Merge branch 'bb/pets' into bb/pets-gen

Carry forward the overlay/waiting-state updates and resolve the gateway merge conflict.
Also tighten the desktop pet-generation flow by cleaning superseded previews,
using the draft's source prompt during hatch, and previewing rows from the
returned sheet taxonomy.

.dockerignore

@@ -63,3 +63,45 @@ data/
 # Compose/profile runtime state (bind-mounted; avoid ownership/secret issues)
 hermes-config/
 runtime/
+
+# ---------- Not needed inside the Docker image ----------
+
+# Desktop app source (Tauri/Electron); never installed in the container
+apps/
+
+# Test suite — not shipped in production images
+tests/
+
+# Documentation site (Docusaurus) and supplementary docs
+website/
+docs/
+
+# Assets only used by the GitHub README
+assets/
+infographic/
+
+# Plugin-level docs (hermes-achievements ships docs/ but the runtime doesn't read them)
+plugins/hermes-achievements/docs/
+
+# Nix / Homebrew / AUR packaging metadata — irrelevant to Docker
+nix/
+flake.nix
+flake.lock
+packaging/
+
+# Design and planning documents
+plans/
+.plans/
+
+# ACP registry manifest (icon + agent.json) — not consumed at runtime
+acp_registry/
+
+# Repo-level dotfiles that are git-only or dev-tooling config
+.env.example
+.envrc
+.gitattributes
+.hadolint.yaml
+.mailmap
+
+# Top-level LICENSE (not matched by *.md); not needed inside the container
+LICENSE

.github/workflows/contributor-check.yml

@@ -1,12 +1,11 @@
 name: Contributor Attribution Check
 
 on:
-  pull_request:
-    branches: [main]
   # No paths filter — the job must always run so the required check
   # reports a status (path-gated workflows leave checks "pending" forever
   # when no matching files change, which blocks merge).
-
+  pull_request:
+    branches: [main]
 permissions:
   contents: read
 

.github/workflows/deploy-site.yml

@@ -11,8 +11,20 @@ on:
       - 'optional-skills/**'
       - '.github/workflows/deploy-site.yml'
   workflow_dispatch:
+    inputs:
+      skills_index_run_id:
+        description: 'Optional Build Skills Index run ID whose skills-index artifact should be deployed'
+        required: false
+        type: string
+      rebuild_skills_index:
+        description: 'Force a fresh multi-source crawl instead of reusing the latest healthy index'
+        required: false
+        default: false
+        type: boolean
 
 permissions:
+  contents: read
+  actions: read
   pages: write
   id-token: write
 
@@ -44,7 +56,7 @@ jobs:
 
       - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020  # v4
         with:
-          node-version: 20
+          node-version: 22
           cache: npm
           cache-dependency-path: website/package-lock.json
 
@@ -55,26 +67,81 @@ jobs:
       - name: Install PyYAML for skill extraction
         run: pip install pyyaml==6.0.2 httpx==0.28.1
 
-      - name: Build skills index (unified multi-source catalog)
+      - name: Prepare skills index (unified multi-source catalog)
         env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GH_TOKEN: ${{ github.token }}
+          GITHUB_TOKEN: ${{ github.token }}
+          SKILLS_INDEX_RUN_ID: ${{ github.event.inputs.skills_index_run_id || '' }}
+          REBUILD_SKILLS_INDEX: ${{ github.event.inputs.rebuild_skills_index || 'false' }}
         run: |
-          # Rebuild the unified catalog. The file is gitignored, so a fresh
-          # checkout starts without it and we want the freshest crawl in
-          # every deploy.
+          # The unified external catalog is expensive to crawl and can burn
+          # through the repository installation's GitHub API quota when several
+          # docs deploys land close together.  Normal docs deploys therefore
+          # reuse the latest healthy catalog: first the artifact from a
+          # scheduled skills-index run, then the currently live index.  Only a
+          # manual force rebuild does a fresh crawl here.
           #
-          # This MUST be fatal. build_skills_index.py runs a health check and
-          # exits non-zero WITHOUT writing the output file when a source
-          # collapses (e.g. a GitHub API rate limit zeroes the github /
-          # claude-marketplace / well-known taps all at once). Letting the
-          # deploy continue would either (a) ship a degenerate index missing
-          # whole hubs — the June 2026 regression where OpenAI/Anthropic/
-          # HuggingFace/NVIDIA tabs vanished — or (b) fall through to a
-          # local-only catalog. Failing here keeps the last good deployment
-          # live (GitHub Pages serves the previous build) instead of
-          # publishing a broken catalog. Re-run the workflow once the
-          # transient rate limit clears.
+          # If we do crawl, the build remains fatal. build_skills_index.py runs
+          # the health check BEFORE writing and exits non-zero on source
+          # collapse, keeping the last good Pages deployment live instead of
+          # publishing a degenerate catalog.
+          set -euo pipefail
+          INDEX_PATH="website/static/api/skills-index.json"
+          mkdir -p "$(dirname "$INDEX_PATH")"
+
+          validate_index() {
+            python3 - "$INDEX_PATH" <<'PY'
+          import json
+          import sys
+          from pathlib import Path
+
+          path = Path(sys.argv[1])
+          try:
+              data = json.loads(path.read_text(encoding="utf-8"))
+          except Exception as exc:
+              print(f"invalid skills index JSON: {exc}", file=sys.stderr)
+              sys.exit(1)
+          skills = data.get("skills")
+          if not isinstance(skills, list) or len(skills) < 1500:
+              count = len(skills) if isinstance(skills, list) else "missing"
+              print(f"skills index too small: {count}", file=sys.stderr)
+              sys.exit(1)
+          print(f"skills index ready: {len(skills)} skills")
+          PY
+          }
+
+          if [ "$REBUILD_SKILLS_INDEX" = "true" ]; then
+            python3 scripts/build_skills_index.py
+            validate_index
+            exit 0
+          fi
+
+          if [ -n "$SKILLS_INDEX_RUN_ID" ]; then
+            tmpdir="$(mktemp -d)"
+            echo "Downloading skills-index artifact from run $SKILLS_INDEX_RUN_ID"
+            if gh run download "$SKILLS_INDEX_RUN_ID" --name skills-index --dir "$tmpdir"; then
+              candidate="$(find "$tmpdir" -name skills-index.json -type f | head -n 1 || true)"
+              if [ -n "$candidate" ]; then
+                cp "$candidate" "$INDEX_PATH"
+                if validate_index; then
+                  exit 0
+                fi
+              fi
+            fi
+            echo "::warning::Could not use skills-index artifact from run $SKILLS_INDEX_RUN_ID; trying live index"
+          fi
+
+          echo "Downloading currently live skills index"
+          if curl -fsSL --retry 3 --retry-delay 5 \
+            "https://hermes-agent.nousresearch.com/docs/api/skills-index.json" \
+            -o "$INDEX_PATH" && validate_index; then
+            exit 0
+          fi
+
+          echo "::warning::Live skills index unavailable or unhealthy; falling back to a fresh crawl"
+          rm -f "$INDEX_PATH"
           python3 scripts/build_skills_index.py
+          validate_index
 
       - name: Extract skill metadata for dashboard
         run: python3 website/scripts/extract-skills.py

.github/workflows/docker-lint.yml

@@ -18,13 +18,12 @@ on:
       - docker/**
       - .hadolint.yaml
       - .github/workflows/docker-lint.yml
+
+  # No paths filter — the job must always run so the required check
+  # reports a status (path-gated workflows leave checks "pending" forever
+  # when no matching files change, which blocks merge).
   pull_request:
     branches: [main]
-    paths:
-      - Dockerfile
-      - docker/**
-      - .hadolint.yaml
-      - .github/workflows/docker-lint.yml
 
 permissions:
   contents: read

.github/workflows/docker-publish.yml

@@ -11,16 +11,13 @@ on:
       - 'docker/**'
       - '.github/workflows/docker-publish.yml'
       - '.github/actions/hermes-smoke-test/**'
+
+  # No paths filter — the job must always run so the required check
+  # reports a status (path-gated workflows leave checks "pending" forever
+  # when no matching files change, which blocks merge).
   pull_request:
     branches: [main]
-    paths:
-      - '**/*.py'
-      - 'pyproject.toml'
-      - 'uv.lock'
-      - 'Dockerfile'
-      - 'docker/**'
-      - '.github/workflows/docker-publish.yml'
-      - '.github/actions/hermes-smoke-test/**'
+
   release:
     types: [published]
 
@@ -90,7 +87,7 @@ jobs:
       # (see `_SKIP_PARTS` in scripts/run_tests_parallel.py) because each
       # shard would otherwise reach the session-scoped ``built_image``
       # fixture in ``tests/docker/conftest.py`` and start a 3-7min
-      # ``docker build`` under a 180s pytest-timeout cap — guaranteed to
+      # ``docker build`` — guaranteed to
       # die in fixture setup.
       #
       # Piggybacking here avoids a second image build: the smoke test
@@ -114,7 +111,7 @@ jobs:
         run: |
           uv venv .venv --python 3.11
           source .venv/bin/activate
-          # ``dev`` extra pulls in pytest, pytest-asyncio, pytest-timeout —
+          # ``dev`` extra pulls in pytest, pytest-asyncio —
           # everything tests/docker/ needs.  We deliberately avoid ``all``
           # here because the docker tests only drive the container via
           # subprocess and don't import hermes_agent's optional deps.

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

@@ -1,10 +1,12 @@
 name: Docs Site Checks
 
 on:
+  # No paths filter — the job must always run so the required check
+  # reports a status (path-gated workflows leave checks "pending" forever
+  # when no matching files change, which blocks merge).
   pull_request:
-    paths:
-      - 'website/**'
-      - '.github/workflows/docs-site-checks.yml'
+    branches: [main]
+
   workflow_dispatch:
 
 permissions:
@@ -14,11 +16,11 @@ jobs:
   docs-site-checks:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
-      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020  # v4
+      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
         with:
-          node-version: 20
+          node-version: 22
           cache: npm
           cache-dependency-path: website/package-lock.json
 
@@ -26,9 +28,9 @@ jobs:
         run: npm ci
         working-directory: website
 
-      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
         with:
-          python-version: '3.11'
+          python-version: "3.11"
 
       - name: Install ascii-guard
         run: python -m pip install ascii-guard==2.3.0 pyyaml==6.0.3

.github/workflows/history-check.yml

@@ -14,6 +14,9 @@ name: History Check
 # the PR head and main to be non-empty.
 
 on:
+  # No paths filter — the job must always run so the required check
+  # reports a status (path-gated workflows leave checks "pending" forever
+  # when no matching files change, which blocks merge).
   pull_request:
     branches: [main]
 
@@ -24,9 +27,9 @@ jobs:
   check-common-ancestor:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
         with:
-          fetch-depth: 0  # full history both sides for merge-base
+          fetch-depth: 0 # full history both sides for merge-base
 
       - name: Reject PRs with no common ancestor on main
         run: |

.github/workflows/lint.yml

@@ -15,12 +15,12 @@ on:
       - "**/*.md"
       - "docs/**"
       - "website/**"
+
+  # No paths filter — the job must always run so the required check
+  # reports a status (path-gated workflows leave checks "pending" forever
+  # when no matching files change, which blocks merge).
   pull_request:
     branches: [main]
-    paths-ignore:
-      - "**/*.md"
-      - "docs/**"
-      - "website/**"
 
 permissions:
   contents: read
@@ -154,7 +154,6 @@ jobs:
               });
             }
 
-
   ruff-blocking:
     # Enforce the rules in pyproject.toml [tool.ruff.lint.select]. Currently
     # PLW1514 (unspecified-encoding) — catches bare ``open()`` /

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

@@ -1,255 +0,0 @@
-name: Nix Lockfile Fix
-
-on:
-  push:
-    branches: [main]
-    paths:
-      - 'package-lock.json'
-      - 'package.json'
-      - 'ui-tui/package.json'
-      - 'apps/desktop/package.json'
-  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:
-  # ── Auto-fix on main ───────────────────────────────────────────────
-  # Fires when a push to main touches package.json or package-lock.json.
-  # Runs fix-lockfiles and pushes the hash update commit directly to main
-  # so Nix builds never stay broken.
-  #
-  # Safety invariants:
-  #   1. The fix commit only touches nix/*.nix files, which are NOT in
-  #      the paths filter above, so this cannot re-trigger itself.
-  #   2. An explicit file-whitelist check before commit aborts if
-  #      fix-lockfiles ever modifies unexpected files.
-  #   3. Job-level concurrency with cancel-in-progress: true ensures
-  #      back-to-back pushes collapse to the newest; ref: main checkout
-  #      always operates on the latest branch state.
-  #   4. Uses a GitHub App token (not GITHUB_TOKEN) so the fix commit
-  #      triggers downstream nix.yml verification.
-  auto-fix-main:
-    if: github.event_name == 'push'
-    runs-on: ubuntu-latest
-    timeout-minutes: 25
-    concurrency:
-      group: auto-fix-main
-      cancel-in-progress: true
-    steps:
-      - name: Generate GitHub App token
-        id: app-token
-        uses: actions/create-github-app-token@7bfa3a4717ef143a604ee0a99d859b8886a96d00  # v1.9.3
-        with:
-          app-id: ${{ secrets.APP_ID }}
-          private-key: ${{ secrets.APP_PRIVATE_KEY }}
-
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-        with:
-          ref: main
-          token: ${{ steps.app-token.outputs.token }}
-
-      - uses: ./.github/actions/nix-setup
-        with:
-          cachix-auth-token: ${{ secrets.CACHIX_AUTH_TOKEN }}
-
-      - name: Apply lockfile hashes
-        id: apply
-        run: nix run .#fix-lockfiles -- --apply
-
-      - name: Commit & push
-        if: steps.apply.outputs.changed == 'true'
-        shell: bash
-        run: |
-          set -euo pipefail
-
-          # Ensure only nix/lib.nix (home of the single npmDepsHash) was
-          # modified — prevents accidental self-triggering if fix-lockfiles
-          # ever touches package files.
-          unexpected="$(git diff --name-only | grep -Ev '^nix/lib\.nix$' || true)"
-          if [ -n "$unexpected" ]; then
-            echo "::error::Unexpected modified files: $unexpected"
-            exit 1
-          fi
-
-          # Record the base SHA before committing — used to detect package
-          # file changes if we need to rebase after a non-fast-forward push.
-          BASE_SHA="$(git rev-parse HEAD)"
-
-          git config user.name 'github-actions[bot]'
-          git config user.email '41898282+github-actions[bot]@users.noreply.github.com'
-          git add nix/lib.nix
-          git commit -m "fix(nix): auto-refresh npm lockfile hashes" \
-            -m "Source: $GITHUB_SHA" \
-            -m "Run: $GITHUB_SERVER_URL/$GITHUB_REPOSITORY/actions/runs/$GITHUB_RUN_ID"
-
-          # Retry push with rebase in case main advanced with an unrelated
-          # commit during the nix build. Without this, a non-fast-forward
-          # rejection silently loses the fix. If package files changed during
-          # the rebase, abort — a fresh auto-fix run will handle the new state.
-          for attempt in 1 2 3; do
-            if git push origin HEAD:main; then
-              exit 0
-            fi
-            echo "::warning::Push attempt $attempt failed (non-fast-forward?), rebasing…"
-            git fetch origin main
-
-            # If package files changed between our base and the new main,
-            # our computed hashes are stale. Abort and let the next triggered
-            # run recompute from the correct package-lock state.
-            pkg_changed="$(git diff --name-only "$BASE_SHA"..origin/main -- \
-              'package-lock.json' 'package.json' \
-              'ui-tui/package.json' 'apps/desktop/package.json' || true)"
-            if [ -n "$pkg_changed" ]; then
-              echo "::warning::Package files changed since hash computation — aborting; a fresh run will recompute"
-              exit 0
-            fi
-
-            git rebase origin/main
-          done
-          echo "::error::Failed to push after 3 rebase attempts"
-          exit 1
-
-  # ── PR fix (manual / checkbox) ─────────────────────────────────────
-  # Existing behavior: run on manual dispatch OR when a task-list
-  # checkbox in the sticky lockfile-check comment flips from [ ] to [x].
-  fix:
-    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@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-        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
-        with:
-          cachix-auth-token: ${{ secrets.CACHIX_AUTH_TOKEN }}
-
-      - name: Apply lockfile hashes
-        id: apply
-        run: nix run .#fix-lockfiles
-
-      - 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/lib.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

@@ -1,105 +0,0 @@
-name: Nix
-
-on:
-  push:
-    branches: [main]
-  pull_request:
-
-permissions:
-  contents: read
-  pull-requests: write
-
-concurrency:
-  group: nix-${{ github.ref }}
-  cancel-in-progress: true
-
-jobs:
-  nix:
-    strategy:
-      matrix:
-        os: [ubuntu-latest, macos-latest]
-    runs-on: ${{ matrix.os }}
-    timeout-minutes: 30
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-      - uses: ./.github/actions/nix-setup
-        with:
-          cachix-auth-token: ${{ secrets.CACHIX_AUTH_TOKEN }}
-
-      - name: Resolve head SHA
-        if: github.event_name == 'pull_request'
-        id: sha
-        shell: bash
-        run: |
-          FULL="${{ github.event.pull_request.head.sha || github.sha }}"
-          echo "full=$FULL" >> "$GITHUB_OUTPUT"
-          echo "short=${FULL:0:7}" >> "$GITHUB_OUTPUT"
-
-      - name: Check flake
-        id: flake
-        continue-on-error: true
-        run: nix flake check --print-build-logs
-
-      # When the flake check fails, run a targeted diagnostic to see if
-      # the failure is specifically a stale npm lockfile hash in one of the
-      # known npm subpackages (tui / web).  This avoids surfacing a generic
-      # "build failed" message when the fix is a single known command.
-      - name: Diagnose npm lockfile hashes
-        id: hash_check
-        if: steps.flake.outcome == 'failure' && runner.os == 'Linux'
-        continue-on-error: true
-        env:
-          LINK_SHA: ${{ steps.sha.outputs.full }}
-        run: nix run .#fix-lockfiles -- --check
-
-      # If fix-lockfiles itself crashes (infrastructure blip, cache throttle,
-      # etc.) it won't set stale=true/false.  Treat that as a distinct failure
-      # mode rather than silently ignoring it.
-      - name: Fail if hash check crashed without reporting
-        if: steps.hash_check.outcome == 'failure' && steps.hash_check.outputs.stale != 'true' && steps.hash_check.outputs.stale != 'false'
-        run: |
-          echo "::error::fix-lockfiles exited without reporting stale status — likely an infrastructure or script failure"
-          exit 1
-
-      - name: Post sticky PR comment (stale hashes)
-        if: steps.hash_check.outputs.stale == 'true' && github.event_name == 'pull_request'
-        uses: marocchino/sticky-pull-request-comment@52423e01640425a022ef5fd42c6fb5f633a02728  # v2.9.1
-        with:
-          header: nix-lockfile-check
-          message: |
-            ### ⚠️ npm lockfile hash out of date
-
-            Checked against commit [`${{ 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`:
-
-            ${{ steps.hash_check.outputs.report }}
-
-            #### Apply the fix
-
-            - [ ] **Apply lockfile fix** — tick to push a commit with the correct hashes to this PR branch
-            - Or [run the Nix Lockfile Fix workflow](${{ 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` and commit the diff
-
-      # Clear the sticky comment when either the flake check passed outright (no
-      # hash check needed) or the hash check explicitly returned stale=false
-      # (check failed for a non-hash reason).
-      - name: Clear sticky PR comment (resolved)
-        if: |
-          github.event_name == 'pull_request' &&
-          (steps.hash_check.outputs.stale == 'false' ||
-           steps.flake.outcome == 'success')
-        uses: marocchino/sticky-pull-request-comment@52423e01640425a022ef5fd42c6fb5f633a02728  # v2.9.1
-        with:
-          header: nix-lockfile-check
-          delete: true
-
-      - name: Final fail if flake check failed
-        if: steps.flake.outcome == 'failure'
-        run: |
-          if [ "${{ steps.hash_check.outputs.stale }}" == "true" ]; then
-            echo "::error::Nix build failed due to stale npm lockfile hash. Run: nix run .#fix-lockfiles"
-          else
-            echo "::error::Nix flake check failed. See logs above."
-          fi
-          exit 1

.github/workflows/osv-scanner.yml

@@ -20,29 +20,23 @@ name: OSV-Scanner
 # vulnerabilities in pinned deps that we may need to patch deliberately.
 
 on:
+  # No paths filter — the job must always run so the required check
+  # reports a status (path-gated workflows leave checks "pending" forever
+  # when no matching files change, which blocks merge).
   pull_request:
     branches: [main]
-    paths:
-      - 'uv.lock'
-      - 'pyproject.toml'
-      - 'package.json'
-      - 'package-lock.json'
-      - 'ui-tui/package.json'
-      - 'website/package.json'
-      - 'website/package-lock.json'
-      - '.github/workflows/osv-scanner.yml'
   push:
     branches: [main]
     paths:
-      - 'uv.lock'
-      - 'pyproject.toml'
-      - 'package.json'
-      - 'package-lock.json'
-      - 'website/package-lock.json'
+      - "uv.lock"
+      - "pyproject.toml"
+      - "package.json"
+      - "package-lock.json"
+      - "website/package-lock.json"
   schedule:
     # Weekly scan against main — catches CVEs published after merge for
     # deps that haven't changed since.
-    - cron: '0 9 * * 1'
+    - cron: "0 9 * * 1"
   workflow_dispatch:
 
 permissions:
@@ -54,7 +48,7 @@ permissions:
 jobs:
   scan:
     name: Scan lockfiles
-    uses: google/osv-scanner-action/.github/workflows/osv-scanner-reusable.yml@9a498708959aeaef5ef730655706c5a1df1edbc2  # v2.3.8
+    uses: google/osv-scanner-action/.github/workflows/osv-scanner-reusable.yml@9a498708959aeaef5ef730655706c5a1df1edbc2 # v2.3.8
     with:
       # Scan explicit lockfiles rather than recursing, so we only look at
       # the three sources of truth and skip vendored / test / worktree dirs.

.github/workflows/skills-index.yml

@@ -53,4 +53,4 @@ jobs:
       - name: Trigger Deploy Site workflow
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: gh workflow run deploy-site.yml --repo ${{ github.repository }}
+        run: gh workflow run deploy-site.yml --repo ${{ github.repository }} -f skills_index_run_id=${{ github.run_id }}

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

@@ -1,11 +1,11 @@
 name: Supply Chain Audit
 
 on:
-  pull_request:
-    types: [opened, synchronize, reopened]
   # No paths filter — the jobs must always run so required checks
   # report a status (path-gated workflows leave checks "pending" forever
   # when no matching files change, which blocks merge).
+  pull_request:
+    types: [opened, synchronize, reopened]
 
 permissions:
   pull-requests: write
@@ -29,8 +29,10 @@ jobs:
       scan: ${{ steps.filter.outputs.scan }}
       # True when pyproject.toml changed in this PR
       deps: ${{ steps.filter.outputs.deps }}
+      # True when the curated MCP catalog / bundled MCP manifests changed.
+      mcp_catalog: ${{ steps.filter.outputs.mcp_catalog }}
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
         with:
           fetch-depth: 0
       - name: Check for relevant file changes
@@ -54,6 +56,14 @@ jobs:
           else
             echo "deps=false" >> "$GITHUB_OUTPUT"
           fi
+          MCP_CATALOG_FILES=$(git diff --name-only "$BASE"..."$HEAD" -- \
+            'optional-mcps/**' \
+            'hermes_cli/mcp_catalog.py' || true)
+          if [ -n "$MCP_CATALOG_FILES" ]; then
+            echo "mcp_catalog=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "mcp_catalog=false" >> "$GITHUB_OUTPUT"
+          fi
 
   scan:
     name: Scan PR for critical supply chain risks
@@ -62,7 +72,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
         with:
           fetch-depth: 0
 
@@ -197,7 +207,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
         with:
           fetch-depth: 0
 
@@ -268,3 +278,50 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - run: echo "No pyproject.toml changes, skipping dependency bounds check."
+
+  mcp-catalog-review:
+    name: MCP catalog security review
+    needs: changes
+    if: needs.changes.outputs.mcp_catalog == 'true'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        with:
+          fetch-depth: 0
+
+      - name: Require explicit MCP catalog review label
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -euo pipefail
+          PR="${{ github.event.pull_request.number }}"
+          LABELS=$(gh pr view "$PR" --json labels --jq '.labels[].name' || true)
+          if echo "$LABELS" | grep -Fxq 'mcp-catalog-reviewed'; then
+            echo "MCP catalog review label present."
+            exit 0
+          fi
+
+          BODY="## ⚠️ MCP catalog security review required
+
+          This PR changes the bundled MCP catalog or MCP catalog installer code. MCP entries can define local commands that users later install into \`mcp_servers\`, so this needs explicit maintainer review before merge.
+
+          A maintainer should verify:
+          - any new/changed \`optional-mcps/**/manifest.yaml\` command and args are expected,
+          - stdio transports do not use shell+egress/exfiltration payloads,
+          - git install refs are pinned and bootstrap commands are minimal,
+          - requested env vars/secrets match the upstream MCP's documented needs.
+
+          After review, add the \`mcp-catalog-reviewed\` label and re-run this check."
+
+          gh pr comment "$PR" --body "$BODY" || echo "::warning::Could not post PR comment (expected for fork PRs)"
+          echo "::error::MCP catalog changes require the mcp-catalog-reviewed label."
+          exit 1
+
+  mcp-catalog-review-gate:
+    name: MCP catalog security review
+    needs: changes
+    if: always() && needs.changes.outputs.mcp_catalog != 'true'
+    runs-on: ubuntu-latest
+    steps:
+      - run: echo "No MCP catalog changes, skipping MCP catalog security review."

.github/workflows/tests.yml

@@ -4,13 +4,13 @@ on:
   push:
     branches: [main]
     paths-ignore:
-      - '**/*.md'
-      - 'docs/**'
+      - "**/*.md"
+      - "docs/**"
+  # No paths filter — the job must always run so the required check
+  # reports a status (path-gated workflows leave checks "pending" forever
+  # when no matching files change, which blocks merge).
   pull_request:
     branches: [main]
-    paths-ignore:
-      - '**/*.md'
-      - 'docs/**'
 
 permissions:
   contents: read
@@ -30,13 +30,17 @@ jobs:
         slice: [1, 2, 3, 4, 5, 6]
     steps:
       - name: Checkout code
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
       - name: Restore duration cache
-        uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae # v5.0.5
         with:
           path: test_durations.json
-          # Single stable key. main always overwrites, PRs always find it.
+          # main always writes a new suffix, but jobs pick the latest one with the same prefix
+          # quote from https://docs.github.com/en/actions/reference/workflows-and-actions/dependency-caching#cache-hits-and-misses
+          # If you provide restore-keys, the cache action sequentially searches for any caches that match the list of restore-keys.
+          # If there are no exact matches, the action searches for partial matches of the restore keys.
+          # When the action finds a partial match, the most recent cache is restored to the path directory.
           key: test-durations
 
       - name: Install ripgrep (prebuilt binary)
@@ -54,16 +58,32 @@ jobs:
           rg --version
 
       - name: Install uv
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86  # v5
+        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5
+        with:
+          # Persist uv's download/wheel cache (~/.cache/uv) across runs.
+          # Keyed on the dependency manifests, so the cache is reused until
+          # pyproject.toml or uv.lock changes. `uv sync` still runs every
+          # time, but resolves from the warm cache instead of re-downloading
+          # and re-building wheels.
+          enable-cache: true
+          cache-dependency-glob: |
+            pyproject.toml
+            uv.lock
 
       - name: Set up Python 3.11
         run: uv python install 3.11
 
       - name: Install dependencies
-        run: |
-          uv venv .venv --python 3.11
-          source .venv/bin/activate
-          uv pip install -e ".[all,dev]"
+        # `uv sync --locked` installs the exact pinned set from uv.lock (and
+        # fails if the lock is out of sync with pyproject.toml), giving a
+        # reproducible env. It also creates .venv itself, so no separate
+        # `uv venv` step is needed.
+        run: uv sync --locked --python 3.11 --extra all --extra dev
+
+      - name: Minimize uv cache
+        # Optimized for CI: prunes pre-built wheels that are cheap to
+        # re-download, keeping the persisted cache small and fast to restore.
+        run: uv cache prune --ci
 
       - name: Run tests (slice ${{ matrix.slice }}/6)
         # Per-file isolation via scripts/run_tests_parallel.py: discovers
@@ -99,7 +119,7 @@ jobs:
           NOUS_API_KEY: ""
 
       - name: Upload per-slice durations
-        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
           name: test-durations-slice-${{ matrix.slice }}
           path: test_durations.json
@@ -109,11 +129,11 @@ jobs:
   # (including PRs) get balanced slicing.
   save-durations:
     needs: test
-    if: always() && github.ref == 'refs/heads/main'
+    if: needs.test.result == 'success' && github.ref == 'refs/heads/main'
     runs-on: ubuntu-latest
     steps:
       - name: Download all slice durations
-        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c  # v8.0.1
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
         with:
           pattern: test-durations-slice-*
           path: durations
@@ -133,17 +153,17 @@ jobs:
           "
 
       - name: Save merged duration cache
-        uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae # v5.0.5
         with:
           path: test_durations.json
-          key: test-durations
+          key: test-durations-${{ github.run_id }}
 
   e2e:
     runs-on: ubuntu-latest
     timeout-minutes: 15
     steps:
       - name: Checkout code
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
       - name: Install ripgrep (prebuilt binary)
         run: |
@@ -160,16 +180,32 @@ jobs:
           rg --version
 
       - name: Install uv
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86  # v5
+        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5
+        with:
+          # Persist uv's download/wheel cache (~/.cache/uv) across runs.
+          # Keyed on the dependency manifests, so the cache is reused until
+          # pyproject.toml or uv.lock changes. `uv sync` still runs every
+          # time, but resolves from the warm cache instead of re-downloading
+          # and re-building wheels.
+          enable-cache: true
+          cache-dependency-glob: |
+            pyproject.toml
+            uv.lock
 
       - name: Set up Python 3.11
         run: uv python install 3.11
 
       - name: Install dependencies
-        run: |
-          uv venv .venv --python 3.11
-          source .venv/bin/activate
-          uv pip install -e ".[all,dev]"
+        # `uv sync --locked` installs the exact pinned set from uv.lock (and
+        # fails if the lock is out of sync with pyproject.toml), giving a
+        # reproducible env. It also creates .venv itself, so no separate
+        # `uv venv` step is needed.
+        run: uv sync --locked --python 3.11 --extra all --extra dev
+
+      - name: Minimize uv cache
+        # Optimized for CI: prunes pre-built wheels that are cheap to
+        # re-download, keeping the persisted cache small and fast to restore.
+        run: uv cache prune --ci
 
       - name: Packaged-wheel i18n smoke test
         run: |
@@ -183,4 +219,4 @@ jobs:
         env:
           OPENROUTER_API_KEY: ""
           OPENAI_API_KEY: ""
-          NOUS_API_KEY: ""
+          NOUS_API_KEY: ""

.github/workflows/typecheck.yml

@@ -0,0 +1,45 @@
+# .github/workflows/typecheck.yml
+name: Typecheck
+
+on:
+  push:
+    branches: [main]
+  # No paths filter — the job must always run so the required check
+  # reports a status (path-gated workflows leave checks "pending" forever
+  # when no matching files change, which blocks merge).
+  pull_request:
+    branches: [main]
+
+jobs:
+  typecheck:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        package:
+          [ui-tui, web, apps/bootstrap-installer, apps/desktop, apps/shared]
+      fail-fast: false # report all failures, not just the first one
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
+        with:
+          node-version: 22
+          cache: npm
+      - run: npm ci
+      - run: npm run --prefix ${{ matrix.package }} typecheck
+
+  # Production build of the desktop renderer. `typecheck` runs `tsc` only,
+  # which does NOT exercise Vite/Rolldown module resolution — so an
+  # unresolvable package export (e.g. a transitive @assistant-ui/tap that no
+  # longer exports "./react-shim") slips past typecheck and only explodes when
+  # users build apps/desktop from source on install/update. Run the real
+  # `vite build` here so that class of break fails in CI instead.
+  desktop-build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
+        with:
+          node-version: 22
+          cache: npm
+      - run: npm ci
+      - run: npm run --prefix apps/desktop build

.github/workflows/uv-lockfile-check.yml

@@ -47,15 +47,15 @@ on:
   push:
     branches: [main]
     paths:
-      - 'pyproject.toml'
-      - 'uv.lock'
-      - '.github/workflows/uv-lockfile-check.yml'
+      - "pyproject.toml"
+      - "uv.lock"
+      - ".github/workflows/uv-lockfile-check.yml"
+
+  # No paths filter — the job must always run so the required check
+  # reports a status (path-gated workflows leave checks "pending" forever
+  # when no matching files change, which blocks merge).
   pull_request:
     branches: [main]
-    paths:
-      - 'pyproject.toml'
-      - 'uv.lock'
-      - '.github/workflows/uv-lockfile-check.yml'
 
 permissions:
   contents: read
@@ -71,10 +71,10 @@ jobs:
     timeout-minutes: 5
     steps:
       - name: Checkout code
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
       - name: Install uv
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86  # v5
+        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5
 
       # `uv lock --check` re-resolves the project from pyproject.toml and
       # compares the result to uv.lock, exiting non-zero if they disagree.

.gitignore

@@ -89,6 +89,9 @@ website/static/api/skills-index.json
 # every build).
 website/static/api/skills.json
 website/static/api/skills-meta.json
+# automation-blueprints-index.json is a build artifact emitted by
+# website/scripts/extract-automation-blueprints.py during prebuild.
+website/static/api/automation-blueprints-index.json
 models-dev-upstream/
 
 # Local editor / agent tooling (machine-specific; keep in global config, not the repo)
@@ -114,6 +117,12 @@ docs/superpowers/*
 # treat it as a local edit and autostash it on every run (#38529).
 .hermes-bootstrap-complete
 
+# Interrupted-update breadcrumb + recovery lock written next to the shared venv
+# by `hermes update` / launch-time self-heal. Runtime state, never a code change
+# — ignore so `git status` stays clean and update's autostash skips them.
+.update-incomplete
+.update-incomplete.lock
+
 # Tool Search live-test harness output — non-deterministic model transcripts,
 # regenerated by scripts/tool_search_livetest.py. Never an artifact of the repo.
 scripts/out/
@@ -123,3 +132,7 @@ scripts/out/
 # stores the published notes. They are not a build artifact and must never be
 # committed to the repo root. See the hermes-release skill.
 RELEASE_v*.md
+
+# Desktop demo-run scratch output (hermes writes demo/*.txt during recorded
+# walkthroughs). Throwaway artifacts, never part of the app.
+apps/desktop/demo/

AGENTS.md

@@ -4,6 +4,201 @@ Instructions for AI coding assistants and developers working on the hermes-agent
 
 **Never give up on the right solution.**
 
+## What Hermes Is
+
+Hermes is a personal AI agent that runs the same agent core across a CLI, a
+messaging gateway (Telegram, Discord, Slack, and ~20 other platforms), a TUI,
+and an Electron desktop app. It learns across sessions (memory + skills),
+delegates to subagents, runs scheduled jobs, and drives a real terminal and
+browser. It is extended primarily through **plugins and skills**, not by
+growing the core.
+
+Two properties shape almost every design decision and are the lens for
+reviewing any change:
+
+- **Per-conversation prompt caching is sacred.** A long-lived conversation
+  reuses a cached prefix every turn. Anything that mutates past context,
+  swaps toolsets, or rebuilds the system prompt mid-conversation invalidates
+  that cache and multiplies the user's cost. We do not do it (the one
+  exception is context compression).
+- **The core is a narrow waist; capability lives at the edges.** Every model
+  tool we add is sent on every API call, so the bar for a new *core* tool is
+  high. Most new capability should arrive as a CLI command + skill, a
+  service-gated tool, or a plugin — not as core surface.
+
+## Contribution Rubric — What We Want / What We Don't
+
+This is the project's intent layer. Use it two ways:
+
+1. **For humans and for your own work** — what gets merged and what gets
+   rejected, so a contribution aims at the target.
+2. **For automated review (the triage sweeper)** — guidance on when a PR is
+   safe to close on the three allowed reasons (`implemented_on_main`,
+   `cannot_reproduce`, `incoherent`) and, just as important, **when NOT to
+   close** one. Taste-based "we don't want this / out of scope" closes are NOT
+   an automated decision — those stay with a human maintainer. The sweeper's
+   job here is to recognize design intent and *avoid wrongly closing a
+   legitimate contribution*, not to make the won't-implement call itself.
+
+Read the balance right: Hermes ships a **lot** — most merges are bug fixes to
+real reported behavior, and the product surface (platforms, channels,
+providers, models, desktop/TUI features) expands aggressively and on purpose.
+The restraint below is aimed squarely at the **core agent + the model tool
+schema**, the one place where every addition is paid for on every API call.
+"Smallest footprint" governs *how a capability is wired into the core*, NOT
+whether the product is allowed to grow. We are expansive at the edges and
+conservative at the waist.
+
+### What we want
+
+- **Fix real bugs, well.** The bulk of what lands is `fix(...)` against an
+  actual reported symptom. A good fix reproduces the symptom on current
+  `main`, points to the exact line where it manifests, and fixes the whole bug
+  class — sibling call paths included — not just the one site the reporter hit.
+- **Expand reach at the edges.** New platform adapters, channels, providers,
+  models, and desktop/TUI/dashboard features are welcome and land routinely,
+  including large ones (a new messaging channel, a session-cap feature, a
+  Windows PTY bridge). Breadth in the product is a goal, not a footprint
+  concern — as long as it integrates with the existing setup/config UX
+  (`hermes tools`, `hermes setup`, auto-install) rather than bolting on a raw
+  env var.
+- **Refactor god-files into clean modules.** Extracting a multi-thousand-line
+  cluster out of `cli.py` / `run_agent.py` / `gateway/run.py` into a focused
+  mixin or module is wanted work, even when the diff is huge and mechanical
+  (large `+N/-N` refactors merge regularly). The "every line traces to the
+  request" test applies to *feature* PRs; a declared refactor's request IS the
+  extraction.
+- **Keep the core narrow.** New *model tools* are the expensive exception —
+  every tool ships on every API call. Prefer, in order: extend existing code →
+  CLI command + skill → service-gated tool (`check_fn`) → plugin → MCP server
+  in the catalog → new core tool (last resort). See "The Footprint Ladder."
+- **Extend, don't duplicate.** Before adding a module/manager/hook, check
+  whether existing infrastructure already covers the use case. When several PRs
+  integrate the same *category*, design one shared interface instead of merging
+  them one at a time (see the ABC + orchestrator note under the Footprint
+  Ladder).
+- **Behavior contracts over snapshots.** Tests should assert how two pieces of
+  data must relate (invariants), not freeze a current value (model lists,
+  config version literals, enumeration counts). See "Don't write
+  change-detector tests."
+- **E2E validation, not just green unit mocks.** For anything touching
+  resolution chains, config propagation, security boundaries, remote
+  backends, or file/network I/O, exercise the real path with real imports
+  against a temp `HERMES_HOME`. Mocks hide integration bugs.
+- **Cache-, alternation-, and invariant-safe.** Preserve prompt caching, strict
+  message role alternation (never two same-role messages in a row; never a
+  synthetic user message injected mid-loop), and a system prompt that is
+  byte-stable for the life of a conversation.
+- **Contributor credit preserved.** Salvage external work by cherry-picking
+  (rebase-merge) so authorship survives in git history; don't reimplement from
+  scratch when you can build on top.
+
+### What we don't want (rejected even when well-built)
+
+- **Speculative infrastructure.** Hooks, callbacks, or extension points with no
+  concrete consumer. Adding a hook is easy; removing one after plugins depend
+  on it is hard. A hook is NOT speculative if a contributor has a real, stated
+  use case — even if the consumer ships separately.
+- **New `HERMES_*` env vars for non-secret config.** `.env` is for secrets
+  only (API keys, tokens, passwords). All behavioral settings — timeouts,
+  thresholds, feature flags, display prefs — go in `config.yaml`. Bridge to an
+  internal env var if the mechanism needs one, but user-facing docs point to
+  `config.yaml`. Reject PRs that tell users to "set X in your .env" unless X
+  is a credential.
+- **A new core tool when terminal + file already do the job, or when a skill
+  would.** If the only barrier is file visibility on a remote backend, fix the
+  mount, not the toolset.
+- **Lazy-reading escape hatches on instructional tools.** No `offset`/`limit`
+  pagination on tools that load content the agent must read fully (skills,
+  prompts, playbooks). Models will read page 1 and skip the rest.
+- **"Fixes" that destroy the feature they secure.** A mitigation that kills the
+  feature's purpose is the wrong mitigation. Read the original commit's intent
+  (`git log -p -S`) before restricting behavior; find a fix that preserves the
+  feature.
+- **Outbound telemetry / usage attribution without opt-in gating.** No new
+  analytics, third-party identifier tagging, or attribution tags until a
+  generic user-facing opt-in (config gate + setup prompt + `hermes tools`
+  toggle) exists. Park behind a label, do not merge.
+- **Change-detector tests, cache-breaking mid-conversation, dead code wired in
+  without E2E proof, and plugins that touch core files.** Plugins live in their
+  own directory and work within the ABCs/hooks we provide; if a plugin needs
+  more, widen the generic plugin surface, don't special-case it in core.
+
+### Before you call it a bug — verify the premise (and when NOT to close)
+
+The most common reason a well-written PR gets closed is not code quality — it
+is that the change is built on a **wrong premise**, or it treats an
+**intentional design as a gap**. These patterns cut both ways: they tell a
+human reviewer what to scrutinize, and they tell the automated sweeper when a
+PR is NOT safe to close as `implemented_on_main` / `cannot_reproduce` (when in
+doubt, leave it open for a human). They are distilled from real closes.
+
+- **"Intentional design, not a gap."** A limitation that looks like an
+  oversight is often deliberate. Before "fixing" a missing link or a
+  restriction, ask whether the isolation IS the design. Example: profiles are
+  independent islands on purpose — a PR adding live config inheritance from the
+  default profile was closed because coupling profiles together is exactly what
+  the design prevents (the copy-at-creation `--clone` path already covers the
+  legitimate "start from my default" case). Read the original commit's intent
+  (`git log -p -S "<symbol>"`) before assuming something is unfinished.
+- **"The premise doesn't hold against how X actually works."** A PR's
+  justification frequently rests on a wrong mental model of an existing
+  mechanism. Trace the real code/runtime before accepting the rationale. Two
+  real closes: a rate-limit "re-probe during cooldown" PR (the breaker only
+  trips on a *confirmed-empty* account bucket, so re-probing just hammers a
+  bucket we've already proven empty); a usage-accumulation fix whose new branch
+  **never executes at runtime** because an earlier guard already popped the
+  state it depended on. If you can't point to the exact line where the bug
+  manifests AND show the fix changes that line's behavior, you haven't verified
+  the premise.
+- **"This fix was wrong — the absence/omission was deliberate."** Adding the
+  obvious-looking missing piece can break things the omission was protecting.
+  Example: restoring "missing" `__init__.py` files made a test tree importable
+  as a dotted package that shadowed the real plugin, deleting its `register()`
+  at import time. The absence was load-bearing.
+- **"Overreached / resurrected an approach we'd moved past."** Scope creep that
+  supersedes an agreed-on base, or revives a direction the maintainers
+  deliberately closed, gets rejected even when the code works. Keep the change
+  to the narrow piece that was actually agreed; offer the rest as a focused
+  follow-up.
+
+The throughline: **verify the claim AND the intent against the codebase before
+writing or merging a fix.** A confirmed reproduction on current `main` plus a
+line-level account of where the fix acts beats a plausible-sounding rationale
+every time. When in doubt about intent, it is cheaper to ask than to ship a
+fix that fights the design.
+
+### The Footprint Ladder (new capability decision)
+
+Each rung adds more permanent surface than the one above. Choose the highest
+(least-footprint) rung that correctly solves the problem:
+
+1. **Extend existing code** — the capability is a variation of something that
+   already exists. Zero new surface.
+2. **CLI command + skill** — manages config/state/infra expressible as shell
+   commands. The agent runs `hermes <subcommand>` guided by a skill. Zero
+   model-tool footprint. Default choice for subscriptions, scheduled tasks,
+   service setup. Examples: `hermes webhook`, `hermes cron`, `hermes tools`.
+3. **Service-gated tool (`check_fn`)** — needs structured params/returns AND
+   only appears when a prerequisite is configured. Zero footprint otherwise.
+   Examples: Home Assistant tools (gated on token), memory-provider tools.
+4. **Plugin** — third-party/niche/user-specific capability that doesn't ship in
+   core. Lives in `~/.hermes/plugins/` or a pip package, discovered at runtime.
+5. **MCP server (in the catalog)** — if the capability genuinely needs to be a
+   tool (structured I/O the agent invokes) but isn't core-fundamental, prefer
+   building it as an MCP server and adding it to the MCP catalog over growing
+   the core toolset. The agent connects to it through the built-in MCP client;
+   zero permanent core-schema footprint, and it's reusable by any MCP host.
+6. **New core tool** — only when the capability is fundamental, broadly useful
+   to nearly every user, and unreachable via terminal + file (or an MCP server).
+   Examples of correct core tools: terminal, read_file, web_search,
+   browser_navigate.
+
+When 3+ open PRs try to integrate the same *category* of thing (memory
+backends, providers, notifiers), don't merge them one at a time — design an
+ABC + orchestrator, wrap the existing built-in as the first provider, and turn
+the competing PRs into plugins against that interface.
+
 ## Development Environment
 
 ```bash
@@ -264,7 +459,7 @@ npm install       # first time
 npm run dev       # watch mode (rebuilds hermes-ink + tsx --watch)
 npm start         # production
 npm run build     # full build (hermes-ink + tsc)
-npm run type-check # typecheck only (tsc --noEmit)
+npm run typecheck # typecheck only (tsc --noEmit)
 npm run lint      # eslint
 npm run fmt       # prettier
 npm test          # vitest
@@ -302,9 +497,11 @@ A **separate** chat surface from both the classic CLI and the dashboard's embedd
 
 ## Adding New Tools
 
-For most custom or local-only tools, do **not** edit Hermes core. Use the plugin
-route instead: create `~/.hermes/plugins/<name>/plugin.yaml` and
-`~/.hermes/plugins/<name>/__init__.py`, then register tools with
+Before adding any tool, settle the footprint question first (see "The
+Footprint Ladder" in the Contribution Rubric): most capabilities should NOT
+be core tools. For custom or local-only tools, do **not** edit Hermes core.
+Use the plugin route instead: create `~/.hermes/plugins/<name>/plugin.yaml`
+and `~/.hermes/plugins/<name>/__init__.py`, then register tools with
 `ctx.register_tool(...)`. Plugin toolsets are discovered automatically and can be
 enabled or disabled without touching `tools/` or `toolsets.py`.
 

CONTRIBUTING.md

@@ -78,7 +78,41 @@ This isn't a quality bar — it's a coupling-and-maintenance decision. Memory pr
 | **uv** | Fast Python package manager ([install](https://docs.astral.sh/uv/)) |
 | **Node.js 20+** | Optional — needed for browser tools and WhatsApp bridge (matches root `package.json` engines) |
 
-### Clone and install
+### Install with the standard installer
+
+For most contributors, the best development bootstrap is the same path users
+take: run the standard installer, then work inside the repository it cloned.
+The installer creates the Hermes venv, wires the `hermes` command, stamps the
+install method for `hermes update`, and clones the full git project into
+`$HERMES_HOME/hermes-agent` (usually `~/.hermes/hermes-agent`). That keeps your
+development environment on the same layout the CLI, updater, lazy dependency
+installer, gateway, and docs assume.
+
+```bash
+curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
+cd "${HERMES_HOME:-$HOME/.hermes}/hermes-agent"
+
+# Add dev/test extras on top of the standard install.
+uv pip install -e ".[all,dev]"
+
+# Optional: browser tools / docs site dependencies.
+npm install
+```
+
+After that, create branches and run tests from that checkout:
+
+```bash
+git checkout -b fix/description
+scripts/run_tests.sh
+```
+
+### Manual clone fallback
+
+Use this only if you intentionally do not want Hermes' managed install layout
+(for example, a throwaway clone inside a container or CI job). If you install
+this way, make sure you run the `hermes` entrypoint from this venv; running the
+system `python3 -m hermes_cli.main` can pick up unrelated system Python
+packages.
 
 ```bash
 git clone https://github.com/NousResearch/hermes-agent.git
@@ -109,15 +143,19 @@ echo "OPENROUTER_API_KEY=***" >> ~/.hermes/.env
 ### Run
 
 ```bash
-# Symlink for global access
-mkdir -p ~/.local/bin
-ln -sf "$(pwd)/venv/bin/hermes" ~/.local/bin/hermes
-
-# Verify
+# The standard installer already put `hermes` on PATH.
 hermes doctor
 hermes chat -q "Hello"
 ```
 
+If you used the manual clone fallback, run `./hermes` from the checkout or
+symlink this clone's venv explicitly:
+
+```bash
+mkdir -p ~/.local/bin
+ln -sf "$(pwd)/venv/bin/hermes" ~/.local/bin/hermes
+```
+
 ### Run tests
 
 ```bash

Dockerfile

@@ -25,7 +25,7 @@ ENV PLAYWRIGHT_BROWSERS_PATH=/opt/hermes/.playwright
 # hermes process, the dashboard, and per-profile gateways.
 RUN apt-get update && \
     apt-get install -y --no-install-recommends \
-    ca-certificates curl iputils-ping python3 python-is-python3 ripgrep ffmpeg gcc python3-dev python3-venv libffi-dev libolm-dev procps git openssh-client docker-cli xz-utils && \
+    ca-certificates curl iputils-ping python3 python-is-python3 ripgrep ffmpeg gcc g++ make cmake python3-dev python3-venv libffi-dev libolm-dev procps git openssh-client docker-cli xz-utils && \
     rm -rf /var/lib/apt/lists/*
 
 # ---------- s6-overlay install ----------
@@ -146,9 +146,9 @@ RUN npm install --prefer-offline --no-audit && \
 #
 # `uv sync --frozen --no-install-project --extra all --extra messaging`
 # installs the deps reachable through the composite `[all]` extra
-# (handpicked set intended for the production image), plus gateway
-# messaging adapters that should work in the published image without a
-# first-boot lazy install.  We do NOT use `--all-extras`:
+# (handpicked set intended for the production image — excludes `[dev]`),
+# plus gateway messaging adapters that should work in the published image
+# without a first-boot lazy install.  We do NOT use `--all-extras`:
 # that would pull in `[rl]` (atroposlib + tinker + torch + wandb from
 # git), `[yc-bench]` (another git dep), and `[termux-all]` (Android
 # redundancy), none of which belong in the published container.
@@ -164,19 +164,30 @@ RUN npm install --prefer-offline --no-audit && \
 # image update and recall/retain then fails with
 # `ModuleNotFoundError: No module named 'hindsight_client'` (#38128).
 #
+# The Matrix gateway's deps ([matrix] extra) are baked in because
+# python-olm (transitive via mautrix[encryption]) builds from source on
+# Python/image combinations without usable wheels.  The Docker image is
+# Linux-only, so keeping the native libolm/build-toolchain packages here
+# avoids the cross-platform failures that kept [matrix] out of [all]
+# while still making Matrix work in the published container. Fixes #30399.
+#
 # The editable link is created after the source copy below.
 COPY pyproject.toml uv.lock ./
 RUN touch ./README.md
-RUN uv sync --frozen --no-install-project --extra all --extra messaging --extra anthropic --extra bedrock --extra azure-identity --extra hindsight
+RUN uv sync --frozen --no-install-project --extra all --extra messaging --extra anthropic --extra bedrock --extra azure-identity --extra hindsight --extra matrix
+
+# ---------- Frontend build (cached independently from Python source) ----------
+# Copy only the frontend source trees first so that Python-only changes don't
+# invalidate the (relatively slow) web + ui-tui build layer.
+COPY web/ web/
+COPY ui-tui/ ui-tui/
+RUN cd web && npm run build && \
+    cd ../ui-tui && npm run build
 
 # ---------- Source code ----------
 # .dockerignore excludes node_modules, so the installs above survive.
 COPY --chown=hermes:hermes . .
 
-# Build browser dashboard and terminal UI assets.
-RUN cd web && npm run build && \
-    cd ../ui-tui && npm run build
-
 # ---------- Permissions ----------
 # Make install dir world-readable so any HERMES_UID can read it at runtime.
 # The venv needs to be traversable too.

MANIFEST.in

@@ -1,5 +1,6 @@
 graft skills
 graft optional-skills
+graft optional-mcps
 graft locales
 # Bundled plugin manifests (plugin.yaml / plugin.yml). Without these the
 # PluginManager scan (hermes_cli/plugins.py) finds zero plugins on installs

README.md

@@ -3,7 +3,9 @@
 </p>
 
 # Hermes Agent ☤
-
+<p align="center">
+  <a href="https://hermes-agent.nousresearch.com/">Hermes Agent</a> | <a href="https://hermes-agent.nousresearch.com/">Hermes Desktop</a>
+</p>
 <p align="center">
   <a href="https://hermes-agent.nousresearch.com/docs/"><img src="https://img.shields.io/badge/Docs-hermes--agent.nousresearch.com-FFD700?style=for-the-badge" alt="Documentation"></a>
   <a href="https://discord.gg/NousResearch"><img src="https://img.shields.io/badge/Discord-5865F2?style=for-the-badge&logo=discord&logoColor=white" alt="Discord"></a>
@@ -179,16 +181,20 @@ See `hermes claw migrate --help` for all options, or use the `openclaw-migration
 
 We welcome contributions! See the [Contributing Guide](https://hermes-agent.nousresearch.com/docs/developer-guide/contributing) for development setup, code style, and PR process.
 
-Quick start for contributors — clone and go with `setup-hermes.sh`:
+Quick start for contributors — use the standard installer, then work from the
+full git checkout it creates at `$HERMES_HOME/hermes-agent` (usually
+`~/.hermes/hermes-agent`). This matches the layout used by `hermes update`, the
+managed venv, lazy dependencies, gateway, and docs tooling.
 
 ```bash
-git clone https://github.com/NousResearch/hermes-agent.git
-cd hermes-agent
-./setup-hermes.sh     # installs uv, creates venv, installs .[all], symlinks ~/.local/bin/hermes
-./hermes              # auto-detects the venv, no need to `source` first
+curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
+cd "${HERMES_HOME:-$HOME/.hermes}/hermes-agent"
+uv pip install -e ".[all,dev]"
+scripts/run_tests.sh
 ```
 
-Manual path (equivalent to the above):
+Manual clone fallback (for throwaway clones/CI where you intentionally do not
+want the managed install layout):
 
 ```bash
 curl -LsSf https://astral.sh/uv/install.sh | sh

README.zh-CN.md

@@ -164,16 +164,18 @@ hermes claw migrate --overwrite  # 覆盖已有冲突
 
 欢迎贡献！请参阅 [贡献指南](https://hermes-agent.nousresearch.com/docs/developer-guide/contributing) 了解开发设置、代码风格和 PR 流程。
 
-贡献者快速开始——克隆并使用 `setup-hermes.sh`：
+贡献者快速开始——使用标准安装器，然后在它创建的完整 git checkout 中开发：
+`$HERMES_HOME/hermes-agent`（通常是 `~/.hermes/hermes-agent`）。这会匹配
+`hermes update`、托管 venv、lazy dependencies、gateway 和 docs tooling 使用的布局。
 
 ```bash
-git clone https://github.com/NousResearch/hermes-agent.git
-cd hermes-agent
-./setup-hermes.sh     # 安装 uv、创建 venv、安装 .[all]、创建符号链接 ~/.local/bin/hermes
-./hermes              # 自动检测 venv，无需先 source
+curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
+cd "${HERMES_HOME:-$HOME/.hermes}/hermes-agent"
+uv pip install -e ".[all,dev]"
+scripts/run_tests.sh
 ```
 
-手动安装（等效于上述命令）：
+手动克隆备用路径（用于一次性 clone / CI，或你明确不想使用 managed install layout 时）：
 
 ```bash
 curl -LsSf https://astral.sh/uv/install.sh | sh

acp_adapter/server.py

@@ -824,6 +824,7 @@ class HermesACPAgent(acp.Agent):
 
         try:
             from model_tools import get_tool_definitions
+            from agent.memory_manager import inject_memory_provider_tools
 
             enabled_toolsets = _expand_acp_enabled_toolsets(
                 getattr(state.agent, "enabled_toolsets", None) or ["hermes-acp"],
@@ -839,6 +840,7 @@ class HermesACPAgent(acp.Agent):
             state.agent.valid_tool_names = {
                 tool["function"]["name"] for tool in state.agent.tools or []
             }
+            inject_memory_provider_tools(state.agent)
             invalidate = getattr(state.agent, "_invalidate_system_prompt", None)
             if callable(invalidate):
                 invalidate()
@@ -1779,10 +1781,25 @@ class HermesACPAgent(acp.Agent):
     def _cmd_tools(self, args: str, state: SessionState) -> str:
         try:
             from model_tools import get_tool_definitions
+            from types import SimpleNamespace
+            from agent.memory_manager import inject_memory_provider_tools
+
             toolsets = _expand_acp_enabled_toolsets(
                 getattr(state.agent, "enabled_toolsets", None) or ["hermes-acp"]
             )
             tools = get_tool_definitions(enabled_toolsets=toolsets, quiet_mode=True)
+            tool_view = SimpleNamespace(
+                tools=list(tools or []),
+                valid_tool_names={
+                    tool.get("function", {}).get("name")
+                    for tool in tools or []
+                    if isinstance(tool, dict)
+                },
+                enabled_toolsets=toolsets,
+                _memory_manager=getattr(state.agent, "_memory_manager", None),
+            )
+            inject_memory_provider_tools(tool_view)
+            tools = tool_view.tools
             if not tools:
                 return "No tools available."
             lines = [f"Available tools ({len(tools)}):"]

agent/account_usage.py

@@ -145,7 +145,7 @@ def build_nous_credits_snapshot(account_info) -> Optional[AccountUsageSnapshot]:
     account info to show (fail-open: caller just shows nothing).
     """
     try:
-        from hermes_cli.nous_account import nous_portal_billing_url
+        from hermes_cli.nous_account import nous_portal_topup_url
 
         if account_info is None or not getattr(account_info, "logged_in", False):
             return None
@@ -213,7 +213,8 @@ def build_nous_credits_snapshot(account_info) -> Optional[AccountUsageSnapshot]:
         if not windows and not details:
             return None
 
-        details.append(f"Manage / top up: {nous_portal_billing_url(account_info)}")
+        details.append(f"Top up: {nous_portal_topup_url(account_info)}")
+        details.append("(or run /credits)")
 
         plan = getattr(sub, "plan", None) if sub is not None else None
         return AccountUsageSnapshot(
@@ -337,6 +338,93 @@ def _snapshot_from_credits_state(state) -> Optional[AccountUsageSnapshot]:
         return None
 
 
+@dataclass(frozen=True)
+class CreditsView:
+    """Surface-agnostic data for the ``/credits`` command.
+
+    One portal fetch, one parse — consumed identically by the CLI panel, the
+    gateway button, and any other money surface. Fail-open: when not logged in
+    or the portal is unreachable, ``logged_in`` is False / ``topup_url`` is None
+    and callers degrade gracefully.
+    """
+
+    logged_in: bool
+    balance_lines: tuple[str, ...] = ()
+    identity_line: Optional[str] = None
+    topup_url: Optional[str] = None
+    depleted: bool = False
+
+
+def build_credits_view(*, markdown: bool = False, timeout: float = 10.0) -> CreditsView:
+    """Build the /credits view: balance block + identity line + top-up URL.
+
+    Reuses the same account fetch + snapshot + URL builder as the /usage credits
+    block, so the numbers always match. The balance block is the rendered
+    snapshot MINUS its trailing top-up/command-hint lines (the /credits surface
+    supplies its own affordance). Fail-open → ``CreditsView(logged_in=False)``.
+    """
+    not_logged_in = CreditsView(logged_in=False)
+    try:
+        from hermes_cli.auth import get_provider_auth_state
+
+        tok = (get_provider_auth_state("nous") or {}).get("access_token")
+        if not (isinstance(tok, str) and tok.strip()):
+            return not_logged_in
+    except Exception:
+        return not_logged_in
+
+    try:
+        import concurrent.futures
+
+        from hermes_cli.nous_account import (
+            get_nous_portal_account_info,
+            nous_portal_topup_url,
+        )
+
+        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
+            account = pool.submit(get_nous_portal_account_info, force_fresh=True).result(
+                timeout=timeout
+            )
+    except Exception:
+        logger.debug("credits ▸ /credits portal fetch failed (fail-open)", exc_info=True)
+        return not_logged_in
+
+    if account is None or not getattr(account, "logged_in", False):
+        return not_logged_in
+
+    snapshot = build_nous_credits_snapshot(account)
+    # Balance lines = the snapshot block minus the two trailing affordance lines
+    # ("Top up: <url>" + "(or run /credits)") that build_nous_credits_snapshot
+    # appends for the /usage surface. /credits renders its own button/panel.
+    balance_lines: list[str] = []
+    if snapshot is not None:
+        rendered = render_account_usage_lines(snapshot, markdown=markdown)
+        balance_lines = [
+            line
+            for line in rendered
+            if not line.lstrip().startswith("Top up:")
+            and not line.lstrip().startswith("(or run")
+        ]
+
+    # Identity line — shown before any open (roadmap §4.4).
+    email = getattr(account, "email", None)
+    org_name = getattr(account, "org_name", None)
+    who: list[str] = []
+    if email:
+        who.append(str(email))
+    if org_name:
+        who.append(f"org {org_name}")
+    identity_line = ("Topping up as " + " / ".join(who)) if who else None
+
+    return CreditsView(
+        logged_in=True,
+        balance_lines=tuple(balance_lines),
+        identity_line=identity_line,
+        topup_url=nous_portal_topup_url(account),
+        depleted=getattr(account, "paid_service_access", None) is False,
+    )
+
+
 def _resolve_codex_usage_url(base_url: str) -> str:
     normalized = (base_url or "").strip().rstrip("/")
     if not normalized:

agent/agent_init.py

@@ -27,7 +27,7 @@ import threading
 import time
 import uuid
 from datetime import datetime
-from typing import Any, Dict, List, Optional
+from typing import Any, Callable, Dict, List, Optional
 from urllib.parse import urlparse, parse_qs, urlunparse
 
 from agent.context_compressor import ContextCompressor
@@ -187,6 +187,7 @@ def init_agent(
     thinking_callback: callable = None,
     reasoning_callback: callable = None,
     clarify_callback: callable = None,
+    read_terminal_callback: callable = None,
     step_callback: callable = None,
     stream_delta_callback: callable = None,
     interim_assistant_callback: callable = None,
@@ -194,6 +195,7 @@ def init_agent(
     status_callback: callable = None,
     notice_callback: callable = None,
     notice_clear_callback: callable = None,
+    event_callback: Optional[Callable[[str, dict], None]] = None,
     max_tokens: int = None,
     reasoning_config: Dict[str, Any] = None,
     service_tier: str = None,
@@ -298,6 +300,7 @@ def init_agent(
     # would mangle the escape sequences.  None = use builtins.print.
     agent._print_fn = None
     agent.background_review_callback = None  # Optional sync callback for gateway delivery
+    agent.memory_notifications = "on"  # Memory update notifications: "off", "on", "verbose"
     agent.skip_context_files = skip_context_files
     agent.load_soul_identity = load_soul_identity
     agent.pass_session_id = pass_session_id
@@ -417,12 +420,14 @@ def init_agent(
     agent.thinking_callback = thinking_callback
     agent.reasoning_callback = reasoning_callback
     agent.clarify_callback = clarify_callback
+    agent.read_terminal_callback = read_terminal_callback
     agent.step_callback = step_callback
     agent.stream_delta_callback = stream_delta_callback
     agent.interim_assistant_callback = interim_assistant_callback
     agent.status_callback = status_callback
     agent.notice_callback = notice_callback
     agent.notice_clear_callback = notice_clear_callback
+    agent.event_callback = event_callback
     agent.tool_gen_callback = tool_gen_callback
 
     
@@ -898,6 +903,9 @@ def init_agent(
         agent.api_key = client_kwargs.get("api_key", "")
         agent.base_url = client_kwargs.get("base_url", agent.base_url)
         try:
+            from agent.ssl_guard import verify_ca_bundle_with_fallback
+
+            verify_ca_bundle_with_fallback()
             agent.client = agent._create_openai_client(client_kwargs, reason="agent_init", shared=True)
             if not agent.quiet_mode:
                 print(f"🤖 AI Agent initialized with model: {agent.model}")
@@ -1191,38 +1199,8 @@ def init_agent(
             _ra().logger.warning("Memory provider plugin init failed: %s", _mpe)
             agent._memory_manager = None
 
-    # Inject memory provider tool schemas into the tool surface.
-    # Skip tools whose names already exist (plugins may register the
-    # same tools via ctx.register_tool(), which lands in agent.tools
-    # through _ra().get_tool_definitions()).  Duplicate function names cause
-    # 400 errors on providers that enforce unique names (e.g. Xiaomi
-    # MiMo via Nous Portal).
-    #
-    # Respect the platform's enabled_toolsets configuration (#5544):
-    #   enabled_toolsets is None        → no filter, inject (backward compat)
-    #   "memory" in enabled_toolsets    → user opted in, inject
-    #   otherwise (incl. [])            → user excluded memory, skip injection
-    #
-    # Without this gate, `platform_toolsets: telegram: []` still leaks memory
-    # provider tools (fact_store, etc.) into the tool surface — a 10x latency
-    # penalty on local models and a frequent trigger of tool-call loops.
-    if agent._memory_manager and agent.tools is not None and (
-        agent.enabled_toolsets is None or "memory" in agent.enabled_toolsets
-    ):
-        _existing_tool_names = {
-            t.get("function", {}).get("name")
-            for t in agent.tools
-            if isinstance(t, dict)
-        }
-        for _schema in agent._memory_manager.get_all_tool_schemas():
-            _tname = _schema.get("name", "")
-            if _tname and _tname in _existing_tool_names:
-                continue  # already registered via plugin path
-            _wrapped = {"type": "function", "function": _schema}
-            agent.tools.append(_wrapped)
-            if _tname:
-                agent.valid_tool_names.add(_tname)
-                _existing_tool_names.add(_tname)
+    from agent.memory_manager import inject_memory_provider_tools as _inject_memory_provider_tools
+    _inject_memory_provider_tools(agent)
 
     # Skills config: nudge interval for skill creation reminders
     agent._skill_nudge_interval = 10

agent/agent_runtime_helpers.py

@@ -49,7 +49,7 @@ def _ra():
 
 
 AGENT_RUNTIME_POST_HOOK_TOOL_NAMES = frozenset(
-    {"todo", "session_search", "memory", "clarify", "delegate_task"}
+    {"todo", "session_search", "memory", "clarify", "read_terminal", "delegate_task"}
 )
 
 
@@ -445,6 +445,45 @@ def repair_message_sequence(agent, messages: List[Dict]) -> int:
     return repairs
 
 
+def repair_message_sequence_with_cursor(agent, messages: List[Dict]) -> int:
+    """Run :func:`repair_message_sequence` and keep the SessionDB flush
+    cursor consistent with the compacted list (#44837).
+
+    ``repair_message_sequence`` merges/drops messages in place, shrinking
+    the list. ``_last_flushed_db_idx`` (the DB-write cursor) indexes into
+    that list, so after compaction it can point past the new end — the
+    turn-end flush would then skip the assistant/tool chain entirely — or
+    past unflushed messages shifted to lower indexes.
+
+    Repair preserves object identity for surviving messages, so counting
+    the survivors from the previously-flushed prefix gives the exact new
+    cursor even when messages are dropped/merged at indexes *before* the
+    cursor — a plain ``min()`` clamp would silently skip that many
+    unflushed rows. Falls back to the clamp when no prefix snapshot is
+    available.
+
+    Returns the number of repairs made (same as ``repair_message_sequence``).
+    """
+    pre_repair_flushed_ids = None
+    flush_cursor = getattr(agent, "_last_flushed_db_idx", None)
+    if isinstance(flush_cursor, int) and flush_cursor > 0:
+        pre_repair_flushed_ids = {id(m) for m in messages[:flush_cursor]}
+
+    repairs = repair_message_sequence(agent, messages)
+
+    if repairs > 0 and hasattr(agent, "_last_flushed_db_idx"):
+        if pre_repair_flushed_ids is not None:
+            agent._last_flushed_db_idx = sum(
+                1 for m in messages if id(m) in pre_repair_flushed_ids
+            )
+        else:
+            agent._last_flushed_db_idx = min(
+                agent._last_flushed_db_idx, len(messages)
+            )
+
+    return repairs
+
+
 
 def strip_think_blocks(agent, content: str) -> str:
     """Remove reasoning/thinking blocks from content, returning only visible text.
@@ -579,12 +618,33 @@ def recover_with_credential_pool(
     current_provider = (getattr(agent, "provider", "") or "").strip().lower()
     pool_provider = (getattr(pool, "provider", "") or "").strip().lower()
     if current_provider and pool_provider and current_provider != pool_provider:
-        _ra().logger.warning(
-            "Credential pool provider mismatch: pool=%s, agent=%s — "
-            "skipping pool mutation to avoid cross-provider contamination",
-            pool_provider, current_provider,
-        )
-        return False, has_retried_429
+        # Custom endpoints use two naming conventions for the SAME provider:
+        # the agent carries the generic ``custom`` label while the pool is
+        # keyed ``custom:<name>`` (see CUSTOM_POOL_PREFIX). A literal string
+        # compare treats them as a mismatch and skips recovery for every
+        # custom-provider user — 401s/429s then burn the full retry cycle
+        # with no rotation or refresh. Accept the pair as matching only when
+        # the agent's CURRENT base_url actually resolves to this pool key,
+        # so a fallback provider (or a different custom endpoint) still
+        # triggers the guard.
+        _custom_match = False
+        if current_provider == "custom" and pool_provider.startswith("custom:"):
+            try:
+                from agent.credential_pool import get_custom_provider_pool_key
+                _agent_base = (getattr(agent, "base_url", "") or "").strip()
+                _custom_match = bool(_agent_base) and (
+                    (get_custom_provider_pool_key(_agent_base) or "").strip().lower()
+                    == pool_provider
+                )
+            except Exception:
+                _custom_match = False
+        if not _custom_match:
+            _ra().logger.warning(
+                "Credential pool provider mismatch: pool=%s, agent=%s — "
+                "skipping pool mutation to avoid cross-provider contamination",
+                pool_provider, current_provider,
+            )
+            return False, has_retried_429
 
     effective_reason = classified_reason
     if effective_reason is None:
@@ -679,15 +739,28 @@ def recover_with_credential_pool(
         # long-running TUI sessions stuck on stale tokens until the user
         # exited and reopened.
         is_entitlement = agent._is_entitlement_failure(error_context, status_code)
+        _auth_haystack = " ".join(
+            str(error_context.get(k) or "").lower()
+            for k in ("message", "reason", "code", "error")
+            if isinstance(error_context, dict)
+        )
+        if (
+            not is_entitlement
+            and status_code == 403
+            and "oauth authentication is currently not allowed for this organization" in _auth_haystack
+        ):
+            is_entitlement = True
+        if (
+            not is_entitlement
+            and status_code == 403
+            and (agent.provider or "") == "anthropic"
+            and getattr(agent, "api_mode", "") == "anthropic_messages"
+        ):
+            is_entitlement = True
         if not is_entitlement and status_code == 403 and (agent.provider or "") == "xai-oauth":
-            _disambiguator_haystack = " ".join(
-                str(error_context.get(k) or "").lower()
-                for k in ("message", "reason", "code", "error")
-                if isinstance(error_context, dict)
-            )
             _is_xai_auth_failure = (
-                "[wke=unauthenticated:" in _disambiguator_haystack
-                or "oauth2 access token could not be validated" in _disambiguator_haystack
+                "[wke=unauthenticated:" in _auth_haystack
+                or "oauth2 access token could not be validated" in _auth_haystack
             )
             if not _is_xai_auth_failure:
                 is_entitlement = True
@@ -808,6 +881,8 @@ def try_recover_primary_transport(
 
 def drop_thinking_only_and_merge_users(
     messages: List[Dict[str, Any]],
+    *,
+    drop_codex_reasoning_items: bool = True,
 ) -> List[Dict[str, Any]]:
     """Drop thinking-only assistant turns; merge any adjacent user messages left behind.
 
@@ -829,7 +904,13 @@ def drop_thinking_only_and_merge_users(
         return messages
 
     # Pass 1: drop thinking-only assistant turns.
-    kept = [m for m in messages if not _ra().AIAgent._is_thinking_only_assistant(m)]
+    kept = [
+        m for m in messages
+        if not _ra().AIAgent._is_thinking_only_assistant(
+            m,
+            drop_codex_reasoning_items=drop_codex_reasoning_items,
+        )
+    ]
     dropped = len(messages) - len(kept)
     if dropped == 0:
         return messages
@@ -1136,12 +1217,23 @@ def dump_api_request_debug(
 
         timestamp = datetime.now().strftime("%Y%m%d_%H%M%S_%f")
         dump_file = agent.logs_dir / f"request_dump_{agent.session_id}_{timestamp}.json"
-        atomic_json_write(dump_file, dump_payload, default=str)
+
+        # Redact secrets before persisting/printing. This dump captures the
+        # full request body (system prompt, tool defs, context-embedded
+        # values), and this path fires unconditionally on API errors — so it
+        # otherwise lands any context-embedded secret in cleartext on disk.
+        # Run the serialized dump through the same scrubber used for logs/tool
+        # output, then hand the resulting payload back to the shared atomic
+        # JSON writer so request dumps keep the same write semantics as before.
+        from agent.redact import redact_sensitive_text
+        _serialized = json.dumps(dump_payload, ensure_ascii=False, indent=2, default=str)
+        _redacted_payload = json.loads(redact_sensitive_text(_serialized, force=True))
+        atomic_json_write(dump_file, _redacted_payload, default=str)
 
         agent._vprint(f"{agent.log_prefix}🧾 Request debug dump written to: {dump_file}")
 
         if env_var_enabled("HERMES_DUMP_REQUEST_STDOUT"):
-            print(json.dumps(dump_payload, ensure_ascii=False, indent=2, default=str))
+            print(json.dumps(_redacted_payload, ensure_ascii=False, indent=2, default=str))
 
         return dump_file
     except Exception as dump_error:
@@ -1784,6 +1876,17 @@ def invoke_tool(agent, function_name: str, function_args: dict, effective_task_i
                 ),
                 next_args,
             )
+    elif function_name == "read_terminal":
+        def _execute(next_args: dict) -> Any:
+            from tools.read_terminal_tool import read_terminal_tool as _read_terminal_tool
+            return _finish_agent_tool(
+                _read_terminal_tool(
+                    start_line=next_args.get("start_line"),
+                    count=next_args.get("count"),
+                    callback=getattr(agent, "read_terminal_callback", None),
+                ),
+                next_args,
+            )
     elif function_name == "delegate_task":
         def _execute(next_args: dict) -> Any:
             return _finish_agent_tool(agent._dispatch_delegate_task(next_args), next_args)

agent/anthropic_adapter.py

@@ -73,20 +73,50 @@ ADAPTIVE_EFFORT_MAP = {
     "minimal": "low",
 }
 
-# Models that accept the "xhigh" output_config.effort level.  Opus 4.7 added
-# xhigh as a distinct level between high and max; older adaptive-thinking
-# models (4.6) reject it with a 400.  Keep this substring list in sync with
-# the Anthropic migration guide as new model families ship.
-_XHIGH_EFFORT_SUBSTRINGS = ("4-7", "4.7", "4-8", "4.8")
+# ── Anthropic thinking-mode classification ────────────────────────────
+# Claude 4.6 replaced budget-based extended thinking with *adaptive* thinking,
+# and 4.7 additionally forbids the manual ``thinking`` block entirely and drops
+# temperature/top_p/top_k.  Newer Claude releases (4.8, and named models like
+# claude-fable-5) follow the same modern contract — but they share no common
+# version substring, so an allowlist of version numbers ("4.6", "4.7", …) goes
+# stale the moment a model ships without a recognized number and silently
+# routes it down the legacy manual-thinking path.
+#
+# Instead we DEFAULT unknown Claude models to the modern contract and keep an
+# explicit *legacy* list of the older Claude families that still require manual
+# thinking.  This mirrors _get_anthropic_max_output's "default to newest" design
+# (future models are unlikely to regress to the older contract), so each new
+# Claude release works without a code change.
+#
+# Non-Claude Anthropic-Messages models (minimax, qwen3, GLM, …) are NOT Claude,
+# so they fall through to the legacy path automatically — exactly what those
+# manual-thinking endpoints need.
+
+# Older Claude families that DON'T support adaptive thinking (manual thinking
+# with budget_tokens only). Substring-matched against the model name.
+_LEGACY_MANUAL_THINKING_CLAUDE_SUBSTRINGS = (
+    "claude-3",          # 3, 3.5, 3.7
+    "claude-opus-4-0", "claude-opus-4.0", "claude-opus-4-1", "claude-opus-4.1",
+    "claude-sonnet-4-0", "claude-sonnet-4.0",
+    "claude-opus-4-2025", "claude-sonnet-4-2025",  # date-stamped 4.0 IDs
+    "claude-opus-4-5", "claude-opus-4.5",
+    "claude-sonnet-4-5", "claude-sonnet-4.5",
+    "claude-haiku-4-5", "claude-haiku-4.5",
+)
+
+# Older Claude families that DON'T accept the "xhigh" effort level (4.6 only
+# supports low/medium/high/max). xhigh arrived with Opus 4.7. Adaptive models
+# not in this list (4.7, 4.8, fable, future) accept xhigh.
+_NO_XHIGH_CLAUDE_SUBSTRINGS = (
+    "claude-opus-4-6", "claude-opus-4.6",
+    "claude-sonnet-4-6", "claude-sonnet-4.6",
+)
+
+
+def _is_claude_model(model: str | None) -> bool:
+    return "claude" in (model or "").lower()
 
-# Models where extended thinking is deprecated/removed (4.6+ behavior: adaptive
-# is the only supported mode; 4.7 additionally forbids manual thinking entirely
-# and drops temperature/top_p/top_k).
-_ADAPTIVE_THINKING_SUBSTRINGS = ("4-6", "4.6", "4-7", "4.7", "4-8", "4.8")
 
-# Models where temperature/top_p/top_k return 400 if set to non-default values.
-# This is the Opus 4.7 contract; future 4.x+ models are expected to follow it.
-_NO_SAMPLING_PARAMS_SUBSTRINGS = ("4-7", "4.7", "4-8", "4.8")
 _FAST_MODE_SUPPORTED_SUBSTRINGS = ("opus-4-6", "opus-4.6")
 
 # ── Max output token limits per Anthropic model ───────────────────────
@@ -94,6 +124,8 @@ _FAST_MODE_SUPPORTED_SUBSTRINGS = ("opus-4-6", "opus-4.6")
 # max_tokens as a mandatory field.  Previously we hardcoded 16384, which
 # starves thinking-enabled models (thinking tokens count toward the limit).
 _ANTHROPIC_OUTPUT_LIMITS = {
+    # Mythos-class named models (claude-fable-5, …) — 1M context, reasoning
+    "claude-fable":      128_000,
     # Claude 4.8
     "claude-opus-4-8":   128_000,
     # Claude 4.7
@@ -208,8 +240,17 @@ def _resolve_anthropic_messages_max_tokens(
 
 
 def _supports_adaptive_thinking(model: str) -> bool:
-    """Return True for Claude 4.6+ models that support adaptive thinking."""
-    return any(v in model for v in _ADAPTIVE_THINKING_SUBSTRINGS)
+    """Return True for Claude models that use adaptive thinking (4.6+).
+
+    Defaults *unknown* Claude models to adaptive (the modern contract) and
+    only returns False for the explicit legacy list of older Claude families
+    that require manual budget-based thinking. Non-Claude Anthropic-Messages
+    models (minimax, qwen3, …) return False so they keep the manual path.
+    """
+    if not _is_claude_model(model):
+        return False
+    m = model.lower()
+    return not any(v in m for v in _LEGACY_MANUAL_THINKING_CLAUDE_SUBSTRINGS)
 
 
 def _supports_xhigh_effort(model: str) -> bool:
@@ -219,18 +260,33 @@ def _supports_xhigh_effort(model: str) -> bool:
     Pre-4.7 adaptive models (Opus/Sonnet 4.6) only accept low/medium/high/max
     and reject xhigh with an HTTP 400. Callers should downgrade xhigh→max
     when this returns False.
+
+    Defaults unknown adaptive Claude models to accepting xhigh (4.7+ contract);
+    only the 4.6 family and legacy manual-thinking models are excluded.
     """
-    return any(v in model for v in _XHIGH_EFFORT_SUBSTRINGS)
+    if not _supports_adaptive_thinking(model):
+        return False
+    m = model.lower()
+    return not any(v in m for v in _NO_XHIGH_CLAUDE_SUBSTRINGS)
 
 
 def _forbids_sampling_params(model: str) -> bool:
     """Return True for models that 400 on any non-default temperature/top_p/top_k.
 
-    Opus 4.7 explicitly rejects sampling parameters; later Claude releases are
-    expected to follow suit.  Callers should omit these fields entirely rather
-    than passing zero/default values (the API rejects anything non-null).
+    Opus 4.7 introduced this restriction; later Claude releases follow it.
+    Defaults unknown Claude models to forbidding sampling params (the modern
+    contract). The 4.6 family still accepts them, and the legacy manual-thinking
+    families (4.5 and older) accept them too, so both are excluded. Non-Claude
+    models are unaffected. Callers should omit these fields entirely rather than
+    passing zero/default values (the API rejects anything non-null).
     """
-    return any(v in model for v in _NO_SAMPLING_PARAMS_SUBSTRINGS)
+    if not _is_claude_model(model):
+        return False
+    m = model.lower()
+    # 4.6 family is adaptive but still accepts sampling params.
+    if any(v in m for v in _NO_XHIGH_CLAUDE_SUBSTRINGS):
+        return False
+    return not any(v in m for v in _LEGACY_MANUAL_THINKING_CLAUDE_SUBSTRINGS)
 
 
 def _supports_fast_mode(model: str) -> bool:
@@ -695,6 +751,9 @@ def build_anthropic_client(
     from httpx import Timeout
 
     normalized_base_url = _normalize_base_url_text(base_url)
+    if normalized_base_url:
+        import re as _re
+        normalized_base_url = _re.sub(r"/v1/?$", "", normalized_base_url.rstrip("/"))
     _read_timeout = timeout if (isinstance(timeout, (int, float)) and timeout > 0) else 900.0
     kwargs = {
         "timeout": Timeout(timeout=float(_read_timeout), connect=10.0),
@@ -821,6 +880,7 @@ def _read_claude_code_credentials_from_keychain() -> Optional[Dict[str, Any]]:
             capture_output=True,
             text=True,
             timeout=5,
+            stdin=subprocess.DEVNULL,
         )
     except (OSError, subprocess.TimeoutExpired):
         logger.debug("Keychain: security command not available or timed out")
@@ -1163,7 +1223,10 @@ def run_oauth_setup_token() -> Optional[str]:
             "Install it with: npm install -g @anthropic-ai/claude-code"
         )
 
-    # Run interactively — stdin/stdout/stderr inherited so user can interact
+    # Run interactively — stdin/stdout/stderr inherited so the user can
+    # complete the OAuth login prompt. Must keep inherited stdin; the TUI-EOF
+    # concern does not apply to an interactive login the user explicitly
+    # invokes.  noqa: subprocess-stdin
     try:
         subprocess.run([claude_path, "setup-token"])
     except (KeyboardInterrupt, EOFError):
@@ -1511,6 +1574,15 @@ def _convert_content_part_to_anthropic(part: Any) -> Optional[Dict[str, Any]]:
 
     if ptype == "input_text":
         block: Dict[str, Any] = {"type": "text", "text": part.get("text", "")}
+    elif ptype == "text":
+        # A stored Anthropic text block. Rebuild from whitelisted fields only —
+        # SDK response text blocks carry output-only siblings (parsed_output,
+        # citations=None) that the Messages INPUT schema rejects with HTTP 400
+        # "Extra inputs are not permitted". Do NOT dict(part) it verbatim.
+        block = {"type": "text", "text": part.get("text", "")}
+        cits = part.get("citations")
+        if isinstance(cits, list) and cits:
+            block["citations"] = cits
     elif ptype in {"image_url", "input_image"}:
         image_value = part.get("image_url", {})
         url = image_value.get("url", "") if isinstance(image_value, dict) else str(image_value or "")
@@ -1625,6 +1697,58 @@ def _content_parts_to_anthropic_blocks(parts: Any) -> List[Dict[str, Any]]:
     return out
 
 
+def _sanitize_replay_block(b: Dict[str, Any]) -> Optional[Dict[str, Any]]:
+    """Strip output-only fields from a stored Anthropic content block so it is
+    valid as REQUEST input on replay.
+
+    The SDK response objects carry output-only attributes that the Messages
+    *input* schema forbids ("Extra inputs are not permitted"): text blocks get
+    ``parsed_output``/``citations`` (when null), tool_use blocks get ``caller``,
+    etc. ``normalize_response`` captured blocks verbatim via ``_to_plain_data``,
+    so these leak back as input on the next turn → HTTP 400.
+
+    Whitelist per type (NOT a blacklist) so future SDK output-only fields can't
+    reintroduce the bug. Returns a clean block, or None to drop it.
+    """
+    if not isinstance(b, dict):
+        return None
+    btype = b.get("type")
+    if btype == "text":
+        out: Dict[str, Any] = {"type": "text", "text": b.get("text", "")}
+        # citations is input-valid ONLY when it's a non-empty list; the SDK
+        # emits citations=None on responses, which the input schema rejects.
+        cits = b.get("citations")
+        if isinstance(cits, list) and cits:
+            out["citations"] = cits
+        if isinstance(b.get("cache_control"), dict):
+            out["cache_control"] = b["cache_control"]
+        return out
+    if btype == "thinking":
+        out = {"type": "thinking", "thinking": b.get("thinking", "")}
+        if b.get("signature"):
+            out["signature"] = b["signature"]
+        return out
+    if btype == "redacted_thinking":
+        # Only valid with its data payload; drop if missing.
+        return {"type": "redacted_thinking", "data": b["data"]} if b.get("data") else None
+    if btype == "tool_use":
+        out = {
+            "type": "tool_use",
+            "id": _sanitize_tool_id(b.get("id", "")),
+            "name": b.get("name", ""),
+            "input": b.get("input", {}),
+        }
+        if isinstance(b.get("cache_control"), dict):
+            out["cache_control"] = b["cache_control"]
+        return out
+    if btype == "image":
+        src = b.get("source")
+        return {"type": "image", "source": src} if isinstance(src, dict) else None
+    # Unknown/unsupported block type on the input path — drop rather than risk
+    # another "Extra inputs are not permitted".
+    return None
+
+
 def _convert_assistant_message(m: Dict[str, Any]) -> Dict[str, Any]:
     """Convert an assistant message to Anthropic content blocks.
 
@@ -1632,6 +1756,55 @@ def _convert_assistant_message(m: Dict[str, Any]) -> Dict[str, Any]:
     reasoning_content injection for Kimi/DeepSeek endpoints.
     """
     content = m.get("content", "")
+    # Anthropic interleaved-thinking fast path: when this turn carries a
+    # verbatim, order-preserving block list (set by normalize_response only
+    # for turns that interleave SIGNED thinking with tool_use), replay it.
+    # Each block is run through _sanitize_replay_block to strip output-only
+    # SDK fields (parsed_output, caller, citations=None, …) that the Messages
+    # INPUT schema forbids — replaying them verbatim caused HTTP 400 "Extra
+    # inputs are not permitted" (text.parsed_output). Block ORDER is preserved
+    # (the reason this channel exists); only forbidden sibling fields are
+    # dropped, leaving thinking signatures and tool_use id/name/input intact.
+    ordered_blocks = m.get("anthropic_content_blocks")
+    if isinstance(ordered_blocks, list) and ordered_blocks:
+        # Re-source each tool_use input from the stored tool_calls map rather
+        # than the captured block. The ordered-blocks list captures tool_use
+        # input from the RAW API response (normalize_response), which is NOT
+        # credential-redacted; tool_calls[].function.arguments IS redacted at
+        # storage time (build_assistant_message, #19798). Replaying the raw
+        # block input would resurrect a secret the model inlined into a tool
+        # call (e.g. terminal(command="curl -H 'Authorization: Bearer sk-...'")
+        # onto the wire, even though the same value is redacted everywhere else
+        # in history. Keying by sanitized tool id preserves interleave order
+        # (the reason this channel exists) while swapping in the redacted
+        # input. Adapted from #36071 (replay-time tool-input re-sourcing).
+        redacted_input_by_id: Dict[str, Any] = {}
+        for tc in m.get("tool_calls", []) or []:
+            if not isinstance(tc, dict):
+                continue
+            fn = tc.get("function", {}) or {}
+            raw_args = fn.get("arguments", "{}")
+            try:
+                parsed_args = json.loads(raw_args) if isinstance(raw_args, str) else raw_args
+            except (json.JSONDecodeError, ValueError):
+                parsed_args = {}
+            redacted_input_by_id[_sanitize_tool_id(tc.get("id", ""))] = parsed_args
+        replayed: List[Dict[str, Any]] = []
+        for b in ordered_blocks:
+            clean = _sanitize_replay_block(b)
+            if clean is None:
+                continue
+            if clean.get("type") == "tool_use":
+                # Override raw (un-redacted) input with the redacted copy when
+                # we have one for this id; fall back to the sanitized block
+                # input only if the tool_call is missing (shape mismatch).
+                redacted = redacted_input_by_id.get(clean.get("id", ""))
+                if redacted is not None:
+                    clean["input"] = redacted
+            replayed.append(clean)
+        if replayed:
+            return {"role": "assistant", "content": replayed}
+
     blocks = _extract_preserved_thinking_blocks(m)
     if content:
         if isinstance(content, list):

agent/auxiliary_client.py

@@ -102,7 +102,7 @@ OpenAI = _OpenAIProxy()  # module-level name, resolves lazily on call/isinstance
 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
+from utils import base_url_host_matches, base_url_hostname, model_forces_max_completion_tokens, normalize_proxy_env_vars
 
 logger = logging.getLogger(__name__)
 
@@ -1144,7 +1144,8 @@ def _endpoint_speaks_anthropic_messages(base_url: str) -> bool:
     normalized = (base_url or "").strip().lower().rstrip("/")
     if not normalized:
         return False
-    if normalized.endswith("/anthropic"):
+    path = urlparse(normalized).path.rstrip("/")
+    if path.endswith("/anthropic") or path.endswith("/anthropic/v1"):
         return True
     hostname = base_url_hostname(normalized)
     if hostname == "api.anthropic.com":
@@ -3078,23 +3079,20 @@ def _try_configured_fallback_chain(
         if not fb_provider or fb_provider.lower() == skip:
             continue
         fb_model = str(entry.get("model", "")).strip() or None
-        fb_base_url = str(entry.get("base_url", "")).strip() or None
-        fb_api_key = str(entry.get("api_key", "")).strip() or None
 
         label = f"fallback_chain[{i}]({fb_provider})"
 
         try:
-            fb_client = _resolve_single_provider(
-                fb_provider, fb_model, fb_base_url, fb_api_key)
+            fb_client, resolved_model = _resolve_fallback_entry(entry)
         except Exception:
-            fb_client = None
+            fb_client, resolved_model = None, None
 
         if fb_client is not None:
             logger.info(
                 "Auxiliary %s: %s on %s — configured fallback to %s (%s)",
-                task, reason, failed_provider, label, fb_model or "default",
+                task, reason, failed_provider, label, resolved_model or fb_model or "default",
             )
-            return fb_client, fb_model, label
+            return fb_client, resolved_model or fb_model, label
         tried.append(label)
 
     if tried:
@@ -3105,6 +3103,103 @@ def _try_configured_fallback_chain(
     return None, None, ""
 
 
+def _fallback_entry_api_key(entry: Dict[str, Any]) -> Optional[str]:
+    """Resolve inline or env-backed API key from a fallback-chain entry."""
+    explicit = str(entry.get("api_key") or "").strip()
+    if explicit:
+        return explicit
+    key_env = str(entry.get("key_env") or entry.get("api_key_env") or "").strip()
+    if key_env:
+        return os.getenv(key_env, "").strip() or None
+    return None
+
+
+def _resolve_fallback_entry(entry: Dict[str, Any]) -> Tuple[Optional[Any], Optional[str]]:
+    """Resolve one fallback entry through the central provider router."""
+    provider = str(entry.get("provider") or "").strip()
+    model = str(entry.get("model") or "").strip() or None
+    if not provider or not model:
+        return None, None
+    base_url = str(entry.get("base_url") or "").strip() or None
+    api_key = _fallback_entry_api_key(entry)
+    api_mode = str(entry.get("api_mode") or entry.get("transport") or "").strip() or None
+    return resolve_provider_client(
+        provider,
+        model=model,
+        explicit_base_url=base_url,
+        explicit_api_key=api_key,
+        api_mode=api_mode,
+    )
+
+
+def _try_main_fallback_chain(
+    task: Optional[str],
+    failed_provider: str = "",
+    reason: str = "error",
+) -> Tuple[Optional[Any], Optional[str], str]:
+    """Try the top-level main-agent fallback chain for an auxiliary call.
+
+    ``provider: auto`` auxiliary tasks should respect the user's declared
+    main fallback policy before dropping into Hermes' built-in discovery
+    chain. The top-level chain is read through ``get_fallback_chain`` so
+    both modern ``fallback_providers`` and legacy ``fallback_model`` entries
+    participate in the same order as the main agent.
+    """
+    try:
+        from hermes_cli.config import load_config
+        from hermes_cli.fallback_config import get_fallback_chain
+
+        chain = get_fallback_chain(load_config())
+    except Exception as exc:
+        logger.debug("Auxiliary %s: could not load main fallback chain: %s", task or "call", exc)
+        return None, None, ""
+
+    if not chain:
+        return None, None, ""
+
+    failed_norm = (failed_provider or "").strip().lower()
+    main_norm = (_read_main_provider() or "").strip().lower()
+    skip = {p for p in (failed_norm, main_norm, "auto") if p}
+    tried: List[str] = []
+
+    for i, entry in enumerate(chain):
+        if not isinstance(entry, dict):
+            continue
+        fb_provider = str(entry.get("provider") or "").strip()
+        fb_model = str(entry.get("model") or "").strip()
+        if not fb_provider or not fb_model:
+            continue
+        fb_norm = fb_provider.lower()
+        label = f"fallback_providers[{i}]({fb_provider})"
+        if fb_norm in skip:
+            tried.append(f"{label} (skipped)")
+            continue
+        if _is_provider_unhealthy(fb_norm):
+            _log_skip_unhealthy(fb_norm, task)
+            tried.append(f"{label} (unhealthy)")
+            continue
+        try:
+            fb_client, resolved_model = _resolve_fallback_entry(entry)
+        except Exception as exc:
+            logger.debug("Auxiliary %s: main fallback %s failed to resolve: %s", task or "call", label, exc)
+            fb_client, resolved_model = None, None
+        if fb_client is not None:
+            logger.info(
+                "Auxiliary %s: %s on %s — main fallback chain to %s (%s)",
+                task or "call", reason, failed_provider or "auto", label,
+                resolved_model or fb_model,
+            )
+            return fb_client, resolved_model or fb_model, fb_provider
+        tried.append(label)
+
+    if tried:
+        logger.debug(
+            "Auxiliary %s: main fallback chain exhausted (tried: %s)",
+            task or "call", ", ".join(tried),
+        )
+    return None, None, ""
+
+
 def _resolve_single_provider(
     provider: str,
     model: Optional[str] = None,
@@ -3115,16 +3210,19 @@ def _resolve_single_provider(
 
     Uses the existing provider resolution infrastructure where possible.
     """
-    # Reuse resolve_provider_client which handles provider→client mapping
+    # Reuse resolve_provider_client which handles provider→client mapping.
     client, resolved_model = resolve_provider_client(
         provider=provider,
         model=model,
-        base_url=base_url,
-        api_key=api_key,
+        explicit_base_url=base_url,
+        explicit_api_key=api_key,
     )
     return client
 
-def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Optional[OpenAI], Optional[str]]:
+def _resolve_auto(
+    main_runtime: Optional[Dict[str, Any]] = None,
+    task: Optional[str] = None,
+) -> Tuple[Optional[OpenAI], Optional[str]]:
     """Full auto-detection chain.
 
     Priority:
@@ -3190,7 +3288,7 @@ def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Option
     if (main_provider and main_model
             and main_provider not in {"auto", ""}):
         resolved_provider = main_provider
-        explicit_base_url = None
+        explicit_base_url = runtime_base_url or None
         explicit_api_key = None
         if runtime_base_url and (main_provider == "custom" or main_provider.startswith("custom:")):
             resolved_provider = "custom"
@@ -3222,7 +3320,22 @@ def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Option
                             main_provider, resolved or main_model)
                 return client, resolved or main_model
 
-    # ── Step 2: aggregator / fallback chain ──────────────────────────────
+    # ── Step 2: user-configured fallback policy ─────────────────────────
+    # In auto mode, respect the task-specific fallback chain first, then the
+    # main agent's top-level fallback_providers/fallback_model chain. The
+    # hardcoded provider discovery chain below is only the convenience default
+    # for users who have not declared a fallback policy.
+    if task:
+        fb_client, fb_model, _fb_label = _try_configured_fallback_chain(
+            task, main_provider or "auto", reason="main provider unavailable")
+        if fb_client is not None:
+            return fb_client, fb_model
+    fb_client, fb_model, _fb_label = _try_main_fallback_chain(
+        task, main_provider or "auto", reason="main provider unavailable")
+    if fb_client is not None:
+        return fb_client, fb_model
+
+    # ── Step 3: aggregator / fallback chain ──────────────────────────────
     tried = []
     for label, try_fn in _get_provider_chain():
         if _is_provider_unhealthy(label):
@@ -3343,6 +3456,7 @@ def resolve_provider_client(
     api_mode: str = None,
     main_runtime: Optional[Dict[str, Any]] = None,
     is_vision: bool = False,
+    task: Optional[str] = None,
 ) -> Tuple[Optional[Any], Optional[str]]:
     """Central router: given a provider name and optional model, return a
     configured client with the correct auth, base URL, and API format.
@@ -3463,7 +3577,7 @@ def resolve_provider_client(
 
     # ── Auto: try all providers in priority order ────────────────────
     if provider == "auto":
-        client, resolved = _resolve_auto(main_runtime=main_runtime)
+        client, resolved = _resolve_auto(main_runtime=main_runtime, task=task)
         if client is None:
             return None, None
         # When auto-detection lands on a non-OpenRouter provider (e.g. a
@@ -4300,13 +4414,15 @@ def get_auxiliary_extra_body() -> dict:
     return _nous_extra_body() if auxiliary_is_nous else {}
 
 
-def auxiliary_max_tokens_param(value: int) -> dict:
+def auxiliary_max_tokens_param(value: int, *, model: Optional[str] = None) -> dict:
     """Return the correct max tokens kwarg for the auxiliary client's provider.
-    
+
     OpenRouter and local models use 'max_tokens'. Direct OpenAI with newer
-    models (gpt-4o, o-series, gpt-5+) requires 'max_completion_tokens'.
+    models (gpt-4o, gpt-4.1, gpt-5+, o-series) requires 'max_completion_tokens'.
     The Codex adapter translates max_tokens internally, so we use max_tokens
-    for it as well.
+    for it as well. Pass ``model`` so third-party OpenAI-compatible endpoints
+    fronting the newer families are also recognised — URL-only detection
+    misses the case where a custom base URL serves e.g. ``gpt-5.4``.
     """
     custom_base = _current_custom_base_url()
     or_key = os.getenv("OPENROUTER_API_KEY")
@@ -4316,6 +4432,9 @@ def auxiliary_max_tokens_param(value: int) -> dict:
             and _read_nous_auth() is None
             and base_url_hostname(custom_base) in {"api.openai.com", "api.githubcopilot.com"}):
         return {"max_completion_tokens": value}
+    # ...and for any caller serving a newer OpenAI-family model by name.
+    if model_forces_max_completion_tokens(model):
+        return {"max_completion_tokens": value}
     return {"max_tokens": value}
 
 
@@ -4351,11 +4470,16 @@ def _client_cache_key(
     api_mode: Optional[str] = None,
     main_runtime: Optional[Dict[str, Any]] = None,
     is_vision: bool = False,
+    task: Optional[str] = None,
 ) -> tuple:
     runtime = _normalize_main_runtime(main_runtime)
     runtime_key = tuple(runtime.get(field, "") for field in _MAIN_RUNTIME_FIELDS) if provider == "auto" else ()
+    # `auto` can now resolve through task-specific or main fallback policy,
+    # so the task participates in the cache key. Non-auto providers keep the
+    # old cache shape because the explicit provider/model tuple is sufficient.
+    task_key = (task or "") if provider == "auto" else ""
     pool_hint = _pool_cache_hint(provider, main_runtime=main_runtime)
-    return (provider, async_mode, base_url or "", api_key or "", api_mode or "", runtime_key, is_vision, pool_hint)
+    return (provider, async_mode, base_url or "", api_key or "", api_mode or "", runtime_key, is_vision, task_key, pool_hint)
 
 
 def _store_cached_client(cache_key: tuple, client: Any, default_model: Optional[str], *, bound_loop: Any = None) -> None:
@@ -4548,6 +4672,7 @@ def _get_cached_client(
     api_mode: str = None,
     main_runtime: Optional[Dict[str, Any]] = None,
     is_vision: bool = False,
+    task: Optional[str] = None,
 ) -> Tuple[Optional[Any], Optional[str]]:
     """Get or create a cached client for the given provider.
 
@@ -4585,6 +4710,7 @@ def _get_cached_client(
         api_mode=api_mode,
         main_runtime=main_runtime,
         is_vision=is_vision,
+        task=task,
     )
     with _client_cache_lock:
         if cache_key in _client_cache:
@@ -4629,6 +4755,7 @@ def _get_cached_client(
         api_mode=api_mode,
         main_runtime=runtime,
         is_vision=is_vision,
+        task=task,
     )
     if client is not None:
         # For async clients, remember which loop they were created on so we
@@ -4999,7 +5126,7 @@ def _build_call_kwargs(
 
     # Provider-specific extra_body
     merged_extra = dict(extra_body or {})
-    if provider == "nous" or auxiliary_is_nous:
+    if provider == "nous":
         merged_extra.setdefault("tags", []).extend(_nous_portal_tags())
     if merged_extra:
         kwargs["extra_body"] = merged_extra
@@ -5134,7 +5261,7 @@ def call_llm(
             if not resolved_base_url:
                 logger.info("Auxiliary %s: provider %s unavailable, trying auto-detection chain",
                             task or "call", resolved_provider)
-                client, final_model = _get_cached_client("auto", main_runtime=main_runtime)
+                client, final_model = _get_cached_client("auto", main_runtime=main_runtime, task=task)
         if client is None:
             raise RuntimeError(
                 f"No LLM provider configured for task={task} provider={resolved_provider}. "
@@ -5460,14 +5587,19 @@ def call_llm(
 
             # Fallback order (#26882, #26803):
             #   1. User-configured fallback_chain (per-task) if set
-            #   2. Main agent model (last-resort safety net)
-            # For auto users (no explicit aux provider), use the full
-            # auto-detection chain instead — its Step 1 IS the main agent
-            # model, so users on `auto` already get main-model fallback.
+            #   2. For auto: top-level main fallback_providers/fallback_model
+            #   3. For auto: built-in auxiliary discovery chain
+            #   4. For explicit aux providers: main agent model safety net
             fb_client, fb_model, fb_label = (None, None, "")
             if is_auto:
-                fb_client, fb_model, fb_label = _try_payment_fallback(
-                    resolved_provider, task, reason=reason)
+                fb_client, fb_model, fb_label = _try_configured_fallback_chain(
+                    task, resolved_provider or "auto", reason=reason)
+                if fb_client is None:
+                    fb_client, fb_model, fb_label = _try_main_fallback_chain(
+                        task, resolved_provider or "auto", reason=reason)
+                if fb_client is None:
+                    fb_client, fb_model, fb_label = _try_payment_fallback(
+                        resolved_provider, task, reason=reason)
             else:
                 fb_client, fb_model, fb_label = _try_configured_fallback_chain(
                     task, resolved_provider or "auto", reason=reason)
@@ -5630,7 +5762,7 @@ async def async_call_llm(
             if not resolved_base_url:
                 logger.info("Auxiliary %s: provider %s unavailable, trying auto-detection chain",
                             task or "call", resolved_provider)
-                client, final_model = _get_cached_client("auto", async_mode=True)
+                client, final_model = _get_cached_client("auto", async_mode=True, main_runtime=main_runtime, task=task)
         if client is None:
             raise RuntimeError(
                 f"No LLM provider configured for task={task} provider={resolved_provider}. "
@@ -5898,13 +6030,19 @@ async def async_call_llm(
 
             # Fallback order (#26882, #26803):
             #   1. User-configured fallback_chain (per-task) if set
-            #   2. Main agent model (last-resort safety net)
-            # Auto users get the full auto-detection chain instead — its
-            # Step 1 IS the main agent model.
+            #   2. For auto: top-level main fallback_providers/fallback_model
+            #   3. For auto: built-in auxiliary discovery chain
+            #   4. For explicit aux providers: main agent model safety net
             fb_client, fb_model, fb_label = (None, None, "")
             if is_auto:
-                fb_client, fb_model, fb_label = _try_payment_fallback(
-                    resolved_provider, task, reason=reason)
+                fb_client, fb_model, fb_label = _try_configured_fallback_chain(
+                    task, resolved_provider or "auto", reason=reason)
+                if fb_client is None:
+                    fb_client, fb_model, fb_label = _try_main_fallback_chain(
+                        task, resolved_provider or "auto", reason=reason)
+                if fb_client is None:
+                    fb_client, fb_model, fb_label = _try_payment_fallback(
+                        resolved_provider, task, reason=reason)
             else:
                 fb_client, fb_model, fb_label = _try_configured_fallback_chain(
                     task, resolved_provider or "auto", reason=reason)

agent/background_review.py

@@ -237,18 +237,25 @@ _COMBINED_REVIEW_PROMPT = (
 def summarize_background_review_actions(
     review_messages: List[Dict],
     prior_snapshot: List[Dict],
+    notification_mode: str = "on",
 ) -> List[str]:
     """Build the human-facing action summary for a background review pass.
 
-    Walks the review agent's session messages and collects "successful tool
-    action" descriptions to surface to the user (e.g. "Memory updated").
-    Tool messages already present in ``prior_snapshot`` are skipped so we
-    don't re-surface stale results from the prior conversation that the
-    review agent inherited via ``conversation_history`` (issue #14944).
+    Walks the review agent's session messages and collects successful memory
+    and skill-management actions to surface to the user. Tool messages already
+    present in ``prior_snapshot`` are skipped so stale inherited results are
+    not re-surfaced as fresh background work (issue #14944).
 
-    Matching is by ``tool_call_id`` when available, with a content-equality
-    fallback for tool messages that lack one.
+    ``notification_mode`` controls display detail:
+    - ``off``: return no actions.
+    - ``on``: generic "Memory updated"/tool messages.
+    - ``verbose``: include compact content previews from tool-call arguments.
     """
+    mode = str(notification_mode or "on").lower()
+    if mode == "off":
+        return []
+    verbose = mode == "verbose"
+
     existing_tool_call_ids = set()
     existing_tool_contents = set()
     for prior in prior_snapshot or []:
@@ -262,6 +269,42 @@ def summarize_background_review_actions(
             if isinstance(content, str):
                 existing_tool_contents.add(content)
 
+    # Map review-agent tool results back to the calls that produced them.  The
+    # result JSON only says "Entry added"; the call arguments contain action,
+    # target, and content previews.  Restricting to notify_tools also prevents
+    # helper tools from surfacing as memory work just because they succeeded.
+    notify_tools = {"memory", "skill_manage"}
+    all_tool_call_ids: set = set()
+    call_details: dict = {}
+    for msg in review_messages or []:
+        if not isinstance(msg, dict) or msg.get("role") != "assistant":
+            continue
+        for tc in msg.get("tool_calls", []) or []:
+            if not isinstance(tc, dict):
+                continue
+            fn = tc.get("function", {}) or {}
+            fn_name = fn.get("name", "")
+            tcid = tc.get("id")
+            if tcid:
+                all_tool_call_ids.add(tcid)
+            if fn_name not in notify_tools:
+                continue
+            try:
+                args = json.loads(fn.get("arguments", "{}"))
+            except (json.JSONDecodeError, TypeError):
+                args = {}
+            if tcid:
+                call_details[tcid] = {
+                    "tool": fn_name,
+                    "action": args.get("action", "?"),
+                    "target": args.get("target", "memory"),
+                    "content": args.get("content", ""),
+                    "old_text": args.get("old_text", ""),
+                    "name": args.get("name", ""),
+                    "old_string": args.get("old_string", ""),
+                    "new_string": args.get("new_string", ""),
+                }
+
     actions: List[str] = []
     for msg in review_messages or []:
         if not isinstance(msg, dict) or msg.get("role") != "tool":
@@ -273,6 +316,8 @@ def summarize_background_review_actions(
             content_str = msg.get("content")
             if isinstance(content_str, str) and content_str in existing_tool_contents:
                 continue
+        if tcid and all_tool_call_ids and tcid not in call_details:
+            continue
         try:
             data = json.loads(msg.get("content", "{}"))
         except (json.JSONDecodeError, TypeError):
@@ -280,19 +325,75 @@ def summarize_background_review_actions(
         if not isinstance(data, dict) or not data.get("success"):
             continue
         message = data.get("message", "")
-        target = data.get("target", "")
-        if "created" in message.lower():
-            actions.append(message)
-        elif "updated" in message.lower():
-            actions.append(message)
-        elif "added" in message.lower() or (target and "add" in message.lower()):
-            label = "Memory" if target == "memory" else "User profile" if target == "user" else target
-            actions.append(f"{label} updated")
-        elif "Entry added" in message:
-            label = "Memory" if target == "memory" else "User profile" if target == "user" else target
-            actions.append(f"{label} updated")
-        elif "removed" in message.lower() or "replaced" in message.lower():
+        detail = call_details.get(tcid, {})
+        target = data.get("target", "") or detail.get("target", "")
+        is_skill = detail.get("tool") == "skill_manage"
+
+        message_lower = message.lower()
+        if not verbose:
+            if "created" in message_lower:
+                actions.append(message)
+                continue
+            if "updated" in message_lower:
+                actions.append(message)
+                continue
+            if is_skill and "patched" in message_lower:
+                actions.append(message)
+                continue
+
+        if is_skill:
+            label = "Skill"
+        elif target:
             label = "Memory" if target == "memory" else "User profile" if target == "user" else target
+        else:
+            continue
+
+        if verbose:
+            action = detail.get("action", "")
+            content = detail.get("content", "")
+            old_text = detail.get("old_text", "")
+            skill_name = detail.get("name", "")
+            max_preview = 120
+            if is_skill:
+                change = data.get("_change", {})
+                old_string = change.get("old", "") or detail.get("old_string", "")
+                new_string = change.get("new", "") or detail.get("new_string", "")
+                description = change.get("description", "")
+                if action == "patch" and (old_string or new_string):
+                    old_preview = old_string[:80].replace("\n", " ") + (
+                        "…" if len(old_string) > 80 else ""
+                    )
+                    new_preview = new_string[:80].replace("\n", " ") + (
+                        "…" if len(new_string) > 80 else ""
+                    )
+                    actions.append(
+                        f"📝 Skill '{skill_name}' patched: "
+                        f"\"{old_preview}\" → \"{new_preview}\""
+                    )
+                elif action == "create" and description:
+                    actions.append(f"📝 Skill '{skill_name}' created: {description}")
+                elif action == "edit" and description:
+                    actions.append(f"📝 Skill '{skill_name}' rewritten: {description}")
+                else:
+                    actions.append(f"📝 {message}" if message else f"Skill {action}")
+            elif action == "add" and content:
+                preview = content[:max_preview] + ("…" if len(content) > max_preview else "")
+                actions.append(f"{label} ➕ {preview}")
+            elif action == "replace" and content:
+                preview = content[:max_preview] + ("…" if len(content) > max_preview else "")
+                actions.append(f"{label} ✏️ {preview}")
+            elif action == "remove" and old_text:
+                preview = old_text[:60] + ("…" if len(old_text) > 60 else "")
+                actions.append(f"{label} ➖ {preview}")
+            else:
+                actions.append(f"{label} updated")
+        elif (
+            "added" in message_lower
+            or "replaced" in message_lower
+            or "removed" in message_lower
+            or (target and "add" in message.lower())
+            or "Entry added" in message
+        ):
             actions.append(f"{label} updated")
     return actions
 
@@ -522,6 +623,7 @@ def _run_review_in_thread(
         actions = summarize_background_review_actions(
             review_messages,
             messages_snapshot,
+            notification_mode=getattr(agent, "memory_notifications", "on"),
         )
 
         if actions:

agent/bedrock_adapter.py

@@ -58,17 +58,34 @@ _bedrock_runtime_client_cache: Dict[str, Any] = {}
 _bedrock_control_client_cache: Dict[str, Any] = {}
 
 
+_MIN_BOTO3_VERSION = (1, 34, 59)
+
+
 def _require_boto3():
-    """Import boto3, raising a clear error if not installed."""
+    """Import boto3, raising a clear error if not installed or too old."""
     try:
         import boto3
-        return boto3
     except ImportError:
         raise ImportError(
             "The 'boto3' package is required for the AWS Bedrock provider. "
             "Install it with: pip install boto3\n"
             "Or install Hermes with Bedrock support: pip install -e '.[bedrock]'"
         )
+    # converse() / converse_stream() were added in boto3 1.34.59.
+    # When Hermes is installed editable into system Python, the system boto3
+    # (e.g. Ubuntu 24.04 ships 1.34.46) may take precedence over the venv
+    # version pinned in pyproject.toml.
+    try:
+        version = tuple(int(x) for x in boto3.__version__.split(".")[:3])
+    except (AttributeError, ValueError):
+        return boto3  # can't parse — don't block on version check
+    if version < _MIN_BOTO3_VERSION:
+        raise RuntimeError(
+            f"boto3 {boto3.__version__} does not support converse_stream "
+            f"(minimum 1.34.59 required). Upgrade with: "
+            f"pip install --upgrade boto3"
+        )
+    return boto3
 
 
 def _get_bedrock_runtime_client(region: str):
@@ -208,6 +225,41 @@ def is_stale_connection_error(exc: BaseException) -> bool:
     return False
 
 
+def is_streaming_access_denied_error(exc: BaseException) -> bool:
+    """Return True when AWS denied the ``bedrock:InvokeModelWithResponseStream`` action.
+
+    IAM policies scoped to ``bedrock:InvokeModel`` only (a common least-privilege
+    setup) reject ``converse_stream()`` with an ``AccessDeniedException`` whose
+    message names the streaming action, e.g.::
+
+        User: arn:aws:iam::123456789012:user/x is not authorized to perform:
+        bedrock:InvokeModelWithResponseStream on resource: ...
+
+    This is permanent for the session — retrying the stream can never succeed —
+    so callers should flip to the non-streaming ``converse()`` path (which maps
+    to ``bedrock:InvokeModel``) instead of burning retries.
+
+    Detection is deliberately message-based: boto3 surfaces this as a
+    ``ClientError`` with ``Error.Code == "AccessDeniedException"``, and the
+    AnthropicBedrock SDK wraps the same AWS response in its own exception
+    types, but both preserve the action name in the message.
+    """
+    msg = str(exc).lower()
+    if "invokemodelwithresponsestream" not in msg:
+        return False
+    # ClientError with an explicit access-denied code is the canonical form.
+    try:
+        from botocore.exceptions import ClientError
+    except ImportError:  # pragma: no cover — botocore always present with boto3
+        ClientError = None  # type: ignore[assignment]
+    if ClientError is not None and isinstance(exc, ClientError):
+        code = (getattr(exc, "response", None) or {}).get("Error", {}).get("Code", "")
+        return code in ("AccessDeniedException", "UnauthorizedException")
+    # Wrapped forms (e.g. AnthropicBedrock SDK PermissionDeniedError) — match
+    # on the authorization-failure phrasing AWS uses.
+    return "not authorized" in msg or "accessdenied" in msg
+
+
 # ---------------------------------------------------------------------------
 # AWS credential detection
 # ---------------------------------------------------------------------------
@@ -900,11 +952,14 @@ def build_converse_kwargs(
     if system_prompt:
         kwargs["system"] = system_prompt
 
-    if temperature is not None:
-        kwargs["inferenceConfig"]["temperature"] = temperature
+    from agent.anthropic_adapter import _forbids_sampling_params
 
-    if top_p is not None:
-        kwargs["inferenceConfig"]["topP"] = top_p
+    if not _forbids_sampling_params(model):
+        if temperature is not None:
+            kwargs["inferenceConfig"]["temperature"] = temperature
+
+        if top_p is not None:
+            kwargs["inferenceConfig"]["topP"] = top_p
 
     if stop_sequences:
         kwargs["inferenceConfig"]["stopSequences"] = stop_sequences
@@ -1003,6 +1058,16 @@ def call_converse_stream(
     try:
         response = client.converse_stream(**kwargs)
     except Exception as exc:
+        if is_streaming_access_denied_error(exc):
+            # IAM allows bedrock:InvokeModel but not
+            # InvokeModelWithResponseStream — permanent for this session.
+            # Fall back to the non-streaming converse() path.
+            logger.info(
+                "bedrock: converse_stream denied by IAM on (region=%s, model=%s) — "
+                "falling back to non-streaming converse().",
+                region, model,
+            )
+            return normalize_converse_response(client.converse(**kwargs))
         if is_stale_connection_error(exc):
             logger.warning(
                 "bedrock: stale-connection error on converse_stream(region=%s, "

agent/chat_completion_helpers.py

@@ -952,6 +952,18 @@ def build_assistant_message(agent, assistant_message, finish_reason: str) -> dic
         if preserved:
             msg["reasoning_details"] = preserved
 
+    # Anthropic interleaved-thinking replay: when a turn interleaves signed
+    # thinking blocks with tool_use, the parallel reasoning_details +
+    # tool_calls fields lose the cross-type ordering, and reconstruction
+    # front-loads thinking — reordering signed blocks and triggering HTTP 400
+    # ("thinking ... blocks in the latest assistant message cannot be
+    # modified"). Carry the verbatim ordered block list so the adapter can
+    # replay the latest assistant message unchanged. See
+    # agent/transports/anthropic.py and agent/anthropic_adapter.py.
+    ordered_blocks = getattr(assistant_message, "anthropic_content_blocks", None)
+    if ordered_blocks:
+        msg["anthropic_content_blocks"] = ordered_blocks
+
     # Codex Responses API: preserve encrypted reasoning items for
     # multi-turn continuity. These get replayed as input on the next turn.
     codex_items = getattr(assistant_message, "codex_reasoning_items", None)
@@ -1603,6 +1615,8 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
                     _get_bedrock_runtime_client,
                     invalidate_runtime_client,
                     is_stale_connection_error,
+                    is_streaming_access_denied_error,
+                    normalize_converse_response,
                     stream_converse_with_callbacks,
                 )
                 region = api_kwargs.pop("__bedrock_region__", "us-east-1")
@@ -1611,6 +1625,29 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
                 try:
                     raw_response = client.converse_stream(**api_kwargs)
                 except Exception as _bedrock_exc:
+                    # IAM policies scoped to bedrock:InvokeModel only (no
+                    # InvokeModelWithResponseStream) reject converse_stream()
+                    # with AccessDeniedException. That denial is permanent for
+                    # the session — fall back to the non-streaming converse()
+                    # inline (it maps to bedrock:InvokeModel) and disable
+                    # streaming for subsequent calls so we don't re-fail every
+                    # turn.
+                    if is_streaming_access_denied_error(_bedrock_exc):
+                        agent._disable_streaming = True
+                        agent._safe_print(
+                            "\n⚠  AWS IAM denied bedrock:InvokeModelWithResponseStream — "
+                            "falling back to non-streaming InvokeModel.\n"
+                            "   Grant that action to restore streaming output.\n"
+                        )
+                        logger.info(
+                            "bedrock: converse_stream denied by IAM (%s) — "
+                            "using non-streaming converse() for this session.",
+                            type(_bedrock_exc).__name__,
+                        )
+                        result["response"] = normalize_converse_response(
+                            client.converse(**api_kwargs)
+                        )
+                        return
                     # Evict the cached client on stale-connection failures
                     # so the outer retry loop builds a fresh client/pool.
                     if is_stale_connection_error(_bedrock_exc):
@@ -1698,6 +1735,14 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
     # poll loop uses this to detect stale connections that keep receiving
     # SSE keep-alive pings but no actual data.
     last_chunk_time = {"t": time.time()}
+    # Stale-stream patience, shared between the httpx socket read timeout
+    # (built in ``_call_chat_completions`` below) and the stale-stream detector
+    # (computed further down, before the worker thread starts).  Initialized
+    # here so the read-timeout builder can floor itself at the stale value and
+    # never fire before the detector.  ``None`` until the detector value is
+    # resolved, so the builder degrades to its plain default if it ever runs
+    # first.
+    _stream_stale_timeout = None
 
     def _fire_first_delta():
         if not first_delta_fired["done"] and on_first_delta:
@@ -1734,6 +1779,26 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
                     "Local provider detected (%s) — stream read timeout raised to %.0fs",
                     agent.base_url, _stream_read_timeout,
                 )
+            elif (
+                _stream_read_timeout == 120.0
+                and _stream_stale_timeout is not None
+                and _stream_stale_timeout != float("inf")
+                and _stream_stale_timeout > _stream_read_timeout
+            ):
+                # Cloud reasoning models (e.g. Opus) routinely pause mid-stream
+                # for minutes during extended thinking.  The stale-stream
+                # detector is deliberately scaled up to tolerate this (180–300s,
+                # see the stale-timeout block below), but the raw httpx socket
+                # read timeout defaulted to a flat 120s and fired *first* —
+                # tearing down a healthy reasoning stream before the stale
+                # detector (which owns retry + diagnostics) could act.  Keep the
+                # socket read timeout in step with the detector so it no longer
+                # preempts it.
+                _stream_read_timeout = _stream_stale_timeout
+                logger.debug(
+                    "Cloud reasoning stream — read timeout raised to %.0fs to "
+                    "match stale-stream detector", _stream_read_timeout,
+                )
         # Cap connect/pool at 60s even when provider timeout is higher.
         # connect/pool cover TCP handshake, not model inference.
         _conn_cap = min(_base_timeout, 60.0) if _provider_timeout_cfg is not None else 30.0
@@ -2384,9 +2449,34 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
                             "stream" in _err_lower
                             and "not supported" in _err_lower
                         )
-                        if _is_stream_unsupported:
+                        # AWS Bedrock (AnthropicBedrock SDK path): IAM policies
+                        # with bedrock:InvokeModel but not
+                        # InvokeModelWithResponseStream reject messages.stream()
+                        # with a permission error naming the streaming action.
+                        # Permanent for the session — flip to non-streaming
+                        # (messages.create() maps to bedrock:InvokeModel).
+                        _is_bedrock_stream_denied = False
+                        if (
+                            not _is_stream_unsupported
+                            and "invokemodelwithresponsestream" in _err_lower
+                        ):
+                            # Cheap message pre-check before importing the
+                            # adapter — bedrock_adapter triggers a lazy boto3
+                            # install at import time, which must not run for
+                            # unrelated providers' stream errors.
+                            from agent.bedrock_adapter import (
+                                is_streaming_access_denied_error,
+                            )
+                            _is_bedrock_stream_denied = (
+                                is_streaming_access_denied_error(e)
+                            )
+                        if _is_stream_unsupported or _is_bedrock_stream_denied:
                             agent._disable_streaming = True
                             agent._safe_print(
+                                "\n⚠  AWS IAM denied bedrock:InvokeModelWithResponseStream. "
+                                "Switching to non-streaming.\n"
+                                "   Grant that action to restore streaming output.\n"
+                                if _is_bedrock_stream_denied else
                                 "\n⚠  Streaming is not supported for this "
                                 "model/provider. Switching to non-streaming.\n"
                                 "   To avoid this delay, set display.streaming: false "

agent/codex_responses_adapter.py

@@ -127,14 +127,21 @@ def _chat_content_to_responses_parts(content: Any, *, role: str = "user") -> Lis
     return converted
 
 
-def _summarize_user_message_for_log(content: Any) -> str:
-    """Return a short text summary of a user message for logging/trajectory.
+def _summarize_user_message_for_log(content: Any, *, sep: str = " ") -> str:
+    """Flatten message content to a plain-text summary.
 
     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.
+    parts from the API server.  Several consumers want a plain string:
+
+    - Logging, spinner previews, and trajectory files (the default ``sep=" "``).
+    - External memory providers, which feed the text to regexes
+      (``sanitize_context``) and text APIs — a raw list crashes the sync with
+      ``expected string or bytes-like object, got 'list'`` (use ``sep="\\n"``).
+
+    Text parts are joined with ``sep``; images become a ``[N image(s)]`` marker
+    so the turn isn't recorded as if the attachment never existed.  Returns an
+    empty string for empty lists and ``str(content)`` for unexpected scalar
+    types.
     """
     if content is None:
         return ""
@@ -157,7 +164,7 @@ def _summarize_user_message_for_log(content: Any) -> str:
                     text_bits.append(text)
             elif ptype in {"image_url", "input_image"}:
                 image_count += 1
-        summary = " ".join(text_bits).strip()
+        summary = sep.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
@@ -1074,6 +1081,7 @@ def _normalize_codex_response(
     message_items_raw: List[Dict[str, Any]] = []
     tool_calls: List[Any] = []
     has_incomplete_items = response_status in {"queued", "in_progress", "incomplete"}
+    saw_streaming_or_item_incomplete = response_status in {"queued", "in_progress"}
     saw_commentary_phase = False
     saw_final_answer_phase = False
     saw_reasoning_item = False
@@ -1088,6 +1096,7 @@ def _normalize_codex_response(
 
         if item_status in {"queued", "in_progress", "incomplete"}:
             has_incomplete_items = True
+            saw_streaming_or_item_incomplete = True
 
         if item_type == "message":
             item_phase = getattr(item, "phase", None)
@@ -1245,7 +1254,9 @@ def _normalize_codex_response(
         finish_reason = "tool_calls"
     elif leaked_tool_call_text:
         finish_reason = "incomplete"
-    elif has_incomplete_items or (saw_commentary_phase and not saw_final_answer_phase):
+    elif saw_streaming_or_item_incomplete:
+        finish_reason = "incomplete"
+    elif (has_incomplete_items or saw_commentary_phase) and not saw_final_answer_phase:
         finish_reason = "incomplete"
     elif (reasoning_items_raw or reasoning_parts or saw_reasoning_item) and not final_text:
         # Response contains only reasoning (encrypted thinking state and/or

agent/codex_runtime.py

@@ -25,6 +25,154 @@ from typing import Any, Dict, List
 logger = logging.getLogger(__name__)
 
 
+def _coerce_usage_int(value: Any) -> int:
+    if isinstance(value, bool):
+        return 0
+    if isinstance(value, int):
+        return max(value, 0)
+    if isinstance(value, float):
+        return max(int(value), 0)
+    if isinstance(value, str):
+        try:
+            return max(int(value), 0)
+        except ValueError:
+            return 0
+    return 0
+
+
+def _record_codex_app_server_usage(agent, turn) -> dict[str, Any]:
+    """Translate Codex app-server token usage into Hermes accounting.
+
+    Codex app-server reports usage via thread/tokenUsage/updated as:
+    inputTokens, cachedInputTokens, outputTokens, reasoningOutputTokens,
+    totalTokens.
+
+    Hermes' canonical prompt bucket includes uncached input + cached input.
+    The Codex app-server protocol does not currently expose cache-write tokens,
+    so that bucket remains zero on this runtime.
+
+    Even when Codex omits usage for a turn, Hermes should still count that turn
+    as one API call for session/status accounting.
+    """
+    agent.session_api_calls += 1
+
+    usage = getattr(turn, "token_usage_last", None)
+    if not isinstance(usage, dict) or not usage:
+        if agent._session_db and agent.session_id:
+            try:
+                if not agent._session_db_created:
+                    agent._ensure_db_session()
+                agent._session_db.update_token_counts(
+                    agent.session_id,
+                    model=agent.model,
+                    api_call_count=1,
+                )
+            except Exception as exc:
+                logger.debug(
+                    "Codex app-server api-call persistence failed (session=%s): %s",
+                    agent.session_id, exc,
+                )
+        return {}
+
+    from agent.usage_pricing import CanonicalUsage, estimate_usage_cost
+
+    input_tokens = _coerce_usage_int(usage.get("inputTokens"))
+    cache_read_tokens = _coerce_usage_int(usage.get("cachedInputTokens"))
+    output_tokens = _coerce_usage_int(usage.get("outputTokens"))
+    reasoning_tokens = _coerce_usage_int(usage.get("reasoningOutputTokens"))
+    reported_total = _coerce_usage_int(usage.get("totalTokens"))
+
+    canonical_usage = CanonicalUsage(
+        input_tokens=input_tokens,
+        output_tokens=output_tokens,
+        cache_read_tokens=cache_read_tokens,
+        cache_write_tokens=0,
+        reasoning_tokens=reasoning_tokens,
+        raw_usage=usage,
+    )
+    prompt_tokens = canonical_usage.prompt_tokens
+    completion_tokens = canonical_usage.output_tokens
+    total_tokens = reported_total or canonical_usage.total_tokens
+    usage_dict = {
+        "prompt_tokens": prompt_tokens,
+        "completion_tokens": completion_tokens,
+        "total_tokens": total_tokens,
+        "input_tokens": canonical_usage.input_tokens,
+        "output_tokens": canonical_usage.output_tokens,
+        "cache_read_tokens": canonical_usage.cache_read_tokens,
+        "cache_write_tokens": canonical_usage.cache_write_tokens,
+        "reasoning_tokens": canonical_usage.reasoning_tokens,
+    }
+
+    compressor = getattr(agent, "context_compressor", None)
+    if compressor is not None:
+        try:
+            compressor.update_from_response(usage_dict)
+            context_window = getattr(turn, "model_context_window", None)
+            if isinstance(context_window, int) and context_window > 0:
+                compressor.context_length = context_window
+        except Exception:
+            logger.debug("codex app-server usage update failed", exc_info=True)
+
+    agent.session_prompt_tokens += prompt_tokens
+    agent.session_completion_tokens += completion_tokens
+    agent.session_total_tokens += total_tokens
+    agent.session_input_tokens += canonical_usage.input_tokens
+    agent.session_output_tokens += canonical_usage.output_tokens
+    agent.session_cache_read_tokens += canonical_usage.cache_read_tokens
+    agent.session_cache_write_tokens += canonical_usage.cache_write_tokens
+    agent.session_reasoning_tokens += canonical_usage.reasoning_tokens
+
+    cost_result = estimate_usage_cost(
+        agent.model,
+        canonical_usage,
+        provider=agent.provider,
+        base_url=agent.base_url,
+        api_key=getattr(agent, "api_key", ""),
+    )
+    if cost_result.amount_usd is not None:
+        agent.session_estimated_cost_usd += float(cost_result.amount_usd)
+    agent.session_cost_status = cost_result.status
+    agent.session_cost_source = cost_result.source
+
+    if agent._session_db and agent.session_id:
+        try:
+            if not agent._session_db_created:
+                agent._ensure_db_session()
+            agent._session_db.update_token_counts(
+                agent.session_id,
+                input_tokens=canonical_usage.input_tokens,
+                output_tokens=canonical_usage.output_tokens,
+                cache_read_tokens=canonical_usage.cache_read_tokens,
+                cache_write_tokens=canonical_usage.cache_write_tokens,
+                reasoning_tokens=canonical_usage.reasoning_tokens,
+                estimated_cost_usd=float(cost_result.amount_usd)
+                if cost_result.amount_usd is not None else None,
+                cost_status=cost_result.status,
+                cost_source=cost_result.source,
+                billing_provider=agent.provider,
+                billing_base_url=agent.base_url,
+                billing_mode="subscription_included"
+                if cost_result.status == "included" else None,
+                model=agent.model,
+                api_call_count=1,
+            )
+        except Exception as exc:
+            logger.debug(
+                "Codex app-server token persistence failed (session=%s, tokens=%d): %s",
+                agent.session_id, total_tokens, exc,
+            )
+
+    return {
+        **usage_dict,
+        "last_prompt_tokens": prompt_tokens,
+        "estimated_cost_usd": float(cost_result.amount_usd)
+        if cost_result.amount_usd is not None else None,
+        "cost_status": cost_result.status,
+        "cost_source": cost_result.source,
+    }
+
+
 def run_codex_app_server_turn(
     agent,
     *,
@@ -120,6 +268,8 @@ def run_codex_app_server_turn(
     agent._iters_since_skill = (
         getattr(agent, "_iters_since_skill", 0) + turn.tool_iterations
     )
+    usage_result = _record_codex_app_server_usage(agent, turn)
+    api_calls = 1
 
     # Now check the skill nudge AFTER iters were incremented — same
     # pattern the chat_completions path uses (line ~15432).
@@ -164,12 +314,13 @@ def run_codex_app_server_turn(
     return {
         "final_response": turn.final_text,
         "messages": messages,
-        "api_calls": 1,  # one app-server "turn" maps to one logical API call
+        "api_calls": api_calls,
         "completed": not turn.interrupted and turn.error is None,
         "partial": turn.interrupted or turn.error is not None,
         "error": turn.error,
         "codex_thread_id": turn.thread_id,
         "codex_turn_id": turn.turn_id,
+        **usage_result,
     }
 
 

agent/coding_context.py

@@ -0,0 +1,738 @@
+"""Coding-context awareness — base Hermes, every interactive surface.
+
+When the user runs Hermes inside a code workspace (CLI, TUI, desktop app, or an
+editor over ACP), Hermes shifts into a **coding posture**. This module is the
+single place that decides whether we're in that posture and what it implies,
+so the rest of the codebase never re-derives "are we coding?" on its own.
+
+Architecture — one seam, many consumers
+----------------------------------------
+The posture is modelled as a frozen :class:`RuntimeMode` selected from a small
+:class:`ContextProfile` registry (today: ``coding`` and ``general``). A profile
+is *data* — it declares the toolset to collapse to, the operating brief to
+inject, and hints for other domains (model routing, memory, subagents). Every
+domain reads the same resolved object instead of probing git/config itself:
+
+  * **System prompt** — ``RuntimeMode.system_blocks()`` → the operating brief +
+    a live git/workspace snapshot (``agent/system_prompt.py``).
+  * **Toolset** — ``RuntimeMode.toolset_selection()`` → the ``coding`` toolset
+    plus the user's enabled MCP servers (``cli.py`` / ``tui_gateway``). Only
+    under the opt-in ``focus`` mode: the default posture is prompt-only and
+    never touches the user's configured toolsets (toolsets like messaging /
+    smart-home / music are off-by-default anyway, and someone who explicitly
+    enabled image-gen or Spotify shouldn't lose it for being in a git repo).
+  * **Delegation** — subagents inherit the parent's toolset and run through the
+    same prompt builder, so the coding posture propagates to children for free.
+  * **Model / memory / compression** — declared on the profile
+    (``model_hint``, ``memory_policy``) as the extension seam; consumers read
+    ``mode.profile`` rather than re-deciding.
+
+Cache safety
+------------
+The mode is resolved **once** and is immutable. The workspace snapshot is built
+once at prompt-build time and baked into the *stable* system-prompt tier — never
+re-probed per turn (that would shatter the prompt cache). Branch and dirty state
+drift mid-session, so the brief tells the model to re-check with ``git`` before
+acting on the snapshot. A ``/coding`` flip therefore only takes effect next
+session (deferred), the same contract as ``/skills install`` vs ``--now``.
+
+Activation (config ``agent.coding_context``):
+
+  * ``auto`` (default) — posture (brief + snapshot) on an interactive coding
+    surface sitting in a code workspace (git repo or recognised project root).
+    Prompt-only; toolsets and the skill index untouched.
+  * ``focus`` — like ``auto``, but additionally collapses the toolset to the
+    ``coding`` set + enabled MCP servers and demotes non-coding skill
+    categories to names-only in the prompt's skill index (no skill is ever
+    hidden). Explicit opt-in for a lean schema.
+  * ``on`` — force the posture anywhere (incl. non-workspaces). Prompt-only.
+  * ``off`` — disable entirely.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+import subprocess
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Optional
+
+logger = logging.getLogger("hermes.coding_context")
+
+CODING_TOOLSET = "coding"
+
+# Surfaces where a coding posture makes sense under ``auto``. Messaging
+# platforms (telegram, discord, slack, …) are intentionally absent — a chat bot
+# in a group is not pair-programming.
+INTERACTIVE_CODING_PLATFORMS = {"cli", "tui", "acp", "desktop", ""}
+
+# Project-root signals that mark a directory as a code workspace even when it
+# isn't (yet) a git repo. Cheap filename checks — no parsing.
+_PROJECT_MARKERS = (
+    "pyproject.toml", "setup.py", "setup.cfg", "requirements.txt",
+    "package.json", "tsconfig.json", "deno.json",
+    "Cargo.toml", "go.mod", "pom.xml", "build.gradle", "build.gradle.kts",
+    "Gemfile", "composer.json", "mix.exs", "pubspec.yaml",
+    "CMakeLists.txt", "Makefile", "Dockerfile",
+    "AGENTS.md", "CLAUDE.md", ".cursorrules",
+)
+
+# Agent-instruction files surfaced separately from manifests in the snapshot.
+_CONTEXT_FILES = ("AGENTS.md", "CLAUDE.md", ".cursorrules")
+
+# Lockfile → package manager, checked in priority order.
+_PY_LOCKFILES = (("uv.lock", "uv"), ("poetry.lock", "poetry"), ("Pipfile.lock", "pipenv"))
+_JS_LOCKFILES = (
+    ("pnpm-lock.yaml", "pnpm"), ("bun.lockb", "bun"), ("bun.lock", "bun"),
+    ("yarn.lock", "yarn"), ("package-lock.json", "npm"),
+)
+
+# package.json scripts / Makefile targets worth surfacing as verify commands.
+_VERIFY_TARGETS = ("test", "tests", "lint", "typecheck", "check", "build", "fmt", "format")
+_MAX_VERIFY_COMMANDS = 8
+_MAX_FACT_FILE_BYTES = 256 * 1024
+
+_GIT_TIMEOUT = 2.5
+
+
+# Per-model edit-format steering. Matching the edit tool format to how a model
+# was trained reduces mistakes and wasted reasoning (OpenAI/Codex handle
+# patch-style diffs best; Anthropic models — and most open-weight coding
+# models, whose RL scaffolds use str_replace-style editors — do best with
+# string-replacement). Our `patch` tool exposes both: mode="patch" (V4A
+# multi-file) and mode="replace" (find-and-swap). We nudge each family toward
+# its native format. Unknown families get nothing (the brief's neutral wording
+# stands). Substrings match the model id; aligned with TOOL_USE_ENFORCEMENT_MODELS.
+#
+# GPT/Codex get V4A for ALL edits, single-file included: in codex-rs,
+# apply_patch (V4A — apply_patch.lark) is the ONLY file editor, no
+# str_replace-style tool exists, and the shipped model prompts say to use
+# apply_patch even "for single file edits" — so a replace-mode nudge would
+# steer those models toward a format their first-party harness never taught
+# them.
+_EDIT_FORMAT_GUIDANCE: dict[str, tuple[tuple[str, ...], str]] = {
+    "patch": (
+        ("gpt", "codex"),
+        "- Edit format: author new files with `write_file`; for edits to "
+        "existing code use `patch` with `mode='patch'` (V4A diff) — including "
+        "single-file edits. It's the edit format you handle most reliably.",
+    ),
+    "replace": (
+        ("claude", "sonnet", "opus", "haiku",
+         "gemini", "gemma", "deepseek", "qwen", "kimi", "glm", "grok",
+         "hermes", "llama", "mistral", "devstral", "minimax"),
+        "- Edit format: author new files with `write_file`; for edits to "
+        "existing code prefer `patch` in `mode='replace'` — match a unique "
+        "snippet and swap it. Reach for `mode='patch'` (V4A) only when an edit "
+        "genuinely spans several files at once.",
+    ),
+}
+
+
+def _model_family(model: Optional[str]) -> Optional[str]:
+    """Classify a model id into an edit-format family key, or ``None``.
+
+    Used to steer the coding posture toward the edit tool format a model was
+    trained on. Family-agnostic by design: an unrecognised model gets ``None``
+    and the operating brief's neutral edit wording applies.
+    """
+    if not model:
+        return None
+    lowered = model.lower()
+    for family, (needles, _line) in _EDIT_FORMAT_GUIDANCE.items():
+        if any(n in lowered for n in needles):
+            return family
+    return None
+
+
+def _edit_format_line(model: Optional[str]) -> str:
+    """The edit-format guidance line for this model's family (``""`` if none)."""
+    family = _model_family(model)
+    if family is None:
+        return ""
+    return _EDIT_FORMAT_GUIDANCE[family][1]
+
+
+# Operating brief for the coding posture. Tool names referenced here (read_file,
+# search_files, patch, write_file, terminal, todo) are in the coding toolset and
+# in _HERMES_CORE_TOOLS, so they're present on every surface this fires on.
+CODING_AGENT_GUIDANCE = (
+    "You are a coding agent pairing with the user inside their codebase. "
+    "Operate like a careful senior engineer.\n"
+    "\n"
+    "Gather context first:\n"
+    "- Read the relevant files with `read_file` and locate code with "
+    "`search_files` before changing anything. Trace a symbol to its definition "
+    "and usages rather than guessing its shape.\n"
+    "- Batch independent lookups: when several reads/searches don't depend on "
+    "each other, issue them together in one turn instead of one at a time.\n"
+    "- Never invent files, symbols, APIs, or imports. If you haven't seen it in "
+    "the repo, go look. Don't assume a library is available — check the project "
+    "manifest (pyproject.toml / package.json / Cargo.toml / go.mod) and how "
+    "neighbouring files import it.\n"
+    "\n"
+    "Make changes through the tools, not the chat:\n"
+    "- Edit with `patch`/`write_file`. Do NOT print code blocks to the user as "
+    "a substitute for editing — apply the change, then summarise it. Only show "
+    "code when the user explicitly asks to see it.\n"
+    "- Match the project's existing style and conventions; AGENTS.md / "
+    "CLAUDE.md / .cursorrules already in context win over your defaults. Touch "
+    "only what the task needs — no drive-by refactors, renames, or reformatting "
+    "— and add any imports/dependencies your code requires.\n"
+    "- If an edit fails to apply, re-read the file to get the current exact "
+    "contents before retrying — don't repeat a stale patch. If the same region "
+    "fails twice, rewrite the enclosing function or file with `write_file` "
+    "instead of attempting a third patch.\n"
+    "\n"
+    "Verify, and know when to stop:\n"
+    "- Use `terminal` for git, builds, tests, and inspection. Run the relevant "
+    "tests/linter/build and confirm they pass before claiming the work is done.\n"
+    "- Terminal state persists across calls: current directory and exported "
+    "environment variables carry forward. Activate a virtualenv or export setup "
+    "vars once, then reuse that state instead of re-sourcing it before every "
+    "test command.\n"
+    "- Fix root causes, not symptoms: when you find a bug, check sibling call "
+    "paths for the same flaw and fix the class, not just the reported site.\n"
+    "- When fixing linter/type errors on a file, stop after about three "
+    "attempts on the same file and ask the user rather than looping.\n"
+    "- Track multi-step work with `todo`. Reference code as `path:line` instead "
+    "of pasting whole files.\n"
+    "\n"
+    "Respect the user's repo: don't commit, push, or rewrite history unless "
+    "asked, and never read, print, or commit secrets — leave `.env` and "
+    "credential files alone unless the user explicitly asks. The Workspace "
+    "block below is a snapshot from session start — re-run `git status`/"
+    "`git branch` before relying on it. Be concise: lead with the change or "
+    "answer, not a preamble."
+)
+
+
+# ── Context profiles (declarative posture definitions) ──────────────────────
+
+
+@dataclass(frozen=True)
+class ContextProfile:
+    """A named operating posture. Pure data — consumers read these fields.
+
+    ``toolset``      — collapse to this toolset (+ enabled MCP) when no explicit
+                       selection is pinned; ``None`` keeps the platform default.
+    ``guidance``     — operating brief injected into the stable system prompt;
+                       ``""`` injects nothing.
+    ``model_hint``   — routing preference key for smart model routing
+                       (extension seam; not yet consumed by the router).
+    ``memory_policy``— memory namespace/weighting hint (extension seam).
+    ``compact_skill_categories`` — skill categories DEMOTED to names-only in
+                       the system-prompt skill index under the opt-in ``focus``
+                       mode. Never hidden: every skill name stays visible
+                       (so memory-anchored recall keeps working) — only the
+                       descriptions are dropped to cut index noise. Deny-list
+                       semantics so unknown/custom categories keep full
+                       entries.
+    """
+
+    name: str
+    toolset: Optional[str] = None
+    guidance: str = ""
+    model_hint: Optional[str] = None
+    memory_policy: str = "default"
+    compact_skill_categories: tuple[str, ...] = ()
+
+
+# Skill categories that are clearly not part of a coding workflow. Demoted to
+# names-only in the prompt's skill index under the opt-in ``focus`` mode only
+# (deny-list — anything not listed here, incl. custom user categories, keeps
+# full entries). Coding-adjacent categories (devops, github, mcp,
+# data-science, diagramming, research, security, …) are intentionally absent.
+_NON_CODING_SKILL_CATEGORIES = (
+    "apple", "communication", "cooking", "creative", "email", "finance",
+    "gaming", "gifs", "health", "media", "music", "note-taking",
+    "productivity", "shopping", "smart-home", "social-media", "travel",
+    "yuanbao",
+)
+
+
+GENERAL_PROFILE = ContextProfile(name="general")
+CODING_PROFILE = ContextProfile(
+    name="coding",
+    toolset=CODING_TOOLSET,
+    guidance=CODING_AGENT_GUIDANCE,
+    model_hint="coding",
+    memory_policy="project",
+    compact_skill_categories=_NON_CODING_SKILL_CATEGORIES,
+)
+
+_PROFILES: dict[str, ContextProfile] = {
+    GENERAL_PROFILE.name: GENERAL_PROFILE,
+    CODING_PROFILE.name: CODING_PROFILE,
+}
+
+
+def get_profile(name: str) -> ContextProfile:
+    """Return a registered profile, falling back to ``general``."""
+    return _PROFILES.get(name, GENERAL_PROFILE)
+
+
+# ── Helpers ─────────────────────────────────────────────────────────────────
+
+
+def _coding_mode(config: Optional[dict[str, Any]]) -> str:
+    """Return the normalized ``agent.coding_context`` mode (auto/focus/on/off)."""
+    if config is None:
+        try:
+            from hermes_cli.config import load_config
+
+            config = load_config()
+        except Exception:
+            config = {}
+    raw = ((config or {}).get("agent", {}) or {}).get("coding_context", "auto")
+    mode = str(raw).strip().lower()
+    if mode in {"focus", "strict", "lean"}:
+        return "focus"
+    if mode in {"on", "true", "yes", "1", "always"}:
+        return "on"
+    if mode in {"off", "false", "no", "0", "never"}:
+        return "off"
+    return "auto"
+
+
+def _resolve_cwd(cwd: Optional[str | Path]) -> Path:
+    if cwd:
+        return Path(cwd).expanduser()
+    try:
+        from agent.runtime_cwd import resolve_agent_cwd
+
+        return resolve_agent_cwd()
+    except Exception:
+        return Path(os.getcwd())
+
+
+def _git_root(cwd: Path) -> Optional[Path]:
+    current = cwd.resolve()
+    for parent in [current, *current.parents]:
+        if (parent / ".git").exists():
+            return parent
+    return None
+
+
+def _home() -> Optional[Path]:
+    try:
+        return Path.home().resolve()
+    except (OSError, RuntimeError):
+        return None
+
+
+def _marker_root(cwd: Path) -> Optional[Path]:
+    """Nearest ancestor that looks like a project root, or ``None``.
+
+    Walks up at most a few levels so a manifest in the workspace root counts
+    even when the user is in a subdirectory. ``$HOME`` itself is skipped — a
+    Makefile or AGENTS.md sitting in the home directory is global user config,
+    not a project-root signal.
+    """
+    current = cwd.resolve()
+    home = _home()
+    for depth, parent in enumerate([current, *current.parents]):
+        if depth > 6:
+            break
+        if parent == home:
+            continue
+        for marker in _PROJECT_MARKERS:
+            if (parent / marker).exists():
+                return parent
+    return None
+
+
+def _detect_profile_name(mode: str, platform: str, cwd_str: str) -> str:
+    """Resolve which profile applies.
+
+    ``auto``/``focus``: coding when the surface is interactive AND the cwd is a
+    code workspace (a git repo or a recognised project root). ``on``: always
+    coding. ``off``: always general.
+
+    A git repo rooted at ``$HOME`` (the dotfiles pattern) is NOT a workspace
+    signal — without the guard, every session anywhere under a dotfiles-managed
+    home directory would silently flip to the coding posture.
+
+    Detection is intentionally not memoized: it's a handful of ``stat`` calls,
+    and callers resolve the mode once per session anyway. Caching here would
+    risk a stale posture if a long-lived process (gateway/TUI) serves sessions
+    from different working directories.
+    """
+    if mode == "off":
+        return GENERAL_PROFILE.name
+    if mode == "on":
+        return CODING_PROFILE.name
+    if platform and platform.strip().lower() not in INTERACTIVE_CODING_PLATFORMS:
+        return GENERAL_PROFILE.name
+    cwd = Path(cwd_str)
+    git_root = _git_root(cwd)
+    if git_root is not None and git_root == _home():
+        git_root = None  # dotfiles repo at $HOME — not a code workspace
+    if git_root is not None or _marker_root(cwd) is not None:
+        return CODING_PROFILE.name
+    return GENERAL_PROFILE.name
+
+
+# ── RuntimeMode (the seam) ──────────────────────────────────────────────────
+
+
+@dataclass(frozen=True)
+class RuntimeMode:
+    """The resolved operating posture for a session. Immutable by construction.
+
+    Built once via :func:`resolve_runtime_mode` and consumed by every domain
+    that cares about the coding/general distinction. Never mutate or re-resolve
+    mid-session — that would break the prompt cache.
+    """
+
+    profile: ContextProfile
+    surface: str
+    cwd: Path
+    # The normalized ``agent.coding_context`` mode this posture was resolved
+    # under (auto/focus/on/off). Toolset collapse is gated on ``focus``.
+    config_mode: str = "auto"
+    # The model id this session runs (e.g. "anthropic/claude-opus-4.8"). Used
+    # only to steer edit-format guidance toward the model's family — see
+    # ``_edit_format_line``. Fixed for the session, so cache-safe.
+    model: Optional[str] = None
+
+    @property
+    def kind(self) -> str:
+        return self.profile.name
+
+    @property
+    def is_coding(self) -> bool:
+        return self.profile.name == CODING_PROFILE.name
+
+    def toolset_selection(self, config: Optional[dict[str, Any]] = None) -> Optional[list[str]]:
+        """Toolset list for this posture, or ``None`` to keep the platform default.
+
+        Non-``None`` only under the opt-in ``focus`` mode. The default posture
+        is prompt-only: most strippable toolsets are off-by-default anyway, and
+        a user who explicitly enabled one (image-gen for frontend/game assets,
+        messaging for build notifications, …) keeps it while coding.
+
+        Callers apply this only when the user hasn't pinned an explicit
+        selection (``--toolsets``, ``HERMES_TUI_TOOLSETS``, …); they never
+        override a pin. Returns the profile's toolset plus enabled MCP servers.
+        """
+        if self.config_mode != "focus":
+            return None
+        if self.profile.toolset is None:
+            return None
+        return [self.profile.toolset, *_enabled_mcp_servers(config)]
+
+    def system_blocks(self) -> list[str]:
+        """Stable system-prompt blocks for this posture (brief + workspace).
+
+        The operating brief carries a model-family edit-format nudge appended
+        to it (one cached string, not a separate block) so the model is steered
+        toward the `patch` mode it handles best — see ``_edit_format_line``.
+        """
+        if not self.is_coding:
+            return []
+        blocks: list[str] = []
+        if self.profile.guidance:
+            brief = self.profile.guidance
+            edit_line = _edit_format_line(self.model)
+            if edit_line:
+                brief = f"{brief}\n{edit_line}"
+            blocks.append(brief)
+        workspace = build_coding_workspace_block(self.cwd)
+        if workspace:
+            blocks.append(workspace)
+        return blocks
+
+    def compact_skill_categories(self) -> frozenset[str]:
+        """Skill categories to demote to names-only in the prompt's skill index.
+
+        Gated on the opt-in ``focus`` mode, like the toolset collapse: the
+        default posture leaves the skill index untouched. Users who didn't ask
+        for a lean prompt keep full entries for every category — index changes
+        under ``auto`` proved too surprising in practice, even names-only ones
+        (a demoted description is information the model no longer weighs when
+        deciding what to load).
+
+        Demoted — never hidden — even under ``focus``. An earlier revision
+        fully pruned these categories from the index, which caused silent
+        capability loss in a real workflow: agent-created skills are the
+        model's accumulated project memory (server-ops runbooks, learned
+        pitfalls, …), and models do not reliably reach for ``skills_list`` to
+        rediscover what the index stopped showing them. Names-only keeps every
+        skill loadable on recall while still cutting the description noise.
+        """
+        if not self.is_coding or self.config_mode != "focus":
+            return frozenset()
+        return frozenset(self.profile.compact_skill_categories)
+
+
+def resolve_runtime_mode(
+    *,
+    platform: Optional[str] = None,
+    cwd: Optional[str | Path] = None,
+    config: Optional[dict[str, Any]] = None,
+    model: Optional[str] = None,
+) -> RuntimeMode:
+    """Resolve the operating posture once. Cheap — a handful of ``stat`` calls.
+
+    This is the single entry point every domain should call. The returned
+    object is immutable and safe to cache for the session. Detection itself is
+    intentionally *not* memoized (see ``_detect_profile_name``) so a long-lived
+    process can't pin a stale posture; callers resolve once per session and
+    hold the result. ``model`` is recorded only to steer edit-format guidance;
+    it never affects detection.
+    """
+    resolved_cwd = _resolve_cwd(cwd)
+    mode = _coding_mode(config)
+    name = _detect_profile_name(
+        mode, (platform or "").strip().lower(), str(resolved_cwd)
+    )
+    return RuntimeMode(
+        profile=get_profile(name),
+        surface=platform or "",
+        cwd=resolved_cwd,
+        config_mode=mode,
+        model=model,
+    )
+
+
+# ── Back-compat surface (thin wrappers over RuntimeMode) ────────────────────
+
+
+def is_coding_context(
+    *,
+    platform: Optional[str] = None,
+    cwd: Optional[str | Path] = None,
+    config: Optional[dict[str, Any]] = None,
+) -> bool:
+    """Whether Hermes should operate in its coding posture right now."""
+    return resolve_runtime_mode(platform=platform, cwd=cwd, config=config).is_coding
+
+
+def coding_selection(
+    *,
+    platform: Optional[str] = None,
+    cwd: Optional[str | Path] = None,
+    config: Optional[dict[str, Any]] = None,
+) -> Optional[list[str]]:
+    """Toolset selection for the coding posture.
+
+    ``None`` unless the user opted into ``focus`` mode AND the posture is
+    active — the default coding posture never overrides configured toolsets.
+    """
+    return resolve_runtime_mode(
+        platform=platform, cwd=cwd, config=config
+    ).toolset_selection(config)
+
+
+def coding_system_blocks(
+    *,
+    platform: Optional[str] = None,
+    cwd: Optional[str | Path] = None,
+    config: Optional[dict[str, Any]] = None,
+    model: Optional[str] = None,
+) -> list[str]:
+    """Stable system-prompt blocks for the current posture (empty when general).
+
+    ``model`` steers the brief's edit-format nudge toward the model's family.
+    """
+    return resolve_runtime_mode(
+        platform=platform, cwd=cwd, config=config, model=model
+    ).system_blocks()
+
+
+def coding_compact_skill_categories(
+    *,
+    platform: Optional[str] = None,
+    cwd: Optional[str | Path] = None,
+    config: Optional[dict[str, Any]] = None,
+) -> frozenset[str]:
+    """Skill categories the active posture demotes to names-only in the index.
+
+    Empty outside the coding posture and outside the opt-in ``focus`` mode —
+    the default posture never touches the skill index. Under ``focus``,
+    demoted — never hidden: every skill name stays in the index and remains
+    loadable via ``skill_view`` / ``skills_list``; only descriptions are
+    dropped.
+    """
+    return resolve_runtime_mode(
+        platform=platform, cwd=cwd, config=config
+    ).compact_skill_categories()
+
+
+def _enabled_mcp_servers(config: Optional[dict[str, Any]]) -> list[str]:
+    """Names of MCP servers the user has enabled — kept in the coding posture.
+
+    MCP servers (figma, browser, tophat, …) are explicitly configured and part
+    of the coding workflow, not noise to strip.
+    """
+    try:
+        from hermes_cli.config import read_raw_config
+        from hermes_cli.tools_config import _parse_enabled_flag
+
+        servers = read_raw_config().get("mcp_servers") or {}
+        return [
+            str(name)
+            for name, cfg in servers.items()
+            if isinstance(cfg, dict)
+            and _parse_enabled_flag(cfg.get("enabled", True), default=True)
+        ]
+    except Exception:
+        return []
+
+
+# ── git/workspace probe ─────────────────────────────────────────────────────
+
+
+def _git(cwd: Path, *args: str) -> str:
+    try:
+        out = subprocess.run(
+            ["git", "-C", str(cwd), *args],
+            capture_output=True,
+            text=True,
+            timeout=_GIT_TIMEOUT,
+        )
+    except (OSError, subprocess.SubprocessError):
+        return ""
+    return out.stdout.strip() if out.returncode == 0 else ""
+
+
+def _parse_status(porcelain: str) -> tuple[dict[str, str], dict[str, int]]:
+    """Parse ``git status --porcelain=2 --branch`` into branch + counts."""
+    branch: dict[str, str] = {}
+    counts = {"staged": 0, "modified": 0, "untracked": 0, "conflicts": 0}
+    for line in porcelain.splitlines():
+        if line.startswith("# branch.head"):
+            branch["head"] = line.split(maxsplit=2)[-1]
+        elif line.startswith("# branch.upstream"):
+            branch["upstream"] = line.split(maxsplit=2)[-1]
+        elif line.startswith("# branch.ab"):
+            parts = line.split()
+            branch["ahead"], branch["behind"] = parts[2].lstrip("+"), parts[3].lstrip("-")
+        elif line.startswith(("1 ", "2 ")):
+            xy = line.split(maxsplit=2)[1]
+            if xy[0] != ".":
+                counts["staged"] += 1
+            if xy[1] != ".":
+                counts["modified"] += 1
+        elif line.startswith("u "):
+            counts["conflicts"] += 1
+        elif line.startswith("? "):
+            counts["untracked"] += 1
+    return branch, counts
+
+
+def _read_small(path: Path) -> str:
+    """Read a small text file, or ``""`` — never raises, never reads huge files."""
+    try:
+        if not path.is_file() or path.stat().st_size > _MAX_FACT_FILE_BYTES:
+            return ""
+        return path.read_text(encoding="utf-8", errors="replace")
+    except OSError:
+        return ""
+
+
+def _project_facts(root: Path) -> list[str]:
+    """Detected project facts for the workspace snapshot.
+
+    The point is to hand the model its *verify loop* up front — which manifest,
+    which package manager, and the exact test/lint/build commands — instead of
+    making it rediscover them every session. Cheap: stat calls plus reads of a
+    couple of small files; built once at prompt-build time (cache-safe).
+    """
+    facts: list[str] = []
+
+    manifests = [m for m in _PROJECT_MARKERS if m not in _CONTEXT_FILES and (root / m).is_file()]
+    package_managers = [
+        pm for lock, pm in (*_PY_LOCKFILES, *_JS_LOCKFILES) if (root / lock).is_file()
+    ]
+    if manifests:
+        line = f"- Project: {', '.join(manifests[:6])}"
+        if package_managers:
+            line += f" ({'/'.join(dict.fromkeys(package_managers))})"
+        facts.append(line)
+
+    verify: list[str] = []
+    if (root / "scripts" / "run_tests.sh").is_file():
+        verify.append("scripts/run_tests.sh")
+    if (root / "package.json").is_file():
+        try:
+            scripts = json.loads(_read_small(root / "package.json") or "{}").get("scripts") or {}
+        except (json.JSONDecodeError, AttributeError):
+            scripts = {}
+        js_pm = next((pm for lock, pm in _JS_LOCKFILES if (root / lock).is_file()), "npm")
+        verify.extend(f"{js_pm} run {name}" for name in _VERIFY_TARGETS if name in scripts)
+    if (root / "pytest.ini").is_file() or "[tool.pytest" in _read_small(root / "pyproject.toml"):
+        verify.append("pytest")
+    makefile = _read_small(root / "Makefile")
+    if makefile:
+        verify.extend(
+            f"make {name}" for name in _VERIFY_TARGETS
+            if re.search(rf"^{re.escape(name)}\s*:", makefile, re.MULTILINE)
+        )
+    if verify:
+        deduped = list(dict.fromkeys(verify))[:_MAX_VERIFY_COMMANDS]
+        facts.append(f"- Verify: {'; '.join(deduped)}")
+
+    context_files = [c for c in _CONTEXT_FILES if (root / c).is_file()]
+    if context_files:
+        facts.append(f"- Context files: {', '.join(context_files)}")
+
+    return facts
+
+
+def build_coding_workspace_block(cwd: Optional[str | Path] = None) -> str:
+    """Workspace snapshot for the system prompt (empty outside a workspace).
+
+    Git state (branch/status/commits) when the cwd is in a repo, plus detected
+    project facts (manifest, package manager, verify commands, context files)
+    — so marker-only (non-git) projects still get a snapshot.
+    """
+    resolved = _resolve_cwd(cwd)
+    git_root = _git_root(resolved)
+    root = git_root or _marker_root(resolved)
+    if root is None:
+        return ""
+
+    lines = ["Workspace (snapshot at session start — re-check with `git` before acting on it):"]
+    lines.append(f"- Root: {root}")
+
+    if git_root is not None:
+        branch, counts = _parse_status(_git(root, "status", "--porcelain=2", "--branch"))
+        head = branch.get("head", "")
+        if head and head != "(detached)":
+            line = f"- Branch: {head}"
+            if branch.get("upstream"):
+                line += f" \u2192 {branch['upstream']}"
+                ahead, behind = branch.get("ahead", "0"), branch.get("behind", "0")
+                if ahead != "0" or behind != "0":
+                    line += f" (ahead {ahead}, behind {behind})"
+            lines.append(line)
+        elif head == "(detached)":
+            lines.append("- Branch: (detached HEAD)")
+
+        # Linked worktree: the per-worktree git dir differs from the shared common dir.
+        # We surface the fact that it's a worktree (so the model knows branches/stashes
+        # are shared state) but deliberately do NOT expose the primary tree path —
+        # giving the model a second absolute path causes it to sometimes run commands
+        # in the wrong directory.
+        git_dir, common_dir = _git(root, "rev-parse", "--git-dir"), _git(root, "rev-parse", "--git-common-dir")
+        if git_dir and common_dir and Path(git_dir).resolve() != Path(common_dir).resolve():
+            lines.append("- Worktree: linked (git state shared with primary tree)")
+
+        dirty = [f"{n} {label}" for label, n in (
+            ("staged", counts["staged"]), ("modified", counts["modified"]),
+            ("untracked", counts["untracked"]), ("conflicts", counts["conflicts"]),
+        ) if n]
+        lines.append(f"- Status: {', '.join(dirty) if dirty else 'clean'}")
+
+        recent = _git(root, "log", "-3", "--pretty=%h %s")
+        if recent:
+            lines.append("- Recent commits:")
+            lines.extend(f"    {c}" for c in recent.splitlines())
+
+    lines.extend(_project_facts(root))
+    return "\n".join(lines)

agent/context_compressor.py

@@ -7,7 +7,7 @@ protecting head and tail context.
 Improvements over v2:
   - Structured summary template with Resolved/Pending question tracking
   - Filter-safe summarizer preamble that treats prior turns as source material
-  - "Remaining Work" replaces "Next Steps" to avoid reading as active instructions
+  - Historical (reference-only) section headings replace "Next Steps"/"Remaining Work" to avoid reading as active instructions
   - Clear separator when summary merges into tail message
   - Iterative summary updates (preserves info across multiple compactions)
   - Token-budget tail protection instead of fixed message count
@@ -34,7 +34,75 @@ from agent.redact import redact_sensitive_text
 
 logger = logging.getLogger(__name__)
 
+HISTORICAL_TASK_HEADING = "## Historical Task Snapshot"
+HISTORICAL_IN_PROGRESS_HEADING = "## Historical In-Progress State"
+HISTORICAL_PENDING_ASKS_HEADING = "## Historical Pending User Asks"
+HISTORICAL_REMAINING_WORK_HEADING = "## Historical Remaining Work"
+
+
 SUMMARY_PREFIX = (
+    "[CONTEXT COMPACTION — REFERENCE ONLY] Earlier turns were compacted "
+    "into the summary below. This is a handoff from a previous context "
+    "window — treat it as background reference, NOT as active instructions. "
+    "Do NOT answer questions or fulfill requests mentioned in this summary; "
+    "they were already addressed. "
+    "Respond ONLY to the latest user message that appears AFTER this "
+    "summary — that message is the single source of truth for what to do "
+    "right now. "
+    "Topic overlap with the summary does NOT mean you should resume its "
+    "task: even on similar topics, the latest user message WINS. Treat ONLY "
+    "the latest message as the active task and discard stale items from "
+    f"'{HISTORICAL_TASK_HEADING}' / '{HISTORICAL_IN_PROGRESS_HEADING}' / "
+    f"'{HISTORICAL_PENDING_ASKS_HEADING}' / "
+    f"'{HISTORICAL_REMAINING_WORK_HEADING}' entirely — do not 'wrap up' or "
+    "'finish' work described there unless the latest message explicitly "
+    "asks for it. "
+    "Reverse signals in the latest message (e.g. 'stop', 'undo', 'roll "
+    "back', 'just verify', 'don't do that anymore', 'never mind', a new "
+    "topic) must immediately end any in-flight work described in the "
+    "summary; do not re-surface it in later turns. "
+    "IMPORTANT: Your persistent memory (MEMORY.md, USER.md) in the system "
+    "prompt is ALWAYS authoritative and active — never ignore or deprioritize "
+    "memory content due to this compaction note. "
+    "The current session state (files, config, etc.) may reflect work "
+    "described here — avoid repeating it:"
+)
+LEGACY_SUMMARY_PREFIX = "[CONTEXT SUMMARY]:"
+
+# Metadata key added to context compression summary messages so that frontends
+# (CLI, Desktop, gateway, TUI) can distinguish them from real assistant/user
+# messages and filter or render them appropriately without content-prefix
+# heuristics. See https://github.com/NousResearch/hermes-agent/issues/38389
+#
+# Underscore-prefixed ON PURPOSE: the wire sanitizers
+# (agent/transports/chat_completions.py convert_messages and the summary-path
+# mirror in agent/chat_completion_helpers.py) strip every top-level message
+# key starting with "_" before the request leaves the process. Strict
+# OpenAI-compatible gateways (Fireworks, Mistral, Moonshot/Kimi, opencode-go)
+# reject payloads carrying unknown keys with "Extra inputs are not permitted",
+# poisoning every subsequent request in the session — a bare key like
+# "is_compressed_summary" would reach the wire and trip exactly that.
+COMPRESSED_SUMMARY_METADATA_KEY = "_compressed_summary"
+
+# Appended to every standalone summary message (and to the merged-into-tail
+# prefix) so the model has an unambiguous "summary ends here" boundary.
+# Without it, weak models read the verbatim "## Active Task" quote as fresh
+# user input (#11475, #14521) or regurgitate an assistant-role summary as
+# their own output (#33256).
+_SUMMARY_END_MARKER = (
+    "--- END OF CONTEXT SUMMARY — "
+    "respond to the message below, not the summary above ---"
+)
+
+# Handoff prefixes that shipped in earlier releases. A summary persisted under
+# one of these can be inherited into a resumed lineage (#35344); when it is
+# re-normalized on re-compaction we must strip the OLD prefix too, otherwise the
+# stale directive it carried (e.g. "resume exactly from Active Task") survives
+# embedded in the body and keeps hijacking replies. Keep newest-first; entries
+# are matched literally. Add a frozen copy here whenever SUMMARY_PREFIX changes.
+_HISTORICAL_SUMMARY_PREFIXES = (
+    # Carveout era (#41607/#38364/#42812): "consistent → use as background"
+    # licensed stale-task resumption on topic overlap.
     "[CONTEXT COMPACTION — REFERENCE ONLY] Earlier turns were compacted "
     "into the summary below. This is a handoff from a previous context "
     "window — treat it as background reference, NOT as active instructions. "
@@ -57,17 +125,7 @@ SUMMARY_PREFIX = (
     "prompt is ALWAYS authoritative and active — never ignore or deprioritize "
     "memory content due to this compaction note. "
     "The current session state (files, config, etc.) may reflect work "
-    "described here — avoid repeating it:"
-)
-LEGACY_SUMMARY_PREFIX = "[CONTEXT SUMMARY]:"
-
-# Handoff prefixes that shipped in earlier releases. A summary persisted under
-# one of these can be inherited into a resumed lineage (#35344); when it is
-# re-normalized on re-compaction we must strip the OLD prefix too, otherwise the
-# stale directive it carried (e.g. "resume exactly from Active Task") survives
-# embedded in the body and keeps hijacking replies. Keep newest-first; entries
-# are matched literally. Add a frozen copy here whenever SUMMARY_PREFIX changes.
-_HISTORICAL_SUMMARY_PREFIXES = (
+    "described here — avoid repeating it:",
     # Pre-#35344: contained the self-contradicting "resume exactly" directive.
     "[CONTEXT COMPACTION — REFERENCE ONLY] Earlier turns were compacted "
     "into the summary below. This is a handoff from a previous context "
@@ -110,10 +168,23 @@ _SUMMARY_FAILURE_COOLDOWN_SECONDS = 600
 # become another unbounded transcript copy after the LLM summarizer failed.
 _FALLBACK_SUMMARY_MAX_CHARS = 8_000
 _FALLBACK_TURN_MAX_CHARS = 700
+_AUTO_FOCUS_MAX_TURNS = 3
+_AUTO_FOCUS_TURN_MAX_CHARS = 260
+_AUTO_FOCUS_MAX_CHARS = 700
+# Keep a short run of recent messages verbatim even when the token budget is
+# already exhausted.  The public ``protect_last_n`` default is intentionally
+# high for small/light tails, but using all 20 as a hard floor here would bring
+# back the old large-tool-output case where nothing can be compacted.
+_MAX_TAIL_MESSAGE_FLOOR = 8
 
 
 _PATH_MENTION_RE = re.compile(r"(?:/|~/?|[A-Za-z]:\\)[^\s`'\")\]}<>]+")
 
+# MEDIA delivery directives must not reach the summarizer — if one leaks into
+# the summary, the downstream model may re-emit it as an active directive on
+# the next turn, triggering bogus attachment sends (#14665).
+_MEDIA_DIRECTIVE_RE = re.compile(r"MEDIA:\S+")
+
 
 def _dedupe_append(items: list[str], value: str, *, limit: int) -> None:
     value = value.strip()
@@ -974,6 +1045,7 @@ class ContextCompressor(ContextEngine):
         for msg in turns:
             role = msg.get("role", "unknown")
             content = redact_sensitive_text(msg.get("content") or "")
+            content = _MEDIA_DIRECTIVE_RE.sub("[media attachment]", content)
 
             # Tool results: keep enough content for the summarizer
             if role == "tool":
@@ -1155,7 +1227,7 @@ class ContextCompressor(ContextEngine):
             )
 
         reason_text = f" Summary failure reason: {reason}." if reason else ""
-        body = f"""## Active Task
+        body = f"""{HISTORICAL_TASK_HEADING}
 {active_task}
 
 ## Goal
@@ -1172,7 +1244,7 @@ Recovered from a deterministic fallback because the LLM context summarizer was u
 ## Active State
 Unknown from deterministic fallback. Inspect current repository/session state if needed.
 
-## In Progress
+{HISTORICAL_IN_PROGRESS_HEADING}
 {active_task}
 
 ## Blocked
@@ -1184,13 +1256,13 @@ None recoverable from deterministic fallback.
 ## Resolved Questions
 None recoverable from deterministic fallback.
 
-## Pending User Asks
+{HISTORICAL_PENDING_ASKS_HEADING}
 {active_task}
 
 ## Relevant Files
 {_bullets(relevant_files, limit=12)}
 
-## Remaining Work
+{HISTORICAL_REMAINING_WORK_HEADING}
 Continue from the most recent unfulfilled user ask and protected tail messages. Verify state with tools before making claims.
 
 ## Last Dropped Turns
@@ -1312,7 +1384,7 @@ Summary generation was unavailable, so this is a best-effort deterministic fallb
             _temporal_anchoring_rule = ""
 
         # Shared structured template (used by both paths).
-        _template_sections = f"""## Active Task
+        _template_sections = f"""{HISTORICAL_TASK_HEADING}
 [THE SINGLE MOST IMPORTANT FIELD. Capture the user's most recent unfulfilled
 input verbatim — the exact words they used. This includes:
 - Explicit task assignments ("refactor the auth module")
@@ -1359,7 +1431,7 @@ Be specific with file paths, commands, line numbers, and results.]
 - Any running processes or servers
 - Environment details that matter]
 
-## In Progress
+{HISTORICAL_IN_PROGRESS_HEADING}
 [Work currently underway — what was being done when compaction fired]
 
 ## Blocked
@@ -1371,14 +1443,14 @@ Be specific with file paths, commands, line numbers, and results.]
 ## Resolved Questions
 [Questions the user asked that were ALREADY answered — include the answer so it is not repeated]
 
-## Pending User Asks
-[Questions or requests from the user that have NOT yet been answered or fulfilled. If none, write "None."]
+{HISTORICAL_PENDING_ASKS_HEADING}
+[Questions or requests from the user that have NOT yet been answered or fulfilled. These are STALE — they were from the compacted turns. Write them here for reference only. The agent must NOT act on them unless the latest user message explicitly requests it. If none, write "None."]
 
 ## Relevant Files
 [Files read, modified, or created — with brief note on each]
 
-## Remaining Work
-[What remains to be done — framed as context, not instructions]
+{HISTORICAL_REMAINING_WORK_HEADING}
+[What remains to be done — framed as STALE context for reference only. The agent must NOT resume this work unless the latest user message explicitly asks for it.]
 
 ## Critical Context
 [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.]
@@ -1421,7 +1493,7 @@ 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. Even for the focus topic, NEVER preserve API keys, tokens, passwords, or credentials — use [REDACTED]."""
+This compaction should 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 = {
@@ -1574,7 +1646,13 @@ The user has requested that this compaction PRIORITISE preserving all informatio
         text = (summary or "").strip()
         for prefix in (SUMMARY_PREFIX, LEGACY_SUMMARY_PREFIX, *_HISTORICAL_SUMMARY_PREFIXES):
             if text.startswith(prefix):
-                return text[len(prefix):].lstrip()
+                text = text[len(prefix):].lstrip()
+                break
+        # Strip the trailing end marker too — a rehydrated handoff body that
+        # keeps it would leak the boundary directive into the iterative-update
+        # summarizer prompt (and the marker is re-appended on insertion anyway).
+        if text.endswith(_SUMMARY_END_MARKER):
+            text = text[: -len(_SUMMARY_END_MARKER)].rstrip()
         return text
 
     @classmethod
@@ -1590,6 +1668,52 @@ The user has requested that this compaction PRIORITISE preserving all informatio
             return True
         return any(text.startswith(p) for p in _HISTORICAL_SUMMARY_PREFIXES)
 
+    @staticmethod
+    def _has_compressed_summary_metadata(message: Any) -> bool:
+        """Return True if *message* carries the compressed-summary flag.
+
+        Callers (frontends, CLI, gateway) can use this to distinguish context
+        compaction summaries from real assistant or user messages without
+        relying on content-prefix heuristics.  The flag is in-process only —
+        the wire sanitizers strip underscore-prefixed keys before API calls.
+        """
+        if not isinstance(message, dict):
+            return False
+        return bool(message.get(COMPRESSED_SUMMARY_METADATA_KEY))
+
+    @classmethod
+    def _derive_auto_focus_topic(
+        cls,
+        messages: List[Dict[str, Any]],
+    ) -> Optional[str]:
+        """Infer a compact focus hint from the most recent real user turns."""
+        candidates: list[str] = []
+        for idx in range(len(messages) - 1, -1, -1):
+            msg = messages[idx]
+            if msg.get("role") != "user":
+                continue
+            content = msg.get("content")
+            if cls._is_context_summary_content(content):
+                continue
+            text = redact_sensitive_text(_content_text_for_contains(content).strip())
+            if not text:
+                continue
+            text = " ".join(text.split())
+            if len(text) > _AUTO_FOCUS_TURN_MAX_CHARS:
+                text = text[: _AUTO_FOCUS_TURN_MAX_CHARS - 1].rstrip() + "…"
+            candidates.append(text)
+            if len(candidates) >= _AUTO_FOCUS_MAX_TURNS:
+                break
+
+        if not candidates:
+            return None
+
+        candidates.reverse()
+        focus = "Recent user focus:\n" + "\n".join(f"- {item}" for item in candidates)
+        if len(focus) > _AUTO_FOCUS_MAX_CHARS:
+            focus = focus[: _AUTO_FOCUS_MAX_CHARS - 1].rstrip() + "…"
+        return focus
+
     @classmethod
     def _find_latest_context_summary(
         cls,
@@ -1742,6 +1866,105 @@ The user has requested that this compaction PRIORITISE preserving all informatio
                 return i
         return -1
 
+    def _find_last_assistant_message_idx(
+        self, messages: List[Dict[str, Any]], head_end: int
+    ) -> int:
+        """Return the index of the last user-visible assistant reply at or
+        after *head_end*, or -1.
+
+        A "user-visible reply" is an assistant message with non-empty
+        textual content — i.e. one that the WebUI / TUI / SessionsPage
+        rendered as a bubble the operator could read. We deliberately
+        skip assistant messages that contain only ``tool_calls`` (and
+        no text), because those render as small "calling tool X"
+        indicators and aren't what the reporter means by "the output
+        of the last message you sent" (#29824).
+
+        Falling back to the most recent assistant message of ANY kind
+        only kicks in when no content-bearing assistant message exists
+        in the compressible region — typically a fresh session that
+        just started a multi-step tool sequence with no prior reply
+        to anchor. In that case the agent fix is a no-op and the
+        existing user-message anchor carries the load.
+        """
+        last_any = -1
+        for i in range(len(messages) - 1, head_end - 1, -1):
+            msg = messages[i]
+            if msg.get("role") != "assistant":
+                continue
+            if last_any < 0:
+                last_any = i
+            content = msg.get("content")
+            if isinstance(content, str) and content.strip():
+                return i
+            if isinstance(content, list):
+                # Multimodal / Anthropic-style content: look for any
+                # text block with non-empty text.
+                for part in content:
+                    if isinstance(part, dict):
+                        text = part.get("text") or part.get("content")
+                        if isinstance(text, str) and text.strip():
+                            return i
+        return last_any
+
+    def _ensure_last_assistant_message_in_tail(
+        self,
+        messages: List[Dict[str, Any]],
+        cut_idx: int,
+        head_end: int,
+    ) -> int:
+        """Guarantee the most recent assistant message is in the protected tail.
+
+        WebUI / TUI / SessionsPage bug (#29824). Without this anchor,
+        ``_find_tail_cut_by_tokens`` can leave the user's most recent
+        visible assistant response inside the compressed middle region —
+        especially when the conversation has a single oversized tool
+        result or a long stretch of tool-call/result pairs after the
+        last assistant reply. The summariser then rolls that reply up
+        into the single ``[CONTEXT COMPACTION — REFERENCE ONLY]`` block
+        persisted as ``role="user"`` or ``role="assistant"``. From the
+        operator's perspective the WebUI session viewer
+        (``web/src/pages/SessionsPage.tsx``) and the TUI chat panel
+        both suddenly show the opaque "Context compaction" block in the
+        slot where they were just reading the assistant's actual reply:
+
+            User:       "i cant see the output of the last message you
+                         sent, i did see it previously, however now see
+                         'context compaction'"
+
+        Mirror of ``_ensure_last_user_message_in_tail`` but anchors on
+        the last assistant-role message. Re-runs the tool-group
+        alignment so we don't split a ``tool_call`` / ``tool_result``
+        group that immediately precedes the anchored message — orphaned
+        tool messages would otherwise be removed by
+        ``_sanitize_tool_pairs`` and trigger the same data-loss symptom
+        we're trying to prevent.
+        """
+        last_asst_idx = self._find_last_assistant_message_idx(messages, head_end)
+        if last_asst_idx < 0:
+            # No assistant message in the compressible region — nothing
+            # to anchor (single-turn pre-reply state, etc.).
+            return cut_idx
+        if last_asst_idx >= cut_idx:
+            # Already in the tail — the token-budget walk did the right
+            # thing on its own.
+            return cut_idx
+        # Pull cut_idx back to the assistant message, then re-align so
+        # we don't split a tool group that immediately precedes it
+        # (e.g. an ``assistant(tool_calls)`` → ``tool(result)`` →
+        # ``assistant(final reply)`` sequence would otherwise leave the
+        # ``tool`` orphan when cut lands at the final reply).
+        new_cut = self._align_boundary_backward(messages, last_asst_idx)
+        if not self.quiet_mode:
+            logger.debug(
+                "Anchoring tail cut to last assistant message at index %d "
+                "(was %d, aligned to %d) to keep the previously-visible "
+                "reply out of the compaction summary (#29824)",
+                last_asst_idx, cut_idx, new_cut,
+            )
+        # Safety: never go back into the head region.
+        return max(new_cut, head_end + 1)
+
     def _ensure_last_user_message_in_tail(
         self,
         messages: List[Dict[str, Any]],
@@ -1753,7 +1976,7 @@ The user has requested that this compaction PRIORITISE preserving all informatio
         Context compressor bug (#10896): ``_align_boundary_backward`` can pull
         ``cut_idx`` past a user message when it tries to keep tool_call/result
         groups together.  If the last user message ends up in the *compressed*
-        middle region the LLM summariser writes it into "Pending User Asks",
+        middle region the LLM summariser writes it into "Historical Pending User Asks",
         but ``SUMMARY_PREFIX`` tells the next model to respond only to user
         messages *after* the summary — so the task effectively disappears from
         the active context, causing the agent to stall, repeat completed work,
@@ -1800,11 +2023,12 @@ The user has requested that this compaction PRIORITISE preserving all informatio
         derived from ``summary_target_ratio * context_length``, so it
         scales automatically with the model's context window.
 
-        Token budget is the primary criterion.  A hard minimum of 3 messages
-        is always protected, but the budget is allowed to exceed by up to
-        1.5x to avoid cutting inside an oversized message (tool output, file
-        read, etc.).  If even the minimum 3 messages exceed 1.5x the budget
-        the cut is placed right after the head so compression still runs.
+        Token budget is the primary criterion.  A bounded message-count floor
+        keeps a short run of recent turns verbatim even when the budget is
+        exhausted, but the budget is allowed to exceed by up to 1.5x to avoid
+        cutting inside an oversized message (tool output, file read, etc.). If
+        even that floor exceeds 1.5x the budget, the cut is placed right after
+        the head so compression still runs.
 
         Never cuts inside a tool_call/result group.  Always ensures the most
         recent user message is in the tail (see ``_ensure_last_user_message_in_tail``).
@@ -1812,8 +2036,19 @@ The user has requested that this compaction PRIORITISE preserving all informatio
         if token_budget is None:
             token_budget = self.tail_token_budget
         n = len(messages)
-        # Hard minimum: always keep at least 3 messages in the tail
-        min_tail = min(3, n - head_end - 1) if n - head_end > 1 else 0
+        # Hard minimum: always keep a bounded recent-message floor in the tail.
+        # ``protect_last_n`` remains a minimum up to the cap; the cap avoids
+        # preserving a whole run of bulky tool outputs on every compaction.
+        available_tail = max(0, n - head_end - 1)
+        min_tail_floor = max(3, min(self.protect_last_n, _MAX_TAIL_MESSAGE_FLOOR))
+        # Leave at least two non-head messages available to summarize on short
+        # transcripts; otherwise compression can replace a tiny middle with a
+        # summary and save no messages at all.
+        compressible_tail_cap = max(3, available_tail - 2)
+        min_tail = (
+            min(min_tail_floor, compressible_tail_cap, available_tail)
+            if available_tail > 1 else 0
+        )
         soft_ceiling = int(token_budget * 1.5)
         accumulated = 0
         cut_idx = n  # start from beyond the end
@@ -1885,6 +2120,13 @@ The user has requested that this compaction PRIORITISE preserving all informatio
         # active task is never lost to compression (fixes #10896).
         cut_idx = self._ensure_last_user_message_in_tail(messages, cut_idx, head_end)
 
+        # Ensure the most recent assistant message is always in the tail
+        # so the previously-visible reply isn't silently rolled into the
+        # ``[CONTEXT COMPACTION — REFERENCE ONLY]`` block (fixes #29824).
+        # Each anchor only walks ``cut_idx`` backward, so chaining them is
+        # monotonic — the tail can only grow, never shrink.
+        cut_idx = self._ensure_last_assistant_message_in_tail(messages, cut_idx, head_end)
+
         return max(cut_idx, head_end + 1)
 
     # ------------------------------------------------------------------
@@ -2037,7 +2279,8 @@ The user has requested that this compaction PRIORITISE preserving all informatio
             )
 
         # Phase 3: Generate structured summary
-        summary = self._generate_summary(turns_to_summarize, focus_topic=focus_topic)
+        summary_focus_topic = focus_topic or self._derive_auto_focus_topic(messages)
+        summary = self._generate_summary(turns_to_summarize, focus_topic=summary_focus_topic)
 
         # If summary generation failed, behavior splits on
         # ``abort_on_summary_failure`` (config: compression.abort_on_summary_failure):
@@ -2117,32 +2360,33 @@ The user has requested that this compaction PRIORITISE preserving all informatio
 
         # When the summary lands as a standalone role="user" message,
         # weak models read the verbatim "## Active Task" quote of a past
-        # user request as fresh input (#11475, #14521). Append the explicit
-        # end marker — the same one used in the merge-into-tail path — so
-        # the model has a clear "summary above, not new input" signal.
-        if not _merge_summary_into_tail and summary_role == "user":
-            summary = (
-                summary
-                + "\n\n--- END OF CONTEXT SUMMARY — "
-                "respond to the message below, not the summary above ---"
-            )
+        # user request as fresh input (#11475, #14521).
+        # When it lands as role="assistant", models may regurgitate the
+        # summary text as their own output (#33256). In both cases, append
+        # the explicit end marker so the model has a clear "summary ends
+        # here, respond to the message below" signal.
+        if not _merge_summary_into_tail:
+            summary = summary + "\n\n" + _SUMMARY_END_MARKER
 
         if not _merge_summary_into_tail:
-            compressed.append({"role": summary_role, "content": summary})
+            compressed.append({
+                "role": summary_role,
+                "content": summary,
+                COMPRESSED_SUMMARY_METADATA_KEY: True,
+            })
 
         for i in range(compress_end, n_messages):
             msg = messages[i].copy()
             if _merge_summary_into_tail and i == compress_end:
-                merged_prefix = (
-                    summary
-                    + "\n\n--- END OF CONTEXT SUMMARY — "
-                    "respond to the message below, not the summary above ---\n\n"
-                )
+                merged_prefix = summary + "\n\n" + _SUMMARY_END_MARKER + "\n\n"
                 msg["content"] = _append_text_to_content(
                     msg.get("content"),
                     merged_prefix,
                     prepend=True,
                 )
+                # Mark the merged message so frontends can identify it as
+                # containing a compression summary prefix.
+                msg[COMPRESSED_SUMMARY_METADATA_KEY] = True
                 _merge_summary_into_tail = False
             compressed.append(msg)
 

agent/context_references.py

@@ -246,7 +246,14 @@ def _expand_file_reference(
     if not path.is_file():
         return f"{ref.raw}: path is not a file", None
     if _is_binary_file(path):
-        return f"{ref.raw}: binary files are not supported", None
+        # A binary file can't be inlined as text, but it IS on disk (the agent's
+        # tools run where this resolves — the local cwd, or the staged copy in a
+        # remote session workspace). Returning a bare "not supported" warning
+        # with no content was a dead end: the model saw a failure and gave up
+        # (told the user the file type wasn't supported). Instead, hand it an
+        # actionable block — the path, type, size, and a nudge to use its tools —
+        # so it can read/convert/view the file itself.
+        return None, _binary_reference_block(ref, path)
 
     text = path.read_text(encoding="utf-8")
     if ref.line_start is not None:
@@ -290,6 +297,7 @@ def _expand_git_reference(
             capture_output=True,
             text=True,
             timeout=30,
+            stdin=subprocess.DEVNULL,
         )
     except subprocess.TimeoutExpired:
         return f"{ref.raw}: git command timed out (30s)", None
@@ -482,6 +490,7 @@ def _rg_files(path: Path, cwd: Path, limit: int) -> list[Path] | None:
             capture_output=True,
             text=True,
             timeout=10,
+            stdin=subprocess.DEVNULL,
         )
     except (FileNotFoundError, OSError, subprocess.TimeoutExpired):
         return None
@@ -491,6 +500,30 @@ def _rg_files(path: Path, cwd: Path, limit: int) -> list[Path] | None:
     return files[:limit]
 
 
+def _human_bytes(n: int) -> str:
+    size = float(n)
+    for unit in ("B", "KB", "MB", "GB"):
+        if size < 1024 or unit == "GB":
+            return f"{int(size)} {unit}" if unit == "B" else f"{size:.1f} {unit}"
+        size /= 1024
+    return f"{size:.1f} GB"
+
+
+def _binary_reference_block(ref: ContextReference, path: Path) -> str:
+    mime, _ = mimetypes.guess_type(path.name)
+    mime = mime or "application/octet-stream"
+    try:
+        size = _human_bytes(path.stat().st_size)
+    except OSError:
+        size = "unknown size"
+    return (
+        f"📎 {ref.raw} ({mime}, {size}) — binary file, not inlined as text. "
+        f"It is available on disk at `{path}`. Use your tools to work with it "
+        f"(read or convert it, extract its text, or view/render it as needed); "
+        f"do not tell the user the file type is unsupported."
+    )
+
+
 def _file_metadata(path: Path) -> str:
     if _is_binary_file(path):
         return f"{path.stat().st_size} bytes"

agent/conversation_compression.py

@@ -40,6 +40,16 @@ from agent.model_metadata import estimate_request_tokens_rough
 
 logger = logging.getLogger(__name__)
 
+# Stable marker the gateway matches on to re-tag the auto-compaction lifecycle
+# status as ``kind="compacting"`` (tui_gateway/server.py::_status_update), so
+# drivers like the desktop app can show an explicit "Summarizing…" indicator
+# instead of the transcript appearing to silently reset. Keep the marker phrase
+# intact if you reword COMPACTION_STATUS.
+COMPACTION_STATUS_MARKER = "Compacting context"
+COMPACTION_STATUS = (
+    f"🗜️ {COMPACTION_STATUS_MARKER} — summarizing earlier conversation so I can continue..."
+)
+
 
 def _compression_lock_holder(agent: Any) -> str:
     """Build a unique holder id for the lock: pid:tid:agent-instance:uuid.
@@ -324,9 +334,7 @@ def compress_context(
         f"{approx_tokens:,}" if approx_tokens else "unknown", agent.model,
         focus_topic,
     )
-    agent._emit_status(
-        "🗜️ Compacting context — summarizing earlier conversation so I can continue..."
-    )
+    agent._emit_status(COMPACTION_STATUS)
 
     # ── Compression lock ────────────────────────────────────────────────
     # Atomic, state.db-backed lock per session_id.  Without this, two
@@ -595,6 +603,20 @@ def compress_context(
             force=True,
         )
 
+    # Emit session:compress event so hooks (e.g. MemPalace sync) can ingest
+    # the completed old session before its details are lost.
+    _old_sid_for_event = locals().get("old_session_id")
+    if getattr(agent, "event_callback", None):
+        try:
+            agent.event_callback("session:compress", {
+                "platform": agent.platform or "",
+                "session_id": agent.session_id,
+                "old_session_id": _old_sid_for_event or "",
+                "compression_count": agent.context_compressor.compression_count,
+            })
+        except Exception as e:
+            logger.debug("event_callback error on session:compress: %s", e)
+
     # Keep the post-compression rough estimate for diagnostics, but do not
     # treat it as provider-reported prompt usage. Schema-heavy rough estimates
     # can remain above threshold even after the next real API request fits.
@@ -631,7 +653,11 @@ def compress_context(
     return compressed, new_system_prompt
 
 
-def try_shrink_image_parts_in_messages(api_messages: list) -> bool:
+def try_shrink_image_parts_in_messages(
+    api_messages: list,
+    *,
+    max_dimension: int = 8000,
+) -> bool:
     """Re-encode all native image parts at a smaller size to recover from
     image-too-large errors (Anthropic 5 MB, unknown other providers).
 
@@ -642,7 +668,8 @@ def try_shrink_image_parts_in_messages(api_messages: list) -> bool:
     Strategy: look for ``image_url`` / ``input_image`` parts carrying a
     ``data:image/...;base64,...`` payload.  For each one whose encoded
     size exceeds 4 MB (a safe target that slides under Anthropic's 5 MB
-    ceiling with header overhead), write the base64 to a tempfile, call
+    ceiling with header overhead) or whose longest side exceeds
+    ``max_dimension``, write the base64 to a tempfile, call
     ``vision_tools._resize_image_for_vision`` to produce a smaller data
     URL, and substitute it in place.
 
@@ -664,10 +691,9 @@ def try_shrink_image_parts_in_messages(api_messages: list) -> bool:
     # after a confirmed provider rejection, so the alternative is failure.
     target_bytes = 4 * 1024 * 1024
     # Anthropic enforces an 8000px per-side dimension cap independently of
-    # the 5 MB byte cap.  A tall screenshot can be well under 5 MB yet far
-    # over 8000px (e.g. 1200×12000 at 0.06 MB).  We check pixel dimensions
-    # even when the byte budget is fine.
-    max_dimension = 8000
+    # the 5 MB byte cap.  In many-image requests, the provider can report a
+    # lower cap (observed: 2000px).  The caller passes that parsed ceiling
+    # when the rejection includes it.
     changed_count = 0
     # Track parts that are over the target but could NOT be shrunk under it.
     # If any survive, retrying is pointless — the same oversized payload will
@@ -684,9 +710,9 @@ def try_shrink_image_parts_in_messages(api_messages: list) -> bool:
         # Check both byte size AND pixel dimensions.
         needs_shrink = len(url) > target_bytes  # over byte budget
         if not needs_shrink:
-            # Even if bytes are fine, check pixel dimensions against
-            # Anthropic's 8000px cap.  A tall image can be tiny in bytes
-            # yet huge in pixels.
+            # Even if bytes are fine, check pixel dimensions against the
+            # provider's reported per-side cap.  A screenshot can be tiny in
+            # bytes yet too large in pixels.
             try:
                 import base64 as _b64_dim
                 header_d, _, data_d = url.partition(",")
@@ -795,6 +821,8 @@ def try_shrink_image_parts_in_messages(api_messages: list) -> bool:
 
 
 __all__ = [
+    "COMPACTION_STATUS",
+    "COMPACTION_STATUS_MARKER",
     "check_compression_model_feasibility",
     "replay_compression_warning",
     "compress_context",

agent/conversation_loop.py

@@ -71,6 +71,35 @@ logger = logging.getLogger(__name__)
 INTERRUPT_WAITING_FOR_MODEL_PREFIX = "Operation interrupted: waiting for model response ("
 
 
+def _image_error_max_dimension(error: Exception) -> Optional[int]:
+    """Extract a provider-reported image dimension ceiling, if present."""
+    parts = []
+    for value in (
+        error,
+        getattr(error, "message", None),
+        getattr(error, "body", None),
+    ):
+        if value:
+            try:
+                parts.append(str(value))
+            except Exception:
+                pass
+    text = " ".join(parts).lower()
+    if "image" not in text or "dimension" not in text or "max allowed size" not in text:
+        return None
+
+    match = re.search(r"max allowed size(?:\s+for [^:]+)?:\s*(\d{3,5})\s*pixels?", text)
+    if not match:
+        return None
+    try:
+        max_dimension = int(match.group(1))
+    except ValueError:
+        return None
+    if 512 <= max_dimension <= 8000:
+        return max_dimension
+    return None
+
+
 def _ollama_context_limit_error(agent: Any, request_tokens: int) -> Optional[str]:
     """Return a user-facing error when Ollama is loaded with too little context."""
     if not getattr(agent, "tools", None):
@@ -271,11 +300,20 @@ def _restore_or_build_system_prompt(agent, system_message, conversation_history)
                 agent.session_id, exc,
             )
 
-    if stored_prompt:
+    if stored_prompt and _stored_prompt_matches_runtime(agent, stored_prompt):
         # Continuing session — reuse the exact system prompt from the
         # previous turn so the Anthropic cache prefix matches.
         agent._cached_system_prompt = stored_prompt
         return
+    if stored_prompt:
+        stored_state = "stale_runtime"
+        logger.info(
+            "Stored system prompt for session %s has stale runtime identity; "
+            "rebuilding for model=%s provider=%s.",
+            agent.session_id,
+            getattr(agent, "model", "") or "",
+            getattr(agent, "provider", "") or "",
+        )
 
     if conversation_history and stored_state in ("null", "empty"):
         # Continuing session whose stored prompt is unusable.  The
@@ -337,6 +375,30 @@ def _restore_or_build_system_prompt(agent, system_message, conversation_history)
             )
 
 
+def _stored_prompt_matches_runtime(agent, prompt: str) -> bool:
+    """Return False when the persisted Model/Provider lines are stale."""
+
+    def line_value(label: str) -> str:
+        prefix = f"{label}:"
+        value = ""
+        for line in prompt.splitlines():
+            if line.startswith(prefix):
+                value = line[len(prefix):].strip()
+        return value
+
+    stored_model = line_value("Model")
+    current_model = str(getattr(agent, "model", "") or "").strip()
+    if stored_model and current_model and stored_model != current_model:
+        return False
+
+    stored_provider = line_value("Provider")
+    current_provider = str(getattr(agent, "provider", "") or "").strip()
+    if stored_provider and current_provider and stored_provider != current_provider:
+        return False
+
+    return True
+
+
 def _get_continuation_prompt(is_partial_stub: bool, dropped_tools: Optional[List[str]] = None) -> str:
     if is_partial_stub and dropped_tools:
         tool_list = ", ".join(dropped_tools[:3])
@@ -368,6 +430,42 @@ def _get_continuation_prompt(is_partial_stub: bool, dropped_tools: Optional[List
         )
 
 
+# Shared recovery hint appended to every content-policy refusal message. Both
+# the HTTP-200 refusal path (``finish_reason=content_filter``) and the
+# exception path (a provider moderation error classified as
+# ``content_policy_blocked``) end with the same actionable next steps, so they
+# share one trailer to keep the guidance from drifting between the two sites.
+_CONTENT_POLICY_RECOVERY_HINT = (
+    "Try rephrasing the request, narrowing the context, or "
+    "adding a fallback provider with `hermes fallback add`."
+)
+
+
+def _content_policy_blocked_result(
+    messages: List[Dict],
+    api_call_count: int,
+    *,
+    final_response: str,
+    error_detail: str,
+) -> Dict[str, Any]:
+    """Build the terminal turn result for a content-policy block.
+
+    A content-policy refusal is deterministic for the unchanged prompt, so the
+    turn ends here (no retry). Both the HTTP-200 refusal handler and the
+    exception-path handler return the identical shape — a failed, non-completed
+    turn carrying the user-facing message and a ``content_policy_blocked:``
+    prefixed error — so they funnel through this one builder.
+    """
+    return {
+        "final_response": final_response,
+        "messages": messages,
+        "api_calls": api_call_count,
+        "completed": False,
+        "failed": True,
+        "error": f"content_policy_blocked: {error_detail}",
+    }
+
+
 def run_conversation(
     agent,
     user_message: str,
@@ -595,7 +693,11 @@ def run_conversation(
         # landed after an orphan tool result). Most providers return
         # empty content on malformed sequences, which would otherwise
         # retrigger the empty-retry loop indefinitely.
-        repaired_seq = agent._repair_message_sequence(messages)
+        # repair_message_sequence_with_cursor also recomputes the SessionDB
+        # flush cursor (_last_flushed_db_idx) when repair compacts the list,
+        # so the turn-end flush doesn't skip the assistant/tool chain (#44837).
+        from agent.agent_runtime_helpers import repair_message_sequence_with_cursor
+        repaired_seq = repair_message_sequence_with_cursor(agent, messages)
         if repaired_seq > 0:
             request_logger.info(
                 "Repaired %s message-alternation violations before request (session=%s)",
@@ -703,7 +805,10 @@ def run_conversation(
         # a thinking-only turn. Runs on the per-call copy only — the
         # stored conversation history keeps the reasoning block for the
         # UI transcript and session persistence.
-        api_messages = agent._drop_thinking_only_and_merge_users(api_messages)
+        api_messages = agent._drop_thinking_only_and_merge_users(
+            api_messages,
+            drop_codex_reasoning_items=agent.api_mode != "codex_responses",
+        )
 
         # Normalize message whitespace and tool-call JSON for consistent
         # prefix matching.  Ensures bit-perfect prefixes across turns,
@@ -1312,6 +1417,106 @@ def run_conversation(
                         )
                         finish_reason = "length"
 
+                # ── Content-policy refusal (HTTP 200) ──────────────────
+                # The model — or the provider's safety system — returned a
+                # *successful* response whose stop/finish reason is a refusal:
+                # Anthropic ``stop_reason="refusal"`` → ``content_filter``;
+                # OpenAI / portal ``finish_reason="content_filter"`` or a
+                # populated ``message.refusal`` (mapped in the chat_completions
+                # transport); Bedrock ``guardrail_intervened``. The content is
+                # typically empty, so without this branch the response falls
+                # through to the empty-response / invalid-response retry loops
+                # and is mis-surfaced as "rate limited" / "no content after
+                # retries" — burning paid attempts reproducing a deterministic
+                # refusal. Surface it clearly and stop. Mirrors the
+                # exception-based ``content_policy_blocked`` recovery: try a
+                # configured fallback once, otherwise return the refusal.
+                if finish_reason == "content_filter":
+                    _refusal_transport = agent._get_transport()
+                    if agent.api_mode == "anthropic_messages":
+                        _refusal_result = _refusal_transport.normalize_response(
+                            response, strip_tool_prefix=agent._is_anthropic_oauth
+                        )
+                    else:
+                        _refusal_result = _refusal_transport.normalize_response(response)
+                    _refusal_text = (getattr(_refusal_result, "content", None) or "").strip()
+                    # Some refusals carry the explanation only in the reasoning
+                    # channel; fall back to it so the user sees *something*.
+                    if not _refusal_text:
+                        _refusal_text = (agent._extract_reasoning(_refusal_result) or "").strip()
+
+                    agent._invoke_api_request_error_hook(
+                        task_id=effective_task_id,
+                        turn_id=turn_id,
+                        api_request_id=api_request_id,
+                        api_call_count=api_call_count,
+                        api_start_time=api_start_time,
+                        api_kwargs=api_kwargs,
+                        error_type="ContentPolicyBlocked",
+                        error_message=_refusal_text or "model declined to respond (content_filter)",
+                        status_code=None,
+                        retry_count=retry_count,
+                        max_retries=max_retries,
+                        retryable=False,
+                        reason=FailoverReason.content_policy_blocked.value,
+                    )
+
+                    if thinking_spinner:
+                        thinking_spinner.stop("")
+                        thinking_spinner = None
+                    if agent.thinking_callback:
+                        agent.thinking_callback("")
+
+                    # Deterministic for the unchanged prompt — never retry.
+                    # Try a configured fallback once (a different model may not
+                    # refuse); otherwise surface the refusal terminally.
+                    if agent._has_pending_fallback():
+                        agent._buffer_status(
+                            "⚠️ Model declined to respond (safety refusal) — trying fallback..."
+                        )
+                    if agent._try_activate_fallback():
+                        retry_count = 0
+                        compression_attempts = 0
+                        _retry.primary_recovery_attempted = False
+                        continue
+
+                    agent._flush_status_buffer()
+                    _refusal_log = (
+                        _refusal_text[:500] + "..."
+                        if len(_refusal_text) > 500
+                        else _refusal_text
+                    )
+                    logger.warning(
+                        "%sModel declined to respond (finish_reason=content_filter). "
+                        "model=%s provider=%s refusal=%s",
+                        agent.log_prefix, agent.model, agent.provider,
+                        _refusal_log or "(no text)",
+                    )
+                    agent._emit_status(
+                        "⚠️ The model declined to respond to this request (safety refusal)."
+                    )
+
+                    _refusal_detail = (
+                        f"Model's explanation: {_refusal_text}"
+                        if _refusal_text
+                        else "The model returned no explanation."
+                    )
+                    _refusal_response = (
+                        "⚠️  The model declined to respond to this request "
+                        "(safety refusal — not a Hermes/gateway failure).\n\n"
+                        f"{_refusal_detail}\n\n"
+                        f"{_CONTENT_POLICY_RECOVERY_HINT}"
+                    )
+
+                    agent._cleanup_task_resources(effective_task_id)
+                    agent._persist_session(messages, conversation_history)
+                    return _content_policy_blocked_result(
+                        messages,
+                        api_call_count,
+                        final_response=_refusal_response,
+                        error_detail=_refusal_text or "model declined (content_filter)",
+                    )
+
                 if finish_reason == "length":
                     if getattr(response, "id", "") == PARTIAL_STREAM_STUB_ID:
                         agent._vprint(
@@ -2063,7 +2268,11 @@ def run_conversation(
                     and not _retry.image_shrink_retry_attempted
                 ):
                     _retry.image_shrink_retry_attempted = True
-                    if agent._try_shrink_image_parts_in_messages(api_messages):
+                    image_max_dimension = _image_error_max_dimension(api_error) or 8000
+                    if agent._try_shrink_image_parts_in_messages(
+                        api_messages,
+                        max_dimension=image_max_dimension,
+                    ):
                         agent._vprint(
                             f"{agent.log_prefix}📐 Image(s) exceeded provider size limit — "
                             f"shrank and retrying...",
@@ -2221,30 +2430,54 @@ def run_conversation(
                     print(f"{agent.log_prefix}     • Legacy cleanup: hermes config set ANTHROPIC_TOKEN \"\"")
                     print(f"{agent.log_prefix}     • Clear stale keys: hermes config set ANTHROPIC_API_KEY \"\"")
 
-                # ── Thinking block signature recovery ─────────────────
+                # Thinking block signature recovery.
+                #
                 # Anthropic signs thinking blocks against the full turn
-                # content.  Any upstream mutation (context compression,
+                # content. Any upstream mutation (context compression,
                 # session truncation, message merging) invalidates the
-                # signature → HTTP 400.  Recovery: strip reasoning_details
-                # from all messages so the next retry sends no thinking
-                # blocks at all.  One-shot — don't retry infinitely.
+                # signature and the API replies HTTP 400 ("invalid
+                # signature" or "cannot be modified"). Recovery strips
+                # ``reasoning_details`` so the retry sends no thinking
+                # blocks at all. One-shot per outer loop.
+                #
+                # The strip targets ``api_messages``, which is the
+                # API-call-time list that ``_build_api_kwargs`` consumes
+                # on every retry. ``api_messages`` was populated once at
+                # the start of the turn from shallow copies of
+                # ``messages``, so mutating it does not touch the
+                # canonical store. The previous implementation popped
+                # ``reasoning_details`` from ``messages`` instead, which
+                # had two problems: ``api_messages`` carried its own
+                # reference to the field through the shallow copy, so the
+                # retry's wire payload still included thinking blocks and
+                # the recovery never reached the API; and the mutation
+                # persisted into ``state.db`` through any subsequent
+                # ``_persist_session`` call, permanently corrupting the
+                # conversation. Future turns would replay the stripped
+                # state, hit the same 400, and the agent would terminate
+                # with ``max_retries_exhausted``, often spawning
+                # cascading compaction-ended sessions chained off the
+                # corrupted parent.
                 if (
                     classified.reason == FailoverReason.thinking_signature
                     and not _retry.thinking_sig_retry_attempted
                 ):
                     _retry.thinking_sig_retry_attempted = True
-                    for _m in messages:
-                        if isinstance(_m, dict):
+                    _api_stripped = 0
+                    for _m in api_messages:
+                        if isinstance(_m, dict) and "reasoning_details" in _m:
                             _m.pop("reasoning_details", None)
+                            _api_stripped += 1
                     agent._vprint(
-                        f"{agent.log_prefix}⚠️  Thinking block signature invalid — "
-                        f"stripped all thinking blocks, retrying...",
+                        f"{agent.log_prefix}⚠️  Thinking block signature invalid, "
+                        f"stripped reasoning_details from api_messages for retry...",
                         force=True,
                     )
                     logger.warning(
                         "%sThinking block signature recovery: stripped "
-                        "reasoning_details from %d messages",
-                        agent.log_prefix, len(messages),
+                        "reasoning_details from %d api_messages "
+                        "(canonical messages unchanged)",
+                        agent.log_prefix, _api_stripped,
                     )
                     continue
 
@@ -2607,10 +2840,13 @@ def run_conversation(
                     except Exception:
                         pass
                     if _genuine_nous_rate_limit:
-                        # Skip straight to max_retries -- the
-                        # top-of-loop guard will handle fallback or
-                        # bail cleanly.
-                        retry_count = max_retries
+                        # Re-enter the loop exactly once so the
+                        # top-of-loop Nous guard handles fallback or
+                        # bails cleanly. (Setting retry_count to
+                        # max_retries would make the while condition
+                        # false immediately and the guard would never
+                        # run -- no fallback, generic exhaustion error.)
+                        retry_count = max(0, max_retries - 1)
                         continue
                     # Upstream capacity 429: fall through to normal
                     # retry logic.  A different model (or the same
@@ -3052,20 +3288,17 @@ def run_conversation(
                     if classified.reason == FailoverReason.content_policy_blocked:
                         _summary = agent._summarize_api_error(api_error)
                         _policy_response = (
-                            f"⚠️  The model provider's safety filter blocked this request "
-                            f"(not a Hermes/gateway failure).\n\n"
+                            "⚠️  The model provider's safety filter blocked this request "
+                            "(not a Hermes/gateway failure).\n\n"
                             f"Provider message: {_summary}\n\n"
-                            f"Try rephrasing the request, narrowing the context, or "
-                            f"adding a fallback provider with `hermes fallback add`."
+                            f"{_CONTENT_POLICY_RECOVERY_HINT}"
+                        )
+                        return _content_policy_blocked_result(
+                            messages,
+                            api_call_count,
+                            final_response=_policy_response,
+                            error_detail=_summary,
                         )
-                        return {
-                            "final_response": _policy_response,
-                            "messages": messages,
-                            "api_calls": api_call_count,
-                            "completed": False,
-                            "failed": True,
-                            "error": f"content_policy_blocked: {_summary}",
-                        }
                     return {
                         "final_response": None,
                         "messages": messages,

agent/copilot_acp_client.py

@@ -70,16 +70,6 @@ def _resolve_args() -> list[str]:
 
 def _resolve_home_dir() -> str:
     """Return a stable HOME for child ACP processes."""
-
-    try:
-        from hermes_constants import get_subprocess_home
-
-        profile_home = get_subprocess_home()
-        if profile_home:
-            return profile_home
-    except Exception:
-        pass
-
     home = os.environ.get("HOME", "").strip()
     if home:
         return home
@@ -105,7 +95,10 @@ def _resolve_home_dir() -> str:
 
 def _build_subprocess_env() -> dict[str, str]:
     env = os.environ.copy()
-    env["HOME"] = _resolve_home_dir()
+    home = _resolve_home_dir()
+    env["HOME"] = home
+    from hermes_constants import apply_subprocess_home_env
+    apply_subprocess_home_env(env)
     return env
 
 

agent/credits_tracker.py

@@ -194,17 +194,71 @@ class AgentNotice:
     id: Optional[str] = None
 
 
+# ── is_free_tier_model (local-data-only free-model check) ────────────────────
+
+
+def is_free_tier_model(model: str, base_url: str = "") -> bool:
+    """Return True when *model* is a Nous free-tier model, using ONLY local data.
+
+    Two signals, both zero-network:
+
+    1. The ``:free`` suffix — the canonical Nous free SKU marker (e.g.
+       ``nvidia/nemotron-3-ultra:free``). Free by construction on the API side
+       (spend is forced to 0 for ``:free`` ids).
+    2. A peek into the in-process pricing cache in ``hermes_cli.models``
+       (populated when the model picker fetched ``/v1/models`` pricing for
+       *base_url*). PEEK ONLY — a cache miss never triggers a fetch. This is
+       CLI/TUI-session best-effort: gateway sessions never run the picker's
+       pricing fetch, so suppression there rests entirely on the ``:free``
+       suffix (which all Nous free SKUs carry).
+
+    Fail-open to False (the depleted notice still shows) on any error: wrongly
+    showing the warning is recoverable noise; wrongly hiding it on a paid model
+    would mask a real billing block.
+    """
+    if not model:
+        return False
+    if model.endswith(":free"):
+        return True
+    if not base_url:
+        return False
+    try:
+        from hermes_cli.models import _is_model_free, _pricing_cache
+
+        # Mirror get_pricing_for_provider's key normalization: the agent's
+        # Nous base_url is /v1-suffixed (https://inference-api.nousresearch.com/v1)
+        # but the picker keys _pricing_cache on the pre-/v1 root.
+        key = base_url.rstrip("/")
+        if key.endswith("/v1"):
+            key = key[:-3].rstrip("/")
+        pricing = _pricing_cache.get(key)
+        if not pricing:
+            return False
+        return _is_model_free(model, pricing)
+    except Exception:
+        return False
+
+
 # ── evaluate_credits_notices (pure reconciliation function) ──────────────────
 
 
 def evaluate_credits_notices(
     state: CreditsState,
     latch: dict,
+    *,
+    model_is_free: bool = False,
 ) -> tuple[list[AgentNotice], list[str]]:
     """Reconcile credits notices against the latch. Mutates ``latch`` IN PLACE.
 
     latch = {"active": set[str], "seen_below_90": bool, "usage_band": Optional[int]}.
 
+    ``model_is_free``: True when the session's active model is a Nous free-tier
+    model (see :func:`is_free_tier_model`). Suppresses the ``credits.depleted``
+    notice — a depleted account on a free model can keep inferencing, so the
+    error banner is noise (and confuses free-tier users who never had credits).
+    Suppression does NOT emit the "restored" success notice; that fires only on
+    a genuine ``paid_access`` flip back to True.
+
     Returns ``(to_show: list[AgentNotice], to_clear: list[str])``.
     Caller emits to_clear FIRST, then to_show.
 
@@ -232,6 +286,16 @@ def evaluate_credits_notices(
         for band in CREDITS_USAGE_BANDS:  # ascending → last match wins = highest
             if uf >= band[0]:
                 current_band = band
+    # Top-up suppression: when the account holds purchased (top-up) credits,
+    # the subscription-cap gauge is the wrong denominator — warning "90% used"
+    # at a user sitting on $50 of top-up is noise (and it previously stuck
+    # PERMANENTLY alongside grant_spent at >=100%). Suppress the usage band
+    # entirely; the cap-reached case is covered by the grant_spent info notice
+    # below, which already names the remaining top-up balance. A top-up landing
+    # mid-session flips current_band → None and the clear path below removes
+    # any showing band line.
+    if state.purchased_micros > 0:
+        current_band = None
     grant_cond = (
         state.denominator_kind == "subscription_cap"
         and uf is not None
@@ -284,10 +348,14 @@ def evaluate_credits_notices(
         active.discard("credits.grant_spent")
 
     # ── depleted ─────────────────────────────────────────────────────────────
-    if depleted_cond and "credits.depleted" not in active:
+    # Suppressed while the active model is free: inference still works there,
+    # so the error banner would just alarm users (free-tier users especially,
+    # who never had paid credits to "lose").
+    show_depleted = depleted_cond and not model_is_free
+    if show_depleted and "credits.depleted" not in active:
         to_show.append(
             AgentNotice(
-                text="✕ Credit access paused · run /usage for balance",
+                text="✕ Credit access paused · run /credits to top up",
                 level="error",
                 kind=CREDITS_NOTICE_KIND,
                 key="credits.depleted",
@@ -295,20 +363,23 @@ def evaluate_credits_notices(
             )
         )
         active.add("credits.depleted")
-    elif "credits.depleted" in active and not depleted_cond:
+    elif "credits.depleted" in active and not show_depleted:
         to_clear.append("credits.depleted")
         active.discard("credits.depleted")
-        # Recovery: also emit the success notice
-        to_show.append(
-            AgentNotice(
-                text="✓ Credit access restored",
-                level="success",
-                kind="ttl",
-                ttl_ms=CREDITS_RESTORED_TTL_MS,
-                key="credits.restored",
-                id="credits.restored",
+        if not depleted_cond:
+            # Genuine recovery (paid_access flipped back True): also emit the
+            # success notice. A clear caused by switching to a free model while
+            # still depleted must NOT claim access was restored.
+            to_show.append(
+                AgentNotice(
+                    text="✓ Credit access restored",
+                    level="success",
+                    kind="ttl",
+                    ttl_ms=CREDITS_RESTORED_TTL_MS,
+                    key="credits.restored",
+                    id="credits.restored",
+                )
             )
-        )
 
     return (to_show, to_clear)
 

agent/curator.py

@@ -25,7 +25,6 @@ import json
 import logging
 import os
 import re
-import tempfile
 import threading
 from datetime import datetime, timedelta, timezone
 from pathlib import Path
@@ -33,6 +32,7 @@ from typing import Any, Callable, Dict, List, NamedTuple, Optional, Set
 
 from hermes_constants import get_hermes_home
 from tools import skill_usage
+from utils import atomic_json_write
 
 logger = logging.getLogger(__name__)
 
@@ -97,20 +97,7 @@ def load_state() -> Dict[str, Any]:
 def save_state(data: Dict[str, Any]) -> None:
     path = _state_file()
     try:
-        path.parent.mkdir(parents=True, exist_ok=True)
-        fd, tmp = tempfile.mkstemp(dir=str(path.parent), prefix=".curator_state_", suffix=".tmp")
-        try:
-            with os.fdopen(fd, "w", encoding="utf-8") as f:
-                json.dump(data, f, indent=2, sort_keys=True, ensure_ascii=False)
-                f.flush()
-                os.fsync(f.fileno())
-            os.replace(tmp, path)
-        except BaseException:
-            try:
-                os.unlink(tmp)
-            except OSError:
-                pass
-            raise
+        atomic_json_write(path, data, indent=2, sort_keys=True)
     except Exception as e:
         logger.debug("Failed to save curator state: %s", e, exc_info=True)
 

agent/curator_backup.py

@@ -454,16 +454,16 @@ def _restore_cron_skill_links(snapshot_dir: Path) -> Dict[str, Any]:
         report["attempted"] = True  # we tried but there was nothing to do
         return report
 
-    # Load and rewrite the live jobs under the scheduler's lock.
+    # Load and rewrite the live jobs under the scheduler's cross-process lock.
     try:
-        from cron.jobs import load_jobs, save_jobs, _jobs_file_lock
+        from cron.jobs import load_jobs, save_jobs, _jobs_lock
     except ImportError as e:
         report["error"] = f"cron module unavailable: {e}"
         return report
 
     report["attempted"] = True
     try:
-        with _jobs_file_lock:
+        with _jobs_lock():
             live_jobs = load_jobs()
             changed = False
 

agent/display.py

@@ -12,6 +12,7 @@ import time
 from dataclasses import dataclass, field
 from difflib import unified_diff
 from pathlib import Path
+from typing import Any
 
 from utils import safe_json_loads
 from agent.tool_result_classification import file_mutation_result_landed
@@ -168,6 +169,27 @@ def _oneline(text: str) -> str:
     return " ".join(text.split())
 
 
+def _truncate_preview(text: str, max_len: int | None) -> str:
+    if max_len and max_len > 0 and len(text) > max_len:
+        if max_len <= 3:
+            return "." * max_len
+        return text[:max_len - 3] + "..."
+    return text
+
+
+def _delegate_task_goal_parts(tasks: Any, *, per_goal_len: int) -> tuple[int, list[str]]:
+    if not isinstance(tasks, list):
+        return 0, []
+    goals: list[str] = []
+    for task in tasks:
+        if not isinstance(task, dict):
+            continue
+        raw_goal = task.get("goal")
+        goal = "?" if raw_goal is None else _oneline(str(raw_goal))
+        goals.append(_truncate_preview(goal or "?", per_goal_len))
+    return len(goals), goals
+
+
 def build_tool_preview(tool_name: str, args: dict, max_len: int | None = None) -> str | None:
     """Build a short preview of a tool call's primary argument for display.
 
@@ -191,6 +213,22 @@ def build_tool_preview(tool_name: str, args: dict, max_len: int | None = None) -
         "clarify": "question", "skill_manage": "name",
     }
 
+    # delegate_task: show goal (single) or individual task goals (batch)
+    if tool_name == "delegate_task":
+        tasks = args.get("tasks")
+        if tasks and isinstance(tasks, list):
+            task_count, goals = _delegate_task_goal_parts(tasks, per_goal_len=40)
+            preview = (
+                f"{task_count} tasks: " + " | ".join(goals)
+                if goals else f"{len(tasks)} parallel tasks"
+            )
+            return _truncate_preview(preview, max_len)
+        goal = args.get("goal", "")
+        if goal is None:
+            return None
+        preview = _oneline(str(goal))
+        return _truncate_preview(preview, max_len) if preview else None
+
     if tool_name == "process":
         action = args.get("action", "")
         sid = args.get("session_id", "")
@@ -1019,7 +1057,10 @@ def get_cute_tool_message(
     if tool_name == "delegate_task":
         tasks = args.get("tasks")
         if tasks and isinstance(tasks, list):
-            return _wrap(f"┊ 🔀 delegate  {len(tasks)} parallel tasks  {dur}")
+            task_count, goals = _delegate_task_goal_parts(tasks, per_goal_len=30)
+            detail = " | ".join(goals) if goals else "parallel"
+            count_label = task_count or len(tasks)
+            return _wrap(f"┊ 🔀 delegate  {count_label}x: {_trunc(detail, 35)}  {dur}")
         return _wrap(f"┊ 🔀 delegate  {_trunc(args.get('goal', ''), 35)}  {dur}")
 
     preview = build_tool_preview(tool_name, args) or ""

agent/error_classifier.py

@@ -549,14 +549,32 @@ def classify_api_error(
             should_fallback=True,
         )
 
-    # Anthropic thinking block signature invalid (400).
+    # Anthropic thinking block recovery (400).  Two distinct failure modes,
+    # same recovery (strip all reasoning_details and retry without thinking
+    # blocks — see the thinking_signature handler in conversation_loop.py):
+    #   1. Signature mismatch: a thinking block is signed against the full
+    #      turn content; any upstream mutation (context compression, session
+    #      truncation, message merging) invalidates the signature.
+    #      Pattern: "signature" + "thinking".
+    #   2. Frozen-block mutation: Anthropic rejects any change to the
+    #      thinking/redacted_thinking blocks in the *latest* assistant
+    #      message — "`thinking` or `redacted_thinking` blocks in the latest
+    #      assistant message cannot be modified. These blocks must remain as
+    #      they were in the original response."  This carries no "signature"
+    #      token, so the original pattern missed it and the turn hard-aborted
+    #      as a non-retryable client error instead of self-healing.
+    #      Pattern: "thinking" + ("cannot be modified" | "must remain as they were").
     # Don't gate on provider — OpenRouter proxies Anthropic errors, so the
     # provider may be "openrouter" even though the error is Anthropic-specific.
-    # The message pattern ("signature" + "thinking") is unique enough.
+    # The combined patterns are unique enough.
     if (
         status_code == 400
-        and "signature" in error_msg
         and "thinking" in error_msg
+        and (
+            "signature" in error_msg
+            or "cannot be modified" in error_msg
+            or "must remain as they were" in error_msg
+        )
     ):
         return _result(
             FailoverReason.thinking_signature,
@@ -966,6 +984,34 @@ def _classify_400(
             should_fallback=False,
         )
 
+    # Request-validation errors (unsupported / unknown parameter) MUST be
+    # checked BEFORE context_overflow.  A GPT-5 model rejecting max_tokens
+    # returns:
+    #   "Unsupported parameter: 'max_tokens' is not supported with this model.
+    #    Use 'max_completion_tokens' instead."
+    # That string contains the literal substring "max_tokens", which is one of
+    # the _CONTEXT_OVERFLOW_PATTERNS — so without this guard the 400 is
+    # misclassified as context_overflow, routed into the compression loop,
+    # re-sent with the same bad parameter, and ends in "Cannot compress
+    # further".  These errors are deterministic (every retry gets the identical
+    # rejection), so classify as a non-retryable format_error and fall back.
+    #
+    # NOTE: we deliberately do NOT key off the generic ``invalid_request_error``
+    # code here — OpenAI stamps that same code on genuine context-overflow 400s,
+    # so matching it would mis-route real overflows away from compression. The
+    # unambiguous signals are the explicit "unsupported/unknown parameter"
+    # message text and the specific parameter-level error codes.
+    if (
+        any(p in error_msg for p in _REQUEST_VALIDATION_PATTERNS
+            if p != "invalid_request_error")
+        or error_code_lower in {"unknown_parameter", "unsupported_parameter"}
+    ):
+        return result_fn(
+            FailoverReason.format_error,
+            retryable=False,
+            should_fallback=True,
+        )
+
     # Context overflow from 400
     if any(p in error_msg for p in _CONTEXT_OVERFLOW_PATTERNS):
         return result_fn(

agent/errors.py

@@ -0,0 +1,3 @@
+class SSLConfigurationError(Exception):
+    """Raised when SSL/TLS certificate bundle configuration fails."""
+    pass

agent/file_safety.py

@@ -46,11 +46,6 @@ def build_write_denied_paths(home: str) -> set[str]:
             # Top-level Anthropic PKCE credential store remains sensitive even
             # when a profile is active; default/non-profile sessions still read it.
             str(hermes_root / ".anthropic_oauth.json"),
-            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"),
@@ -104,12 +99,6 @@ def is_write_denied(path: str) -> bool:
         if resolved.startswith(prefix):
             return True
 
-    # Hermes control-plane files: block both the ACTIVE profile's view
-    # (hermes_home) AND the global root view. Without the root pass, a
-    # profile-mode session leaves <root>/auth.json + <root>/config.yaml
-    # writable — letting a prompt-injected write_file overwrite the global
-    # files that every profile inherits from (same shape as #15981).
-    control_file_names = ("auth.json", "config.yaml", "webhook_subscriptions.json")
     mcp_tokens_dir_name = "mcp-tokens"
 
     hermes_dirs = []
@@ -122,12 +111,6 @@ def is_write_denied(path: str) -> bool:
             continue
 
     for base_real in hermes_dirs:
-        for name in control_file_names:
-            try:
-                if resolved == os.path.realpath(os.path.join(base_real, name)):
-                    return True
-            except Exception:
-                continue
         try:
             mcp_real = os.path.realpath(os.path.join(base_real, mcp_tokens_dir_name))
             if resolved == mcp_real or resolved.startswith(mcp_real + os.sep):

agent/gemini_native_adapter.py

@@ -41,6 +41,16 @@ DEFAULT_GEMINI_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"
 GEMINI_DEFAULT_MAX_OUTPUT_TOKENS = 65535
 
 
+def bare_gemini_model_id(model: str) -> str:
+    """Strip Gemini's own provider prefix from an aggregator-style model id."""
+    name = (model or "").strip()
+    lowered = name.lower()
+    for prefix in ("google/", "gemini/"):
+        if lowered.startswith(prefix):
+            return name[len(prefix):].strip() or name
+    return name
+
+
 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()
@@ -330,7 +340,7 @@ def _build_gemini_contents(messages: List[Dict[str, Any]]) -> tuple[List[Dict[st
     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}]}
+        system_instruction = {"role": "system", "parts": [{"text": joined_system}]}
     return contents, system_instruction
 
 
@@ -914,6 +924,7 @@ class GeminiNativeClient:
             thinking_config=thinking_config,
         )
 
+        model = bare_gemini_model_id(model)
         if stream:
             return self._stream_completion(model=model, request=request, timeout=timeout)
 

agent/lsp/install.py

@@ -262,6 +262,7 @@ def _install_npm(
             capture_output=True,
             text=True,
             timeout=300,
+            stdin=subprocess.DEVNULL,
         )
         if proc.returncode != 0:
             logger.warning(
@@ -310,6 +311,7 @@ def _install_go(pkg: str, bin_name: str) -> Optional[str]:
             text=True,
             timeout=600,
             env=env,
+            stdin=subprocess.DEVNULL,
         )
         if proc.returncode != 0:
             logger.warning(
@@ -347,6 +349,7 @@ def _install_pip(pkg: str, bin_name: str) -> Optional[str]:
             capture_output=True,
             text=True,
             timeout=300,
+            stdin=subprocess.DEVNULL,
         )
         if proc.returncode != 0:
             logger.warning(

agent/memory_manager.py

@@ -33,6 +33,7 @@ from concurrent.futures import ThreadPoolExecutor
 from typing import Any, Dict, List, Optional
 
 from agent.memory_provider import MemoryProvider
+from agent.skill_commands import extract_user_instruction_from_skill_message
 from tools.registry import tool_error
 
 logger = logging.getLogger(__name__)
@@ -44,6 +45,66 @@ logger = logging.getLogger(__name__)
 _SYNC_DRAIN_TIMEOUT_S = 5.0
 
 
+def memory_provider_tools_enabled(enabled_toolsets: Optional[List[str]]) -> bool:
+    """Return whether external memory-provider tools should be exposed."""
+    if enabled_toolsets is None:
+        return True
+    if not enabled_toolsets:
+        return False
+    if "memory" in enabled_toolsets:
+        return True
+
+    try:
+        from toolsets import resolve_toolset
+
+        return any("memory" in resolve_toolset(name) for name in enabled_toolsets)
+    except Exception:
+        logger.debug("Failed to resolve enabled toolsets for memory-provider tools", exc_info=True)
+        return False
+
+
+def inject_memory_provider_tools(agent: Any) -> int:
+    """Append external memory-provider tool schemas to an agent tool surface."""
+    memory_manager = getattr(agent, "_memory_manager", None)
+    tools = getattr(agent, "tools", None)
+    if not memory_manager or tools is None:
+        return 0
+
+    existing_tool_names = {
+        tool.get("function", {}).get("name")
+        for tool in tools
+        if isinstance(tool, dict)
+    }
+    if (
+        "memory" not in existing_tool_names
+        and not memory_provider_tools_enabled(getattr(agent, "enabled_toolsets", None))
+    ):
+        return 0
+
+    get_schemas = getattr(memory_manager, "get_all_tool_schemas", None)
+    if not callable(get_schemas):
+        return 0
+
+    valid_tool_names = getattr(agent, "valid_tool_names", None)
+    if valid_tool_names is None:
+        valid_tool_names = set()
+        agent.valid_tool_names = valid_tool_names
+
+    added = 0
+    for schema in get_schemas():
+        if not isinstance(schema, dict):
+            continue
+        tool_name = schema.get("name", "")
+        if not tool_name or tool_name in existing_tool_names:
+            continue
+        tools.append({"type": "function", "function": schema})
+        valid_tool_names.add(tool_name)
+        existing_tool_names.add(tool_name)
+        added += 1
+
+    return added
+
+
 # ---------------------------------------------------------------------------
 # Context fencing helpers
 # ---------------------------------------------------------------------------
@@ -370,16 +431,37 @@ class MemoryManager:
 
     # -- Prefetch / recall ---------------------------------------------------
 
+    @staticmethod
+    def _strip_skill_scaffolding(text: str) -> Optional[str]:
+        """Return memory-worthy user text, or None to skip the turn.
+
+        When a user invokes a /skill or /bundle, Hermes expands the turn into
+        a model-facing message that embeds the entire skill body. Feeding that
+        verbatim to memory providers pollutes their stores/embeddings with
+        prompt scaffolding instead of what the user actually asked. We recover
+        just the user's instruction here, once, for every provider — so this
+        is fixed for the whole provider fan-out, not per backend.
+
+        - Non-skill messages pass through unchanged.
+        - Skill turns with a user instruction return that instruction.
+        - Bare skill invocations (no instruction) return None → callers skip
+          the turn, since there is no user content worth remembering.
+        """
+        return extract_user_instruction_from_skill_message(text)
+
     def prefetch_all(self, query: str, *, session_id: str = "") -> str:
         """Collect prefetch context from all providers.
 
         Returns merged context text labeled by provider. Empty providers
         are skipped. Failures in one provider don't block others.
         """
+        clean_query = self._strip_skill_scaffolding(query)
+        if not clean_query:
+            return ""
         parts = []
         for provider in self._providers:
             try:
-                result = provider.prefetch(query, session_id=session_id)
+                result = provider.prefetch(clean_query, session_id=session_id)
                 if result and result.strip():
                     parts.append(result)
             except Exception as e:
@@ -400,10 +482,14 @@ class MemoryManager:
         if not providers:
             return
 
+        clean_query = self._strip_skill_scaffolding(query)
+        if not clean_query:
+            return
+
         def _run() -> None:
             for provider in providers:
                 try:
-                    provider.queue_prefetch(query, session_id=session_id)
+                    provider.queue_prefetch(clean_query, session_id=session_id)
                 except Exception as e:
                     logger.debug(
                         "Memory provider '%s' queue_prefetch failed (non-fatal): %s",
@@ -455,6 +541,11 @@ class MemoryManager:
         if not providers:
             return
 
+        clean_user_content = self._strip_skill_scaffolding(user_content)
+        if not clean_user_content:
+            return
+        user_content = clean_user_content
+
         def _run() -> None:
             for provider in providers:
                 try:

agent/model_metadata.py

@@ -5,6 +5,7 @@ and run_agent.py for pre-flight context checks.
 """
 
 import ipaddress
+import json
 import logging
 import os
 import re
@@ -16,7 +17,7 @@ from urllib.parse import urlparse
 import requests
 import yaml
 
-from utils import base_url_host_matches, base_url_hostname
+from utils import atomic_json_write, base_url_host_matches, base_url_hostname
 
 from hermes_constants import OPENROUTER_MODELS_URL
 
@@ -111,6 +112,57 @@ _endpoint_model_metadata_cache: Dict[str, Dict[str, Dict[str, Any]]] = {}
 _endpoint_model_metadata_cache_time: Dict[str, float] = {}
 _ENDPOINT_MODEL_CACHE_TTL = 300
 
+
+def _get_model_metadata_cache_path() -> Path:
+    """Return path to the OpenRouter model metadata disk cache."""
+    from hermes_constants import get_hermes_home
+    return get_hermes_home() / "cache" / "openrouter_model_metadata.json"
+
+
+def _model_metadata_disk_cache_age_seconds() -> Optional[float]:
+    """Return disk-cache age in seconds, or None if freshness is unknown."""
+    try:
+        cache_path = _get_model_metadata_cache_path()
+        if not cache_path.exists():
+            return None
+        age = time.time() - cache_path.stat().st_mtime
+        if age < 0:
+            return None
+        return age
+    except Exception:
+        return None
+
+
+def _load_model_metadata_disk_cache() -> Dict[str, Dict[str, Any]]:
+    """Load processed OpenRouter metadata cache from disk."""
+    try:
+        cache_path = _get_model_metadata_cache_path()
+        with cache_path.open("r", encoding="utf-8") as f:
+            data = json.load(f)
+        if not isinstance(data, dict):
+            return {}
+        return {
+            str(key): value
+            for key, value in data.items()
+            if isinstance(value, dict)
+        }
+    except Exception as e:
+        logger.debug("Failed to load OpenRouter model metadata disk cache: %s", e)
+        return {}
+
+
+def _save_model_metadata_disk_cache(data: Dict[str, Dict[str, Any]]) -> None:
+    """Save processed OpenRouter metadata cache to disk atomically."""
+    try:
+        atomic_json_write(
+            _get_model_metadata_cache_path(),
+            data,
+            indent=0,
+            separators=(",", ":"),
+        )
+    except Exception as e:
+        logger.debug("Failed to save OpenRouter model metadata disk cache: %s", e)
+
 # Descending tiers for context length probing when the model is unknown.
 # We start at 256K (covers GPT-5.x, many current large-context models) and
 # step down on context-length errors until one works.  Tier[0] is also the
@@ -141,6 +193,8 @@ DEFAULT_CONTEXT_LENGTHS = {
     # fuzzy-match collisions (e.g. "anthropic/claude-sonnet-4" is a
     # substring of "anthropic/claude-sonnet-4.6").
     # OpenRouter-prefixed models resolve via OpenRouter live API or models.dev.
+    "claude-fable-5": 1000000,
+    "claude-fable": 1000000,
     "claude-opus-4-8": 1000000,
     "claude-opus-4.8": 1000000,
     "claude-opus-4-7": 1000000,
@@ -207,7 +261,13 @@ DEFAULT_CONTEXT_LENGTHS = {
     # https://platform.minimax.io/docs/api-reference/text-chat-openai
     "minimax-m3": 1000000,
     "minimax": 204800,
-    # GLM
+    # GLM — GLM-5.2 ships with a 1M context window (verified empirically:
+    # needle-in-a-haystack retrieval at 789K prompt tokens succeeded with
+    # zero errors on api.z.ai/api/coding/paas/v4).  Older GLM models
+    # (5, 5.1, 5-turbo) are ~202K.  Longest-key-first substring matching
+    # ensures "glm-5.2" resolves to 1M while older variants still hit the
+    # generic 202K fallback.
+    "glm-5.2": 1_048_576,
     "glm": 202752,
     # xAI Grok — xAI /v1/models does not return context_length metadata,
     # so these hardcoded fallbacks prevent Hermes from probing-down to
@@ -625,6 +685,15 @@ def fetch_model_metadata(force_refresh: bool = False) -> Dict[str, Dict[str, Any
     if not force_refresh and _model_metadata_cache and (time.time() - _model_metadata_cache_time) < _MODEL_CACHE_TTL:
         return _model_metadata_cache
 
+    if not force_refresh:
+        disk_age = _model_metadata_disk_cache_age_seconds()
+        if disk_age is not None and disk_age < _MODEL_CACHE_TTL:
+            disk_cache = _load_model_metadata_disk_cache()
+            if disk_cache:
+                _model_metadata_cache = disk_cache
+                _model_metadata_cache_time = time.time() - disk_age
+                return _model_metadata_cache
+
     try:
         response = requests.get(OPENROUTER_MODELS_URL, timeout=10, verify=_resolve_requests_verify())
         response.raise_for_status()
@@ -646,12 +715,24 @@ def fetch_model_metadata(force_refresh: bool = False) -> Dict[str, Dict[str, Any
 
         _model_metadata_cache = cache
         _model_metadata_cache_time = time.time()
+        _save_model_metadata_disk_cache(cache)
         logger.debug("Fetched metadata for %s models from OpenRouter", len(cache))
         return cache
 
     except Exception as e:
         logger.warning(f"Failed to fetch model metadata from OpenRouter: {e}")
-        return _model_metadata_cache or {}
+        if _model_metadata_cache:
+            return _model_metadata_cache
+        disk_cache = _load_model_metadata_disk_cache()
+        if disk_cache:
+            _model_metadata_cache = disk_cache
+            disk_age = _model_metadata_disk_cache_age_seconds()
+            if disk_age is not None:
+                _model_metadata_cache_time = time.time() - min(disk_age, _MODEL_CACHE_TTL)
+            else:
+                _model_metadata_cache_time = time.time() - _MODEL_CACHE_TTL + 1
+            return _model_metadata_cache
+        return {}
 
 
 def fetch_endpoint_model_metadata(
@@ -968,6 +1049,16 @@ def parse_available_output_tokens_from_error(error_msg: str) -> Optional[int]:
         # OpenRouter/Nous phrasing of the same condition.
         "in the output" in error_lower
         and "maximum context length" in error_lower
+    ) or (
+        # LM Studio / llama.cpp / some OpenAI-compatible servers:
+        #   "This model's maximum context length is 65536 tokens. However, you
+        #    requested 65536 output tokens and your prompt contains 77409
+        #    characters ..."
+        # The "requested N output tokens" phrasing means the OUTPUT cap is the
+        # problem (the input itself fits) — reduce max_tokens, don't compress.
+        "maximum context length" in error_lower
+        and "requested" in error_lower
+        and "output tokens" in error_lower
     )
     if not is_output_cap_error:
         return None
@@ -999,6 +1090,22 @@ def parse_available_output_tokens_from_error(error_msg: str) -> Optional[int]:
         if _available >= 1:
             return _available
 
+    # LM Studio / llama.cpp style: context window is reported in tokens but the
+    # prompt size is reported in CHARACTERS, e.g.
+    #   "maximum context length is 65536 tokens ... your prompt contains 77409
+    #    characters ...".
+    # Estimate the input tokens conservatively (~3 chars/token, which
+    # over-reserves the input so the retried output cap stays safely inside the
+    # window) and leave the remainder of the window for output.
+    _m_ctx_tok = re.search(r'maximum context length is (\d+)\s*token', error_lower)
+    _m_chars = re.search(r'prompt contains (\d+)\s*character', error_lower)
+    if _m_ctx_tok and _m_chars:
+        _ctx = int(_m_ctx_tok.group(1))
+        _est_input = (int(_m_chars.group(1)) + 2) // 3
+        _available = _ctx - _est_input
+        if _available >= 1:
+            return _available
+
     return None
 
 
@@ -1784,10 +1891,43 @@ def get_model_context_length(
         if ctx is not None:
             save_context_length(model, base_url, ctx)
             return ctx
+    # 5f. OpenRouter live /models metadata — authoritative for OpenRouter-routed
+    # models. OpenRouter's catalog carries per-model context_length (e.g.
+    # anthropic/claude-fable-5 -> 1M) and refreshes as new slugs ship, so it
+    # must win over both models.dev (step 5g) and the hardcoded family catch-all
+    # (step 8). Before this branch, an OpenRouter selection set
+    # effective_provider="openrouter", which (a) made the models.dev lookup miss
+    # brand-new slugs and (b) skipped the step-6 OR fallback (gated on `not
+    # effective_provider`), so a fresh slug like claude-fable-5 fell through to
+    # the generic "claude": 200K entry and under-reported a 1M window. Mirrors
+    # the dedicated Nous/Copilot/GMI branches above.
+    if effective_provider == "openrouter":
+        metadata = fetch_model_metadata()
+        entry = metadata.get(model)
+        if entry:
+            or_ctx = entry.get("context_length")
+            # Guard against the known OpenRouter Kimi-family 32k underreport
+            # (same class the hardcoded overrides exist to mitigate).
+            if isinstance(or_ctx, int) and or_ctx > 0 and not (
+                or_ctx == 32768 and _model_name_suggests_kimi(model)
+            ):
+                return or_ctx
+
     if effective_provider:
         from agent.models_dev import lookup_models_dev_context
         ctx = lookup_models_dev_context(effective_provider, model)
         if ctx:
+            # MiniMax M3: models.dev reports 512K but actual context is 1M.
+            # Prefer hardcoded catalog over stale probe value.
+            if _model_name_suggests_minimax_m3(model):
+                catalog = DEFAULT_CONTEXT_LENGTHS.get("minimax-m3")
+                if catalog and ctx < catalog:
+                    logger.info(
+                        "Rejecting models.dev context=%s for %r "
+                        "(MiniMax-M3 underreport); using hardcoded default %s",
+                        ctx, model, f"{catalog:,}",
+                    )
+                    ctx = catalog
             return ctx
 
     # 6. OpenRouter live API metadata — provider-unaware fallback.

agent/moonshot_schema.py

@@ -135,7 +135,14 @@ def _repair_schema(node: Any, is_schema: bool = True) -> Any:
 
 def _fill_missing_type(node: Dict[str, Any]) -> Dict[str, Any]:
     """Infer a reasonable ``type`` if this schema node has none."""
-    if "type" in node and node["type"] not in {None, ""}:
+    node_type = node.get("type")
+    if isinstance(node_type, list):
+        concrete = next(
+            (t for t in node_type if isinstance(t, str) and t not in {"", "null"}),
+            "string",
+        )
+        return {**node, "type": concrete}
+    if "type" in node and node_type not in {None, ""}:
         return node
 
     # Heuristic: presence of ``properties`` → object, ``items`` → array, ``enum``

agent/pet/__init__.py

@@ -0,0 +1,51 @@
+"""Petdex pet engine — shared core for the CLI, TUI, and desktop surfaces.
+
+Petdex (https://github.com/crafter-station/petdex) is a public gallery of
+animated sprite "pets" for coding agents.  Each pet is a ``pet.json`` plus a
+``spritesheet.{webp,png}`` of 192×208 px cells. Current Codex/petdex sheets use
+an 8-column × 9-row atlas; older Hermes/petdex sheets used an 8-row atlas.
+Hermes infers the row taxonomy from the sheet and maps agent activity onto
+idle/run/review/failed/wave/jump.
+
+This package is the **single source of truth** for the feature so the base
+CLI (Python) and TUI (Ink, via ``tui_gateway``) never duplicate the hard
+parts:
+
+- :mod:`agent.pet.constants` — frame geometry + the :class:`PetState` enum.
+- :mod:`agent.pet.state`     — map agent activity → a :class:`PetState`.
+- :mod:`agent.pet.manifest`  — fetch the public petdex manifest.
+- :mod:`agent.pet.store`     — install / list / resolve pets on disk
+                               (profile-aware via ``get_hermes_home()``).
+- :mod:`agent.pet.render`    — decode a spritesheet and encode frames for a
+                               terminal (kitty / iTerm2 / sixel graphics
+                               protocols, with a Unicode half-block
+                               fallback).
+
+Rendering in the Electron desktop is necessarily TypeScript (canvas), but it
+reuses the same on-disk store and the same state semantics.
+
+The whole feature is a *display* concern: it adds no model tool, mutates no
+system prompt or toolset, and therefore has zero effect on prompt caching.
+"""
+
+from agent.pet.constants import (
+    DEFAULT_SCALE,
+    FRAME_H,
+    FRAME_W,
+    FRAMES_PER_STATE,
+    LOOP_MS,
+    STATE_ROWS,
+    PetState,
+)
+from agent.pet.state import derive_pet_state
+
+__all__ = [
+    "DEFAULT_SCALE",
+    "FRAME_H",
+    "FRAME_W",
+    "FRAMES_PER_STATE",
+    "LOOP_MS",
+    "STATE_ROWS",
+    "PetState",
+    "derive_pet_state",
+]

agent/pet/constants.py

@@ -0,0 +1,167 @@
+"""Pet sprite geometry + animation-state taxonomy.
+
+These values are the common petdex/Codex pet geometry. The real ``pet.json``
+usually only carries ``id``/``displayName``/``description``/``spritesheetPath``;
+row taxonomy is inferred from the atlas shape so Hermes can render both legacy
+8-row sheets and current 9-row Codex sheets.
+"""
+
+from __future__ import annotations
+
+from enum import Enum
+
+# Frame geometry (pixels). Current Codex/petdex spritesheets are 8 columns x 9
+# rows (1536x1872), while older Hermes/petdex sheets used 9 columns x 8 rows
+# (1728x1664). Renderers derive both row taxonomy and real column count from the
+# concrete sheet, so either shape works.
+FRAME_W = 192
+FRAME_H = 208
+
+# Frames consumed per animation state (the petdex web app uses CSS
+# ``steps(6)``).  A sheet may physically contain more columns; we only step
+# through the first ``FRAMES_PER_STATE``.
+FRAMES_PER_STATE = 6
+
+# Full-loop duration for one state, milliseconds (petdex default).
+LOOP_MS = 1100
+
+# Default on-screen scale relative to native frame size.  ``display.pet.scale``
+# is the single master scalar: the desktop canvas multiplies its native pixels
+# by it and every terminal surface derives its half-block/kitty column width
+# from it (see :func:`cols_for_scale`), so one number shrinks all three
+# interfaces together.  (petdex's own clients render at 0.7; we default smaller
+# so the kitty/GUI mascot stays a glanceable corner sprite.  The half-block
+# fallback can't shrink as far — see ``UNICODE_MIN_COLS`` — and clamps to its
+# legibility floor instead.)
+DEFAULT_SCALE = 0.33
+
+# User-settable scale bounds (``/pet scale``, desktop slider).  Floor keeps the
+# pet clickable/visible; ceiling stops a fat-fingered value from filling the
+# screen.  The unicode fallback additionally clamps to ``UNICODE_MIN_COLS``.
+MIN_SCALE = 0.1
+MAX_SCALE = 3.0
+
+
+def clamp_scale(scale: float) -> float:
+    """Clamp *scale* to ``[MIN_SCALE, MAX_SCALE]`` (the single validation point)."""
+    return max(MIN_SCALE, min(MAX_SCALE, scale))
+
+# Terminal cells one native frame spans at ``scale == 1.0``.  A cell is ~8px
+# wide, a frame is ``FRAME_W`` (192) px → 24 cells.  This mirrors the kitty
+# graphics placement (``scaled_px // 8``) so at full scale every renderer agrees.
+BASE_UNICODE_COLS = FRAME_W // 8
+
+# Legibility floor for the half-block fallback.  A half-block cell samples the
+# sprite at only 1 horizontal + 2 vertical taps, so below this width a 192×208
+# pet collapses into an unreadable blob *regardless* of scale.  kitty/GUI draw
+# true pixels and have no such floor — that's why the same ``scale: 0.33`` is
+# crisp there but mush in half-blocks.  ``scale`` shrinks the unicode pet down
+# TO this floor (and grows it above), instead of past it into noise.
+UNICODE_MIN_COLS = 16
+
+
+def cols_for_scale(scale: float) -> int:
+    """Half-block width implied by *scale*, clamped to the legibility floor.
+
+    Above the floor it tracks the kitty cell box (``scaled_px // 8``) so the two
+    renderers converge at larger sizes; below it the floor keeps the sprite
+    readable rather than letting it devolve into a blob.
+    """
+    return max(UNICODE_MIN_COLS, round(BASE_UNICODE_COLS * (scale or DEFAULT_SCALE)))
+
+
+def resolve_cols(scale: float, unicode_cols: int = 0) -> int:
+    """Resolve terminal width: explicit *unicode_cols* override, else from *scale*."""
+    return int(unicode_cols) if unicode_cols and int(unicode_cols) > 0 else cols_for_scale(scale)
+
+
+class PetState(str, Enum):
+    """Animation state a pet can be shown in.
+
+    These are Hermes' activity state names. They are not always identical to the
+    source atlas row names: Codex-format pets use rows like ``jumping`` /
+    ``running`` while the UI keeps the shorter ``jump`` / ``run`` names.
+    """
+
+    IDLE = "idle"
+    WAVE = "wave"
+    RUN = "run"
+    FAILED = "failed"
+    REVIEW = "review"
+    JUMP = "jump"
+    WAITING = "waiting"
+
+
+# Legacy Hermes/petdex row order (top -> bottom) used by the older 8-row,
+# 9-column atlas shape.
+LEGACY_STATE_ROWS: list[str] = [
+    PetState.IDLE.value,
+    PetState.WAVE.value,
+    PetState.RUN.value,
+    PetState.FAILED.value,
+    PetState.REVIEW.value,
+    PetState.JUMP.value,
+    "extra1",
+    "extra2",
+]
+
+# Current Petdex row order (top -> bottom) used by 1536x1872 atlases:
+# 8 columns x 9 rows of 192x208 cells.
+CODEX_STATE_ROWS: list[str] = [
+    PetState.IDLE.value,
+    "running-right",
+    "running-left",
+    "waving",
+    "jumping",
+    PetState.FAILED.value,
+    PetState.WAITING.value,
+    "running",
+    PetState.REVIEW.value,
+]
+
+# Default/fallback for callers without a sheet. Prefer the current 9-row Codex
+# format because generated pets and the public Codex pet contract use it.
+STATE_ROWS: list[str] = CODEX_STATE_ROWS
+
+# Canonical Hermes activity names -> accepted row-name aliases in descending
+# preference. This keeps our internal state names stable (`wave`/`jump`/`run`)
+# while matching Petdex's current `waving`/`jumping`/`running` taxonomy.
+STATE_ALIASES: dict[str, tuple[str, ...]] = {
+    PetState.IDLE.value: (PetState.IDLE.value,),
+    PetState.WAVE.value: (PetState.WAVE.value, "waving"),
+    PetState.JUMP.value: (PetState.JUMP.value, "jumping"),
+    PetState.RUN.value: (PetState.RUN.value, "running"),
+    PetState.FAILED.value: (PetState.FAILED.value,),
+    PetState.REVIEW.value: (PetState.REVIEW.value,),
+    PetState.WAITING.value: (PetState.WAITING.value,),
+}
+
+
+def state_aliases_for(state: "PetState | str") -> tuple[str, ...]:
+    """Return accepted row-name aliases for *state* (always non-empty)."""
+    value = state.value if isinstance(state, PetState) else str(state)
+    aliases = STATE_ALIASES.get(value)
+    return aliases if aliases else (value,)
+
+
+def state_rows_for_grid(row_count: int | None) -> list[str]:
+    """Return the row taxonomy for a spritesheet with *row_count* rows."""
+    try:
+        rows = int(row_count or 0)
+    except (TypeError, ValueError):
+        rows = 0
+
+    if rows >= len(CODEX_STATE_ROWS):
+        return CODEX_STATE_ROWS
+    return LEGACY_STATE_ROWS
+
+
+def state_row_index(state: "PetState | str", row_count: int | None = None) -> int:
+    """Return the spritesheet row index for *state* (clamped, never raises)."""
+    rows = state_rows_for_grid(row_count)
+    for name in state_aliases_for(state):
+        try:
+            return rows.index(name)
+        except ValueError:
+            continue
+    return 0  # fall back to the idle row

agent/pet/generate/__init__.py

@@ -0,0 +1,29 @@
+"""Pet generation — base-draft → hatch pipeline.
+
+Public surface used by the gateway RPCs, the CLI ``hermes pets generate``
+command, and tests:
+
+- :func:`generate_base_drafts` / :func:`hatch_pet` — the two-step flow.
+- :class:`HatchResult`, :class:`GenerationError`.
+- :mod:`atlas` — deterministic frame extraction + atlas composition/validation.
+
+Image generation is delegated to the active reference-capable
+:class:`~agent.image_gen_provider.ImageGenProvider` (OpenAI gpt-image-2 or Krea);
+atlas assembly is fully deterministic so it's testable without any API calls.
+"""
+
+from __future__ import annotations
+
+from agent.pet.generate.imagegen import GenerationError
+from agent.pet.generate.orchestrate import (
+    HatchResult,
+    generate_base_drafts,
+    hatch_pet,
+)
+
+__all__ = [
+    "GenerationError",
+    "HatchResult",
+    "generate_base_drafts",
+    "hatch_pet",
+]

agent/pet/generate/atlas.py

@@ -0,0 +1,400 @@
+"""Deterministic spritesheet assembly — generated row strips → Hermes atlas.
+
+Image-generation models are good at *drawing* a row of poses but bad at exact
+grid geometry, so the model never owns the atlas layout: it produces one loose
+horizontal strip per state, and these deterministic ops slice that strip into
+clean, centered, transparent ``192x208`` cells and pack them into the sheet our
+renderer reads.
+
+The atlas is **Hermes-native**, not the petdex/Codex format. Our renderer
+(:mod:`agent.pet.render`) keys frames as ``rows = states, cols = frames`` using
+:data:`agent.pet.constants.STATE_ROWS`, so we emit exactly the six states the
+engine drives — idle, wave, run, failed, review, jump — left-packed with
+trailing transparent cells (which the renderer trims). Sheet is
+``COLUMNS*192 x ROWS*208`` (1152x1248).
+
+The frame-segmentation, fit-to-cell, and transparency-residue logic is adapted
+from OpenAI's ``hatch-pet`` skill (openai/skills, Apache-2.0).
+"""
+
+from __future__ import annotations
+
+import io
+import logging
+import math
+from pathlib import Path
+
+from agent.pet.constants import FRAME_H, FRAME_W
+
+logger = logging.getLogger(__name__)
+
+CELL_WIDTH = FRAME_W
+CELL_HEIGHT = FRAME_H
+
+# (state, row index, frame count). Order/row indices MUST match
+# ``STATE_ROWS`` so the renderer crops the right row for each driven state.
+# Frame counts are the petdex-ish per-state lengths; the renderer trims any
+# trailing blank columns, so rows shorter than ``COLUMNS`` just leave the tail
+# transparent.
+ROW_SPECS: list[tuple[str, int, int]] = [
+    ("idle", 0, 6),
+    ("wave", 1, 4),
+    ("run", 2, 6),
+    ("failed", 3, 6),
+    ("review", 4, 6),
+    ("jump", 5, 5),
+]
+
+ROWS = len(ROW_SPECS)
+COLUMNS = max(count for _, _, count in ROW_SPECS)
+ATLAS_WIDTH = COLUMNS * CELL_WIDTH
+ATLAS_HEIGHT = ROWS * CELL_HEIGHT
+
+FRAME_COUNTS: dict[str, int] = {state: count for state, _, count in ROW_SPECS}
+
+# Alpha at/below which a pixel is "background" for component detection.
+_ALPHA_FLOOR = 16
+# Cell padding kept around a fitted sprite so poses never touch the edge.
+_CELL_PAD = 10
+
+
+# ───────────────────────── background removal ─────────────────────────
+
+
+def _color_distance(r: int, g: int, b: int, key: tuple[int, int, int]) -> float:
+    return math.sqrt((r - key[0]) ** 2 + (g - key[1]) ** 2 + (b - key[2]) ** 2)
+
+
+def _has_transparency(image) -> bool:
+    """True if the strip already carries a real alpha background."""
+    extrema = image.getchannel("A").getextrema()
+    # Min alpha 0 somewhere and a meaningful share of fully-transparent pixels.
+    if extrema[0] > _ALPHA_FLOOR:
+        return False
+    hist = image.getchannel("A").histogram()
+    transparent = sum(hist[: _ALPHA_FLOOR + 1])
+    total = image.width * image.height
+    return transparent > total * 0.05
+
+
+def _dominant_corner_color(image) -> tuple[int, int, int]:
+    """Sample the four corners and return the most common opaque color."""
+    from collections import Counter
+
+    w, h = image.width, image.height
+    px = image.load()
+    counter: Counter = Counter()
+    for x, y in ((0, 0), (w - 1, 0), (0, h - 1), (w - 1, h - 1)):
+        r, g, b, a = px[x, y]
+        if a > _ALPHA_FLOOR:
+            counter[(r, g, b)] += 1
+    if not counter:
+        return (0, 255, 0)
+    return counter.most_common(1)[0][0]
+
+
+def remove_background(image, *, chroma_key: tuple[int, int, int] | None = None, threshold: float = 110.0):
+    """Return *image* (RGBA) with its flat background keyed out to transparent.
+
+    If the strip already has a transparent background we leave it alone; else we
+    key out *chroma_key* (or the dominant corner color when not given). This
+    handles both providers that emit transparency natively and those that paint
+    a solid backdrop.
+    """
+    rgba = image.convert("RGBA")
+    if _has_transparency(rgba):
+        return rgba
+
+    key = chroma_key or _dominant_corner_color(rgba)
+    px = rgba.load()
+    for y in range(rgba.height):
+        for x in range(rgba.width):
+            r, g, b, a = px[x, y]
+            if a > _ALPHA_FLOOR and _color_distance(r, g, b, key) <= threshold:
+                px[x, y] = (0, 0, 0, 0)
+    return rgba
+
+
+# ───────────────────────── frame extraction ─────────────────────────
+
+
+def _fit_to_cell(image):
+    """Crop to content, scale to fit a padded cell, and center on transparent."""
+    from PIL import Image
+
+    target = Image.new("RGBA", (CELL_WIDTH, CELL_HEIGHT), (0, 0, 0, 0))
+    bbox = image.getbbox()
+    if bbox is None:
+        return target
+
+    sprite = image.crop(bbox)
+    max_w = CELL_WIDTH - _CELL_PAD
+    max_h = CELL_HEIGHT - _CELL_PAD
+    scale = min(max_w / sprite.width, max_h / sprite.height, 1.0)
+    if scale != 1.0:
+        sprite = sprite.resize(
+            (max(1, round(sprite.width * scale)), max(1, round(sprite.height * scale))),
+            Image.Resampling.LANCZOS,
+        )
+    left = (CELL_WIDTH - sprite.width) // 2
+    top = (CELL_HEIGHT - sprite.height) // 2
+    target.alpha_composite(sprite, (left, top))
+    return target
+
+
+def _connected_components(image) -> list[dict]:
+    """Flood-fill the alpha mask into connected blobs (4-connectivity)."""
+    alpha = image.getchannel("A")
+    w, h = image.size
+    data = alpha.tobytes()
+    visited = bytearray(w * h)
+    out: list[dict] = []
+
+    for start, a in enumerate(data):
+        if a <= _ALPHA_FLOOR or visited[start]:
+            continue
+        stack = [start]
+        visited[start] = 1
+        pixels: list[int] = []
+        min_x = w
+        min_y = h
+        max_x = 0
+        max_y = 0
+        while stack:
+            cur = stack.pop()
+            pixels.append(cur)
+            x = cur % w
+            y = cur // w
+            min_x = min(min_x, x)
+            min_y = min(min_y, y)
+            max_x = max(max_x, x)
+            max_y = max(max_y, y)
+            for nb, ok in (
+                (cur - 1, x > 0),
+                (cur + 1, x + 1 < w),
+                (cur - w, y > 0),
+                (cur + w, y + 1 < h),
+            ):
+                if ok and not visited[nb] and data[nb] > _ALPHA_FLOOR:
+                    visited[nb] = 1
+                    stack.append(nb)
+        out.append(
+            {
+                "pixels": pixels,
+                "area": len(pixels),
+                "bbox": (min_x, min_y, max_x + 1, max_y + 1),
+                "center_x": (min_x + max_x + 1) / 2,
+            }
+        )
+    return out
+
+
+def _group_image(source, components: list[dict], padding: int = 4):
+    from PIL import Image
+
+    w, h = source.size
+    min_x = max(0, min(c["bbox"][0] for c in components) - padding)
+    min_y = max(0, min(c["bbox"][1] for c in components) - padding)
+    max_x = min(w, max(c["bbox"][2] for c in components) + padding)
+    max_y = min(h, max(c["bbox"][3] for c in components) + padding)
+
+    out = Image.new("RGBA", (max_x - min_x, max_y - min_y), (0, 0, 0, 0))
+    src_px = source.load()
+    out_px = out.load()
+    for c in components:
+        for idx in c["pixels"]:
+            x = idx % w
+            y = idx // w
+            out_px[x - min_x, y - min_y] = src_px[x, y]
+    return out
+
+
+def _component_frames(strip, frame_count: int) -> list | None:
+    """Segment a strip into *frame_count* sprites by connected components.
+
+    Picks the ``frame_count`` largest blobs as seeds (left→right), attaches
+    smaller blobs to the nearest seed, and returns one fitted cell per group.
+    Returns ``None`` when it can't find enough distinct sprites (caller falls
+    back to equal slicing).
+    """
+    components = _connected_components(strip)
+    if not components:
+        return None
+
+    largest = max(c["area"] for c in components)
+    seed_threshold = max(120, largest * 0.20)
+    seeds = [c for c in components if c["area"] >= seed_threshold]
+    if len(seeds) < frame_count:
+        seeds = sorted(components, key=lambda c: c["area"], reverse=True)[:frame_count]
+    if len(seeds) < frame_count:
+        return None
+
+    seeds = sorted(
+        sorted(seeds, key=lambda c: c["area"], reverse=True)[:frame_count],
+        key=lambda c: c["center_x"],
+    )
+    seed_ids = {id(s) for s in seeds}
+    groups: list[list[dict]] = [[s] for s in seeds]
+    noise_threshold = max(12, largest * 0.002)
+    for c in components:
+        if id(c) in seed_ids or c["area"] < noise_threshold:
+            continue
+        nearest = min(range(len(seeds)), key=lambda i: abs(seeds[i]["center_x"] - c["center_x"]))
+        groups[nearest].append(c)
+
+    return [_fit_to_cell(_group_image(strip, g)) for g in groups]
+
+
+def _slot_frames(strip, frame_count: int) -> list:
+    """Fallback: slice the strip into *frame_count* equal columns."""
+    slot = strip.width / frame_count
+    frames = []
+    for i in range(frame_count):
+        left = round(i * slot)
+        right = round((i + 1) * slot)
+        frames.append(_fit_to_cell(strip.crop((left, 0, right, strip.height))))
+    return frames
+
+
+def extract_strip_frames(
+    strip,
+    frame_count: int,
+    *,
+    chroma_key: tuple[int, int, int] | None = None,
+    method: str = "auto",
+) -> list:
+    """Turn one generated row strip into *frame_count* clean 192x208 cells.
+
+    *strip* is a PIL image (or path). Background is keyed out, then frames are
+    found by connected components (``auto``) with an equal-slot fallback.
+    """
+    from PIL import Image
+
+    if isinstance(strip, (str, Path)):
+        with Image.open(strip) as opened:
+            strip = opened.convert("RGBA")
+    else:
+        strip = strip.convert("RGBA")
+
+    strip = remove_background(strip, chroma_key=chroma_key)
+
+    if method in ("auto", "components"):
+        frames = _component_frames(strip, frame_count)
+        if frames is not None:
+            return frames
+        if method == "components":
+            raise ValueError(f"could not segment {frame_count} sprites from strip")
+    return _slot_frames(strip, frame_count)
+
+
+# ───────────────────────── atlas composition ─────────────────────────
+
+
+def single_frame(image):
+    """One fitted 192x208 cell from a standalone image (e.g. the base look).
+
+    Used as an idle fallback so a pet always renders even if the idle row
+    generation failed.
+    """
+    from PIL import Image
+
+    if isinstance(image, (str, Path)):
+        with Image.open(image) as opened:
+            image = opened.convert("RGBA")
+    return _fit_to_cell(remove_background(image))
+
+
+def _clear_transparent_rgb(image):
+    """Zero the RGB of fully-transparent pixels (no colored-halo residue)."""
+    from PIL import Image
+
+    rgba = image.convert("RGBA")
+    data = bytearray(rgba.tobytes())
+    for i in range(0, len(data), 4):
+        if data[i + 3] == 0:
+            data[i] = data[i + 1] = data[i + 2] = 0
+    return Image.frombytes("RGBA", rgba.size, bytes(data))
+
+
+def compose_atlas(frames_by_state: dict[str, list]):
+    """Pack per-state frame lists into the Hermes atlas (RGBA, residue-cleared).
+
+    Missing/short states leave their trailing cells transparent; extra frames
+    beyond a state's spec are dropped.
+    """
+    from PIL import Image
+
+    atlas = Image.new("RGBA", (ATLAS_WIDTH, ATLAS_HEIGHT), (0, 0, 0, 0))
+    for state, row, count in ROW_SPECS:
+        frames = frames_by_state.get(state) or []
+        for col, frame in enumerate(frames[:count]):
+            cell = frame.convert("RGBA")
+            if cell.size != (CELL_WIDTH, CELL_HEIGHT):
+                cell = _fit_to_cell(cell)
+            atlas.alpha_composite(cell, (col * CELL_WIDTH, row * CELL_HEIGHT))
+    return _clear_transparent_rgb(atlas)
+
+
+def atlas_to_webp_bytes(atlas) -> bytes:
+    """Encode an atlas image to lossless WebP bytes (the on-disk pet format)."""
+    buf = io.BytesIO()
+    atlas.save(buf, format="WEBP", lossless=True, quality=100, method=6, exact=True)
+    return buf.getvalue()
+
+
+def validate_atlas(atlas) -> dict:
+    """Check geometry, per-cell occupancy, and transparency invariants.
+
+    Returns ``{ok, width, height, errors, warnings, filled_states}``. Errors are
+    blockers (wrong size, empty used cell, opaque/dirty transparency); warnings
+    are soft (a whole state row blank — generation likely dropped a row).
+    """
+    from PIL import Image
+
+    if isinstance(atlas, (str, Path)):
+        with Image.open(atlas) as opened:
+            atlas = opened.convert("RGBA")
+    else:
+        atlas = atlas.convert("RGBA")
+
+    errors: list[str] = []
+    warnings: list[str] = []
+
+    if atlas.size != (ATLAS_WIDTH, ATLAS_HEIGHT):
+        errors.append(f"expected {ATLAS_WIDTH}x{ATLAS_HEIGHT}, got {atlas.width}x{atlas.height}")
+        return {"ok": False, "width": atlas.width, "height": atlas.height, "errors": errors, "warnings": warnings, "filled_states": []}
+
+    filled_states: list[str] = []
+    for state, row, count in ROW_SPECS:
+        row_pixels = 0
+        for col in range(count):
+            left = col * CELL_WIDTH
+            top = row * CELL_HEIGHT
+            cell = atlas.crop((left, top, left + CELL_WIDTH, top + CELL_HEIGHT))
+            nonblank = sum(cell.getchannel("A").histogram()[1:])
+            row_pixels += nonblank
+        if row_pixels > 0:
+            filled_states.append(state)
+        else:
+            warnings.append(f"state '{state}' has no frames")
+
+    if not filled_states:
+        errors.append("atlas is empty — no state produced any frames")
+
+    # Transparent pixels must carry zero RGB (no halo residue).
+    data = atlas.tobytes()
+    residue = 0
+    for i in range(0, len(data), 4):
+        if data[i + 3] == 0 and (data[i] or data[i + 1] or data[i + 2]):
+            residue += 1
+    if residue:
+        errors.append(f"{residue} transparent pixels retain RGB residue")
+
+    return {
+        "ok": not errors,
+        "width": atlas.width,
+        "height": atlas.height,
+        "errors": errors,
+        "warnings": warnings,
+        "filled_states": filled_states,
+    }

agent/pet/generate/imagegen.py

@@ -0,0 +1,168 @@
+"""Thin image-generation layer for pet sprites.
+
+Wraps the active :class:`~agent.image_gen_provider.ImageGenProvider` with the
+two things sprite generation needs that the agent-facing ``image_generate`` tool
+doesn't expose: **N variants** (loop) and **reference-image grounding** (so each
+animation row stays the same character as the chosen base).
+
+Reference grounding only works on providers that support it — currently OpenAI
+``gpt-image-2`` (image edits) and Krea (style references). We resolve to one of
+those and surface a clear, actionable error otherwise rather than silently
+producing an ungrounded, drifting pet.
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+# Providers that can ground generation on a reference image.
+_REF_CAPABLE = ("openai", "openai-codex", "krea")
+
+
+class GenerationError(RuntimeError):
+    """Raised on any image-generation failure (no provider, API error, IO)."""
+
+
+@dataclass(frozen=True)
+class SpriteProvider:
+    """Resolved provider plus whether it can take reference images."""
+
+    name: str
+    provider: object
+    supports_references: bool
+
+
+def _discover() -> None:
+    try:
+        from hermes_cli.plugins import _ensure_plugins_discovered
+
+        _ensure_plugins_discovered()
+    except Exception as exc:  # noqa: BLE001 - discovery is best-effort
+        logger.debug("image-gen plugin discovery failed: %s", exc)
+
+
+def resolve_provider(*, require_references: bool = True) -> SpriteProvider:
+    """Pick the image provider to use for sprite work.
+
+    Preference: the configured provider when it's reference-capable, else the
+    first available reference-capable provider. With *require_references* off we
+    fall back to any available provider (used for prompt-only base drafts).
+    """
+    _discover()
+    from agent.image_gen_registry import get_active_provider, get_provider
+
+    # Configured / active provider first.
+    active = None
+    try:
+        active = get_active_provider()
+    except Exception:  # noqa: BLE001
+        active = None
+    if active is not None:
+        name = getattr(active, "name", "")
+        if name in _REF_CAPABLE and active.is_available():
+            return SpriteProvider(name=name, provider=active, supports_references=True)
+
+    # Any available reference-capable provider.
+    for name in _REF_CAPABLE:
+        provider = get_provider(name)
+        if provider is not None and provider.is_available():
+            return SpriteProvider(name=name, provider=provider, supports_references=True)
+
+    if not require_references and active is not None and active.is_available():
+        return SpriteProvider(
+            name=getattr(active, "name", "unknown"), provider=active, supports_references=False
+        )
+
+    raise GenerationError(
+        "Pet generation needs a reference-capable image backend. "
+        "Run `hermes tools` → Image Generation → OpenAI (gpt-image-2) and add an "
+        "OpenAI API key (or configure Krea)."
+    )
+
+
+def _save_local(image_ref: str, *, prefix: str) -> Path:
+    """Return a local path for *image_ref*, downloading it if it's a URL."""
+    if image_ref.startswith(("http://", "https://")):
+        from agent.image_gen_provider import save_url_image
+
+        return Path(save_url_image(image_ref, prefix=prefix))
+    return Path(image_ref)
+
+
+def _rejected_background(error: str) -> bool:
+    """True when a provider error is specifically about the ``background`` param.
+
+    Transparent backgrounds are a per-model capability (e.g. some gpt-image tiers
+    reject ``background=transparent`` outright). We detect that one rejection so
+    we can retry without the flag rather than failing the whole pet — our chroma
+    key pass makes the result transparent regardless.
+    """
+    lowered = (error or "").lower()
+    return "background" in lowered and ("not supported" in lowered or "transparent" in lowered)
+
+
+def generate(
+    prompt: str,
+    *,
+    n: int = 1,
+    reference_images: list[Path] | None = None,
+    provider: SpriteProvider | None = None,
+    prefix: str = "pet_gen",
+) -> list[Path]:
+    """Generate *n* square sprite images and return their local paths.
+
+    *reference_images* grounds the output on a base image (required for rows).
+    We *ask* for a transparent background, but fall back to an opaque generation
+    (cleaned up downstream by the chroma-key pass) on models that reject the
+    flag. Raises :class:`GenerationError` if nothing usable comes back.
+    """
+    sprite = provider or resolve_provider(require_references=bool(reference_images))
+    if reference_images and not sprite.supports_references:
+        raise GenerationError(
+            f"image backend '{sprite.name}' cannot use reference images; "
+            "configure OpenAI gpt-image-2 or Krea for pet generation"
+        )
+
+    refs = [str(p) for p in (reference_images or [])]
+
+    def _run(extra: dict) -> tuple[Path | None, str]:
+        kwargs: dict = {"aspect_ratio": "square", **extra}
+        if refs:
+            kwargs["reference_images"] = refs
+        try:
+            result = sprite.provider.generate(prompt, **kwargs)
+        except Exception as exc:  # noqa: BLE001 - normalize provider crashes
+            logger.debug("provider.generate crashed: %s", exc)
+            return None, str(exc)
+        if not isinstance(result, dict) or not result.get("success"):
+            return None, (result or {}).get("error", "unknown error") if isinstance(result, dict) else "no result"
+        image_ref = result.get("image")
+        if not image_ref:
+            return None, "provider returned no image"
+        try:
+            return _save_local(str(image_ref), prefix=prefix), ""
+        except Exception as exc:  # noqa: BLE001
+            return None, f"could not save generated image: {exc}"
+
+    out: list[Path] = []
+    last_error = ""
+    allow_transparent = True
+    for _ in range(max(1, n)):
+        path, err = _run({"background": "transparent"} if allow_transparent else {})
+        # Model doesn't support the transparent flag → drop it for this and every
+        # remaining variant (no point re-probing a capability we just disproved).
+        if path is None and allow_transparent and _rejected_background(err):
+            allow_transparent = False
+            path, err = _run({})
+        if path is not None:
+            out.append(path)
+        else:
+            last_error = err
+
+    if not out:
+        raise GenerationError(last_error or "image generation produced no output")
+    return out

agent/pet/generate/orchestrate.py

@@ -0,0 +1,149 @@
+"""Pet generation orchestration — the base-draft → hatch flow.
+
+Two steps, mirroring the UX across every surface:
+
+1. :func:`generate_base_drafts` — a handful of prompt-only "what should this pet
+   look like" variants. Cheap; the user picks one (or retries for a fresh set).
+2. :func:`hatch_pet` — takes the chosen base and generates one grounded row
+   strip per Hermes state, slices each into frames, composes the atlas, validates
+   it, and writes the pet into the store.
+
+Splitting it this way bounds cost (4 cheap base calls per round; the ~6 row
+calls happen once, on the pet you actually keep) and gives each UI a natural
+preview/loading point.
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Callable
+
+from agent.pet.generate import atlas, imagegen, prompts
+from agent.pet.generate.imagegen import GenerationError, SpriteProvider
+
+logger = logging.getLogger(__name__)
+
+# (event, detail) — e.g. ("row", "idle"), ("compose", ""), ("save", "<slug>").
+ProgressFn = Callable[[str, str], None]
+
+
+@dataclass(frozen=True)
+class HatchResult:
+    """Outcome of a successful :func:`hatch_pet`."""
+
+    slug: str
+    display_name: str
+    spritesheet: Path
+    states: list[str]
+    validation: dict
+
+
+def _harden_transparency(path: Path) -> Path:
+    """Key out any solid backdrop the provider painted; save as an RGBA PNG.
+
+    ``background=transparent`` is requested on every call, but image models honor
+    it inconsistently — some still paint a flat (often near-white) backdrop. We
+    run the same chroma-key pass the row extractor uses so every base draft the
+    user picks between (and the reference the rows are grounded on) is a clean
+    cutout. Best-effort: a decode failure leaves the original untouched.
+    """
+    from PIL import Image
+
+    try:
+        with Image.open(path) as opened:
+            keyed = atlas.remove_background(opened.convert("RGBA"))
+        out = path.with_suffix(".png")
+        keyed.save(out, format="PNG")
+        return out
+    except Exception as exc:  # noqa: BLE001 - cosmetic; fall back to the raw image
+        logger.debug("base draft transparency hardening failed for %s: %s", path, exc)
+        return path
+
+
+def generate_base_drafts(
+    concept: str,
+    *,
+    n: int = 4,
+    style: str = "auto",
+    provider: SpriteProvider | None = None,
+) -> list[Path]:
+    """Generate *n* candidate base looks for *concept*; returns image paths.
+
+    Each draft is hardened to a transparent cutout (see :func:`_harden_transparency`).
+    """
+    prompt = prompts.build_base_prompt(concept, style=style)
+    sprite = provider or imagegen.resolve_provider(require_references=False)
+    raw = imagegen.generate(prompt, n=n, provider=sprite, prefix="pet_base")
+    return [_harden_transparency(p) for p in raw]
+
+
+def hatch_pet(
+    *,
+    base_image: str | Path,
+    slug: str,
+    display_name: str = "",
+    description: str = "",
+    concept: str = "",
+    style: str = "auto",
+    on_progress: ProgressFn | None = None,
+    provider: SpriteProvider | None = None,
+) -> HatchResult:
+    """Turn an approved base image into a full, installed Hermes pet.
+
+    Generates a grounded row strip per state, extracts frames, composes +
+    validates the atlas, and registers it. The idle row falls back to the base
+    look so the pet always renders. Raises :class:`GenerationError` on failure.
+    """
+    base = Path(base_image)
+    if not base.is_file():
+        raise GenerationError(f"base image not found: {base}")
+
+    sprite = provider or imagegen.resolve_provider(require_references=True)
+    progress = on_progress or (lambda *_: None)
+    label = concept or display_name or slug
+
+    frames_by_state: dict[str, list] = {}
+    for state, _row, count in atlas.ROW_SPECS:
+        progress("row", state)
+        row_prompt = prompts.build_row_prompt(state, count, label, style=style)
+        try:
+            strips = imagegen.generate(
+                row_prompt,
+                n=1,
+                reference_images=[base],
+                provider=sprite,
+                prefix=f"pet_row_{state}",
+            )
+            frames_by_state[state] = atlas.extract_strip_frames(strips[0], count, method="auto")
+        except Exception as exc:  # noqa: BLE001 - a single row may fail; keep going
+            logger.warning("pet row '%s' failed: %s", state, exc)
+
+    # Idle is the resting state the renderer falls back to — guarantee it.
+    if not frames_by_state.get("idle"):
+        progress("row", "idle-fallback")
+        frames_by_state["idle"] = [atlas.single_frame(base)]
+
+    progress("compose", "")
+    sheet = atlas.compose_atlas(frames_by_state)
+    validation = atlas.validate_atlas(sheet)
+    if not validation["ok"]:
+        raise GenerationError("; ".join(validation["errors"]) or "atlas validation failed")
+
+    from agent.pet import store
+
+    progress("save", slug)
+    pet = store.register_local_pet(
+        sheet,
+        slug=slug,
+        display_name=display_name or slug,
+        description=description,
+    )
+    return HatchResult(
+        slug=pet.slug,
+        display_name=pet.display_name,
+        spritesheet=pet.spritesheet,
+        states=validation["filled_states"],
+        validation=validation,
+    )

agent/pet/generate/prompts.py

@@ -0,0 +1,74 @@
+"""Prompt builders for pet generation.
+
+Two prompt shapes: a *base* prompt (prompt-only, produces the canonical look the
+user picks between) and per-*state* *row* prompts (grounded on the chosen base,
+produce one horizontal strip of N poses). Prompts stay concise and
+sprite-production oriented; the identity lock and "one transparent row" framing
+matter more than flowery description.
+
+Hermes drives six states (see :data:`agent.pet.generate.atlas.ROW_SPECS`); these
+mirror that set rather than the petdex/Codex nine.
+"""
+
+from __future__ import annotations
+
+# What each Hermes state should depict (kept short — these go straight into the
+# row prompt). Phrased to avoid the common sprite-gen failure modes (detached
+# effects, motion lines, shadows).
+STATE_ACTIONS: dict[str, str] = {
+    "idle": "a calm idle loop: subtle breathing, a tiny blink or gentle bob, no big gestures",
+    "wave": "a friendly greeting: raising a paw/hand/limb to wave, clear up-and-down gesture",
+    "run": "focused active work: leaning in, concentrating, busy 'thinking/processing' energy (NOT foot-running)",
+    "failed": "a sad or deflated reaction: slumped, dejected, small frown — readable but not noisy",
+    "review": "careful inspection: a focused lean, head tilt, studying something intently",
+    "jump": "a happy celebration jump: anticipation, lift off the ground, peak, and land",
+}
+
+_STYLE_HINTS: dict[str, str] = {
+    "auto": "",
+    "pixel": " Render in clean pixel-art style.",
+    "plush": " Render as a soft plush toy.",
+    "clay": " Render as a claymation / soft 3D clay figure.",
+    "sticker": " Render as a glossy die-cut sticker.",
+    "flat-vector": " Render in flat vector mascot style.",
+    "3d-toy": " Render as a glossy 3D toy.",
+    "painterly": " Render in a soft painterly style.",
+}
+
+_BACKGROUND = (
+    "Center one full-body character on a fully transparent background. "
+    "No text, no labels, no shadow, no ground line, no scenery, no frame, no border."
+)
+
+
+def style_hint(style: str | None) -> str:
+    return _STYLE_HINTS.get((style or "auto").strip().lower(), "")
+
+
+def build_base_prompt(concept: str, *, style: str | None = "auto") -> str:
+    """The base look: a single, clean, centered full-body mascot."""
+    concept = (concept or "a cute friendly mascot creature").strip()
+    return (
+        f"A cute, characterful mascot pet: {concept}. "
+        "Compact, whole-body silhouette that reads clearly at small size, "
+        "appealing face, simple consistent palette. "
+        f"{_BACKGROUND}{style_hint(style)}"
+    )
+
+
+def build_row_prompt(state: str, frame_count: int, concept: str, *, style: str | None = "auto") -> str:
+    """A row strip: *frame_count* poses of the SAME character, left→right.
+
+    The attached base image is the identity source of truth; the prompt locks
+    species, palette, face, and props to it.
+    """
+    action = STATE_ACTIONS.get(state, "a simple idle pose")
+    concept = (concept or "the mascot").strip()
+    return (
+        f"Using the attached reference image as the exact same character "
+        f"(same species, face, colors, markings, proportions, and props), "
+        f"draw a single horizontal strip of {frame_count} animation frames showing {action}. "
+        f"The {frame_count} poses must be evenly spaced left to right, each fully separated "
+        "(not overlapping), same size and baseline, forming a smooth loop. "
+        f"Keep the character identical across all frames. {_BACKGROUND}{style_hint(style)}"
+    )

agent/pet/manifest.py

@@ -0,0 +1,128 @@
+"""Fetch the public petdex manifest.
+
+``https://petdex.dev/api/manifest`` 307-redirects to a JSON document on R2:
+
+    {
+      "generatedAt": "...",
+      "total": 2926,
+      "pets": [
+        {"slug": "boba", "displayName": "Boba", "kind": "creature",
+         "submittedBy": "railly",
+         "spritesheetUrl": "https://assets.petdex.dev/.../spritesheet.webp",
+         "petJsonUrl": "https://assets.petdex.dev/.../pet.json",
+         "zipUrl": "https://assets.petdex.dev/.../boba.zip"},
+        ...
+      ]
+    }
+
+Read-only and unauthenticated; no credentials involved.
+"""
+
+from __future__ import annotations
+
+import logging
+import time
+from dataclasses import dataclass
+
+logger = logging.getLogger(__name__)
+
+MANIFEST_URL = "https://petdex.dev/api/manifest"
+
+_DEFAULT_TIMEOUT = 20.0
+
+# In-process cache for the (large, slow, identical-per-call) manifest. The list
+# is a static CDN object that barely changes, yet a single session can ask for
+# it many times — every gallery open, plus a full re-fetch per install/select
+# (``find_entry``). A short TTL collapses those into one network hit without
+# going stale for long. Cleared by :func:`clear_cache` (tests).
+_MANIFEST_TTL = 300.0
+_cache: tuple[float, list[ManifestEntry]] | None = None
+
+
+def clear_cache() -> None:
+    """Drop the cached manifest (forces the next fetch to hit the network)."""
+    global _cache
+    _cache = None
+
+
+@dataclass(frozen=True)
+class ManifestEntry:
+    """A single pet's row in the manifest."""
+
+    slug: str
+    display_name: str
+    kind: str
+    submitted_by: str
+    spritesheet_url: str
+    pet_json_url: str
+    zip_url: str
+
+    @classmethod
+    def from_dict(cls, data: dict) -> "ManifestEntry":
+        return cls(
+            slug=str(data.get("slug", "")).strip(),
+            display_name=str(data.get("displayName", "") or data.get("slug", "")),
+            kind=str(data.get("kind", "") or "pet"),
+            submitted_by=str(data.get("submittedBy", "") or ""),
+            spritesheet_url=str(data.get("spritesheetUrl", "") or ""),
+            pet_json_url=str(data.get("petJsonUrl", "") or ""),
+            zip_url=str(data.get("zipUrl", "") or ""),
+        )
+
+
+class ManifestError(RuntimeError):
+    """Raised when the manifest can't be fetched or parsed."""
+
+
+def fetch_manifest(*, timeout: float = _DEFAULT_TIMEOUT, force: bool = False) -> list[ManifestEntry]:
+    """Return every approved pet from the public manifest.
+
+    Cached in-process for ``_MANIFEST_TTL`` seconds (pass ``force=True`` to
+    bypass). Follows the 307 redirect to R2.  Raises :class:`ManifestError` on
+    any network/parse failure so callers can surface a clean message.
+    """
+    global _cache
+
+    if not force and _cache is not None and time.monotonic() - _cache[0] < _MANIFEST_TTL:
+        return _cache[1]
+
+    try:
+        import httpx
+    except ImportError as exc:  # pragma: no cover - httpx is a core dep
+        raise ManifestError("httpx is required to fetch the petdex manifest") from exc
+
+    try:
+        resp = httpx.get(
+            MANIFEST_URL,
+            timeout=timeout,
+            follow_redirects=True,
+            headers={"User-Agent": "hermes-agent-petdex"},
+        )
+        resp.raise_for_status()
+        payload = resp.json()
+    except Exception as exc:  # noqa: BLE001 - normalize to one error type
+        raise ManifestError(f"could not fetch petdex manifest: {exc}") from exc
+
+    pets = payload.get("pets") if isinstance(payload, dict) else None
+    if not isinstance(pets, list):
+        raise ManifestError("petdex manifest had no 'pets' array")
+
+    entries: list[ManifestEntry] = []
+    for raw in pets:
+        if not isinstance(raw, dict):
+            continue
+        entry = ManifestEntry.from_dict(raw)
+        if entry.slug and entry.spritesheet_url:
+            entries.append(entry)
+
+    _cache = (time.monotonic(), entries)
+    return entries
+
+
+def find_entry(slug: str, *, timeout: float = _DEFAULT_TIMEOUT) -> ManifestEntry | None:
+    """Return the manifest entry for *slug*, or ``None`` if not listed."""
+    slug = slug.strip().lower()
+    for entry in fetch_manifest(timeout=timeout):
+        if entry.slug.lower() == slug:
+            return entry
+    return None

agent/pet/render.py

@@ -0,0 +1,618 @@
+"""Decode a pet spritesheet and encode frames for a terminal.
+
+Shared by the base CLI (writes the escape bytes to its own stdout) and the
+TUI (``tui_gateway`` ships the encoded bytes to Ink, which writes them) so the
+decode + capability-detection + protocol-encoding logic exists exactly once.
+
+Supported output modes, in fidelity order:
+
+- ``kitty``   — the kitty graphics protocol (kitty, Ghostty, WezTerm).
+- ``iterm``   — iTerm2 inline images (iTerm2, WezTerm).
+- ``sixel``   — DEC sixel (xterm -ti vt340, foot, mlterm, WezTerm, …).
+- ``unicode`` — 24-bit half-block downscale; works in any truecolor terminal.
+
+Frame decoding requires Pillow (a core Hermes dependency).  If Pillow or the
+spritesheet is unavailable the renderer degrades to ``unicode`` text or an
+empty string rather than raising.
+"""
+
+from __future__ import annotations
+
+import base64
+import io
+import logging
+import os
+import sys
+from functools import lru_cache
+from pathlib import Path
+
+from agent.pet.constants import (
+    DEFAULT_SCALE,
+    FRAME_H,
+    FRAME_W,
+    FRAMES_PER_STATE,
+    PetState,
+    state_row_index,
+)
+
+logger = logging.getLogger(__name__)
+
+# Public render-mode names accepted by ``display.pet.render_mode``.
+RENDER_MODES = ("auto", "kitty", "iterm", "sixel", "unicode", "off")
+
+
+# ─────────────────────────────────────────────────────────────────────────
+# Terminal capability detection
+# ─────────────────────────────────────────────────────────────────────────
+
+def detect_terminal_graphics() -> str:
+    """Best-effort detection of the richest graphics protocol available.
+
+    Env-based (non-blocking — we never issue a DA1/terminal query that could
+    hang a pipe).  Returns one of ``kitty`` / ``iterm`` / ``sixel`` /
+    ``unicode``.  Conservative: unknown terminals get ``unicode``, which works
+    anywhere with truecolor.
+    """
+    term = os.environ.get("TERM", "").lower()
+    term_program = os.environ.get("TERM_PROGRAM", "").lower()
+
+    # The VS Code / Cursor integrated terminal sets TERM_PROGRAM=vscode
+    # authoritatively but does NOT scrub the terminal env vars it inherits when
+    # launched from another emulator (ITERM_SESSION_ID, KITTY_WINDOW_ID, …).
+    # Trusting those leaks emits an image protocol the embedded xterm.js can't
+    # display — you get a blank frame. Inline images there are opt-in
+    # (terminal.integrated.enableImages), so default to half-blocks, which
+    # always render in its truecolor grid. Users who enabled images can pin
+    # display.pet.render_mode explicitly.
+    if term_program == "vscode":
+        return "unicode"
+
+    # kitty graphics protocol
+    if os.environ.get("KITTY_WINDOW_ID") or "kitty" in term or "ghostty" in term:
+        return "kitty"
+    if term_program in {"ghostty"}:
+        return "kitty"
+
+    # WezTerm speaks both kitty and iterm; prefer kitty (richer placement).
+    if term_program == "wezterm" or os.environ.get("WEZTERM_PANE"):
+        return "kitty"
+
+    # iTerm2 inline images
+    if term_program == "iterm.app" or os.environ.get("ITERM_SESSION_ID"):
+        return "iterm"
+
+    # sixel-capable terminals (env heuristics only)
+    if term_program in {"mintty"} or "foot" in term or "mlterm" in term:
+        return "sixel"
+    if "sixel" in term:
+        return "sixel"
+
+    return "unicode"
+
+
+def resolve_mode(configured: str | None, *, stream=None) -> str:
+    """Resolve the effective render mode from config + the environment.
+
+    ``configured`` is ``display.pet.render_mode`` (``auto`` → detect).  Returns
+    ``off`` when not attached to a TTY (no point emitting graphics into a pipe
+    or logfile).
+    """
+    mode = (configured or "auto").strip().lower()
+    if mode not in RENDER_MODES:
+        mode = "auto"
+    if mode == "off":
+        return "off"
+
+    stream = stream or sys.stdout
+    try:
+        if not (hasattr(stream, "isatty") and stream.isatty()):
+            return "off"
+    except (ValueError, OSError):
+        return "off"
+
+    if mode == "auto":
+        return detect_terminal_graphics()
+    return mode
+
+
+# ─────────────────────────────────────────────────────────────────────────
+# Frame decoding
+# ─────────────────────────────────────────────────────────────────────────
+
+def _open_sheet(path: Path):
+    from PIL import Image
+
+    img = Image.open(path)
+    return img.convert("RGBA")
+
+
+# Max alpha at/below which a frame counts as blank padding.  petdex sheets are
+# left-packed: a state with fewer real frames than ``FRAMES_PER_STATE`` fills
+# the trailing columns with fully transparent cells.  Animating into one flashes
+# the pet blank, so we stop the row at the first such gap.
+_BLANK_ALPHA = 8
+
+
+def _frame_is_blank(frame) -> bool:
+    """True if *frame* has no meaningfully opaque pixel (transparent padding)."""
+    return frame.getchannel("A").getextrema()[1] <= _BLANK_ALPHA
+
+
+@lru_cache(maxsize=16)
+def _raw_frames(
+    sheet_path: str,
+    state_value: str,
+    frame_w: int,
+    frame_h: int,
+    frames_per_state: int,
+) -> tuple:
+    """Cropped, padding-trimmed RGBA frames for one state row (unscaled).
+
+    Steps across the row until the first blank column so pets with ragged
+    per-state frame counts never animate into empty padding.  Cached; returns
+    ``()`` on any decode failure.
+    """
+    try:
+        sheet = _open_sheet(Path(sheet_path))
+        cols = max(1, sheet.width // frame_w)
+        rows = max(1, sheet.height // frame_h)
+        row = state_row_index(state_value, rows)
+        top = row * frame_h
+        # Clamp the row to the sheet (some pets ship fewer rows than the 8 the
+        # taxonomy reserves).
+        if top + frame_h > sheet.height:
+            top = max(0, sheet.height - frame_h)
+
+        frames = []
+        for i in range(min(frames_per_state, cols)):
+            left = i * frame_w
+            frame = sheet.crop((left, top, left + frame_w, top + frame_h))
+            if _frame_is_blank(frame):
+                break  # trailing transparent padding — real frames end here
+            frames.append(frame)
+        return tuple(frames)
+    except Exception as exc:  # noqa: BLE001 - cosmetic feature, never fatal
+        logger.debug("pet frame decode failed (%s, %s): %s", sheet_path, state_value, exc)
+        return ()
+
+
+@lru_cache(maxsize=8)
+def _frames_for(
+    sheet_path: str,
+    state_value: str,
+    frame_w: int,
+    frame_h: int,
+    frames_per_state: int,
+    scale_w: int,
+    scale_h: int,
+):
+    """Return padding-trimmed RGBA frames for one state row, scaled.
+
+    Thin scaling layer over :func:`_raw_frames`; both are cached so repeated
+    frame requests during animation are free.
+    """
+    raw = _raw_frames(sheet_path, state_value, frame_w, frame_h, frames_per_state)
+    if not raw or (scale_w, scale_h) == (frame_w, frame_h):
+        return list(raw)
+    from PIL import Image
+
+    return [f.resize((scale_w, scale_h), Image.LANCZOS) for f in raw]
+
+
+def state_frame_counts(
+    sheet_path: str | Path,
+    *,
+    frame_w: int = FRAME_W,
+    frame_h: int = FRAME_H,
+    frames_per_state: int = FRAMES_PER_STATE,
+) -> dict[str, int]:
+    """Map each driven :class:`PetState` → its real (padding-trimmed) frame count.
+
+    The single source of truth for "how many frames does this state actually
+    have?".  The CLI/TUI consume the trimmed frame lists directly; the gateway
+    ships this map to the desktop canvas, which steps its own loop.
+    """
+    return {
+        state.value: len(
+            _raw_frames(str(sheet_path), state.value, frame_w, frame_h, frames_per_state)
+        )
+        for state in PetState
+    }
+
+
+# ─────────────────────────────────────────────────────────────────────────
+# Encoders
+# ─────────────────────────────────────────────────────────────────────────
+
+def _png_bytes(frame) -> bytes:
+    buf = io.BytesIO()
+    frame.save(buf, format="PNG")
+    return buf.getvalue()
+
+
+def _kitty_apc(ctrl: str, data: str) -> str:
+    """Emit a kitty APC escape for *data*, chunked into ≤4096-byte ``m`` pieces."""
+    chunk = 4096
+    if len(data) <= chunk:
+        return f"\x1b_G{ctrl},m=0;{data}\x1b\\"
+    out = [f"\x1b_G{ctrl},m=1;{data[:chunk]}\x1b\\"]
+    rest = data[chunk:]
+    while rest:
+        piece, rest = rest[:chunk], rest[chunk:]
+        out.append(f"\x1b_Gm={1 if rest else 0};{piece}\x1b\\")
+    return "".join(out)
+
+
+def _encode_kitty(frame, *, cell_cols: int | None = None, cell_rows: int | None = None) -> str:
+    """Encode one frame via the kitty graphics protocol (transmit + display).
+
+    ``a=T`` transmits & displays at the cursor; ``c``/``r`` request a display
+    box in terminal cells so successive frames overwrite the same area.
+    """
+    ctrl = "f=100,a=T,q=2"
+    if cell_cols:
+        ctrl += f",c={cell_cols}"
+    if cell_rows:
+        ctrl += f",r={cell_rows}"
+    return _kitty_apc(ctrl, base64.standard_b64encode(_png_bytes(frame)).decode("ascii"))
+
+
+# ─────────────────────────────────────────────────────────────────────────
+# kitty Unicode placeholders
+#
+# Ink (the TUI's React-for-terminal layer) owns the screen and measures every
+# cell's width, so it can't host raw kitty image escapes (no width to count,
+# clobbered on the next repaint). kitty's *Unicode placeholder* protocol is the
+# grid-safe path: transmit the image once (q=2, virtual placement U=1), then the
+# host app prints ordinary-width placeholder cells (U+10EEEE + diacritics) whose
+# foreground color encodes the image id. Ink counts those as width-1 text, so
+# layout stays correct and the terminal paints the image underneath.
+#   https://sw.kovidgoyal.net/kitty/graphics-protocol/#unicode-placeholders
+# ─────────────────────────────────────────────────────────────────────────
+
+_KITTY_PLACEHOLDER = "\U0010eeee"
+
+# Row/column diacritics, in order (index → diacritic). Verbatim from kitty's
+# gen/rowcolumn-diacritics.txt (Unicode 6.0.0, combining class 230). Index i is
+# the diacritic that encodes the number i; we only ever need the row index.
+_ROWCOL_DIACRITICS: tuple[int, ...] = (
+    0x0305, 0x030D, 0x030E, 0x0310, 0x0312, 0x033D, 0x033E, 0x033F, 0x0346, 0x034A,
+    0x034B, 0x034C, 0x0350, 0x0351, 0x0352, 0x0357, 0x035B, 0x0363, 0x0364, 0x0365,
+    0x0366, 0x0367, 0x0368, 0x0369, 0x036A, 0x036B, 0x036C, 0x036D, 0x036E, 0x036F,
+    0x0483, 0x0484, 0x0485, 0x0486, 0x0487, 0x0592, 0x0593, 0x0594, 0x0595, 0x0597,
+    0x0598, 0x0599, 0x059C, 0x059D, 0x059E, 0x059F, 0x05A0, 0x05A1, 0x05A8, 0x05A9,
+    0x05AB, 0x05AC, 0x05AF, 0x05C4, 0x0610, 0x0611, 0x0612, 0x0613, 0x0614, 0x0615,
+    0x0616, 0x0617, 0x0657, 0x0658, 0x0659, 0x065A, 0x065B, 0x065D, 0x065E, 0x06D6,
+    0x06D7, 0x06D8, 0x06D9, 0x06DA, 0x06DB, 0x06DC, 0x06DF, 0x06E0, 0x06E1, 0x06E2,
+    0x06E4, 0x06E7, 0x06E8, 0x06EB, 0x06EC, 0x0730, 0x0732, 0x0733, 0x0735, 0x0736,
+    0x073A, 0x073D, 0x073F, 0x0740, 0x0741, 0x0743, 0x0745, 0x0747, 0x0749, 0x074A,
+    0x07EB, 0x07EC, 0x07ED, 0x07EE, 0x07EF, 0x07F0, 0x07F1, 0x07F3, 0x0816, 0x0817,
+    0x0818, 0x0819, 0x081B, 0x081C, 0x081D, 0x081E, 0x081F, 0x0820, 0x0821, 0x0822,
+    0x0823, 0x0825, 0x0826, 0x0827, 0x0829, 0x082A, 0x082B, 0x082C, 0x082D, 0x0951,
+    0x0953, 0x0954, 0x0F82, 0x0F83, 0x0F86, 0x0F87, 0x135D, 0x135E, 0x135F, 0x17DD,
+    0x193A, 0x1A17, 0x1A75, 0x1A76, 0x1A77, 0x1A78, 0x1A79, 0x1A7A, 0x1A7B, 0x1A7C,
+    0x1B6B, 0x1B6D, 0x1B6E, 0x1B6F, 0x1B70, 0x1B71, 0x1B72, 0x1B73, 0x1CD0, 0x1CD1,
+    0x1CD2, 0x1CDA, 0x1CDB, 0x1CE0, 0x1DC0, 0x1DC1, 0x1DC3, 0x1DC4, 0x1DC5, 0x1DC6,
+    0x1DC7, 0x1DC8, 0x1DC9, 0x1DCB, 0x1DCC, 0x1DD1, 0x1DD2, 0x1DD3, 0x1DD4, 0x1DD5,
+    0x1DD6, 0x1DD7, 0x1DD8, 0x1DD9, 0x1DDA, 0x1DDB, 0x1DDC, 0x1DDD, 0x1DDE, 0x1DDF,
+    0x1DE0, 0x1DE1, 0x1DE2, 0x1DE3, 0x1DE4, 0x1DE5, 0x1DE6, 0x1DFE, 0x20D0, 0x20D1,
+    0x20D4, 0x20D5, 0x20D6, 0x20D7, 0x20DB, 0x20DC, 0x20E1, 0x20E7, 0x20E9, 0x20F0,
+    0x2CEF, 0x2CF0, 0x2CF1, 0x2DE0, 0x2DE1, 0x2DE2, 0x2DE3, 0x2DE4, 0x2DE5, 0x2DE6,
+    0x2DE7, 0x2DE8, 0x2DE9, 0x2DEA, 0x2DEB, 0x2DEC, 0x2DED, 0x2DEE, 0x2DEF, 0x2DF0,
+    0x2DF1, 0x2DF2, 0x2DF3, 0x2DF4, 0x2DF5, 0x2DF6, 0x2DF7, 0x2DF8, 0x2DF9, 0x2DFA,
+    0x2DFB, 0x2DFC, 0x2DFD, 0x2DFE, 0x2DFF, 0xA66F, 0xA67C, 0xA67D, 0xA6F0, 0xA6F1,
+    0xA8E0, 0xA8E1, 0xA8E2, 0xA8E3, 0xA8E4, 0xA8E5, 0xA8E6, 0xA8E7, 0xA8E8, 0xA8E9,
+    0xA8EA, 0xA8EB, 0xA8EC, 0xA8ED, 0xA8EE, 0xA8EF, 0xA8F0, 0xA8F1, 0xAAB0, 0xAAB2,
+    0xAAB3, 0xAAB7, 0xAAB8, 0xAABE, 0xAABF, 0xAAC1, 0xFE20, 0xFE21, 0xFE22, 0xFE23,
+    0xFE24, 0xFE25, 0xFE26, 0x10A0F, 0x10A38, 0x1D185, 0x1D186, 0x1D187, 0x1D188,
+    0x1D189, 0x1D1AA, 0x1D1AB, 0x1D1AC, 0x1D1AD, 0x1D242, 0x1D243, 0x1D244,
+)
+
+
+def kitty_image_id(slug: str) -> int:
+    """Stable per-pet image id in ``[1, 0x7FFF]``.
+
+    The id is encoded in the placeholder's 24-bit foreground color, so it must
+    be non-zero and fit comfortably under ``0xFFFFFF``. A small CRC keeps it
+    deterministic per slug (so re-renders reuse the same terminal-side image)
+    while making collisions between two different pets unlikely.
+    """
+    import zlib
+
+    return (zlib.crc32(slug.encode("utf-8")) % 0x7FFE) + 1
+
+
+def kitty_color_hex(image_id: int) -> str:
+    """Hex foreground color (``#rrggbb``) that encodes *image_id* for kitty."""
+    return "#%06x" % (image_id & 0xFFFFFF)
+
+
+def kitty_placeholder_rows(cols: int, rows: int) -> list[str]:
+    """Build the placeholder text grid for an *rows*×*cols* image.
+
+    Each line is one row of the grid: the first cell carries the row diacritic
+    (column defaults to 0), and the remaining ``cols-1`` bare placeholders let
+    the terminal auto-increment the column. The foreground color (the image id)
+    is applied by the caller / Ink, not embedded here.
+    """
+    cols = max(1, cols)
+    out: list[str] = []
+    for r in range(max(1, rows)):
+        idx = min(r, len(_ROWCOL_DIACRITICS) - 1)
+        first = _KITTY_PLACEHOLDER + chr(_ROWCOL_DIACRITICS[idx])
+        out.append(first + _KITTY_PLACEHOLDER * (cols - 1))
+    return out
+
+
+def _encode_kitty_virtual(frame, *, image_id: int, cols: int, rows: int) -> str:
+    """Transmit a frame as a kitty *virtual* placement for Unicode placeholders.
+
+    ``a=T`` transmits and creates the placement in one shot; ``U=1`` marks it
+    virtual (no on-screen output, cursor untouched); ``q=2`` suppresses the
+    terminal's OK/error replies that would otherwise corrupt the host app's
+    output. Re-sending with the same ``i`` replaces the image, so the static
+    placeholder cells animate underneath.
+    """
+    ctrl = f"a=T,U=1,i={image_id},c={cols},r={rows},f=100,q=2"
+    return _kitty_apc(ctrl, base64.standard_b64encode(_png_bytes(frame)).decode("ascii"))
+
+
+def _encode_iterm(frame, *, cell_cols: int | None = None, cell_rows: int | None = None) -> str:
+    """Encode one frame as an iTerm2 inline image (OSC 1337 File)."""
+    payload = base64.standard_b64encode(_png_bytes(frame)).decode("ascii")
+    size = len(payload)
+    args = [f"inline=1", f"size={size}", "preserveAspectRatio=1"]
+    if cell_cols:
+        args.append(f"width={cell_cols}")
+    if cell_rows:
+        args.append(f"height={cell_rows}")
+    return f"\x1b]1337;File={';'.join(args)}:{payload}\x07"
+
+
+def _encode_sixel(frame) -> str:
+    """Encode one frame as DEC sixel.
+
+    Quantizes to an adaptive palette (≤255 colors) and emits the sixel band
+    stream.  Pillow has no sixel writer, so this is a compact hand-rolled
+    encoder.  Transparent pixels render as background (color register skipped).
+    """
+    from PIL import Image
+
+    rgba = frame
+    # Composite onto transparent-as-skip: track alpha to decide background.
+    pal = rgba.convert("RGB").quantize(colors=255, method=Image.MEDIANCUT)
+    palette = pal.getpalette() or []
+    px = pal.load()
+    alpha = rgba.getchannel("A").load()
+    w, h = pal.size
+
+    out = ["\x1bP0;1;0q", '"1;1;%d;%d' % (w, h)]
+    # Color register definitions (sixel uses 0..100 scale).
+    used = sorted({px[x, y] for y in range(h) for x in range(w)})
+    for idx in used:
+        r = palette[idx * 3] if idx * 3 < len(palette) else 0
+        g = palette[idx * 3 + 1] if idx * 3 + 1 < len(palette) else 0
+        b = palette[idx * 3 + 2] if idx * 3 + 2 < len(palette) else 0
+        out.append("#%d;2;%d;%d;%d" % (idx, r * 100 // 255, g * 100 // 255, b * 100 // 255))
+
+    # Emit in 6-row bands.
+    for band in range(0, h, 6):
+        for color_idx in used:
+            line = ["#%d" % color_idx]
+            run_char = None
+            run_len = 0
+
+            def flush():
+                nonlocal run_char, run_len
+                if run_char is None:
+                    return
+                if run_len > 3:
+                    line.append("!%d%s" % (run_len, run_char))
+                else:
+                    line.append(run_char * run_len)
+                run_char, run_len = None, 0
+
+            for x in range(w):
+                bits = 0
+                for bit in range(6):
+                    y = band + bit
+                    if y < h and alpha[x, y] > 32 and px[x, y] == color_idx:
+                        bits |= 1 << bit
+                ch = chr(63 + bits)
+                if ch == run_char:
+                    run_len += 1
+                else:
+                    flush()
+                    run_char, run_len = ch, 1
+            flush()
+            out.append("".join(line) + "$")  # carriage return within band
+        out.append("-")  # next band
+    out.append("\x1b\\")
+    return "".join(out)
+
+
+_HALF_BLOCK = "▀"
+
+# A single half-block cell: top pixel + bottom pixel as (r, g, b, a) tuples.
+Cell = tuple[tuple[int, int, int, int], tuple[int, int, int, int]]
+
+
+def _downscale_cells(frame, *, target_cols: int) -> list[list[Cell]]:
+    """Downscale a frame to a grid of half-block cells.
+
+    Each cell pairs a top and bottom pixel so one terminal row encodes two
+    pixel rows.  Returns rows of ``((tr,tg,tb,ta),(br,bg,bb,ba))`` — the
+    framework-neutral representation shared by the ANSI encoder (CLI) and the
+    structured ``cells`` API (Ink).
+    """
+    from PIL import Image
+
+    target_cols = max(4, target_cols)
+    aspect = frame.height / max(1, frame.width)
+    target_rows = max(2, int(round(target_cols * aspect * 0.5)) * 2)
+    small = frame.resize((target_cols, target_rows), Image.LANCZOS).convert("RGBA")
+    px = small.load()
+
+    grid: list[list[Cell]] = []
+    for y in range(0, target_rows, 2):
+        row: list[Cell] = []
+        for x in range(target_cols):
+            top = px[x, y]
+            bottom = px[x, y + 1] if y + 1 < target_rows else (0, 0, 0, 0)
+            row.append((top, bottom))
+        grid.append(row)
+    return grid
+
+
+def _encode_unicode(frame, *, target_cols: int) -> str:
+    """Downscale to truecolor ANSI half-blocks (one char = 2 vertical pixels)."""
+    lines: list[str] = []
+    for row in _downscale_cells(frame, target_cols=target_cols):
+        cells: list[str] = []
+        for (tr, tg, tb, ta), (br, bg, bb, ba) in row:
+            if ta < 32 and ba < 32:
+                cells.append("\x1b[0m ")  # fully transparent → blank
+                continue
+            cells.append(f"\x1b[38;2;{tr};{tg};{tb}m\x1b[48;2;{br};{bg};{bb}m{_HALF_BLOCK}")
+        lines.append("".join(cells) + "\x1b[0m")
+    return "\n".join(lines)
+
+
+# ─────────────────────────────────────────────────────────────────────────
+# Public renderer
+# ─────────────────────────────────────────────────────────────────────────
+
+class PetRenderer:
+    """Holds a pet's spritesheet and yields encoded frames per (state, index).
+
+    Construct once per pet, then call :meth:`frame` on an animation timer.
+    Cheap to call repeatedly — decoded frames are cached.
+    """
+
+    def __init__(
+        self,
+        spritesheet: str | Path,
+        *,
+        mode: str = "unicode",
+        scale: float = DEFAULT_SCALE,
+        unicode_cols: int = 20,
+        frame_w: int = FRAME_W,
+        frame_h: int = FRAME_H,
+        frames_per_state: int = FRAMES_PER_STATE,
+    ) -> None:
+        self.spritesheet = str(spritesheet)
+        self.mode = mode if mode in RENDER_MODES else "unicode"
+        self.scale = scale
+        self.unicode_cols = unicode_cols
+        self.frame_w = frame_w
+        self.frame_h = frame_h
+        self.frames_per_state = frames_per_state
+
+    @property
+    def available(self) -> bool:
+        return self.mode != "off" and Path(self.spritesheet).is_file()
+
+    def frame_count(self, state: PetState | str) -> int:
+        return len(self._frames(state))
+
+    def _frames(self, state: PetState | str):
+        value = state.value if isinstance(state, PetState) else str(state)
+        scale_w = max(1, int(self.frame_w * self.scale))
+        scale_h = max(1, int(self.frame_h * self.scale))
+        return _frames_for(
+            self.spritesheet,
+            value,
+            self.frame_w,
+            self.frame_h,
+            self.frames_per_state,
+            scale_w,
+            scale_h,
+        )
+
+    def cells(self, state: PetState | str, index: int, *, cols: int | None = None) -> list[list[Cell]]:
+        """Return one frame as a half-block cell grid (framework-neutral).
+
+        Used by the TUI, which renders the grid with native Ink color props
+        instead of raw ANSI.  Returns ``[]`` when no frame is available.
+        """
+        frames = self._frames(state)
+        if not frames:
+            return []
+        frame = frames[index % len(frames)]
+        return _downscale_cells(frame, target_cols=cols or self.unicode_cols)
+
+    def _cell_box(self, frame) -> tuple[int, int]:
+        """Terminal cell box for a scaled frame (~8×16 px per cell).
+
+        Must match :meth:`frame` graphics sizing — kitty stretches the image to
+        fill ``c``×``r`` cells, so these must reflect the scaled pixel
+        dimensions, not a native-aspect column count (that upscales small pets).
+        """
+        return max(1, frame.width // 8), max(1, frame.height // 16)
+
+    def kitty_payload(self, state: PetState | str, *, image_id: int) -> dict | None:
+        """Build the kitty Unicode-placeholder payload for one state.
+
+        Returns ``{cols, rows, placeholder, frames}`` where ``frames`` is a
+        list of transmit escapes (one per animation frame, all reusing
+        ``image_id``) and ``placeholder`` is the static text grid Ink paints.
+        Placement geometry is derived from the scaled frame pixels (via
+        :meth:`_cell_box`), not ``unicode_cols`` — kitty upscales to fill
+        ``c``×``r`` cells. ``None`` when no frame is available.
+        """
+        frames = self._frames(state)
+        if not frames:
+            return None
+        cols, rows = self._cell_box(frames[0])
+        return {
+            "cols": cols,
+            "rows": rows,
+            "placeholder": kitty_placeholder_rows(cols, rows),
+            "frames": [
+                _encode_kitty_virtual(f, image_id=image_id, cols=cols, rows=rows) for f in frames
+            ],
+        }
+
+    def frame(self, state: PetState | str, index: int) -> str:
+        """Return the encoded escape string for one frame, or ``""``.
+
+        ``index`` is taken modulo the available frame count so callers can pass
+        a free-running counter.
+        """
+        if self.mode == "off":
+            return ""
+        frames = self._frames(state)
+        if not frames:
+            return ""
+        frame = frames[index % len(frames)]
+        cell_cols, cell_rows = self._cell_box(frame)
+
+        try:
+            if self.mode == "kitty":
+                return _encode_kitty(frame, cell_cols=cell_cols, cell_rows=cell_rows)
+            if self.mode == "iterm":
+                return _encode_iterm(frame, cell_cols=cell_cols, cell_rows=cell_rows)
+            if self.mode == "sixel":
+                return _encode_sixel(frame)
+            return _encode_unicode(frame, target_cols=self.unicode_cols)
+        except Exception as exc:  # noqa: BLE001 - degrade silently
+            logger.debug("pet frame encode failed (mode=%s): %s", self.mode, exc)
+            return ""
+
+
+def build_renderer(
+    spritesheet: str | Path,
+    *,
+    configured_mode: str | None = None,
+    scale: float = DEFAULT_SCALE,
+    unicode_cols: int = 20,
+    stream=None,
+) -> PetRenderer:
+    """Convenience factory: resolve the mode from config+env, then construct."""
+    mode = resolve_mode(configured_mode, stream=stream)
+    return PetRenderer(
+        spritesheet,
+        mode=mode,
+        scale=scale,
+        unicode_cols=unicode_cols,
+    )

agent/pet/state.py

@@ -0,0 +1,81 @@
+"""Map agent activity → a :class:`PetState`.
+
+This is the one place the "what is the agent doing right now?" → "which
+animation row?" decision lives.  Each surface feeds it the signals it already
+tracks:
+
+- CLI    — ``KawaiiSpinner`` waiting/thinking state + tool outcomes.
+- TUI    — gateway ``tool.start/complete`` + ``message.delta/complete`` events.
+- Desktop — the ``$busy``/``$awaitingResponse``/tool-event nanostores
+            (re-implemented in TS, but mirroring this priority order).
+
+Keeping the priority order here (and documenting it) lets the TypeScript
+mirror stay faithful without a second design.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Iterable
+from typing import Any
+
+from agent.pet.constants import PetState
+
+
+def todos_all_done(todos: Iterable[Any] | None) -> bool:
+    """True iff there's ≥1 todo and every one is completed/cancelled.
+
+    The "celebrate" beat (``JUMP``) fires when a plan finishes; this mirrors
+    the TUI's ``isTodoDone`` so the trigger is defined once across surfaces.
+    Accepts dicts (``{"status": ...}``) or objects with a ``status`` attr.
+    """
+    items = list(todos or [])
+    if not items:
+        return False
+
+    def _status(t: Any) -> Any:
+        return t.get("status") if isinstance(t, dict) else getattr(t, "status", None)
+
+    return all(_status(t) in ("completed", "cancelled") for t in items)
+
+
+def derive_pet_state(
+    *,
+    busy: bool = False,
+    awaiting_input: bool = False,
+    error: bool = False,
+    celebrate: bool = False,
+    just_completed: bool = False,
+    tool_running: bool = False,
+    reasoning: bool = False,
+) -> PetState:
+    """Resolve the animation state from coarse activity signals.
+
+    Priority (highest first) — only one row can show at a time, so the most
+    salient signal wins:
+
+    1. ``error``          → ``FAILED``  (a tool/turn just failed)
+    2. ``celebrate``      → ``JUMP``    (explicit success beat, e.g. todos done)
+    3. ``just_completed`` → ``WAVE``    (turn finished cleanly / greeting)
+    4. ``awaiting_input`` → ``WAITING`` (blocked on the user — a clarify/approval
+       prompt is open; this outranks the in-flight signals below because the turn
+       is paused on *you*, even though a tool is technically mid-call)
+    5. ``tool_running``   → ``RUN``     (a tool is executing)
+    6. ``reasoning``      → ``REVIEW``  (model is thinking / reading)
+    7. ``busy``           → ``RUN``     (turn in flight, unspecified work)
+    8. otherwise          → ``IDLE``
+    """
+    if error:
+        return PetState.FAILED
+    if celebrate:
+        return PetState.JUMP
+    if just_completed:
+        return PetState.WAVE
+    if awaiting_input:
+        return PetState.WAITING
+    if tool_running:
+        return PetState.RUN
+    if reasoning:
+        return PetState.REVIEW
+    if busy:
+        return PetState.RUN
+    return PetState.IDLE

agent/pet/store.py

@@ -0,0 +1,388 @@
+"""On-disk pet store — install / list / resolve pets.
+
+Pets live under ``get_hermes_home()/pets/<slug>/`` so every profile gets its
+own set (we deliberately do **not** reuse petdex's ``~/.codex/pets`` default —
+that's owned by the petdex npm CLI and isn't profile-aware).  Each installed
+pet directory holds:
+
+    pets/<slug>/
+        pet.json            # {id, displayName, description, spritesheetPath}
+        spritesheet.webp    # (or .png)
+
+The active pet is resolved from the caller-supplied ``display.pet.slug`` config
+value (falling back to the first installed pet), so this module stays free of
+the config loader.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import re
+from dataclasses import dataclass
+from pathlib import Path
+
+from hermes_constants import get_hermes_home
+
+logger = logging.getLogger(__name__)
+
+_DOWNLOAD_TIMEOUT = 60.0
+
+
+class PetStoreError(RuntimeError):
+    """Raised on install/IO failures."""
+
+
+@dataclass(frozen=True)
+class InstalledPet:
+    """A pet present on disk."""
+
+    slug: str
+    display_name: str
+    description: str
+    directory: Path
+    spritesheet: Path
+
+    @property
+    def exists(self) -> bool:
+        return self.spritesheet.is_file()
+
+
+def pets_dir() -> Path:
+    """Return the profile-scoped pets directory (created on demand)."""
+    path = get_hermes_home() / "pets"
+    path.mkdir(parents=True, exist_ok=True)
+    return path
+
+
+def _read_pet_json(directory: Path) -> dict:
+    pet_json = directory / "pet.json"
+    if not pet_json.is_file():
+        return {}
+    try:
+        return json.loads(pet_json.read_text(encoding="utf-8"))
+    except (OSError, ValueError) as exc:
+        logger.debug("unreadable pet.json in %s: %s", directory, exc)
+        return {}
+
+
+def _resolve_spritesheet(directory: Path, meta: dict) -> Path:
+    """Find the spritesheet for a pet dir.
+
+    Honors ``spritesheetPath`` from pet.json, else probes the conventional
+    filenames (``spritesheet.{webp,png}`` and petdex R2's ``sprite.webp``).
+    """
+    declared = str(meta.get("spritesheetPath", "") or "").strip()
+    if declared:
+        candidate = directory / declared
+        if candidate.is_file():
+            return candidate
+    for name in ("spritesheet.webp", "spritesheet.png", "sprite.webp", "sprite.png"):
+        candidate = directory / name
+        if candidate.is_file():
+            return candidate
+    # Default expectation even if missing, so callers get a stable path.
+    return directory / "spritesheet.webp"
+
+
+def load_pet(slug: str) -> InstalledPet | None:
+    """Return the :class:`InstalledPet` for *slug*, or ``None`` if absent."""
+    slug = slug.strip()
+    directory = pets_dir() / slug
+    if not directory.is_dir():
+        return None
+    meta = _read_pet_json(directory)
+    return InstalledPet(
+        slug=slug,
+        display_name=str(meta.get("displayName", "") or slug),
+        description=str(meta.get("description", "") or ""),
+        directory=directory,
+        spritesheet=_resolve_spritesheet(directory, meta),
+    )
+
+
+def installed_pets() -> list[InstalledPet]:
+    """Return every installed pet (dirs containing a usable spritesheet)."""
+    out: list[InstalledPet] = []
+    for child in sorted(pets_dir().iterdir()):
+        if not child.is_dir():
+            continue
+        pet = load_pet(child.name)
+        if pet and pet.exists:
+            out.append(pet)
+    return out
+
+
+def resolve_active_pet(configured_slug: str | None = None) -> InstalledPet | None:
+    """Resolve which pet to display.
+
+    Precedence: the configured slug (``display.pet.slug``) if it's installed,
+    otherwise the first installed pet alphabetically, otherwise ``None``.
+    """
+    if configured_slug:
+        pet = load_pet(configured_slug.strip())
+        if pet and pet.exists:
+            return pet
+    pets = installed_pets()
+    return pets[0] if pets else None
+
+
+def install_pet(slug: str, *, force: bool = False, timeout: float = _DOWNLOAD_TIMEOUT) -> InstalledPet:
+    """Download *slug* from the manifest into the pets directory.
+
+    Idempotent: a fully-installed pet is returned as-is unless *force*.  Raises
+    :class:`PetStoreError` / :class:`~agent.pet.manifest.ManifestError` on
+    failure.
+    """
+    from agent.pet.manifest import find_entry
+
+    slug = slug.strip()
+    existing = load_pet(slug)
+    if existing and existing.exists and not force:
+        return existing
+
+    entry = find_entry(slug, timeout=timeout)
+    if entry is None:
+        raise PetStoreError(f"pet '{slug}' is not in the petdex manifest")
+
+    directory = pets_dir() / slug
+    directory.mkdir(parents=True, exist_ok=True)
+
+    sprite_ext = ".png" if entry.spritesheet_url.lower().split("?")[0].endswith(".png") else ".webp"
+    sprite_path = directory / f"spritesheet{sprite_ext}"
+
+    _download(entry.spritesheet_url, sprite_path, timeout=timeout)
+
+    # Fetch the upstream pet.json if present; otherwise synthesize a minimal
+    # one so the local layout is self-describing.
+    meta: dict = {}
+    if entry.pet_json_url:
+        try:
+            meta = _download_json(entry.pet_json_url, timeout=timeout)
+        except Exception as exc:  # noqa: BLE001 - non-fatal, fall back below
+            logger.debug("pet.json fetch failed for %s: %s", slug, exc)
+    if not isinstance(meta, dict) or not meta:
+        meta = {"id": slug, "displayName": entry.display_name, "description": ""}
+    meta["spritesheetPath"] = sprite_path.name
+    meta.setdefault("id", slug)
+    meta.setdefault("displayName", entry.display_name)
+    (directory / "pet.json").write_text(json.dumps(meta, indent=2), encoding="utf-8")
+
+    pet = load_pet(slug)
+    if pet is None or not pet.exists:
+        raise PetStoreError(f"install of '{slug}' did not produce a spritesheet")
+    return pet
+
+
+def slugify(name: str) -> str:
+    """Lowercase, hyphenate, and strip a display name into a filesystem slug."""
+    slug = re.sub(r"[^a-z0-9]+", "-", (name or "").strip().lower()).strip("-")
+    return slug or "pet"
+
+
+def unique_slug(name: str) -> str:
+    """A :func:`slugify` result that doesn't collide with an existing pet dir."""
+    base = slugify(name)
+    slug = base
+    counter = 2
+    while (pets_dir() / slug).exists():
+        slug = f"{base}-{counter}"
+        counter += 1
+    return slug
+
+
+def _write_spritesheet(source, dest: Path) -> None:
+    """Write *source* (PIL image, bytes, or path) as a lossless WebP at *dest*."""
+    if isinstance(source, (bytes, bytearray)):
+        dest.write_bytes(bytes(source))
+        return
+
+    from PIL import Image
+
+    if isinstance(source, (str, Path)):
+        with Image.open(source) as opened:
+            image = opened.convert("RGBA")
+    else:
+        image = source.convert("RGBA")
+    image.save(dest, format="WEBP", lossless=True, quality=100, method=6, exact=True)
+
+
+def register_local_pet(
+    spritesheet,
+    *,
+    slug: str,
+    display_name: str = "",
+    description: str = "",
+) -> InstalledPet:
+    """Write a locally-generated pet into the store and return it.
+
+    *spritesheet* may be a PIL image, raw WebP/PNG bytes, or a path. The pet
+    appears in :func:`installed_pets` immediately, and because :func:`install_pet`
+    returns an already-on-disk pet before consulting the manifest, it can be
+    adopted (``pet.select`` / ``/pet <slug>``) without a manifest entry.
+    """
+    slug = slugify(slug)
+    directory = pets_dir() / slug
+    directory.mkdir(parents=True, exist_ok=True)
+    sprite_path = directory / "spritesheet.webp"
+    try:
+        _write_spritesheet(spritesheet, sprite_path)
+    except Exception as exc:  # noqa: BLE001 - normalize to one error type
+        raise PetStoreError(f"could not write spritesheet for '{slug}': {exc}") from exc
+
+    meta = {
+        "id": slug,
+        "displayName": display_name or slug,
+        "description": description or "",
+        "spritesheetPath": sprite_path.name,
+        "createdBy": "generator",
+    }
+    (directory / "pet.json").write_text(json.dumps(meta, indent=2), encoding="utf-8")
+
+    pet = load_pet(slug)
+    if pet is None or not pet.exists:
+        raise PetStoreError(f"register of generated pet '{slug}' did not produce a spritesheet")
+    return pet
+
+
+_THUMB_FRAME_W = 192
+_THUMB_FRAME_H = 208
+_THUMB_W = 96  # rendered ~40px; 2x+ keeps it crisp on HiDPI
+
+
+def _thumbs_dir() -> Path:
+    path = pets_dir() / ".thumbs"
+    path.mkdir(parents=True, exist_ok=True)
+    return path
+
+
+def _is_petdex_host(url: str) -> bool:
+    """True only for petdex.dev hosts — bounds server-side fetch (anti-SSRF)."""
+    from urllib.parse import urlparse
+
+    try:
+        host = (urlparse(url).hostname or "").lower()
+    except ValueError:
+        return False
+    return host == "petdex.dev" or host.endswith(".petdex.dev")
+
+
+def thumbnail_png(slug: str, *, source_url: str = "", timeout: float = 30.0) -> bytes | None:
+    """Return a small idle-frame PNG for *slug*, cached on disk.
+
+    Crops the top-left (idle, frame 0) cell of the spritesheet and downsamples
+    it to a thumbnail. Source preference: an installed spritesheet on disk, else
+    *source_url* — but only when it points at petdex (so the gateway never
+    fetches an arbitrary client-supplied URL). Returns ``None`` when there's no
+    usable source or Pillow/network fails; callers render a placeholder.
+
+    Doing this server-side sidesteps the renderer's CSP / R2 hotlink limits that
+    break a direct ``<img src=cdn>`` and lets the result ride the authenticated
+    gateway as a same-origin data URL.
+    """
+    slug = slug.strip()
+    if not slug:
+        return None
+
+    cache = _thumbs_dir() / f"{slug}.png"
+    if cache.is_file():
+        try:
+            return cache.read_bytes()
+        except OSError:
+            pass
+
+    sheet_bytes: bytes | None = None
+    pet = load_pet(slug)
+    if pet and pet.exists:
+        try:
+            sheet_bytes = pet.spritesheet.read_bytes()
+        except OSError:
+            sheet_bytes = None
+
+    if sheet_bytes is None and source_url and _is_petdex_host(source_url):
+        try:
+            import httpx
+
+            resp = httpx.get(
+                source_url,
+                timeout=timeout,
+                follow_redirects=True,
+                headers={"User-Agent": "hermes-agent-petdex"},
+            )
+            resp.raise_for_status()
+            sheet_bytes = resp.content
+        except Exception as exc:  # noqa: BLE001 - cosmetic, degrade to placeholder
+            logger.debug("thumb fetch failed for %s: %s", slug, exc)
+
+    if not sheet_bytes:
+        return None
+
+    try:
+        import io
+
+        from PIL import Image
+
+        with Image.open(io.BytesIO(sheet_bytes)) as im:
+            frame = im.convert("RGBA").crop(
+                (0, 0, min(_THUMB_FRAME_W, im.width), min(_THUMB_FRAME_H, im.height))
+            )
+            height = round(_THUMB_W * _THUMB_FRAME_H / _THUMB_FRAME_W)
+            frame = frame.resize((_THUMB_W, height), Image.NEAREST)
+            buf = io.BytesIO()
+            frame.save(buf, format="PNG")
+            data = buf.getvalue()
+    except Exception as exc:  # noqa: BLE001
+        logger.debug("thumb crop failed for %s: %s", slug, exc)
+        return None
+
+    try:
+        cache.write_bytes(data)
+    except OSError:
+        pass
+    return data
+
+
+def remove_pet(slug: str) -> bool:
+    """Delete an installed pet directory.  Returns True if anything was removed."""
+    import shutil
+
+    directory = pets_dir() / slug.strip()
+    if not directory.is_dir():
+        return False
+    shutil.rmtree(directory, ignore_errors=True)
+    return not directory.exists()
+
+
+def _download(url: str, dest: Path, *, timeout: float) -> None:
+    import httpx
+
+    try:
+        with httpx.stream(
+            "GET",
+            url,
+            timeout=timeout,
+            follow_redirects=True,
+            headers={"User-Agent": "hermes-agent-petdex"},
+        ) as resp:
+            resp.raise_for_status()
+            tmp = dest.with_suffix(dest.suffix + ".part")
+            with tmp.open("wb") as fh:
+                for chunk in resp.iter_bytes():
+                    fh.write(chunk)
+            tmp.replace(dest)
+    except Exception as exc:  # noqa: BLE001
+        raise PetStoreError(f"download failed for {url}: {exc}") from exc
+
+
+def _download_json(url: str, *, timeout: float) -> dict:
+    import httpx
+
+    resp = httpx.get(
+        url,
+        timeout=timeout,
+        follow_redirects=True,
+        headers={"User-Agent": "hermes-agent-petdex"},
+    )
+    resp.raise_for_status()
+    data = resp.json()
+    return data if isinstance(data, dict) else {}

agent/prompt_builder.py

@@ -8,6 +8,7 @@ import json
 import logging
 import os
 import threading
+import contextvars
 from collections import OrderedDict
 from pathlib import Path
 
@@ -489,15 +490,41 @@ PLATFORM_HINTS = {
         "files arrive as downloadable documents. You can also include image "
         "URLs in markdown format ![alt](url) and they will be sent as photos."
     ),
+    "whatsapp_cloud": (
+        "You are on a text messaging communication platform, WhatsApp "
+        "(via Meta's official Business Cloud API). Standard markdown "
+        "(**bold**, ~~strike~~, # headers, [links](url)) is auto-converted "
+        "to WhatsApp's native syntax (*bold*, ~strike~, etc.) — feel free "
+        "to write in markdown. Tables are NOT supported — prefer bullet "
+        "lists or labeled key:value pairs. "
+        "You can send media files natively: include MEDIA:/absolute/path/to/file "
+        "in your response. Images (.jpg, .png) become photo attachments, "
+        "videos (.mp4) play inline, audio (.mp3, .ogg) sends as voice/audio "
+        "messages, other files arrive as documents. Image URLs in markdown "
+        "format ![alt](url) also work. "
+        "IMPORTANT: this platform has a 24-hour conversation window — if the "
+        "user hasn't messaged in 24h, free-form replies are refused by Meta "
+        "(error 131047). This rarely matters for live chat, but is worth "
+        "knowing if you're scheduling a delayed message."
+    ),
     "telegram": (
         "You are on a text messaging communication platform, Telegram. "
-        "Standard markdown is automatically converted to Telegram format. "
+        "Standard Markdown is automatically converted to Telegram formatting. "
         "Supported: **bold**, *italic*, ~~strikethrough~~, ||spoiler||, "
         "`inline code`, ```code blocks```, [links](url), and ## headers. "
-        "Telegram has NO table syntax — prefer bullet lists or labeled "
-        "key: value pairs over pipe tables (any tables you do emit are "
-        "auto-rewritten into row-group bullets, which you can produce "
-        "directly for cleaner output). "
+        "Telegram now supports rich Markdown, so lean into it: whenever it "
+        "makes the answer clearer or easier to scan, actively reach for real "
+        "Markdown tables (pipe `| col | col |` syntax), bullet and numbered "
+        "lists, task lists (`- [ ]` / `- [x]`), headings, nested blockquotes, "
+        "collapsible details, footnotes/references, math/formulas (`$...$`, "
+        "`$$...$$`), underline, subscript/superscript, marked (highlighted) "
+        "text, and anchors. Default to structured formatting over dense "
+        "paragraphs for any comparison, set of steps, key/value summary, or "
+        "tabular data. Prefer real Markdown tables and task lists over "
+        "hand-built bullet substitutes when presenting structured data; these "
+        "degrade gracefully (tables become readable bullet groups) when rich "
+        "rendering is unavailable, but advanced constructs like math and "
+        "collapsible details may render as plain source text in that case. "
         "You can send media files natively: to deliver a file to the user, "
         "include MEDIA:/absolute/path/to/file in your response. Images "
         "(.png, .jpg, .webp) appear as photos, audio (.ogg) sends as voice "
@@ -885,6 +912,22 @@ def build_environment_hints() -> str:
                 f"`uname -a && whoami && pwd`."
             )
 
+    # Hermes desktop GUI — any agent running under the desktop app should know
+    # it. HERMES_DESKTOP marks the backend powering the chat; HERMES_DESKTOP_TERMINAL
+    # marks a hermes launched in the embedded terminal pane. Both set by main.cjs.
+    _truthy = ("1", "true", "yes")
+    _in_desktop = (os.getenv("HERMES_DESKTOP") or "").strip().lower() in _truthy
+    _in_desktop_term = (os.getenv("HERMES_DESKTOP_TERMINAL") or "").strip().lower() in _truthy
+    if _in_desktop or _in_desktop_term:
+        _desktop_hint = "Runtime surface: you're running inside the Hermes desktop GUI app."
+        if _in_desktop_term:
+            _desktop_hint += (
+                " You're in its embedded terminal pane, beside the GUI chat — the user can "
+                "select your output (⌥-drag on macOS, Shift-drag elsewhere) and press "
+                "⌘/Ctrl+L to send it to the chat composer."
+            )
+        hints.append(_desktop_hint)
+
     if is_wsl():
         hints.append(WSL_ENVIRONMENT_HINT)
 
@@ -916,6 +959,52 @@ CONTEXT_TRUNCATE_HEAD_RATIO = 0.7
 CONTEXT_TRUNCATE_TAIL_RATIO = 0.2
 
 
+def _get_context_file_max_chars() -> int:
+    """Return the configured context-file truncation limit.
+
+    ``CONTEXT_FILE_MAX_CHARS`` remains the upstream-compatible default and
+    fallback. Users with larger context windows can raise
+    ``context_file_max_chars`` in config.yaml without patching Hermes.
+    """
+    try:
+        from hermes_cli.config import load_config
+
+        val = load_config().get("context_file_max_chars")
+        if isinstance(val, (int, float)) and val > 0:
+            return int(val)
+    except Exception as e:
+        logger.debug("Could not read context_file_max_chars from config: %s", e)
+    return CONTEXT_FILE_MAX_CHARS
+
+# Collect truncation warnings so the caller (run_agent) can surface them.
+# A ContextVar (not a module-global list) isolates accumulation per thread /
+# per async task, so concurrent gateway-session prompt builds can't drain or
+# clear each other's pending warnings (cross-session leak). Each build runs in
+# its own context, collects its own warnings, and drains them synchronously.
+_truncation_warnings: "contextvars.ContextVar[Optional[list]]" = contextvars.ContextVar(
+    "context_file_truncation_warnings", default=None
+)
+
+
+def _record_truncation_warning(msg: str) -> None:
+    """Append a truncation warning to the current context's accumulator."""
+    warnings = _truncation_warnings.get()
+    if warnings is None:
+        warnings = []
+        _truncation_warnings.set(warnings)
+    warnings.append(msg)
+
+
+def drain_truncation_warnings() -> list:
+    """Return and clear any truncation warnings accumulated in this context."""
+    warnings = _truncation_warnings.get()
+    if not warnings:
+        return []
+    drained = list(warnings)
+    warnings.clear()
+    return drained
+
+
 # =========================================================================
 # Skills prompt cache
 # =========================================================================
@@ -1085,11 +1174,12 @@ def _skill_should_show(
 def build_skills_system_prompt(
     available_tools: "set[str] | None" = None,
     available_toolsets: "set[str] | None" = None,
+    compact_categories: "frozenset[str] | None" = None,
 ) -> str:
     """Build a compact skill index for the system prompt.
 
     Two-layer cache:
-      1. In-process LRU dict keyed by (skills_dir, tools, toolsets)
+      1. In-process LRU dict keyed by (skills_dir, tools, toolsets, hidden)
       2. Disk snapshot (``.skills_prompt_snapshot.json``) validated by
          mtime/size manifest — survives process restarts
 
@@ -1099,6 +1189,12 @@ def build_skills_system_prompt(
     scanned alongside the local ``~/.hermes/skills/`` directory.  External dirs
     are read-only — they appear in the index but new skills are always created
     in the local dir.  Local skills take precedence when names collide.
+
+    ``compact_categories`` (e.g. from the coding posture — see
+    agent/coding_context.py) demotes whole categories to a names-only line in
+    the rendered index. Nothing is ever hidden: every skill name stays
+    visible and loadable via ``skill_view`` / ``skills_list``; only the
+    descriptions are dropped, and a footer note explains the demotion.
     """
     skills_dir = get_skills_dir()
     external_dirs = get_all_skills_dirs()[1:]  # skip local (index 0)
@@ -1115,7 +1211,7 @@ def build_skills_system_prompt(
         or get_session_env("HERMES_SESSION_PLATFORM")
         or ""
     )
-    disabled = get_disabled_skill_names()
+    disabled = get_disabled_skill_names(_platform_hint or None)
     cache_key = (
         str(skills_dir.resolve()),
         tuple(str(d) for d in external_dirs),
@@ -1123,6 +1219,7 @@ def build_skills_system_prompt(
         tuple(sorted(str(ts) for ts in (available_toolsets or set()))),
         _platform_hint,
         tuple(sorted(disabled)),
+        tuple(sorted(compact_categories or ())),
     )
     with _SKILLS_PROMPT_CACHE_LOCK:
         cached = _SKILLS_PROMPT_CACHE.get(cache_key)
@@ -1256,18 +1353,44 @@ def build_skills_system_prompt(
             except Exception as e:
                 logger.debug("Could not read external skill description %s: %s", desc_file, e)
 
+    # Posture-driven category demotion (e.g. non-coding skills while pairing
+    # on code). Demoted categories stay in the index as a single names-only
+    # line — descriptions are dropped to cut noise, but every skill name
+    # remains visible so memory-anchored recall ("load <name>") keeps working.
+    # NEVER remove entries entirely: agent-created skills are the model's
+    # project memory, and models don't reach for skills_list to rediscover
+    # what the index stops showing them. Match on the top-level category
+    # segment so nested categories ("social-media/twitter") are demoted with
+    # their parent.
+    demoted = frozenset(
+        cat for cat in skills_by_category
+        if cat.split("/", 1)[0] in (compact_categories or frozenset())
+    )
+
+    hidden_note = ""
+    if demoted:
+        hidden_note = (
+            "\n(Categories marked [names only] are outside the current coding "
+            "context, so their descriptions are omitted — the skills work "
+            "normally and load with skill_view(name) as usual.)"
+        )
+
     if not skills_by_category:
         result = ""
     else:
         index_lines = []
         for category in sorted(skills_by_category.keys()):
+            # Deduplicate and sort skills within each category
+            seen = set()
+            if category in demoted:
+                names = sorted({name for name, _ in skills_by_category[category]})
+                index_lines.append(f"  {category} [names only]: {', '.join(names)}")
+                continue
             cat_desc = category_descriptions.get(category, "")
             if cat_desc:
                 index_lines.append(f"  {category}: {cat_desc}")
             else:
                 index_lines.append(f"  {category}:")
-            # Deduplicate and sort skills within each category
-            seen = set()
             for name, desc in sorted(skills_by_category[category], key=lambda x: x[0]):
                 if name in seen:
                     continue
@@ -1304,6 +1427,7 @@ def build_skills_system_prompt(
             "</available_skills>\n"
             "\n"
             "Only proceed without loading a skill if genuinely none are relevant to the task."
+            + hidden_note
         )
 
     # ── Store in LRU cache ────────────────────────────────────────────
@@ -1367,13 +1491,13 @@ def build_nous_subscription_prompt(valid_tool_names: "set[str] | None" = None) -
 
     lines = [
         "# Nous Subscription",
-        "Nous subscription includes managed web tools (Firecrawl), image generation (FAL), OpenAI TTS, and browser automation (Browser Use) by default. Modal execution is optional.",
+        "Nous subscription includes managed web tools (Firecrawl), image generation (FAL), OpenAI TTS, OpenAI Whisper STT, and browser automation (Browser Use) by default. Modal execution is optional.",
         "Current capability status:",
     ]
     lines.extend(_status_line(feature) for feature in features.items())
     lines.extend(
         [
-            "When a Nous-managed feature is active, do not ask the user for Firecrawl, FAL, OpenAI TTS, or Browser-Use API keys.",
+            "When a Nous-managed feature is active, do not ask the user for Firecrawl, FAL, OpenAI TTS, OpenAI Whisper, or Browser-Use API keys.",
             "If the user is not subscribed and asks for a capability that Nous subscription would unlock or simplify, suggest Nous subscription as one option alongside direct setup or local alternatives.",
             "Do not mention subscription unless the user asks about it or it directly solves the current missing capability.",
             "Useful commands: hermes setup, hermes setup tools, hermes setup terminal, hermes status.",
@@ -1386,10 +1510,19 @@ def build_nous_subscription_prompt(valid_tool_names: "set[str] | None" = None) -
 # Context files (SOUL.md, AGENTS.md, .cursorrules)
 # =========================================================================
 
-def _truncate_content(content: str, filename: str, max_chars: int = CONTEXT_FILE_MAX_CHARS) -> str:
+def _truncate_content(content: str, filename: str, max_chars: Optional[int] = None) -> str:
     """Head/tail truncation with a marker in the middle."""
+    if max_chars is None:
+        max_chars = _get_context_file_max_chars()
     if len(content) <= max_chars:
         return content
+    msg = (
+        f"⚠️  Context file {filename} TRUNCATED: "
+        f"{len(content)} chars exceeds limit of {max_chars} — "
+        f"increase context_file_max_chars or trim the file!"
+    )
+    logger.warning(msg)
+    _record_truncation_warning(msg)
     head_chars = int(max_chars * CONTEXT_TRUNCATE_HEAD_RATIO)
     tail_chars = int(max_chars * CONTEXT_TRUNCATE_TAIL_RATIO)
     head = content[:head_chars]

agent/redact.py

@@ -104,6 +104,7 @@ _PREFIX_PATTERNS = [
     r"mem0_[A-Za-z0-9]{10,}",           # Mem0 Platform API key
     r"brv_[A-Za-z0-9]{10,}",            # ByteRover API key
     r"xai-[A-Za-z0-9]{30,}",            # xAI (Grok) API key
+    r"ntn_[A-Za-z0-9]{10,}",            # Notion internal integration token
 ]
 
 # ENV assignment patterns: KEY=value where KEY contains a secret-like name

agent/secret_sources/bitwarden.py

@@ -274,6 +274,7 @@ def _platform_asset_name() -> str:
                 capture_output=True,
                 text=True,
                 timeout=2,
+                stdin=subprocess.DEVNULL,
             )
             if "musl" in (res.stdout + res.stderr).lower():
                 libc = "musl"
@@ -525,6 +526,7 @@ def _run_bws_list(
             capture_output=True,
             text=True,
             timeout=_BWS_RUN_TIMEOUT,
+            stdin=subprocess.DEVNULL,
         )
     except subprocess.TimeoutExpired as exc:
         raise RuntimeError(

agent/skill_commands.py

@@ -26,6 +26,91 @@ _skill_commands_platform: Optional[str] = None
 _SKILL_INVALID_CHARS = re.compile(r"[^a-z0-9-]")
 _SKILL_MULTI_HYPHEN = re.compile(r"-{2,}")
 
+# ---------------------------------------------------------------------------
+# Skill-scaffolding markers and the canonical extractor.
+#
+# When a user invokes a /skill (or /bundle), Hermes expands the turn into a
+# model-facing message that embeds the full skill body plus scaffolding. That
+# expanded text is what flows into the agent loop — and into memory providers
+# via MemoryManager. Providers that store or embed the raw user turn (mem0,
+# openviking, hindsight, retaindb, byterover, honcho, supermemory) would
+# otherwise capture the entire skill body instead of what the user actually
+# asked. ``extract_user_instruction_from_skill_message`` recovers just the
+# user's instruction so memory stays clean.
+#
+# These markers MUST stay byte-identical to the builders below
+# (``_build_skill_message`` here, ``build_bundle_invocation_message`` in
+# agent/skill_bundles.py). They are co-located with the single-skill builder
+# on purpose, and the bundle markers are asserted against the bundle builder in
+# tests/openviking_plugin/test_openviking.py::test_skill_markers_match_hermes_scaffolding.
+# ---------------------------------------------------------------------------
+_SKILL_INVOCATION_PREFIX = "[IMPORTANT: The user has invoked the "
+_SINGLE_SKILL_MARKER = "The full skill content is loaded below.]"
+_SINGLE_SKILL_INSTRUCTION = (
+    "The user has provided the following instruction alongside the skill invocation: "
+)
+_RUNTIME_NOTE = "\n\n[Runtime note:"
+_BUNDLE_MARKER = " skill bundle,"
+_BUNDLE_USER_INSTRUCTION = "\nUser instruction: "
+_BUNDLE_FIRST_SKILL_BLOCK = "\n\n[Loaded as part of the "
+
+
+def extract_user_instruction_from_skill_message(content: Any) -> Optional[str]:
+    """Recover the user's instruction from a slash-skill-expanded turn.
+
+    Returns:
+        - The original string unchanged when it is NOT skill scaffolding
+          (a normal user message passes straight through).
+        - The extracted user instruction when the scaffolding carried one.
+        - ``None`` when the content is skill scaffolding with no user
+          instruction (i.e. a bare ``/skill`` invocation). Callers that feed
+          memory providers should skip the turn in that case — there is no
+          user content worth storing.
+    """
+    if not isinstance(content, str):
+        return None
+
+    if not content.startswith(_SKILL_INVOCATION_PREFIX):
+        return content
+
+    if _BUNDLE_MARKER in content:
+        return _extract_bundle_user_instruction(content)
+
+    if _SINGLE_SKILL_MARKER in content:
+        return _extract_single_skill_user_instruction(content)
+
+    return None
+
+
+def _extract_single_skill_user_instruction(message: str) -> Optional[str]:
+    # Single-skill format appends the user instruction after the skill body, so
+    # the last occurrence is the user-provided one; the body may quote this text.
+    marker_idx = message.rfind(_SINGLE_SKILL_INSTRUCTION)
+    if marker_idx < 0:
+        return None
+
+    instruction = message[marker_idx + len(_SINGLE_SKILL_INSTRUCTION):]
+    runtime_idx = instruction.find(_RUNTIME_NOTE)
+    if runtime_idx >= 0:
+        instruction = instruction[:runtime_idx]
+    instruction = instruction.strip()
+    return instruction or None
+
+
+def _extract_bundle_user_instruction(message: str) -> Optional[str]:
+    # Bundle format puts the user instruction before the loaded skills, so the
+    # first occurrence is the user-provided one.
+    marker_idx = message.find(_BUNDLE_USER_INSTRUCTION)
+    if marker_idx < 0:
+        return None
+
+    instruction = message[marker_idx + len(_BUNDLE_USER_INSTRUCTION):]
+    first_skill_idx = instruction.find(_BUNDLE_FIRST_SKILL_BLOCK)
+    if first_skill_idx >= 0:
+        instruction = instruction[:first_skill_idx]
+    instruction = instruction.strip()
+    return instruction or None
+
 
 def _resolve_skill_commands_platform() -> Optional[str]:
     """Return the current platform scope used for disabled-skill filtering.

agent/skill_preprocessing.py

@@ -74,6 +74,7 @@ def run_inline_shell(command: str, cwd: Path | None, timeout: int) -> str:
             text=True,
             timeout=max(1, int(timeout)),
             check=False,
+            stdin=subprocess.DEVNULL,
         )
     except subprocess.TimeoutExpired:
         return f"[inline-shell timeout after {timeout}s: {command}]"

agent/skill_utils.py

@@ -43,14 +43,20 @@ EXCLUDED_SKILL_DIRS = frozenset(
     )
 )
 
+# Supporting files live inside a skill package and are loaded explicitly via
+# skill_view(skill, file_path=...). They are not standalone skills and must not
+# be scanned for active SKILL.md/DESCRIPTION.md entries, even if a Curator or
+# archive workflow preserves a complete old skill package under references/.
+SKILL_SUPPORT_DIRS = frozenset(("references", "templates", "assets", "scripts"))
+
 
 def is_excluded_skill_path(path) -> bool:
-    """True if any component of *path* is in EXCLUDED_SKILL_DIRS.
+    """True if *path* should be skipped by active skill scanners.
 
-    Use this on every SKILL.md path produced by ``rglob`` to prune
-    dependency, virtualenv, VCS, and cache directories. Centralising the
-    check here keeps every skill-scanning site in sync with the shared
-    exclusion set.
+    Use this on every ``SKILL.md`` path produced by direct ``rglob`` scans to
+    prune dependency, virtualenv, VCS, cache, and progressive-disclosure
+    support-package paths. Centralising the check here keeps every
+    skill-scanning site in sync with the shared exclusion set.
 
     Accepts a Path or string.
     """
@@ -59,7 +65,36 @@ def is_excluded_skill_path(path) -> bool:
     except AttributeError:
         from pathlib import PurePath
         parts = PurePath(str(path)).parts
-    return any(part in EXCLUDED_SKILL_DIRS for part in parts)
+    return any(part in EXCLUDED_SKILL_DIRS for part in parts) or is_skill_support_path(
+        path
+    )
+
+
+def is_skill_support_path(path) -> bool:
+    """True if *path* is under a support dir of an actual skill root.
+
+    ``references/``, ``templates/``, ``assets/``, and ``scripts/`` are
+    progressive-disclosure support areas when they sit directly inside a skill
+    directory containing ``SKILL.md``. They are not active discovery roots for
+    standalone skills. A preserved package such as
+    ``some-skill/references/old-skill-package/SKILL.md`` is documentation data
+    unless the caller explicitly loads it via ``file_path``.
+
+    Legitimate categories or skill names such as ``skills/scripts/foo`` remain
+    discoverable because their ``scripts`` component is not directly under a
+    directory that contains ``SKILL.md``.
+    """
+    path_obj = path if isinstance(path, Path) else Path(str(path))
+    parts = path_obj.parts
+    # Last component may be a file or candidate skill directory name. Only
+    # components before the leaf can be containing support directories.
+    for idx, part in enumerate(parts[:-1]):
+        if part not in SKILL_SUPPORT_DIRS or idx == 0:
+            continue
+        skill_root = Path(*parts[:idx])
+        if (skill_root / "SKILL.md").exists():
+            return True
+    return False
 
 
 # ── Lazy YAML loader ─────────────────────────────────────────────────────
@@ -272,27 +307,65 @@ def skill_matches_environment(frontmatter: Dict[str, Any]) -> bool:
 # ── Disabled skills ───────────────────────────────────────────────────────
 
 
+_RAW_CONFIG_CACHE: Dict[Tuple[str, int, int], Dict[str, Any]] = {}
+
+
+def _raw_config_cache_clear() -> None:
+    """Test hook — drop the shared raw config cache."""
+    _RAW_CONFIG_CACHE.clear()
+
+
+def _load_raw_config() -> Dict[str, Any]:
+    """Read config.yaml with a shared mtime+size keyed cache.
+
+    This module intentionally avoids importing ``hermes_cli.config`` on the
+    skill prompt/build path. A tiny local cache gives the same repeated-read
+    win without pulling the heavier CLI config stack into startup.
+    """
+    config_path = get_config_path()
+    if not config_path.exists():
+        return {}
+    try:
+        stat = config_path.stat()
+        cache_key = (str(config_path), stat.st_mtime_ns, stat.st_size)
+    except OSError:
+        cache_key = None
+
+    if cache_key is not None:
+        cached = _RAW_CONFIG_CACHE.get(cache_key)
+        if cached is not None:
+            return cached
+
+    try:
+        parsed = yaml_load(config_path.read_text(encoding="utf-8"))
+    except Exception as e:
+        logger.debug("Could not read skill config %s: %s", config_path, e)
+        return {}
+    if not isinstance(parsed, dict):
+        return {}
+
+    if cache_key is not None:
+        _RAW_CONFIG_CACHE.clear()
+        _RAW_CONFIG_CACHE[cache_key] = parsed
+    return parsed
+
+
 def get_disabled_skill_names(platform: str | None = None) -> Set[str]:
     """Read disabled skill names from config.yaml.
 
     Args:
         platform: Explicit platform name (e.g. ``"telegram"``).  When
             *None*, resolves from ``HERMES_PLATFORM`` or
-            ``HERMES_SESSION_PLATFORM`` env vars.  Falls back to the
-            global disabled list when no platform is determined.
+            ``HERMES_SESSION_PLATFORM`` env vars.  Returns the global
+            disabled list, unioned with the platform-specific list when a
+            platform is resolved (a globally-disabled skill stays disabled
+            on every platform).
 
     Reads the config file directly (no CLI config imports) to stay
     lightweight.
     """
-    config_path = get_config_path()
-    if not config_path.exists():
-        return set()
-    try:
-        parsed = yaml_load(config_path.read_text(encoding="utf-8"))
-    except Exception as e:
-        logger.debug("Could not read skill config %s: %s", config_path, e)
-        return set()
-    if not isinstance(parsed, dict):
+    parsed = _load_raw_config()
+    if not parsed:
         return set()
 
     skills_cfg = parsed.get("skills")
@@ -305,13 +378,14 @@ def get_disabled_skill_names(platform: str | None = None) -> Set[str]:
         or os.getenv("HERMES_PLATFORM")
         or get_session_env("HERMES_SESSION_PLATFORM")
     )
+    global_disabled = _normalize_string_set(skills_cfg.get("disabled"))
     if resolved_platform:
         platform_disabled = (skills_cfg.get("platform_disabled") or {}).get(
             resolved_platform
         )
         if platform_disabled is not None:
-            return _normalize_string_set(platform_disabled)
-    return _normalize_string_set(skills_cfg.get("disabled"))
+            return global_disabled | _normalize_string_set(platform_disabled)
+    return global_disabled
 
 
 def _normalize_string_set(values) -> Set[str]:
@@ -336,6 +410,7 @@ _EXTERNAL_DIRS_CACHE: Dict[Tuple[str, int], List[Path]] = {}
 def _external_dirs_cache_clear() -> None:
     """Test hook — drop the in-process cache."""
     _EXTERNAL_DIRS_CACHE.clear()
+    _raw_config_cache_clear()
 
 
 def get_external_skills_dirs() -> List[Path]:
@@ -368,11 +443,8 @@ def get_external_skills_dirs() -> List[Path]:
             # Return a copy so callers can't mutate the cached list.
             return list(cached)
 
-    try:
-        parsed = yaml_load(config_path.read_text(encoding="utf-8"))
-    except Exception:
-        return []
-    if not isinstance(parsed, dict):
+    parsed = _load_raw_config()
+    if not parsed:
         return []
 
     skills_cfg = parsed.get("skills")
@@ -584,15 +656,7 @@ def resolve_skill_config_values(
     current values (or the declared default if the key isn't set).
     Path values are expanded via ``os.path.expanduser``.
     """
-    config_path = get_config_path()
-    config: Dict[str, Any] = {}
-    if config_path.exists():
-        try:
-            parsed = yaml_load(config_path.read_text(encoding="utf-8"))
-            if isinstance(parsed, dict):
-                config = parsed
-        except Exception:
-            pass
+    config = _load_raw_config()
 
     resolved: Dict[str, Any] = {}
     for var in config_vars:
@@ -632,12 +696,21 @@ def extract_skill_description(frontmatter: Dict[str, Any]) -> str:
 def iter_skill_index_files(skills_dir: Path, filename: str):
     """Walk skills_dir yielding sorted paths matching *filename*.
 
-    Excludes Hermes metadata, VCS, virtualenv/dependency, and cache
-    directories so dependencies cannot register nested skills.
+    Excludes Hermes metadata, VCS, virtualenv/dependency, cache, and skill
+    support directories. Support directories (references/templates/assets/
+    scripts) can contain arbitrary markdown and even archived package
+    ``SKILL.md`` files, but they are progressive-disclosure data loaded through
+    ``skill_view(..., file_path=...)`` rather than active skill roots.
     """
     matches = []
     for root, dirs, files in os.walk(skills_dir, followlinks=True):
-        dirs[:] = [d for d in dirs if d not in EXCLUDED_SKILL_DIRS]
+        has_skill_md = "SKILL.md" in files
+        dirs[:] = [
+            d
+            for d in dirs
+            if d not in EXCLUDED_SKILL_DIRS
+            and not (has_skill_md and d in SKILL_SUPPORT_DIRS)
+        ]
         if filename in files:
             matches.append(Path(root) / filename)
     for path in sorted(matches, key=lambda p: str(p.relative_to(skills_dir))):

agent/ssl_guard.py

@@ -0,0 +1,94 @@
+"""Preventive SSL CA certificate checks for Hermes Agent.
+
+This module catches broken CA bundle paths before OpenAI/httpx turns them into
+opaque ``FileNotFoundError: [Errno 2] No such file or directory`` failures.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import ssl
+from pathlib import Path
+
+from agent.errors import SSLConfigurationError
+
+logger = logging.getLogger(__name__)
+
+_CA_BUNDLE_ENV_VARS = (
+    "HERMES_CA_BUNDLE",
+    "SSL_CERT_FILE",
+    "REQUESTS_CA_BUNDLE",
+    "CURL_CA_BUNDLE",
+)
+
+_SKIP_VALUES = {"1", "true", "yes", "on"}
+
+
+def _skip_ssl_guard_enabled() -> bool:
+    return os.getenv("HERMES_SKIP_SSL_GUARD", "").strip().lower() in _SKIP_VALUES
+
+
+def _repair_hint() -> str:
+    return (
+        "Repair: python -m pip install --force-reinstall certifi openai httpx\n"
+        "If you configured a custom corporate CA bundle, fix or unset the "
+        "broken CA bundle environment variable."
+    )
+
+
+def _ssl_err(message: str) -> SSLConfigurationError:
+    """Create a consistent, user-actionable SSL configuration error."""
+    return SSLConfigurationError(f"{message}\n{_repair_hint()}")
+
+
+def _validate_bundle_path(label: str, value: str, *, require_substantial: bool = False) -> None:
+    path = Path(value).expanduser()
+    if not path.exists():
+        raise _ssl_err(f"{label} points to a missing CA bundle: {value}")
+    if not path.is_file():
+        raise _ssl_err(f"{label} does not point to a CA bundle file: {value}")
+    if require_substantial and path.stat().st_size < 1024:
+        raise _ssl_err(f"{label} at {value} appears corrupted (too small)")
+    try:
+        ctx = ssl.create_default_context(cafile=str(path))
+    except Exception as exc:
+        raise _ssl_err(f"{label} CA bundle at {value} cannot be loaded: {exc}") from exc
+    if not ctx.get_ca_certs():
+        raise _ssl_err(f"{label} CA bundle at {value} did not load any certificates")
+
+
+def verify_ca_bundle() -> None:
+    """Verify configured and bundled CA certificates are present and loadable.
+
+    Raises:
+        SSLConfigurationError: If an explicit CA-bundle environment variable
+            points at a bad path, or if certifi's bundled ``cacert.pem`` is
+            missing/corrupt.
+    """
+    if _skip_ssl_guard_enabled():
+        logger.debug("SSL CA bundle guard skipped via HERMES_SKIP_SSL_GUARD")
+        return
+
+    for env_var in _CA_BUNDLE_ENV_VARS:
+        value = os.getenv(env_var)
+        if value:
+            _validate_bundle_path(env_var, value)
+
+    try:
+        import certifi
+    except Exception as exc:
+        raise _ssl_err(f"certifi is not importable: {exc}") from exc
+
+    ca_bundle = str(certifi.where())
+    _validate_bundle_path("certifi", ca_bundle, require_substantial=True)
+
+
+def verify_ca_bundle_with_fallback() -> None:
+    """Backward-compatible wrapper for older call sites.
+
+    The old PR name mentioned a platform fallback, but allowing startup with a
+    broken certifi bundle still leaves httpx/OpenAI and requests call sites
+    failing later. Keep the wrapper name but enforce the same check.
+    """
+    verify_ca_bundle()

agent/system_prompt.py

@@ -40,6 +40,7 @@ from agent.prompt_builder import (
     TASK_COMPLETION_GUIDANCE,
     TOOL_USE_ENFORCEMENT_GUIDANCE,
     TOOL_USE_ENFORCEMENT_MODELS,
+    drain_truncation_warnings,
 )
 from agent.runtime_cwd import resolve_context_cwd
 
@@ -191,9 +192,23 @@ def build_system_prompt_parts(agent: Any, system_message: Optional[str] = None)
             )
             if toolset
         }
+        # Focus mode (opt-in) demotes non-coding skill categories to
+        # names-only in the index (never hidden — skill_view/skills_list
+        # reach everything, and every name stays visible for recall). The
+        # default coding posture leaves the index untouched.
+        _compact_cats = frozenset()
+        try:
+            from agent.coding_context import coding_compact_skill_categories
+
+            _compact_cats = coding_compact_skill_categories(
+                platform=agent.platform, cwd=resolve_context_cwd()
+            )
+        except Exception:
+            _compact_cats = frozenset()
         skills_prompt = _r.build_skills_system_prompt(
             available_tools=agent.valid_tool_names,
             available_toolsets=avail_toolsets,
+            compact_categories=_compact_cats or None,
         )
     else:
         skills_prompt = ""
@@ -221,6 +236,26 @@ def build_system_prompt_parts(agent: Any, system_message: Optional[str] = None)
     if _env_hints:
         stable_parts.append(_env_hints)
 
+    # Coding posture (base Hermes, any interactive coding surface in a code
+    # workspace — see agent/coding_context.py). The operating brief + the live
+    # git/workspace snapshot are built once here and cached for the session;
+    # the snapshot is never re-probed per turn (that would break the prompt
+    # cache), so the brief tells the model to re-check git before relying on it.
+    if agent.valid_tool_names:
+        try:
+            from agent.coding_context import coding_system_blocks
+
+            stable_parts.extend(
+                coding_system_blocks(
+                    platform=agent.platform,
+                    cwd=resolve_context_cwd(),
+                    model=agent.model,
+                )
+            )
+        except Exception:
+            # Coding-context probing must never block prompt build.
+            pass
+
     # Local Python toolchain probe — names python/pip/uv/PEP-668 state when
     # something is non-default so the model can pick the right install
     # strategy without discovering by failure.  Emits a single line; emits
@@ -366,7 +401,14 @@ def build_system_prompt(agent: Any, system_message: Optional[str] = None) -> str
     warm across turns.
     """
     parts = build_system_prompt_parts(agent, system_message=system_message)
-    return "\n\n".join(p for p in (parts["stable"], parts["context"], parts["volatile"]) if p)
+    joined = "\n\n".join(p for p in (parts["stable"], parts["context"], parts["volatile"]) if p)
+
+    # Surface context-file truncation warnings through the normal agent status
+    # channel so gateway/CLI users see them in chat instead of only in logs.
+    for warning in drain_truncation_warnings():
+        agent._emit_status(warning)
+
+    return joined
 
 
 def invalidate_system_prompt(agent: Any) -> None:

agent/tool_executor.py

@@ -417,7 +417,7 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
 
     # ── Logging / callbacks ──────────────────────────────────────────
     tool_names_str = ", ".join(name for _, name, _, _, _, _ in parsed_calls)
-    if not agent.quiet_mode:
+    if not agent.quiet_mode and getattr(agent, "tool_progress_mode", "all") != "off":
         print(f"  ⚡ Concurrent: {num_tools} tool calls — {tool_names_str}")
         for i, (tc, name, args, middleware_trace, block_result, blocked_by_guardrail) in enumerate(parsed_calls, 1):
             args_str = json.dumps(args, ensure_ascii=False)
@@ -702,7 +702,7 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
         if agent._should_emit_quiet_tool_messages():
             cute_msg = _get_cute_tool_message_impl(name, args, tool_duration, result=function_result)
             agent._safe_print(f"  {cute_msg}")
-        elif getattr(agent, "tool_progress_mode", "all") != "off":
+        elif not agent.quiet_mode and getattr(agent, "tool_progress_mode", "all") != "off":
             _preview_str = _multimodal_text_summary(function_result)
             if agent.verbose_logging:
                 print(f"  ✅ Tool {i+1} completed in {tool_duration:.2f}s")
@@ -866,7 +866,7 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
         elif function_name == "skill_manage":
             agent._iters_since_skill = 0
 
-        if not agent.quiet_mode:
+        if not agent.quiet_mode and getattr(agent, "tool_progress_mode", "all") != "off":
             args_str = json.dumps(function_args, ensure_ascii=False)
             if agent.verbose_logging:
                 print(f"  📞 Tool {i}: {function_name}({list(function_args.keys())})")
@@ -1065,6 +1065,25 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
             tool_duration = time.time() - tool_start_time
             if agent._should_emit_quiet_tool_messages():
                 agent._vprint(f"  {_get_cute_tool_message_impl('clarify', function_args, tool_duration, result=function_result)}")
+        elif function_name == "read_terminal":
+            def _execute(next_args: dict) -> Any:
+                from tools.read_terminal_tool import read_terminal_tool as _read_terminal_tool
+                return _read_terminal_tool(
+                    start_line=next_args.get("start_line"),
+                    count=next_args.get("count"),
+                    callback=getattr(agent, "read_terminal_callback", None),
+                )
+            function_result, function_args = _run_agent_tool_execution_middleware(
+                agent,
+                function_name=function_name,
+                function_args=function_args,
+                effective_task_id=effective_task_id,
+                tool_call_id=getattr(tool_call, "id", "") or "",
+                execute=_execute,
+            )
+            tool_duration = time.time() - tool_start_time
+            if agent._should_emit_quiet_tool_messages():
+                agent._vprint(f"  {_get_cute_tool_message_impl('read_terminal', function_args, tool_duration, result=function_result)}")
         elif function_name == "delegate_task":
             tasks_arg = function_args.get("tasks")
             if tasks_arg and isinstance(tasks_arg, list):
@@ -1365,7 +1384,7 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
         # entire batch.  The model sees it on the next API iteration.
         agent._apply_pending_steer_to_tool_results(messages, 1)
 
-        if not agent.quiet_mode:
+        if not agent.quiet_mode and getattr(agent, "tool_progress_mode", "all") != "off":
             if agent.verbose_logging:
                 print(f"  ✅ Tool {i} completed in {tool_duration:.2f}s")
                 print(agent._wrap_verbose("Result: ", function_result))

agent/transports/anthropic.py

@@ -84,7 +84,7 @@ class AnthropicTransport(ProviderTransport):
         to OpenAI finish_reason, and collects reasoning_details in provider_data.
         """
         import json
-        from agent.anthropic_adapter import _to_plain_data
+        from agent.anthropic_adapter import _to_plain_data, _sanitize_replay_block
         from agent.transports.types import ToolCall
 
         strip_tool_prefix = kwargs.get("strip_tool_prefix", False)
@@ -94,14 +94,40 @@ class AnthropicTransport(ProviderTransport):
         reasoning_parts = []
         reasoning_details = []
         tool_calls = []
+        # Verbatim, order-preserving copy of every content block in the turn.
+        # Anthropic signs each thinking block against the turn content that
+        # PRECEDES it at its position; when a turn interleaves thinking and
+        # tool_use (adaptive/interleaved thinking, Claude 4.6+), the parallel
+        # reasoning_details + tool_calls lists below lose that cross-type
+        # ordering. Replaying the latest assistant message in the wrong order
+        # invalidates the signatures -> HTTP 400 "thinking ... blocks in the
+        # latest assistant message cannot be modified". Preserve the exact
+        # block sequence here so the adapter can replay it unchanged. See
+        # tests/agent/test_anthropic_thinking_block_order.py.
+        ordered_blocks = []
 
         for block in response.content:
+            block_dict = _to_plain_data(block)
+            clean_block = None
+            if isinstance(block_dict, dict):
+                # Sanitize at capture so output-only SDK fields (parsed_output,
+                # caller, citations=None, …) never persist to state.db and leak
+                # back as request input on replay → HTTP 400 "Extra inputs are
+                # not permitted". Defence-in-depth with the replay-side sanitize.
+                clean_block = _sanitize_replay_block(block_dict)
+                if clean_block is not None:
+                    ordered_blocks.append(clean_block)
             if block.type == "text":
                 text_parts.append(block.text)
-            elif block.type == "thinking":
-                reasoning_parts.append(block.thinking)
-                block_dict = _to_plain_data(block)
-                if isinstance(block_dict, dict):
+            elif block.type in ("thinking", "redacted_thinking"):
+                if block.type == "thinking":
+                    reasoning_parts.append(block.thinking)
+                # Use the sanitized block (clean_block) for reasoning_details too,
+                # since _extract_preserved_thinking_blocks replays these on the
+                # non-ordered path. Falls back to raw only if sanitize dropped it.
+                if isinstance(clean_block, dict):
+                    reasoning_details.append(clean_block)
+                elif isinstance(block_dict, dict):
                     reasoning_details.append(block_dict)
             elif block.type == "tool_use":
                 name = block.name
@@ -130,6 +156,23 @@ class AnthropicTransport(ProviderTransport):
         provider_data = {}
         if reasoning_details:
             provider_data["reasoning_details"] = reasoning_details
+        # Only worth carrying the ordered-blocks channel when the turn
+        # actually interleaves signed thinking with tool_use — that's the
+        # only shape the parallel lists reconstruct incorrectly. A turn that
+        # is purely text, or thinking-then-tools with a single leading
+        # thinking block, replays correctly without it.
+        _has_signed_thinking = any(
+            isinstance(b, dict)
+            and b.get("type") in ("thinking", "redacted_thinking")
+            and (b.get("signature") or b.get("data"))
+            for b in ordered_blocks
+        )
+        _has_tool_use = any(
+            isinstance(b, dict) and b.get("type") == "tool_use"
+            for b in ordered_blocks
+        )
+        if _has_signed_thinking and _has_tool_use:
+            provider_data["anthropic_content_blocks"] = ordered_blocks
 
         return NormalizedResponse(
             content="\n".join(text_parts) if text_parts else None,
@@ -143,10 +186,21 @@ class AnthropicTransport(ProviderTransport):
     def validate_response(self, response: Any) -> bool:
         """Check Anthropic response structure is valid.
 
-        An empty content list is legitimate when ``stop_reason == "end_turn"``
-        — the model's canonical way of signalling "nothing more to add" after
-        a tool turn that already delivered the user-facing text. Treating it
-        as invalid falsely retries a completed response.
+        An empty content list is legitimate for terminal stop reasons that
+        carry no text payload:
+
+        - ``end_turn`` — the model's canonical "nothing more to add" after a
+          tool turn that already delivered the user-facing text.
+        - ``refusal`` — the model declined to respond (Claude 4.5+). The
+          Messages API returns an empty ``content`` list with this stop
+          reason. Treating it as invalid sends a deterministic refusal into
+          the invalid-response retry loop, which reproduces the refusal on
+          every attempt and surfaces a misleading "rate limited / invalid
+          response" error instead of the refusal. ``normalize_response`` maps
+          ``refusal`` → ``content_filter`` so the agent loop's refusal handler
+          can surface it.
+
+        Treating either as invalid falsely retries a completed response.
         """
         if response is None:
             return False
@@ -154,7 +208,7 @@ class AnthropicTransport(ProviderTransport):
         if not isinstance(content_blocks, list):
             return False
         if not content_blocks:
-            return getattr(response, "stop_reason", None) == "end_turn"
+            return getattr(response, "stop_reason", None) in {"end_turn", "refusal"}
         return True
 
     def extract_cache_stats(self, response: Any) -> Optional[Dict[str, int]]:

agent/transports/chat_completions.py

@@ -531,6 +531,7 @@ class ChatCompletionsTransport(ProviderTransport):
                 supports_reasoning=params.get("supports_reasoning", False),
                 qwen_session_metadata=params.get("qwen_session_metadata"),
                 model=model,
+                base_url=params.get("base_url"),
                 ollama_num_ctx=params.get("ollama_num_ctx"),
                 session_id=params.get("session_id"),
             )
@@ -664,8 +665,42 @@ class ChatCompletionsTransport(ProviderTransport):
         if rd:
             provider_data["reasoning_details"] = rd
 
+        # OpenAI structured-refusal field. When a model declines, the SDK
+        # populates ``message.refusal`` with the explanation and leaves
+        # ``content`` empty. OpenAI-compatible proxies that front Anthropic /
+        # Bedrock (e.g. Nous Portal) surface a Claude refusal this way — or via
+        # ``finish_reason="content_filter"`` — instead of the native
+        # ``stop_reason="refusal"``. Without capturing it the refusal looks
+        # like an empty response, so the agent loop retries a deterministic
+        # refusal three times and gives up with "no content after retries".
+        # Promote it to content + a ``content_filter`` finish reason so the
+        # loop's refusal handler surfaces it clearly and stops. ``refusal`` is
+        # ``None`` for normal responses, so this is a no-op in the common case.
+        content = msg.content
+        refusal = getattr(msg, "refusal", None)
+        if refusal is None and hasattr(msg, "model_extra"):
+            _msg_extra = getattr(msg, "model_extra", None) or {}
+            if isinstance(_msg_extra, dict):
+                refusal = _msg_extra.get("refusal")
+        if isinstance(refusal, str) and refusal.strip():
+            # Record the refusal explanation regardless — it's useful provider
+            # metadata even when the model also returned a usable payload.
+            provider_data["refusal"] = refusal
+            _has_text = isinstance(content, str) and content.strip()
+            _has_tool_calls = bool(tool_calls)
+            # Only promote to a terminal ``content_filter`` when the refusal is
+            # the *sole* payload — no visible text and no tool calls. A response
+            # that carries real content (or tool calls) alongside a refusal note
+            # is a normal, usable turn: surfacing it as a failed safety refusal
+            # would discard the model's actual work. In the empty-payload case,
+            # adopt the refusal as content so the loop has something to show.
+            if not _has_text and not _has_tool_calls:
+                content = refusal
+                if finish_reason in (None, "stop"):
+                    finish_reason = "content_filter"
+
         return NormalizedResponse(
-            content=msg.content,
+            content=content,
             tool_calls=tool_calls,
             finish_reason=finish_reason,
             reasoning=reasoning,

agent/transports/codex.py

@@ -218,22 +218,10 @@ class ResponsesApiTransport(ProviderTransport):
             kwargs.pop("timeout", None)
 
         if is_codex_backend:
-            prompt_cache_key = kwargs.get("prompt_cache_key")
-            cache_scope_id = str(prompt_cache_key or session_id or "").strip()
-            if cache_scope_id:
-                existing_extra_headers = kwargs.get("extra_headers")
-                merged_extra_headers: Dict[str, str] = {}
-                if isinstance(existing_extra_headers, dict):
-                    merged_extra_headers.update(
-                        {
-                            str(key): str(value)
-                            for key, value in existing_extra_headers.items()
-                            if key and value is not None
-                        }
-                    )
-                merged_extra_headers["session_id"] = cache_scope_id
-                merged_extra_headers["x-client-request-id"] = cache_scope_id
-                kwargs["extra_headers"] = merged_extra_headers
+            # chatgpt.com/backend-api/codex rejects body-level
+            # ``extra_headers`` with HTTP 400. Correlation/cache routing for
+            # this backend must not be sent through the Responses payload.
+            kwargs.pop("extra_headers", None)
 
         max_tokens = params.get("max_tokens")
         if max_tokens is not None and not is_codex_backend:

agent/transports/codex_app_server.py

@@ -378,6 +378,7 @@ def check_codex_binary(
             capture_output=True,
             text=True,
             timeout=10,
+            stdin=subprocess.DEVNULL,
         )
     except FileNotFoundError:
         return False, (

agent/transports/codex_app_server_session.py

@@ -72,6 +72,9 @@ class TurnResult:
     error: Optional[str] = None  # Set if turn ended in a non-recoverable error
     turn_id: Optional[str] = None
     thread_id: Optional[str] = None
+    token_usage_last: Optional[dict[str, Any]] = None
+    token_usage_total: Optional[dict[str, Any]] = None
+    model_context_window: Optional[int] = None
     # Hint to the caller that the underlying codex subprocess is likely
     # wedged (turn-level timeout fired, post-tool watchdog tripped, or
     # token-refresh failure killed the child). The caller should retire
@@ -501,6 +504,7 @@ class CodexAppServerSession:
                     pending = self._client.take_notification(timeout=0)
                     if pending is None:
                         break
+                    _apply_token_usage_notification(result, pending)
                     self._track_pending_file_change(pending)
                     proj = projector.project(pending)
                     if proj.messages:
@@ -536,6 +540,8 @@ class CodexAppServerSession:
                 except Exception:  # pragma: no cover - display callback
                     logger.debug("on_event callback raised", exc_info=True)
 
+            _apply_token_usage_notification(result, note)
+
             # Track in-progress fileChange items so the approval bridge
             # can surface a real change summary when codex requests
             # approval (the approval params themselves don't carry the
@@ -802,6 +808,30 @@ class CodexAppServerSession:
         return cached
 
 
+def _apply_token_usage_notification(result: TurnResult, note: dict) -> None:
+    """Capture Codex app-server token usage updates for caller accounting.
+
+    Codex does not put token usage on turn/completed. It emits a separate
+    thread/tokenUsage/updated notification containing cumulative totals and
+    the latest turn breakdown.
+    """
+    if not isinstance(note, dict) or note.get("method") != "thread/tokenUsage/updated":
+        return
+    params = note.get("params") or {}
+    token_usage = params.get("tokenUsage") or {}
+    if not isinstance(token_usage, dict):
+        return
+    last = token_usage.get("last")
+    total = token_usage.get("total")
+    if isinstance(last, dict):
+        result.token_usage_last = dict(last)
+    if isinstance(total, dict):
+        result.token_usage_total = dict(total)
+    window = token_usage.get("modelContextWindow")
+    if isinstance(window, int) and window > 0:
+        result.model_context_window = window
+
+
 def _approval_choice_to_codex_decision(choice: str) -> str:
     """Map Hermes approval choices onto codex's CommandExecutionApprovalDecision
     / FileChangeApprovalDecision wire values.

agent/transports/types.py

@@ -121,6 +121,18 @@ class NormalizedResponse:
         pd = self.provider_data or {}
         return pd.get("reasoning_details")
 
+    @property
+    def anthropic_content_blocks(self):
+        """Verbatim, order-preserving Anthropic content blocks for a turn.
+
+        Present only when an Anthropic turn interleaves signed thinking with
+        tool_use — the one shape the parallel reasoning_details + tool_calls
+        lists reconstruct in the wrong order, invalidating thinking-block
+        signatures on replay. See agent/transports/anthropic.py.
+        """
+        pd = self.provider_data or {}
+        return pd.get("anthropic_content_blocks")
+
     @property
     def codex_reasoning_items(self):
         pd = self.provider_data or {}

agent/usage_pricing.py

@@ -13,6 +13,7 @@ DEFAULT_PRICING = {"input": 0.0, "output": 0.0}
 
 _ZERO = Decimal("0")
 _ONE_MILLION = Decimal("1000000")
+_NOUS_DEFAULT_BASE_URL = "https://inference-api.nousresearch.com/v1"
 
 CostStatus = Literal["actual", "estimated", "included", "unknown"]
 CostSource = Literal[
@@ -570,6 +571,8 @@ def resolve_billing_route(
         return BillingRoute(provider="openai-codex", model=model, base_url=base_url or "", billing_mode="subscription_included")
     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 == "nous" or base_url_host_matches(base_url or "", "inference-api.nousresearch.com"):
+        return BillingRoute(provider="nous", model=model, base_url=base_url or _NOUS_DEFAULT_BASE_URL, billing_mode="official_models_api")
     if provider_name == "anthropic":
         return BillingRoute(provider="anthropic", model=model.split("/")[-1], base_url=base_url or "", billing_mode="official_docs_snapshot")
     if provider_name == "openai":

apps/bootstrap-installer/package.json

@@ -11,11 +11,12 @@
     "tauri": "tauri",
     "tauri:dev": "tauri dev",
     "tauri:build": "tauri build",
-    "tauri:build:debug": "tauri build --debug"
+    "tauri:build:debug": "tauri build --debug",
+    "typecheck": "tsc -p . --noEmit"
   },
   "dependencies": {
     "@nous-research/ui": "0.16.0",
-    "@tailwindcss/vite": "^4.2.1",
+    "@tailwindcss/vite": "^4.2.4",
     "@tailwindcss/typography": "^0.5.19",
     "@tauri-apps/api": "^2.0.0",
     "@tauri-apps/plugin-dialog": "^2.0.0",
@@ -39,8 +40,8 @@
     "@tauri-apps/cli": "^2.0.0",
     "@types/react": "^19.2.14",
     "@types/react-dom": "^19.2.3",
-    "@vitejs/plugin-react": "^5.2.0",
-    "typescript": "~5.9.3",
-    "vite": "^7.3.1"
+    "@vitejs/plugin-react": "^6.0.2",
+    "typescript": "^6.0.3",
+    "vite": "^8.0.16"
   }
 }

apps/bootstrap-installer/src-tauri/src/update.rs

@@ -3,8 +3,9 @@
 //! Driven when the installer is launched as `Hermes-Setup.exe --update` (see
 //! `AppMode` in lib.rs). The desktop app hands off to us — it exits, then we:
 //!
-//!   1. wait for the old Hermes desktop process to fully exit (so the venv
-//!      shim is free; otherwise `hermes update` aborts with exit code 2),
+//!   1. wait for the old Hermes desktop process to fully exit (so both the
+//!      venv shim and packaged app.asar are free; otherwise `hermes update`
+//!      or repair bootstrap can race locked files),
 //!   2. run `hermes update --yes --gateway` (Python/repo update; this does NOT
 //!      rebuild apps/desktop by design — see cmd_update in hermes_cli/main.py),
 //!   3. run `hermes desktop --build-only` (the rebuild step update skips),
@@ -38,8 +39,8 @@ use crate::events::{BootstrapEvent, LogStream, StageInfo, StageState};
 /// hermes_cli/main.py (sys.exit(2)). We surface a targeted message for this.
 const UPDATE_EXIT_CONCURRENT: i32 = 2;
 
-/// How long to wait for the old desktop process to release the venv shim
-/// before giving up and letting `hermes update`'s own guard decide.
+/// How long to wait for the old desktop process to release files under the
+/// install tree before giving up and letting `hermes update`'s own guard decide.
 const DESKTOP_EXIT_WAIT: Duration = Duration::from_secs(20);
 const DESKTOP_EXIT_POLL: Duration = Duration::from_millis(500);
 
@@ -150,8 +151,10 @@ async fn run_update(app: AppHandle) -> Result<()> {
     // ---- pre-step: wait for the old desktop to die -----------------------
     // The desktop exec'd us then called app.exit(), but process teardown is
     // async on Windows. If it still holds the venv shim, `hermes update`
-    // aborts with exit 2. Give it a bounded window to clear.
-    wait_for_venv_free(&install_root, &app).await;
+    // aborts with exit 2. If it still holds the packaged app.asar,
+    // install.ps1's repair/re-clone path cannot move/remove the install tree.
+    // Give both handles a bounded window to clear.
+    wait_for_install_locks_free(&install_root, &app, "update").await;
 
     // ---- stage 1: hermes update -----------------------------------------
     // Pass --branch so `hermes update` targets the branch this installer was
@@ -173,8 +176,8 @@ async fn run_update(app: AppHandle) -> Result<()> {
         vec!["update".into(), "--yes".into(), "--gateway".into()];
     // --force skips `hermes update`'s Windows running-exe guard (which would
     // `sys.exit(2)` and dead-end the handoff). By contract the desktop has
-    // already exited and waited for the venv shim to unlock before launching
-    // us, and wait_for_venv_free below force-kills any straggler — so by the
+    // already exited and waited for the install locks to clear before launching
+    // us, and wait_for_install_locks_free below force-kills any straggler — so by the
     // time `hermes update` runs there is no legitimate hermes.exe to protect,
     // and the guard would only produce a false "Hermes is still running" stop.
     update_args.push("--force".into());
@@ -391,48 +394,57 @@ async fn run_update(app: AppHandle) -> Result<()> {
     Ok(())
 }
 
-/// Poll until the venv shim is no longer locked (Windows) or a bounded timeout
-/// elapses. On non-Windows this is a short fixed grace since file locking
-/// isn't the failure mode there.
-async fn wait_for_venv_free(install_root: &Path, app: &AppHandle) {
-    let shim = venv_hermes(install_root);
+/// Poll until the venv shim AND packaged desktop app bundle are no longer locked
+/// (Windows) or a bounded timeout elapses. On non-Windows this is a short fixed
+/// grace since file locking isn't the failure mode there.
+pub(crate) async fn wait_for_install_locks_free(install_root: &Path, app: &AppHandle, stage: &str) {
+    let lock_targets = install_lock_probe_paths(install_root);
     let deadline = Instant::now() + DESKTOP_EXIT_WAIT;
 
-    emit_log(app, Some("update"), LogStream::Stdout, "[update] waiting for Hermes to exit…");
+    emit_log(app, Some(stage), LogStream::Stdout, "[handoff] waiting for Hermes to exit…");
 
     loop {
-        if !is_locked(&shim) {
+        let locked = locked_paths(&lock_targets);
+        if locked.is_empty() {
             return;
         }
         if Instant::now() >= deadline {
-            // Last resort: a backend hermes.exe (or a grandchild it spawned)
-            // is still holding the shim. The desktop should have reaped its
-            // tree before handing off, but SIGTERM races / detached
-            // grandchildren / AV handles can leave a straggler. Rather than
-            // "proceed anyway" straight into uv's "Access is denied", force-kill
-            // every hermes.exe except ourselves, then give the OS a beat to
-            // unload the image.
+            // Last resort: a backend hermes.exe (or the desktop Hermes.exe
+            // itself) is still holding one of the update-sensitive files. The
+            // desktop should have reaped its tree before handing off, but
+            // SIGTERM races / detached grandchildren / AV handles can leave a
+            // straggler. Rather than "proceed anyway" straight into uv's
+            // "Access is denied" or install.ps1's locked app.asar failure,
+            // force-kill every Hermes.exe except ourselves, then give the OS a
+            // beat to unload the image.
             emit_log(
                 app,
-                Some("update"),
+                Some(stage),
                 LogStream::Stdout,
-                "[update] Hermes still holding the venv shim; force-killing stragglers…",
+                &format!(
+                    "[handoff] Hermes still holding install files ({}); force-killing stragglers…",
+                    format_locked_paths(&locked)
+                ),
             );
             force_kill_other_hermes();
             tokio::time::sleep(Duration::from_millis(800)).await;
-            if !is_locked(&shim) {
+            let locked_after_kill = locked_paths(&lock_targets);
+            if locked_after_kill.is_empty() {
                 emit_log(
                     app,
-                    Some("update"),
+                    Some(stage),
                     LogStream::Stdout,
-                    "[update] venv shim freed after force-kill",
+                    "[handoff] install files freed after force-kill",
                 );
             } else {
                 emit_log(
                     app,
-                    Some("update"),
+                    Some(stage),
                     LogStream::Stdout,
-                    "[update] venv shim still locked; proceeding (--force + quarantine will handle it)",
+                    &format!(
+                        "[handoff] install files still locked ({}); proceeding (--force + quarantine will handle it)",
+                        format_locked_paths(&locked_after_kill)
+                    ),
                 );
             }
             return;
@@ -441,13 +453,44 @@ async fn wait_for_venv_free(install_root: &Path, app: &AppHandle) {
     }
 }
 
+fn install_lock_probe_paths(install_root: &Path) -> Vec<PathBuf> {
+    let mut paths = vec![venv_hermes(install_root)];
+    paths.extend(desktop_app_payload_paths(install_root));
+    paths
+}
+
+fn desktop_app_payload_paths(install_root: &Path) -> Vec<PathBuf> {
+    let release = install_root.join("apps").join("desktop").join("release");
+    if cfg!(target_os = "windows") {
+        vec![
+            release.join("win-unpacked").join("resources").join("app.asar"),
+            release.join("win-arm64-unpacked").join("resources").join("app.asar"),
+        ]
+    } else if cfg!(target_os = "macos") {
+        vec![
+            release.join("mac").join("Hermes.app").join("Contents").join("Resources").join("app.asar"),
+            release.join("mac-arm64").join("Hermes.app").join("Contents").join("Resources").join("app.asar"),
+        ]
+    } else {
+        vec![release.join("linux-unpacked").join("resources").join("app.asar")]
+    }
+}
+
+fn locked_paths(paths: &[PathBuf]) -> Vec<PathBuf> {
+    paths.iter().filter(|p| is_locked(p)).cloned().collect()
+}
+
+fn format_locked_paths(paths: &[PathBuf]) -> String {
+    paths.iter().map(|p| p.display().to_string()).collect::<Vec<_>>().join(", ")
+}
+
 /// Force-kill any `hermes.exe` other than this process. Windows-only; a no-op
 /// elsewhere (POSIX has no mandatory-lock contention). We can't selectively
 /// target "the backend" by PID here — the desktop already exited and we never
 /// knew its children — so we kill the whole `hermes.exe` image tree via
 /// taskkill, excluding our own PID.
 ///
-/// Safe w.r.t. our own update child: this runs inside `wait_for_venv_free`,
+/// Safe w.r.t. our own update child: this runs inside the install-lock wait,
 /// which completes BEFORE we spawn `venv\Scripts\hermes.exe update`. At this
 /// point no update-driven hermes.exe exists yet, so the only hermes.exe images
 /// are stragglers from the old desktop — exactly what we want gone. (`/FI PID
@@ -891,6 +934,29 @@ mod tests {
         assert!(!is_locked(Path::new("/nonexistent/does/not/exist/xyz")));
     }
 
+    #[test]
+    fn lock_probe_paths_include_desktop_app_payload() {
+        let root = Path::new("/x/hermes-agent");
+        let probes = install_lock_probe_paths(root);
+
+        assert!(
+            probes.iter().any(|p| p == &venv_hermes(root)),
+            "venv shim remains part of the update lock probe"
+        );
+        assert!(
+            probes.iter().any(|p| p.ends_with(Path::new("resources/app.asar"))),
+            "packaged app.asar must be probed so repair/re-clone waits for the old desktop to exit"
+        );
+    }
+
+    #[test]
+    fn locked_paths_ignores_missing_payloads() {
+        let root = Path::new("/nonexistent/hermes-agent");
+        let probes = install_lock_probe_paths(root);
+
+        assert!(locked_paths(&probes).is_empty());
+    }
+
     #[test]
     fn parses_update_branch_from_space_or_equals_args() {
         assert_eq!(

apps/bootstrap-installer/tsconfig.json

@@ -1,8 +1,8 @@
 {
   "compilerOptions": {
-    "target": "ES2022",
+    "target": "ES2023",
     "useDefineForClassFields": true,
-    "lib": ["ES2022", "DOM", "DOM.Iterable"],
+    "lib": ["ES2023", "DOM", "DOM.Iterable"],
     "module": "ESNext",
     "skipLibCheck": true,
     "moduleResolution": "bundler",
@@ -16,9 +16,8 @@
     "noUnusedParameters": true,
     "esModuleInterop": true,
     "noFallthroughCasesInSwitch": true,
-    "baseUrl": ".",
     "paths": {
-      "@/*": ["src/*"]
+      "@/*": ["./src/*"]
     }
   },
   "include": ["src"],

apps/desktop/README.md

@@ -34,7 +34,7 @@ It builds and launches the GUI against your existing install — same config, ke
 
 ### Prebuilt installers
 
-Prebuilt installers are built and distributed via [the Hermes Desktop website.](https://hermes-agent.nousresearch.com/desktop).
+Prebuilt installers are built and distributed via [the Hermes Desktop website.](https://hermes-agent.nousresearch.com/).
 
 ---
 
@@ -93,7 +93,7 @@ Run before opening a PR (lint may surface pre-existing warnings but must exit cl
 
 ```bash
 npm run fix
-npm run type-check
+npm run typecheck
 npm run lint
 npm run test:desktop:all
 ```

apps/desktop/electron/backend-env.cjs

@@ -0,0 +1,112 @@
+const path = require('node:path')
+
+// Match the POSIX fallback surface used by the Python terminal environment.
+// macOS apps launched from Finder/Dock often inherit only /usr/bin:/bin:/usr/sbin:/sbin,
+// which misses Apple Silicon Homebrew and user-installed CLI tools such as codex.
+const POSIX_SANE_PATH_ENTRIES = Object.freeze([
+  '/opt/homebrew/bin',
+  '/opt/homebrew/sbin',
+  '/usr/local/sbin',
+  '/usr/local/bin',
+  '/usr/sbin',
+  '/usr/bin',
+  '/sbin',
+  '/bin'
+])
+
+function delimiterForPlatform(platform = process.platform) {
+  return platform === 'win32' ? ';' : ':'
+}
+
+function pathModuleForPlatform(platform = process.platform) {
+  return platform === 'win32' ? path.win32 : path.posix
+}
+
+function pathEnvKey(env = process.env, platform = process.platform) {
+  if (platform !== 'win32') return 'PATH'
+  return Object.keys(env || {}).find(key => key.toUpperCase() === 'PATH') || 'PATH'
+}
+
+function currentPathValue(env = process.env, platform = process.platform) {
+  const key = pathEnvKey(env, platform)
+  return env?.[key] || ''
+}
+
+function appendUniquePathEntries(entries, { delimiter = path.delimiter } = {}) {
+  const seen = new Set()
+  const ordered = []
+
+  for (const entry of entries) {
+    if (!entry) continue
+    const parts = Array.isArray(entry) ? entry : String(entry).split(delimiter)
+    for (const part of parts) {
+      if (!part || seen.has(part)) continue
+      seen.add(part)
+      ordered.push(part)
+    }
+  }
+
+  return ordered.join(delimiter)
+}
+
+function buildDesktopBackendPath({
+  hermesHome,
+  venvRoot,
+  currentPath = '',
+  platform = process.platform,
+  pathModule = pathModuleForPlatform(platform)
+} = {}) {
+  const delimiter = delimiterForPlatform(platform)
+  const hermesNodeBin = hermesHome ? pathModule.join(hermesHome, 'node', 'bin') : null
+  const venvBin = venvRoot ? pathModule.join(venvRoot, platform === 'win32' ? 'Scripts' : 'bin') : null
+  const saneEntries = platform === 'win32' ? [] : POSIX_SANE_PATH_ENTRIES
+
+  return appendUniquePathEntries(
+    [hermesNodeBin, venvBin, currentPath, saneEntries],
+    { delimiter }
+  )
+}
+
+function normalizeHermesHomeRoot(hermesHome, { pathModule = pathModuleForPlatform(process.platform) } = {}) {
+  if (!hermesHome) return hermesHome
+  const resolved = pathModule.resolve(String(hermesHome))
+  const parent = pathModule.dirname(resolved)
+  if (pathModule.basename(parent).toLowerCase() === 'profiles') {
+    return pathModule.dirname(parent)
+  }
+  return resolved
+}
+
+function buildDesktopBackendEnv({
+  hermesHome,
+  pythonPathEntries = [],
+  venvRoot,
+  currentEnv = process.env,
+  platform = process.platform,
+  pathModule = pathModuleForPlatform(platform)
+} = {}) {
+  const delimiter = delimiterForPlatform(platform)
+  const currentPythonPath = currentEnv?.PYTHONPATH || ''
+  const key = pathEnvKey(currentEnv, platform)
+
+  return {
+    PYTHONPATH: appendUniquePathEntries([...pythonPathEntries, currentPythonPath], { delimiter }),
+    [key]: buildDesktopBackendPath({
+      hermesHome,
+      venvRoot,
+      currentPath: currentPathValue(currentEnv, platform),
+      platform,
+      pathModule
+    })
+  }
+}
+
+module.exports = {
+  POSIX_SANE_PATH_ENTRIES,
+  appendUniquePathEntries,
+  buildDesktopBackendEnv,
+  buildDesktopBackendPath,
+  delimiterForPlatform,
+  normalizeHermesHomeRoot,
+  pathEnvKey
+}

apps/desktop/electron/backend-env.test.cjs

@@ -0,0 +1,111 @@
+const test = require('node:test')
+const assert = require('node:assert/strict')
+const path = require('node:path')
+
+const {
+  POSIX_SANE_PATH_ENTRIES,
+  appendUniquePathEntries,
+  buildDesktopBackendEnv,
+  buildDesktopBackendPath,
+  normalizeHermesHomeRoot,
+  pathEnvKey
+} = require('./backend-env.cjs')
+
+test('desktop backend PATH adds Hermes-managed bins and missing POSIX sane entries', () => {
+  const result = buildDesktopBackendPath({
+    hermesHome: '/Users/test/.hermes',
+    venvRoot: '/Users/test/.hermes/hermes-agent/venv',
+    currentPath: '/usr/bin:/bin:/usr/sbin:/sbin:/usr/local/bin',
+    platform: 'darwin',
+    pathModule: path.posix
+  })
+
+  const entries = result.split(':')
+  assert.equal(entries[0], '/Users/test/.hermes/node/bin')
+  assert.equal(entries[1], '/Users/test/.hermes/hermes-agent/venv/bin')
+  assert.ok(entries.includes('/opt/homebrew/bin'), 'Apple Silicon Homebrew bin is added')
+  assert.ok(entries.includes('/opt/homebrew/sbin'), 'Apple Silicon Homebrew sbin is added')
+  assert.ok(entries.includes('/usr/local/sbin'), 'missing standard sbin is added')
+
+  for (const expected of POSIX_SANE_PATH_ENTRIES) {
+    assert.ok(entries.includes(expected), `${expected} should be present`)
+  }
+})
+
+test('desktop backend PATH preserves first occurrence and avoids duplicates', () => {
+  const result = buildDesktopBackendPath({
+    hermesHome: '/Users/test/.hermes',
+    venvRoot: '/Users/test/.hermes/hermes-agent/venv',
+    currentPath: '/opt/homebrew/bin:/usr/bin:/opt/homebrew/bin:/bin',
+    platform: 'darwin',
+    pathModule: path.posix
+  })
+
+  const entries = result.split(':')
+  assert.equal(entries.filter(entry => entry === '/opt/homebrew/bin').length, 1)
+  assert.ok(
+    entries.indexOf('/opt/homebrew/bin') < entries.indexOf('/opt/homebrew/sbin'),
+    'existing Homebrew bin keeps its precedence over appended missing sane entries'
+  )
+})
+
+test('buildDesktopBackendEnv extends PYTHONPATH and backend PATH together', () => {
+  const env = buildDesktopBackendEnv({
+    hermesHome: '/Users/test/.hermes',
+    pythonPathEntries: ['/repo/hermes-agent'],
+    venvRoot: '/Users/test/.hermes/hermes-agent/venv',
+    currentEnv: {
+      PATH: '/usr/bin:/bin',
+      PYTHONPATH: '/existing/pythonpath'
+    },
+    platform: 'darwin',
+    pathModule: path.posix
+  })
+
+  assert.equal(env.PYTHONPATH, '/repo/hermes-agent:/existing/pythonpath')
+  assert.ok(env.PATH.startsWith('/Users/test/.hermes/node/bin:/Users/test/.hermes/hermes-agent/venv/bin:'))
+  assert.ok(env.PATH.includes('/opt/homebrew/bin'))
+})
+
+test('normalizeHermesHomeRoot maps profile homes back to the global Hermes root', () => {
+  assert.equal(
+    normalizeHermesHomeRoot('/Users/test/.hermes/profiles/oracle', { pathModule: path.posix }),
+    '/Users/test/.hermes'
+  )
+  assert.equal(
+    normalizeHermesHomeRoot('C:\\Users\\test\\AppData\\Local\\hermes\\profiles\\oracle', { pathModule: path.win32 }),
+    'C:\\Users\\test\\AppData\\Local\\hermes'
+  )
+  assert.equal(
+    normalizeHermesHomeRoot('/Users/test/.hermes', { pathModule: path.posix }),
+    '/Users/test/.hermes'
+  )
+})
+
+test('Windows PATH casing and delimiter are preserved without POSIX sane entries', () => {
+  const env = buildDesktopBackendEnv({
+    hermesHome: 'C:\\Users\\test\\AppData\\Local\\hermes',
+    pythonPathEntries: ['C:\\repo\\hermes-agent'],
+    venvRoot: 'C:\\Users\\test\\AppData\\Local\\hermes\\hermes-agent\\venv',
+    currentEnv: {
+      Path: 'C:\\Windows\\System32;C:\\Windows',
+      PYTHONPATH: 'C:\\existing\\pythonpath'
+    },
+    platform: 'win32',
+    pathModule: path.win32
+  })
+
+  assert.equal(pathEnvKey({ Path: 'x' }, 'win32'), 'Path')
+  assert.equal(env.PATH, undefined)
+  assert.ok(env.Path.startsWith('C:\\Users\\test\\AppData\\Local\\hermes\\node\\bin;'))
+  assert.ok(env.Path.includes('\\venv\\Scripts;'))
+  assert.ok(env.Path.includes(';C:\\Windows\\System32;C:\\Windows'))
+  assert.equal(env.Path.includes('/opt/homebrew/bin'), false)
+})
+
+test('appendUniquePathEntries drops empty entries and keeps first occurrence', () => {
+  assert.equal(
+    appendUniquePathEntries([':/a::/b', ['/a', '/c']], { delimiter: ':' }),
+    '/a:/b:/c'
+  )
+})

apps/desktop/electron/backend-ready.cjs

@@ -0,0 +1,66 @@
+const _READY_RE = /^HERMES_DASHBOARD_READY port=(\d+)/m
+
+/**
+ * Watch a child process's stdout for the `HERMES_DASHBOARD_READY port=<N>`
+ * line that web_server.py prints after uvicorn binds its socket.
+ *
+ * Returns the parsed port. Rejects if:
+ *   - the child exits before emitting the line
+ *   - the child emits an `error` event
+ *   - no line arrives within the timeout
+ *
+ * A single `cleanup()` tears down every listener (data/exit/error/timeout)
+ * on every terminal path — resolve, reject, or timeout — so repeated
+ * backend spawns don't leak listener slots on the child.
+ */
+function waitForDashboardPort(child, timeoutMs = 45_000) {
+  return new Promise((resolve, reject) => {
+    let buf = ''
+    let done = false
+
+    function cleanup() {
+      if (done) return
+      done = true
+      clearTimeout(timer)
+      child.stdout.off('data', onData)
+      child.off('exit', onExit)
+      child.off('error', onError)
+    }
+
+    function onData(chunk) {
+      buf += chunk.toString()
+      let nl
+      while ((nl = buf.indexOf('\n')) !== -1) {
+        const line = buf.slice(0, nl)
+        buf = buf.slice(nl + 1)
+        const m = line.match(_READY_RE)
+        if (m) {
+          cleanup()
+          resolve(parseInt(m[1], 10))
+          return
+        }
+      }
+    }
+
+    function onExit(code, signal) {
+      cleanup()
+      reject(new Error(`Hermes backend: exited before port announcement (${signal || code})`))
+    }
+
+    function onError(err) {
+      cleanup()
+      reject(err)
+    }
+
+    const timer = setTimeout(() => {
+      cleanup()
+      reject(new Error(`Timed out waiting for Hermes backend port announcement (${timeoutMs}ms)`))
+    }, timeoutMs)
+
+    child.stdout.on('data', onData)
+    child.on('exit', onExit)
+    child.on('error', onError)
+  })
+}
+
+module.exports = { waitForDashboardPort }

apps/desktop/electron/bootstrap-runner.cjs

@@ -40,6 +40,15 @@ const path = require('node:path')
 const https = require('node:https')
 const { spawn } = require('node:child_process')
 
+const IS_WINDOWS = process.platform === 'win32'
+
+function hiddenWindowsChildOptions(options = {}) {
+  if (!IS_WINDOWS || Object.prototype.hasOwnProperty.call(options, 'windowsHide')) {
+    return options
+  }
+  return { ...options, windowsHide: true }
+}
+
 const STAMP_COMMIT_RE = /^[0-9a-f]{7,40}$/i
 
 // Stages flagged needs_user_input=true in the manifest are skipped by the
@@ -284,7 +293,7 @@ function spawnPowerShell(scriptPath, args, { emit, stageName, abortSignal, herme
     const ps = process.platform === 'win32' ? resolveWindowsPowerShell() : 'pwsh'
     const fullArgs = ['-NoProfile', '-ExecutionPolicy', 'Bypass', '-File', scriptPath, ...args]
 
-    const child = spawn(ps, fullArgs, {
+    const child = spawn(ps, fullArgs, hiddenWindowsChildOptions({
       stdio: ['ignore', 'pipe', 'pipe'],
       env: {
         ...process.env,
@@ -292,7 +301,7 @@ function spawnPowerShell(scriptPath, args, { emit, stageName, abortSignal, herme
         // choice rather than re-computing the default.
         HERMES_HOME: hermesHome || process.env.HERMES_HOME || ''
       }
-    })
+    }))
 
     let stdout = ''
     let stderr = ''

apps/desktop/electron/connection-config.cjs

@@ -166,6 +166,39 @@ function profileRemoteOverride(config, profile) {
   return { url, authMode: normAuthMode(entry.authMode), token: entry.token }
 }
 
+/**
+ * In global-remote mode one backend serves every Desktop profile, so REST calls
+ * that are scoped by renderer-side `request.profile` must carry that scope as a
+ * query parameter. Local pooled backends and per-profile remote overrides do not
+ * need this: they already run against a backend scoped to the target profile.
+ */
+function pathWithGlobalRemoteProfile(path, profile, opts = {}) {
+  const scopedProfile = connectionScopeKey(profile)
+  if (!scopedProfile || !opts.globalRemote || opts.profileRemoteOverride) {
+    return path
+  }
+
+  const rawPath = String(path || '')
+  if (!rawPath) {
+    return path
+  }
+
+  let parsed
+  try {
+    parsed = new URL(rawPath, 'http://hermes.local')
+  } catch {
+    return path
+  }
+
+  if (parsed.searchParams.has('profile')) {
+    return path
+  }
+
+  parsed.searchParams.set('profile', scopedProfile)
+
+  return `${parsed.pathname}${parsed.search}${parsed.hash}`
+}
+
 function tokenPreview(value) {
   const raw = String(value || '')
 
@@ -247,6 +280,7 @@ module.exports = {
   cookiesHaveLiveSession,
   normAuthMode,
   normalizeRemoteBaseUrl,
+  pathWithGlobalRemoteProfile,
   profileRemoteOverride,
   resolveAuthMode,
   resolveTestWsUrl,

apps/desktop/electron/connection-config.test.cjs

@@ -24,6 +24,7 @@ const {
   cookiesHaveLiveSession,
   normAuthMode,
   normalizeRemoteBaseUrl,
+  pathWithGlobalRemoteProfile,
   profileRemoteOverride,
   resolveAuthMode,
   resolveTestWsUrl,
@@ -90,6 +91,72 @@ test('profileRemoteOverride tolerates a missing/!object profiles map', () => {
   assert.equal(profileRemoteOverride(null, 'coder'), null)
 })
 
+// --- pathWithGlobalRemoteProfile ---
+
+test('pathWithGlobalRemoteProfile appends profile in global remote mode', () => {
+  assert.equal(
+    pathWithGlobalRemoteProfile('/api/model/info', 'iris', {
+      globalRemote: true,
+      profileRemoteOverride: false
+    }),
+    '/api/model/info?profile=iris'
+  )
+})
+
+test('pathWithGlobalRemoteProfile preserves existing query params', () => {
+  assert.equal(
+    pathWithGlobalRemoteProfile('/api/model/options?force=1', 'iris', {
+      globalRemote: true,
+      profileRemoteOverride: false
+    }),
+    '/api/model/options?force=1&profile=iris'
+  )
+})
+
+test('pathWithGlobalRemoteProfile does not replace an explicit profile query', () => {
+  assert.equal(
+    pathWithGlobalRemoteProfile('/api/model/info?profile=default', 'iris', {
+      globalRemote: true,
+      profileRemoteOverride: false
+    }),
+    '/api/model/info?profile=default'
+  )
+})
+
+test('pathWithGlobalRemoteProfile skips local and per-profile remote override paths', () => {
+  assert.equal(
+    pathWithGlobalRemoteProfile('/api/model/info', 'iris', {
+      globalRemote: false,
+      profileRemoteOverride: false
+    }),
+    '/api/model/info'
+  )
+  assert.equal(
+    pathWithGlobalRemoteProfile('/api/model/info', 'iris', {
+      globalRemote: true,
+      profileRemoteOverride: true
+    }),
+    '/api/model/info'
+  )
+})
+
+test('pathWithGlobalRemoteProfile skips empty profile/path safely', () => {
+  assert.equal(
+    pathWithGlobalRemoteProfile('/api/model/info', '', {
+      globalRemote: true,
+      profileRemoteOverride: false
+    }),
+    '/api/model/info'
+  )
+  assert.equal(
+    pathWithGlobalRemoteProfile('', 'iris', {
+      globalRemote: true,
+      profileRemoteOverride: false
+    }),
+    ''
+  )
+})
+
 // --- normalizeRemoteBaseUrl ---
 
 test('normalizeRemoteBaseUrl strips trailing slashes, hash, and query', () => {

apps/desktop/electron/dashboard-token.cjs

@@ -0,0 +1,99 @@
+/**
+ * Helpers for local dashboard session-token discovery.
+ *
+ * The desktop main process can pass HERMES_DASHBOARD_SESSION_TOKEN when it
+ * spawns the local dashboard, but the dashboard is the source of truth for the
+ * token it actually serves to the renderer. If those drift, HTTP readiness
+ * probes still pass while /api/ws rejects the renderer's token.
+ */
+
+const DEFAULT_TOKEN_FETCH_TIMEOUT_MS = 3_000
+
+async function fetchPublicText(url, options = {}) {
+  const { protocol } = new URL(url)
+  if (protocol !== 'http:' && protocol !== 'https:') {
+    throw new Error(`Unsupported Hermes backend URL protocol: ${protocol}`)
+  }
+
+  const timeoutMs = options.timeoutMs ?? DEFAULT_TOKEN_FETCH_TIMEOUT_MS
+  const res = await fetch(url, { signal: AbortSignal.timeout(timeoutMs) }).catch(error => {
+    if (error.name === 'TimeoutError') {
+      throw new Error(`Timed out connecting to Hermes backend after ${timeoutMs}ms`)
+    }
+    throw error
+  })
+  const text = await res.text()
+
+  if (!res.ok) throw new Error(`${res.status}: ${text || res.statusText}`)
+
+  return text
+}
+
+function extractInjectedDashboardToken(html) {
+  const match = /window\.__HERMES_SESSION_TOKEN__\s*=\s*("(?:\\.|[^"\\])*")/.exec(String(html || ''))
+  if (!match) return null
+  try {
+    return JSON.parse(match[1])
+  } catch {
+    return null
+  }
+}
+
+function dashboardIndexUrl(baseUrl) {
+  return `${String(baseUrl || '').replace(/\/+$/, '')}/`
+}
+
+async function resolveServedDashboardToken(baseUrl, fallbackToken, options = {}) {
+  const fetchText = options.fetchText || fetchPublicText
+  const html = await fetchText(dashboardIndexUrl(baseUrl), {
+    timeoutMs: options.timeoutMs ?? DEFAULT_TOKEN_FETCH_TIMEOUT_MS
+  })
+  const servedToken = extractInjectedDashboardToken(html)
+
+  if (servedToken && servedToken !== fallbackToken && typeof options.rememberLog === 'function') {
+    options.rememberLog('[boot] dashboard served a different session token; using served token for WebSocket auth')
+  }
+
+  return servedToken || fallbackToken
+}
+
+/**
+ * A served token that differs from our spawn token while our child is DEAD
+ * came from a process we did not spawn (orphan/port squatter that satisfied
+ * the public /api/status readiness probe). With a live child the mismatch is
+ * benign: our own backend regenerated the token because the env pin did not
+ * survive the spawn.
+ */
+function isForeignBackendToken({ servedToken, spawnToken, childAlive }) {
+  return Boolean(servedToken) && servedToken !== spawnToken && !childAlive
+}
+
+/**
+ * Resolve the token the backend actually serves, adopting benign drift and
+ * failing loudly on a foreign backend. `childAlive` is a thunk so liveness is
+ * sampled after the fetch, not before.
+ */
+async function adoptServedDashboardToken(baseUrl, spawnToken, { childAlive, label = 'Hermes backend', ...options }) {
+  const servedToken = await resolveServedDashboardToken(baseUrl, spawnToken, options).catch(error => {
+    options.rememberLog?.(`[boot] could not read served dashboard token (${label}): ${error.message}`)
+    return spawnToken
+  })
+
+  if (isForeignBackendToken({ servedToken, spawnToken, childAlive: childAlive() })) {
+    throw new Error(
+      `${label} exited and ${dashboardIndexUrl(baseUrl)} is served by a process we did not spawn; refusing its session token.`
+    )
+  }
+
+  return servedToken
+}
+
+module.exports = {
+  DEFAULT_TOKEN_FETCH_TIMEOUT_MS,
+  adoptServedDashboardToken,
+  dashboardIndexUrl,
+  extractInjectedDashboardToken,
+  fetchPublicText,
+  isForeignBackendToken,
+  resolveServedDashboardToken
+}

apps/desktop/electron/dashboard-token.test.cjs

@@ -0,0 +1,142 @@
+/**
+ * Tests for electron/dashboard-token.cjs.
+ *
+ * Run with: node --test electron/dashboard-token.test.cjs
+ * (Wired into npm test:desktop:platforms in package.json.)
+ */
+
+const test = require('node:test')
+const assert = require('node:assert/strict')
+
+const {
+  adoptServedDashboardToken,
+  dashboardIndexUrl,
+  extractInjectedDashboardToken,
+  fetchPublicText,
+  isForeignBackendToken,
+  resolveServedDashboardToken
+} = require('./dashboard-token.cjs')
+
+test('extractInjectedDashboardToken reads the JSON-encoded dashboard token', () => {
+  const html = '<script>window.__HERMES_SESSION_TOKEN__="served-token";window.__HERMES_BASE_PATH__=""</script>'
+  assert.equal(extractInjectedDashboardToken(html), 'served-token')
+})
+
+test('extractInjectedDashboardToken handles escaped token strings', () => {
+  const html = '<script>window.__HERMES_SESSION_TOKEN__="served\\\\token\\"quoted";</script>'
+  assert.equal(extractInjectedDashboardToken(html), 'served\\token"quoted')
+})
+
+test('extractInjectedDashboardToken returns null for missing or malformed values', () => {
+  assert.equal(extractInjectedDashboardToken('<html></html>'), null)
+  assert.equal(extractInjectedDashboardToken('<script>window.__HERMES_SESSION_TOKEN__={bad}</script>'), null)
+})
+
+test('dashboardIndexUrl preserves dashboard path prefixes', () => {
+  assert.equal(dashboardIndexUrl('http://127.0.0.1:9120'), 'http://127.0.0.1:9120/')
+  assert.equal(dashboardIndexUrl('https://host.example/hermes/'), 'https://host.example/hermes/')
+})
+
+test('resolveServedDashboardToken uses the served token and logs when it differs', async () => {
+  const logs = []
+  const token = await resolveServedDashboardToken('http://127.0.0.1:9120', 'spawn-token', {
+    fetchText: async url => {
+      assert.equal(url, 'http://127.0.0.1:9120/')
+      return '<script>window.__HERMES_SESSION_TOKEN__="served-token";</script>'
+    },
+    rememberLog: line => logs.push(line)
+  })
+
+  assert.equal(token, 'served-token')
+  assert.equal(logs.length, 1)
+  assert.match(logs[0], /served a different session token/)
+})
+
+test('resolveServedDashboardToken falls back when the served HTML has no token', async () => {
+  const token = await resolveServedDashboardToken('http://127.0.0.1:9120', 'spawn-token', {
+    fetchText: async () => '<html></html>',
+    rememberLog: () => {
+      throw new Error('should not log when no served token is present')
+    }
+  })
+
+  assert.equal(token, 'spawn-token')
+})
+
+test('resolveServedDashboardToken does not log when served token matches fallback', async () => {
+  const token = await resolveServedDashboardToken('http://127.0.0.1:9120', 'same-token', {
+    fetchText: async () => '<script>window.__HERMES_SESSION_TOKEN__="same-token";</script>',
+    rememberLog: () => {
+      throw new Error('should not log when token already matches')
+    }
+  })
+
+  assert.equal(token, 'same-token')
+})
+
+test('resolveServedDashboardToken propagates fetch errors so callers can fall back explicitly', async () => {
+  await assert.rejects(
+    () =>
+      resolveServedDashboardToken('http://127.0.0.1:9120', 'spawn-token', {
+        fetchText: async () => {
+          throw new Error('boom')
+        }
+      }),
+    /boom/
+  )
+})
+
+test('fetchPublicText rejects unsupported protocols', async () => {
+  await assert.rejects(() => fetchPublicText('file:///tmp/index.html'), /Unsupported Hermes backend URL protocol/)
+})
+
+test('isForeignBackendToken only flags a mismatched token from a dead child', () => {
+  const cases = [
+    [{ servedToken: 'other', spawnToken: 'mine', childAlive: false }, true],
+    // Live child + drift = our backend regenerated the token (env pin lost).
+    [{ servedToken: 'other', spawnToken: 'mine', childAlive: true }, false],
+    [{ servedToken: 'mine', spawnToken: 'mine', childAlive: false }, false],
+    [{ servedToken: 'mine', spawnToken: 'mine', childAlive: true }, false],
+    [{ servedToken: null, spawnToken: 'mine', childAlive: false }, false],
+    [{ servedToken: '', spawnToken: 'mine', childAlive: false }, false]
+  ]
+  for (const [input, expected] of cases) {
+    assert.equal(isForeignBackendToken(input), expected, JSON.stringify(input))
+  }
+})
+
+test('adoptServedDashboardToken adopts drift from a live child', async () => {
+  const token = await adoptServedDashboardToken('http://127.0.0.1:9120', 'spawn-token', {
+    childAlive: () => true,
+    fetchText: async () => '<script>window.__HERMES_SESSION_TOKEN__="served-token";</script>'
+  })
+
+  assert.equal(token, 'served-token')
+})
+
+test('adoptServedDashboardToken refuses a foreign token when our child is dead', async () => {
+  await assert.rejects(
+    () =>
+      adoptServedDashboardToken('http://127.0.0.1:9120', 'spawn-token', {
+        childAlive: () => false,
+        fetchText: async () => '<script>window.__HERMES_SESSION_TOKEN__="squatter-token";</script>',
+        label: 'Hermes backend for profile "work"'
+      }),
+    /profile "work".*process we did not spawn/
+  )
+})
+
+test('adoptServedDashboardToken falls back to the spawn token when the fetch fails', async () => {
+  const logs = []
+  const token = await adoptServedDashboardToken('http://127.0.0.1:9120', 'spawn-token', {
+    childAlive: () => true,
+    fetchText: async () => {
+      throw new Error('boom')
+    },
+    rememberLog: line => logs.push(line)
+  })
+
+  assert.equal(token, 'spawn-token')
+  assert.equal(logs.length, 1)
+  assert.match(logs[0], /could not read served dashboard token \(Hermes backend\): boom/)
+})

apps/desktop/electron/fs-read-dir.cjs

@@ -0,0 +1,109 @@
+'use strict'
+
+const fs = require('node:fs')
+const path = require('node:path')
+const { resolveDirectoryForIpc } = require('./hardening.cjs')
+
+const FS_READDIR_STAT_CONCURRENCY = 16
+
+// Always-hidden noise (covers non-git projects too; gitignore catches many of
+// these, but the project tree should keep the same hygiene without one).
+const FS_READDIR_HIDDEN = new Set([
+  '.git',
+  '.hg',
+  '.svn',
+  '.cache',
+  '.next',
+  '.turbo',
+  '.venv',
+  '__pycache__',
+  'build',
+  'dist',
+  'node_modules',
+  'target',
+  'venv'
+])
+
+function direntIsDirectory(dirent) {
+  return typeof dirent.isDirectory === 'function' && dirent.isDirectory()
+}
+
+function direntIsFile(dirent) {
+  return typeof dirent.isFile === 'function' && dirent.isFile()
+}
+
+function direntIsSymbolicLink(dirent) {
+  return typeof dirent.isSymbolicLink === 'function' && dirent.isSymbolicLink()
+}
+
+function shouldStatDirent(dirent) {
+  if (direntIsDirectory(dirent)) return false
+
+  return direntIsSymbolicLink(dirent) || !direntIsFile(dirent)
+}
+
+async function entryForDirent(dirent, resolved, fsImpl) {
+  const fullPath = path.join(resolved, dirent.name)
+  let isDirectory = direntIsDirectory(dirent)
+
+  if (!isDirectory && shouldStatDirent(dirent)) {
+    try {
+      isDirectory = (await fsImpl.promises.stat(fullPath)).isDirectory()
+    } catch {
+      isDirectory = false
+    }
+  }
+
+  return { name: dirent.name, path: fullPath, isDirectory }
+}
+
+async function mapWithStatConcurrency(items, mapper) {
+  const results = new Array(items.length)
+  let nextIndex = 0
+
+  async function runWorker() {
+    while (nextIndex < items.length) {
+      const index = nextIndex
+      nextIndex += 1
+      results[index] = await mapper(items[index])
+    }
+  }
+
+  const workerCount = Math.min(FS_READDIR_STAT_CONCURRENCY, items.length)
+  const workers = Array.from({ length: workerCount }, () => runWorker())
+  await Promise.all(workers)
+
+  return results
+}
+
+async function readDirForIpc(dirPath, options = {}) {
+  const fsImpl = options.fs || fs
+  let resolved
+
+  try {
+    ;({ resolvedPath: resolved } = await resolveDirectoryForIpc(dirPath, {
+      fs: fsImpl,
+      purpose: 'Directory read'
+    }))
+  } catch (error) {
+    return { entries: [], error: error?.code || 'read-error' }
+  }
+
+  try {
+    const dirents = await fsImpl.promises.readdir(resolved, { withFileTypes: true })
+    const visibleDirents = dirents.filter(dirent => !FS_READDIR_HIDDEN.has(dirent.name))
+    const entries = await mapWithStatConcurrency(visibleDirents, dirent =>
+      entryForDirent(dirent, resolved, fsImpl)
+    )
+
+    entries.sort((a, b) => Number(b.isDirectory) - Number(a.isDirectory) || a.name.localeCompare(b.name))
+
+    return { entries }
+  } catch (error) {
+    return { entries: [], error: error?.code || 'read-error' }
+  }
+}
+
+module.exports = {
+  readDirForIpc
+}

apps/desktop/electron/fs-read-dir.test.cjs

@@ -0,0 +1,364 @@
+'use strict'
+
+const assert = require('node:assert/strict')
+const fs = require('node:fs')
+const os = require('node:os')
+const path = require('node:path')
+const test = require('node:test')
+const { pathToFileURL } = require('node:url')
+
+const { readDirForIpc } = require('./fs-read-dir.cjs')
+
+function mkTmpDir() {
+  return fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-fs-read-dir-'))
+}
+
+function fakeDirent(name, flags = {}) {
+  return {
+    name,
+    isDirectory: () => Boolean(flags.directory),
+    isFile: () => Boolean(flags.file),
+    isSymbolicLink: () => Boolean(flags.symlink)
+  }
+}
+
+test('readDirForIpc hides noisy directories and files from the project tree', async () => {
+  const root = mkTmpDir()
+
+  try {
+    fs.mkdirSync(path.join(root, 'node_modules'))
+    fs.mkdirSync(path.join(root, 'src'))
+    fs.writeFileSync(path.join(root, 'target'), 'hidden file')
+    fs.writeFileSync(path.join(root, 'README.md'), 'visible file')
+
+    const result = await readDirForIpc(root)
+
+    assert.equal(result.error, undefined)
+    assert.deepEqual(
+      result.entries.map(entry => entry.name),
+      ['src', 'README.md']
+    )
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc filters a hidden basename whether it is a file or directory', async () => {
+  const dirRoot = mkTmpDir()
+  const fileRoot = mkTmpDir()
+
+  try {
+    fs.mkdirSync(path.join(dirRoot, 'node_modules'))
+    fs.writeFileSync(path.join(dirRoot, 'visible.txt'), 'visible')
+    fs.writeFileSync(path.join(fileRoot, 'node_modules'), 'hidden file')
+    fs.writeFileSync(path.join(fileRoot, 'visible.txt'), 'visible')
+
+    assert.deepEqual(
+      (await readDirForIpc(dirRoot)).entries.map(entry => entry.name),
+      ['visible.txt']
+    )
+    assert.deepEqual(
+      (await readDirForIpc(fileRoot)).entries.map(entry => entry.name),
+      ['visible.txt']
+    )
+  } finally {
+    fs.rmSync(dirRoot, { recursive: true, force: true })
+    fs.rmSync(fileRoot, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc returns directories before files and sorts by name within groups', async () => {
+  const root = mkTmpDir()
+
+  try {
+    fs.writeFileSync(path.join(root, 'z.txt'), 'z')
+    fs.mkdirSync(path.join(root, 'src'))
+    fs.writeFileSync(path.join(root, 'a.txt'), 'a')
+    fs.mkdirSync(path.join(root, 'lib'))
+
+    const result = await readDirForIpc(root)
+
+    assert.equal(result.error, undefined)
+    assert.deepEqual(
+      result.entries.map(entry => entry.name),
+      ['lib', 'src', 'a.txt', 'z.txt']
+    )
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc accepts file URLs for directories', async () => {
+  const root = mkTmpDir()
+
+  try {
+    fs.mkdirSync(path.join(root, 'src'))
+    fs.writeFileSync(path.join(root, 'README.md'), 'visible file')
+
+    const result = await readDirForIpc(pathToFileURL(root).toString())
+
+    assert.equal(result.error, undefined)
+    assert.deepEqual(
+      result.entries.map(entry => entry.name),
+      ['src', 'README.md']
+    )
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc returns invalid-path for blank or non-string input', async () => {
+  let readdirCalls = 0
+  const fsImpl = {
+    promises: {
+      readdir: async () => {
+        readdirCalls += 1
+        return []
+      }
+    }
+  }
+
+  assert.deepEqual(await readDirForIpc('', { fs: fsImpl }), { entries: [], error: 'invalid-path' })
+  assert.deepEqual(await readDirForIpc('   ', { fs: fsImpl }), { entries: [], error: 'invalid-path' })
+  assert.deepEqual(await readDirForIpc(null, { fs: fsImpl }), { entries: [], error: 'invalid-path' })
+  assert.equal(readdirCalls, 0)
+})
+
+test('readDirForIpc rejects Windows device paths before readdir', async () => {
+  let readdirCalls = 0
+  const fsImpl = {
+    promises: {
+      readdir: async () => {
+        readdirCalls += 1
+        return []
+      }
+    }
+  }
+
+  assert.deepEqual(await readDirForIpc('\\\\?\\C:\\secret', { fs: fsImpl }), {
+    entries: [],
+    error: 'device-path'
+  })
+  assert.equal(readdirCalls, 0)
+})
+
+test('readDirForIpc returns filesystem error codes instead of throwing', async () => {
+  const root = mkTmpDir()
+
+  try {
+    const result = await readDirForIpc(path.join(root, 'missing'))
+
+    assert.deepEqual(result, { entries: [], error: 'ENOENT' })
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc marks a symlink to a directory as a directory', async t => {
+  const root = mkTmpDir()
+
+  try {
+    fs.mkdirSync(path.join(root, 'actual-dir'))
+
+    try {
+      fs.symlinkSync(path.join(root, 'actual-dir'), path.join(root, 'linked-dir'), 'dir')
+    } catch (error) {
+      if (error?.code === 'EPERM' || error?.code === 'EACCES') {
+        t.skip(`symlink creation is not permitted on this platform (${error.code})`)
+
+        return
+      }
+
+      throw error
+    }
+
+    const result = await readDirForIpc(root)
+    const linked = result.entries.find(entry => entry.name === 'linked-dir')
+
+    assert.equal(result.error, undefined)
+    assert.equal(linked?.isDirectory, true)
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc marks a Windows junction to a directory as a directory', async t => {
+  if (process.platform !== 'win32') {
+    t.skip('junctions are a Windows-specific symlink type')
+
+    return
+  }
+
+  const root = mkTmpDir()
+
+  try {
+    fs.mkdirSync(path.join(root, 'actual-dir'))
+
+    try {
+      fs.symlinkSync(path.join(root, 'actual-dir'), path.join(root, 'junction-dir'), 'junction')
+    } catch (error) {
+      if (error?.code === 'EPERM' || error?.code === 'EACCES') {
+        t.skip(`junction creation is not permitted on this platform (${error.code})`)
+
+        return
+      }
+
+      throw error
+    }
+
+    const result = await readDirForIpc(root)
+    const junction = result.entries.find(entry => entry.name === 'junction-dir')
+
+    assert.equal(result.error, undefined)
+    assert.equal(junction?.isDirectory, true)
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc allows expanding symlink or junction directories outside the project root', async t => {
+  const root = mkTmpDir()
+  const outside = mkTmpDir()
+
+  try {
+    fs.writeFileSync(path.join(outside, 'outside.txt'), 'ok')
+
+    const linkPath = path.join(root, 'outside-link')
+    try {
+      fs.symlinkSync(outside, linkPath, process.platform === 'win32' ? 'junction' : 'dir')
+    } catch (error) {
+      if (error?.code === 'EPERM' || error?.code === 'EACCES') {
+        t.skip(`directory symlink creation is not permitted on this platform (${error.code})`)
+
+        return
+      }
+
+      throw error
+    }
+
+    const result = await readDirForIpc(linkPath)
+
+    assert.equal(result.error, undefined)
+    assert.deepEqual(result.entries, [
+      { name: 'outside.txt', path: path.join(linkPath, 'outside.txt'), isDirectory: false }
+    ])
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+    fs.rmSync(outside, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc stats symbolic links and unknown entries without dropping the whole listing', async () => {
+  const input = path.join('virtual-root')
+  const resolved = path.resolve(input)
+  const statCalls = []
+  const fsImpl = {
+    promises: {
+      readdir: async () => [
+        fakeDirent('unknown-entry'),
+        fakeDirent('linked-dir', { symlink: true }),
+        fakeDirent('broken-link', { symlink: true }),
+        fakeDirent('plain.txt', { file: true })
+      ],
+      stat: async fullPath => {
+        if (fullPath === resolved) {
+          return { isDirectory: () => true }
+        }
+
+        statCalls.push(fullPath)
+        if (fullPath.endsWith(`${path.sep}linked-dir`)) {
+          return { isDirectory: () => true }
+        }
+        throw Object.assign(new Error('gone'), { code: 'ENOENT' })
+      }
+    }
+  }
+
+  const result = await readDirForIpc(input, { fs: fsImpl })
+
+  assert.equal(result.error, undefined)
+  assert.deepEqual(
+    statCalls.sort(),
+    [path.join(resolved, 'broken-link'), path.join(resolved, 'linked-dir'), path.join(resolved, 'unknown-entry')].sort()
+  )
+  assert.deepEqual(result.entries, [
+    { name: 'linked-dir', path: path.join(resolved, 'linked-dir'), isDirectory: true },
+    { name: 'broken-link', path: path.join(resolved, 'broken-link'), isDirectory: false },
+    { name: 'plain.txt', path: path.join(resolved, 'plain.txt'), isDirectory: false },
+    { name: 'unknown-entry', path: path.join(resolved, 'unknown-entry'), isDirectory: false }
+  ])
+})
+
+test('readDirForIpc bounds concurrent stats while preserving complete sorted output', async () => {
+  const input = path.join('virtual-root')
+  const resolved = path.resolve(input)
+  const names = Array.from({ length: 105 }, (_, index) => `entry-${String(104 - index).padStart(3, '0')}`)
+  const failedName = 'entry-100'
+  const directoryNames = new Set(names.filter((_, index) => index % 10 === 4))
+  const successfulDirectoryNames = new Set([...directoryNames].filter(name => name !== failedName))
+  const statCalls = []
+  let active = 0
+  let peak = 0
+  let releaseStats
+  let markFirstStatStarted
+  const statsReleased = new Promise(resolve => {
+    releaseStats = resolve
+  })
+  const firstStatStarted = new Promise(resolve => {
+    markFirstStatStarted = resolve
+  })
+  const fsImpl = {
+    promises: {
+      readdir: async () => [
+        fakeDirent('node_modules', { symlink: true }),
+        ...names.map((name, index) => fakeDirent(name, { symlink: index % 2 === 0 }))
+      ],
+      stat: async fullPath => {
+        if (fullPath === resolved) {
+          return { isDirectory: () => true }
+        }
+
+        statCalls.push(fullPath)
+        active += 1
+        peak = Math.max(peak, active)
+        markFirstStatStarted()
+        await statsReleased
+        active -= 1
+
+        const name = path.basename(fullPath)
+        if (name === failedName) {
+          throw Object.assign(new Error('gone'), { code: 'ENOENT' })
+        }
+
+        return { isDirectory: () => successfulDirectoryNames.has(name) }
+      }
+    }
+  }
+
+  const resultPromise = readDirForIpc(input, { fs: fsImpl })
+  await firstStatStarted
+  await new Promise(resolve => setImmediate(resolve))
+  releaseStats()
+  const result = await resultPromise
+
+  const expectedNames = [
+    ...names.filter(name => successfulDirectoryNames.has(name)).sort(),
+    ...names.filter(name => !successfulDirectoryNames.has(name)).sort()
+  ]
+
+  assert.equal(result.error, undefined)
+  assert.equal(result.entries.length, names.length)
+  assert.equal(statCalls.length, names.length)
+  assert.equal(statCalls.some(fullPath => fullPath.endsWith(`${path.sep}node_modules`)), false)
+  assert.ok(peak > 1, `expected concurrent stats, observed peak ${peak}`)
+  assert.ok(peak <= 16, `expected at most 16 concurrent stats, observed peak ${peak}`)
+  assert.deepEqual(
+    result.entries.map(entry => entry.name),
+    expectedNames
+  )
+  assert.equal(result.entries.find(entry => entry.name === failedName)?.isDirectory, false)
+  assert.equal(
+    result.entries.filter(entry => entry.isDirectory).length,
+    successfulDirectoryNames.size
+  )
+})

apps/desktop/electron/git-root.cjs

@@ -0,0 +1,54 @@
+'use strict'
+
+const fs = require('node:fs')
+const path = require('node:path')
+const { resolveRequestedPathForIpc } = require('./hardening.cjs')
+
+function findGitRoot(start, fsImpl = fs) {
+  let dir = start
+
+  for (let i = 0; i < 50; i += 1) {
+    try {
+      if (fsImpl.existsSync(path.join(dir, '.git'))) {
+        return dir
+      }
+    } catch {
+      return null
+    }
+
+    const parent = path.dirname(dir)
+
+    if (parent === dir) {
+      return null
+    }
+
+    dir = parent
+  }
+
+  return null
+}
+
+async function gitRootForIpc(startPath, options = {}) {
+  const fsImpl = options.fs || fs
+  let resolved
+
+  try {
+    resolved = resolveRequestedPathForIpc(startPath, { purpose: 'Git root' })
+  } catch {
+    return null
+  }
+
+  try {
+    const stat = await fsImpl.promises.stat(resolved)
+    const start = stat.isDirectory() ? resolved : path.dirname(resolved)
+
+    return findGitRoot(start, fsImpl)
+  } catch {
+    return findGitRoot(resolved, fsImpl)
+  }
+}
+
+module.exports = {
+  findGitRoot,
+  gitRootForIpc
+}

apps/desktop/electron/git-root.test.cjs

@@ -0,0 +1,40 @@
+'use strict'
+
+const assert = require('node:assert/strict')
+const fs = require('node:fs')
+const os = require('node:os')
+const path = require('node:path')
+const test = require('node:test')
+const { pathToFileURL } = require('node:url')
+
+const { gitRootForIpc } = require('./git-root.cjs')
+
+function mkTmpDir() {
+  return fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-git-root-'))
+}
+
+test('gitRootForIpc returns null for invalid and device paths', async () => {
+  assert.equal(await gitRootForIpc(''), null)
+  assert.equal(await gitRootForIpc('   '), null)
+  assert.equal(await gitRootForIpc(null), null)
+  assert.equal(await gitRootForIpc('\\\\?\\C:\\secret'), null)
+  assert.equal(await gitRootForIpc('file:///%E0%A4%A'), null)
+})
+
+test('gitRootForIpc resolves directories files missing descendants and file URLs', async t => {
+  const root = mkTmpDir()
+  t.after(() => fs.rmSync(root, { recursive: true, force: true }))
+
+  const gitDir = path.join(root, '.git')
+  const srcDir = path.join(root, 'src')
+  const filePath = path.join(srcDir, 'index.ts')
+  fs.mkdirSync(gitDir)
+  fs.mkdirSync(srcDir)
+  fs.writeFileSync(filePath, 'export {}\n', 'utf8')
+
+  assert.equal(await gitRootForIpc(root), root)
+  assert.equal(await gitRootForIpc(srcDir), root)
+  assert.equal(await gitRootForIpc(filePath), root)
+  assert.equal(await gitRootForIpc(pathToFileURL(filePath).toString()), root)
+  assert.equal(await gitRootForIpc(path.join(srcDir, 'missing.ts')), root)
+})

apps/desktop/electron/git-worktrees.cjs

@@ -0,0 +1,174 @@
+'use strict'
+
+// Resolve git-worktree relationships for a set of session cwds, reading git's
+// on-disk metadata directly (no `git` spawn per path):
+//
+//   - A normal checkout has a `.git` DIRECTORY at its root → it's the main
+//     worktree; its repo root IS that directory's parent.
+//   - A linked worktree has a `.git` FILE: `gitdir: <repo>/.git/worktrees/<name>`.
+//     That admin dir's `commondir` points back at the shared `<repo>/.git`, whose
+//     parent is the main repo root.
+//
+// Grouping by repoRoot therefore clusters a repo's main checkout with all of its
+// linked worktrees, regardless of how the worktree directories are named. The
+// branch (read from the worktree's own HEAD) gives each worktree a meaningful
+// label.
+
+const fs = require('node:fs')
+const path = require('node:path')
+const { resolveRequestedPathForIpc } = require('./hardening.cjs')
+
+// Walk up from `start` to the nearest ancestor that carries a `.git` entry
+// (file for a linked worktree, dir for the main checkout). Capped so a stray
+// path can't loop forever.
+function findGitHost(start, fsImpl) {
+  let dir = start
+
+  for (let i = 0; i < 64; i += 1) {
+    const dotgit = path.join(dir, '.git')
+
+    try {
+      if (fsImpl.existsSync(dotgit)) {
+        return dir
+      }
+    } catch {
+      return null
+    }
+
+    const parent = path.dirname(dir)
+
+    if (parent === dir) {
+      return null
+    }
+
+    dir = parent
+  }
+
+  return null
+}
+
+function readBranch(gitDir, fsImpl) {
+  try {
+    const head = fsImpl.readFileSync(path.join(gitDir, 'HEAD'), 'utf8').trim()
+    const ref = head.match(/^ref:\s*refs\/heads\/(.+)$/)
+
+    if (ref) {
+      return ref[1]
+    }
+
+    // Detached HEAD: surface a short sha so the worktree still gets a label.
+    return /^[0-9a-f]{7,40}$/i.test(head) ? head.slice(0, 8) : null
+  } catch {
+    return null
+  }
+}
+
+// Given the directory that owns the `.git` entry, resolve its worktree identity.
+function resolveFromHost(host, fsImpl) {
+  const dotgit = path.join(host, '.git')
+  let stat
+
+  try {
+    stat = fsImpl.statSync(dotgit)
+  } catch {
+    return null
+  }
+
+  if (stat.isDirectory()) {
+    return {
+      repoRoot: host,
+      worktreeRoot: host,
+      isMainWorktree: true,
+      branch: readBranch(dotgit, fsImpl)
+    }
+  }
+
+  // Linked worktree: `.git` is a file pointing at the admin dir.
+  let contents
+
+  try {
+    contents = fsImpl.readFileSync(dotgit, 'utf8').trim()
+  } catch {
+    return null
+  }
+
+  const match = contents.match(/^gitdir:\s*(.+)$/m)
+
+  if (!match) {
+    return null
+  }
+
+  const adminDir = path.resolve(host, match[1].trim())
+
+  // `commondir` resolves to the shared `<repo>/.git`; fall back to walking two
+  // levels up from `<repo>/.git/worktrees/<name>` if it's missing.
+  let commonDir
+
+  try {
+    const rel = fsImpl.readFileSync(path.join(adminDir, 'commondir'), 'utf8').trim()
+    commonDir = path.resolve(adminDir, rel)
+  } catch {
+    commonDir = path.dirname(path.dirname(adminDir))
+  }
+
+  return {
+    repoRoot: path.dirname(commonDir),
+    worktreeRoot: host,
+    isMainWorktree: false,
+    branch: readBranch(adminDir, fsImpl)
+  }
+}
+
+function resolveWorktree(startPath, fsImpl = fs) {
+  let resolved
+
+  try {
+    resolved = resolveRequestedPathForIpc(startPath, { purpose: 'Worktree lookup' })
+  } catch {
+    return null
+  }
+
+  let start = resolved
+
+  try {
+    const stat = fsImpl.statSync(resolved)
+
+    if (!stat.isDirectory()) {
+      start = path.dirname(resolved)
+    }
+  } catch {
+    return null
+  }
+
+  const host = findGitHost(start, fsImpl)
+
+  if (!host) {
+    return null
+  }
+
+  return resolveFromHost(host, fsImpl)
+}
+
+// Batch entry point for the renderer: maps each requested cwd to its worktree
+// info (or null when it isn't inside a git checkout / can't be read). Dedupes so
+// many sessions sharing a cwd cost one lookup.
+async function worktreesForIpc(cwds, options = {}) {
+  const fsImpl = options.fs || fs
+  const list = Array.isArray(cwds) ? cwds : []
+  const out = {}
+
+  for (const cwd of list) {
+    if (typeof cwd !== 'string' || !cwd.trim() || cwd in out) {
+      continue
+    }
+
+    out[cwd] = resolveWorktree(cwd, fsImpl)
+  }
+
+  return out
+}
+
+module.exports = {
+  resolveWorktree,
+  worktreesForIpc
+}