ -70,8 +71,10 @@ from gateway.platforms.base import (
     SendResult,
     cache_image_from_bytes,
     cache_audio_from_bytes,
+    cache_video_from_bytes,
     cache_document_from_bytes,
     resolve_proxy_url,
+    SUPPORTED_VIDEO_TYPES,
     SUPPORTED_DOCUMENT_TYPES,
     utf16_len,
     _prefix_within_utf16_limit,
@@ -493,6 +496,13 @@ class TelegramAdapter(BasePlatformAdapter):
                     "[%s] DM topic '%s' already exists in chat %s (will be mapped from incoming messages)",
                     self.name, name, chat_id,
                 )
+            elif "not a forum" in error_text or "forums_disabled" in error_text:
+                logger.warning(
+                    "[%s] Cannot create DM topic '%s' in chat %s: Topics mode is not enabled. "
+                    "The user must open the DM with this bot in Telegram, tap the bot name "
+                    "at the top, and enable 'Topics' in chat settings before topics can be created.",
+                    self.name, name, chat_id,
+                )
             else:
                 logger.warning(
                     "[%s] Failed to create DM topic '%s' in chat %s: %s",
@@ -534,8 +544,23 @@ 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,
@@ -769,8 +794,28 @@ class TelegramAdapter(BasePlatformAdapter):
                 # Telegram pushes updates to our HTTP endpoint.  This
                 # enables cloud platforms (Fly.io, Railway) to auto-wake
                 # suspended machines on inbound HTTP traffic.
+                #
+                # SECURITY: TELEGRAM_WEBHOOK_SECRET is REQUIRED. Without it,
+                # python-telegram-bot passes secret_token=None and the
+                # webhook endpoint accepts any HTTP POST — attackers can
+                # inject forged updates as if from Telegram. Refuse to
+                # start rather than silently run in fail-open mode.
+                # See GHSA-3vpc-7q5r-276h.
                 webhook_port = int(os.getenv("TELEGRAM_WEBHOOK_PORT", "8443"))
-                webhook_secret = os.getenv("TELEGRAM_WEBHOOK_SECRET", "").strip() or None
+                webhook_secret = os.getenv("TELEGRAM_WEBHOOK_SECRET", "").strip()
+                if not webhook_secret:
+                    raise RuntimeError(
+                        "TELEGRAM_WEBHOOK_SECRET is required when "
+                        "TELEGRAM_WEBHOOK_URL is set. Without it, the "
+                        "webhook endpoint accepts forged updates from "
+                        "anyone who can reach it — see "
+                        "https://github.com/NousResearch/hermes-agent/"
+                        "security/advisories/GHSA-3vpc-7q5r-276h.\n\n"
+                        "Generate a secret and set it in your .env:\n"
+                        "  export TELEGRAM_WEBHOOK_SECRET=\"$(openssl rand -hex 32)\"\n\n"
+                        "Then register it with Telegram when setting the "
+                        "webhook via setWebhook's secret_token parameter."
+                    )
                 from urllib.parse import urlparse
                 webhook_path = urlparse(webhook_url).path or "/telegram"
 
@@ -1081,6 +1126,8 @@ 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:
@@ -1686,7 +1733,6 @@ class TelegramAdapter(BasePlatformAdapter):
             return SendResult(success=False, error="Not connected")
         
         try:
-            import os
             if not os.path.exists(audio_path):
                 return SendResult(success=False, error=self._missing_media_path_error("Audio", audio_path))
             
@@ -1735,7 +1781,6 @@ class TelegramAdapter(BasePlatformAdapter):
             return SendResult(success=False, error="Not connected")
 
         try:
-            import os
             if not os.path.exists(image_path):
                 return SendResult(success=False, error=self._missing_media_path_error("Image", image_path))
 
@@ -2048,7 +2093,7 @@ class TelegramAdapter(BasePlatformAdapter):
             url = m.group(2).replace('\\', '\\\\').replace(')', '\\)')
             return _ph(f'[{display}]({url})')
 
-        text = re.sub(r'\[([^\]]+)\]\(([^)]+)\)', _convert_link, text)
+        text = re.sub(r'\[([^\]]+)\]\(([^()]*(?:\([^()]*\)[^()]*)*)\)', _convert_link, text)
 
         # 4) Convert markdown headers (## Title) → bold *Title*
         def _convert_header(m):
@@ -2256,22 +2301,27 @@ 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)
@@ -2303,10 +2353,16 @@ class TelegramAdapter(BasePlatformAdapter):
         DMs remain unrestricted. Group/supergroup messages are accepted when:
         - the chat is explicitly allowlisted in ``free_response_chats``
         - ``require_mention`` is disabled
-        - the message is a command
         - the message replies to the bot
         - the bot is @mentioned
         - the text/caption matches a configured regex wake-word pattern
+
+        When ``require_mention`` is enabled, slash commands are not given
+        special treatment — they must pass the same mention/reply checks
+        as any other group message.  Users can still trigger commands via
+        the Telegram bot menu (``/command@botname``) or by explicitly
+        mentioning the bot (``@botname /command``), both of which are
+        recognised as mentions by :meth:`_message_mentions_bot`.
         """
         if not self._is_group_chat(message):
             return True
@@ -2321,8 +2377,6 @@ class TelegramAdapter(BasePlatformAdapter):
             return True
         if not self._telegram_require_mention():
             return True
-        if is_command:
-            return True
         if self._is_reply_to_bot(message):
             return True
         if self._message_mentions_bot(message):
@@ -2605,6 +2659,23 @@ class TelegramAdapter(BasePlatformAdapter):
             except Exception as e:
                 logger.warning("[Telegram] Failed to cache audio: %s", e, exc_info=True)
 
+        elif msg.video:
+            try:
+                file_obj = await msg.video.get_file()
+                video_bytes = await file_obj.download_as_bytearray()
+                ext = ".mp4"
+                if getattr(file_obj, "file_path", None):
+                    for candidate in SUPPORTED_VIDEO_TYPES:
+                        if file_obj.file_path.lower().endswith(candidate):
+                            ext = candidate
+                            break
+                cached_path = cache_video_from_bytes(bytes(video_bytes), ext=ext)
+                event.media_urls = [cached_path]
+                event.media_types = [SUPPORTED_VIDEO_TYPES.get(ext, "video/mp4")]
+                logger.info("[Telegram] Cached user video at %s", cached_path)
+            except Exception as e:
+                logger.warning("[Telegram] Failed to cache video: %s", e, exc_info=True)
+
         # Download document files to cache for agent processing
         elif msg.document:
             doc = msg.document
@@ -2621,6 +2692,21 @@ class TelegramAdapter(BasePlatformAdapter):
                     mime_to_ext = {v: k for k, v in SUPPORTED_DOCUMENT_TYPES.items()}
                     ext = mime_to_ext.get(doc.mime_type, "")
 
+                if not ext and doc.mime_type:
+                    video_mime_to_ext = {v: k for k, v in SUPPORTED_VIDEO_TYPES.items()}
+                    ext = video_mime_to_ext.get(doc.mime_type, "")
+
+                if ext in SUPPORTED_VIDEO_TYPES:
+                    file_obj = await doc.get_file()
+                    video_bytes = await file_obj.download_as_bytearray()
+                    cached_path = cache_video_from_bytes(bytes(video_bytes), ext=ext)
+                    event.media_urls = [cached_path]
+                    event.media_types = [SUPPORTED_VIDEO_TYPES[ext]]
+                    event.message_type = MessageType.VIDEO
+                    logger.info("[Telegram] Cached user video document at %s", cached_path)
+                    await self.handle_message(event)
+                    return
+
                 # Check if supported
                 if ext not in SUPPORTED_DOCUMENT_TYPES:
                     supported_list = ", ".join(sorted(SUPPORTED_DOCUMENT_TYPES.keys()))
@@ -2759,13 +2845,11 @@ class TelegramAdapter(BasePlatformAdapter):
             logger.info("[Telegram] Analyzing sticker at %s", cached_path)
 
             from tools.vision_tools import vision_analyze_tool
-            import json as _json
-
             result_json = await vision_analyze_tool(
                 image_url=cached_path,
                 user_prompt=STICKER_VISION_PROMPT,
             )
-            result = _json.loads(result_json)
+            result = json.loads(result_json)
 
             if result.get("success"):
                 description = result.get("analysis", "a sticker")

gateway/platforms/webhook.py

@@ -313,24 +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):
@@ -341,6 +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)

gateway/platforms/wecom.py

@@ -624,13 +624,16 @@ class WeComAdapter(BasePlatformAdapter):
         msgtype = str(body.get("msgtype") or "").lower()
 
         if msgtype == "mixed":
-            mixed = body.get("mixed") if isinstance(body.get("mixed"), dict) else {}
-            items = mixed.get("msg_item") if isinstance(mixed.get("msg_item"), list) else []
+            _raw_mixed = body.get("mixed")
+            mixed = _raw_mixed if isinstance(_raw_mixed, dict) else {}
+            _raw_items = mixed.get("msg_item")
+            items = _raw_items if isinstance(_raw_items, list) else []
             for item in items:
                 if not isinstance(item, dict):
                     continue
                 if str(item.get("msgtype") or "").lower() == "text":
-                    text_block = item.get("text") if isinstance(item.get("text"), dict) else {}
+                    _raw_text = item.get("text")
+                    text_block = _raw_text if isinstance(_raw_text, dict) else {}
                     content = str(text_block.get("content") or "").strip()
                     if content:
                         text_parts.append(content)
@@ -672,8 +675,10 @@ class WeComAdapter(BasePlatformAdapter):
         msgtype = str(body.get("msgtype") or "").lower()
 
         if msgtype == "mixed":
-            mixed = body.get("mixed") if isinstance(body.get("mixed"), dict) else {}
-            items = mixed.get("msg_item") if isinstance(mixed.get("msg_item"), list) else []
+            _raw_mixed = body.get("mixed")
+            mixed = _raw_mixed if isinstance(_raw_mixed, dict) else {}
+            _raw_items = mixed.get("msg_item")
+            items = _raw_items if isinstance(_raw_items, list) else []
             for item in items:
                 if not isinstance(item, dict):
                     continue

gateway/platforms/whatsapp.py

@@ -66,6 +66,37 @@ def _kill_port_process(port: int) -> None:
     except Exception:
         pass
 
+
+def _terminate_bridge_process(proc, *, force: bool = False) -> None:
+    """Terminate the bridge process using process-tree semantics where possible."""
+    if _IS_WINDOWS:
+        cmd = ["taskkill", "/PID", str(proc.pid), "/T"]
+        if force:
+            cmd.append("/F")
+        try:
+            result = subprocess.run(
+                cmd,
+                capture_output=True,
+                text=True,
+                timeout=10,
+            )
+        except FileNotFoundError:
+            if force:
+                proc.kill()
+            else:
+                proc.terminate()
+            return
+
+        if result.returncode != 0:
+            details = (result.stderr or result.stdout or "").strip()
+            raise OSError(details or f"taskkill failed for PID {proc.pid}")
+        return
+
+    import signal
+
+    sig = signal.SIGTERM if not force else signal.SIGKILL
+    os.killpg(os.getpgid(proc.pid), sig)
+
 import sys
 sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
 
@@ -118,6 +149,10 @@ class WhatsAppAdapter(BasePlatformAdapter):
     - bridge_script: Path to the Node.js bridge script
     - bridge_port: Port for HTTP communication (default: 3000)
     - session_path: Path to store WhatsApp session data
+    - dm_policy: "open" | "allowlist" | "disabled" — how DMs are handled (default: "open")
+    - allow_from: List of sender IDs allowed in DMs (when dm_policy="allowlist")
+    - group_policy: "open" | "allowlist" | "disabled" — which groups are processed (default: "open")
+    - group_allow_from: List of group JIDs allowed (when group_policy="allowlist")
     """
     
     # WhatsApp message limits — practical UX limit, not protocol max.
@@ -140,6 +175,10 @@ class WhatsAppAdapter(BasePlatformAdapter):
             get_hermes_dir("platforms/whatsapp/session", "whatsapp/session")
         ))
         self._reply_prefix: Optional[str] = config.extra.get("reply_prefix")
+        self._dm_policy = str(config.extra.get("dm_policy") or os.getenv("WHATSAPP_DM_POLICY", "open")).strip().lower()
+        self._allow_from = self._coerce_allow_list(config.extra.get("allow_from") or config.extra.get("allowFrom"))
+        self._group_policy = str(config.extra.get("group_policy") or os.getenv("WHATSAPP_GROUP_POLICY", "open")).strip().lower()
+        self._group_allow_from = self._coerce_allow_list(config.extra.get("group_allow_from") or config.extra.get("groupAllowFrom"))
         self._mention_patterns = self._compile_mention_patterns()
         self._message_queue: asyncio.Queue = asyncio.Queue()
         self._bridge_log_fh = None
@@ -163,6 +202,33 @@ class WhatsAppAdapter(BasePlatformAdapter):
             return {str(part).strip() for part in raw if str(part).strip()}
         return {part.strip() for part in str(raw).split(",") if part.strip()}
 
+    @staticmethod
+    def _coerce_allow_list(raw) -> set[str]:
+        """Parse allow_from / group_allow_from from config or env var."""
+        if raw is None:
+            return set()
+        if isinstance(raw, list):
+            return {str(part).strip() for part in raw if str(part).strip()}
+        return {part.strip() for part in str(raw).split(",") if part.strip()}
+
+    def _is_dm_allowed(self, sender_id: str) -> bool:
+        """Check whether a DM from the given sender should be processed."""
+        if self._dm_policy == "disabled":
+            return False
+        if self._dm_policy == "allowlist":
+            return sender_id in self._allow_from
+        # "open" — all DMs allowed
+        return True
+
+    def _is_group_allowed(self, chat_id: str) -> bool:
+        """Check whether a group chat should be processed."""
+        if self._group_policy == "disabled":
+            return False
+        if self._group_policy == "allowlist":
+            return chat_id in self._group_allow_from
+        # "open" — all groups allowed
+        return True
+
     def _compile_mention_patterns(self):
         patterns = self.config.extra.get("mention_patterns")
         if patterns is None:
@@ -255,8 +321,18 @@ class WhatsAppAdapter(BasePlatformAdapter):
         return cleaned.strip() or text
 
     def _should_process_message(self, data: Dict[str, Any]) -> bool:
-        if not data.get("isGroup"):
+        is_group = data.get("isGroup", False)
+        if is_group:
+            chat_id = str(data.get("chatId") or "")
+            if not self._is_group_allowed(chat_id):
+                return False
+        else:
+            sender_id = str(data.get("senderId") or data.get("from") or "")
+            if not self._is_dm_allowed(sender_id):
+                return False
+            # DMs that pass the policy gate are always processed
             return True
+        # Group messages: check mention / free-response settings
         chat_id = str(data.get("chatId") or "")
         if chat_id in self._whatsapp_free_response_chats():
             return True
@@ -289,39 +365,40 @@ 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)
             
             # Check if bridge is already running and connected
             import aiohttp
-            import asyncio
             try:
                 async with aiohttp.ClientSession() as session:
                     async with session.get(
@@ -452,10 +529,13 @@ 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."""
@@ -487,22 +567,14 @@ class WhatsAppAdapter(BasePlatformAdapter):
         """Stop the WhatsApp bridge and clean up any orphaned processes."""
         if self._bridge_process:
             try:
-                # Kill the entire process group so child node processes die too
-                import signal
                 try:
-                    if _IS_WINDOWS:
-                        self._bridge_process.terminate()
-                    else:
-                        os.killpg(os.getpgid(self._bridge_process.pid), signal.SIGTERM)
+                    _terminate_bridge_process(self._bridge_process, force=False)
                 except (ProcessLookupError, PermissionError):
                     self._bridge_process.terminate()
                 await asyncio.sleep(1)
                 if self._bridge_process.poll() is None:
                     try:
-                        if _IS_WINDOWS:
-                            self._bridge_process.kill()
-                        else:
-                            os.killpg(os.getpgid(self._bridge_process.pid), signal.SIGKILL)
+                        _terminate_bridge_process(self._bridge_process, force=True)
                     except (ProcessLookupError, PermissionError):
                         self._bridge_process.kill()
             except Exception as e:
@@ -655,6 +727,8 @@ 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:
@@ -766,6 +840,17 @@ class WhatsAppAdapter(BasePlatformAdapter):
         """Send a video natively via bridge — plays inline in WhatsApp."""
         return await self._send_media_to_bridge(chat_id, video_path, "video", caption)
 
+    async def send_voice(
+        self,
+        chat_id: str,
+        audio_path: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        **kwargs,
+    ) -> SendResult:
+        """Send an audio file as a WhatsApp voice message via bridge."""
+        return await self._send_media_to_bridge(chat_id, audio_path, "audio", caption)
+
     async def send_document(
         self,
         chat_id: str,

gateway/run.py

@@ -30,6 +30,8 @@ from pathlib import Path
 from datetime import datetime
 from typing import Dict, Optional, Any, List
 
+from agent.account_usage import fetch_account_usage, render_account_usage_lines
+
 # --- Agent cache tuning ---------------------------------------------------
 # Bounds the per-session AIAgent cache to prevent unbounded growth in
 # long-lived gateways (each AIAgent holds LLM clients, tool schemas,
@@ -86,7 +88,7 @@ sys.path.insert(0, str(Path(__file__).parent.parent))
 
 # Resolve Hermes home directory (respects HERMES_HOME override)
 from hermes_constants import get_hermes_home
-from utils import atomic_yaml_write, is_truthy_value
+from utils import atomic_yaml_write, base_url_host_matches, is_truthy_value
 _hermes_home = get_hermes_home()
 
 # Load environment variables from ~/.hermes/.env first.
@@ -279,6 +281,7 @@ from gateway.session import (
     build_session_context,
     build_session_context_prompt,
     build_session_key,
+    is_shared_multi_user_session,
 )
 from gateway.delivery import DeliveryRouter
 from gateway.platforms.base import (
@@ -629,7 +632,6 @@ class GatewayRunner:
         self._restart_drain_timeout = self._load_restart_drain_timeout()
         self._provider_routing = self._load_provider_routing()
         self._fallback_model = self._load_fallback_model()
-        self._smart_model_routing = self._load_smart_model_routing()
 
         # Wire process registry into session store for reset protection
         from tools.process_registry import process_registry
@@ -787,6 +789,10 @@ class GatewayRunner:
 
     _VOICE_MODE_PATH = _hermes_home / "gateway_voice_mode.json"
 
+    def _voice_key(self, platform: Platform, chat_id: str) -> str:
+        """Return a platform-namespaced key for voice mode state."""
+        return f"{platform.value}:{chat_id}"
+
     def _load_voice_modes(self) -> Dict[str, str]:
         try:
             data = json.loads(self._VOICE_MODE_PATH.read_text())
@@ -797,11 +803,21 @@ class GatewayRunner:
             return {}
 
         valid_modes = {"off", "voice_only", "all"}
-        return {
-            str(chat_id): mode
-            for chat_id, mode in data.items()
-            if mode in valid_modes
-        }
+        result = {}
+        for chat_id, mode in data.items():
+            if mode not in valid_modes:
+                continue
+            key = str(chat_id)
+            # Skip legacy unprefixed keys (warn and skip)
+            if ":" not in key:
+                logger.warning(
+                    "Skipping legacy unprefixed voice mode key %r during migration. "
+                    "Re-enable voice mode on that chat to rebuild the prefixed key.",
+                    key,
+                )
+                continue
+            result[key] = mode
+        return result
 
     def _save_voice_modes(self) -> None:
         try:
@@ -827,9 +843,14 @@ class GatewayRunner:
         disabled_chats = getattr(adapter, "_auto_tts_disabled_chats", None)
         if not isinstance(disabled_chats, set):
             return
+        platform = getattr(adapter, "platform", None)
+        if not isinstance(platform, Platform):
+            return
         disabled_chats.clear()
+        prefix = f"{platform.value}:"
         disabled_chats.update(
-            chat_id for chat_id, mode in self._voice_mode.items() if mode == "off"
+            key[len(prefix):] for key, mode in self._voice_mode.items()
+            if mode == "off" and key.startswith(prefix)
         )
 
     async def _safe_adapter_disconnect(self, adapter, platform) -> None:
@@ -1082,11 +1103,16 @@ class GatewayRunner:
         return model, runtime_kwargs
 
     def _resolve_turn_agent_config(self, user_message: str, model: str, runtime_kwargs: dict) -> dict:
-        from agent.smart_model_routing import resolve_turn_route
+        """Build the effective model/runtime config for a single turn.
+
+        Always uses the session's primary model/provider.  If `/fast` is
+        enabled and the model supports Priority Processing / Anthropic fast
+        mode, attach `request_overrides` so the API call is marked
+        accordingly.
+        """
         from hermes_cli.models import resolve_fast_mode_overrides
 
-        primary = {
-            "model": model,
+        runtime = {
             "api_key": runtime_kwargs.get("api_key"),
             "base_url": runtime_kwargs.get("base_url"),
             "provider": runtime_kwargs.get("provider"),
@@ -1095,7 +1121,18 @@ class GatewayRunner:
             "args": list(runtime_kwargs.get("args") or []),
             "credential_pool": runtime_kwargs.get("credential_pool"),
         }
-        route = resolve_turn_route(user_message, getattr(self, "_smart_model_routing", {}), primary)
+        route = {
+            "model": model,
+            "runtime": runtime,
+            "signature": (
+                model,
+                runtime["provider"],
+                runtime["base_url"],
+                runtime["api_mode"],
+                runtime["command"],
+                tuple(runtime["args"]),
+            ),
+        }
 
         service_tier = getattr(self, "_service_tier", None)
         if not service_tier:
@@ -1103,7 +1140,7 @@ class GatewayRunner:
             return route
 
         try:
-            overrides = resolve_fast_mode_overrides(route.get("model"))
+            overrides = resolve_fast_mode_overrides(route["model"])
         except Exception:
             overrides = None
         route["request_overrides"] = overrides
@@ -1232,7 +1269,6 @@ class GatewayRunner:
         the prefill_messages_file key in ~/.hermes/config.yaml.
         Relative paths are resolved from ~/.hermes/.
         """
-        import json as _json
         file_path = os.getenv("HERMES_PREFILL_MESSAGES_FILE", "")
         if not file_path:
             try:
@@ -1254,7 +1290,7 @@ class GatewayRunner:
             return []
         try:
             with open(path, "r", encoding="utf-8") as f:
-                data = _json.load(f)
+                data = json.load(f)
             if not isinstance(data, list):
                 logger.warning("Prefill messages file must contain a JSON array: %s", path)
                 return []
@@ -1461,20 +1497,6 @@ class GatewayRunner:
             pass
         return None
 
-    @staticmethod
-    def _load_smart_model_routing() -> dict:
-        """Load optional smart cheap-vs-strong model routing config."""
-        try:
-            import yaml as _y
-            cfg_path = _hermes_home / "config.yaml"
-            if cfg_path.exists():
-                with open(cfg_path, encoding="utf-8") as _f:
-                    cfg = _y.safe_load(_f) or {}
-                return cfg.get("smart_model_routing", {}) or {}
-        except Exception:
-            pass
-        return {}
-
     def _snapshot_running_agents(self) -> Dict[str, Any]:
         return {
             session_key: agent
@@ -1647,12 +1669,32 @@ class GatewayRunner:
 
         notified: set = set()
         for session_key in active:
-            # Parse platform + chat_id from the session key.
-            _parsed = _parse_session_key(session_key)
-            if not _parsed:
-                continue
-            platform_str = _parsed["platform"]
-            chat_id = _parsed["chat_id"]
+            source = None
+            try:
+                if getattr(self, "session_store", None) is not None:
+                    self.session_store._ensure_loaded()
+                    entry = self.session_store._entries.get(session_key)
+                    source = getattr(entry, "origin", None) if entry else None
+            except Exception as e:
+                logger.debug(
+                    "Failed to load session origin for shutdown notification %s: %s",
+                    session_key,
+                    e,
+                )
+
+            if source is not None:
+                platform_str = source.platform.value
+                chat_id = source.chat_id
+                thread_id = source.thread_id
+            else:
+                # Fall back to parsing the session key when no persisted
+                # origin is available (legacy sessions/tests).
+                _parsed = _parse_session_key(session_key)
+                if not _parsed:
+                    continue
+                platform_str = _parsed["platform"]
+                chat_id = _parsed["chat_id"]
+                thread_id = _parsed.get("thread_id")
 
             # Deduplicate: one notification per chat, even if multiple
             # sessions (different users/threads) share the same chat.
@@ -1668,7 +1710,6 @@ class GatewayRunner:
 
                 # Include thread_id if present so the message lands in the
                 # correct forum topic / thread.
-                thread_id = _parsed.get("thread_id")
                 metadata = {"thread_id": thread_id} if thread_id else None
 
                 await adapter.send(chat_id, msg, metadata=metadata)
@@ -1921,6 +1962,39 @@ class GatewayRunner:
                 "or configure platform allowlists (e.g., TELEGRAM_ALLOWED_USERS=your_id)."
             )
         
+        # Discover Python plugins before shell hooks so plugin block
+        # decisions take precedence in tie cases.  The CLI startup path
+        # does this via an explicit call in hermes_cli/main.py; the
+        # gateway lazily imports run_agent inside per-request handlers,
+        # so the discover_plugins() side-effect in model_tools.py is NOT
+        # guaranteed to have run by the time we reach this point.
+        try:
+            from hermes_cli.plugins import discover_plugins
+            discover_plugins()
+        except Exception:
+            logger.debug(
+                "plugin discovery failed at gateway startup", exc_info=True,
+            )
+
+        # Register declarative shell hooks from cli-config.yaml.  Gateway
+        # has no TTY, so consent has to come from one of the three opt-in
+        # channels (--accept-hooks on launch, HERMES_ACCEPT_HOOKS env var,
+        # or hooks_auto_accept: true in config.yaml).  We pass
+        # accept_hooks=False here and let register_from_config resolve
+        # the effective value from env + config itself — the CLI-side
+        # registration already honored --accept-hooks, and re-reading
+        # hooks_auto_accept here would just duplicate that lookup.
+        # Failures are logged but must never block gateway startup.
+        try:
+            from hermes_cli.config import load_config
+            from agent.shell_hooks import register_from_config
+            register_from_config(load_config(), accept_hooks=False)
+        except Exception:
+            logger.debug(
+                "shell-hook registration failed at gateway startup",
+                exc_info=True,
+            )
+
         # Discover and load event hooks
         self.hooks.discover_and_load()
         
@@ -2942,10 +3016,59 @@ class GatewayRunner:
         return bool(check_ids & allowed_ids)
 
     def _get_unauthorized_dm_behavior(self, platform: Optional[Platform]) -> str:
-        """Return how unauthorized DMs should be handled for a platform."""
+        """Return how unauthorized DMs should be handled for a platform.
+
+        Resolution order:
+        1. Explicit per-platform ``unauthorized_dm_behavior`` in config — always wins.
+        2. Explicit global ``unauthorized_dm_behavior`` in config — wins when no per-platform.
+        3. When an allowlist (``PLATFORM_ALLOWED_USERS`` or ``GATEWAY_ALLOWED_USERS``) is
+           configured, default to ``"ignore"`` — the allowlist signals that the owner has
+           deliberately restricted access; spamming unknown contacts with pairing codes
+           is both noisy and a potential info-leak. (#9337)
+        4. No allowlist and no explicit config → ``"pair"`` (open-gateway default).
+        """
         config = getattr(self, "config", None)
-        if config and hasattr(config, "get_unauthorized_dm_behavior"):
-            return config.get_unauthorized_dm_behavior(platform)
+
+        # Check for an explicit per-platform override first.
+        if config and hasattr(config, "get_unauthorized_dm_behavior") and platform:
+            platform_cfg = config.platforms.get(platform) if hasattr(config, "platforms") else None
+            if platform_cfg and "unauthorized_dm_behavior" in getattr(platform_cfg, "extra", {}):
+                # Operator explicitly configured behavior for this platform — respect it.
+                return config.get_unauthorized_dm_behavior(platform)
+
+        # Check for an explicit global config override.
+        if config and hasattr(config, "unauthorized_dm_behavior"):
+            if config.unauthorized_dm_behavior != "pair":  # non-default → explicit override
+                return config.unauthorized_dm_behavior
+
+        # No explicit override.  Fall back to allowlist-aware default:
+        # if any allowlist is configured for this platform, silently drop
+        # unauthorized messages instead of sending pairing codes.
+        if platform:
+            platform_env_map = {
+                Platform.TELEGRAM: "TELEGRAM_ALLOWED_USERS",
+                Platform.DISCORD:  "DISCORD_ALLOWED_USERS",
+                Platform.WHATSAPP: "WHATSAPP_ALLOWED_USERS",
+                Platform.SLACK:    "SLACK_ALLOWED_USERS",
+                Platform.SIGNAL:   "SIGNAL_ALLOWED_USERS",
+                Platform.EMAIL:    "EMAIL_ALLOWED_USERS",
+                Platform.SMS:      "SMS_ALLOWED_USERS",
+                Platform.MATTERMOST: "MATTERMOST_ALLOWED_USERS",
+                Platform.MATRIX:   "MATRIX_ALLOWED_USERS",
+                Platform.DINGTALK: "DINGTALK_ALLOWED_USERS",
+                Platform.FEISHU:   "FEISHU_ALLOWED_USERS",
+                Platform.WECOM:    "WECOM_ALLOWED_USERS",
+                Platform.WECOM_CALLBACK: "WECOM_CALLBACK_ALLOWED_USERS",
+                Platform.WEIXIN:   "WEIXIN_ALLOWED_USERS",
+                Platform.BLUEBUBBLES: "BLUEBUBBLES_ALLOWED_USERS",
+                Platform.QQBOT:    "QQ_ALLOWED_USERS",
+            }
+            if os.getenv(platform_env_map.get(platform, ""), "").strip():
+                return "ignore"
+
+        if os.getenv("GATEWAY_ALLOWED_USERS", "").strip():
+            return "ignore"
+
         return "pair"
     
     async def _handle_message(self, event: MessageEvent) -> Optional[str]:
@@ -3154,10 +3277,9 @@ class GatewayRunner:
                     return "Usage: /queue <prompt>"
                 adapter = self.adapters.get(source.platform)
                 if adapter:
-                    from gateway.platforms.base import MessageEvent as _ME, MessageType as _MT
-                    queued_event = _ME(
+                    queued_event = MessageEvent(
                         text=queued_text,
-                        message_type=_MT.TEXT,
+                        message_type=MessageType.TEXT,
                         source=event.source,
                         message_id=event.message_id,
                         channel_prompt=event.channel_prompt,
@@ -3179,10 +3301,9 @@ class GatewayRunner:
                     # Agent hasn't started yet — queue as turn-boundary fallback.
                     adapter = self.adapters.get(source.platform)
                     if adapter:
-                        from gateway.platforms.base import MessageEvent as _ME, MessageType as _MT
-                        queued_event = _ME(
+                        queued_event = MessageEvent(
                             text=steer_text,
-                            message_type=_MT.TEXT,
+                            message_type=MessageType.TEXT,
                             source=event.source,
                             message_id=event.message_id,
                             channel_prompt=event.channel_prompt,
@@ -3202,10 +3323,9 @@ class GatewayRunner:
                 # Running agent is missing or lacks steer() — fall back to queue.
                 adapter = self.adapters.get(source.platform)
                 if adapter:
-                    from gateway.platforms.base import MessageEvent as _ME, MessageType as _MT
-                    queued_event = _ME(
+                    queued_event = MessageEvent(
                         text=steer_text,
-                        message_type=_MT.TEXT,
+                        message_type=MessageType.TEXT,
                         source=event.source,
                         message_id=event.message_id,
                         channel_prompt=event.channel_prompt,
@@ -3235,6 +3355,20 @@ class GatewayRunner:
             if _cmd_def_inner and _cmd_def_inner.name == "background":
                 return await self._handle_background_command(event)
 
+            # Session-level toggles that are safe to run mid-agent —
+            # /yolo can unblock a pending approval prompt, /verbose cycles
+            # the tool-progress display mode for the ongoing stream.
+            # Both modify session state without needing agent interaction
+            # and must not be queued (the safety net would discard them).
+            # /fast and /reasoning are config-only and take effect next
+            # message, so they fall through to the catch-all busy response
+            # below — users should wait and set them between turns.
+            if _cmd_def_inner and _cmd_def_inner.name in ("yolo", "verbose"):
+                if _cmd_def_inner.name == "yolo":
+                    return await self._handle_yolo_command(event)
+                if _cmd_def_inner.name == "verbose":
+                    return await self._handle_verbose_command(event)
+
             # Gateway-handled info/control commands with dedicated
             # running-agent handlers.
             if _cmd_def_inner and _cmd_def_inner.name in _DEDICATED_HANDLERS:
@@ -3540,9 +3674,8 @@ class GatewayRunner:
                 plugin_handler = get_plugin_command_handler(command.replace("_", "-"))
                 if plugin_handler:
                     user_args = event.get_command_args().strip()
-                    import asyncio as _aio
                     result = plugin_handler(user_args)
-                    if _aio.iscoroutine(result):
+                    if asyncio.iscoroutine(result):
                         result = await result
                     return str(result) if result else None
             except Exception as e:
@@ -3659,12 +3792,12 @@ class GatewayRunner:
         history = history or []
         message_text = event.text or ""
 
-        _is_shared_thread = (
-            source.chat_type != "dm"
-            and source.thread_id
-            and not getattr(self.config, "thread_sessions_per_user", False)
+        _is_shared_multi_user = is_shared_multi_user_session(
+            source,
+            group_sessions_per_user=getattr(self.config, "group_sessions_per_user", True),
+            thread_sessions_per_user=getattr(self.config, "thread_sessions_per_user", False),
         )
-        if _is_shared_thread and source.user_name:
+        if _is_shared_multi_user and source.user_name:
             message_text = f"[{source.user_name}] {message_text}"
 
         if event.media_urls:
@@ -3724,9 +3857,7 @@ class GatewayRunner:
             for i, path in enumerate(event.media_urls):
                 mtype = event.media_types[i] if i < len(event.media_types) else ""
                 if mtype in ("", "application/octet-stream"):
-                    import os as _os2
-
-                    _ext = _os2.path.splitext(path)[1].lower()
+                    _ext = os.path.splitext(path)[1].lower()
                     if _ext in _TEXT_EXTENSIONS:
                         mtype = "text/plain"
                     else:
@@ -3736,13 +3867,10 @@ class GatewayRunner:
                 if not mtype.startswith(("application/", "text/")):
                     continue
 
-                import os as _os
-                import re as _re
-
-                basename = _os.path.basename(path)
+                basename = os.path.basename(path)
                 parts = basename.split("_", 2)
                 display_name = parts[2] if len(parts) >= 3 else basename
-                display_name = _re.sub(r'[^\w.\- ]', '_', display_name)
+                display_name = re.sub(r'[^\w.\- ]', '_', display_name)
 
                 if mtype.startswith("text/"):
                     context_note = (
@@ -3759,14 +3887,14 @@ class GatewayRunner:
                 message_text = f"{context_note}\n\n{message_text}"
 
         if getattr(event, "reply_to_text", None) and event.reply_to_message_id:
+            # Always inject the reply-to pointer — even when the quoted text
+            # already appears in history. The prefix isn't deduplication, it's
+            # disambiguation: it tells the agent *which* prior message the user
+            # is referencing. History can contain the same or similar text
+            # multiple times, and without an explicit pointer the agent has to
+            # guess (or answer for both subjects). Token overhead is minimal.
             reply_snippet = event.reply_to_text[:500]
-            found_in_history = any(
-                reply_snippet[:200] in (msg.get("content") or "")
-                for msg in history
-                if msg.get("role") in ("assistant", "user", "tool")
-            )
-            if not found_in_history:
-                message_text = f'[Replying to: "{reply_snippet}"]\n\n{message_text}'
+            message_text = f'[Replying to: "{reply_snippet}"]\n\n{message_text}'
 
         if "@" in message_text:
             try:
@@ -3774,9 +3902,11 @@ class GatewayRunner:
                 from agent.model_metadata import get_model_context_length
 
                 _msg_cwd = os.environ.get("TERMINAL_CWD", os.path.expanduser("~"))
+                _msg_runtime = _resolve_runtime_agent_kwargs()
                 _msg_ctx_len = get_model_context_length(
                     self._model,
-                    base_url=self._base_url or "",
+                    base_url=self._base_url or _msg_runtime.get("base_url") or "",
+                    api_key=_msg_runtime.get("api_key") or "",
                 )
                 _ctx_result = await preprocess_context_references_async(
                     message_text,
@@ -5038,7 +5168,6 @@ class GatewayRunner:
         # Save the requester's routing info so the new gateway process can
         # notify them once it comes back online.
         try:
-            import json as _json
             notify_data = {
                 "platform": event.source.platform.value if event.source.platform else None,
                 "chat_id": event.source.chat_id,
@@ -5046,7 +5175,7 @@ class GatewayRunner:
             if event.source.thread_id:
                 notify_data["thread_id"] = event.source.thread_id
             (_hermes_home / ".restart_notify.json").write_text(
-                _json.dumps(notify_data)
+                json.dumps(notify_data)
             )
         except Exception as e:
             logger.debug("Failed to write restart notify file: %s", e)
@@ -5057,16 +5186,14 @@ class GatewayRunner:
         # marker persists so the new gateway can still detect a delayed
         # /restart redelivery from Telegram.  Overwritten on every /restart.
         try:
-            import json as _json
-            import time as _time
             dedup_data = {
                 "platform": event.source.platform.value if event.source.platform else None,
-                "requested_at": _time.time(),
+                "requested_at": time.time(),
             }
             if event.platform_update_id is not None:
                 dedup_data["update_id"] = event.platform_update_id
             (_hermes_home / ".restart_last_processed.json").write_text(
-                _json.dumps(dedup_data)
+                json.dumps(dedup_data)
             )
         except Exception as e:
             logger.debug("Failed to write restart dedup marker: %s", e)
@@ -5114,12 +5241,10 @@ class GatewayRunner:
             return False
 
         try:
-            import json as _json
-            import time as _time
             marker_path = _hermes_home / ".restart_last_processed.json"
             if not marker_path.exists():
                 return False
-            data = _json.loads(marker_path.read_text())
+            data = json.loads(marker_path.read_text())
         except Exception:
             return False
 
@@ -5133,7 +5258,7 @@ class GatewayRunner:
         # swallow a fresh /restart from the user.
         requested_at = data.get("requested_at")
         if isinstance(requested_at, (int, float)):
-            if _time.time() - requested_at > 300:
+            if time.time() - requested_at > 300:
                 return False
         return event.platform_update_id <= recorded_uid
 
@@ -5524,7 +5649,7 @@ class GatewayRunner:
 
         # Cache notice
         cache_enabled = (
-            ("openrouter" in (result.base_url or "").lower() and "claude" in result.new_model.lower())
+            (base_url_host_matches(result.base_url or "", "openrouter.ai") and "claude" in result.new_model.lower())
             or result.api_mode == "anthropic_messages"
         )
         if cache_enabled:
@@ -5780,11 +5905,13 @@ class GatewayRunner:
         """Handle /voice [on|off|tts|channel|leave|status] command."""
         args = event.get_command_args().strip().lower()
         chat_id = event.source.chat_id
+        platform = event.source.platform
+        voice_key = self._voice_key(platform, chat_id)
 
-        adapter = self.adapters.get(event.source.platform)
+        adapter = self.adapters.get(platform)
 
         if args in ("on", "enable"):
-            self._voice_mode[chat_id] = "voice_only"
+            self._voice_mode[voice_key] = "voice_only"
             self._save_voice_modes()
             if adapter:
                 self._set_adapter_auto_tts_disabled(adapter, chat_id, disabled=False)
@@ -5794,13 +5921,13 @@ class GatewayRunner:
                 "Use /voice tts to get voice replies for all messages."
             )
         elif args in ("off", "disable"):
-            self._voice_mode[chat_id] = "off"
+            self._voice_mode[voice_key] = "off"
             self._save_voice_modes()
             if adapter:
                 self._set_adapter_auto_tts_disabled(adapter, chat_id, disabled=True)
             return "Voice mode disabled. Text-only replies."
         elif args == "tts":
-            self._voice_mode[chat_id] = "all"
+            self._voice_mode[voice_key] = "all"
             self._save_voice_modes()
             if adapter:
                 self._set_adapter_auto_tts_disabled(adapter, chat_id, disabled=False)
@@ -5813,7 +5940,7 @@ class GatewayRunner:
         elif args == "leave":
             return await self._handle_voice_channel_leave(event)
         elif args == "status":
-            mode = self._voice_mode.get(chat_id, "off")
+            mode = self._voice_mode.get(voice_key, "off")
             labels = {
                 "off": "Off (text only)",
                 "voice_only": "On (voice reply to voice messages)",
@@ -5837,15 +5964,15 @@ class GatewayRunner:
             return f"Voice mode: {labels.get(mode, mode)}"
         else:
             # Toggle: off → on, on/all → off
-            current = self._voice_mode.get(chat_id, "off")
+            current = self._voice_mode.get(voice_key, "off")
             if current == "off":
-                self._voice_mode[chat_id] = "voice_only"
+                self._voice_mode[voice_key] = "voice_only"
                 self._save_voice_modes()
                 if adapter:
                     self._set_adapter_auto_tts_disabled(adapter, chat_id, disabled=False)
                 return "Voice mode enabled."
             else:
-                self._voice_mode[chat_id] = "off"
+                self._voice_mode[voice_key] = "off"
                 self._save_voice_modes()
                 if adapter:
                     self._set_adapter_auto_tts_disabled(adapter, chat_id, disabled=True)
@@ -5891,7 +6018,7 @@ class GatewayRunner:
             adapter._voice_text_channels[guild_id] = int(event.source.chat_id)
             if hasattr(adapter, "_voice_sources"):
                 adapter._voice_sources[guild_id] = event.source.to_dict()
-            self._voice_mode[event.source.chat_id] = "all"
+            self._voice_mode[self._voice_key(event.source.platform, event.source.chat_id)] = "all"
             self._save_voice_modes()
             self._set_adapter_auto_tts_disabled(adapter, event.source.chat_id, disabled=False)
             return (
@@ -5918,7 +6045,7 @@ class GatewayRunner:
         except Exception as e:
             logger.warning("Error leaving voice channel: %s", e)
         # Always clean up state even if leave raised an exception
-        self._voice_mode[event.source.chat_id] = "off"
+        self._voice_mode[self._voice_key(event.source.platform, event.source.chat_id)] = "off"
         self._save_voice_modes()
         self._set_adapter_auto_tts_disabled(adapter, event.source.chat_id, disabled=True)
         if hasattr(adapter, "_voice_input_callback"):
@@ -5930,7 +6057,7 @@ class GatewayRunner:
 
         Cleans up runner-side voice_mode state that the adapter cannot reach.
         """
-        self._voice_mode[chat_id] = "off"
+        self._voice_mode[self._voice_key(Platform.DISCORD, chat_id)] = "off"
         self._save_voice_modes()
         adapter = self.adapters.get(Platform.DISCORD)
         self._set_adapter_auto_tts_disabled(adapter, chat_id, disabled=True)
@@ -6016,7 +6143,7 @@ class GatewayRunner:
             return False
 
         chat_id = event.source.chat_id
-        voice_mode = self._voice_mode.get(chat_id, "off")
+        voice_mode = self._voice_mode.get(self._voice_key(event.source.platform, chat_id), "off")
         is_voice_input = (event.message_type == MessageType.VOICE)
 
         should = (
@@ -7137,6 +7264,38 @@ class GatewayRunner:
                     if cached:
                         agent = cached[0]
 
+        # Resolve provider/base_url/api_key for the account-usage fetch.
+        # Prefer the live agent; fall back to persisted billing data on the
+        # SessionDB row so `/usage` still returns account info between turns
+        # when no agent is resident.
+        provider = getattr(agent, "provider", None) if agent and agent is not _AGENT_PENDING_SENTINEL else None
+        base_url = getattr(agent, "base_url", None) if agent and agent is not _AGENT_PENDING_SENTINEL else None
+        api_key = getattr(agent, "api_key", None) if agent and agent is not _AGENT_PENDING_SENTINEL else None
+        if not provider and getattr(self, "_session_db", None) is not None:
+            try:
+                _entry_for_billing = self.session_store.get_or_create_session(source)
+                persisted = self._session_db.get_session(_entry_for_billing.session_id) or {}
+            except Exception:
+                persisted = {}
+            provider = provider or persisted.get("billing_provider")
+            base_url = base_url or persisted.get("billing_base_url")
+
+        # Fetch account usage off the event loop so slow provider APIs don't
+        # block the gateway. Failures are non-fatal -- account_lines stays [].
+        account_lines: list[str] = []
+        if provider:
+            try:
+                account_snapshot = await asyncio.to_thread(
+                    fetch_account_usage,
+                    provider,
+                    base_url=base_url,
+                    api_key=api_key,
+                )
+            except Exception:
+                account_snapshot = None
+            if account_snapshot:
+                account_lines = render_account_usage_lines(account_snapshot, markdown=True)
+
         if agent and hasattr(agent, "session_total_tokens") and agent.session_api_calls > 0:
             lines = []
 
@@ -7194,6 +7353,10 @@ class GatewayRunner:
             if ctx.compression_count:
                 lines.append(f"Compressions: {ctx.compression_count}")
 
+            if account_lines:
+                lines.append("")
+                lines.extend(account_lines)
+
             return "\n".join(lines)
 
         # No agent at all -- check session history for a rough count
@@ -7203,23 +7366,26 @@ class GatewayRunner:
             from agent.model_metadata import estimate_messages_tokens_rough
             msgs = [m for m in history if m.get("role") in ("user", "assistant") and m.get("content")]
             approx = estimate_messages_tokens_rough(msgs)
-            return (
-                f"📊 **Session Info**\n"
-                f"Messages: {len(msgs)}\n"
-                f"Estimated context: ~{approx:,} tokens\n"
-                f"_(Detailed usage available after the first agent response)_"
-            )
+            lines = [
+                "📊 **Session Info**",
+                f"Messages: {len(msgs)}",
+                f"Estimated context: ~{approx:,} tokens",
+                "_(Detailed usage available after the first agent response)_",
+            ]
+            if account_lines:
+                lines.append("")
+                lines.extend(account_lines)
+            return "\n".join(lines)
+        if account_lines:
+            return "\n".join(account_lines)
         return "No usage data available for this session."
 
     async def _handle_insights_command(self, event: MessageEvent) -> str:
         """Handle /insights command -- show usage insights and analytics."""
-        import asyncio as _asyncio
-
         args = event.get_command_args().strip()
 
         # Normalize Unicode dashes (Telegram/iOS auto-converts -- to em/en dash)
-        import re as _re
-        args = _re.sub(r'[\u2012\u2013\u2014\u2015](days|source)', r'--\1', args)
+        args = re.sub(r'[\u2012\u2013\u2014\u2015](days|source)', r'--\1', args)
 
         days = 30
         source = None
@@ -7248,7 +7414,7 @@ class GatewayRunner:
             from hermes_state import SessionDB
             from agent.insights import InsightsEngine
 
-            loop = _asyncio.get_running_loop()
+            loop = asyncio.get_running_loop()
 
             def _run_insights():
                 db = SessionDB()
@@ -7606,9 +7772,6 @@ class GatewayRunner:
         the messenger.  The user's next message is intercepted by
         ``_handle_message`` and written to ``.update_response``.
         """
-        import json
-        import re as _re
-
         pending_path = _hermes_home / ".update_pending.json"
         claimed_path = _hermes_home / ".update_pending.claimed.json"
         output_path = _hermes_home / ".update_output.txt"
@@ -7653,7 +7816,7 @@ class GatewayRunner:
             return
 
         def _strip_ansi(text: str) -> str:
-            return _re.sub(r'\x1b\[[0-9;]*[A-Za-z]', '', text)
+            return re.sub(r'\x1b\[[0-9;]*[A-Za-z]', '', text)
 
         bytes_sent = 0
         last_stream_time = loop.time()
@@ -7801,9 +7964,6 @@ class GatewayRunner:
         cannot resolve the adapter (e.g. after a gateway restart where the
         platform hasn't reconnected yet).
         """
-        import json
-        import re as _re
-
         pending_path = _hermes_home / ".update_pending.json"
         claimed_path = _hermes_home / ".update_pending.claimed.json"
         output_path = _hermes_home / ".update_output.txt"
@@ -7849,7 +8009,7 @@ class GatewayRunner:
 
             if adapter and chat_id:
                 # Strip ANSI escape codes for clean display
-                output = _re.sub(r'\x1b\[[0-9;]*m', '', output).strip()
+                output = re.sub(r'\x1b\[[0-9;]*m', '', output).strip()
                 if output:
                     if len(output) > 3500:
                         output = "…" + output[-3500:]
@@ -7882,14 +8042,12 @@ class GatewayRunner:
 
     async def _send_restart_notification(self) -> None:
         """Notify the chat that initiated /restart that the gateway is back."""
-        import json as _json
-
         notify_path = _hermes_home / ".restart_notify.json"
         if not notify_path.exists():
             return
 
         try:
-            data = _json.loads(notify_path.read_text())
+            data = json.loads(notify_path.read_text())
             platform_str = data.get("platform")
             chat_id = data.get("chat_id")
             thread_id = data.get("thread_id")
@@ -7975,7 +8133,6 @@ class GatewayRunner:
             The enriched message string with vision descriptions prepended.
         """
         from tools.vision_tools import vision_analyze_tool
-        import json as _json
 
         analysis_prompt = (
             "Describe everything visible in this image in thorough detail. "
@@ -7991,7 +8148,7 @@ class GatewayRunner:
                     image_url=path,
                     user_prompt=analysis_prompt,
                 )
-                result = _json.loads(result_json)
+                result = json.loads(result_json)
                 if result.get("success"):
                     description = result.get("analysis", "")
                     enriched_parts.append(
@@ -8050,7 +8207,6 @@ class GatewayRunner:
             return disabled_note
 
         from tools.transcription_tools import transcribe_audio
-        import asyncio
 
         enriched_parts = []
         for path in audio_paths:
@@ -8186,7 +8342,6 @@ class GatewayRunner:
         if not adapter:
             return
         try:
-            from gateway.platforms.base import MessageEvent, MessageType
             synth_event = MessageEvent(
                 text=synth_text,
                 message_type=MessageType.TEXT,
@@ -8291,7 +8446,6 @@ class GatewayRunner:
                             break
                     if adapter and source.chat_id:
                         try:
-                            from gateway.platforms.base import MessageEvent, MessageType
                             synth_event = MessageEvent(
                                 text=synth_text,
                                 message_type=MessageType.TEXT,
@@ -8813,7 +8967,6 @@ class GatewayRunner:
         if _streaming_enabled:
             try:
                 from gateway.stream_consumer import GatewayStreamConsumer, StreamConsumerConfig
-                from gateway.config import Platform
                 _adapter = self.adapters.get(source.platform)
                 if _adapter:
                     _adapter_supports_edit = getattr(_adapter, "SUPPORTS_MESSAGE_EDITING", True)
@@ -9097,8 +9250,7 @@ class GatewayRunner:
                 if args:
                     from agent.display import get_tool_preview_max_len
                     _pl = get_tool_preview_max_len()
-                    import json as _json
-                    args_str = _json.dumps(args, ensure_ascii=False, default=str)
+                    args_str = json.dumps(args, ensure_ascii=False, default=str)
                     # When tool_preview_length is 0 (default), don't truncate
                     # in verbose mode — the user explicitly asked for full
                     # detail.  Platform message-length limits handle the rest.
@@ -9164,8 +9316,7 @@ class GatewayRunner:
             # Skip tool progress for platforms that don't support message
             # editing (e.g. iMessage/BlueBubbles) — each progress update
             # would become a separate message bubble, which is noisy.
-            from gateway.platforms.base import BasePlatformAdapter as _BaseAdapter
-            if type(adapter).edit_message is _BaseAdapter.edit_message:
+            if type(adapter).edit_message is BasePlatformAdapter.edit_message:
                 while not progress_queue.empty():
                     try:
                         progress_queue.get_nowait()
@@ -10299,6 +10450,16 @@ class GatewayRunner:
                     pending = pending_event.text or _build_media_placeholder(pending_event)
                     logger.debug("Processing queued message after agent completion: '%s...'", pending[:40])
 
+            # Leftover /steer: if a steer arrived after the last tool batch
+            # (e.g. during the final API call), the agent couldn't inject it
+            # and returned it in result["pending_steer"]. Deliver it as the
+            # next user turn so it isn't silently dropped.
+            if result and not pending and not pending_event:
+                _leftover_steer = result.get("pending_steer")
+                if _leftover_steer:
+                    pending = _leftover_steer
+                    logger.debug("Delivering leftover /steer as next turn: '%s...'", pending[:40])
+
             # Safety net: if the pending text is a slash command (e.g. "/stop",
             # "/new"), discard it — commands should never be passed to the agent
             # as user input.  The primary fix is in base.py (commands bypass the
@@ -10603,7 +10764,6 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
     # The PID file is scoped to HERMES_HOME, so future multi-profile
     # setups (each profile using a distinct HERMES_HOME) will naturally
     # allow concurrent instances without tripping this guard.
-    import time as _time
     from gateway.status import get_running_pid, remove_pid_file, terminate_pid
     existing_pid = get_running_pid()
     if existing_pid is not None and existing_pid != os.getpid():
@@ -10643,7 +10803,7 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
             for _ in range(20):
                 try:
                     os.kill(existing_pid, 0)
-                    _time.sleep(0.5)
+                    time.sleep(0.5)
                 except (ProcessLookupError, PermissionError):
                     break  # Process is gone
             else:
@@ -10654,10 +10814,16 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
                 )
                 try:
                     terminate_pid(existing_pid, force=True)
-                    _time.sleep(0.5)
+                    time.sleep(0.5)
                 except (ProcessLookupError, PermissionError, OSError):
                     pass
             remove_pid_file()
+            # remove_pid_file() is a no-op when the PID doesn't match.
+            # Force-unlink to cover the old-process-crashed case.
+            try:
+                (get_hermes_home() / "gateway.pid").unlink(missing_ok=True)
+            except Exception:
+                pass
             # Clean up any takeover marker the old process didn't consume
             # (e.g. SIGKILL'd before its shutdown handler could read it).
             try:
@@ -10796,6 +10962,30 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
     else:
         logger.info("Skipping signal handlers (not running in main thread).")
     
+    # Claim the PID file BEFORE bringing up any platform adapters.
+    # This closes the --replace race window: two concurrent `gateway run
+    # --replace` invocations both pass the termination-wait above, but
+    # only the winner of the O_CREAT|O_EXCL race below will ever open
+    # Telegram polling, Discord gateway sockets, etc. The loser exits
+    # cleanly before touching any external service.
+    import atexit
+    from gateway.status import write_pid_file, remove_pid_file, get_running_pid
+    _current_pid = get_running_pid()
+    if _current_pid is not None and _current_pid != os.getpid():
+        logger.error(
+            "Another gateway instance (PID %d) started during our startup. "
+            "Exiting to avoid double-running.", _current_pid
+        )
+        return False
+    try:
+        write_pid_file()
+    except FileExistsError:
+        logger.error(
+            "PID file race lost to another gateway instance. Exiting."
+        )
+        return False
+    atexit.register(remove_pid_file)
+
     # Start the gateway
     success = await runner.start()
     if not success:
@@ -10805,12 +10995,6 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
             logger.error("Gateway exiting cleanly: %s", runner.exit_reason)
         return True
     
-    # Write PID file so CLI can detect gateway is running
-    import atexit
-    from gateway.status import write_pid_file, remove_pid_file
-    write_pid_file()
-    atexit.register(remove_pid_file)
-    
     # Start background cron ticker so scheduled jobs fire automatically.
     # Pass the event loop so cron delivery can use live adapters (E2EE support).
     cron_stop = threading.Event()

gateway/session.py

@@ -152,6 +152,7 @@ class SessionContext:
     source: SessionSource
     connected_platforms: List[Platform]
     home_channels: Dict[Platform, HomeChannel]
+    shared_multi_user_session: bool = False
     
     # Session metadata
     session_key: str = ""
@@ -166,6 +167,7 @@ class SessionContext:
             "home_channels": {
                 p.value: hc.to_dict() for p, hc in self.home_channels.items()
             },
+            "shared_multi_user_session": self.shared_multi_user_session,
             "session_key": self.session_key,
             "session_id": self.session_id,
             "created_at": self.created_at.isoformat() if self.created_at else None,
@@ -240,18 +242,16 @@ def build_session_context_prompt(
         lines.append(f"**Channel Topic:** {context.source.chat_topic}")
 
     # User identity.
-    # In shared thread sessions (non-DM with thread_id), multiple users
-    # contribute to the same conversation.  Don't pin a single user name
-    # in the system prompt — it changes per-turn and would bust the prompt
-    # cache.  Instead, note that this is a multi-user thread; individual
-    # sender names are prefixed on each user message by the gateway.
-    _is_shared_thread = (
-        context.source.chat_type != "dm"
-        and context.source.thread_id
-    )
-    if _is_shared_thread:
+    # In shared multi-user sessions (shared threads OR shared non-thread groups
+    # when group_sessions_per_user=False), multiple users contribute to the same
+    # conversation.  Don't pin a single user name in the system prompt — it
+    # changes per-turn and would bust the prompt cache.  Instead, note that
+    # this is a multi-user session; individual sender names are prefixed on
+    # each user message by the gateway.
+    if context.shared_multi_user_session:
+        session_label = "Multi-user thread" if context.source.thread_id else "Multi-user session"
         lines.append(
-            "**Session type:** Multi-user thread — messages are prefixed "
+            f"**Session type:** {session_label} — messages are prefixed "
             "with [sender name]. Multiple users may participate."
         )
     elif context.source.user_name:
@@ -467,6 +467,27 @@ class SessionEntry:
         )
 
 
+def is_shared_multi_user_session(
+    source: SessionSource,
+    *,
+    group_sessions_per_user: bool = True,
+    thread_sessions_per_user: bool = False,
+) -> bool:
+    """Return True when a non-DM session is shared across participants.
+
+    Mirrors the isolation rules in :func:`build_session_key`:
+      - DMs are never shared.
+      - Threads are shared unless ``thread_sessions_per_user`` is True.
+      - Non-thread group/channel sessions are shared unless
+        ``group_sessions_per_user`` is True (default: True = isolated).
+    """
+    if source.chat_type == "dm":
+        return False
+    if source.thread_id:
+        return not thread_sessions_per_user
+    return not group_sessions_per_user
+
+
 def build_session_key(
     source: SessionSource,
     group_sessions_per_user: bool = True,
@@ -926,12 +947,18 @@ class SessionStore:
                     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_id):
+                        if self._has_active_processes_fn(entry.session_key):
                             continue
-                    except Exception:
-                        pass
+                    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:
@@ -1232,6 +1259,11 @@ def build_session_context(
         source=source,
         connected_platforms=connected,
         home_channels=home_channels,
+        shared_multi_user_session=is_shared_multi_user_session(
+            source,
+            group_sessions_per_user=getattr(config, "group_sessions_per_user", True),
+            thread_sessions_per_user=getattr(config, "thread_sessions_per_user", False),
+        ),
     )
     
     if session_entry:

gateway/session_context.py

@@ -56,6 +56,12 @@ _SESSION_USER_ID: ContextVar = ContextVar("HERMES_SESSION_USER_ID", default=_UNS
 _SESSION_USER_NAME: ContextVar = ContextVar("HERMES_SESSION_USER_NAME", default=_UNSET)
 _SESSION_KEY: ContextVar = ContextVar("HERMES_SESSION_KEY", default=_UNSET)
 
+# Cron auto-delivery vars — set per-job in run_job() so concurrent jobs
+# don't clobber each other's delivery targets.
+_CRON_AUTO_DELIVER_PLATFORM: ContextVar = ContextVar("HERMES_CRON_AUTO_DELIVER_PLATFORM", default=_UNSET)
+_CRON_AUTO_DELIVER_CHAT_ID: ContextVar = ContextVar("HERMES_CRON_AUTO_DELIVER_CHAT_ID", default=_UNSET)
+_CRON_AUTO_DELIVER_THREAD_ID: ContextVar = ContextVar("HERMES_CRON_AUTO_DELIVER_THREAD_ID", default=_UNSET)
+
 _VAR_MAP = {
     "HERMES_SESSION_PLATFORM": _SESSION_PLATFORM,
     "HERMES_SESSION_CHAT_ID": _SESSION_CHAT_ID,
@@ -64,6 +70,9 @@ _VAR_MAP = {
     "HERMES_SESSION_USER_ID": _SESSION_USER_ID,
     "HERMES_SESSION_USER_NAME": _SESSION_USER_NAME,
     "HERMES_SESSION_KEY": _SESSION_KEY,
+    "HERMES_CRON_AUTO_DELIVER_PLATFORM": _CRON_AUTO_DELIVER_PLATFORM,
+    "HERMES_CRON_AUTO_DELIVER_CHAT_ID": _CRON_AUTO_DELIVER_CHAT_ID,
+    "HERMES_CRON_AUTO_DELIVER_THREAD_ID": _CRON_AUTO_DELIVER_THREAD_ID,
 }
 
 

gateway/status.py

@@ -225,8 +225,28 @@ def _cleanup_invalid_pid_path(pid_path: Path, *, cleanup_stale: bool) -> None:
 
 
 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())
+    """Write the current process PID and metadata to the gateway PID file.
+
+    Uses atomic O_CREAT | O_EXCL creation so that concurrent --replace
+    invocations race: exactly one process wins and the rest get
+    FileExistsError.
+    """
+    path = _get_pid_path()
+    path.parent.mkdir(parents=True, exist_ok=True)
+    record = json.dumps(_build_pid_record())
+    try:
+        fd = os.open(path, os.O_CREAT | os.O_EXCL | os.O_WRONLY)
+    except FileExistsError:
+        raise  # Let caller decide: another gateway is racing us
+    try:
+        with os.fdopen(fd, "w", encoding="utf-8") as f:
+            f.write(record)
+    except Exception:
+        try:
+            path.unlink(missing_ok=True)
+        except OSError:
+            pass
+        raise
 
 
 def write_runtime_status(

hermes_cli/auth.py

@@ -20,6 +20,7 @@ import logging
 import os
 import shutil
 import shlex
+import ssl
 import stat
 import base64
 import hashlib
@@ -151,7 +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",
     ),
@@ -353,6 +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
@@ -480,6 +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 {}
@@ -1652,7 +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 {}
 
@@ -1672,13 +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
 
 
@@ -2721,6 +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.

hermes_cli/auth_commands.py

@@ -152,6 +152,23 @@ def auth_add_command(args) -> None:
 
     pool = load_pool(provider)
 
+    # Clear ALL suppressions for this provider — re-adding a credential is
+    # a strong signal the user wants auth re-enabled.  This covers env:*
+    # (shell-exported vars), gh_cli (copilot), claude_code, qwen-cli,
+    # device_code (codex), etc.  One consistent re-engagement pattern.
+    # Matches the Codex device_code re-link pattern that predates this.
+    if not provider.startswith(CUSTOM_POOL_PREFIX):
+        try:
+            from hermes_cli.auth import (
+                _load_auth_store,
+                unsuppress_credential_source,
+            )
+            suppressed = _load_auth_store().get("suppressed_sources", {})
+            for src in list(suppressed.get(provider, []) or []):
+                unsuppress_credential_source(provider, src)
+        except Exception:
+            pass
+
     if requested_type == AUTH_TYPE_API_KEY:
         token = (getattr(args, "api_key", None) or "").strip()
         if not token:
@@ -338,71 +355,28 @@ def auth_remove_command(args) -> None:
         raise SystemExit(f'No credential matching "{target}" for provider {provider}.')
     print(f"Removed {provider} credential #{index} ({removed.label})")
 
-    # If this was an env-seeded credential, also clear the env var from .env
-    # so it doesn't get re-seeded on the next load_pool() call.
-    if removed.source.startswith("env:"):
-        env_var = removed.source[len("env:"):]
-        if env_var:
-            from hermes_cli.config import remove_env_value
-            cleared = remove_env_value(env_var)
-            if cleared:
-                print(f"Cleared {env_var} from .env")
+    # Unified removal dispatch.  Every credential source Hermes reads from
+    # (env vars, external OAuth files, auth.json blocks, custom config)
+    # has a RemovalStep registered in agent.credential_sources.  The step
+    # handles its source-specific cleanup and we centralise suppression +
+    # user-facing output here so every source behaves identically from
+    # the user's perspective.
+    from agent.credential_sources import find_removal_step
+    from hermes_cli.auth import suppress_credential_source
 
-    # 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 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.")
+    step = find_removal_step(provider, removed.source)
+    if step is None:
+        # Unregistered source — e.g. "manual", which has nothing external
+        # to clean up.  The pool entry is already gone; we're done.
+        return
 
-    elif removed.source == "device_code" and provider == "nous":
-        from hermes_cli.auth import (
-            _load_auth_store, _save_auth_store, _auth_store_lock,
-        )
-        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")
-
-    elif removed.source == "hermes_pkce" and provider == "anthropic":
-        from hermes_constants import get_hermes_home
-        oauth_file = get_hermes_home() / ".anthropic_oauth.json"
-        if oauth_file.exists():
-            oauth_file.unlink()
-            print("Cleared Hermes Anthropic OAuth credentials")
-
-    elif removed.source == "claude_code" and provider == "anthropic":
-        from hermes_cli.auth import suppress_credential_source
-        suppress_credential_source(provider, "claude_code")
-        print("Suppressed claude_code credential — it will not be re-seeded.")
-        print("Note: Claude Code credentials still live in ~/.claude/.credentials.json")
-        print("Run `hermes auth add anthropic` to re-enable if needed.")
+    result = step.remove_fn(provider, removed)
+    for line in result.cleaned:
+        print(line)
+    if result.suppress:
+        suppress_credential_source(provider, removed.source)
+    for line in result.hints:
+        print(line)
 
 
 def auth_reset_command(args) -> None:

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/codex_models.py

@@ -24,7 +24,6 @@ _FORWARD_COMPAT_TEMPLATE_MODELS: List[tuple[str, tuple[str, ...]]] = [
     ("gpt-5.4-mini", ("gpt-5.3-codex", "gpt-5.2-codex")),
     ("gpt-5.4", ("gpt-5.3-codex", "gpt-5.2-codex")),
     ("gpt-5.3-codex", ("gpt-5.2-codex",)),
-    ("gpt-5.3-codex-spark", ("gpt-5.3-codex", "gpt-5.2-codex")),
 ]
 
 

hermes_cli/commands.py

@@ -497,9 +497,8 @@ def _collect_gateway_skill_entries(
     # --- Tier 1: Plugin slash commands (never trimmed) ---------------------
     plugin_pairs: list[tuple[str, str]] = []
     try:
-        from hermes_cli.plugins import get_plugin_manager
-        pm = get_plugin_manager()
-        plugin_cmds = getattr(pm, "_plugin_commands", {})
+        from hermes_cli.plugins import get_plugin_commands
+        plugin_cmds = get_plugin_commands()
         for cmd_name in sorted(plugin_cmds):
             name = sanitize_name(cmd_name) if sanitize_name else cmd_name
             if not name:
@@ -925,12 +924,22 @@ class SlashCommandCompleter(Completer):
                     display_meta=meta,
                 )
 
-        # If the user typed @file: or @folder:, delegate to path completions
+        # If the user typed @file: / @folder: (or just @file / @folder with
+        # no colon yet), delegate to path completions.  Accepting the bare
+        # form lets the picker surface directories as soon as the user has
+        # typed `@folder`, without requiring them to first accept the static
+        # `@folder:` hint and re-trigger completion.
         for prefix in ("@file:", "@folder:"):
-            if word.startswith(prefix):
-                path_part = word[len(prefix):] or "."
+            bare = prefix[:-1]
+
+            if word == bare or word.startswith(prefix):
+                want_dir = prefix == "@folder:"
+                path_part = '' if word == bare else word[len(prefix):]
                 expanded = os.path.expanduser(path_part)
-                if expanded.endswith("/"):
+
+                if not expanded or expanded == ".":
+                    search_dir, match_prefix = ".", ""
+                elif expanded.endswith("/"):
                     search_dir, match_prefix = expanded, ""
                 else:
                     search_dir = os.path.dirname(expanded) or "."
@@ -946,15 +955,21 @@ class SlashCommandCompleter(Completer):
                 for entry in sorted(entries):
                     if match_prefix and not entry.lower().startswith(prefix_lower):
                         continue
-                    if count >= limit:
-                        break
                     full_path = os.path.join(search_dir, entry)
                     is_dir = os.path.isdir(full_path)
+                    # `@folder:` must only surface directories; `@file:` only
+                    # regular files.  Without this filter `@folder:` listed
+                    # every .env / .gitignore in the cwd, defeating the
+                    # explicit prefix and confusing users expecting a
+                    # directory picker.
+                    if want_dir != is_dir:
+                        continue
+                    if count >= limit:
+                        break
                     display_path = os.path.relpath(full_path)
                     suffix = "/" if is_dir else ""
-                    kind = "folder" if is_dir else "file"
                     meta = "dir" if is_dir else _file_size_label(full_path)
-                    completion = f"@{kind}:{display_path}{suffix}"
+                    completion = f"{prefix}{display_path}{suffix}"
                     yield Completion(
                         completion,
                         start_position=-len(word),

hermes_cli/config.py

@@ -13,6 +13,7 @@ This module provides:
 """
 
 import copy
+import logging
 import os
 import platform
 import re
@@ -24,6 +25,7 @@ from dataclasses import dataclass
 from pathlib import Path
 from typing import Dict, Any, Optional, List, Tuple
 
+logger = logging.getLogger(__name__)
 
 _IS_WINDOWS = platform.system() == "Windows"
 _ENV_VAR_NAME_RE = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*$")
@@ -385,6 +387,26 @@ DEFAULT_CONFIG = {
         # (terminal and execute_code).  Skill-declared required_environment_variables
         # are passed through automatically; this list is for non-skill use cases.
         "env_passthrough": [],
+        # Extra files to source in the login shell when building the
+        # per-session environment snapshot.  Use this when tools like nvm,
+        # pyenv, asdf, or custom PATH entries are registered by files that
+        # a bash login shell would skip — most commonly ``~/.bashrc``
+        # (bash doesn't source bashrc in non-interactive login mode) or
+        # zsh-specific files like ``~/.zshrc`` / ``~/.zprofile``.
+        # Paths support ``~`` / ``${VAR}``. Missing files are silently
+        # skipped. When empty, Hermes auto-appends ``~/.bashrc`` if the
+        # snapshot shell is bash (this is the ``auto_source_bashrc``
+        # behaviour — disable with that key if you want strict login-only
+        # semantics).
+        "shell_init_files": [],
+        # When true (default), Hermes sources ``~/.bashrc`` in the login
+        # shell used to build the environment snapshot.  This captures
+        # PATH additions, shell functions, and aliases defined in the
+        # user's bashrc — which a plain ``bash -l -c`` would otherwise
+        # miss because bash skips bashrc in non-interactive login mode.
+        # Turn this off if you have a bashrc that misbehaves when sourced
+        # non-interactively (e.g. one that hard-exits on TTY checks).
+        "auto_source_bashrc": True,
         "docker_image": "nikolaik/python-nodejs:python3.11-nodejs20",
         "docker_forward_env": [],
         # Explicit environment variables to set inside Docker containers.
@@ -474,13 +496,6 @@ DEFAULT_CONFIG = {
         },
     },
 
-    "smart_model_routing": {
-        "enabled": False,
-        "max_simple_chars": 160,
-        "max_simple_words": 28,
-        "cheap_model": {},
-    },
-    
     # 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.
@@ -494,6 +509,7 @@ 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": {
@@ -502,6 +518,7 @@ 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",
@@ -509,6 +526,7 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 120,        # seconds — compression summarises large contexts; increase for local models
+            "extra_body": {},
         },
         "session_search": {
             "provider": "auto",
@@ -516,6 +534,8 @@ 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",
@@ -523,6 +543,7 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 30,
+            "extra_body": {},
         },
         "approval": {
             "provider": "auto",
@@ -530,6 +551,7 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 30,
+            "extra_body": {},
         },
         "mcp": {
             "provider": "auto",
@@ -537,6 +559,7 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 30,
+            "extra_body": {},
         },
         "flush_memories": {
             "provider": "auto",
@@ -544,6 +567,7 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 30,
+            "extra_body": {},
         },
         "title_generation": {
             "provider": "auto",
@@ -551,6 +575,7 @@ DEFAULT_CONFIG = {
             "base_url": "",
             "api_key": "",
             "timeout": 30,
+            "extra_body": {},
         },
     },
     
@@ -562,9 +587,14 @@ 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
@@ -635,6 +665,7 @@ DEFAULT_CONFIG = {
         "record_key": "ctrl+b",
         "max_recording_seconds": 120,
         "auto_tts": False,
+        "beep_enabled": True,         # Play record start/stop beeps in CLI voice mode
         "silence_threshold": 200,     # RMS below this = silence (0-32767)
         "silence_duration": 3.0,      # Seconds of silence before auto-stop
     },
@@ -681,6 +712,12 @@ DEFAULT_CONFIG = {
                                # independent of the parent's max_iterations)
         "reasoning_effort": "",  # reasoning effort for subagents: "xhigh", "high", "medium",
                                  # "low", "minimal", "none" (empty = inherit parent's level)
+        "max_concurrent_children": 3,  # max parallel children per batch; floor of 1 enforced, no ceiling
+        # Orchestrator role controls (see tools/delegate_tool.py:_get_max_spawn_depth
+        # and _get_orchestrator_enabled).  Values are clamped to [1, 3] with a
+        # warning log if out of range.
+        "max_spawn_depth": 1,        # depth cap (1 = flat [default], 2 = orchestrator→leaf, 3 = three-level)
+        "orchestrator_enabled": True,  # kill switch for role="orchestrator"
     },
 
     # Ephemeral prefill messages file — JSON list of {role, content} dicts
@@ -693,6 +730,20 @@ DEFAULT_CONFIG = {
     # always goes to ~/.hermes/skills/.
     "skills": {
         "external_dirs": [],   # e.g. ["~/.agents/skills", "/shared/team-skills"]
+        # Substitute ${HERMES_SKILL_DIR} and ${HERMES_SESSION_ID} in SKILL.md
+        # content with the absolute skill directory and the active session id
+        # before the agent sees it.  Lets skill authors reference bundled
+        # scripts without the agent having to join paths.
+        "template_vars": True,
+        # Pre-execute inline shell snippets written as !`cmd` in SKILL.md
+        # body.  Their stdout is inlined into the skill message before the
+        # agent reads it, so skills can inject dynamic context (dates, git
+        # state, detected tool versions, …).  Off by default because any
+        # content from the skill author runs on the host without approval;
+        # only enable for skill sources you trust.
+        "inline_shell": False,
+        # Timeout (seconds) for each !`cmd` snippet when inline_shell is on.
+        "inline_shell_timeout": 10,
     },
 
     # Honcho AI-native memory -- reads ~/.honcho/config.json as single source of truth.
@@ -712,6 +763,14 @@ DEFAULT_CONFIG = {
         "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)
@@ -755,6 +814,21 @@ DEFAULT_CONFIG = {
     "command_allowlist": [],
     # User-defined quick commands that bypass the agent loop (type: exec only)
     "quick_commands": {},
+
+    # Shell-script hooks — declarative bridge that invokes shell scripts
+    # on plugin-hook events (pre_tool_call, post_tool_call, pre_llm_call,
+    # subagent_stop, etc.).  Each entry maps an event name to a list of
+    # {matcher, command, timeout} dicts.  First registration of a new
+    # command prompts the user for consent; subsequent runs reuse the
+    # stored approval from ~/.hermes/shell-hooks-allowlist.json.
+    # See `website/docs/user-guide/features/hooks.md` for schema + examples.
+    "hooks": {},
+
+    # Auto-accept shell-hook registrations without a TTY prompt.  Also
+    # toggleable per-invocation via --accept-hooks or HERMES_ACCEPT_HOOKS=1.
+    # Gateway / cron / non-interactive runs need this (or one of the other
+    # channels) to pick up newly-added hooks.
+    "hooks_auto_accept": False,
     # Custom personalities — add your own entries here
     # Supports string format: {"name": "system prompt"}
     # Or dict format: {"name": {"description": "...", "system_prompt": "...", "tone": "...", "style": "..."}}
@@ -778,6 +852,11 @@ DEFAULT_CONFIG = {
         # Wrap delivered cron responses with a header (task name) and footer
         # ("The agent cannot see this message").  Set to false for clean output.
         "wrap_response": True,
+        # Maximum number of due jobs to run in parallel per tick.
+        # null/0 = unbounded (limited only by thread count).
+        # 1 = serial (pre-v0.9 behaviour).
+        # Also overridable via HERMES_CRON_MAX_PARALLEL env var.
+        "max_parallel_jobs": None,
     },
 
     # execute_code settings — controls the tool used for programmatic tool calls.
@@ -811,7 +890,7 @@ DEFAULT_CONFIG = {
     },
 
     # Config schema version - bump this when adding new required fields
-    "_config_version": 19,
+    "_config_version": 22,
 }
 
 # =============================================================================
@@ -1834,12 +1913,53 @@ def _normalize_custom_provider_entry(
     if not isinstance(entry, dict):
         return None
 
+    # Accept camelCase aliases commonly used in hand-written configs.
+    _CAMEL_ALIASES: Dict[str, str] = {
+        "apiKey": "api_key",
+        "baseUrl": "base_url",
+        "apiMode": "api_mode",
+        "keyEnv": "key_env",
+        "defaultModel": "default_model",
+        "contextLength": "context_length",
+        "rateLimitDelay": "rate_limit_delay",
+    }
+    _KNOWN_KEYS = {
+        "name", "api", "url", "base_url", "api_key", "key_env",
+        "api_mode", "transport", "model", "default_model", "models",
+        "context_length", "rate_limit_delay",
+    }
+    for camel, snake in _CAMEL_ALIASES.items():
+        if camel in entry and snake not in entry:
+            logger.warning(
+                "providers.%s: camelCase key '%s' auto-mapped to '%s' "
+                "(use snake_case to avoid this warning)",
+                provider_key or "?", camel, snake,
+            )
+            entry[snake] = entry[camel]
+    unknown = set(entry.keys()) - _KNOWN_KEYS - set(_CAMEL_ALIASES.keys())
+    if unknown:
+        logger.warning(
+            "providers.%s: unknown config keys ignored: %s",
+            provider_key or "?", ", ".join(sorted(unknown)),
+        )
+
+    from urllib.parse import urlparse
+
     base_url = ""
-    for url_key in ("api", "url", "base_url"):
+    for url_key in ("base_url", "url", "api"):
         raw_url = entry.get(url_key)
         if isinstance(raw_url, str) and raw_url.strip():
-            base_url = raw_url.strip()
-            break
+            candidate = raw_url.strip()
+            parsed = urlparse(candidate)
+            if parsed.scheme and parsed.netloc:
+                base_url = candidate
+                break
+            else:
+                logger.warning(
+                    "providers.%s: '%s' value '%s' is not a valid URL "
+                    "(no scheme or host) — skipped",
+                    provider_key or "?", url_key, candidate,
+                )
     if not base_url:
         return None
 
@@ -2135,7 +2255,6 @@ def print_config_warnings(config: Optional[Dict[str, Any]] = None) -> None:
     if not issues:
         return
 
-    import sys
     lines = ["\033[33m⚠ Config issues detected in config.yaml:\033[0m"]
     for ci in issues:
         marker = "\033[31m✗\033[0m" if ci.severity == "error" else "\033[33m⚠\033[0m"
@@ -2150,7 +2269,6 @@ def warn_deprecated_cwd_env_vars(config: Optional[Dict[str, Any]] = None) -> Non
     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")
 
@@ -2468,6 +2586,71 @@ def migrate_config(interactive: bool = True, quiet: bool = False) -> Dict[str, A
                     else:
                         print("  ✓ Removed unused compression.summary_* keys")
 
+    # ── Version 20 → 21: plugins are now opt-in; grandfather existing user plugins ──
+    # The loader now requires plugins to appear in ``plugins.enabled`` before
+    # loading. Existing installs had all discovered plugins loading by default
+    # (minus anything in ``plugins.disabled``). To avoid silently breaking
+    # those setups on upgrade, populate ``plugins.enabled`` with the set of
+    # currently-installed user plugins that aren't already disabled.
+    #
+    # Bundled plugins (shipped in the repo itself) are NOT grandfathered —
+    # they ship off for everyone, including existing users, so any user who
+    # wants one has to opt in explicitly.
+    if current_ver < 21:
+        config = read_raw_config()
+        plugins_cfg = config.get("plugins")
+        if not isinstance(plugins_cfg, dict):
+            plugins_cfg = {}
+        # Only migrate if the enabled allow-list hasn't been set yet.
+        if "enabled" not in plugins_cfg:
+            disabled = plugins_cfg.get("disabled", []) or []
+            if not isinstance(disabled, list):
+                disabled = []
+            disabled_set = set(disabled)
+
+            # Scan ``$HERMES_HOME/plugins/`` for currently installed user plugins.
+            grandfathered: List[str] = []
+            try:
+                user_plugins_dir = get_hermes_home() / "plugins"
+                if user_plugins_dir.is_dir():
+                    for child in sorted(user_plugins_dir.iterdir()):
+                        if not child.is_dir():
+                            continue
+                        manifest_file = child / "plugin.yaml"
+                        if not manifest_file.exists():
+                            manifest_file = child / "plugin.yml"
+                        if not manifest_file.exists():
+                            continue
+                        try:
+                            with open(manifest_file) as _mf:
+                                manifest = yaml.safe_load(_mf) or {}
+                        except Exception:
+                            manifest = {}
+                        name = manifest.get("name") or child.name
+                        if name in disabled_set:
+                            continue
+                        grandfathered.append(name)
+            except Exception:
+                grandfathered = []
+
+            plugins_cfg["enabled"] = grandfathered
+            config["plugins"] = plugins_cfg
+            save_config(config)
+            results["config_added"].append(
+                f"plugins.enabled (opt-in allow-list, {len(grandfathered)} grandfathered)"
+            )
+            if not quiet:
+                if grandfathered:
+                    print(
+                        f"  ✓ Plugins now opt-in: grandfathered "
+                        f"{len(grandfathered)} existing plugin(s) into plugins.enabled"
+                    )
+                else:
+                    print(
+                        "  ✓ Plugins now opt-in: no existing plugins to grandfather. "
+                        "Use `hermes plugins enable <name>` to activate."
+                    )
+
     if current_ver < latest_ver and not quiet:
         print(f"Config version: {current_ver} → {latest_ver}")
     
@@ -2870,19 +3053,6 @@ _FALLBACK_COMMENT = """
 # 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
 """
 
 
@@ -2914,19 +3084,6 @@ _COMMENTED_SECTIONS = """
 # 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
 """
 
 
@@ -3119,7 +3276,6 @@ def _check_non_ascii_credential(key: str, value: str) -> str:
             bad_chars.append(f"  position {i}: {ch!r} (U+{ord(ch):04X})")
     sanitized = value.encode("ascii", errors="ignore").decode("ascii")
 
-    import sys
     print(
         f"\n  Warning: {key} contains non-ASCII characters that will break API requests.\n"
         f"  This usually happens when copy-pasting from a PDF, rich-text editor,\n"
@@ -3389,6 +3545,10 @@ 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/doctor.py

@@ -30,6 +30,7 @@ load_dotenv(PROJECT_ROOT / ".env", override=False, encoding="utf-8")
 
 from hermes_cli.colors import Colors, color
 from hermes_constants import OPENROUTER_MODELS_URL
+from utils import base_url_host_matches
 
 
 _PROVIDER_ENV_HINTS = (
@@ -277,6 +278,86 @@ 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():
@@ -778,6 +859,16 @@ 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:
@@ -862,7 +953,7 @@ def run_doctor(args):
                     _base = _to_openai_base_url(_base)
                 _url = (_base.rstrip("/") + "/models") if _base else _default_url
                 _headers = {"Authorization": f"Bearer {_key}"}
-                if "api.kimi.com" in _url.lower():
+                if base_url_host_matches(_base, "api.kimi.com"):
                     _headers["User-Agent"] = "KimiCLI/1.30.0"
                 _resp = httpx.get(
                     _url,

hermes_cli/dump.py

@@ -160,7 +160,6 @@ def _config_overrides(config: dict) -> dict[str, str]:
         ("display", "streaming"),
         ("display", "skin"),
         ("display", "show_reasoning"),
-        ("smart_model_routing", "enabled"),
         ("privacy", "redact_pii"),
         ("tts", "provider"),
     ]

hermes_cli/env_loader.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import os
+import sys
 from pathlib import Path
 
 from dotenv import load_dotenv
@@ -14,6 +15,26 @@ from dotenv import load_dotenv
 # pure ASCII (they become HTTP header values).
 _CREDENTIAL_SUFFIXES = ("_API_KEY", "_TOKEN", "_SECRET", "_KEY")
 
+# Names we've already warned about during this process, so repeated
+# load_hermes_dotenv() calls (user env + project env, gateway hot-reload,
+# tests) don't spam the same warning multiple times.
+_WARNED_KEYS: set[str] = set()
+
+
+def _format_offending_chars(value: str, limit: int = 3) -> str:
+    """Return a compact 'U+XXXX ('c'), ...' summary of non-ASCII codepoints."""
+    seen: list[str] = []
+    for ch in value:
+        if ord(ch) > 127:
+            label = f"U+{ord(ch):04X}"
+            if ch.isprintable():
+                label += f" ({ch!r})"
+            if label not in seen:
+                seen.append(label)
+            if len(seen) >= limit:
+                break
+    return ", ".join(seen)
+
 
 def _sanitize_loaded_credentials() -> None:
     """Strip non-ASCII characters from credential env vars in os.environ.
@@ -21,14 +42,42 @@ def _sanitize_loaded_credentials() -> None:
     Called after dotenv loads so the rest of the codebase never sees
     non-ASCII API keys.  Only touches env vars whose names end with
     known credential suffixes (``_API_KEY``, ``_TOKEN``, etc.).
+
+    Emits a one-line warning to stderr when characters are stripped.
+    Silent stripping would mask copy-paste corruption (Unicode lookalike
+    glyphs from PDFs / rich-text editors, ZWSP from web pages) as opaque
+    provider-side "invalid API key" errors (see #6843).
     """
     for key, value in list(os.environ.items()):
         if not any(key.endswith(suffix) for suffix in _CREDENTIAL_SUFFIXES):
             continue
         try:
             value.encode("ascii")
+            continue
         except UnicodeEncodeError:
-            os.environ[key] = value.encode("ascii", errors="ignore").decode("ascii")
+            pass
+        cleaned = value.encode("ascii", errors="ignore").decode("ascii")
+        os.environ[key] = cleaned
+        if key in _WARNED_KEYS:
+            continue
+        _WARNED_KEYS.add(key)
+        stripped = len(value) - len(cleaned)
+        detail = _format_offending_chars(value) or "non-printable"
+        print(
+            f"  Warning: {key} contained {stripped} non-ASCII character"
+            f"{'s' if stripped != 1 else ''} ({detail}) — stripped so the "
+            f"key can be sent as an HTTP header.",
+            file=sys.stderr,
+        )
+        print(
+            "  This usually means the key was copy-pasted from a PDF, "
+            "rich-text editor, or web page that substituted lookalike\n"
+            "  Unicode glyphs for ASCII letters. If authentication fails "
+            "(e.g. \"API key not valid\"), re-copy the key from the\n"
+            "  provider's dashboard and run `hermes setup` (or edit the "
+            ".env file in a plain-text editor).",
+            file=sys.stderr,
+        )
 
 
 def _load_dotenv_with_fallback(path: Path, *, override: bool) -> None:

hermes_cli/gateway.py

@@ -994,8 +994,6 @@ def get_systemd_linger_status() -> tuple[bool | None, str]:
     if not is_linux():
         return None, "not supported on this platform"
 
-    import shutil
-
     if not shutil.which("loginctl"):
         return None, "loginctl not found"
 
@@ -1347,7 +1345,6 @@ def _ensure_linger_enabled() -> None:
         return
 
     import getpass
-    import shutil
 
     username = getpass.getuser()
     linger_file = Path(f"/var/lib/systemd/linger/{username}")
@@ -1656,7 +1653,6 @@ def get_launchd_label() -> str:
 
 
 def _launchd_domain() -> str:
-    import os
     return f"gui/{os.getuid()}"
 
 

hermes_cli/hooks.py

@@ -0,0 +1,385 @@
+"""hermes hooks — inspect and manage shell-script hooks.
+
+Usage::
+
+    hermes hooks list
+    hermes hooks test <event> [--for-tool X] [--payload-file F]
+    hermes hooks revoke <command>
+    hermes hooks doctor
+
+Consent records live under ``~/.hermes/shell-hooks-allowlist.json`` and
+hook definitions come from the ``hooks:`` block in ``~/.hermes/config.yaml``
+(the same config read by the CLI / gateway at startup).
+
+This module is a thin CLI shell over :mod:`agent.shell_hooks`; every
+shared concern (payload serialisation, response parsing, allowlist
+format) lives there.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+
+def hooks_command(args) -> None:
+    """Entry point for ``hermes hooks`` — dispatches to the requested action."""
+    sub = getattr(args, "hooks_action", None)
+
+    if not sub:
+        print("Usage: hermes hooks {list|test|revoke|doctor}")
+        print("Run 'hermes hooks --help' for details.")
+        return
+
+    if sub in ("list", "ls"):
+        _cmd_list(args)
+    elif sub == "test":
+        _cmd_test(args)
+    elif sub in ("revoke", "remove", "rm"):
+        _cmd_revoke(args)
+    elif sub == "doctor":
+        _cmd_doctor(args)
+    else:
+        print(f"Unknown hooks subcommand: {sub}")
+
+
+# ---------------------------------------------------------------------------
+# list
+# ---------------------------------------------------------------------------
+
+def _cmd_list(_args) -> None:
+    from hermes_cli.config import load_config
+    from agent import shell_hooks
+
+    specs = shell_hooks.iter_configured_hooks(load_config())
+
+    if not specs:
+        print("No shell hooks configured in ~/.hermes/config.yaml.")
+        print("See `hermes hooks --help` or")
+        print("    website/docs/user-guide/features/hooks.md")
+        print("for the config schema and worked examples.")
+        return
+
+    by_event: Dict[str, List] = {}
+    for spec in specs:
+        by_event.setdefault(spec.event, []).append(spec)
+
+    allowlist = shell_hooks.load_allowlist()
+    approved = {
+        (e.get("event"), e.get("command"))
+        for e in allowlist.get("approvals", [])
+        if isinstance(e, dict)
+    }
+
+    print(f"Configured shell hooks ({len(specs)} total):\n")
+
+    for event in sorted(by_event.keys()):
+        print(f"  [{event}]")
+        for spec in by_event[event]:
+            is_approved = (spec.event, spec.command) in approved
+            status = "✓ allowed" if is_approved else "✗ not allowlisted"
+            matcher_part = f" matcher={spec.matcher!r}" if spec.matcher else ""
+            print(
+                f"    - {spec.command}{matcher_part} "
+                f"(timeout={spec.timeout}s, {status})"
+            )
+
+            if is_approved:
+                entry = shell_hooks.allowlist_entry_for(spec.event, spec.command)
+                if entry and entry.get("approved_at"):
+                    print(f"      approved_at: {entry['approved_at']}")
+                    mtime_now = shell_hooks.script_mtime_iso(spec.command)
+                    mtime_at = entry.get("script_mtime_at_approval")
+                    if mtime_now and mtime_at and mtime_now > mtime_at:
+                        print(
+                            f"      ⚠ script modified since approval "
+                            f"(was {mtime_at}, now {mtime_now}) — "
+                            f"run `hermes hooks doctor` to re-validate"
+                        )
+        print()
+
+
+# ---------------------------------------------------------------------------
+# test
+# ---------------------------------------------------------------------------
+
+# Synthetic kwargs matching the real invoke_hook() call sites — these are
+# passed verbatim to agent.shell_hooks.run_once(), which routes them through
+# the same _serialize_payload() that production firings use.  That way the
+# stdin a script sees under `hermes hooks test` and `hermes hooks doctor`
+# is identical in shape to what it will see at runtime.
+_DEFAULT_PAYLOADS = {
+    "pre_tool_call": {
+        "tool_name": "terminal",
+        "args": {"command": "echo hello"},
+        "session_id": "test-session",
+        "task_id": "test-task",
+        "tool_call_id": "test-call",
+    },
+    "post_tool_call": {
+        "tool_name": "terminal",
+        "args": {"command": "echo hello"},
+        "session_id": "test-session",
+        "task_id": "test-task",
+        "tool_call_id": "test-call",
+        "result": '{"output": "hello"}',
+    },
+    "pre_llm_call": {
+        "session_id": "test-session",
+        "user_message": "What is the weather?",
+        "conversation_history": [],
+        "is_first_turn": True,
+        "model": "gpt-4",
+        "platform": "cli",
+    },
+    "post_llm_call": {
+        "session_id": "test-session",
+        "model": "gpt-4",
+        "platform": "cli",
+    },
+    "on_session_start": {"session_id": "test-session"},
+    "on_session_end": {"session_id": "test-session"},
+    "on_session_finalize": {"session_id": "test-session"},
+    "on_session_reset": {"session_id": "test-session"},
+    "pre_api_request": {
+        "session_id": "test-session",
+        "task_id": "test-task",
+        "platform": "cli",
+        "model": "claude-sonnet-4-6",
+        "provider": "anthropic",
+        "base_url": "https://api.anthropic.com",
+        "api_mode": "anthropic_messages",
+        "api_call_count": 1,
+        "message_count": 4,
+        "tool_count": 12,
+        "approx_input_tokens": 2048,
+        "request_char_count": 8192,
+        "max_tokens": 4096,
+    },
+    "post_api_request": {
+        "session_id": "test-session",
+        "task_id": "test-task",
+        "platform": "cli",
+        "model": "claude-sonnet-4-6",
+        "provider": "anthropic",
+        "base_url": "https://api.anthropic.com",
+        "api_mode": "anthropic_messages",
+        "api_call_count": 1,
+        "api_duration": 1.234,
+        "finish_reason": "stop",
+        "message_count": 4,
+        "response_model": "claude-sonnet-4-6",
+        "usage": {"input_tokens": 2048, "output_tokens": 512},
+        "assistant_content_chars": 1200,
+        "assistant_tool_call_count": 0,
+    },
+    "subagent_stop": {
+        "parent_session_id": "parent-sess",
+        "child_role": None,
+        "child_summary": "Synthetic summary for hooks test",
+        "child_status": "completed",
+        "duration_ms": 1234,
+    },
+}
+
+
+def _cmd_test(args) -> None:
+    from hermes_cli.config import load_config
+    from hermes_cli.plugins import VALID_HOOKS
+    from agent import shell_hooks
+
+    event = args.event
+    if event not in VALID_HOOKS:
+        print(f"Unknown event: {event!r}")
+        print(f"Valid events: {', '.join(sorted(VALID_HOOKS))}")
+        return
+
+    # Synthetic kwargs in the same shape invoke_hook() would pass.  Merged
+    # with --for-tool (overrides tool_name) and --payload-file (extra kwargs).
+    payload = dict(_DEFAULT_PAYLOADS.get(event, {"session_id": "test-session"}))
+
+    if getattr(args, "for_tool", None):
+        payload["tool_name"] = args.for_tool
+
+    if getattr(args, "payload_file", None):
+        try:
+            custom = json.loads(Path(args.payload_file).read_text())
+            if isinstance(custom, dict):
+                payload.update(custom)
+            else:
+                print(f"Warning: {args.payload_file} is not a JSON object; ignoring")
+        except Exception as exc:
+            print(f"Error reading payload file: {exc}")
+            return
+
+    specs = shell_hooks.iter_configured_hooks(load_config())
+    specs = [s for s in specs if s.event == event]
+
+    if getattr(args, "for_tool", None):
+        specs = [
+            s for s in specs
+            if s.event not in ("pre_tool_call", "post_tool_call")
+            or s.matches_tool(args.for_tool)
+        ]
+
+    if not specs:
+        print(f"No shell hooks configured for event: {event}")
+        if getattr(args, "for_tool", None):
+            print(f"(with matcher filter --for-tool={args.for_tool})")
+        return
+
+    print(f"Firing {len(specs)} hook(s) for event '{event}':\n")
+    for spec in specs:
+        print(f"  → {spec.command}")
+        result = shell_hooks.run_once(spec, payload)
+        _print_run_result(result)
+        print()
+
+
+def _print_run_result(result: Dict[str, Any]) -> None:
+    if result.get("error"):
+        print(f"      ✗ error: {result['error']}")
+        return
+    if result.get("timed_out"):
+        print(f"      ✗ timed out after {result['elapsed_seconds']}s")
+        return
+
+    rc = result.get("returncode")
+    elapsed = result.get("elapsed_seconds", 0)
+    print(f"      exit={rc}  elapsed={elapsed}s")
+
+    stdout = (result.get("stdout") or "").strip()
+    stderr = (result.get("stderr") or "").strip()
+    if stdout:
+        print(f"      stdout: {_truncate(stdout, 400)}")
+    if stderr:
+        print(f"      stderr: {_truncate(stderr, 400)}")
+
+    parsed = result.get("parsed")
+    if parsed:
+        print(f"      parsed (Hermes wire shape): {json.dumps(parsed)}")
+    else:
+        print("      parsed: <none — hook contributed nothing to the dispatcher>")
+
+
+def _truncate(s: str, n: int) -> str:
+    return s if len(s) <= n else s[: n - 3] + "..."
+
+
+# ---------------------------------------------------------------------------
+# revoke
+# ---------------------------------------------------------------------------
+
+def _cmd_revoke(args) -> None:
+    from agent import shell_hooks
+
+    removed = shell_hooks.revoke(args.command)
+    if removed == 0:
+        print(f"No allowlist entry found for command: {args.command}")
+        return
+    print(f"Removed {removed} allowlist entry/entries for: {args.command}")
+    print(
+        "Note: currently running CLI / gateway processes keep their "
+        "already-registered callbacks until they restart."
+    )
+
+
+# ---------------------------------------------------------------------------
+# doctor
+# ---------------------------------------------------------------------------
+
+def _cmd_doctor(_args) -> None:
+    from hermes_cli.config import load_config
+    from agent import shell_hooks
+
+    specs = shell_hooks.iter_configured_hooks(load_config())
+
+    if not specs:
+        print("No shell hooks configured — nothing to check.")
+        return
+
+    print(f"Checking {len(specs)} configured shell hook(s)...\n")
+
+    problems = 0
+    for spec in specs:
+        print(f"  [{spec.event}] {spec.command}")
+        problems += _doctor_one(spec, shell_hooks)
+        print()
+
+    if problems:
+        print(f"{problems} issue(s) found.  Fix before relying on these hooks.")
+    else:
+        print("All shell hooks look healthy.")
+
+
+def _doctor_one(spec, shell_hooks) -> int:
+    problems = 0
+
+    # 1. Script exists and is executable
+    if shell_hooks.script_is_executable(spec.command):
+        print("      ✓ script exists and is executable")
+    else:
+        problems += 1
+        print("      ✗ script missing or not executable "
+              "(chmod +x the file, or fix the path)")
+
+    # 2. Allowlist status
+    entry = shell_hooks.allowlist_entry_for(spec.event, spec.command)
+    if entry:
+        print(f"      ✓ allowlisted (approved {entry.get('approved_at', '?')})")
+    else:
+        problems += 1
+        print("      ✗ not allowlisted — hook will NOT fire at runtime "
+              "(run with --accept-hooks once, or confirm at the TTY prompt)")
+
+    # 3. Mtime drift
+    if entry and entry.get("script_mtime_at_approval"):
+        mtime_now = shell_hooks.script_mtime_iso(spec.command)
+        mtime_at = entry["script_mtime_at_approval"]
+        if mtime_now and mtime_at and mtime_now > mtime_at:
+            problems += 1
+            print(f"      ⚠ script modified since approval "
+                  f"(was {mtime_at}, now {mtime_now}) — review changes, "
+                  f"then `hermes hooks revoke` + re-approve to refresh")
+        elif mtime_now and mtime_at and mtime_now == mtime_at:
+            print("      ✓ script unchanged since approval")
+
+    # 4. Produces valid JSON for a synthetic payload — only when the entry
+    # is already allowlisted.  Otherwise `hermes hooks doctor` would execute
+    # every script listed in a freshly-pulled config before the user has
+    # reviewed them, which directly contradicts the documented workflow
+    # ("spot newly-added hooks *before they register*").
+    if not entry:
+        print("      ℹ skipped JSON smoke test — not allowlisted yet. "
+              "Approve the hook first (via TTY prompt or --accept-hooks), "
+              "then re-run `hermes hooks doctor`.")
+    elif shell_hooks.script_is_executable(spec.command):
+        payload = _DEFAULT_PAYLOADS.get(spec.event, {"extra": {}})
+        result = shell_hooks.run_once(spec, payload)
+        if result.get("timed_out"):
+            problems += 1
+            print(f"      ✗ timed out after {result['elapsed_seconds']}s "
+                  f"on synthetic payload (timeout={spec.timeout}s)")
+        elif result.get("error"):
+            problems += 1
+            print(f"      ✗ execution error: {result['error']}")
+        else:
+            rc = result.get("returncode")
+            elapsed = result.get("elapsed_seconds", 0)
+            stdout = (result.get("stdout") or "").strip()
+            if stdout:
+                try:
+                    json.loads(stdout)
+                    print(f"      ✓ produced valid JSON on synthetic payload "
+                          f"(exit={rc}, {elapsed}s)")
+                except json.JSONDecodeError:
+                    problems += 1
+                    print(f"      ✗ stdout was not valid JSON (exit={rc}, "
+                          f"{elapsed}s): {_truncate(stdout, 120)}")
+            else:
+                print(f"      ✓ ran clean with empty stdout "
+                      f"(exit={rc}, {elapsed}s) — hook is observer-only")
+
+    return problems

hermes_cli/main.py

@@ -51,6 +51,19 @@ import sys
 from pathlib import Path
 from typing import Optional
 
+def _add_accept_hooks_flag(parser) -> None:
+    """Attach the ``--accept-hooks`` flag.  Shared across every agent
+    subparser so the flag works regardless of CLI position."""
+    parser.add_argument(
+        "--accept-hooks",
+        action="store_true",
+        default=argparse.SUPPRESS,
+        help=(
+            "Auto-approve unseen shell hooks without a TTY prompt "
+            "(equivalent to HERMES_ACCEPT_HOOKS=1 / hooks_auto_accept: true)."
+        ),
+    )
+
 
 def _require_tty(command_name: str) -> None:
     """Exit with a clear error if stdin is not a terminal.
@@ -180,7 +193,7 @@ import time as _time
 from datetime import datetime
 
 from hermes_cli import __version__, __release_date__
-from hermes_constants import OPENROUTER_BASE_URL
+from hermes_constants import AI_GATEWAY_BASE_URL, OPENROUTER_BASE_URL
 
 logger = logging.getLogger(__name__)
 
@@ -605,7 +618,6 @@ def _exec_in_container(container_info: dict, cli_args: list):
         container_info: dict with backend, container_name, exec_user, hermes_bin
         cli_args: the original CLI arguments (everything after 'hermes')
     """
-    import shutil
 
     backend = container_info["backend"]
     container_name = container_info["container_name"]
@@ -693,6 +705,10 @@ def _resolve_session_by_name_or_id(name_or_id: str) -> Optional[str]:
     - If it looks like a session ID (contains underscore + hex), try direct lookup first.
     - Otherwise, treat it as a title and use resolve_session_by_title (auto-latest).
     - Falls back to the other method if the first doesn't match.
+    - If the resolved session is a compression root, follow the chain forward
+      to the latest continuation. Users who remember the old root ID (e.g.
+      from an exit summary printed before the bug fix, or from notes) get
+      resumed at the live tip instead of a stale parent with no messages.
     """
     try:
         from hermes_state import SessionDB
@@ -701,14 +717,23 @@ def _resolve_session_by_name_or_id(name_or_id: str) -> Optional[str]:
 
         # Try as exact session ID first
         session = db.get_session(name_or_id)
+        resolved_id: Optional[str] = None
         if session:
-            db.close()
-            return session["id"]
+            resolved_id = session["id"]
+        else:
+            # Try as title (with auto-latest for lineage)
+            resolved_id = db.resolve_session_by_title(name_or_id)
+
+        if resolved_id:
+            # Project forward through compression chain so resumes land on
+            # the live tip instead of a dead compressed parent.
+            try:
+                resolved_id = db.get_compression_tip(resolved_id) or resolved_id
+            except Exception:
+                pass
 
-        # Try as title (with auto-latest for lineage)
-        session_id = db.resolve_session_by_title(name_or_id)
         db.close()
-        return session_id
+        return resolved_id
     except Exception:
         pass
     return None
@@ -990,6 +1015,17 @@ def _launch_tui(resume_session_id: Optional[str] = None, tui_dev: bool = False):
     )
     env.setdefault("HERMES_PYTHON", sys.executable)
     env.setdefault("HERMES_CWD", os.getcwd())
+    # Guarantee an 8GB V8 heap + exposed GC for the TUI. Default node cap is
+    # ~1.5–4GB depending on version and can fatal-OOM on long sessions with
+    # large transcripts / reasoning blobs. Token-level merge: respect any
+    # user-supplied --max-old-space-size (they may have set it higher) and
+    # avoid duplicating --expose-gc.
+    _tokens = env.get("NODE_OPTIONS", "").split()
+    if not any(t.startswith("--max-old-space-size=") for t in _tokens):
+        _tokens.append("--max-old-space-size=8192")
+    if "--expose-gc" not in _tokens:
+        _tokens.append("--expose-gc")
+    env["NODE_OPTIONS"] = " ".join(_tokens)
     if resume_session_id:
         env["HERMES_TUI_RESUME"] = resume_session_id
 
@@ -1144,8 +1180,6 @@ def cmd_gateway(args):
 def cmd_whatsapp(args):
     """Set up WhatsApp: choose mode, configure, install bridge, pair via QR."""
     _require_tty("whatsapp")
-    import subprocess
-    from pathlib import Path
     from hermes_cli.config import get_env_value, save_env_value
 
     print()
@@ -1254,16 +1288,27 @@ def cmd_whatsapp(args):
         return
 
     if not (bridge_dir / "node_modules").exists():
-        print("\n→ Installing WhatsApp bridge dependencies...")
-        result = subprocess.run(
-            ["npm", "install"],
-            cwd=str(bridge_dir),
-            capture_output=True,
-            text=True,
-            timeout=120,
-        )
+        print("\n→ Installing WhatsApp bridge dependencies (this can take a few minutes)...")
+        npm = shutil.which("npm")
+        if not npm:
+            print("  ✗ npm not found on PATH — install Node.js first")
+            return
+        try:
+            result = subprocess.run(
+                [npm, "install", "--no-fund", "--no-audit", "--progress=false"],
+                cwd=str(bridge_dir),
+                stdout=subprocess.DEVNULL,
+                stderr=subprocess.PIPE,
+                text=True,
+            )
+        except KeyboardInterrupt:
+            print("\n  ✗ Install cancelled")
+            return
         if result.returncode != 0:
-            print(f"  ✗ npm install failed: {result.stderr}")
+            err = (result.stderr or "").strip()
+            preview = "\n".join(err.splitlines()[-30:]) if err else "(no output)"
+            print("  ✗ npm install failed:")
+            print(preview)
             return
         print("  ✓ Dependencies installed")
     else:
@@ -1282,8 +1327,6 @@ def cmd_whatsapp(args):
         except (EOFError, KeyboardInterrupt):
             response = "n"
         if response.lower() in ("y", "yes"):
-            import shutil
-
             shutil.rmtree(session_dir, ignore_errors=True)
             session_dir.mkdir(parents=True, exist_ok=True)
             print("  ✓ Session cleared")
@@ -1379,8 +1422,6 @@ def select_provider_and_model(args=None):
 
     # Read effective provider the same way the CLI does at startup:
     # config.yaml model.provider > env var > auto-detect
-    import os
-
     config_provider = None
     model_cfg = config.get("model")
     if isinstance(model_cfg, dict):
@@ -1491,6 +1532,8 @@ def select_provider_and_model(args=None):
     # Step 2: Provider-specific setup + model selection
     if selected_provider == "openrouter":
         _model_flow_openrouter(config, current_model)
+    elif selected_provider == "ai-gateway":
+        _model_flow_ai_gateway(config, current_model)
     elif selected_provider == "nous":
         _model_flow_nous(config, current_model, args=args)
     elif selected_provider == "openai-codex":
@@ -1536,7 +1579,6 @@ def select_provider_and_model(args=None):
         "kilocode",
         "opencode-zen",
         "opencode-go",
-        "ai-gateway",
         "alibaba",
         "huggingface",
         "xiaomi",
@@ -2008,6 +2050,63 @@ def _model_flow_openrouter(config, current_model=""):
         print("No change.")
 
 
+def _model_flow_ai_gateway(config, current_model=""):
+    """Vercel AI Gateway provider: ensure API key, then pick model with pricing."""
+    from hermes_cli.auth import (
+        _prompt_model_selection,
+        _save_model_choice,
+        deactivate_provider,
+    )
+    from hermes_cli.config import get_env_value, save_env_value
+
+    api_key = get_env_value("AI_GATEWAY_API_KEY")
+    if not api_key:
+        print("No Vercel AI Gateway API key configured.")
+        print("Create API key here: https://vercel.com/d?to=%2F%5Bteam%5D%2F%7E%2Fai-gateway&title=AI+Gateway")
+        print("Add a payment method to get $5 in free credits.")
+        print()
+        try:
+            import getpass
+
+            key = getpass.getpass("AI Gateway API key (or Enter to cancel): ").strip()
+        except (KeyboardInterrupt, EOFError):
+            print()
+            return
+        if not key:
+            print("Cancelled.")
+            return
+        save_env_value("AI_GATEWAY_API_KEY", key)
+        print("API key saved.")
+        print()
+
+    from hermes_cli.models import ai_gateway_model_ids, get_pricing_for_provider
+
+    models_list = ai_gateway_model_ids(force_refresh=True)
+    pricing = get_pricing_for_provider("ai-gateway", force_refresh=True)
+
+    selected = _prompt_model_selection(
+        models_list, current_model=current_model, pricing=pricing
+    )
+    if selected:
+        _save_model_choice(selected)
+
+        from hermes_cli.config import load_config, save_config
+
+        cfg = load_config()
+        model = cfg.get("model")
+        if not isinstance(model, dict):
+            model = {"default": model} if model else {}
+            cfg["model"] = model
+        model["provider"] = "ai-gateway"
+        model["base_url"] = AI_GATEWAY_BASE_URL
+        model["api_mode"] = "chat_completions"
+        save_config(cfg)
+        deactivate_provider()
+        print(f"Default model set to: {selected} (via Vercel AI Gateway)")
+    else:
+        print("No change.")
+
+
 def _model_flow_nous(config, current_model="", args=None):
     """Nous Portal provider: ensure logged in, then pick model."""
     from hermes_cli.auth import (
@@ -2028,7 +2127,6 @@ def _model_flow_nous(config, current_model="", args=None):
         save_env_value,
     )
     from hermes_cli.nous_subscription import prompt_enable_tool_gateway
-    import argparse
 
     state = get_provider_auth_state("nous")
     if not state or not state.get("access_token"):
@@ -2196,7 +2294,6 @@ def _model_flow_openai_codex(config, current_model=""):
         DEFAULT_CODEX_BASE_URL,
     )
     from hermes_cli.codex_models import get_codex_model_ids
-    import argparse
 
     status = get_codex_auth_status()
     if not status.get("logged_in"):
@@ -2351,7 +2448,7 @@ def _model_flow_google_gemini_cli(_config, current_model=""):
         return
 
     models = list(_PROVIDER_MODELS.get("google-gemini-cli") or [])
-    default = current_model or (models[0] if models else "gemini-2.5-flash")
+    default = current_model or (models[0] if models else "gemini-3-flash-preview")
     selected = _prompt_model_selection(models, current_model=default)
     if selected:
         _save_model_choice(selected)
@@ -3327,8 +3424,9 @@ def _model_flow_kimi(config, current_model=""):
 
     # Step 3: Model selection — show appropriate models for the endpoint
     if is_coding_plan:
-        # Coding Plan models (kimi-k2.5 first)
+        # Coding Plan models (kimi-k2.6 first)
         model_list = [
+            "kimi-k2.6",
             "kimi-k2.5",
             "kimi-for-coding",
             "kimi-k2-thinking",
@@ -4067,6 +4165,12 @@ def cmd_webhook(args):
     webhook_command(args)
 
 
+def cmd_hooks(args):
+    """Shell-hook inspection and management."""
+    from hermes_cli.hooks import hooks_command
+    hooks_command(args)
+
+
 def cmd_doctor(args):
     """Check configuration and dependencies."""
     from hermes_cli.doctor import run_doctor
@@ -4176,9 +4280,7 @@ def _clear_bytecode_cache(root: Path) -> int:
         ]
         if os.path.basename(dirpath) == "__pycache__":
             try:
-                import shutil as _shutil
-
-                _shutil.rmtree(dirpath)
+                shutil.rmtree(dirpath)
                 removed += 1
             except OSError:
                 pass
@@ -4217,8 +4319,6 @@ def _gateway_prompt(prompt_text: str, default: str = "", timeout: float = 300.0)
     tmp.replace(prompt_path)
 
     # Poll for response
-    import time as _time
-
     deadline = _time.monotonic() + timeout
     while _time.monotonic() < deadline:
         if response_path.exists():
@@ -4250,7 +4350,6 @@ def _build_web_ui(web_dir: Path, *, fatal: bool = False) -> bool:
     """
     if not (web_dir / "package.json").exists():
         return True
-    import shutil
 
     npm = shutil.which("npm")
     if not npm:
@@ -4287,7 +4386,6 @@ def _update_via_zip(args):
     Used on Windows when git file I/O is broken (antivirus, NTFS filter
     drivers causing 'Invalid argument' errors on file creation).
     """
-    import shutil
     import tempfile
     import zipfile
     from urllib.request import urlretrieve
@@ -4364,7 +4462,6 @@ def _update_via_zip(args):
     # breaks on this machine, keep base deps and reinstall the remaining extras
     # individually so update does not silently strip working capabilities.
     print("→ Updating Python dependencies...")
-    import subprocess
 
     uv_bin = shutil.which("uv")
     if uv_bin:
@@ -5115,9 +5212,11 @@ def _install_hangup_protection(gateway_mode: bool = False):
     # (2) Mirror output to update.log and wrap stdio for broken-pipe
     # tolerance.  Any failure here is non-fatal; we just skip the wrap.
     try:
-        from hermes_cli.config import get_hermes_home
+        # Late-bound import so tests can monkeypatch
+        # hermes_cli.config.get_hermes_home to simulate setup failure.
+        from hermes_cli.config import get_hermes_home as _get_hermes_home
 
-        logs_dir = get_hermes_home() / "logs"
+        logs_dir = _get_hermes_home() / "logs"
         logs_dir.mkdir(parents=True, exist_ok=True)
         log_path = logs_dir / "update.log"
         log_file = open(log_path, "a", buffering=1, encoding="utf-8")
@@ -5692,8 +5791,6 @@ def _cmd_update_impl(args, gateway_mode: bool):
                                     # Verify the service actually survived the
                                     # restart.  systemctl restart returns 0 even
                                     # if the new process crashes immediately.
-                                    import time as _time
-
                                     _time.sleep(3)
                                     verify = subprocess.run(
                                         scope_cmd + ["is-active", svc_name],
@@ -6346,6 +6443,17 @@ For more help on a command:
         default=False,
         help="Run in an isolated git worktree (for parallel agents)",
     )
+    parser.add_argument(
+        "--accept-hooks",
+        action="store_true",
+        default=False,
+        help=(
+            "Auto-approve any unseen shell hooks declared in config.yaml "
+            "without a TTY prompt.  Equivalent to HERMES_ACCEPT_HOOKS=1 or "
+            "hooks_auto_accept: true in config.yaml.  Use on CI / headless "
+            "runs that can't prompt."
+        ),
+    )
     parser.add_argument(
         "--skills",
         "-s",
@@ -6468,6 +6576,16 @@ For more help on a command:
         default=argparse.SUPPRESS,
         help="Run in an isolated git worktree (for parallel agents on the same repo)",
     )
+    chat_parser.add_argument(
+        "--accept-hooks",
+        action="store_true",
+        default=argparse.SUPPRESS,
+        help=(
+            "Auto-approve any unseen shell hooks declared in config.yaml "
+            "without a TTY prompt (see also HERMES_ACCEPT_HOOKS env var and "
+            "hooks_auto_accept: in config.yaml)."
+        ),
+    )
     chat_parser.add_argument(
         "--checkpoints",
         action="store_true",
@@ -6587,6 +6705,8 @@ For more help on a command:
         action="store_true",
         help="Replace any existing gateway instance (useful for systemd)",
     )
+    _add_accept_hooks_flag(gateway_run)
+    _add_accept_hooks_flag(gateway_parser)
 
     # gateway start
     gateway_start = gateway_subparsers.add_parser(
@@ -6951,6 +7071,7 @@ For more help on a command:
         "run", help="Run a job on the next scheduler tick"
     )
     cron_run.add_argument("job_id", help="Job ID to trigger")
+    _add_accept_hooks_flag(cron_run)
 
     cron_remove = cron_subparsers.add_parser(
         "remove", aliases=["rm", "delete"], help="Remove a scheduled job"
@@ -6961,8 +7082,9 @@ For more help on a command:
     cron_subparsers.add_parser("status", help="Check if cron scheduler is running")
 
     # cron tick (mostly for debugging)
-    cron_subparsers.add_parser("tick", help="Run due jobs once and exit")
-
+    cron_tick = cron_subparsers.add_parser("tick", help="Run due jobs once and exit")
+    _add_accept_hooks_flag(cron_tick)
+    _add_accept_hooks_flag(cron_parser)
     cron_parser.set_defaults(func=cmd_cron)
 
     # =========================================================================
@@ -7029,6 +7151,67 @@ For more help on a command:
 
     webhook_parser.set_defaults(func=cmd_webhook)
 
+    # =========================================================================
+    # hooks command — shell-hook inspection and management
+    # =========================================================================
+    hooks_parser = subparsers.add_parser(
+        "hooks",
+        help="Inspect and manage shell-script hooks",
+        description=(
+            "Inspect shell-script hooks declared in ~/.hermes/config.yaml, "
+            "test them against synthetic payloads, and manage the first-use "
+            "consent allowlist at ~/.hermes/shell-hooks-allowlist.json."
+        ),
+    )
+    hooks_subparsers = hooks_parser.add_subparsers(dest="hooks_action")
+
+    hooks_subparsers.add_parser(
+        "list", aliases=["ls"],
+        help="List configured hooks with matcher, timeout, and consent status",
+    )
+
+    _hk_test = hooks_subparsers.add_parser(
+        "test",
+        help="Fire every hook matching <event> against a synthetic payload",
+    )
+    _hk_test.add_argument(
+        "event",
+        help="Hook event name (e.g. pre_tool_call, pre_llm_call, subagent_stop)",
+    )
+    _hk_test.add_argument(
+        "--for-tool", dest="for_tool", default=None,
+        help=(
+            "Only fire hooks whose matcher matches this tool name "
+            "(used for pre_tool_call / post_tool_call)"
+        ),
+    )
+    _hk_test.add_argument(
+        "--payload-file", dest="payload_file", default=None,
+        help=(
+            "Path to a JSON file whose contents are merged into the "
+            "synthetic payload before execution"
+        ),
+    )
+
+    _hk_revoke = hooks_subparsers.add_parser(
+        "revoke", aliases=["remove", "rm"],
+        help="Remove a command's allowlist entries (takes effect on next restart)",
+    )
+    _hk_revoke.add_argument(
+        "command",
+        help="The exact command string to revoke (as declared in config.yaml)",
+    )
+
+    hooks_subparsers.add_parser(
+        "doctor",
+        help=(
+            "Check each configured hook: exec bit, allowlist, mtime drift, "
+            "JSON validity, and synthetic run timing"
+        ),
+    )
+
+    hooks_parser.set_defaults(func=cmd_hooks)
+
     # =========================================================================
     # doctor command
     # =========================================================================
@@ -7436,6 +7619,17 @@ Examples:
         action="store_true",
         help="Remove existing plugin and reinstall",
     )
+    _install_enable_group = plugins_install.add_mutually_exclusive_group()
+    _install_enable_group.add_argument(
+        "--enable",
+        action="store_true",
+        help="Auto-enable the plugin after install (skip confirmation prompt)",
+    )
+    _install_enable_group.add_argument(
+        "--no-enable",
+        action="store_true",
+        help="Install disabled (skip confirmation prompt); enable later with `hermes plugins enable <name>`",
+    )
 
     plugins_update = plugins_subparsers.add_parser(
         "update", help="Pull latest changes for an installed plugin"
@@ -7483,9 +7677,7 @@ Examples:
             )
             cmd_info["setup_fn"](plugin_parser)
     except Exception as _exc:
-        import logging as _log
-
-        _log.getLogger(__name__).debug("Plugin CLI discovery failed: %s", _exc)
+        logging.getLogger(__name__).debug("Plugin CLI discovery failed: %s", _exc)
 
     # =========================================================================
     # memory command
@@ -7691,6 +7883,7 @@ Examples:
         action="store_true",
         help="Enable verbose logging on stderr",
     )
+    _add_accept_hooks_flag(mcp_serve_p)
 
     mcp_add_p = mcp_sub.add_parser(
         "add", help="Add an MCP server (discovery-first install)"
@@ -7729,6 +7922,8 @@ Examples:
     )
     mcp_login_p.add_argument("name", help="Server name to re-authenticate")
 
+    _add_accept_hooks_flag(mcp_parser)
+
     def cmd_mcp(args):
         from hermes_cli.mcp_config import mcp_command
 
@@ -7867,7 +8062,6 @@ Examples:
                     return
                 line = _json.dumps(data, ensure_ascii=False) + "\n"
                 if args.output == "-":
-                    import sys
 
                     sys.stdout.write(line)
                 else:
@@ -7877,7 +8071,6 @@ Examples:
             else:
                 sessions = db.export_all(source=args.source)
                 if args.output == "-":
-                    import sys
 
                     for s in sessions:
                         sys.stdout.write(_json.dumps(s, ensure_ascii=False) + "\n")
@@ -7948,8 +8141,6 @@ Examples:
 
             # Launch hermes --resume <id> by replacing the current process
             print(f"Resuming session: {selected_id}")
-            import shutil
-
             hermes_bin = shutil.which("hermes")
             if hermes_bin:
                 os.execvp(hermes_bin, ["hermes", "--resume", selected_id])
@@ -8140,6 +8331,7 @@ Examples:
         help="Run Hermes Agent as an ACP (Agent Client Protocol) server",
         description="Start Hermes Agent in ACP mode for editor integration (VS Code, Zed, JetBrains)",
     )
+    _add_accept_hooks_flag(acp_parser)
 
     def cmd_acp(args):
         """Launch Hermes Agent as an ACP server."""
@@ -8413,6 +8605,42 @@ Examples:
         cmd_version(args)
         return
 
+    # Discover Python plugins and register shell hooks once, before any
+    # command that can fire lifecycle hooks.  Both are idempotent; gated
+    # so introspection/management commands (hermes hooks list, cron
+    # list, gateway status, mcp add, ...) don't pay discovery cost or
+    # trigger consent prompts for hooks the user is still inspecting.
+    # Groups with mixed admin/CRUD vs. agent-running entries narrow via
+    # the nested subcommand (dest varies by parser).
+    _AGENT_COMMANDS = {None, "chat", "acp", "rl"}
+    _AGENT_SUBCOMMANDS = {
+        "cron":    ("cron_command",    {"run", "tick"}),
+        "gateway": ("gateway_command", {"run"}),
+        "mcp":     ("mcp_action",      {"serve"}),
+    }
+    _sub_attr, _sub_set = _AGENT_SUBCOMMANDS.get(args.command, (None, None))
+    if (
+        args.command in _AGENT_COMMANDS
+        or (_sub_attr and getattr(args, _sub_attr, None) in _sub_set)
+    ):
+        _accept_hooks = bool(getattr(args, "accept_hooks", False))
+        try:
+            from hermes_cli.plugins import discover_plugins
+            discover_plugins()
+        except Exception:
+            logger.debug(
+                "plugin discovery failed at CLI startup", exc_info=True,
+            )
+        try:
+            from hermes_cli.config import load_config
+            from agent.shell_hooks import register_from_config
+            register_from_config(load_config(), accept_hooks=_accept_hooks)
+        except Exception:
+            logger.debug(
+                "shell-hook registration failed at CLI startup",
+                exc_info=True,
+            )
+
     # Handle top-level --resume / --continue as shortcut to chat
     if (args.resume or args.continue_last) and args.command is None:
         args.command = "chat"

hermes_cli/model_switch.py

@@ -1035,21 +1035,49 @@ def list_authenticated_providers(
         seen_slugs.add(_cp.slug.lower())
 
     # --- 3. User-defined endpoints from config ---
+    # Track (name, base_url) of what section 3 emits so section 4 can skip
+    # any overlapping ``custom_providers:`` entries.  Callers typically pass
+    # both (gateway/CLI invoke ``get_compatible_custom_providers()`` which
+    # merges ``providers:`` into the list) — without this, the same endpoint
+    # produces two picker rows: one bare-slug ("openrouter") from section 3
+    # and one "custom:openrouter" from section 4, both labelled identically.
+    _section3_emitted_pairs: set = set()
     if user_providers and isinstance(user_providers, dict):
         for ep_name, ep_cfg in user_providers.items():
             if not isinstance(ep_cfg, dict):
                 continue
+            # Skip if this slug was already emitted (e.g. canonical provider
+            # with the same name) or will be picked up by section 4.
+            if ep_name.lower() in seen_slugs:
+                continue
             display_name = ep_cfg.get("name", "") or ep_name
-            api_url = ep_cfg.get("api", "") or ep_cfg.get("url", "") or ""
-            default_model = ep_cfg.get("default_model", "")
+            # ``base_url`` is Hermes's canonical write key (matches
+            # custom_providers and _save_custom_provider); ``api`` / ``url``
+            # remain as fallbacks for hand-edited / legacy configs.
+            api_url = (
+                ep_cfg.get("base_url", "")
+                or ep_cfg.get("api", "")
+                or ep_cfg.get("url", "")
+                or ""
+            )
+            # ``default_model`` is the legacy key; ``model`` matches what
+            # custom_providers entries use, so accept either.
+            default_model = ep_cfg.get("default_model", "") or ep_cfg.get("model", "")
 
             # Build models list from both default_model and full models array
             models_list = []
             if default_model:
                 models_list.append(default_model)
-            # Also include the full models list from config
+            # Also include the full models list from config.
+            # Hermes writes ``models:`` as a dict keyed by model id
+            # (see hermes_cli/main.py::_save_custom_provider); older
+            # configs or hand-edited files may still use a list.
             cfg_models = ep_cfg.get("models", [])
-            if isinstance(cfg_models, list):
+            if isinstance(cfg_models, dict):
+                for m in cfg_models:
+                    if m and m not in models_list:
+                        models_list.append(m)
+            elif isinstance(cfg_models, list):
                 for m in cfg_models:
                     if m and m not in models_list:
                         models_list.append(m)
@@ -1066,6 +1094,14 @@ def list_authenticated_providers(
                 "source": "user-config",
                 "api_url": api_url,
             })
+            seen_slugs.add(ep_name.lower())
+            seen_slugs.add(custom_provider_slug(display_name).lower())
+            _pair = (
+                str(display_name).strip().lower(),
+                str(api_url).strip().rstrip("/").lower(),
+            )
+            if _pair[0] and _pair[1]:
+                _section3_emitted_pairs.add(_pair)
 
     # --- 4. Saved custom providers from config ---
     # Each ``custom_providers`` entry represents one model under a named
@@ -1100,13 +1136,41 @@ def list_authenticated_providers(
                     "api_url": api_url,
                     "models": [],
                 }
+            # The singular ``model:`` field only holds the currently
+            # active model. Hermes's own writer (main.py::_save_custom_provider)
+            # stores every configured model as a dict under ``models:``;
+            # downstream readers (agent/models_dev.py, gateway/run.py,
+            # run_agent.py, hermes_cli/config.py) already consume that dict.
+            # The /model picker previously ignored it, so multi-model
+            # custom providers appeared to have only the active model.
             default_model = (entry.get("model") or "").strip()
             if default_model and default_model not in groups[slug]["models"]:
                 groups[slug]["models"].append(default_model)
 
+            cfg_models = entry.get("models", {})
+            if isinstance(cfg_models, dict):
+                for m in cfg_models:
+                    if m and m not in groups[slug]["models"]:
+                        groups[slug]["models"].append(m)
+            elif isinstance(cfg_models, list):
+                for m in cfg_models:
+                    if m and m not in groups[slug]["models"]:
+                        groups[slug]["models"].append(m)
+
         for slug, grp in groups.items():
             if slug.lower() in seen_slugs:
                 continue
+            # Skip if section 3 already emitted this endpoint under its
+            # ``providers:`` dict key — matches on (display_name, base_url),
+            # the tuple section 4 groups by.  Prevents two picker rows
+            # labelled identically when callers pass both ``user_providers``
+            # and a compatibility-merged ``custom_providers`` list.
+            _pair_key = (
+                str(grp["name"]).strip().lower(),
+                str(grp["api_url"]).strip().rstrip("/").lower(),
+            )
+            if _pair_key[0] and _pair_key[1] and _pair_key in _section3_emitted_pairs:
+                continue
             results.append({
                 "slug": slug,
                 "name": grp["name"],

hermes_cli/models.py

@@ -16,6 +16,12 @@ from difflib import get_close_matches
 from pathlib import Path
 from typing import Any, NamedTuple, Optional
 
+from hermes_cli import __version__ as _HERMES_VERSION
+
+# Identify ourselves so endpoints fronted by Cloudflare's Browser Integrity
+# Check (error 1010) don't reject the default ``Python-urllib/*`` signature.
+_HERMES_USER_AGENT = f"hermes-cli/{_HERMES_VERSION}"
+
 COPILOT_BASE_URL = "https://api.githubcopilot.com"
 COPILOT_MODELS_URL = f"{COPILOT_BASE_URL}/models"
 COPILOT_EDITOR_VERSION = "vscode/1.104.1"
@@ -26,7 +32,7 @@ COPILOT_REASONING_EFFORTS_O_SERIES = ["low", "medium", "high"]
 # Fallback OpenRouter snapshot used when the live catalog is unavailable.
 # (model_id, display description shown in menus)
 OPENROUTER_MODELS: list[tuple[str, str]] = [
-    ("moonshotai/kimi-k2.5",            "recommended"),
+    ("moonshotai/kimi-k2.6",            "recommended"),
     ("anthropic/claude-opus-4.7",       ""),
     ("anthropic/claude-opus-4.6",       ""),
     ("anthropic/claude-sonnet-4.6",     ""),
@@ -62,6 +68,31 @@ OPENROUTER_MODELS: list[tuple[str, str]] = [
 _openrouter_catalog_cache: list[tuple[str, str]] | None = None
 
 
+# Fallback Vercel AI Gateway snapshot used when the live catalog is unavailable.
+# OSS / open-weight models prioritized first, then closed-source by family.
+# Slugs match Vercel's actual /v1/models catalog (e.g. alibaba/ for Qwen,
+# zai/ and xai/ without hyphens).
+VERCEL_AI_GATEWAY_MODELS: list[tuple[str, str]] = [
+    ("moonshotai/kimi-k2.6",                 "recommended"),
+    ("alibaba/qwen3.6-plus",                 ""),
+    ("zai/glm-5.1",                          ""),
+    ("minimax/minimax-m2.7",                 ""),
+    ("anthropic/claude-sonnet-4.6",          ""),
+    ("anthropic/claude-opus-4.7",            ""),
+    ("anthropic/claude-opus-4.6",            ""),
+    ("anthropic/claude-haiku-4.5",           ""),
+    ("openai/gpt-5.4",                       ""),
+    ("openai/gpt-5.4-mini",                  ""),
+    ("openai/gpt-5.3-codex",                 ""),
+    ("google/gemini-3.1-pro-preview",        ""),
+    ("google/gemini-3-flash",                ""),
+    ("google/gemini-3.1-flash-lite-preview", ""),
+    ("xai/grok-4.20-reasoning",              ""),
+]
+
+_ai_gateway_catalog_cache: list[tuple[str, str]] | None = None
+
+
 def _codex_curated_models() -> list[str]:
     """Derive the openai-codex curated list from codex_models.py.
 
@@ -75,7 +106,7 @@ def _codex_curated_models() -> list[str]:
 
 _PROVIDER_MODELS: dict[str, list[str]] = {
     "nous": [
-        "moonshotai/kimi-k2.5",
+        "moonshotai/kimi-k2.6",
         "xiaomi/mimo-v2-pro",
         "anthropic/claude-opus-4.7",
         "anthropic/claude-opus-4.6",
@@ -128,16 +159,14 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
     ],
     "gemini": [
         "gemini-3.1-pro-preview",
+        "gemini-3-pro-preview",
         "gemini-3-flash-preview",
         "gemini-3.1-flash-lite-preview",
-        "gemini-2.5-pro",
-        "gemini-2.5-flash",
-        "gemini-2.5-flash-lite",
     ],
     "google-gemini-cli": [
-        "gemini-2.5-pro",
-        "gemini-2.5-flash",
-        "gemini-2.5-flash-lite",
+        "gemini-3.1-pro-preview",
+        "gemini-3-pro-preview",
+        "gemini-3-flash-preview",
     ],
     "zai": [
         "glm-5.1",
@@ -161,12 +190,13 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         # (map to OpenRouter defaults — users get familiar picks on NIM)
         "qwen/qwen3.5-397b-a17b",
         "deepseek-ai/deepseek-v3.2",
-        "moonshotai/kimi-k2.5",
+        "moonshotai/kimi-k2.6",
         "minimaxai/minimax-m2.5",
         "z-ai/glm5",
         "openai/gpt-oss-120b",
     ],
     "kimi-coding": [
+        "kimi-k2.6",
         "kimi-k2.5",
         "kimi-for-coding",
         "kimi-k2-thinking",
@@ -175,12 +205,14 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "kimi-k2-0905-preview",
     ],
     "kimi-coding-cn": [
+        "kimi-k2.6",
         "kimi-k2.5",
         "kimi-k2-thinking",
         "kimi-k2-turbo-preview",
         "kimi-k2-0905-preview",
     ],
     "moonshot": [
+        "kimi-k2.6",
         "kimi-k2.5",
         "kimi-k2-thinking",
         "kimi-k2-turbo-preview",
@@ -227,7 +259,6 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "gpt-5.4-pro",
         "gpt-5.4",
         "gpt-5.3-codex",
-        "gpt-5.3-codex-spark",
         "gpt-5.2",
         "gpt-5.2-codex",
         "gpt-5.1",
@@ -261,6 +292,7 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "big-pickle",
     ],
     "opencode-go": [
+        "kimi-k2.6",
         "kimi-k2.5",
         "glm-5.1",
         "glm-5",
@@ -268,20 +300,8 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "mimo-v2-omni",
         "minimax-m2.7",
         "minimax-m2.5",
-    ],
-    "ai-gateway": [
-        "anthropic/claude-opus-4.6",
-        "anthropic/claude-sonnet-4.6",
-        "anthropic/claude-sonnet-4.5",
-        "anthropic/claude-haiku-4.5",
-        "openai/gpt-5",
-        "openai/gpt-4.1",
-        "openai/gpt-4.1-mini",
-        "google/gemini-3-pro-preview",
-        "google/gemini-3-flash",
-        "google/gemini-2.5-pro",
-        "google/gemini-2.5-flash",
-        "deepseek/deepseek-v3.2",
+        "qwen3.6-plus",
+        "qwen3.5-plus",
     ],
     "kilocode": [
         "anthropic/claude-opus-4.6",
@@ -315,6 +335,7 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "zai-org/GLM-5",
         "XiaomiMiMo/MiMo-V2-Flash",
         "moonshotai/Kimi-K2-Thinking",
+        "moonshotai/Kimi-K2.6",
     ],
     # AWS Bedrock — static fallback list used when dynamic discovery is
     # unavailable (no boto3, no credentials, or API error).  The agent
@@ -334,6 +355,12 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
     ],
 }
 
+# Vercel AI Gateway: derive the bare-model-id catalog from the curated
+# ``VERCEL_AI_GATEWAY_MODELS`` snapshot so both the picker (tuples with descriptions)
+# and the static fallback catalog (bare ids) stay in sync from a single
+# source of truth.
+_PROVIDER_MODELS["ai-gateway"] = [mid for mid, _ in VERCEL_AI_GATEWAY_MODELS]
+
 # ---------------------------------------------------------------------------
 # Nous Portal free-model filtering
 # ---------------------------------------------------------------------------
@@ -491,8 +518,6 @@ def check_nous_free_tier() -> bool:
     Returns False (assume paid) on any error — never blocks paying users.
     """
     global _free_tier_cache
-    import time
-
     now = time.monotonic()
     if _free_tier_cache is not None:
         cached_result, cached_at = _free_tier_cache
@@ -544,6 +569,7 @@ class ProviderEntry(NamedTuple):
 CANONICAL_PROVIDERS: list[ProviderEntry] = [
     ProviderEntry("nous",           "Nous Portal",              "Nous Portal (Nous Research subscription)"),
     ProviderEntry("openrouter",     "OpenRouter",               "OpenRouter (100+ models, pay-per-use)"),
+    ProviderEntry("ai-gateway",     "Vercel AI Gateway",        "Vercel AI Gateway (200+ models, $5 free credit, no markup)"),
     ProviderEntry("anthropic",      "Anthropic",                "Anthropic (Claude models — API key or Claude Code)"),
     ProviderEntry("openai-codex",   "OpenAI Codex",             "OpenAI Codex"),
     ProviderEntry("xiaomi",         "Xiaomi MiMo",              "Xiaomi MiMo (MiMo-V2 models — pro, omni, flash)"),
@@ -552,7 +578,7 @@ CANONICAL_PROVIDERS: list[ProviderEntry] = [
     ProviderEntry("copilot",        "GitHub Copilot",           "GitHub Copilot (uses GITHUB_TOKEN or gh auth token)"),
     ProviderEntry("copilot-acp",    "GitHub Copilot ACP",       "GitHub Copilot ACP (spawns `copilot --acp --stdio`)"),
     ProviderEntry("huggingface",    "Hugging Face",             "Hugging Face Inference Providers (20+ open models)"),
-    ProviderEntry("gemini",         "Google AI Studio",         "Google AI Studio (Gemini models — OpenAI-compatible endpoint)"),
+    ProviderEntry("gemini",         "Google AI Studio",         "Google AI Studio (Gemini models — native Gemini API)"),
     ProviderEntry("google-gemini-cli", "Google Gemini (OAuth)",   "Google Gemini via OAuth + Code Assist (free tier supported; no API key needed)"),
     ProviderEntry("deepseek",       "DeepSeek",                 "DeepSeek (DeepSeek-V3, R1, coder — direct API)"),
     ProviderEntry("xai",            "xAI",                      "xAI (Grok models — direct API)"),
@@ -567,7 +593,6 @@ CANONICAL_PROVIDERS: list[ProviderEntry] = [
     ProviderEntry("kilocode",       "Kilo Code",                "Kilo Code (Kilo Gateway API)"),
     ProviderEntry("opencode-zen",   "OpenCode Zen",             "OpenCode Zen (35+ curated models, pay-as-you-go)"),
     ProviderEntry("opencode-go",    "OpenCode Go",              "OpenCode Go (open models, $10/month subscription)"),
-    ProviderEntry("ai-gateway",     "Vercel AI Gateway",        "Vercel AI Gateway (200+ models, pay-per-use)"),
     ProviderEntry("bedrock",        "AWS Bedrock",              "AWS Bedrock (Claude, Nova, Llama, DeepSeek — IAM or API key)"),
 ]
 
@@ -663,6 +688,31 @@ def _openrouter_model_is_free(pricing: Any) -> bool:
         return False
 
 
+def _openrouter_model_supports_tools(item: Any) -> bool:
+    """Return True when the model's ``supported_parameters`` advertise tool calling.
+
+    hermes-agent is tool-calling-first — every provider path assumes the model
+    can invoke tools. Models that don't advertise ``tools`` in their
+    ``supported_parameters`` (e.g. image-only or completion-only models) cannot
+    be driven by the agent loop and would fail at the first tool call.
+
+    **Permissive when the field is missing.** Some OpenRouter-compatible gateways
+    (Nous Portal, private mirrors, older catalog snapshots) don't populate
+    ``supported_parameters`` at all. Treat that as "unknown capability → allow"
+    so the picker doesn't silently empty for those users. Only hide models
+    whose ``supported_parameters`` is an explicit list that omits ``tools``.
+
+    Ported from Kilo-Org/kilocode#9068.
+    """
+    if not isinstance(item, dict):
+        return True
+    params = item.get("supported_parameters")
+    if not isinstance(params, list):
+        # Field absent / malformed / None — be permissive.
+        return True
+    return "tools" in params
+
+
 def fetch_openrouter_models(
     timeout: float = 8.0,
     *,
@@ -705,6 +755,11 @@ def fetch_openrouter_models(
         live_item = live_by_id.get(preferred_id)
         if live_item is None:
             continue
+        # Hide models that don't advertise tool-calling support — hermes-agent
+        # requires it and surfacing them leads to immediate runtime failures
+        # when the user selects them. Ported from Kilo-Org/kilocode#9068.
+        if not _openrouter_model_supports_tools(live_item):
+            continue
         desc = "free" if _openrouter_model_is_free(live_item.get("pricing")) else ""
         curated.append((preferred_id, desc))
 
@@ -722,6 +777,93 @@ def model_ids(*, force_refresh: bool = False) -> list[str]:
     return [mid for mid, _ in fetch_openrouter_models(force_refresh=force_refresh)]
 
 
+def _ai_gateway_model_is_free(pricing: Any) -> bool:
+    """Return True if an AI Gateway model has $0 input AND output pricing."""
+    if not isinstance(pricing, dict):
+        return False
+    try:
+        return float(pricing.get("input", "0")) == 0 and float(pricing.get("output", "0")) == 0
+    except (TypeError, ValueError):
+        return False
+
+
+def fetch_ai_gateway_models(
+    timeout: float = 8.0,
+    *,
+    force_refresh: bool = False,
+) -> list[tuple[str, str]]:
+    """Return the curated AI Gateway picker list, refreshed from the live catalog when possible."""
+    global _ai_gateway_catalog_cache
+
+    if _ai_gateway_catalog_cache is not None and not force_refresh:
+        return list(_ai_gateway_catalog_cache)
+
+    from hermes_constants import AI_GATEWAY_BASE_URL
+
+    fallback = list(VERCEL_AI_GATEWAY_MODELS)
+    preferred_ids = [mid for mid, _ in fallback]
+
+    try:
+        req = urllib.request.Request(
+            f"{AI_GATEWAY_BASE_URL.rstrip('/')}/models",
+            headers={"Accept": "application/json"},
+        )
+        with urllib.request.urlopen(req, timeout=timeout) as resp:
+            payload = json.loads(resp.read().decode())
+    except Exception:
+        return list(_ai_gateway_catalog_cache or fallback)
+
+    live_items = payload.get("data", [])
+    if not isinstance(live_items, list):
+        return list(_ai_gateway_catalog_cache or fallback)
+
+    live_by_id: dict[str, dict[str, Any]] = {}
+    for item in live_items:
+        if not isinstance(item, dict):
+            continue
+        mid = str(item.get("id") or "").strip()
+        if not mid:
+            continue
+        live_by_id[mid] = item
+
+    curated: list[tuple[str, str]] = []
+    for preferred_id in preferred_ids:
+        live_item = live_by_id.get(preferred_id)
+        if live_item is None:
+            continue
+        desc = "free" if _ai_gateway_model_is_free(live_item.get("pricing")) else ""
+        curated.append((preferred_id, desc))
+
+    if not curated:
+        return list(_ai_gateway_catalog_cache or fallback)
+
+    # If the live catalog offers a free Moonshot model, auto-promote it to
+    # position #1 as "recommended" — dynamic discovery without a PR.
+    free_moonshot = next(
+        (
+            mid
+            for mid, item in live_by_id.items()
+            if mid.startswith("moonshotai/")
+            and _ai_gateway_model_is_free(item.get("pricing"))
+        ),
+        None,
+    )
+    if free_moonshot:
+        curated = [(mid, desc) for mid, desc in curated if mid != free_moonshot]
+        curated.insert(0, (free_moonshot, "recommended"))
+    else:
+        first_id, _ = curated[0]
+        curated[0] = (first_id, "recommended")
+
+    _ai_gateway_catalog_cache = curated
+    return list(curated)
+
+
+def ai_gateway_model_ids(*, force_refresh: bool = False) -> list[str]:
+    """Return just the AI Gateway model-id strings."""
+    return [mid for mid, _ in fetch_ai_gateway_models(force_refresh=force_refresh)]
+
+
 
 
 # ---------------------------------------------------------------------------
@@ -866,6 +1008,56 @@ def fetch_models_with_pricing(
     return result
 
 
+def fetch_ai_gateway_pricing(
+    timeout: float = 8.0,
+    *,
+    force_refresh: bool = False,
+) -> dict[str, dict[str, str]]:
+    """Fetch Vercel AI Gateway /v1/models and return hermes-shaped pricing.
+
+    Vercel uses ``input`` / ``output`` field names; hermes's picker expects
+    ``prompt`` / ``completion``. This translates. Cache read/write field names
+    already match.
+    """
+    from hermes_constants import AI_GATEWAY_BASE_URL
+
+    cache_key = AI_GATEWAY_BASE_URL.rstrip("/")
+    if not force_refresh and cache_key in _pricing_cache:
+        return _pricing_cache[cache_key]
+
+    try:
+        req = urllib.request.Request(
+            f"{cache_key}/models",
+            headers={"Accept": "application/json"},
+        )
+        with urllib.request.urlopen(req, timeout=timeout) as resp:
+            payload = json.loads(resp.read().decode())
+    except Exception:
+        _pricing_cache[cache_key] = {}
+        return {}
+
+    result: dict[str, dict[str, str]] = {}
+    for item in payload.get("data", []):
+        if not isinstance(item, dict):
+            continue
+        mid = item.get("id")
+        pricing = item.get("pricing")
+        if not (mid and isinstance(pricing, dict)):
+            continue
+        entry: dict[str, str] = {
+            "prompt": str(pricing.get("input", "")),
+            "completion": str(pricing.get("output", "")),
+        }
+        if pricing.get("input_cache_read"):
+            entry["input_cache_read"] = str(pricing["input_cache_read"])
+        if pricing.get("input_cache_write"):
+            entry["input_cache_write"] = str(pricing["input_cache_write"])
+        result[mid] = entry
+
+    _pricing_cache[cache_key] = result
+    return result
+
+
 def _resolve_openrouter_api_key() -> str:
     """Best-effort OpenRouter API key for pricing fetch."""
     return os.getenv("OPENROUTER_API_KEY", "").strip()
@@ -884,7 +1076,7 @@ def _resolve_nous_pricing_credentials() -> tuple[str, str]:
 
 
 def get_pricing_for_provider(provider: str, *, force_refresh: bool = False) -> dict[str, dict[str, str]]:
-    """Return live pricing for providers that support it (openrouter, nous)."""
+    """Return live pricing for providers that support it (openrouter, nous, ai-gateway)."""
     normalized = normalize_provider(provider)
     if normalized == "openrouter":
         return fetch_models_with_pricing(
@@ -892,6 +1084,8 @@ def get_pricing_for_provider(provider: str, *, force_refresh: bool = False) -> d
             base_url="https://openrouter.ai/api",
             force_refresh=force_refresh,
         )
+    if normalized == "ai-gateway":
+        return fetch_ai_gateway_pricing(force_refresh=force_refresh)
     if normalized == "nous":
         api_key, base_url = _resolve_nous_pricing_credentials()
         if base_url:
@@ -1096,7 +1290,6 @@ def detect_provider_for_model(
             from hermes_cli.auth import PROVIDER_REGISTRY
             pconfig = PROVIDER_REGISTRY.get(direct_match)
             if pconfig:
-                import os
                 for env_var in pconfig.api_key_env_vars:
                     if os.getenv(env_var, "").strip():
                         has_creds = True
@@ -1771,7 +1964,7 @@ def probe_api_models(
         candidates.append((alternate_base, True))
 
     tried: list[str] = []
-    headers: dict[str, str] = {}
+    headers: dict[str, str] = {"User-Agent": _HERMES_USER_AGENT}
     if api_key:
         headers["Authorization"] = f"Bearer {api_key}"
     if normalized.startswith(COPILOT_BASE_URL):
@@ -2106,6 +2299,51 @@ def validate_requested_model(
                 ),
             }
 
+    # MiniMax providers don't expose a /models endpoint — validate against
+    # the static catalog instead, similar to openai-codex.
+    if normalized in ("minimax", "minimax-cn"):
+        try:
+            catalog_models = provider_model_ids(normalized)
+        except Exception:
+            catalog_models = []
+        if catalog_models:
+            # Case-insensitive lookup (catalog uses mixed case like MiniMax-M2.7)
+            catalog_lower = {m.lower(): m for m in catalog_models}
+            if requested_for_lookup.lower() in catalog_lower:
+                return {
+                    "accepted": True,
+                    "persist": True,
+                    "recognized": True,
+                    "message": None,
+                }
+            # Auto-correct close matches (case-insensitive)
+            catalog_lower_list = list(catalog_lower.keys())
+            auto = get_close_matches(requested_for_lookup.lower(), catalog_lower_list, n=1, cutoff=0.9)
+            if auto:
+                corrected = catalog_lower[auto[0]]
+                return {
+                    "accepted": True,
+                    "persist": True,
+                    "recognized": True,
+                    "corrected_model": corrected,
+                    "message": f"Auto-corrected `{requested}` → `{corrected}`",
+                }
+            suggestions = get_close_matches(requested_for_lookup.lower(), catalog_lower_list, n=3, cutoff=0.5)
+            suggestion_text = ""
+            if suggestions:
+                suggestion_text = "\n  Similar models: " + ", ".join(f"`{catalog_lower[s]}`" for s in suggestions)
+            return {
+                "accepted": True,
+                "persist": True,
+                "recognized": False,
+                "message": (
+                    f"Note: `{requested}` was not found in the MiniMax catalog."
+                    f"{suggestion_text}"
+                    "\n  MiniMax does not expose a /models endpoint, so Hermes cannot verify the model name."
+                    "\n  The model may still work if it exists on the server."
+                ),
+            }
+
     # Probe the live API to check if the model actually exists
     api_models = fetch_api_models(api_key, base_url)
 
@@ -2188,13 +2426,70 @@ def validate_requested_model(
         except Exception:
             pass  # Fall through to generic warning
 
+    # Static-catalog fallback: when the /models probe was unreachable,
+    # validate against the curated list from provider_model_ids() — same
+    # pattern as the openai-codex and minimax branches above.  This fixes
+    # /model switches in the gateway for providers like opencode-go and
+    # opencode-zen whose /models endpoint returns 404 against the HTML
+    # marketing site.  Without this block, validate_requested_model would
+    # reject every model on such providers, switch_model() would return
+    # success=False, and the gateway would never write to
+    # _session_model_overrides.
     provider_label = _PROVIDER_LABELS.get(normalized, normalized)
+    try:
+        catalog_models = provider_model_ids(normalized)
+    except Exception:
+        catalog_models = []
+
+    if catalog_models:
+        catalog_lower = {m.lower(): m for m in catalog_models}
+        if requested_for_lookup.lower() in catalog_lower:
+            return {
+                "accepted": True,
+                "persist": True,
+                "recognized": True,
+                "message": None,
+            }
+        catalog_lower_list = list(catalog_lower.keys())
+        auto = get_close_matches(
+            requested_for_lookup.lower(), catalog_lower_list, n=1, cutoff=0.9
+        )
+        if auto:
+            corrected = catalog_lower[auto[0]]
+            return {
+                "accepted": True,
+                "persist": True,
+                "recognized": True,
+                "corrected_model": corrected,
+                "message": f"Auto-corrected `{requested}` → `{corrected}`",
+            }
+        suggestions = get_close_matches(
+            requested_for_lookup.lower(), catalog_lower_list, n=3, cutoff=0.5
+        )
+        suggestion_text = ""
+        if suggestions:
+            suggestion_text = "\n  Similar models: " + ", ".join(
+                f"`{catalog_lower[s]}`" for s in suggestions
+            )
+        return {
+            "accepted": True,
+            "persist": True,
+            "recognized": False,
+            "message": (
+                f"Note: `{requested}` was not found in the {provider_label} curated catalog "
+                f"and the /models endpoint was unreachable.{suggestion_text}"
+                f"\n  The model may still work if it exists on the provider."
+            ),
+        }
+
+    # No catalog available — accept with a warning, matching the comment's
+    # stated intent ("Accept and persist, but warn").
     return {
-        "accepted": False,
-        "persist": False,
+        "accepted": True,
+        "persist": True,
         "recognized": False,
         "message": (
-            f"Could not reach the {provider_label} API to validate `{requested}`. "
+            f"Note: could not reach the {provider_label} API to validate `{requested}`. "
             f"If the service isn't down, this model may not be valid."
         ),
     }

hermes_cli/nous_subscription.py

@@ -10,6 +10,7 @@ from hermes_cli.auth import get_nous_auth_status
 from hermes_cli.config import get_env_value, load_config
 from tools.managed_tool_gateway import is_managed_tool_gateway_ready
 from tools.tool_backend_helpers import (
+    fal_key_is_configured,
     has_direct_modal_credentials,
     managed_nous_tools_enabled,
     normalize_browser_cloud_provider,
@@ -271,7 +272,7 @@ def get_nous_subscription_features(
     direct_firecrawl = bool(get_env_value("FIRECRAWL_API_KEY") or get_env_value("FIRECRAWL_API_URL"))
     direct_parallel = bool(get_env_value("PARALLEL_API_KEY"))
     direct_tavily = bool(get_env_value("TAVILY_API_KEY"))
-    direct_fal = bool(get_env_value("FAL_KEY"))
+    direct_fal = fal_key_is_configured()
     direct_openai_tts = bool(resolve_openai_audio_api_key())
     direct_elevenlabs = bool(get_env_value("ELEVENLABS_API_KEY"))
     direct_camofox = bool(get_env_value("CAMOFOX_URL"))
@@ -520,7 +521,7 @@ def apply_nous_managed_defaults(
         browser_cfg["cloud_provider"] = "browser-use"
         changed.add("browser")
 
-    if "image_gen" in selected_toolsets and not get_env_value("FAL_KEY"):
+    if "image_gen" in selected_toolsets and not fal_key_is_configured():
         changed.add("image_gen")
 
     return changed
@@ -548,7 +549,7 @@ def _get_gateway_direct_credentials() -> Dict[str, bool]:
             or get_env_value("TAVILY_API_KEY")
             or get_env_value("EXA_API_KEY")
         ),
-        "image_gen": bool(get_env_value("FAL_KEY")),
+        "image_gen": fal_key_is_configured(),
         "tts": bool(
             resolve_openai_audio_api_key()
             or get_env_value("ELEVENLABS_API_KEY")
@@ -586,7 +587,6 @@ def get_gateway_eligible_tools(
         return [], [], []
 
     if config is None:
-        from hermes_cli.config import load_config
         config = load_config() or {}
 
     # Quick provider check without the heavy get_nous_subscription_features call

hermes_cli/plugins.py

@@ -2,14 +2,20 @@
 Hermes Plugin System
 ====================
 
-Discovers, loads, and manages plugins from three sources:
+Discovers, loads, and manages plugins from four sources:
 
-1. **User plugins**   – ``~/.hermes/plugins/<name>/``
-2. **Project plugins** – ``./.hermes/plugins/<name>/`` (opt-in via
+1. **Bundled plugins** – ``<repo>/plugins/<name>/`` (shipped with hermes-agent;
+   ``memory/`` and ``context_engine/`` subdirs are excluded — they have their
+   own discovery paths)
+2. **User plugins**   – ``~/.hermes/plugins/<name>/``
+3. **Project plugins** – ``./.hermes/plugins/<name>/`` (opt-in via
    ``HERMES_ENABLE_PROJECT_PLUGINS``)
-3. **Pip plugins**     – packages that expose the ``hermes_agent.plugins``
+4. **Pip plugins**     – packages that expose the ``hermes_agent.plugins``
    entry-point group.
 
+Later sources override earlier ones on name collision, so a user or project
+plugin with the same name as a bundled plugin replaces it.
+
 Each directory plugin must contain a ``plugin.yaml`` manifest **and** an
 ``__init__.py`` with a ``register(ctx)`` function.
 
@@ -54,6 +60,8 @@ logger = logging.getLogger(__name__)
 VALID_HOOKS: Set[str] = {
     "pre_tool_call",
     "post_tool_call",
+    "transform_terminal_output",
+    "transform_tool_result",
     "pre_llm_call",
     "post_llm_call",
     "pre_api_request",
@@ -62,6 +70,7 @@ VALID_HOOKS: Set[str] = {
     "on_session_end",
     "on_session_finalize",
     "on_session_reset",
+    "subagent_stop",
 }
 
 ENTRY_POINTS_GROUP = "hermes_agent.plugins"
@@ -75,7 +84,12 @@ def _env_enabled(name: str) -> bool:
 
 
 def _get_disabled_plugins() -> set:
-    """Read the disabled plugins list from config.yaml."""
+    """Read the disabled plugins list from config.yaml.
+
+    Kept for backward compat and explicit deny-list semantics. A plugin
+    name in this set will never load, even if it appears in
+    ``plugins.enabled``.
+    """
     try:
         from hermes_cli.config import load_config
         config = load_config()
@@ -85,6 +99,36 @@ def _get_disabled_plugins() -> set:
         return set()
 
 
+def _get_enabled_plugins() -> Optional[set]:
+    """Read the enabled-plugins allow-list from config.yaml.
+
+    Plugins are opt-in by default — only plugins whose name appears in
+    this set are loaded. Returns:
+
+    * ``None`` — the key is missing or malformed. Callers should treat
+      this as "nothing enabled yet" (the opt-in default); the first
+      ``migrate_config`` run populates the key with a grandfathered set
+      of currently-installed user plugins so existing setups don't
+      break on upgrade.
+    * ``set()`` — an empty list was explicitly set; nothing loads.
+    * ``set(...)`` — the concrete allow-list.
+    """
+    try:
+        from hermes_cli.config import load_config
+        config = load_config()
+        plugins_cfg = config.get("plugins")
+        if not isinstance(plugins_cfg, dict):
+            return None
+        if "enabled" not in plugins_cfg:
+            return None
+        enabled = plugins_cfg.get("enabled")
+        if not isinstance(enabled, list):
+            return None
+        return set(enabled)
+    except Exception:
+        return None
+
+
 # ---------------------------------------------------------------------------
 # Data classes
 # ---------------------------------------------------------------------------
@@ -420,27 +464,66 @@ class PluginManager:
 
         manifests: List[PluginManifest] = []
 
-        # 1. User plugins (~/.hermes/plugins/)
+        # 1. Bundled plugins (<repo>/plugins/<name>/)
+        # Repo-shipped generic plugins live next to hermes_cli/.  Memory and
+        # context_engine subdirs are handled by their own discovery paths, so
+        # skip those names here.  Bundled plugins are discovered (so they
+        # show up in `hermes plugins`) but only loaded when added to
+        # `plugins.enabled` in config.yaml — opt-in like any other plugin.
+        repo_plugins = Path(__file__).resolve().parent.parent / "plugins"
+        manifests.extend(
+            self._scan_directory(
+                repo_plugins,
+                source="bundled",
+                skip_names={"memory", "context_engine"},
+            )
+        )
+
+        # 2. User plugins (~/.hermes/plugins/)
         user_dir = get_hermes_home() / "plugins"
         manifests.extend(self._scan_directory(user_dir, source="user"))
 
-        # 2. Project plugins (./.hermes/plugins/)
+        # 3. Project plugins (./.hermes/plugins/)
         if _env_enabled("HERMES_ENABLE_PROJECT_PLUGINS"):
             project_dir = Path.cwd() / ".hermes" / "plugins"
             manifests.extend(self._scan_directory(project_dir, source="project"))
 
-        # 3. Pip / entry-point plugins
+        # 4. Pip / entry-point plugins
         manifests.extend(self._scan_entry_points())
 
-        # Load each manifest (skip user-disabled plugins)
+        # Load each manifest (skip user-disabled plugins).
+        # Later sources override earlier ones on name collision — user plugins
+        # take precedence over bundled, project plugins take precedence over
+        # user.  Dedup here so we only load the final winner.
         disabled = _get_disabled_plugins()
+        enabled = _get_enabled_plugins()  # None = opt-in default (nothing enabled)
+        winners: Dict[str, PluginManifest] = {}
         for manifest in manifests:
+            winners[manifest.name] = manifest
+        for manifest in winners.values():
+            # Explicit disable always wins.
             if manifest.name in disabled:
                 loaded = LoadedPlugin(manifest=manifest, enabled=False)
                 loaded.error = "disabled via config"
                 self._plugins[manifest.name] = loaded
                 logger.debug("Skipping disabled plugin '%s'", manifest.name)
                 continue
+            # Opt-in gate: plugins must be in the enabled allow-list.
+            # If the allow-list is missing (None), treat as "nothing enabled"
+            # — users have to explicitly enable plugins to load them.
+            # Memory and context_engine providers are excluded from this gate
+            # since they have their own single-select config (memory.provider
+            # / context.engine), not the enabled list.
+            if enabled is None or manifest.name not in enabled:
+                loaded = LoadedPlugin(manifest=manifest, enabled=False)
+                loaded.error = "not enabled in config (run `hermes plugins enable {}` to activate)".format(
+                    manifest.name
+                )
+                self._plugins[manifest.name] = loaded
+                logger.debug(
+                    "Skipping '%s' (not in plugins.enabled)", manifest.name
+                )
+                continue
             self._load_plugin(manifest)
 
         if manifests:
@@ -454,8 +537,18 @@ class PluginManager:
     # Directory scanning
     # -----------------------------------------------------------------------
 
-    def _scan_directory(self, path: Path, source: str) -> List[PluginManifest]:
-        """Read ``plugin.yaml`` manifests from subdirectories of *path*."""
+    def _scan_directory(
+        self,
+        path: Path,
+        source: str,
+        skip_names: Optional[Set[str]] = None,
+    ) -> List[PluginManifest]:
+        """Read ``plugin.yaml`` manifests from subdirectories of *path*.
+
+        *skip_names* is an optional allow-list of names to ignore (used
+        for the bundled scan to exclude ``memory`` / ``context_engine``
+        subdirs that have their own discovery path).
+        """
         manifests: List[PluginManifest] = []
         if not path.is_dir():
             return manifests
@@ -463,6 +556,8 @@ class PluginManager:
         for child in sorted(path.iterdir()):
             if not child.is_dir():
                 continue
+            if skip_names and child.name in skip_names:
+                continue
             manifest_file = child / "plugin.yaml"
             if not manifest_file.exists():
                 manifest_file = child / "plugin.yml"
@@ -530,7 +625,7 @@ class PluginManager:
         loaded = LoadedPlugin(manifest=manifest)
 
         try:
-            if manifest.source in ("user", "project"):
+            if manifest.source in ("user", "project", "bundled"):
                 module = self._load_directory_module(manifest)
             else:
                 module = self._load_entrypoint_module(manifest)
@@ -779,23 +874,31 @@ def get_pre_tool_call_block_message(
     return None
 
 
+def _ensure_plugins_discovered() -> PluginManager:
+    """Return the global manager after running idempotent plugin discovery."""
+    manager = get_plugin_manager()
+    manager.discover_and_load()
+    return manager
+
+
 def get_plugin_context_engine():
     """Return the plugin-registered context engine, or None."""
-    return get_plugin_manager()._context_engine
+    return _ensure_plugins_discovered()._context_engine
 
 
 def get_plugin_command_handler(name: str) -> Optional[Callable]:
     """Return the handler for a plugin-registered slash command, or ``None``."""
-    entry = get_plugin_manager()._plugin_commands.get(name)
+    entry = _ensure_plugins_discovered()._plugin_commands.get(name)
     return entry["handler"] if entry else None
 
 
 def get_plugin_commands() -> Dict[str, dict]:
     """Return the full plugin commands dict (name → {handler, description, plugin}).
 
-    Safe to call before discovery — returns an empty dict if no plugins loaded.
+    Triggers idempotent plugin discovery so callers can use plugin commands
+    before any explicit discover_plugins() call.
     """
-    return get_plugin_manager()._plugin_commands
+    return _ensure_plugins_discovered()._plugin_commands
 
 
 def get_plugin_toolsets() -> List[tuple]:

hermes_cli/plugins_cmd.py

@@ -15,6 +15,7 @@ import shutil
 import subprocess
 import sys
 from pathlib import Path
+from typing import Optional
 
 from hermes_constants import get_hermes_home
 
@@ -281,8 +282,16 @@ def _require_installed_plugin(name: str, plugins_dir: Path, console) -> Path:
 # ---------------------------------------------------------------------------
 
 
-def cmd_install(identifier: str, force: bool = False) -> None:
-    """Install a plugin from a Git URL or owner/repo shorthand."""
+def cmd_install(
+    identifier: str,
+    force: bool = False,
+    enable: Optional[bool] = None,
+) -> None:
+    """Install a plugin from a Git URL or owner/repo shorthand.
+
+    After install, prompt "Enable now? [y/N]" unless *enable* is provided
+    (True = auto-enable without prompting, False = install disabled).
+    """
     import tempfile
     from rich.console import Console
 
@@ -391,6 +400,40 @@ def cmd_install(identifier: str, force: bool = False) -> None:
 
     _display_after_install(target, identifier)
 
+    # Determine the canonical plugin name for enable-list bookkeeping.
+    installed_name = installed_manifest.get("name") or target.name
+
+    # Decide whether to enable: explicit flag > interactive prompt > default off
+    should_enable = enable
+    if should_enable is None:
+        # Interactive prompt unless stdin isn't a TTY (scripted install).
+        if sys.stdin.isatty() and sys.stdout.isatty():
+            try:
+                answer = input(
+                    f"  Enable '{installed_name}' now? [y/N]: "
+                ).strip().lower()
+                should_enable = answer in ("y", "yes")
+            except (EOFError, KeyboardInterrupt):
+                should_enable = False
+        else:
+            should_enable = False
+
+    if should_enable:
+        enabled = _get_enabled_set()
+        disabled = _get_disabled_set()
+        enabled.add(installed_name)
+        disabled.discard(installed_name)
+        _save_enabled_set(enabled)
+        _save_disabled_set(disabled)
+        console.print(
+            f"[green]✓[/green] Plugin [bold]{installed_name}[/bold] enabled."
+        )
+    else:
+        console.print(
+            f"[dim]Plugin installed but not enabled. "
+            f"Run `hermes plugins enable {installed_name}` to activate.[/dim]"
+        )
+
     console.print("[dim]Restart the gateway for the plugin to take effect:[/dim]")
     console.print("[dim]  hermes gateway restart[/dim]")
     console.print()
@@ -468,7 +511,11 @@ def cmd_remove(name: str) -> None:
 
 
 def _get_disabled_set() -> set:
-    """Read the disabled plugins set from config.yaml."""
+    """Read the disabled plugins set from config.yaml.
+
+    An explicit deny-list. A plugin name here never loads, even if also
+    listed in ``plugins.enabled``.
+    """
     try:
         from hermes_cli.config import load_config
         config = load_config()
@@ -488,103 +535,196 @@ def _save_disabled_set(disabled: set) -> None:
     save_config(config)
 
 
+def _get_enabled_set() -> set:
+    """Read the enabled plugins allow-list from config.yaml.
+
+    Plugins are opt-in: only names here are loaded. Returns ``set()`` if
+    the key is missing (same behaviour as "nothing enabled yet").
+    """
+    try:
+        from hermes_cli.config import load_config
+        config = load_config()
+        plugins_cfg = config.get("plugins", {})
+        if not isinstance(plugins_cfg, dict):
+            return set()
+        enabled = plugins_cfg.get("enabled", [])
+        return set(enabled) if isinstance(enabled, list) else set()
+    except Exception:
+        return set()
+
+
+def _save_enabled_set(enabled: set) -> None:
+    """Write the enabled plugins list to config.yaml."""
+    from hermes_cli.config import load_config, save_config
+    config = load_config()
+    if "plugins" not in config:
+        config["plugins"] = {}
+    config["plugins"]["enabled"] = sorted(enabled)
+    save_config(config)
+
+
 def cmd_enable(name: str) -> None:
-    """Enable a previously disabled plugin."""
+    """Add a plugin to the enabled allow-list (and remove it from disabled)."""
     from rich.console import Console
 
     console = Console()
-    plugins_dir = _plugins_dir()
-
-    # Verify the plugin exists
-    target = plugins_dir / name
-    if not target.is_dir():
-        console.print(f"[red]Plugin '{name}' is not installed.[/red]")
+    # Discover the plugin — check installed (user) AND bundled.
+    if not _plugin_exists(name):
+        console.print(f"[red]Plugin '{name}' is not installed or bundled.[/red]")
         sys.exit(1)
 
+    enabled = _get_enabled_set()
     disabled = _get_disabled_set()
-    if name not in disabled:
+
+    if name in enabled and name not in disabled:
         console.print(f"[dim]Plugin '{name}' is already enabled.[/dim]")
         return
 
+    enabled.add(name)
     disabled.discard(name)
+    _save_enabled_set(enabled)
     _save_disabled_set(disabled)
-    console.print(f"[green]✓[/green] Plugin [bold]{name}[/bold] enabled. Takes effect on next session.")
+    console.print(
+        f"[green]✓[/green] Plugin [bold]{name}[/bold] enabled. "
+        "Takes effect on next session."
+    )
 
 
 def cmd_disable(name: str) -> None:
-    """Disable a plugin without removing it."""
+    """Remove a plugin from the enabled allow-list (and add to disabled)."""
     from rich.console import Console
 
     console = Console()
-    plugins_dir = _plugins_dir()
-
-    # Verify the plugin exists
-    target = plugins_dir / name
-    if not target.is_dir():
-        console.print(f"[red]Plugin '{name}' is not installed.[/red]")
+    if not _plugin_exists(name):
+        console.print(f"[red]Plugin '{name}' is not installed or bundled.[/red]")
         sys.exit(1)
 
+    enabled = _get_enabled_set()
     disabled = _get_disabled_set()
-    if name in disabled:
+
+    if name not in enabled and name in disabled:
         console.print(f"[dim]Plugin '{name}' is already disabled.[/dim]")
         return
 
+    enabled.discard(name)
     disabled.add(name)
+    _save_enabled_set(enabled)
     _save_disabled_set(disabled)
-    console.print(f"[yellow]\u2298[/yellow] Plugin [bold]{name}[/bold] disabled. Takes effect on next session.")
+    console.print(
+        f"[yellow]\u2298[/yellow] Plugin [bold]{name}[/bold] disabled. "
+        "Takes effect on next session."
+    )
 
 
-def cmd_list() -> None:
-    """List installed plugins."""
-    from rich.console import Console
-    from rich.table import Table
+def _plugin_exists(name: str) -> bool:
+    """Return True if a plugin with *name* is installed (user) or bundled."""
+    # Installed: directory name or manifest name match in user plugins dir
+    user_dir = _plugins_dir()
+    if user_dir.is_dir():
+        if (user_dir / name).is_dir():
+            return True
+        for child in user_dir.iterdir():
+            if not child.is_dir():
+                continue
+            manifest = _read_manifest(child)
+            if manifest.get("name") == name:
+                return True
+    # Bundled: <repo>/plugins/<name>/
+    from pathlib import Path as _P
+    import hermes_cli
+    repo_plugins = _P(hermes_cli.__file__).resolve().parent.parent / "plugins"
+    if repo_plugins.is_dir():
+        candidate = repo_plugins / name
+        if candidate.is_dir() and (
+            (candidate / "plugin.yaml").exists()
+            or (candidate / "plugin.yml").exists()
+        ):
+            return True
+    return False
 
+
+def _discover_all_plugins() -> list:
+    """Return a list of (name, version, description, source, dir_path) for
+    every plugin the loader can see — user + bundled + project.
+
+    Matches the ordering/dedup of ``PluginManager.discover_and_load``:
+    bundled first, then user, then project; user overrides bundled on
+    name collision.
+    """
     try:
         import yaml
     except ImportError:
         yaml = None
 
-    console = Console()
-    plugins_dir = _plugins_dir()
+    seen: dict = {}  # name -> (name, version, description, source, path)
 
-    dirs = sorted(d for d in plugins_dir.iterdir() if d.is_dir())
-    if not dirs:
+    # Bundled (<repo>/plugins/<name>/), excluding memory/ and context_engine/
+    import hermes_cli
+    repo_plugins = Path(hermes_cli.__file__).resolve().parent.parent / "plugins"
+    for base, source in ((repo_plugins, "bundled"), (_plugins_dir(), "user")):
+        if not base.is_dir():
+            continue
+        for d in sorted(base.iterdir()):
+            if not d.is_dir():
+                continue
+            if source == "bundled" and d.name in ("memory", "context_engine"):
+                continue
+            manifest_file = d / "plugin.yaml"
+            if not manifest_file.exists():
+                manifest_file = d / "plugin.yml"
+            if not manifest_file.exists():
+                continue
+            name = d.name
+            version = ""
+            description = ""
+            if yaml:
+                try:
+                    with open(manifest_file) as f:
+                        manifest = yaml.safe_load(f) or {}
+                    name = manifest.get("name", d.name)
+                    version = manifest.get("version", "")
+                    description = manifest.get("description", "")
+                except Exception:
+                    pass
+            # User plugins override bundled on name collision.
+            if name in seen and source == "bundled":
+                continue
+            src_label = source
+            if source == "user" and (d / ".git").exists():
+                src_label = "git"
+            seen[name] = (name, version, description, src_label, d)
+    return list(seen.values())
+
+
+def cmd_list() -> None:
+    """List all plugins (bundled + user) with enabled/disabled state."""
+    from rich.console import Console
+    from rich.table import Table
+
+    console = Console()
+    entries = _discover_all_plugins()
+    if not entries:
         console.print("[dim]No plugins installed.[/dim]")
         console.print("[dim]Install with:[/dim] hermes plugins install owner/repo")
         return
 
+    enabled = _get_enabled_set()
     disabled = _get_disabled_set()
 
-    table = Table(title="Installed Plugins", show_lines=False)
+    table = Table(title="Plugins", show_lines=False)
     table.add_column("Name", style="bold")
     table.add_column("Status")
     table.add_column("Version", style="dim")
     table.add_column("Description")
     table.add_column("Source", style="dim")
 
-    for d in dirs:
-        manifest_file = d / "plugin.yaml"
-        name = d.name
-        version = ""
-        description = ""
-        source = "local"
-
-        if manifest_file.exists() and yaml:
-            try:
-                with open(manifest_file) as f:
-                    manifest = yaml.safe_load(f) or {}
-                name = manifest.get("name", d.name)
-                version = manifest.get("version", "")
-                description = manifest.get("description", "")
-            except Exception:
-                pass
-
-        # Check if it's a git repo (installed via hermes plugins install)
-        if (d / ".git").exists():
-            source = "git"
-
-        is_disabled = name in disabled or d.name in disabled
-        status = "[red]disabled[/red]" if is_disabled else "[green]enabled[/green]"
+    for name, version, description, source, _dir in entries:
+        if name in disabled:
+            status = "[red]disabled[/red]"
+        elif name in enabled:
+            status = "[green]enabled[/green]"
+        else:
+            status = "[yellow]not enabled[/yellow]"
         table.add_row(name, status, str(version), description, source)
 
     console.print()
@@ -592,6 +732,7 @@ def cmd_list() -> None:
     console.print()
     console.print("[dim]Interactive toggle:[/dim] hermes plugins")
     console.print("[dim]Enable/disable:[/dim] hermes plugins enable/disable <name>")
+    console.print("[dim]Plugins are opt-in by default — only 'enabled' plugins load.[/dim]")
 
 
 # ---------------------------------------------------------------------------
@@ -742,41 +883,25 @@ def cmd_toggle() -> None:
     """Interactive composite UI — general plugins + provider plugin categories."""
     from rich.console import Console
 
-    try:
-        import yaml
-    except ImportError:
-        yaml = None
-
     console = Console()
-    plugins_dir = _plugins_dir()
 
-    # -- General plugins discovery --
-    dirs = sorted(d for d in plugins_dir.iterdir() if d.is_dir())
-    disabled = _get_disabled_set()
+    # -- General plugins discovery (bundled + user) --
+    entries = _discover_all_plugins()
+    enabled_set = _get_enabled_set()
+    disabled_set = _get_disabled_set()
 
     plugin_names = []
     plugin_labels = []
     plugin_selected = set()
 
-    for i, d in enumerate(dirs):
-        manifest_file = d / "plugin.yaml"
-        name = d.name
-        description = ""
-
-        if manifest_file.exists() and yaml:
-            try:
-                with open(manifest_file) as f:
-                    manifest = yaml.safe_load(f) or {}
-                name = manifest.get("name", d.name)
-                description = manifest.get("description", "")
-            except Exception:
-                pass
-
-        plugin_names.append(name)
+    for i, (name, _version, description, source, _d) in enumerate(entries):
         label = f"{name} \u2014 {description}" if description else name
+        if source == "bundled":
+            label = f"{label} [bundled]"
+        plugin_names.append(name)
         plugin_labels.append(label)
-
-        if name not in disabled and d.name not in disabled:
+        # Selected (enabled) when in enabled-set AND not in disabled-set
+        if name in enabled_set and name not in disabled_set:
             plugin_selected.add(i)
 
     # -- Provider categories --
@@ -804,10 +929,10 @@ def cmd_toggle() -> None:
     try:
         import curses
         _run_composite_ui(curses, plugin_names, plugin_labels, plugin_selected,
-                          disabled, categories, console)
+                          disabled_set, categories, console)
     except ImportError:
         _run_composite_fallback(plugin_names, plugin_labels, plugin_selected,
-                                disabled, categories, console)
+                                disabled_set, categories, console)
 
 
 def _run_composite_ui(curses, plugin_names, plugin_labels, plugin_selected,
@@ -1020,18 +1145,29 @@ def _run_composite_ui(curses, plugin_names, plugin_labels, plugin_selected,
     curses.wrapper(_draw)
     flush_stdin()
 
-    # Persist general plugin changes
-    new_disabled = set()
+    # Persist general plugin changes. The new allow-list is the set of
+    # plugin names that were checked; anything not checked is explicitly
+    # disabled (written to disabled-list) so it remains off even if the
+    # plugin code does something clever like auto-enable in the future.
+    new_enabled: set = set()
+    new_disabled: set = set(disabled)  # preserve existing disabled state for unseen plugins
     for i, name in enumerate(plugin_names):
-        if i not in chosen:
+        if i in chosen:
+            new_enabled.add(name)
+            new_disabled.discard(name)
+        else:
             new_disabled.add(name)
 
-    if new_disabled != disabled:
+    prev_enabled = _get_enabled_set()
+    enabled_changed = new_enabled != prev_enabled
+    disabled_changed = new_disabled != disabled
+
+    if enabled_changed or disabled_changed:
+        _save_enabled_set(new_enabled)
         _save_disabled_set(new_disabled)
-        enabled_count = len(plugin_names) - len(new_disabled)
         console.print(
-            f"\n[green]\u2713[/green] General plugins: {enabled_count} enabled, "
-            f"{len(new_disabled)} disabled."
+            f"\n[green]\u2713[/green] General plugins: {len(new_enabled)} enabled, "
+            f"{len(plugin_names) - len(new_enabled)} disabled."
         )
     elif n_plugins > 0:
         console.print("\n[dim]General plugins unchanged.[/dim]")
@@ -1078,11 +1214,17 @@ def _run_composite_fallback(plugin_names, plugin_labels, plugin_selected,
                 return
             print()
 
-        new_disabled = set()
+        new_enabled: set = set()
+        new_disabled: set = set(disabled)
         for i, name in enumerate(plugin_names):
-            if i not in chosen:
+            if i in chosen:
+                new_enabled.add(name)
+                new_disabled.discard(name)
+            else:
                 new_disabled.add(name)
-        if new_disabled != disabled:
+        prev_enabled = _get_enabled_set()
+        if new_enabled != prev_enabled or new_disabled != disabled:
+            _save_enabled_set(new_enabled)
             _save_disabled_set(new_disabled)
 
     # Provider categories
@@ -1108,7 +1250,17 @@ def plugins_command(args) -> None:
     action = getattr(args, "plugins_action", None)
 
     if action == "install":
-        cmd_install(args.identifier, force=getattr(args, "force", False))
+        # Map argparse tri-state: --enable=True, --no-enable=False, neither=None (prompt)
+        enable_arg = None
+        if getattr(args, "enable", False):
+            enable_arg = True
+        elif getattr(args, "no_enable", False):
+            enable_arg = False
+        cmd_install(
+            args.identifier,
+            force=getattr(args, "force", False),
+            enable=enable_arg,
+        )
     elif action == "update":
         cmd_update(args.name)
     elif action in ("remove", "rm", "uninstall"):

hermes_cli/providers.py

@@ -23,6 +23,8 @@ import logging
 from dataclasses import dataclass
 from typing import Any, Dict, List, Optional, Tuple
 
+from utils import base_url_host_matches, base_url_hostname
+
 logger = logging.getLogger(__name__)
 
 
@@ -322,12 +324,16 @@ def normalize_provider(name: str) -> str:
 
 
 def get_provider(name: str) -> Optional[ProviderDef]:
-    """Look up a provider by id or alias, merging all data sources.
+    """Look up a built-in provider by id or alias.
 
     Resolution order:
       1. Hermes overlays (for providers not in models.dev: nous, openai-codex, etc.)
       2. models.dev catalog + Hermes overlay
-      3. User-defined providers from config (TODO: Phase 4)
+
+    User-defined providers from config.yaml (``providers:`` / ``custom_providers:``)
+    are resolved by :func:`resolve_provider_full`, which layers ``resolve_user_provider``
+    and ``resolve_custom_provider`` on top of this function. Callers that need
+    user-config support should use ``resolve_provider_full`` instead.
 
     Returns a fully-resolved ProviderDef or None.
     """
@@ -430,11 +436,12 @@ def determine_api_mode(provider: str, base_url: str = "") -> str:
     # URL-based heuristics for custom / unknown providers
     if base_url:
         url_lower = base_url.rstrip("/").lower()
-        if url_lower.endswith("/anthropic") or "api.anthropic.com" in url_lower:
+        hostname = base_url_hostname(base_url)
+        if url_lower.endswith("/anthropic") or hostname == "api.anthropic.com":
             return "anthropic_messages"
-        if "api.openai.com" in url_lower:
+        if hostname == "api.openai.com":
             return "codex_responses"
-        if "bedrock-runtime" in url_lower and "amazonaws.com" in url_lower:
+        if hostname.startswith("bedrock-runtime.") and base_url_host_matches(base_url, "amazonaws.com"):
             return "bedrock_converse"
 
     return "chat_completions"

hermes_cli/runtime_provider.py

@@ -29,6 +29,7 @@ from hermes_cli.auth import (
 )
 from hermes_cli.config import get_compatible_custom_providers, load_config
 from hermes_constants import OPENROUTER_BASE_URL
+from utils import base_url_host_matches, base_url_hostname
 
 
 def _normalize_custom_provider_name(value: str) -> str:
@@ -38,14 +39,22 @@ def _normalize_custom_provider_name(value: str) -> str:
 def _detect_api_mode_for_url(base_url: str) -> Optional[str]:
     """Auto-detect api_mode from the resolved base URL.
 
-    Direct api.openai.com endpoints need the Responses API for GPT-5.x
-    tool calls with reasoning (chat/completions returns 400).
+    - Direct api.openai.com endpoints need the Responses API for GPT-5.x
+      tool calls with reasoning (chat/completions returns 400).
+    - Third-party Anthropic-compatible gateways (MiniMax, Zhipu GLM,
+      LiteLLM proxies, etc.) conventionally expose the native Anthropic
+      protocol under a ``/anthropic`` suffix — treat those as
+      ``anthropic_messages`` transport instead of the default
+      ``chat_completions``.
     """
     normalized = (base_url or "").strip().lower().rstrip("/")
-    if "api.x.ai" in normalized:
+    hostname = base_url_hostname(base_url)
+    if hostname == "api.x.ai":
         return "codex_responses"
-    if "api.openai.com" in normalized and "openrouter" not in normalized:
+    if hostname == "api.openai.com":
         return "codex_responses"
+    if normalized.endswith("/anthropic"):
+        return "anthropic_messages"
     return None
 
 
@@ -194,8 +203,12 @@ def _resolve_runtime_from_pool_entry(
         elif provider in ("opencode-zen", "opencode-go"):
             from hermes_cli.models import opencode_model_api_mode
             api_mode = opencode_model_api_mode(provider, model_cfg.get("default", ""))
-        elif base_url.rstrip("/").endswith("/anthropic"):
-            api_mode = "anthropic_messages"
+        else:
+            # Auto-detect Anthropic-compatible endpoints (/anthropic suffix,
+            # api.openai.com → codex_responses, api.x.ai → codex_responses).
+            detected = _detect_api_mode_for_url(base_url)
+            if detected:
+                api_mode = detected
 
     # OpenCode base URLs end with /v1 for OpenAI-compatible models, but the
     # Anthropic SDK prepends its own /v1/messages to the base_url.  Strip the
@@ -469,7 +482,7 @@ def _resolve_openrouter_runtime(
     # When hitting a custom endpoint (e.g. Z.ai, local LLM), prefer
     # OPENAI_API_KEY so the OpenRouter key doesn't leak to an unrelated
     # provider (issues #420, #560).
-    _is_openrouter_url = "openrouter.ai" in base_url
+    _is_openrouter_url = base_url_host_matches(base_url, "openrouter.ai")
     if _is_openrouter_url:
         api_key_candidates = [
             explicit_api_key,
@@ -479,8 +492,12 @@ def _resolve_openrouter_runtime(
     else:
         # Custom endpoint: use api_key from config when using config base_url (#1760).
         # When the endpoint is Ollama Cloud, check OLLAMA_API_KEY — it's
-        # the canonical env var for ollama.com authentication.
-        _is_ollama_url = "ollama.com" in base_url.lower()
+        # the canonical env var for ollama.com authentication. Match on
+        # HOST, not substring — a custom base_url whose path contains
+        # "ollama.com" (e.g. http://127.0.0.1/ollama.com/v1) or whose
+        # hostname is a look-alike (ollama.com.attacker.test) must not
+        # receive the Ollama credential. See GHSA-76xc-57q6-vm5m.
+        _is_ollama_url = base_url_host_matches(base_url, "ollama.com")
         api_key_candidates = [
             explicit_api_key,
             (cfg_api_key if use_config_base_url else ""),
@@ -642,8 +659,11 @@ def _resolve_explicit_runtime(
             configured_mode = _parse_api_mode(model_cfg.get("api_mode"))
             if configured_mode:
                 api_mode = configured_mode
-            elif base_url.rstrip("/").endswith("/anthropic"):
-                api_mode = "anthropic_messages"
+            else:
+                # Auto-detect Anthropic-compatible endpoints (/anthropic suffix).
+                detected = _detect_api_mode_for_url(base_url)
+                if detected:
+                    api_mode = detected
 
         return {
             "provider": provider,
@@ -890,8 +910,7 @@ def resolve_runtime_provider(
                 code="no_aws_credentials",
             )
         # Read bedrock-specific config from config.yaml
-        from hermes_cli.config import load_config as _load_bedrock_config
-        _bedrock_cfg = _load_bedrock_config().get("bedrock", {})
+        _bedrock_cfg = load_config().get("bedrock", {})
         # Region priority: config.yaml bedrock.region → env var → us-east-1
         region = (_bedrock_cfg.get("region") or "").strip() or resolve_bedrock_region()
         auth_source = resolve_aws_auth_env_var() or "aws-sdk-default-chain"
@@ -965,10 +984,13 @@ def resolve_runtime_provider(
             elif provider in ("opencode-zen", "opencode-go"):
                 from hermes_cli.models import opencode_model_api_mode
                 api_mode = opencode_model_api_mode(provider, model_cfg.get("default", ""))
-            # Auto-detect Anthropic-compatible endpoints by URL convention
-            # (e.g. https://api.minimax.io/anthropic, https://dashscope.../anthropic)
-            elif base_url.rstrip("/").endswith("/anthropic"):
-                api_mode = "anthropic_messages"
+            else:
+                # Auto-detect Anthropic-compatible endpoints by URL convention
+                # (e.g. https://api.minimax.io/anthropic, https://dashscope.../anthropic)
+                # plus api.openai.com → codex_responses and api.x.ai → codex_responses.
+                detected = _detect_api_mode_for_url(base_url)
+                if detected:
+                    api_mode = detected
         # Strip trailing /v1 for OpenCode Anthropic models (see comment above).
         if api_mode == "anthropic_messages" and provider in ("opencode-zen", "opencode-go"):
             base_url = re.sub(r"/v1/?$", "", base_url)

hermes_cli/setup.py

@@ -22,6 +22,7 @@ from typing import Optional, Dict, Any
 
 from hermes_cli.nous_subscription import get_nous_subscription_features
 from tools.tool_backend_helpers import managed_nous_tools_enabled
+from utils import base_url_hostname
 from hermes_constants import get_optional_skills_dir
 
 logger = logging.getLogger(__name__)
@@ -89,19 +90,19 @@ _DEFAULT_PROVIDER_MODELS = {
         "grok-code-fast-1",
     ],
     "gemini": [
-        "gemini-3.1-pro-preview", "gemini-3-flash-preview", "gemini-3.1-flash-lite-preview",
-        "gemini-2.5-pro", "gemini-2.5-flash", "gemini-2.5-flash-lite",
+        "gemini-3.1-pro-preview", "gemini-3-pro-preview",
+        "gemini-3-flash-preview", "gemini-3.1-flash-lite-preview",
     ],
     "zai": ["glm-5.1", "glm-5", "glm-4.7", "glm-4.5", "glm-4.5-flash"],
-    "kimi-coding": ["kimi-k2.5", "kimi-k2-thinking", "kimi-k2-turbo-preview"],
-    "kimi-coding-cn": ["kimi-k2.5", "kimi-k2-thinking", "kimi-k2-turbo-preview"],
+    "kimi-coding": ["kimi-k2.6", "kimi-k2.5", "kimi-k2-thinking", "kimi-k2-turbo-preview"],
+    "kimi-coding-cn": ["kimi-k2.6", "kimi-k2.5", "kimi-k2-thinking", "kimi-k2-turbo-preview"],
     "arcee": ["trinity-large-thinking", "trinity-large-preview", "trinity-mini"],
     "minimax": ["MiniMax-M2.7", "MiniMax-M2.5", "MiniMax-M2.1", "MiniMax-M2"],
     "minimax-cn": ["MiniMax-M2.7", "MiniMax-M2.5", "MiniMax-M2.1", "MiniMax-M2"],
     "ai-gateway": ["anthropic/claude-opus-4.6", "anthropic/claude-sonnet-4.6", "openai/gpt-5", "google/gemini-3-flash"],
     "kilocode": ["anthropic/claude-opus-4.6", "anthropic/claude-sonnet-4.6", "openai/gpt-5.4", "google/gemini-3-pro-preview", "google/gemini-3-flash-preview"],
     "opencode-zen": ["gpt-5.4", "gpt-5.3-codex", "claude-sonnet-4-6", "gemini-3-flash", "glm-5", "kimi-k2.5", "minimax-m2.7"],
-    "opencode-go": ["glm-5.1", "glm-5", "kimi-k2.5", "mimo-v2-pro", "mimo-v2-omni", "minimax-m2.5", "minimax-m2.7"],
+    "opencode-go": ["kimi-k2.6", "kimi-k2.5", "glm-5.1", "glm-5", "mimo-v2-pro", "mimo-v2-omni", "minimax-m2.5", "minimax-m2.7", "qwen3.6-plus", "qwen3.5-plus"],
     "huggingface": [
         "Qwen/Qwen3.5-397B-A17B", "Qwen/Qwen3-235B-A22B-Thinking-2507",
         "Qwen/Qwen3-Coder-480B-A35B-Instruct", "deepseek-ai/DeepSeek-R1-0528",
@@ -433,7 +434,6 @@ def _print_setup_summary(config: dict, hermes_home):
         tool_status.append(("Text-to-Speech (Google Gemini)", True, None))
     elif tts_provider == "neutts":
         try:
-            import importlib.util
             neutts_ok = importlib.util.find_spec("neutts") is not None
         except Exception:
             neutts_ok = False
@@ -441,6 +441,16 @@ def _print_setup_summary(config: dict, hermes_home):
             tool_status.append(("Text-to-Speech (NeuTTS local)", True, None))
         else:
             tool_status.append(("Text-to-Speech (NeuTTS — not installed)", False, "run 'hermes setup tts'"))
+    elif tts_provider == "kittentts":
+        try:
+            import importlib.util
+            kittentts_ok = importlib.util.find_spec("kittentts") is not None
+        except Exception:
+            kittentts_ok = False
+        if kittentts_ok:
+            tool_status.append(("Text-to-Speech (KittenTTS local)", True, None))
+        else:
+            tool_status.append(("Text-to-Speech (KittenTTS — not installed)", False, "run 'hermes setup tts'"))
     else:
         tool_status.append(("Text-to-Speech (Edge TTS)", True, None))
 
@@ -803,7 +813,8 @@ def setup_model_provider(config: dict, *, quick: bool = False):
         elif _vision_idx == 1:  # OpenAI-compatible endpoint
             _base_url = prompt("  Base URL (blank for OpenAI)").strip() or "https://api.openai.com/v1"
             _api_key_label = "  API key"
-            if "api.openai.com" in _base_url.lower():
+            _is_native_openai = base_url_hostname(_base_url) == "api.openai.com"
+            if _is_native_openai:
                 _api_key_label = "  OpenAI API key"
             _oai_key = prompt(_api_key_label, password=True).strip()
             if _oai_key:
@@ -811,7 +822,7 @@ def setup_model_provider(config: dict, *, quick: bool = False):
                 # Save vision base URL to config (not .env — only secrets go there)
                 _vaux = config.setdefault("auxiliary", {}).setdefault("vision", {})
                 _vaux["base_url"] = _base_url
-                if "api.openai.com" in _base_url.lower():
+                if _is_native_openai:
                     _oai_vision_models = ["gpt-4o", "gpt-4o-mini", "gpt-4.1", "gpt-4.1-mini", "gpt-4.1-nano"]
                     _vm_choices = _oai_vision_models + ["Use default (gpt-4o-mini)"]
                     _vm_idx = prompt_choice("Select vision model:", _vm_choices, 0)
@@ -847,7 +858,6 @@ def setup_model_provider(config: dict, *, quick: bool = False):
 
 def _check_espeak_ng() -> bool:
     """Check if espeak-ng is installed."""
-    import shutil
     return shutil.which("espeak-ng") is not None or shutil.which("espeak") is not None
 
 
@@ -901,6 +911,31 @@ def _install_neutts_deps() -> bool:
         return False
 
 
+def _install_kittentts_deps() -> bool:
+    """Install KittenTTS dependencies with user approval. Returns True on success."""
+    import subprocess
+    import sys
+
+    wheel_url = (
+        "https://github.com/KittenML/KittenTTS/releases/download/"
+        "0.8.1/kittentts-0.8.1-py3-none-any.whl"
+    )
+    print()
+    print_info("Installing kittentts Python package (~25-80MB model downloaded on first use)...")
+    print()
+    try:
+        subprocess.run(
+            [sys.executable, "-m", "pip", "install", "-U", wheel_url, "soundfile", "--quiet"],
+            check=True, timeout=300,
+        )
+        print_success("kittentts installed successfully")
+        return True
+    except (subprocess.CalledProcessError, subprocess.TimeoutExpired) as e:
+        print_error(f"Failed to install kittentts: {e}")
+        print_info(f"Try manually: python -m pip install -U '{wheel_url}' soundfile")
+        return False
+
+
 def _setup_tts_provider(config: dict):
     """Interactive TTS provider selection with install flow for NeuTTS."""
     tts_config = config.get("tts", {})
@@ -916,6 +951,7 @@ def _setup_tts_provider(config: dict):
         "mistral": "Mistral Voxtral TTS",
         "gemini": "Google Gemini TTS",
         "neutts": "NeuTTS",
+        "kittentts": "KittenTTS",
     }
     current_label = provider_labels.get(current_provider, current_provider)
 
@@ -939,9 +975,10 @@ def _setup_tts_provider(config: dict):
             "Mistral Voxtral TTS (multilingual, native Opus, needs API key)",
             "Google Gemini TTS (30 prebuilt voices, prompt-controllable, needs API key)",
             "NeuTTS (local on-device, free, ~300MB model download)",
+            "KittenTTS (local on-device, free, lightweight ~25-80MB ONNX)",
         ]
     )
-    providers.extend(["edge", "elevenlabs", "openai", "xai", "minimax", "mistral", "gemini", "neutts"])
+    providers.extend(["edge", "elevenlabs", "openai", "xai", "minimax", "mistral", "gemini", "neutts", "kittentts"])
     choices.append(f"Keep current ({current_label})")
     keep_current_idx = len(choices) - 1
     idx = prompt_choice("Select TTS provider:", choices, keep_current_idx)
@@ -962,7 +999,6 @@ def _setup_tts_provider(config: dict):
     if selected == "neutts":
         # Check if already installed
         try:
-            import importlib.util
             already_installed = importlib.util.find_spec("neutts") is not None
         except Exception:
             already_installed = False
@@ -1061,6 +1097,29 @@ def _setup_tts_provider(config: dict):
                 print_warning("No API key provided. Falling back to Edge TTS.")
                 selected = "edge"
 
+    elif selected == "kittentts":
+        # Check if already installed
+        try:
+            import importlib.util
+            already_installed = importlib.util.find_spec("kittentts") is not None
+        except Exception:
+            already_installed = False
+
+        if already_installed:
+            print_success("KittenTTS is already installed")
+        else:
+            print()
+            print_info("KittenTTS is lightweight (~25-80MB, CPU-only, no API key required).")
+            print_info("Voices: Jasper, Bella, Luna, Bruno, Rosie, Hugo, Kiki, Leo")
+            print()
+            if prompt_yes_no("Install KittenTTS now?", True):
+                if not _install_kittentts_deps():
+                    print_warning("KittenTTS installation incomplete. Falling back to Edge TTS.")
+                    selected = "edge"
+            else:
+                print_info("Skipping install. Set tts.provider to 'kittentts' after installing manually.")
+                selected = "edge"
+
     # Save the selection
     if "tts" not in config:
         config["tts"] = {}
@@ -1082,8 +1141,6 @@ def setup_tts(config: dict):
 def setup_terminal_backend(config: dict):
     """Configure the terminal execution backend."""
     import platform as _platform
-    import shutil
-
     print_header("Terminal Backend")
     print_info("Choose where Hermes runs shell commands and code.")
     print_info("This affects tool execution, file access, and isolation.")
@@ -2358,6 +2415,74 @@ def setup_tools(config: dict, first_install: bool = False):
 # =============================================================================
 
 
+def _model_section_has_credentials(config: dict) -> bool:
+    """Return True when any known inference provider has usable credentials.
+
+    Sources of truth:
+      * ``PROVIDER_REGISTRY`` in ``hermes_cli.auth`` — lists every supported
+        provider along with its ``api_key_env_vars``.
+      * ``active_provider`` in the auth store — covers OAuth device-code /
+        external-OAuth providers (Nous, Codex, Qwen, Gemini CLI, ...).
+      * The legacy OpenRouter aggregator env vars, which route generic
+        ``OPENAI_API_KEY`` / ``OPENROUTER_API_KEY`` values through OpenRouter.
+    """
+    try:
+        from hermes_cli.auth import get_active_provider
+        if get_active_provider():
+            return True
+    except Exception:
+        pass
+
+    try:
+        from hermes_cli.auth import PROVIDER_REGISTRY
+    except Exception:
+        PROVIDER_REGISTRY = {}  # type: ignore[assignment]
+
+    def _has_key(pconfig) -> bool:
+        for env_var in pconfig.api_key_env_vars:
+            # CLAUDE_CODE_OAUTH_TOKEN is set by Claude Code itself, not by
+            # the user — mirrors is_provider_explicitly_configured in auth.py.
+            if env_var == "CLAUDE_CODE_OAUTH_TOKEN":
+                continue
+            if get_env_value(env_var):
+                return True
+        return False
+
+    # Prefer the provider declared in config.yaml, avoids false positives
+    # from stray env vars (GH_TOKEN, etc.) when the user has already picked
+    # a different provider.
+    model_cfg = config.get("model") if isinstance(config, dict) else None
+    if isinstance(model_cfg, dict):
+        provider_id = (model_cfg.get("provider") or "").strip().lower()
+        if provider_id in PROVIDER_REGISTRY:
+            if _has_key(PROVIDER_REGISTRY[provider_id]):
+                return True
+        if provider_id == "openrouter":
+            for env_var in ("OPENROUTER_API_KEY", "OPENAI_API_KEY"):
+                if get_env_value(env_var):
+                    return True
+
+    # OpenRouter aggregator fallback (no provider declared in config).
+    for env_var in ("OPENROUTER_API_KEY", "OPENAI_API_KEY"):
+        if get_env_value(env_var):
+            return True
+
+    for pid, pconfig in PROVIDER_REGISTRY.items():
+        # Skip copilot in auto-detect: GH_TOKEN / GITHUB_TOKEN are
+        # commonly set for git tooling.  Mirrors resolve_provider in auth.py.
+        if pid == "copilot":
+            continue
+        if _has_key(pconfig):
+            return True
+    return False
+
+
+def _gateway_platform_short_label(label: str) -> str:
+    """Strip trailing parenthetical qualifiers from a gateway platform label."""
+    base = label.split("(", 1)[0].strip()
+    return base or label
+
+
 def _get_section_config_summary(config: dict, section_key: str) -> Optional[str]:
     """Return a short summary if a setup section is already configured, else None.
 
@@ -2366,20 +2491,7 @@ def _get_section_config_summary(config: dict, section_key: str) -> Optional[str]
     so that test patches on ``setup_mod.get_env_value`` take effect.
     """
     if section_key == "model":
-        has_key = bool(
-            get_env_value("OPENROUTER_API_KEY")
-            or get_env_value("OPENAI_API_KEY")
-            or get_env_value("ANTHROPIC_API_KEY")
-        )
-        if not has_key:
-            # Check for OAuth providers
-            try:
-                from hermes_cli.auth import get_active_provider
-                if get_active_provider():
-                    has_key = True
-            except Exception:
-                pass
-        if not has_key:
+        if not _model_section_has_credentials(config):
             return None
         model = config.get("model")
         if isinstance(model, str) and model.strip():
@@ -2397,37 +2509,11 @@ def _get_section_config_summary(config: dict, section_key: str) -> Optional[str]
         return f"max turns: {max_turns}"
 
     elif section_key == "gateway":
-        platforms = []
-        if get_env_value("TELEGRAM_BOT_TOKEN"):
-            platforms.append("Telegram")
-        if get_env_value("DISCORD_BOT_TOKEN"):
-            platforms.append("Discord")
-        if get_env_value("SLACK_BOT_TOKEN"):
-            platforms.append("Slack")
-        if get_env_value("SIGNAL_ACCOUNT"):
-            platforms.append("Signal")
-        if get_env_value("EMAIL_ADDRESS"):
-            platforms.append("Email")
-        if get_env_value("TWILIO_ACCOUNT_SID"):
-            platforms.append("SMS")
-        if get_env_value("MATRIX_ACCESS_TOKEN") or get_env_value("MATRIX_PASSWORD"):
-            platforms.append("Matrix")
-        if get_env_value("MATTERMOST_TOKEN"):
-            platforms.append("Mattermost")
-        if get_env_value("WHATSAPP_PHONE_NUMBER_ID"):
-            platforms.append("WhatsApp")
-        if get_env_value("DINGTALK_CLIENT_ID"):
-            platforms.append("DingTalk")
-        if get_env_value("FEISHU_APP_ID"):
-            platforms.append("Feishu")
-        if get_env_value("WECOM_BOT_ID"):
-            platforms.append("WeCom")
-        if get_env_value("WEIXIN_ACCOUNT_ID"):
-            platforms.append("Weixin")
-        if get_env_value("BLUEBUBBLES_SERVER_URL"):
-            platforms.append("BlueBubbles")
-        if get_env_value("WEBHOOK_ENABLED"):
-            platforms.append("Webhooks")
+        platforms = [
+            _gateway_platform_short_label(label)
+            for label, env_var, _ in _GATEWAY_PLATFORMS
+            if get_env_value(env_var)
+        ]
         if platforms:
             return ", ".join(platforms)
         return None  # No platforms configured — section must run

hermes_cli/timeouts.py

@@ -0,0 +1,82 @@
+from __future__ import annotations
+
+
+def _coerce_timeout(raw: object) -> float | None:
+    try:
+        timeout = float(raw)
+    except (TypeError, ValueError):
+        return None
+    if timeout <= 0:
+        return None
+    return timeout
+
+
+def get_provider_request_timeout(
+    provider_id: str, model: str | None = None
+) -> float | None:
+    """Return a configured provider request timeout in seconds, if any."""
+    if not provider_id:
+        return None
+
+    try:
+        from hermes_cli.config import load_config
+    except ImportError:
+        return None
+
+    config = load_config()
+    providers = config.get("providers", {}) if isinstance(config, dict) else {}
+    provider_config = (
+        providers.get(provider_id, {}) if isinstance(providers, dict) else {}
+    )
+    if not isinstance(provider_config, dict):
+        return None
+
+    model_config = _get_model_config(provider_config, model)
+    if model_config is not None:
+        timeout = _coerce_timeout(model_config.get("timeout_seconds"))
+        if timeout is not None:
+            return timeout
+
+    return _coerce_timeout(provider_config.get("request_timeout_seconds"))
+
+
+def get_provider_stale_timeout(
+    provider_id: str, model: str | None = None
+) -> float | None:
+    """Return a configured non-stream stale timeout in seconds, if any."""
+    if not provider_id:
+        return None
+
+    try:
+        from hermes_cli.config import load_config
+    except ImportError:
+        return None
+
+    config = load_config()
+    providers = config.get("providers", {}) if isinstance(config, dict) else {}
+    provider_config = (
+        providers.get(provider_id, {}) if isinstance(providers, dict) else {}
+    )
+    if not isinstance(provider_config, dict):
+        return None
+
+    model_config = _get_model_config(provider_config, model)
+    if model_config is not None:
+        timeout = _coerce_timeout(model_config.get("stale_timeout_seconds"))
+        if timeout is not None:
+            return timeout
+
+    return _coerce_timeout(provider_config.get("stale_timeout_seconds"))
+
+
+def _get_model_config(
+    provider_config: dict[str, object], model: str | None
+) -> dict[str, object] | None:
+    if not model:
+        return None
+
+    models = provider_config.get("models", {})
+    model_config = models.get(model, {}) if isinstance(models, dict) else {}
+    if isinstance(model_config, dict):
+        return model_config
+    return None

hermes_cli/tips.py

@@ -127,7 +127,7 @@ TIPS = [
 
     # --- Tools & Capabilities ---
     "execute_code runs Python scripts that call Hermes tools programmatically — results stay out of context.",
-    "delegate_task spawns up to 3 concurrent sub-agents with isolated contexts for parallel work.",
+    "delegate_task spawns up to 3 concurrent sub-agents by default (configurable via delegation.max_concurrent_children) with isolated contexts for parallel work.",
     "web_extract works on PDF URLs — pass any PDF link and it converts to markdown.",
     "search_files is ripgrep-backed and faster than grep — use it instead of terminal grep.",
     "patch uses 9 fuzzy matching strategies so minor whitespace differences won't break edits.",
@@ -245,7 +245,7 @@ TIPS = [
     "Three plugin types: general (tools/hooks), memory providers, and context engines.",
     "hermes plugins install owner/repo installs plugins directly from GitHub.",
     "8 external memory providers available: Honcho, OpenViking, Mem0, Hindsight, and more.",
-    "Plugin hooks include pre_tool_call, post_tool_call, pre_llm_call, and post_llm_call.",
+    "Plugin hooks include pre/post_tool_call, pre/post_llm_call, and transform_terminal_output for output canonicalization.",
 
     # --- Miscellaneous ---
     "Prompt caching (Anthropic) reduces costs by reusing cached system prompt prefixes.",
@@ -323,7 +323,6 @@ TIPS = [
     "GPT-5 and Codex use 'developer' role instead of 'system' in the message format.",
     "Per-task auxiliary overrides: auxiliary.vision.provider, auxiliary.compression.model, etc. in config.yaml.",
     "The auxiliary client treats 'main' as a provider alias — resolves to your actual primary provider + model.",
-    "Smart routing can auto-route simple queries to a cheaper model — set smart_model_routing.enabled: true.",
     "hermes claw migrate --dry-run previews OpenClaw migration without writing anything.",
     "File paths pasted with quotes or escaped spaces are handled automatically — no manual cleanup needed.",
     "Slash commands never trigger the large-paste collapse — /command with big arguments works correctly.",
@@ -346,4 +345,3 @@ def get_random_tip(exclude_recent: int = 0) -> str:
     return random.choice(TIPS)
 
 
-

hermes_cli/tools_config.py

@@ -24,7 +24,8 @@ from hermes_cli.nous_subscription import (
     apply_nous_managed_defaults,
     get_nous_subscription_features,
 )
-from tools.tool_backend_helpers import managed_nous_tools_enabled
+from tools.tool_backend_helpers import fal_key_is_configured, managed_nous_tools_enabled
+from utils import base_url_hostname
 
 logger = logging.getLogger(__name__)
 
@@ -181,6 +182,14 @@ TOOL_CATEGORIES = {
                 ],
                 "tts_provider": "gemini",
             },
+            {
+                "name": "KittenTTS",
+                "badge": "local · free",
+                "tag": "Lightweight local ONNX TTS (~25MB), no API key",
+                "env_vars": [],
+                "tts_provider": "kittentts",
+                "post_setup": "kittentts",
+            },
         ],
     },
     "web": {
@@ -422,6 +431,36 @@ def _run_post_setup(post_setup_key: str):
             _print_warning("    Node.js not found. Install Camofox via Docker:")
             _print_info("      docker run -p 9377:9377 -e CAMOFOX_PORT=9377 jo-inc/camofox-browser")
 
+    elif post_setup_key == "kittentts":
+        try:
+            __import__("kittentts")
+            _print_success("    kittentts is already installed")
+            return
+        except ImportError:
+            pass
+        import subprocess
+        _print_info("    Installing kittentts (~25-80MB model, CPU-only)...")
+        wheel_url = (
+            "https://github.com/KittenML/KittenTTS/releases/download/"
+            "0.8.1/kittentts-0.8.1-py3-none-any.whl"
+        )
+        try:
+            result = subprocess.run(
+                [sys.executable, "-m", "pip", "install", "-U", wheel_url, "soundfile", "--quiet"],
+                capture_output=True, text=True, timeout=300,
+            )
+            if result.returncode == 0:
+                _print_success("    kittentts installed")
+                _print_info("    Voices: Jasper, Bella, Luna, Bruno, Rosie, Hugo, Kiki, Leo")
+                _print_info("    Models: KittenML/kitten-tts-nano-0.8-int8 (25MB), micro (41MB), mini (80MB)")
+            else:
+                _print_warning("    kittentts install failed:")
+                _print_info(f"      {result.stderr.strip()[:300]}")
+                _print_info(f"    Run manually: python -m pip install -U '{wheel_url}' soundfile")
+        except subprocess.TimeoutExpired:
+            _print_warning("    kittentts install timed out (>5min)")
+            _print_info(f"    Run manually: python -m pip install -U '{wheel_url}' soundfile")
+
     elif post_setup_key == "rl_training":
         try:
             __import__("tinker_atropos")
@@ -546,6 +585,10 @@ def _get_platform_tools(
             ts_tools = set(resolve_toolset(ts_key))
             if ts_tools and ts_tools.issubset(all_tool_names):
                 enabled_toolsets.add(ts_key)
+        default_off = set(_DEFAULT_OFF_TOOLSETS)
+        if platform in default_off:
+            default_off.remove(platform)
+        enabled_toolsets -= default_off
 
     # Plugin toolsets: enabled by default unless explicitly disabled.
     # A plugin toolset is "known" for a platform once `hermes tools`
@@ -833,7 +876,7 @@ def _toolset_needs_configuration_prompt(ts_key: str, config: dict) -> bool:
         browser_cfg = config.get("browser", {})
         return not isinstance(browser_cfg, dict) or "cloud_provider" not in browser_cfg
     if ts_key == "image_gen":
-        return not get_env_value("FAL_KEY")
+        return not fal_key_is_configured()
 
     return not _toolset_has_keys(ts_key, config)
 
@@ -1175,17 +1218,17 @@ def _configure_simple_requirements(ts_key: str):
                 _print_warning("    Skipped")
         elif idx == 1:
             base_url = _prompt("    OPENAI_BASE_URL (blank for OpenAI)").strip() or "https://api.openai.com/v1"
-            key_label = "    OPENAI_API_KEY" if "api.openai.com" in base_url.lower() else "    API key"
+            is_native_openai = base_url_hostname(base_url) == "api.openai.com"
+            key_label = "    OPENAI_API_KEY" if is_native_openai else "    API key"
             api_key = _prompt(key_label, password=True)
             if api_key and api_key.strip():
                 save_env_value("OPENAI_API_KEY", api_key.strip())
                 # Save vision base URL to config (not .env — only secrets go there)
-                from hermes_cli.config import load_config, save_config
                 _cfg = load_config()
                 _aux = _cfg.setdefault("auxiliary", {}).setdefault("vision", {})
                 _aux["base_url"] = base_url
                 save_config(_cfg)
-                if "api.openai.com" in base_url.lower():
+                if is_native_openai:
                     save_env_value("AUXILIARY_VISION_MODEL", "gpt-4o-mini")
                 _print_success("    Saved")
             else:

hermes_cli/web_server.py

@@ -16,6 +16,7 @@ import json
 import logging
 import os
 import secrets
+import subprocess
 import sys
 import threading
 import time
@@ -114,6 +115,91 @@ def _require_token(request: Request) -> None:
         raise HTTPException(status_code=401, detail="Unauthorized")
 
 
+# Accepted Host header values for loopback binds. DNS rebinding attacks
+# point a victim browser at an attacker-controlled hostname (evil.test)
+# which resolves to 127.0.0.1 after a TTL flip — bypassing same-origin
+# checks because the browser now considers evil.test and our dashboard
+# "same origin". Validating the Host header at the app layer rejects any
+# request whose Host isn't one we bound for. See GHSA-ppp5-vxwm-4cf7.
+_LOOPBACK_HOST_VALUES: frozenset = frozenset({
+    "localhost", "127.0.0.1", "::1",
+})
+
+
+def _is_accepted_host(host_header: str, bound_host: str) -> bool:
+    """True if the Host header targets the interface we bound to.
+
+    Accepts:
+    - Exact bound host (with or without port suffix)
+    - Loopback aliases when bound to loopback
+    - Any host when bound to 0.0.0.0 (explicit opt-in to non-loopback,
+      no protection possible at this layer)
+    """
+    if not host_header:
+        return False
+    # Strip port suffix. IPv6 addresses use bracket notation:
+    #   [::1]         — no port
+    #   [::1]:9119    — with port
+    # Plain hosts/v4:
+    #   localhost:9119
+    #   127.0.0.1:9119
+    h = host_header.strip()
+    if h.startswith("["):
+        # IPv6 bracketed — port (if any) follows "]:"
+        close = h.find("]")
+        if close != -1:
+            host_only = h[1:close]  # strip brackets
+        else:
+            host_only = h.strip("[]")
+    else:
+        host_only = h.rsplit(":", 1)[0] if ":" in h else h
+    host_only = host_only.lower()
+
+    # 0.0.0.0 bind means operator explicitly opted into all-interfaces
+    # (requires --insecure per web_server.start_server). No Host-layer
+    # defence can protect that mode; rely on operator network controls.
+    if bound_host in ("0.0.0.0", "::"):
+        return True
+
+    # Loopback bind: accept the loopback names
+    bound_lc = bound_host.lower()
+    if bound_lc in _LOOPBACK_HOST_VALUES:
+        return host_only in _LOOPBACK_HOST_VALUES
+
+    # Explicit non-loopback bind: require exact host match
+    return host_only == bound_lc
+
+
+@app.middleware("http")
+async def host_header_middleware(request: Request, call_next):
+    """Reject requests whose Host header doesn't match the bound interface.
+
+    Defends against DNS rebinding: a victim browser on a localhost
+    dashboard is tricked into fetching from an attacker hostname that
+    TTL-flips to 127.0.0.1. CORS and same-origin checks don't help —
+    the browser now treats the attacker origin as same-origin with the
+    dashboard. Host-header validation at the app layer catches it.
+
+    See GHSA-ppp5-vxwm-4cf7.
+    """
+    # Store the bound host on app.state so this middleware can read it —
+    # set by start_server() at listen time.
+    bound_host = getattr(app.state, "bound_host", None)
+    if bound_host:
+        host_header = request.headers.get("host", "")
+        if not _is_accepted_host(host_header, bound_host):
+            return JSONResponse(
+                status_code=400,
+                content={
+                    "detail": (
+                        "Invalid Host header. Dashboard requests must use "
+                        "the hostname the server was bound to."
+                    ),
+                },
+            )
+    return await call_next(request)
+
+
 @app.middleware("http")
 async def auth_middleware(request: Request, call_next):
     """Require the session token on all /api/ routes except the public list."""
@@ -232,8 +318,8 @@ _CATEGORY_MERGE: Dict[str, str] = {
     "checkpoints": "agent",
     "approvals": "security",
     "human_delay": "display",
-    "smart_model_routing": "agent",
     "dashboard": "display",
+    "code_execution": "agent",
 }
 
 # Display order for tabs — unlisted categories sort alphabetically after these.
@@ -476,6 +562,138 @@ async def get_status():
     }
 
 
+# ---------------------------------------------------------------------------
+# Gateway + update actions (invoked from the Status page).
+#
+# Both commands are spawned as detached subprocesses so the HTTP request
+# returns immediately.  stdin is closed (``DEVNULL``) so any stray ``input()``
+# calls fail fast with EOF rather than hanging forever.  stdout/stderr are
+# streamed to a per-action log file under ``~/.hermes/logs/<action>.log`` so
+# the dashboard can tail them back to the user.
+# ---------------------------------------------------------------------------
+
+_ACTION_LOG_DIR: Path = get_hermes_home() / "logs"
+
+# Short ``name`` (from the URL) → absolute log file path.
+_ACTION_LOG_FILES: Dict[str, str] = {
+    "gateway-restart": "gateway-restart.log",
+    "hermes-update": "hermes-update.log",
+}
+
+# ``name`` → most recently spawned Popen handle.  Used so ``status`` can
+# report liveness and exit code without shelling out to ``ps``.
+_ACTION_PROCS: Dict[str, subprocess.Popen] = {}
+
+
+def _spawn_hermes_action(subcommand: List[str], name: str) -> subprocess.Popen:
+    """Spawn ``hermes <subcommand>`` detached and record the Popen handle.
+
+    Uses the running interpreter's ``hermes_cli.main`` module so the action
+    inherits the same venv/PYTHONPATH the web server is using.
+    """
+    log_file_name = _ACTION_LOG_FILES[name]
+    _ACTION_LOG_DIR.mkdir(parents=True, exist_ok=True)
+    log_path = _ACTION_LOG_DIR / log_file_name
+    log_file = open(log_path, "ab", buffering=0)
+    log_file.write(
+        f"\n=== {name} started {time.strftime('%Y-%m-%d %H:%M:%S')} ===\n".encode()
+    )
+
+    cmd = [sys.executable, "-m", "hermes_cli.main", *subcommand]
+
+    popen_kwargs: Dict[str, Any] = {
+        "cwd": str(PROJECT_ROOT),
+        "stdin": subprocess.DEVNULL,
+        "stdout": log_file,
+        "stderr": subprocess.STDOUT,
+        "env": {**os.environ, "HERMES_NONINTERACTIVE": "1"},
+    }
+    if sys.platform == "win32":
+        popen_kwargs["creationflags"] = (
+            subprocess.CREATE_NEW_PROCESS_GROUP  # type: ignore[attr-defined]
+            | getattr(subprocess, "DETACHED_PROCESS", 0)
+        )
+    else:
+        popen_kwargs["start_new_session"] = True
+
+    proc = subprocess.Popen(cmd, **popen_kwargs)
+    _ACTION_PROCS[name] = proc
+    return proc
+
+
+def _tail_lines(path: Path, n: int) -> List[str]:
+    """Return the last ``n`` lines of ``path``.  Reads the whole file — fine
+    for our small per-action logs.  Binary-decoded with ``errors='replace'``
+    so log corruption doesn't 500 the endpoint."""
+    if not path.exists():
+        return []
+    try:
+        text = path.read_text(errors="replace")
+    except OSError:
+        return []
+    lines = text.splitlines()
+    return lines[-n:] if n > 0 else lines
+
+
+@app.post("/api/gateway/restart")
+async def restart_gateway():
+    """Kick off a ``hermes gateway restart`` in the background."""
+    try:
+        proc = _spawn_hermes_action(["gateway", "restart"], "gateway-restart")
+    except Exception as exc:
+        _log.exception("Failed to spawn gateway restart")
+        raise HTTPException(status_code=500, detail=f"Failed to restart gateway: {exc}")
+    return {
+        "ok": True,
+        "pid": proc.pid,
+        "name": "gateway-restart",
+    }
+
+
+@app.post("/api/hermes/update")
+async def update_hermes():
+    """Kick off ``hermes update`` in the background."""
+    try:
+        proc = _spawn_hermes_action(["update"], "hermes-update")
+    except Exception as exc:
+        _log.exception("Failed to spawn hermes update")
+        raise HTTPException(status_code=500, detail=f"Failed to start update: {exc}")
+    return {
+        "ok": True,
+        "pid": proc.pid,
+        "name": "hermes-update",
+    }
+
+
+@app.get("/api/actions/{name}/status")
+async def get_action_status(name: str, lines: int = 200):
+    """Tail an action log and report whether the process is still running."""
+    log_file_name = _ACTION_LOG_FILES.get(name)
+    if log_file_name is None:
+        raise HTTPException(status_code=404, detail=f"Unknown action: {name}")
+
+    log_path = _ACTION_LOG_DIR / log_file_name
+    tail = _tail_lines(log_path, min(max(lines, 1), 2000))
+
+    proc = _ACTION_PROCS.get(name)
+    if proc is None:
+        running = False
+        exit_code: Optional[int] = None
+        pid: Optional[int] = None
+    else:
+        exit_code = proc.poll()
+        running = exit_code is None
+        pid = proc.pid
+
+    return {
+        "name": name,
+        "running": running,
+        "exit_code": exit_code,
+        "pid": pid,
+        "lines": tail,
+    }
+
+
 @app.get("/api/sessions")
 async def get_sessions(limit: int = 20, offset: int = 0):
     try:
@@ -1958,6 +2176,8 @@ async def update_config_raw(body: RawConfigUpdate):
 @app.get("/api/analytics/usage")
 async def get_usage_analytics(days: int = 30):
     from hermes_state import SessionDB
+    from agent.insights import InsightsEngine
+
     db = SessionDB()
     try:
         cutoff = time.time() - (days * 86400)
@@ -1997,8 +2217,24 @@ async def get_usage_analytics(days: int = 30):
             FROM sessions WHERE started_at > ?
         """, (cutoff,))
         totals = dict(cur3.fetchone())
+        insights_report = InsightsEngine(db).generate(days=days)
+        skills = insights_report.get("skills", {
+            "summary": {
+                "total_skill_loads": 0,
+                "total_skill_edits": 0,
+                "total_skill_actions": 0,
+                "distinct_skills_used": 0,
+            },
+            "top_skills": [],
+        })
 
-        return {"daily": daily, "by_model": by_model, "totals": totals, "period_days": days}
+        return {
+            "daily": daily,
+            "by_model": by_model,
+            "totals": totals,
+            "period_days": days,
+            "skills": skills,
+        }
     finally:
         db.close()
 
@@ -2305,13 +2541,15 @@ def start_server(
             "authentication. Only use on trusted networks.", host,
         )
 
+    # Record the bound host so host_header_middleware can validate incoming
+    # Host headers against it. Defends against DNS rebinding (GHSA-ppp5-vxwm-4cf7).
+    app.state.bound_host = host
+
     if open_browser:
-        import threading
         import webbrowser
 
         def _open():
-            import time as _t
-            _t.sleep(1.0)
+            time.sleep(1.0)
             webbrowser.open(f"http://{host}:{port}")
 
         threading.Thread(target=_open, daemon=True).start()

hermes_state.py

@@ -383,10 +383,19 @@ class SessionDB:
         return session_id
 
     def end_session(self, session_id: str, end_reason: str) -> None:
-        """Mark a session as ended."""
+        """Mark a session as ended.
+
+        No-ops when the session is already ended. The first end_reason wins:
+        compression-split sessions must keep their ``end_reason = 'compression'``
+        record even if a later stale ``end_session()`` call (e.g. from a
+        desynced CLI session_id after ``/resume`` or ``/branch``) targets them
+        with a different reason. Use ``reopen_session()`` first if you
+        intentionally need to re-end a closed session with a new reason.
+        """
         def _do(conn):
             conn.execute(
-                "UPDATE sessions SET ended_at = ?, end_reason = ? WHERE id = ?",
+                "UPDATE sessions SET ended_at = ?, end_reason = ? "
+                "WHERE id = ? AND ended_at IS NULL",
                 (time.time(), end_reason, session_id),
             )
         self._execute_write(_do)
@@ -714,6 +723,42 @@ class SessionDB:
 
         return f"{base} #{max_num + 1}"
 
+    def get_compression_tip(self, session_id: str) -> Optional[str]:
+        """Walk the compression-continuation chain forward and return the tip.
+
+        A compression continuation is a child session where:
+        1. The parent's ``end_reason = 'compression'``
+        2. The child was created AFTER the parent was ended (started_at >= ended_at)
+
+        The second condition distinguishes compression continuations from
+        delegate subagents or branch children, which can also have a
+        ``parent_session_id`` but were created while the parent was still live.
+
+        Returns the session_id of the latest continuation in the chain, or the
+        input ``session_id`` if it isn't part of a compression chain (or if the
+        input itself doesn't exist).
+        """
+        current = session_id
+        # Bound the walk defensively — compression chains this deep are
+        # pathological and shouldn't happen in practice. 100 = plenty.
+        for _ in range(100):
+            with self._lock:
+                cursor = self._conn.execute(
+                    "SELECT id FROM sessions "
+                    "WHERE parent_session_id = ? "
+                    "  AND started_at >= ("
+                    "      SELECT ended_at FROM sessions "
+                    "      WHERE id = ? AND end_reason = 'compression'"
+                    "  ) "
+                    "ORDER BY started_at DESC LIMIT 1",
+                    (current, current),
+                )
+                row = cursor.fetchone()
+            if row is None:
+                return current
+            current = row["id"]
+        return current
+
     def list_sessions_rich(
         self,
         source: str = None,
@@ -721,6 +766,7 @@ class SessionDB:
         limit: int = 20,
         offset: int = 0,
         include_children: bool = False,
+        project_compression_tips: bool = True,
     ) -> List[Dict[str, Any]]:
         """List sessions with preview (first user message) and last active timestamp.
 
@@ -732,6 +778,14 @@ class SessionDB:
 
         By default, child sessions (subagent runs, compression continuations)
         are excluded.  Pass ``include_children=True`` to include them.
+
+        With ``project_compression_tips=True`` (default), sessions that are
+        roots of compression chains are projected forward to their latest
+        continuation — one logical conversation = one list entry, showing the
+        live continuation's id/message_count/title/last_active. This prevents
+        compressed continuations from being invisible to users while keeping
+        delegate subagents and branches hidden. Pass ``False`` to return the
+        raw root rows (useful for admin/debug UIs).
         """
         where_clauses = []
         params = []
@@ -782,8 +836,77 @@ class SessionDB:
                 s["preview"] = ""
             sessions.append(s)
 
+        # Project compression roots forward to their tips. Each row whose
+        # end_reason is 'compression' has a continuation child; replace the
+        # surfaced fields (id, message_count, title, last_active, ended_at,
+        # end_reason, preview) with the tip's values so the list entry acts
+        # as the live conversation. Keep the root's started_at to preserve
+        # chronological ordering by original conversation start.
+        if project_compression_tips and not include_children:
+            projected = []
+            for s in sessions:
+                if s.get("end_reason") != "compression":
+                    projected.append(s)
+                    continue
+                tip_id = self.get_compression_tip(s["id"])
+                if tip_id == s["id"]:
+                    projected.append(s)
+                    continue
+                tip_row = self._get_session_rich_row(tip_id)
+                if not tip_row:
+                    projected.append(s)
+                    continue
+                # Preserve the root's started_at for stable sort order, but
+                # surface the tip's identity and activity data.
+                merged = dict(s)
+                for key in (
+                    "id", "ended_at", "end_reason", "message_count",
+                    "tool_call_count", "title", "last_active", "preview",
+                    "model", "system_prompt",
+                ):
+                    if key in tip_row:
+                        merged[key] = tip_row[key]
+                merged["_lineage_root_id"] = s["id"]
+                projected.append(merged)
+            sessions = projected
+
         return sessions
 
+    def _get_session_rich_row(self, session_id: str) -> Optional[Dict[str, Any]]:
+        """Fetch a single session with the same enriched columns as
+        ``list_sessions_rich`` (preview + last_active). Returns None if the
+        session doesn't exist.
+        """
+        query = """
+            SELECT s.*,
+                COALESCE(
+                    (SELECT SUBSTR(REPLACE(REPLACE(m.content, X'0A', ' '), X'0D', ' '), 1, 63)
+                     FROM messages m
+                     WHERE m.session_id = s.id AND m.role = 'user' AND m.content IS NOT NULL
+                     ORDER BY m.timestamp, m.id LIMIT 1),
+                    ''
+                ) AS _preview_raw,
+                COALESCE(
+                    (SELECT MAX(m2.timestamp) FROM messages m2 WHERE m2.session_id = s.id),
+                    s.started_at
+                ) AS last_active
+            FROM sessions s
+            WHERE s.id = ?
+        """
+        with self._lock:
+            cursor = self._conn.execute(query, (session_id,))
+            row = cursor.fetchone()
+        if not row:
+            return None
+        s = dict(row)
+        raw = s.pop("_preview_raw", "").strip()
+        if raw:
+            text = raw[:60]
+            s["preview"] = text + ("..." if len(raw) > 60 else "")
+        else:
+            s["preview"] = ""
+        return s
+
     # =========================================================================
     # Message storage
     # =========================================================================
@@ -1126,10 +1249,37 @@ class SessionDB:
             try:
                 with self._lock:
                     ctx_cursor = self._conn.execute(
-                        """SELECT role, content FROM messages
-                           WHERE session_id = ? AND id >= ? - 1 AND id <= ? + 1
-                           ORDER BY id""",
-                        (match["session_id"], match["id"], match["id"]),
+                        """WITH target AS (
+                               SELECT session_id, timestamp, id
+                               FROM messages
+                               WHERE id = ?
+                           )
+                           SELECT role, content
+                           FROM (
+                               SELECT m.id, m.timestamp, m.role, m.content
+                               FROM messages m
+                               JOIN target t ON t.session_id = m.session_id
+                               WHERE (m.timestamp < t.timestamp)
+                                  OR (m.timestamp = t.timestamp AND m.id < t.id)
+                               ORDER BY m.timestamp DESC, m.id DESC
+                               LIMIT 1
+                           )
+                           UNION ALL
+                           SELECT role, content
+                           FROM messages
+                           WHERE id = ?
+                           UNION ALL
+                           SELECT role, content
+                           FROM (
+                               SELECT m.id, m.timestamp, m.role, m.content
+                               FROM messages m
+                               JOIN target t ON t.session_id = m.session_id
+                               WHERE (m.timestamp > t.timestamp)
+                                  OR (m.timestamp = t.timestamp AND m.id > t.id)
+                               ORDER BY m.timestamp ASC, m.id ASC
+                               LIMIT 1
+                           )""",
+                        (match["id"], match["id"]),
                     )
                     context_msgs = [
                         {"role": r["role"], "content": (r["content"] or "")[:200]}

mini_swe_runner.py

@@ -43,13 +43,23 @@ from dotenv import load_dotenv
 load_dotenv()
 
 
-def _effective_temperature_for_model(model: str) -> Optional[float]:
-    """Return a fixed temperature for models with strict sampling contracts."""
+def _effective_temperature_for_model(
+    model: str,
+    base_url: Optional[str] = None,
+) -> Optional[float]:
+    """Return a fixed temperature for models with strict sampling contracts.
+
+    Returns ``None`` when the model manages temperature server-side (Kimi);
+    callers must omit the ``temperature`` kwarg entirely in that case.
+    """
     try:
-        from agent.auxiliary_client import _fixed_temperature_for_model
+        from agent.auxiliary_client import _fixed_temperature_for_model, OMIT_TEMPERATURE
     except Exception:
         return None
-    return _fixed_temperature_for_model(model)
+    result = _fixed_temperature_for_model(model, base_url)
+    if result is OMIT_TEMPERATURE:
+        return None  # caller must omit temperature
+    return result
 
 
 
@@ -457,7 +467,10 @@ Complete the user's task step by step."""
                         "tools": self.tools,
                         "timeout": 300.0,
                     }
-                    fixed_temperature = _effective_temperature_for_model(self.model)
+                    fixed_temperature = _effective_temperature_for_model(
+                        self.model,
+                        str(getattr(self.client, "base_url", "") or ""),
+                    )
                     if fixed_temperature is not None:
                         api_kwargs["temperature"] = fixed_temperature
 

model_tools.py

@@ -282,6 +282,31 @@ def get_tool_definitions(
                 filtered_tools[i] = {"type": "function", "function": dynamic_schema}
                 break
 
+    # Rebuild discord_server schema based on the bot's privileged intents
+    # (detected from GET /applications/@me) and the user's action allowlist
+    # in config.  Hides actions the bot's intents don't support so the
+    # model never attempts them, and annotates fetch_messages when the
+    # MESSAGE_CONTENT intent is missing.
+    if "discord_server" in available_tool_names:
+        try:
+            from tools.discord_tool import get_dynamic_schema
+            dynamic = get_dynamic_schema()
+        except Exception:  # pragma: no cover — defensive, fall back to static
+            dynamic = None
+        if dynamic is None:
+            # Tool filtered out entirely (empty allowlist or detection disabled
+            # the only remaining actions).  Drop it from the schema list.
+            filtered_tools = [
+                t for t in filtered_tools
+                if t.get("function", {}).get("name") != "discord_server"
+            ]
+            available_tool_names.discard("discord_server")
+        else:
+            for i, td in enumerate(filtered_tools):
+                if td.get("function", {}).get("name") == "discord_server":
+                    filtered_tools[i] = {"type": "function", "function": dynamic}
+                    break
+
     # Strip web tool cross-references from browser_navigate description when
     # web_search / web_extract are not available.  The static schema says
     # "prefer web_search or web_extract" which causes the model to hallucinate
@@ -525,6 +550,30 @@ def handle_function_call(
         except Exception:
             pass
 
+        # Generic tool-result canonicalization seam: plugins receive the
+        # final result string (JSON, usually) and may replace it by
+        # returning a string from transform_tool_result. Runs after
+        # post_tool_call (which stays observational) and before the result
+        # is appended back into conversation context. Fail-open; the first
+        # valid string return wins; non-string returns are ignored.
+        try:
+            from hermes_cli.plugins import invoke_hook
+            hook_results = invoke_hook(
+                "transform_tool_result",
+                tool_name=function_name,
+                args=function_args,
+                result=result,
+                task_id=task_id or "",
+                session_id=session_id or "",
+                tool_call_id=tool_call_id or "",
+            )
+            for hook_result in hook_results:
+                if isinstance(hook_result, str):
+                    result = hook_result
+                    break
+        except Exception:
+            pass
+
         return result
 
     except Exception as e: