docs(infographic): flaky fallback-cooldown timing fix

test_non_retryable_exhaustion_arms_cooldown captured `before` ahead of three
_try_activate_fallback() calls, then asserted the armed cooldown was
<= before + 5.0 + 1.0. The cooldown is set relative to the *final* call, and
the activation work (agent init, resolve_provider_client) can take >1s on a
loaded CI worker — so `before`-anchored upper bound overshot by ~0.5s and
failed reliably (slice 8/8, observed twice: 230.21 vs 229.73 bound,
905.80 vs 905.31 bound).

Anchor the upper bound to `after = time.monotonic()` captured once the
cooldown is armed. cooldown is final_call_time + 5.0 and after >= final_call_time,
so cooldown <= after + 5.0 + 1.0 holds regardless of activation latency. Still
proves the short 5s window vs the 60s rate-limit one (well under the +50s
discriminator in the sibling test). Test added in #53909.

.envrc

@@ -1,5 +1,5 @@
 watch_file pyproject.toml uv.lock
 watch_file package-lock.json package.json web/package.json ui-tui/package.json website/package.json apps/shared/package.json apps/desktop/package.json ui-tui/packages/hermes-ink/package.json
-watch_file flake.nix flake.lock nix/devShell.nix nix/tui.nix nix/package.nix nix/python.nix
+watch_file flake.nix flake.lock nix/devShell.nix nix/tui.nix nix/package.nix nix/python.nix nix/hermes-agent.nix nix/desktop.nix
 
 use flake

.github/actions/hermes-smoke-test/action.yml

@@ -1,50 +0,0 @@
-name: Hermes smoke test
-description: >
-  Run the image's built-in entrypoint against `--help` and `dashboard --help`
-  to catch basic runtime regressions before publishing.  Requires the image
-  to already be loaded into the local Docker daemon under `image`.
-
-  Works identically on amd64 and arm64 runners.
-
-inputs:
-  image:
-    description: Fully-qualified image tag (e.g. nousresearch/hermes-agent:test)
-    required: true
-
-runs:
-  using: composite
-  steps:
-    - name: Ensure /tmp/hermes-test is hermes-writable
-      shell: bash
-      run: |
-        # The image runs as the hermes user (UID 10000).  GitHub Actions
-        # creates /tmp/hermes-test root-owned by default, which hermes
-        # can't write to — chown it to match the in-container UID before
-        # bind-mounting.  Real users doing `docker run -v ~/.hermes:...`
-        # with their own UID hit the same issue and have their own
-        # remediations (HERMES_UID env var, or chown locally).
-        mkdir -p /tmp/hermes-test
-        sudo chown -R 10000:10000 /tmp/hermes-test
-
-    - name: hermes --help
-      shell: bash
-      run: |
-        # Use the image's real ENTRYPOINT (/init + main-wrapper.sh) so
-        # this exercises the actual production startup path. PR #30136
-        # review caught that an --entrypoint override here had been
-        # silently neutered by the s6-overlay migration — stage2-hook
-        # ignores its CMD args, so the smoke test was a no-op.
-        docker run --rm \
-          -v /tmp/hermes-test:/opt/data \
-          "${{ inputs.image }}" --help
-
-    - name: hermes dashboard --help
-      shell: bash
-      run: |
-        # Regression guard for #9153: dashboard was present in source but
-        # missing from the published image.  If this fails, something in
-        # the Dockerfile is excluding the dashboard subcommand from the
-        # installed package.
-        docker run --rm \
-          -v /tmp/hermes-test:/opt/data \
-          "${{ inputs.image }}" dashboard --help

.github/workflows/ci.yml

@@ -20,6 +20,7 @@ permissions:
   pull-requests: write # needed by lint (PR comment) + supply-chain (PR comment)
   actions: read # needed by osv-scanner (SARIF upload)
   security-events: write # needed by osv-scanner (SARIF upload)
+  packages: write # needed by docker build
 
 concurrency:
   group: ci-${{ github.ref }}
@@ -32,6 +33,7 @@ jobs:
   # (all lanes true) so post-merge validation is never weakened.
   # ─────────────────────────────────────────────────────────────────────
   detect:
+    name: Detect affected areas
     runs-on: ubuntu-latest
     outputs:
       python: ${{ steps.classify.outputs.python }}
@@ -53,11 +55,15 @@ jobs:
   # Skipped workflows (if condition is false) don't spin up runners.
   # ─────────────────────────────────────────────────────────────────────
   tests:
+    name: Python tests
     needs: detect
     if: needs.detect.outputs.python == 'true'
     uses: ./.github/workflows/tests.yml
+    with:
+      slice_count: 8
 
   lint:
+    name: Python lints
     needs: detect
     if: needs.detect.outputs.python == 'true'
     uses: ./.github/workflows/lint.yml
@@ -65,35 +71,49 @@ jobs:
       event_name: ${{ needs.detect.outputs.event_name }}
 
   typecheck:
+    name: TypeScript
     needs: detect
     if: needs.detect.outputs.frontend == 'true'
     uses: ./.github/workflows/typecheck.yml
 
   docs-site:
+    name: Docs Site
     needs: detect
     if: needs.detect.outputs.site == 'true'
     uses: ./.github/workflows/docs-site-checks.yml
 
   history-check:
+    name: Deny unrelated histories
     needs: detect
     if: needs.detect.outputs.event_name == 'pull_request'
     uses: ./.github/workflows/history-check.yml
 
   contributor-check:
+    name: Check contributors
     needs: detect
     if: needs.detect.outputs.python == 'true'
     uses: ./.github/workflows/contributor-check.yml
 
   uv-lockfile:
+    name: Check uv.lock
     needs: detect
     uses: ./.github/workflows/uv-lockfile-check.yml
 
   docker-lint:
+    name: Lint Docker scripts
     needs: detect
     if: needs.detect.outputs.docker_meta == 'true'
     uses: ./.github/workflows/docker-lint.yml
 
+  docker:
+    name: Build&Test Docker image
+    needs: detect
+    if: needs.detect.outputs.python == 'true' || needs.detect.outputs.frontend == 'true' || needs.detect.outputs.docker_meta == 'true'
+    uses: ./.github/workflows/docker.yml
+    secrets: inherit
+
   supply-chain:
+    name: Supply-chain scan
     needs: detect
     if: needs.detect.outputs.event_name == 'pull_request' && (needs.detect.outputs.scan == 'true' || needs.detect.outputs.deps == 'true' || needs.detect.outputs.mcp_catalog == 'true')
     uses: ./.github/workflows/supply-chain-audit.yml
@@ -104,7 +124,7 @@ jobs:
       mcp_catalog: ${{ needs.detect.outputs.mcp_catalog == 'true' }}
 
   osv-scanner:
-    needs: detect
+    name: OSV scan
     uses: ./.github/workflows/osv-scanner.yml
 
   # ─────────────────────────────────────────────────────────────────────
@@ -127,6 +147,8 @@ jobs:
       - docker-lint
       - supply-chain
       - osv-scanner
+      # we don't require docker to pass rn because it's so slow lol
+      # - docker
     if: always()
     runs-on: ubuntu-latest
     steps:

.github/workflows/docker-lint.yml

@@ -2,7 +2,7 @@ name: Docker / shell lint
 
 # Lints the container build inputs: Dockerfile (via hadolint) and any shell
 # scripts under docker/ (via shellcheck). These catch the class of regression
-# the behavioral docker-publish smoke test can't — unquoted variable
+# the behavioral docker smoke test can't — unquoted variable
 # expansions, silently-failing RUN commands, etc.
 #
 # Rules and ignores are documented in .hadolint.yaml at the repo root.

.github/workflows/docker.yml

@@ -1,24 +1,9 @@
-name: Docker Build and Publish
+name: Docker Build, Test, and Publish
 
 on:
-  push:
-    branches: [main]
-    paths:
-      - '**/*.py'
-      - 'pyproject.toml'
-      - 'uv.lock'
-      - 'Dockerfile'
-      - '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:
-
   release:
     types: [published]
+  workflow_call:
 
 permissions:
   contents: read
@@ -39,11 +24,7 @@ env:
   IMAGE_NAME: nousresearch/hermes-agent
 
 jobs:
-  # ---------------------------------------------------------------------------
-  # Build amd64 natively.  This job also runs the smoke tests (basic --help
-  # and the dashboard subcommand regression guard from #9153), because amd64
-  # is the only arch we can `load` into the local daemon on an amd64 runner.
-  # ---------------------------------------------------------------------------
+  # Build, test, and optionally push the amd64 image.
   build-amd64:
     # Only run on the upstream repository, not on forks
     if: github.repository == 'NousResearch/hermes-agent'
@@ -53,24 +34,19 @@ jobs:
       digest: ${{ steps.push.outputs.digest }}
     steps:
       - name: Checkout code
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
-      # The image build + smoke test + integration tests run ONLY on
-      # push-to-main and release — never on PRs. They are the heaviest jobs
-      # in CI (~15-45 min) and a broken build surfaces on the main push (and
-      # is gated pre-merge by docker-lint + uv-lockfile-check). Every step
-      # below is skipped on PRs, so the job still reports green and the
-      # required check never hangs.
+      # The image build + integration tests run on every event
+      # (PRs, push-to-main, release). Publish steps below are gated to
+      # push-to-main / release only.
       - name: Set up Docker Buildx
-        if: github.event_name != 'pull_request'
-        uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f  # v3
+        uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f # v3
 
-      # Build once, load into the local daemon for smoke testing.  Cached
+      # Build once, load into the local daemon for testing.  Cached
       # to gha with a per-arch scope; the push step below reuses every
       # layer from this build.
-      - name: Build image (amd64, smoke test)
-        if: github.event_name != 'pull_request'
-        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f  # v7.1.0
+      - name: Build image (amd64)
+        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
         with:
           context: .
           file: Dockerfile
@@ -82,25 +58,12 @@ jobs:
           cache-from: type=gha,scope=docker-amd64
           cache-to: type=gha,mode=max,scope=docker-amd64
 
-      - name: Smoke test image
-        if: github.event_name != 'pull_request'
-        uses: ./.github/actions/hermes-smoke-test
-        with:
-          image: ${{ env.IMAGE_NAME }}:test
-
-      # ---------------------------------------------------------------------
       # Run the docker-integration test suite against the freshly-built
-      # image already loaded into the local daemon (`:test`).  These tests
-      # are excluded from the sharded `tests.yml :: test` matrix on purpose
-      # (see `_SKIP_PARTS` in scripts/run_tests_parallel.py) because each
-      # shard would otherwise reach the session-scoped ``built_image``
-      # fixture in ``tests/docker/conftest.py`` and start a 3-7min
-      # ``docker build`` — guaranteed to
-      # die in fixture setup.
+      # image already loaded into the local daemon (`:test`).
       #
-      # Piggybacking here avoids a second image build: the smoke test
-      # already proved the image loads + runs, so the daemon has it under
-      # `${IMAGE_NAME}:test` and we just point ``HERMES_TEST_IMAGE`` at
+      # Piggybacking here avoids a second image build: the build step
+      # already loaded the image into the daemon under
+      # `${IMAGE_NAME}:test`, so we just point ``HERMES_TEST_IMAGE`` at
       # that.  The fixture's ``HERMES_TEST_IMAGE`` branch (see
       # tests/docker/conftest.py:62-63) short-circuits the rebuild.
       #
@@ -110,26 +73,20 @@ jobs:
       # cheapest path to coverage on every PR that touches docker code.
       # ---------------------------------------------------------------------
       - name: Install uv (for docker tests)
-        if: github.event_name != 'pull_request'
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86  # v5
+        uses: astral-sh/setup-uv@fac544c07dec837d0ccb6301d7b5580bf5edae39 # 8.2.0
 
       - name: Set up Python 3.11 (for docker tests)
-        if: github.event_name != 'pull_request'
         run: uv python install 3.11
 
       - name: Install Python dependencies (for docker tests)
-        if: github.event_name != 'pull_request'
         run: |
-          uv venv .venv --python 3.11
-          source .venv/bin/activate
           # ``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.
-          uv pip install -e ".[dev]"
+          uv sync --locked --python 3.11 --extra dev
 
       - name: Run docker integration tests
-        if: github.event_name != 'pull_request'
         env:
           # Skip rebuild; use the image already loaded by the build step.
           HERMES_TEST_IMAGE: ${{ env.IMAGE_NAME }}:test
@@ -139,12 +96,11 @@ jobs:
           OPENAI_API_KEY: ""
           NOUS_API_KEY: ""
         run: |
-          source .venv/bin/activate
-          python -m pytest tests/docker/ -v --tb=short
+          scripts/run_tests.sh tests/docker/ --file-timeout 600
 
       - name: Log in to Docker Hub
         if: github.event_name == 'push' && github.ref == 'refs/heads/main' || github.event_name == 'release'
-        uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121  # v4.1.0
+        uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121 # v4.1.0
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
@@ -155,7 +111,7 @@ jobs:
       - name: Push amd64 by digest
         id: push
         if: github.event_name == 'push' && github.ref == 'refs/heads/main' || github.event_name == 'release'
-        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f  # v7.1.0
+        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
         with:
           context: .
           file: Dockerfile
@@ -179,7 +135,7 @@ jobs:
 
       - name: Upload digest artifact
         if: github.event_name == 'push' && github.ref == 'refs/heads/main' || github.event_name == 'release'
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02  # v4
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7
         with:
           name: digest-amd64
           path: /tmp/digests/*
@@ -187,10 +143,7 @@ jobs:
           retention-days: 1
 
   # ---------------------------------------------------------------------------
-  # Build arm64 natively on GitHub's free arm64 runner.  This replaces the
-  # previous QEMU-emulated arm64 build, which was ~5-10x slower and shared
-  # a cache scope with amd64.  Matches the amd64 job's shape: build+load,
-  # smoke test, then on push/release push by digest.
+  # Build, test, and optionally push the arm64 image.
   # ---------------------------------------------------------------------------
   build-arm64:
     if: github.repository == 'NousResearch/hermes-agent'
@@ -200,29 +153,26 @@ jobs:
       digest: ${{ steps.push.outputs.digest }}
     steps:
       - name: Checkout code
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
-      # arm64 build runs only on push-to-main and release (see build-amd64).
       - name: Set up Docker Buildx
-        if: github.event_name != 'pull_request'
-        uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f  # v3
+        uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f # v3
 
       # Log in to ghcr.io so the registry-backed build cache below can be
       # read (cache-from) on every event and written (cache-to) on
       # push/release.  Uses the workflow's GITHUB_TOKEN, which is valid for
       # the whole job — unlike the gha cache backend's short-lived Azure SAS
       # token, which expired mid-build on slow cold-cache arm64 runs and
-      # crashed the build before the smoke test (the reason the gha cache
+      # crashed the build before the tests ran (the reason the gha cache
       # was removed from arm64 PRs in the first place).
       - name: Log in to ghcr.io (build cache)
-        if: github.event_name != 'pull_request'
-        uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121  # v4.1.0
+        uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121 # v4.1.0
         with:
           registry: ghcr.io
           username: ${{ github.actor }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
-      # Build once, load into the local daemon for smoke testing, then push
+      # Build once, load into the local daemon for testing, then push
       # by digest below. Reads AND writes the registry-backed cache so the
       # push reuses layers from this build and the next build starts warm.
       #
@@ -230,9 +180,8 @@ jobs:
       # cache that previously broke here: its credential is the job-lifetime
       # GITHUB_TOKEN, not a short-lived SAS token, so the cold-build-outlives-
       # token failure mode cannot recur.
-      - name: Build image (arm64, smoke test, cached publish)
-        if: github.event_name != 'pull_request'
-        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f  # v7.1.0
+      - name: Build image (arm64, cached publish)
+        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
         with:
           context: .
           file: Dockerfile
@@ -244,15 +193,29 @@ jobs:
           cache-from: type=registry,ref=ghcr.io/nousresearch/hermes-agent:buildcache-arm64
           cache-to: type=registry,ref=ghcr.io/nousresearch/hermes-agent:buildcache-arm64,mode=max
 
-      - name: Smoke test image
-        if: github.event_name != 'pull_request'
-        uses: ./.github/actions/hermes-smoke-test
-        with:
-          image: ${{ env.IMAGE_NAME }}:test
+      - name: Install uv for docker tests
+        uses: astral-sh/setup-uv@fac544c07dec837d0ccb6301d7b5580bf5edae39 # 8.2.0
+
+      - name: Set up Python 3.11 for docker tests
+        run: uv python install 3.11
+
+      - name: Install Python dependencies for docker tests
+        run: |
+          uv sync --locked --python 3.11 --extra dev
+
+      - name: Run docker tests
+        env:
+          # Skip rebuild; use the image already loaded by the build step.
+          HERMES_TEST_IMAGE: ${{ env.IMAGE_NAME }}:test
+          OPENROUTER_API_KEY: ""
+          OPENAI_API_KEY: ""
+          NOUS_API_KEY: ""
+        run: |
+          scripts/run_tests.sh tests/docker/ --file-timeout 600
 
       - name: Log in to Docker Hub
         if: github.event_name == 'push' && github.ref == 'refs/heads/main' || github.event_name == 'release'
-        uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121  # v4.1.0
+        uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121 # v4.1.0
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
@@ -260,7 +223,7 @@ jobs:
       - name: Push arm64 by digest
         id: push
         if: github.event_name == 'push' && github.ref == 'refs/heads/main' || github.event_name == 'release'
-        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f  # v7.1.0
+        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
         with:
           context: .
           file: Dockerfile
@@ -282,7 +245,7 @@ jobs:
 
       - name: Upload digest artifact
         if: github.event_name == 'push' && github.ref == 'refs/heads/main' || github.event_name == 'release'
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02  # v4
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7
         with:
           name: digest-arm64
           path: /tmp/digests/*
@@ -304,17 +267,17 @@ jobs:
     timeout-minutes: 10
     steps:
       - name: Download digests
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093  # v4
+        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4
         with:
           path: /tmp/digests
           pattern: digest-*
           merge-multiple: true
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f  # v3
+        uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f # v3
 
       - name: Log in to Docker Hub
-        uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121  # v4.1.0
+        uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121 # v4.1.0
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}

.github/workflows/lint.yml

@@ -37,7 +37,7 @@ jobs:
           fetch-depth: 0 # need full history for merge-base + worktree
 
       - name: Install uv
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5
+        uses: astral-sh/setup-uv@fac544c07dec837d0ccb6301d7b5580bf5edae39 # 8.2.0
 
       - name: Install ruff + ty
         uses: ./.github/actions/retry
@@ -110,7 +110,7 @@ jobs:
           cat .lint-reports/summary.md >> "$GITHUB_STEP_SUMMARY"
 
       - name: Upload reports as artifact
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7
         with:
           name: lint-reports
           path: .lint-reports/
@@ -164,7 +164,7 @@ jobs:
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
       - name: Install uv
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5
+        uses: astral-sh/setup-uv@fac544c07dec837d0ccb6301d7b5580bf5edae39 # 8.2.0
 
       - name: Install ruff
         uses: ./.github/actions/retry

.github/workflows/skills-index.yml

@@ -3,17 +3,17 @@ name: Build Skills Index
 on:
   schedule:
     # Run twice daily: 6 AM and 6 PM UTC
-    - cron: '0 6,18 * * *'
-  workflow_dispatch:  # Manual trigger
+    - cron: "0 6,18 * * *"
+  workflow_dispatch: # Manual trigger
   push:
     branches: [main]
     paths:
-      - 'scripts/build_skills_index.py'
-      - '.github/workflows/skills-index.yml'
+      - "scripts/build_skills_index.py"
+      - ".github/workflows/skills-index.yml"
 
 permissions:
   contents: read
-  actions: write   # to trigger deploy-site.yml on schedule
+  actions: write # to trigger deploy-site.yml on schedule
 
 jobs:
   build-index:
@@ -21,11 +21,11 @@ jobs:
     if: github.repository == 'NousResearch/hermes-agent'
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
-      - 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 dependencies
         run: pip install httpx==0.28.1 pyyaml==6.0.2
@@ -36,7 +36,7 @@ jobs:
         run: python scripts/build_skills_index.py
 
       - name: Upload index artifact
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02  # v4
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7
         with:
           name: skills-index
           path: website/static/api/skills-index.json

.github/workflows/tests.yml

@@ -2,6 +2,11 @@ name: Tests
 
 on:
   workflow_call:
+    inputs:
+      slice_count:
+        description: Number of parallel test slices
+        type: number
+        default: 8
 
 permissions:
   contents: read
@@ -12,13 +17,11 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  test:
+  generate:
+    name: "Generate slices"
     runs-on: ubuntu-latest
-    timeout-minutes: 30
-    strategy:
-      fail-fast: false
-      matrix:
-        slice: [1, 2, 3, 4, 5, 6]
+    outputs:
+      matrix: ${{ steps.matrix.outputs.matrix }}
     steps:
       - name: Checkout code
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -27,13 +30,26 @@ jobs:
         uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae # v5.0.5
         with:
           path: test_durations.json
-          # 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: Generate test slices
+        id: matrix
+        run: |
+          MATRIX=$(python3 scripts/run_tests_parallel.py --generate-slices ${{ inputs.slice_count }})
+          echo "matrix=$MATRIX" >> "$GITHUB_OUTPUT"
+
+  test:
+    name: Run tests slice ${{ matrix.slice.index }}/${{ inputs.slice_count }}
+    needs: generate
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    strategy:
+      fail-fast: false
+      matrix: ${{ fromJSON(needs.generate.outputs.matrix) }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+
       - name: Install ripgrep (prebuilt binary)
         run: |
           set -euo pipefail
@@ -49,7 +65,7 @@ jobs:
           rg --version
 
       - name: Install uv
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5
+        uses: astral-sh/setup-uv@fac544c07dec837d0ccb6301d7b5580bf5edae39 # 8.2.0
         with:
           # Persist uv's download/wheel cache (~/.cache/uv) across runs.
           # Keyed on the dependency manifests, so the cache is reused until
@@ -78,33 +94,19 @@ jobs:
         # 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
-        # every test_*.py file under tests/ (excluding integration/ + e2e/),
-        # then runs `python -m pytest <file>` in a freshly-spawned subprocess
+      - name: Run tests (slice ${{ matrix.slice.index }}/${{ inputs.slice_count }})
+        # Per-file isolation via scripts/run_tests.sh: each test file runs
+        # in its own freshly-spawned `python -m pytest <file>` subprocess
         # with bounded parallelism. No xdist, no shared workers, no
         # module-level state leakage between files.
         #
-        # Why per-file (not per-test): per-test spawn cost (~250ms × 17k
-        # tests = 70min CPU minimum) blew the wall-clock budget. Per-file
-        # spawn (~250ms × ~850 files = ~3.5min) fits while still giving
-        # every file a fresh interpreter — the only isolation boundary
-        # that matters in practice (cross-file leakage was the original
-        # flake source; intra-file is the test author's responsibility).
-        #
-        # Why drop xdist entirely: xdist's persistent workers accumulate
-        # state across files, which is exactly the leakage we wanted to
-        # fix. ThreadPoolExecutor + subprocess.run is ~60 lines and does
-        # the job with cleaner semantics.
-        #
-        # Matrix slicing (--slice I/N): files are distributed across 6
-        # jobs by cached duration (LPT algorithm) so each job gets
-        # roughly equal wall time. Without a cache, files default to 2s
-        # estimate and get split roughly evenly by count — still correct,
-        # just not perfectly balanced.
+        # File list is pre-computed by the generate job (--generate-slices)
+        # which runs LPT distribution once and passes the file list to each
+        # matrix job via --files. Previously each job re-discovered files and
+        # re-ran LPT independently — redundant N times.
         run: |
           source .venv/bin/activate
-          python scripts/run_tests_parallel.py --slice ${{ matrix.slice }}/6
+          scripts/run_tests.sh --files '${{ matrix.slice.files }}'
         env:
           # Ensure tests don't accidentally call real APIs
           OPENROUTER_API_KEY: ""
@@ -114,7 +116,7 @@ jobs:
       - name: Upload per-slice durations
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
-          name: test-durations-slice-${{ matrix.slice }}
+          name: test-durations-slice-${{ matrix.slice.index }}
           path: test_durations.json
           retention-days: 1
 
@@ -173,7 +175,7 @@ jobs:
           rg --version
 
       - name: Install uv
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5
+        uses: astral-sh/setup-uv@fac544c07dec837d0ccb6301d7b5580bf5edae39 # 8.2.0
         with:
           # Persist uv's download/wheel cache (~/.cache/uv) across runs.
           # Keyed on the dependency manifests, so the cache is reused until

.github/workflows/typecheck.yml

@@ -6,6 +6,7 @@ on:
 
 jobs:
   typecheck:
+    name: Check TypeScript
     runs-on: ubuntu-latest
     strategy:
       matrix:
@@ -22,8 +23,7 @@ jobs:
       # native builds. Skipping install scripts drops node-pty's node-gyp
       # header fetch — the transient flake that killed this job pre-`tsc` — and
       # is faster. retry covers the remaining registry blips.
-      - 
-        uses: ./.github/actions/retry
+      - uses: ./.github/actions/retry
         with:
           command: npm ci --ignore-scripts
       - run: npm run --prefix ${{ matrix.package }} typecheck
@@ -35,6 +35,7 @@ jobs:
   # 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:
+    name: Build desktop app
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -44,8 +45,7 @@ jobs:
           cache: npm
       # Keep install scripts here: the production build may need node-pty's
       # native binary. retry handles the transient install-time fetch flakes.
-      - 
-        uses: ./.github/actions/retry
+      - uses: ./.github/actions/retry
         with:
           command: npm ci
       - run: npm run --prefix apps/desktop build

.github/workflows/upload_to_pypi.yml

@@ -5,11 +5,11 @@ name: Publish to PyPI
 on:
   push:
     tags:
-      - 'v20*'  # CalVer tags: v2026.5.15, v2026.5.15.2, etc.
+      - "v20*" # CalVer tags: v2026.5.15, v2026.5.15.2, etc.
   workflow_dispatch:
     inputs:
       confirm_tag:
-        description: 'Tag to publish (e.g. v2026.5.15). Must already exist.'
+        description: "Tag to publish (e.g. v2026.5.15). Must already exist."
         required: true
         type: string
 
@@ -27,7 +27,7 @@ jobs:
     name: Build distribution 📦
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
         with:
           persist-credentials: false
           # On workflow_dispatch, check out the confirmed tag.
@@ -43,17 +43,17 @@ jobs:
           fi
 
       - name: Set up Python
-        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
         with:
-          python-version: '3.13'
+          python-version: "3.13"
 
       - name: Install uv
-        uses: astral-sh/setup-uv@d0cc045d04ccac9d8b7881df0226f9e82c39688e  # v6
+        uses: astral-sh/setup-uv@fac544c07dec837d0ccb6301d7b5580bf5edae39 # 8.2.0
 
       - name: Set up Node.js
-        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020  # v4
+        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
         with:
-          node-version: '22'
+          node-version: "22"
 
       - name: Build web dashboard
         run: cd web && npm ci && npm run build
@@ -81,7 +81,7 @@ jobs:
         run: uv build --sdist --wheel
 
       - name: Upload distribution artifacts
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02  # v4
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7
         with:
           name: python-package-distributions
           path: dist/
@@ -94,17 +94,17 @@ jobs:
       name: pypi
       url: https://pypi.org/p/hermes-agent
     permissions:
-      id-token: write  # OIDC trusted publishing
+      id-token: write # OIDC trusted publishing
 
     steps:
       - name: Download distribution artifacts
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093  # v4
+        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4
         with:
           name: python-package-distributions
           path: dist/
 
       - name: Publish to PyPI
-        uses: pypa/gh-action-pypi-publish@cef221092ed1bacb1cc03d23a2d87d1d172e277b  # v1.14.0
+        uses: pypa/gh-action-pypi-publish@cef221092ed1bacb1cc03d23a2d87d1d172e277b # v1.14.0
         with:
           skip-existing: true
 
@@ -116,12 +116,12 @@ jobs:
     needs: publish
     runs-on: ubuntu-latest
     permissions:
-      contents: write   # attach assets to the existing release
-      id-token: write   # sigstore signing
+      contents: write # attach assets to the existing release
+      id-token: write # sigstore signing
 
     steps:
       - name: Download distribution artifacts
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093  # v4
+        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4
         with:
           name: python-package-distributions
           path: dist/
@@ -145,7 +145,7 @@ jobs:
 
       - name: Sign with Sigstore
         if: env.skip_sign != 'true'
-        uses: sigstore/gh-action-sigstore-python@04cffa1d795717b140764e8b640de88853c92acc  # v3.3.0
+        uses: sigstore/gh-action-sigstore-python@04cffa1d795717b140764e8b640de88853c92acc # v3.3.0
         with:
           inputs: >-
             ./dist/*.tar.gz

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

@@ -4,7 +4,7 @@ name: uv.lock check
 # that modify pyproject.toml without regenerating uv.lock (or vice versa)
 # must not merge, because the Docker build's `uv sync --frozen` step will
 # fail on a stale lockfile and we'd rather catch it here than in the
-# docker-publish workflow on main.
+# docker workflow on main.
 #
 # ─────────────────────────────────────────────────────────────────────────
 # IMPORTANT: this check runs against the MERGED state, not just your branch
@@ -63,7 +63,7 @@ jobs:
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
       - name: Install uv
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5
+        uses: astral-sh/setup-uv@fac544c07dec837d0ccb6301d7b5580bf5edae39 # 8.2.0
 
       # `uv lock --check` re-resolves the project from pyproject.toml and
       # compares the result to uv.lock, exiting non-zero if they disagree.
@@ -100,7 +100,7 @@ jobs:
 
           This check is blocking because the Docker image build uses
           `uv sync --frozen --extra all`, which rejects stale lockfiles
-          — catching it here avoids a ~15 min failed docker-publish run
+          — catching it here avoids a ~15 min failed docker run
           on `main` post-merge.
           EOF
             echo "::error title=uv.lock out of sync::Run \`uv lock\` locally and commit the result. If on a PR, sync with main first."

Dockerfile

@@ -189,7 +189,13 @@ RUN cd web && npm run build && \
 
 # ---------- Source code ----------
 # .dockerignore excludes node_modules, so the installs above survive.
-COPY . .
+# --link decouples this layer from parents for cache purposes; --chmod bakes
+# the final read-only permissions at copy time so we skip the separate
+# `chmod -R` pass that previously walked ~30k files across the venv +
+# node_modules + source (21s amd64 / 222s arm64 — #49113).  `a+rX,go-w`
+# gives the non-root hermes user read + traverse but no write; root retains
+# write so the build steps below don't need chmod u+w dances.
+COPY --link --chmod=a+rX,go-w . .
 
 # ---------- Permissions ----------
 # Link hermes-agent itself (editable). Deps are already installed in the
@@ -197,19 +203,15 @@ COPY . .
 # resolution or downloads.
 RUN uv pip install --no-cache-dir --no-deps -e "."
 
-# Keep /opt/hermes immutable for the runtime hermes user. Hosted/container
-# instances must not be able to self-edit the installed source or venv; user
-# data, skills, plugins, config, logs, and dashboard uploads live under
-# /opt/data instead. Root can still repair the image during build/boot, but
-# supervised Hermes processes drop to the non-root hermes user.
+# Wire the exec shim and install-method stamp.  Files under /opt/hermes are
+# already root-owned (COPY, uv sync, npm install all run as root) and
+# read-only for the hermes user (go-w from the --chmod above).
+
 USER root
 RUN mkdir -p /opt/hermes/bin && \
     cp /opt/hermes/docker/hermes-exec-shim.sh /opt/hermes/bin/hermes && \
     chmod 0755 /opt/hermes/bin/hermes && \
-    printf 'docker\n' > /opt/hermes/.install_method && \
-    chown -R root:root /opt/hermes && \
-    chmod -R a+rX /opt/hermes && \
-    chmod -R a-w /opt/hermes
+    printf 'docker\n' > /opt/hermes/.install_method
 # The ``.install_method`` stamp is baked next to the running code (the install
 # tree), NOT into $HERMES_HOME. $HERMES_HOME (/opt/data) is a shared data
 # volume that is commonly bind-mounted from the host and even shared with a
@@ -236,13 +238,11 @@ RUN mkdir -p /opt/hermes/bin && \
 #
 # The arg is optional — local `docker build` without --build-arg simply
 # omits the file, and the runtime falls back to live-git lookup.  CI
-# (.github/workflows/docker-publish.yml) passes ${{ github.sha }} so
+# (.github/workflows/docker.yml) passes ${{ github.sha }} so
 # every published image has it.
 ARG HERMES_GIT_SHA=
 RUN if [ -n "${HERMES_GIT_SHA}" ]; then \
-        chmod u+w /opt/hermes && \
-        printf '%s\n' "${HERMES_GIT_SHA}" > /opt/hermes/.hermes_build_sha && \
-        chmod a-w /opt/hermes /opt/hermes/.hermes_build_sha; \
+        printf '%s\n' "${HERMES_GIT_SHA}" > /opt/hermes/.hermes_build_sha; \
     fi
 
 # ---------- s6-overlay service wiring ----------

acp_adapter/tools.py

@@ -74,7 +74,7 @@ _POLISHED_TOOLS = {
     "kanban_create", "kanban_show", "kanban_comment", "kanban_complete",
     "kanban_block", "kanban_link", "kanban_heartbeat",
     "yb_query_group_info", "yb_query_group_members", "yb_search_sticker",
-    "yb_send_dm", "yb_send_sticker", "mixture_of_agents",
+    "yb_send_dm", "yb_send_sticker",
 }
 
 

agent/agent_init.py

@@ -719,6 +719,55 @@ def init_agent(
                     print("🔑 Using credentials: Microsoft Entra ID")
                 elif isinstance(effective_key, str) and len(effective_key) > 12:
                     print(f"🔑 Using token: {effective_key[:8]}...{effective_key[-4:]}")
+    elif agent.provider == "moa":
+        from agent.moa_loop import MoAClient
+        agent.api_mode = "chat_completions"
+
+        # Route reference-model outputs to the agent's tool_progress_callback so
+        # every surface that already consumes it (CLI spinner/scrollback, TUI,
+        # desktop, gateway) can show each reference's answer as a labelled block
+        # before the aggregator acts. The facade emits "moa.reference" and
+        # "moa.aggregating" events; we forward them through the same callback
+        # the tool lifecycle uses. Best-effort and cache-safe — these are
+        # display-only events, they never touch the message history.
+        def _moa_reference_relay(event: str, **kwargs: Any) -> None:
+            cb = getattr(agent, "tool_progress_callback", None)
+            if cb is None:
+                return
+            try:
+                if event == "moa.reference":
+                    label = str(kwargs.get("label") or "")
+                    text = str(kwargs.get("text") or "")
+                    idx = kwargs.get("index")
+                    count = kwargs.get("count")
+                    cb(
+                        "moa.reference",
+                        label,
+                        text,
+                        None,
+                        moa_index=idx,
+                        moa_count=count,
+                    )
+                elif event == "moa.aggregating":
+                    cb(
+                        "moa.aggregating",
+                        str(kwargs.get("aggregator") or ""),
+                        None,
+                        None,
+                        moa_ref_count=kwargs.get("ref_count"),
+                    )
+            except Exception:
+                pass
+
+        agent.client = MoAClient(
+            agent.model or "default",
+            reference_callback=_moa_reference_relay,
+        )
+        agent._client_kwargs = {}
+        agent.api_key = api_key or "moa-virtual-provider"
+        agent.base_url = "moa://local"
+        if not agent.quiet_mode:
+            print(f"🤖 AI Agent initialized with MoA preset: {agent.model}")
     elif agent.api_mode == "bedrock_converse":
         # AWS Bedrock — uses boto3 directly, no OpenAI client needed.
         # Region is extracted from the base_url or defaults to us-east-1.
@@ -1621,8 +1670,10 @@ def init_agent(
             f"Model {agent.model} has a context window of {_ctx:,} tokens, "
             f"which is below the minimum {MINIMUM_CONTEXT_LENGTH:,} required "
             f"by Hermes Agent.  Choose a model with at least "
-            f"{MINIMUM_CONTEXT_LENGTH // 1000}K context, or set "
-            f"model.context_length in config.yaml to override."
+            f"{MINIMUM_CONTEXT_LENGTH // 1000}K context.  If your server "
+            f"reports a window smaller than the model's true window, set "
+            f"model.context_length in config.yaml to the real value "
+            f"(this must be at least {MINIMUM_CONTEXT_LENGTH // 1000}K)."
         )
 
     # Inject context engine tool schemas (e.g. lcm_grep, lcm_describe, lcm_expand).

agent/agent_runtime_helpers.py

@@ -42,6 +42,14 @@ from utils import base_url_host_matches, base_url_hostname, env_var_enabled, ato
 logger = logging.getLogger(__name__)
 
 
+# Max consecutive successful credential-pool token refreshes of the SAME entry
+# on a persistent auth failure before we give up and let the fallback chain
+# activate. A single-entry OAuth pool can re-mint a fresh token indefinitely
+# even when the upstream keeps rejecting it, so without this cap the retry loop
+# spins forever and never reaches ``_try_activate_fallback``. See #26080.
+_MAX_AUTH_REFRESH_ATTEMPTS = 2
+
+
 def _ra():
     """Lazy ``run_agent`` reference for test-patch routing."""
     import run_agent
@@ -775,6 +783,30 @@ def recover_with_credential_pool(
             return False, has_retried_429
         refreshed = pool.try_refresh_current()
         if refreshed is not None:
+            # ``try_refresh_current()`` re-mints a fresh OAuth token and reports
+            # success even when the upstream keeps rejecting it — a single-entry
+            # pool (common for OAuth/Max subscribers) has nothing to rotate to,
+            # so a bare "refreshed → retry" loop spins forever on the same dead
+            # token and the configured fallback never activates. Cap consecutive
+            # same-entry refreshes and fall through to fallback once exceeded.
+            # See #26080.
+            refreshed_id = getattr(refreshed, "id", None)
+            if refreshed_id is not None:
+                refresh_counts = getattr(agent, "_auth_pool_refresh_counts", None)
+                if refresh_counts is None:
+                    refresh_counts = {}
+                    agent._auth_pool_refresh_counts = refresh_counts
+                refresh_key = (agent.provider, refreshed_id)
+                refresh_counts[refresh_key] = refresh_counts.get(refresh_key, 0) + 1
+                if refresh_counts[refresh_key] > _MAX_AUTH_REFRESH_ATTEMPTS:
+                    _ra().logger.warning(
+                        "Credential auth failure persists after %s refreshes for "
+                        "pool entry %s — treating as unrecoverable and allowing "
+                        "fallback to activate.",
+                        refresh_counts[refresh_key] - 1,
+                        refreshed_id,
+                    )
+                    return False, has_retried_429
             _ra().logger.info(f"Credential auth failure — refreshed pool entry {getattr(refreshed, 'id', '?')}")
             agent._swap_credential(refreshed)
             return True, has_retried_429
@@ -1420,6 +1452,15 @@ def create_openai_client(agent, client_kwargs: dict, *, reason: str, shared: boo
         keepalive_http = agent._build_keepalive_http_client(client_kwargs.get("base_url", ""))
         if keepalive_http is not None:
             client_kwargs["http_client"] = keepalive_http
+    # Delegate all rate-limit / 5xx retry to hermes's outer conversation loop,
+    # which honors Retry-After and applies adaptive/jittered backoff. The OpenAI
+    # SDK default (max_retries=2) uses its own 1-2s backoff that ignores
+    # Retry-After and double-retries inside our loop — the same deadlock the
+    # Anthropic clients hit (#26293). This is the single chokepoint every primary
+    # OpenAI/aggregator client passes through (init, switch_model, recovery,
+    # restore, request-scoped); auxiliary_client builds its own clients and keeps
+    # SDK retries because it is NOT wrapped by the conversation loop.
+    client_kwargs.setdefault("max_retries", 0)
     # Uses the module-level `OpenAI` name, resolved lazily on first
     # access via __getattr__ below. Tests patch via `run_agent.OpenAI`.
     client = _ra().OpenAI(**client_kwargs)
@@ -1499,6 +1540,10 @@ def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mo
     # _client_kwargs is a dict — snapshot a shallow copy so mutating the
     # live dict doesn't poison the rollback target.
     _snapshot["_client_kwargs"] = dict(getattr(agent, "_client_kwargs", {}) or {})
+    # Snapshot the credential pool reference so a failed client rebuild can
+    # restore the original pool (issue #52727: pool reload is part of this
+    # switch and must be reversible on rollback).
+    _snapshot["_credential_pool"] = getattr(agent, "_credential_pool", _MISSING)
 
     try:
         # Clear the per-config context_length override so the new model's
@@ -1523,8 +1568,36 @@ def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mo
         if api_key:
             agent.api_key = api_key
 
+        # ── Reload credential pool for the new provider (issue #52727) ──
+        # Without this, ``recover_with_credential_pool`` sees a
+        # ``pool.provider != agent.provider`` mismatch and short-circuits,
+        # leaving the new provider with no rotation/recovery on 401/429 and
+        # burning the original pool's entries. Only reload when the provider
+        # actually changed (or the pool was missing) — re-selecting the same
+        # provider must not churn the pool reference. A reload failure is
+        # logged + swallowed: the switch itself must still complete.
+        old_norm = (old_provider or "").strip().lower()
+        new_norm = (new_provider or "").strip().lower()
+        if old_norm != new_norm or getattr(agent, "_credential_pool", None) is None:
+            try:
+                from agent.credential_pool import load_pool
+                agent._credential_pool = load_pool(new_provider)
+            except Exception as _pool_exc:  # noqa: BLE001
+                logger.warning(
+                    "switch_model: credential pool reload failed for %s (%s); "
+                    "continuing without pool rotation this turn",
+                    new_provider, _pool_exc,
+                )
+
         # ── Build new client ──
-        if api_mode == "anthropic_messages":
+        if (new_provider or "").strip().lower() == "moa":
+            from agent.moa_loop import MoAClient
+
+            agent.api_key = api_key or "moa-virtual-provider"
+            agent.base_url = "moa://local"
+            agent._client_kwargs = {}
+            agent.client = MoAClient(agent.model or "default")
+        elif api_mode == "anthropic_messages":
             from agent.anthropic_adapter import (
                 build_anthropic_client,
                 resolve_anthropic_token,
@@ -1697,6 +1770,27 @@ def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mo
         old_model, old_provider, new_model, new_provider,
     )
 
+    # ── Persist billing route to session DB ──
+    # The agent's _session_db / session_id may not be set in all contexts
+    # (tests, bare agents without a session DB, etc.).  This ensures the
+    # dashboard Model cards show the actual provider after a mid-session
+    # /model switch instead of the stale session-creation provider.
+    # See #48248 for the full bug description.
+    _session_db = getattr(agent, "_session_db", None)
+    _session_id = getattr(agent, "session_id", None)
+    if _session_db is not None and _session_id:
+        try:
+            _session_db.update_session_billing_route(
+                _session_id,
+                provider=agent.provider,
+                base_url=agent.base_url,
+                billing_mode=getattr(agent, "api_mode", None),
+            )
+        except Exception:
+            logger.warning(
+                "Failed to persist billing route after model switch",
+                exc_info=True,
+            )
 
 
 def invoke_tool(agent, function_name: str, function_args: dict, effective_task_id: str,

agent/anthropic_adapter.py

@@ -673,6 +673,9 @@ def _build_anthropic_client_with_bearer_hook(
     kwargs = {
         "timeout": timeout_obj,
         "http_client": http_client,
+        # Delegate retry to hermes's outer loop (honors Retry-After); the SDK
+        # default max_retries=2 ignores it and double-retries. (#26293)
+        "max_retries": 0,
         # The SDK requires *something* for api_key/auth_token. Our
         # event hook overrides Authorization per request so this value
         # is never sent. The sentinel string makes accidental leaks
@@ -757,6 +760,12 @@ def build_anthropic_client(
     _read_timeout = timeout if (isinstance(timeout, (int, float)) and timeout > 0) else 900.0
     kwargs = {
         "timeout": Timeout(timeout=float(_read_timeout), connect=10.0),
+        # Delegate all rate-limit / 5xx retry to hermes's outer conversation
+        # loop, which honors Retry-After. The SDK default (max_retries=2) uses
+        # its own 1-2s backoff that ignores Retry-After and double-retries
+        # inside our loop — burning request slots against a bucket that won't
+        # refill for minutes. (#26293)
+        "max_retries": 0,
     }
     if normalized_base_url:
         # Azure Anthropic endpoints require an ``api-version`` query parameter.
@@ -852,6 +861,9 @@ def build_anthropic_bedrock_client(region: str):
     return _anthropic_sdk.AnthropicBedrock(
         aws_region=region,
         timeout=Timeout(timeout=900.0, connect=10.0),
+        # Delegate retry to hermes's outer loop (honors Retry-After); the SDK
+        # default max_retries=2 ignores it and double-retries. (#26293)
+        max_retries=0,
         default_headers={"anthropic-beta": ",".join([*_COMMON_BETAS, _CONTEXT_1M_BETA])},
     )
 
@@ -914,44 +926,72 @@ def _read_claude_code_credentials_from_keychain() -> Optional[Dict[str, Any]]:
     return None
 
 
+def _read_claude_code_credentials_from_file() -> Optional[Dict[str, Any]]:
+    """Read Claude Code OAuth credentials from ~/.claude/.credentials.json.
+
+    Returns dict with {accessToken, refreshToken?, expiresAt?, source} or None.
+    """
+    cred_path = Path.home() / ".claude" / ".credentials.json"
+    if not cred_path.exists():
+        return None
+    try:
+        data = json.loads(cred_path.read_text(encoding="utf-8"))
+    except (json.JSONDecodeError, OSError, IOError) as e:
+        logger.debug("Failed to read ~/.claude/.credentials.json: %s", e)
+        return None
+
+    oauth_data = data.get("claudeAiOauth")
+    if not (oauth_data and isinstance(oauth_data, dict)):
+        return None
+    access_token = oauth_data.get("accessToken", "")
+    if not access_token:
+        return None
+    return {
+        "accessToken": access_token,
+        "refreshToken": oauth_data.get("refreshToken", ""),
+        "expiresAt": oauth_data.get("expiresAt", 0),
+        "source": "claude_code_credentials_file",
+    }
+
+
 def read_claude_code_credentials() -> Optional[Dict[str, Any]]:
     """Read refreshable Claude Code OAuth credentials.
 
-    Checks two sources in order:
+    Reads from two possible sources and reconciles them:
       1. macOS Keychain (Darwin only) — "Claude Code-credentials" entry
       2. ~/.claude/.credentials.json file
 
+    Selection rules when both are present:
+      - If exactly one is non-expired, prefer that one. (Handles the case
+        where Claude Code refreshes one source but not the other — observed
+        in the wild on Claude Code 2.1.x.)
+      - Otherwise, prefer the source with the later ``expiresAt`` so that
+        any subsequent refresh uses the most recent ``refreshToken``.
+
     This intentionally excludes ~/.claude.json primaryApiKey. Opencode's
     subscription flow is OAuth/setup-token based with refreshable credentials,
     and native direct Anthropic provider usage should follow that path rather
     than auto-detecting Claude's first-party managed key.
 
-    Returns dict with {accessToken, refreshToken?, expiresAt?} or None.
+    Returns dict with {accessToken, refreshToken?, expiresAt?, source} or None.
     """
-    # Try macOS Keychain first (covers Claude Code >=2.1.114)
     kc_creds = _read_claude_code_credentials_from_keychain()
-    if kc_creds:
-        return kc_creds
+    file_creds = _read_claude_code_credentials_from_file()
 
-    # Fall back to JSON file
-    cred_path = Path.home() / ".claude" / ".credentials.json"
-    if cred_path.exists():
-        try:
-            data = json.loads(cred_path.read_text(encoding="utf-8"))
-            oauth_data = data.get("claudeAiOauth")
-            if oauth_data and isinstance(oauth_data, dict):
-                access_token = oauth_data.get("accessToken", "")
-                if access_token:
-                    return {
-                        "accessToken": access_token,
-                        "refreshToken": oauth_data.get("refreshToken", ""),
-                        "expiresAt": oauth_data.get("expiresAt", 0),
-                        "source": "claude_code_credentials_file",
-                    }
-        except (json.JSONDecodeError, OSError, IOError) as e:
-            logger.debug("Failed to read ~/.claude/.credentials.json: %s", e)
+    if kc_creds and file_creds:
+        kc_valid = is_claude_code_token_valid(kc_creds)
+        file_valid = is_claude_code_token_valid(file_creds)
+        if kc_valid and not file_valid:
+            return kc_creds
+        if file_valid and not kc_valid:
+            return file_creds
+        # Both valid or both expired: prefer the later expiresAt so the
+        # downstream refresh path uses the freshest refresh_token.
+        kc_exp = kc_creds.get("expiresAt", 0) or 0
+        file_exp = file_creds.get("expiresAt", 0) or 0
+        return kc_creds if kc_exp >= file_exp else file_creds
 
-    return None
+    return kc_creds or file_creds
 
 
 def is_claude_code_token_valid(creds: Dict[str, Any]) -> bool:
@@ -1034,8 +1074,40 @@ def refresh_anthropic_oauth_pure(refresh_token: str, *, use_json: bool = False)
 
 
 def _refresh_oauth_token(creds: Dict[str, Any]) -> Optional[str]:
-    """Attempt to refresh an expired Claude Code OAuth token."""
-    refresh_token = creds.get("refreshToken", "")
+    """Attempt to refresh an expired Claude Code OAuth token.
+
+    Claude Code's OAuth refresh tokens are single-use: a successful refresh
+    rotates the pair and invalidates the old refresh token. Claude Code itself
+    also refreshes on its own schedule (IDE/CLI activity), so by the time
+    Hermes notices an expired token, Claude Code may have already rotated it.
+    POSTing our now-stale refresh token in that window races Claude Code and
+    fails with ``invalid_grant``.
+
+    So before refreshing, re-read the live credential sources. If Claude Code
+    has already produced a valid token, adopt it and skip the POST entirely.
+    Only fall back to refreshing ourselves when no fresh credential is found.
+    """
+    # Claude Code may have already refreshed — adopt its token rather than
+    # racing it with our (possibly already-rotated) refresh token. Only adopt
+    # when the live re-read produced a DIFFERENT token with a real future
+    # expiry: re-adopting the same credential we were just handed would be a
+    # no-op, and a 0/absent ``expiresAt`` means "managed key / unknown expiry"
+    # (see is_claude_code_token_valid) which must NOT be treated as a fresh
+    # refresh here.
+    current = read_claude_code_credentials()
+    if current:
+        current_token = current.get("accessToken", "")
+        current_exp = current.get("expiresAt", 0) or 0
+        if (
+            current_token
+            and current_token != creds.get("accessToken", "")
+            and current_exp > 0
+            and is_claude_code_token_valid(current)
+        ):
+            logger.debug("Adopted Claude Code's already-refreshed OAuth token")
+            return current_token
+
+    refresh_token = (current or {}).get("refreshToken", "") or creds.get("refreshToken", "")
     if not refresh_token:
         logger.debug("No refresh token available — cannot refresh")
         return None

agent/auxiliary_client.py

@@ -666,6 +666,28 @@ def _pool_runtime_base_url(entry: Any, fallback: str = "") -> str:
     return str(url or "").strip().rstrip("/")
 
 
+# Hostnames (lowercase, exact) that the auxiliary Anthropic path is allowed to
+# be pointed at via config.yaml model.base_url. Anything else falls back to the
+# Anthropic default — operators routing main-session traffic through a
+# non-Anthropic host (e.g. OpenRouter, OpenAI) with provider=anthropic in config
+# must NOT have that foreign host leak into the auxiliary client. See #52608.
+_ANTHROPIC_COMPATIBLE_HOSTS = frozenset({
+    "api.anthropic.com",
+})
+
+
+def _is_anthropic_compatible_host(url: str) -> bool:
+    """Return True if ``url``'s hostname is an Anthropic endpoint we trust for aux calls."""
+    if not url:
+        return False
+    try:
+        from urllib.parse import urlparse
+        host = (urlparse(url).hostname or "").strip().lower().rstrip(".")
+        return host in _ANTHROPIC_COMPATIBLE_HOSTS
+    except Exception:
+        return False
+
+
 def _nous_min_key_ttl_seconds() -> int:
     try:
         return max(60, int(os.getenv("HERMES_NOUS_MIN_KEY_TTL_SECONDS", "1800")))
@@ -1647,13 +1669,14 @@ def _try_openrouter(explicit_api_key: str = None, model: str = None) -> Tuple[Op
     pool_present, entry = _select_pool_entry("openrouter")
     if pool_present:
         or_key = explicit_api_key or _pool_runtime_api_key(entry)
-        if not or_key:
-            _mark_provider_unhealthy("openrouter", ttl=60)
-            return None, None
-        base_url = _pool_runtime_base_url(entry, OPENROUTER_BASE_URL) or OPENROUTER_BASE_URL
-        logger.debug("Auxiliary client: OpenRouter via pool")
-        return OpenAI(api_key=or_key, base_url=base_url,
-                       default_headers=build_or_headers()), model or _OPENROUTER_MODEL
+        if or_key:
+            base_url = _pool_runtime_base_url(entry, OPENROUTER_BASE_URL) or OPENROUTER_BASE_URL
+            logger.debug("Auxiliary client: OpenRouter via pool")
+            return OpenAI(api_key=or_key, base_url=base_url,
+                           default_headers=build_or_headers()), model or _OPENROUTER_MODEL
+        # Pool exists but is exhausted (no usable runtime key) — fall through to
+        # the OPENROUTER_API_KEY env-var path rather than failing outright.
+        logger.debug("Auxiliary client: OpenRouter pool exhausted, trying OPENROUTER_API_KEY")
 
     or_key = explicit_api_key or os.getenv("OPENROUTER_API_KEY")
     if not or_key:
@@ -2256,9 +2279,16 @@ def _try_anthropic(explicit_api_key: str = None) -> Tuple[Optional[Any], Optiona
     if not token:
         return None, None
 
-    # Allow base URL override from config.yaml model.base_url, but only
-    # when the configured provider is anthropic — otherwise a non-Anthropic
-    # base_url (e.g. Codex endpoint) would leak into Anthropic requests.
+    # Allow base URL override from config.yaml model.base_url, but only when:
+    #   1. the configured provider is anthropic (otherwise a non-Anthropic
+    #      base_url, e.g. Codex endpoint, would leak into Anthropic requests), AND
+    #   2. the override URL actually points at an Anthropic-compatible endpoint.
+    # Without gate (2), operators who route main-session traffic through a
+    # non-Anthropic provider that accepts Anthropic-format requests (e.g.
+    # OpenRouter at openrouter.ai/api/v1, with provider=anthropic in config.yaml)
+    # would have every auxiliary side-channel call (memory extractors,
+    # reflection, vision, title generation) 401 from the foreign host —
+    # see issue #52608.
     base_url = _pool_runtime_base_url(entry, _ANTHROPIC_DEFAULT_BASE_URL) if pool_present else _ANTHROPIC_DEFAULT_BASE_URL
     try:
         from hermes_cli.config import load_config
@@ -2268,7 +2298,7 @@ def _try_anthropic(explicit_api_key: str = None) -> Tuple[Optional[Any], Optiona
             cfg_provider = str(model_cfg.get("provider") or "").strip().lower()
             if cfg_provider == "anthropic":
                 cfg_base_url = (model_cfg.get("base_url") or "").strip().rstrip("/")
-                if cfg_base_url:
+                if cfg_base_url and _is_anthropic_compatible_host(cfg_base_url):
                     base_url = cfg_base_url
     except Exception:
         pass
@@ -2754,6 +2784,25 @@ def _is_model_incompatible_error(exc: Exception) -> bool:
     ))
 
 
+def _is_invalid_aux_response_error(exc: Exception) -> bool:
+    """Detect provider responses that authenticated but cannot serve aux shape.
+
+    Some OpenAI-compatible routes return HTTP 200 with an empty/malformed
+    ChatCompletion instead of a normal provider error.  That is still a
+    provider/model capability failure for auxiliary tasks: downstream callers
+    need ``choices[0].message`` and should be able to continue through the
+    same fallback path as explicit model-incompatibility errors.
+    """
+    if not isinstance(exc, RuntimeError):
+        return False
+    msg = str(exc).lower()
+    return (
+        "auxiliary " in msg
+        and "llm returned invalid response" in msg
+        and "choices[0].message" in msg
+    )
+
+
 def _evict_cached_clients(provider: str) -> None:
     """Drop cached auxiliary clients for a provider so fresh creds are used."""
     normalized = _normalize_aux_provider(provider)
@@ -3576,6 +3625,37 @@ def _resolve_auto(
     # config.yaml (auxiliary.<task>.provider) still win over this.
     main_provider = str(runtime_provider or _read_main_provider() or "")
     main_model = str(runtime_model or _read_main_model() or "")
+
+    # MoA virtual provider: the "model" is a preset name (e.g. "opus-gpt") and
+    # there is no real "moa" HTTP endpoint, so resolving an aux client against
+    # provider="moa"/model=<preset> sends the preset name as the model id and
+    # the provider 400s ("opus-gpt is not a valid model ID"). Auxiliary tasks
+    # (title generation, compression, vision, …) don't need the reference
+    # fan-out — they should run on the aggregator, which is the preset's acting
+    # model. Resolve the MoA preset to its aggregator slot and continue Step 1
+    # with that real provider+model. Mirrors the MoA context-length resolution.
+    if main_provider == "moa":
+        try:
+            from hermes_cli.config import load_config
+            from hermes_cli.moa_config import resolve_moa_preset
+
+            _preset = resolve_moa_preset(load_config().get("moa") or {}, main_model)
+            _agg = _preset.get("aggregator") or {}
+            _agg_provider = str(_agg.get("provider") or "").strip()
+            _agg_model = str(_agg.get("model") or "").strip()
+            if _agg_provider and _agg_model and _agg_provider.lower() != "moa":
+                main_provider = _agg_provider
+                main_model = _agg_model
+                # The MoA virtual runtime carries a non-HTTP base_url
+                # ("moa://local") and a placeholder api_key; they belong to the
+                # facade, not the aggregator's real provider. Drop them so the
+                # aggregator resolves through its own provider credentials.
+                runtime_base_url = ""
+                runtime_api_key = ""
+                runtime_api_mode = ""
+        except Exception:
+            logger.debug("MoA aux resolution to aggregator failed", exc_info=True)
+
     if (main_provider and main_model
             and main_provider not in {"auto", ""}):
         resolved_provider = main_provider
@@ -5445,6 +5525,9 @@ def _validate_llm_response(response: Any, task: str = None) -> Any:
         if not choices or not hasattr(choices[0], "message"):
             raise AttributeError("missing choices[0].message")
     except (AttributeError, TypeError, IndexError) as exc:
+        recovered = _recover_aux_response_message(response)
+        if recovered is not None:
+            return recovered
         response_type = type(response).__name__
         response_preview = str(response)[:120]
         raise RuntimeError(
@@ -5456,6 +5539,64 @@ def _validate_llm_response(response: Any, task: str = None) -> Any:
     return response
 
 
+def _recover_aux_response_message(response: Any) -> Optional[Any]:
+    """Synthesize chat-completions shape from Responses-style text fields.
+
+    Auxiliary callers consume ``choices[0].message``.  Some compatible
+    endpoints return text outside ``choices`` (for example ``output_text`` or
+    ``output`` items).  Preserve that response before declaring it malformed.
+    """
+    text = _extract_aux_response_text(response)
+    if not text:
+        return None
+
+    choice = SimpleNamespace(
+        message=SimpleNamespace(content=text),
+        finish_reason=getattr(response, "finish_reason", None) or "stop",
+    )
+    try:
+        response.choices = [choice]
+        return response
+    except Exception:
+        return SimpleNamespace(
+            id=getattr(response, "id", ""),
+            model=getattr(response, "model", ""),
+            object=getattr(response, "object", "chat.completion"),
+            choices=[choice],
+            usage=getattr(response, "usage", None),
+        )
+
+
+def _extract_aux_response_text(response: Any) -> str:
+    output_text = _obj_get(response, "output_text")
+    if isinstance(output_text, str) and output_text.strip():
+        return output_text.strip()
+
+    output = _obj_get(response, "output")
+    if not isinstance(output, list):
+        return ""
+
+    parts: List[str] = []
+    for item in output:
+        item_type = _obj_get(item, "type")
+        if item_type and item_type != "message":
+            continue
+        for part in (_obj_get(item, "content") or []):
+            part_type = _obj_get(part, "type")
+            if part_type in {"output_text", "text", None}:
+                text = _obj_get(part, "text")
+                if isinstance(text, str) and text.strip():
+                    parts.append(text.strip())
+    return "\n".join(parts).strip()
+
+
+def _obj_get(obj: Any, key: str, default: Any = None) -> Any:
+    value = getattr(obj, key, default)
+    if value is default and isinstance(obj, dict):
+        value = obj.get(key, default)
+    return value
+
+
 def call_llm(
     task: str = None,
     *,
@@ -5853,11 +5994,21 @@ def call_llm(
         # When the provider returns a 429 rate-limit (not billing), fall
         # back to an alternative provider instead of exhausting retries
         # against the same rate-limited endpoint.
+        #
+        # ── Auth error fallback (#21165) ─────────────────────────────
+        # When the resolved provider returns 401 and neither the Nous
+        # refresh path nor explicit provider credential refresh applies,
+        # fall back to an alternative provider instead of dropping the
+        # auxiliary task on the floor (silent compression failure /
+        # message loss). Auth is NOT a capacity error: it only bypasses
+        # the explicit-provider gate when the user is in auto mode.
         should_fallback = (
-            _is_payment_error(first_err)
+            _is_auth_error(first_err)
+            or _is_payment_error(first_err)
             or _is_connection_error(first_err)
             or _is_rate_limit_error(first_err)
             or _is_model_incompatible_error(first_err)
+            or _is_invalid_aux_response_error(first_err)
         )
         # Respect explicit provider choice for transient errors (auth, request
         # validation, etc.) but allow fallback when the provider clearly cannot
@@ -5880,9 +6031,12 @@ def call_llm(
             or _is_connection_error(first_err)
             or _is_rate_limit_error(first_err)
             or _is_model_incompatible_error(first_err)
+            or _is_invalid_aux_response_error(first_err)
         )
         if should_fallback and (is_auto or is_capacity_error):
-            if _is_payment_error(first_err):
+            if _is_auth_error(first_err):
+                reason = "auth error"
+            elif _is_payment_error(first_err):
                 reason = "payment error"
                 # Resolve the actual provider label (resolved_provider may be
                 # "auto"; the client's base_url tells us which backend got the
@@ -5895,6 +6049,8 @@ def call_llm(
                 reason = "rate limit"
             elif _is_model_incompatible_error(first_err):
                 reason = "model incompatible with route"
+            elif _is_invalid_aux_response_error(first_err):
+                reason = "invalid provider response"
             else:
                 reason = "connection error"
             logger.info("Auxiliary %s: %s on %s (%s), trying fallback",
@@ -6329,11 +6485,17 @@ async def async_call_llm(
                         raise
 
         # ── Payment / connection / rate-limit fallback (mirrors sync call_llm) ──
+        # Auth error fallback (#21165): a 401 that survived the refresh path
+        # falls back in auto mode just like the sync call_llm() path. Auth is
+        # NOT a capacity error, so on an explicit provider it still respects
+        # the user's choice (handled by the is_auto/is_capacity_error gate).
         should_fallback = (
-            _is_payment_error(first_err)
+            _is_auth_error(first_err)
+            or _is_payment_error(first_err)
             or _is_connection_error(first_err)
             or _is_rate_limit_error(first_err)
             or _is_model_incompatible_error(first_err)
+            or _is_invalid_aux_response_error(first_err)
         )
         # Capacity errors (payment/quota/connection/rate-limit) bypass the
         # explicit-provider gate — the provider cannot serve the request
@@ -6348,9 +6510,12 @@ async def async_call_llm(
             or _is_connection_error(first_err)
             or _is_rate_limit_error(first_err)
             or _is_model_incompatible_error(first_err)
+            or _is_invalid_aux_response_error(first_err)
         )
         if should_fallback and (is_auto or is_capacity_error):
-            if _is_payment_error(first_err):
+            if _is_auth_error(first_err):
+                reason = "auth error"
+            elif _is_payment_error(first_err):
                 reason = "payment error"
                 _mark_provider_unhealthy(
                     _recoverable_pool_provider(resolved_provider, client) or resolved_provider
@@ -6359,6 +6524,8 @@ async def async_call_llm(
                 reason = "rate limit"
             elif _is_model_incompatible_error(first_err):
                 reason = "model incompatible with route"
+            elif _is_invalid_aux_response_error(first_err):
+                reason = "invalid provider response"
             else:
                 reason = "connection error"
             logger.info("Auxiliary %s (async): %s on %s (%s), trying fallback",

agent/chat_completion_helpers.py

@@ -37,6 +37,18 @@ from tools.terminal_tool import is_persistent_env
 from utils import base_url_host_matches, base_url_hostname, env_float, env_int
 
 logger = logging.getLogger(__name__)
+_OPENROUTER_PROVIDER_SORT_VALUES = {"throughput", "latency", "price"}
+
+# When the fallback chain is fully exhausted on a non-rate-limit failure
+# (e.g. every provider returns a non-retryable client error like HTTP 400),
+# arm a short cooldown so the NEXT turn's restore_primary_runtime stays gated
+# and does not reset _fallback_index=0 to replay the entire chain again.
+# Without this, a client/gateway that re-submits immediately would re-marshal
+# the full (potentially 80k-token) context once per provider every turn and
+# can drive a constrained host into memory/swap exhaustion.  Rate-limit /
+# billing reasons keep their own 60s cooldown (set above); this is the
+# narrower non-rate-limit case.  See issue #24996.
+_FALLBACK_EXHAUSTED_COOLDOWN_S = 5.0
 
 
 def _ra():
@@ -115,6 +127,23 @@ def _is_openai_codex_backend(agent) -> bool:
     )
 
 
+def _validated_openrouter_provider_sort(raw_sort: Any) -> Optional[str]:
+    """Return a normalized OpenRouter provider.sort value or None."""
+    if not isinstance(raw_sort, str):
+        return None
+    sort_value = raw_sort.strip().lower()
+    if not sort_value:
+        return None
+    if sort_value in _OPENROUTER_PROVIDER_SORT_VALUES:
+        return sort_value
+    logger.warning(
+        "Ignoring invalid OpenRouter provider.sort value %r (allowed: %s)",
+        raw_sort,
+        ", ".join(sorted(_OPENROUTER_PROVIDER_SORT_VALUES)),
+    )
+    return None
+
+
 def _env_float(name: str, default: float) -> float:
     try:
         return float(os.getenv(name, str(default)))
@@ -229,6 +258,11 @@ def interruptible_api_call(agent, api_kwargs: dict):
                         invalidate_runtime_client(region)
                     raise
                 result["response"] = normalize_converse_response(raw_response)
+            elif agent.provider == "moa":
+                # MoA is a virtual chat-completions provider backed by the
+                # in-process MoAClient facade. Do not rebuild a request-local
+                # OpenAI client from the virtual runtime metadata.
+                result["response"] = agent.client.chat.completions.create(**api_kwargs)
             else:
                 request_client = _set_request_client(
                     agent._create_request_openai_client(
@@ -698,8 +732,9 @@ def build_api_kwargs(agent, api_messages: list) -> dict:
         _prefs["ignore"] = agent.providers_ignored
     if agent.providers_order:
         _prefs["order"] = agent.providers_order
-    if agent.provider_sort:
-        _prefs["sort"] = agent.provider_sort
+    _provider_sort = _validated_openrouter_provider_sort(agent.provider_sort)
+    if _provider_sort:
+        _prefs["sort"] = _provider_sort
     if agent.provider_require_parameters:
         _prefs["require_parameters"] = True
     if agent.provider_data_collection:
@@ -1093,8 +1128,22 @@ def try_activate_fallback(agent, reason: "FailoverReason | None" = None) -> bool
         if (not fallback_already_active) or (primary_provider and current_provider == primary_provider):
             agent._rate_limited_until = time.monotonic() + 60
     if agent._fallback_index >= len(agent._fallback_chain):
+        # Chain exhausted.  If we actually walked a non-empty chain and the
+        # failure was NOT a rate-limit/billing event (those already armed
+        # their own 60s cooldown above), arm a short cooldown so the next
+        # turn's restore_primary_runtime stays gated instead of resetting
+        # _fallback_index=0 and re-marshaling the whole context across every
+        # provider again.  Guards the cross-turn replay storm in #24996.
+        if (
+            len(agent._fallback_chain) > 0
+            and reason not in {FailoverReason.rate_limit, FailoverReason.billing}
+        ):
+            _existing_cooldown = getattr(agent, "_rate_limited_until", 0) or 0
+            agent._rate_limited_until = max(
+                _existing_cooldown,
+                time.monotonic() + _FALLBACK_EXHAUSTED_COOLDOWN_S,
+            )
         return False
-
     fb = agent._fallback_chain[agent._fallback_index]
     agent._fallback_index += 1
     fb_provider = (fb.get("provider") or "").strip().lower()
@@ -1210,14 +1259,16 @@ def try_activate_fallback(agent, reason: "FailoverReason | None" = None) -> bool
             agent._transport_cache.clear()
         agent._fallback_activated = True
 
-        # Clear the credential pool when the fallback provider doesn't match
-        # the pool's provider.  The pool was seeded for the primary provider;
-        # leaving it attached means downstream recovery (rate_limit / billing /
-        # auth) calls ``_swap_credential`` with a primary entry which overwrites
-        # the agent's ``base_url`` back to the primary's endpoint — every
-        # fallback request then 404s against the wrong host.  See #33163.
+        # Rebind the credential pool to the fallback provider when the provider
+        # changes.  Keeping the primary pool attached would make downstream
+        # recovery (rate_limit / billing / auth) mutate the wrong credential
+        # set and can overwrite the fallback's base_url back to the primary
+        # endpoint.  See #33163.
+        #
         # When the fallback shares the pool's provider (e.g. both openrouter
-        # entries with different routing) the pool is preserved.
+        # entries with different routing) the pool is preserved.  When the
+        # providers differ, load the fallback provider's own pool if one exists
+        # so provider-specific rotation continues to work after the switch.
         _existing_pool = getattr(agent, "_credential_pool", None)
         if _existing_pool is not None:
             _pool_provider = (getattr(_existing_pool, "provider", "") or "").strip().lower()
@@ -1228,6 +1279,22 @@ def try_activate_fallback(agent, reason: "FailoverReason | None" = None) -> bool
                     fb_provider, fb_model, _pool_provider,
                 )
                 agent._credential_pool = None
+        if getattr(agent, "_credential_pool", None) is None:
+            try:
+                from agent.credential_pool import load_pool
+
+                fallback_pool = load_pool(fb_provider)
+                if fallback_pool and fallback_pool.has_credentials():
+                    agent._credential_pool = fallback_pool
+                    logger.info(
+                        "Fallback to %s/%s: attached fallback credential pool",
+                        fb_provider, fb_model,
+                    )
+            except Exception as exc:
+                logger.debug(
+                    "Fallback to %s/%s: could not attach credential pool: %s",
+                    fb_provider, fb_model, exc,
+                )
 
         # Honor per-provider / per-model request_timeout_seconds for the
         # fallback target (same knob the primary client uses).  None = use
@@ -1458,8 +1525,9 @@ def handle_max_iterations(agent, messages: list, api_call_count: int) -> str:
                 provider_preferences["ignore"] = agent.providers_ignored
             if agent.providers_order:
                 provider_preferences["order"] = agent.providers_order
-            if agent.provider_sort:
-                provider_preferences["sort"] = agent.provider_sort
+            _provider_sort = _validated_openrouter_provider_sort(agent.provider_sort)
+            if _provider_sort:
+                provider_preferences["sort"] = _provider_sort
             if provider_preferences and (
                 (agent.provider or "").strip().lower() == "openrouter"
                 or agent._is_openrouter_url()
@@ -2391,12 +2459,19 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
                             diag=request_client_holder.get("diag"),
                         )
                         _close_request_client_once("stream_mid_tool_retry_cleanup")
-                        try:
-                            agent._replace_primary_openai_client(
-                                reason="stream_mid_tool_retry_pool_cleanup"
-                            )
-                        except Exception:
-                            pass
+                        if agent.api_mode == "anthropic_messages":
+                            try:
+                                agent._anthropic_client.close()
+                                agent._rebuild_anthropic_client()
+                            except Exception:
+                                pass
+                        else:
+                            try:
+                                agent._replace_primary_openai_client(
+                                    reason="stream_mid_tool_retry_pool_cleanup"
+                                )
+                            except Exception:
+                                pass
                         continue
 
                     # SSE error events from proxies (e.g. OpenRouter sends
@@ -2444,12 +2519,19 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
                             _close_request_client_once("stream_retry_cleanup")
                             # Also rebuild the primary client to purge
                             # any dead connections from the pool.
-                            try:
-                                agent._replace_primary_openai_client(
-                                    reason="stream_retry_pool_cleanup"
-                                )
-                            except Exception:
-                                pass
+                            if agent.api_mode == "anthropic_messages":
+                                try:
+                                    agent._anthropic_client.close()
+                                    agent._rebuild_anthropic_client()
+                                except Exception:
+                                    pass
+                            else:
+                                try:
+                                    agent._replace_primary_openai_client(
+                                        reason="stream_retry_pool_cleanup"
+                                    )
+                                except Exception:
+                                    pass
                             continue
                         # Retries exhausted. Log the final failure with
                         # full diagnostic detail (chain, headers,
@@ -2561,6 +2643,17 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
             _stream_stale_timeout = max(_stream_stale_timeout_base, 240.0)
         else:
             _stream_stale_timeout = _stream_stale_timeout_base
+        # Reasoning-model floor: known reasoning models (Nemotron 3 Ultra,
+        # OpenAI o1/o3, Anthropic Opus 4.x thinking, DeepSeek R1, Qwen QwQ,
+        # xAI Grok reasoning, etc.) routinely exceed the default 180s chat-
+        # model threshold during their thinking phase.  The cloud gateway
+        # upstream kills the socket first, surfacing as BrokenPipeError.
+        # Raises the floor only — never overrides explicit user config
+        # (handled by get_provider_stale_timeout above).
+        from agent.reasoning_timeouts import get_reasoning_stale_timeout_floor
+        _reasoning_floor = get_reasoning_stale_timeout_floor(api_kwargs.get("model"))
+        if _reasoning_floor is not None:
+            _stream_stale_timeout = max(_stream_stale_timeout, _reasoning_floor)
 
     t = threading.Thread(target=_call, daemon=True)
     t.start()
@@ -2609,10 +2702,17 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
                 pass
             # Rebuild the primary client too — its connection pool
             # may hold dead sockets from the same provider outage.
-            try:
-                agent._replace_primary_openai_client(reason="stale_stream_pool_cleanup")
-            except Exception:
-                pass
+            if agent.api_mode == "anthropic_messages":
+                try:
+                    agent._anthropic_client.close()
+                    agent._rebuild_anthropic_client()
+                except Exception:
+                    pass
+            else:
+                try:
+                    agent._replace_primary_openai_client(reason="stale_stream_pool_cleanup")
+                except Exception:
+                    pass
             # Reset the timer so we don't kill repeatedly while
             # the inner thread processes the closure.
             last_chunk_time["t"] = time.time()

agent/coding_context.py

@@ -83,6 +83,59 @@ _PROJECT_MARKERS = (
 # Agent-instruction files surfaced separately from manifests in the snapshot.
 _CONTEXT_FILES = ("AGENTS.md", "CLAUDE.md", ".cursorrules")
 
+# Source-file extensions that make a git repo a *code* workspace even with no
+# manifest. Without this, `git init` on a notes/writing/research folder (a huge
+# non-coding use case) would flip the whole session into the coding posture just
+# for having a `.git`. A manifest still wins on its own (see `_PROJECT_MARKERS`).
+_CODE_EXTENSIONS = frozenset({
+    ".py", ".pyi", ".ipynb", ".js", ".jsx", ".ts", ".tsx", ".mjs", ".cjs",
+    ".go", ".rs", ".java", ".kt", ".kts", ".scala", ".rb", ".php", ".c", ".h",
+    ".cc", ".cpp", ".hpp", ".cs", ".swift", ".m", ".mm", ".dart", ".ex", ".exs",
+    ".lua", ".sh", ".bash", ".zsh", ".sql", ".vue", ".svelte", ".r", ".jl",
+    ".hs", ".clj", ".erl", ".pl",
+})
+
+# Dirs never worth scanning for the code check (deps/build/vcs/venv noise).
+_CODE_SCAN_SKIP_DIRS = frozenset({
+    ".git", "node_modules", "venv", ".venv", "__pycache__", "dist", "build",
+    "target", ".next", ".turbo", "vendor",
+})
+
+# Bounded sweep: a code workspace reveals itself in the first handful of entries.
+_CODE_SCAN_MAX_ENTRIES = 500
+
+
+def _has_code_files(root: Path) -> bool:
+    """Cheap, bounded check for source files in a repo's top two levels.
+
+    Lets a git repo of loose scripts (no manifest) still read as a code
+    workspace while a bare notes/writing repo does not. Scans the root and its
+    immediate subdirectories only, capped at ``_CODE_SCAN_MAX_ENTRIES`` stats —
+    a handful of readdirs at session start, not a full walk.
+    """
+    seen = 0
+    stack = [(root, True)]
+    while stack:
+        directory, is_root = stack.pop()
+        try:
+            with os.scandir(directory) as entries:
+                for entry in entries:
+                    seen += 1
+                    if seen > _CODE_SCAN_MAX_ENTRIES:
+                        return False
+                    name = entry.name
+                    try:
+                        if entry.is_file():
+                            if os.path.splitext(name)[1].lower() in _CODE_EXTENSIONS:
+                                return True
+                        elif is_root and entry.is_dir() and name not in _CODE_SCAN_SKIP_DIRS and not name.startswith("."):
+                            stack.append((Path(entry.path), False))
+                    except OSError:
+                        continue
+        except OSError:
+            continue
+    return False
+
 # Lockfile → package manager, checked in priority order.
 _PY_LOCKFILES = (("uv.lock", "uv"), ("poetry.lock", "poetry"), ("Pipfile.lock", "pipenv"))
 _JS_LOCKFILES = (
@@ -368,10 +421,16 @@ def _detect_profile_name(mode: str, platform: str, cwd_str: str) -> str:
     if platform and platform.strip().lower() not in INTERACTIVE_CODING_PLATFORMS:
         return GENERAL_PROFILE.name
     cwd = Path(cwd_str)
+    # A recognized project root (manifest / AGENTS.md / .cursorrules) is a code
+    # workspace on its own — cheap stat checks, no scan.
+    if _marker_root(cwd) is not None:
+        return CODING_PROFILE.name
     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:
+    # A bare git repo only counts when it actually holds code, so `git init` on a
+    # notes/writing/research folder stays in the general posture.
+    if git_root is not None and _has_code_files(git_root):
         return CODING_PROFILE.name
     return GENERAL_PROFILE.name
 

agent/conversation_compression.py

@@ -288,6 +288,29 @@ def replay_compression_warning(agent: Any) -> None:
             pass
 
 
+def conversation_history_after_compression(agent: Any, messages: list) -> Optional[list]:
+    """Return the correct flush baseline after a compression boundary.
+
+    Legacy compression rotates to a fresh child session. That child has not
+    seen the compacted transcript through the normal same-turn flush path yet,
+    so callers must clear ``conversation_history`` to ``None`` and let the next
+    persistence call write the whole compacted list.
+
+    In-place compaction is different: ``archive_and_compact()`` has already
+    soft-archived the previous active rows and inserted ``messages`` as the new
+    active live transcript under the same session id. If the same agent turn
+    continues with ``conversation_history=None``, the identity-based flush path
+    treats those already-persisted compacted dicts as new and appends them a
+    second time, doubling the active context and retriggering compression.
+
+    A shallow copy is intentional: it captures the current compacted dict
+    identities as history while allowing later same-turn appends to remain new.
+    """
+    if bool(getattr(agent, "_last_compaction_in_place", False)):
+        return list(messages)
+    return None
+
+
 def compress_context(
     agent: Any,
     messages: list,

agent/conversation_loop.py

@@ -28,6 +28,7 @@ import uuid
 from typing import Any, Dict, List, Optional
 
 from agent.codex_responses_adapter import _summarize_user_message_for_log
+from agent.conversation_compression import conversation_history_after_compression
 from agent.display import KawaiiSpinner
 from agent.error_classifier import FailoverReason, classify_api_error
 from agent.iteration_budget import IterationBudget
@@ -502,6 +503,7 @@ def run_conversation(
     stream_callback: Optional[callable] = None,
     persist_user_message: Optional[str] = None,
     persist_user_timestamp: Optional[float] = None,
+    moa_config: Optional[dict[str, Any]] = None,
 ) -> Dict[str, Any]:
     """
     Run a complete conversation with tool calling until completion.
@@ -524,6 +526,19 @@ def run_conversation(
     Returns:
         Dict: Complete conversation result with final response and message history
     """
+    if moa_config is None:
+        try:
+            from hermes_cli.moa_config import decode_moa_turn
+
+            _decoded_message, _decoded_moa_config = decode_moa_turn(user_message)
+            if _decoded_moa_config is not None:
+                user_message = _decoded_message
+                moa_config = _decoded_moa_config
+                if persist_user_message is None:
+                    persist_user_message = _decoded_message
+        except Exception:
+            pass
+
     # ── Per-turn setup (the prologue) ──
     # All once-per-turn setup — stdio guarding, retry-counter resets, user
     # message sanitization, todo/nudge hydration, system-prompt restore-or-
@@ -573,6 +588,13 @@ def run_conversation(
     compression_attempts = 0
     _turn_exit_reason = "unknown"  # Diagnostic: why the loop ended
 
+    # Per-turn tally of consecutive successful credential-pool token refreshes,
+    # keyed by (provider, pool-entry-id). A persistent upstream 401 lets
+    # ``try_refresh_current()`` "succeed" forever on a single-entry OAuth pool,
+    # so this tally caps same-entry refreshes and lets the fallback chain take
+    # over instead of spinning. Reset here so each turn starts fresh. See #26080.
+    agent._auth_pool_refresh_counts = {}
+
     # Optional opt-in runtime: if api_mode == codex_app_server, hand the
     # turn to the codex app-server subprocess (terminal/file ops/patching
     # all run inside Codex). Default Hermes path is bypassed entirely.
@@ -802,6 +824,28 @@ def run_conversation(
         if effective_system:
             api_messages = [{"role": "system", "content": effective_system}] + api_messages
 
+        if moa_config:
+            try:
+                from agent.moa_loop import aggregate_moa_context
+
+                _moa_context = aggregate_moa_context(
+                    user_prompt=original_user_message if isinstance(original_user_message, str) else str(original_user_message),
+                    api_messages=api_messages,
+                    reference_models=moa_config.get("reference_models") or [],
+                    aggregator=moa_config.get("aggregator") or {},
+                    temperature=float(moa_config.get("reference_temperature", 0.6) or 0.6),
+                    aggregator_temperature=float(moa_config.get("aggregator_temperature", 0.4) or 0.4),
+                )
+                if _moa_context:
+                    for _msg in reversed(api_messages):
+                        if _msg.get("role") == "user":
+                            _base = _msg.get("content", "")
+                            if isinstance(_base, str):
+                                _msg["content"] = _base + "\n\n" + _moa_context
+                            break
+            except Exception as _moa_exc:
+                logger.warning("MoA context aggregation failed: %s", _moa_exc)
+
         # Inject ephemeral prefill messages right after the system prompt
         # but before conversation history. Same API-call-time-only pattern.
         if agent.prefill_messages:
@@ -1123,7 +1167,7 @@ def run_conversation(
                 # stream.  Mirror the ACP exclusion used for Responses
                 # API upgrade (lines ~1083-1085).
                 elif (
-                    agent.provider == "copilot-acp"
+                    agent.provider in {"copilot-acp", "moa"}
                     or str(agent.base_url or "").lower().startswith("acp://copilot")
                     or str(agent.base_url or "").lower().startswith("acp+tcp://")
                 ):
@@ -1974,9 +2018,21 @@ def run_conversation(
                     agent.thinking_callback("")
                 api_elapsed = time.time() - api_start_time
                 agent._vprint(f"{agent.log_prefix}⚡ Interrupted during API call.", force=True)
-                agent._persist_session(messages, conversation_history)
                 interrupted = True
-                final_response = f"{INTERRUPT_WAITING_FOR_MODEL_PREFIX}{api_elapsed:.1f}s elapsed)."
+                # Preserve any assistant text already streamed to the user
+                # before the stop landed. Dropping it leaves history with no
+                # record of the half-finished reply on screen, so the next turn
+                # the model "forgets" what it just said — exactly what users hit
+                # when they stop to redirect mid-response.
+                _partial = agent._strip_think_blocks(
+                    getattr(agent, "_current_streamed_assistant_text", "") or ""
+                ).strip()
+                if _partial:
+                    messages.append({"role": "assistant", "content": _partial})
+                    final_response = _partial
+                else:
+                    final_response = f"{INTERRUPT_WAITING_FOR_MODEL_PREFIX}{api_elapsed:.1f}s elapsed)."
+                agent._persist_session(messages, conversation_history)
                 break
 
             except Exception as api_error:
@@ -2210,6 +2266,15 @@ def run_conversation(
                     # "unknown variant `image_url`, expected `text`".
                     "unknown variant `image_url`, expected `text`",
                     "unknown variant image_url, expected text",
+                    # OpenRouter routes a request to upstream endpoints and,
+                    # when none of the candidate endpoints for the model accept
+                    # image input, returns HTTP 404 "No endpoints found that
+                    # support image input". Without this phrase the agent never
+                    # strips the images, the retry loop re-sends the same
+                    # rejected request until exhaustion, and the gateway leaves
+                    # every subsequent message queued behind the stuck turn —
+                    # the P1 in issue #21160. The 404 passes the 4xx gate below.
+                    "no endpoints found that support image input",
                 )
                 _err_lower = _err_body.lower()
                 _looks_like_image_rejection = any(
@@ -2781,10 +2846,9 @@ def run_conversation(
                             approx_tokens=approx_tokens,
                             task_id=effective_task_id,
                         )
-                        # Compression created a new session — clear history
-                        # so _flush_messages_to_session_db writes compressed
-                        # messages to the new session, not skipping them.
-                        conversation_history = None
+                        conversation_history = conversation_history_after_compression(
+                            agent, messages
+                        )
                         if len(messages) < original_len or old_ctx > _reduced_ctx:
                             agent._buffer_status(
                                 f"🗜️ Context reduced to {_reduced_ctx:,} tokens "
@@ -2796,15 +2860,25 @@ def run_conversation(
                     # Fall through to normal error handling if compression
                     # is exhausted or didn't help.
 
-                # Eager fallback for rate-limit errors (429 or quota exhaustion).
-                # When a fallback model is configured, switch immediately instead
-                # of burning through retries with exponential backoff -- the
-                # primary provider won't recover within the retry window.
+                # Eager fallback for rate-limit errors (429 or quota exhaustion)
+                # and transport errors (connection failure / timeout / provider
+                # overloaded).  Rate limits and billing: switch immediately —
+                # the primary provider won't recover within the retry window.
+                # Transport errors: allow 1 retry first (transient hiccups
+                # recover), then fall back if the provider is truly unreachable.
                 is_rate_limited = classified.reason in {
                     FailoverReason.rate_limit,
                     FailoverReason.billing,
                 }
-                if is_rate_limited and agent._fallback_index < len(agent._fallback_chain):
+                _is_transport_failure = classified.reason in {
+                    FailoverReason.timeout,
+                    FailoverReason.overloaded,
+                }
+                _should_fallback = (
+                    is_rate_limited
+                    or (_is_transport_failure and retry_count >= 2)
+                )
+                if _should_fallback and agent._fallback_index < len(agent._fallback_chain):
                     # Don't eagerly fallback if credential pool rotation may
                     # still recover.  See _pool_may_recover_from_rate_limit
                     # for the single-credential-pool and CloudCode-quota
@@ -2819,6 +2893,10 @@ def run_conversation(
                             agent._buffer_status(
                                 "⚠️ Billing or credits exhausted — switching to fallback provider..."
                             )
+                        elif _is_transport_failure:
+                            agent._buffer_status(
+                                "⚠️ Provider unreachable — switching to fallback provider..."
+                            )
                         else:
                             agent._buffer_status("⚠️ Rate limited — switching to fallback provider...")
                         if agent._try_activate_fallback(reason=classified.reason):
@@ -2993,10 +3071,9 @@ def run_conversation(
                         messages, system_message, approx_tokens=approx_tokens,
                         task_id=effective_task_id,
                     )
-                    # Compression created a new session — clear history
-                    # so _flush_messages_to_session_db writes compressed
-                    # messages to the new session, not skipping them.
-                    conversation_history = None
+                    conversation_history = conversation_history_after_compression(
+                        agent, messages
+                    )
 
                     # Re-estimate tokens after compression.  Same-message-count
                     # compression (tool-result pruning, in-place summarization)
@@ -3160,10 +3237,9 @@ def run_conversation(
                         messages, system_message, approx_tokens=approx_tokens,
                         task_id=effective_task_id,
                     )
-                    # Compression created a new session — clear history
-                    # so _flush_messages_to_session_db writes compressed
-                    # messages to the new session, not skipping them.
-                    conversation_history = None
+                    conversation_history = conversation_history_after_compression(
+                        agent, messages
+                    )
 
                     # Re-estimate tokens after compression.  Same-message-count
                     # compression (tool-result pruning, in-place summarization)
@@ -3490,6 +3566,65 @@ def run_conversation(
                             force=True,
                         )
 
+                    # Detect thinking-timeout pattern: a known reasoning model
+                    # hit a transport-layer error before the first content
+                    # token arrived.  Distinct from _is_stream_drop above
+                    # (which fires for large file-write stream drops) and
+                    # from any classifier reason that's not a transport
+                    # timeout.  Reuses the reasoning-model allowlist from
+                    # agent/reasoning_timeouts.py (Fixes #52217) so the
+                    # trigger is consistent with what the per-model
+                    # stale-timeout floor covers.  After the classifier
+                    # override at agent/error_classifier.py:720-738 (this
+                    # PR), transport disconnects on reasoning models route
+                    # to FailoverReason.timeout rather than
+                    # context_overflow, so this branch actually fires.
+                    # Detection and message text live in
+                    # agent.thinking_timeout_guidance so they're
+                    # unit-testable without driving the full retry loop.
+                    # (Part 2 of Fixes #52310.)
+                    from agent.thinking_timeout_guidance import (
+                        is_thinking_timeout,
+                    )
+                    _is_thinking_timeout = is_thinking_timeout(
+                        classified,
+                        _model,
+                        error_msg,
+                    )
+                    if _is_thinking_timeout:
+                        agent._vprint(
+                            f"{agent.log_prefix}   💡 The model's thinking "
+                            f"phase exceeded the upstream proxy's idle "
+                            f"timeout before the first content token "
+                            f"arrived. This is a known issue with "
+                            f"reasoning models behind cloud gateways "
+                            f"(NVIDIA NIM, OpenAI, Anthropic, DeepSeek).",
+                            force=True,
+                        )
+                        agent._vprint(
+                            f"{agent.log_prefix}      Workarounds in priority order:",
+                            force=True,
+                        )
+                        agent._vprint(
+                            f"{agent.log_prefix}      1. Set "
+                            f"`providers.{_provider}.models.{_model}.stale_timeout_seconds: 900` "
+                            f"in `~/.hermes/config.yaml` to extend the per-call "
+                            f"timeout. (Hermes's built-in floor is 600s for "
+                            f"known reasoning models — if you still see this "
+                            f"after raising, the upstream cap is even shorter.)",
+                            force=True,
+                        )
+                        agent._vprint(
+                            f"{agent.log_prefix}      2. Lower `reasoning_budget` or set "
+                            f"`reasoning_effort: medium` on this model if the provider supports it.",
+                            force=True,
+                        )
+                        agent._vprint(
+                            f"{agent.log_prefix}      3. Use a smaller / faster reasoning "
+                            f"model if the task doesn't require deep thinking.",
+                            force=True,
+                        )
+
                     logger.error(
                         "%sAPI call failed after %s retries. %s | provider=%s model=%s msgs=%s tokens=~%s",
                         agent.log_prefix, max_retries, _final_summary,
@@ -3506,7 +3641,22 @@ def run_conversation(
                             _final_response += f"\n\n{_billing_guidance}"
                     else:
                         _final_response = f"API call failed after {max_retries} retries: {_final_summary}"
-                    if _is_stream_drop:
+                    if _is_thinking_timeout:
+                        # Thinking-timeout guidance overrides the generic
+                        # stream-drop guidance — the latter is wrong for
+                        # this case (it suggests splitting large file
+                        # writes, which isn't what happened).  See the
+                        # reasoning-model override at
+                        # agent/error_classifier.py:720-738 and the
+                        # detection block above for context.
+                        from agent.thinking_timeout_guidance import (
+                            build_thinking_timeout_guidance,
+                        )
+                        _final_response += build_thinking_timeout_guidance(
+                            provider=_provider,
+                            model=_model,
+                        )
+                    elif _is_stream_drop:
                         _final_response += (
                             "\n\nThe provider's stream connection keeps "
                             "dropping — this often happens when generating "
@@ -3538,7 +3688,12 @@ def run_conversation(
                         _ra_raw = _resp_headers.get("retry-after") or _resp_headers.get("Retry-After")
                         if _ra_raw:
                             try:
-                                _retry_after = min(float(_ra_raw), 120)  # Cap at 2 minutes
+                                # Cap at 10 minutes. Anthropic Tier 1 input-token
+                                # buckets reset in ~171s, so a 120s cap caused us to
+                                # retry before the actual reset window and re-trip the
+                                # limit. 600s covers all realistic provider reset
+                                # windows while still rejecting pathological values. (#26293)
+                                _retry_after = min(float(_ra_raw), 600)
                             except (TypeError, ValueError):
                                 pass
                 wait_time = _retry_after if _retry_after else jittered_backoff(retry_count, base_delay=2.0, max_delay=60.0)
@@ -4193,10 +4348,9 @@ def run_conversation(
                         approx_tokens=agent.context_compressor.last_prompt_tokens,
                         task_id=effective_task_id,
                     )
-                    # Compression created a new session — clear history so
-                    # _flush_messages_to_session_db writes compressed messages
-                    # to the new session (see preflight compression comment).
-                    conversation_history = None
+                    conversation_history = conversation_history_after_compression(
+                        agent, messages
+                    )
                 
                 # Save session log incrementally (so progress is visible even if interrupted)
                 agent._session_messages = messages
@@ -4571,7 +4725,11 @@ def run_conversation(
                         "_verification_stop_synthetic": True,
                     })
                     agent._session_messages = messages
-                    agent._emit_status("↻ Verification required before finishing")
+                    # Run the verification-stop loop silently — the nudge is an
+                    # internal turn that should not add noise to the user's
+                    # terminal. Keep a debug breadcrumb in agent.log for tracing.
+                    logger.debug("verification stop-loop nudge issued (attempt %d)",
+                                 agent._verification_stop_nudges)
                     continue
 
                 messages.append(final_msg)

agent/credential_pool.py

@@ -11,6 +11,7 @@ import uuid
 import re
 from dataclasses import dataclass, fields, replace
 from datetime import datetime, timezone
+from pathlib import Path
 from typing import Any, Dict, List, Optional, Set, Tuple
 
 from hermes_constants import OPENROUTER_BASE_URL
@@ -447,6 +448,63 @@ def get_pool_strategy(provider: str) -> str:
 DEFAULT_MAX_CONCURRENT_PER_CREDENTIAL = 1
 
 
+def _write_through_provider_state_to_global_root(
+    provider_id: str, state: Dict[str, Any]
+) -> None:
+    """Persist a rotated OAuth ``state`` into the global-root auth.json.
+
+    Best-effort write-through for the multi-profile rotation hazard
+    (#48415 / #43589): nous, openai-codex, and xai-oauth rotate the
+    refresh_token on refresh, so when a profile pool refresh rotates a grant
+    it resolved from the root fallback, the rotated chain must land back in
+    root. Otherwise root keeps a now-revoked refresh token and every other
+    profile reading the stale root grant dies with ``refresh_token_reused`` /
+    ``invalid_grant`` once its access token expires.
+
+    Only updates ``providers.<provider_id>`` in the root store; never touches
+    the profile store (the caller already saved that). Swallows all errors — a
+    failed write-through degrades to the pre-existing behavior (root stale), it
+    must never break the profile's own successful save. Mirrors
+    ``hermes_cli.auth._write_through_xai_oauth_to_global_root`` (which covers
+    the non-pool xAI refresh path) for the credential-pool refresh path.
+    """
+    try:
+        global_path = auth_mod._global_auth_file_path()
+    except Exception:
+        return
+    if global_path is None:
+        # Classic mode (profile == root); the profile save already hit root.
+        return
+    # Seat belt: under pytest, refuse to write the real user's
+    # ~/.hermes/auth.json even when HERMES_HOME points at a profile path
+    # (mirrors the read-side guard in _load_global_auth_store). Uses the
+    # unmodified HOME env, not Path.home() which fixtures may monkeypatch.
+    if os.environ.get("PYTEST_CURRENT_TEST"):
+        real_home_env = os.environ.get("HOME", "")
+        if real_home_env:
+            real_root = Path(real_home_env) / ".hermes" / "auth.json"
+            try:
+                if global_path.resolve(strict=False) == real_root.resolve(strict=False):
+                    return
+            except Exception:
+                return
+    try:
+        if global_path.exists():
+            global_store = _load_auth_store(global_path)
+        else:
+            global_store = {}
+        if not isinstance(global_store, dict):
+            return
+        _store_provider_state(global_store, provider_id, dict(state), set_active=False)
+        auth_mod._save_auth_store(global_store, global_path)
+    except Exception as exc:  # pragma: no cover - best effort
+        logger.debug(
+            "%s pool refresh: write-through to global root failed: %s",
+            provider_id,
+            exc,
+        )
+
+
 class CredentialPool:
     def __init__(self, provider: str, entries: List[PooledCredential]):
         self.provider = provider
@@ -479,10 +537,11 @@ class CredentialPool:
                 self._entries[idx] = new
                 return
 
-    def _persist(self) -> None:
+    def _persist(self, *, removed_ids: Optional[List[str]] = None) -> None:
         write_credential_pool(
             self.provider,
             [entry.to_dict() for entry in self._entries],
+            removed_ids=removed_ids,
         )
 
     def _is_terminal_auth_failure(
@@ -800,6 +859,28 @@ class CredentialPool:
         try:
             with _auth_store_lock():
                 auth_store = _load_auth_store()
+                # Decide BEFORE writing whether this profile is reading the
+                # grant from the global root (no own providers.<id> block) vs.
+                # genuinely shadowing it. A pool refresh rotates single-use
+                # OAuth refresh tokens, so a profile that resolved the grant
+                # from root MUST write the rotated chain back to root too —
+                # otherwise root keeps a revoked refresh token and every other
+                # profile reading the stale root grant dies with
+                # refresh_token_reused / invalid_grant once its access token
+                # expires. This mirrors the xAI write-through in
+                # hermes_cli.auth._save_xai_oauth_tokens (#43589); the pool
+                # refresh path is the Codex/xAI analog reported in #48415.
+                _wt_provider_id = {
+                    "nous": "nous",
+                    "openai-codex": "openai-codex",
+                    "xai-oauth": "xai-oauth",
+                }.get(self.provider)
+                write_through_to_root = bool(_wt_provider_id) and not (
+                    isinstance(auth_store.get("providers"), dict)
+                    and isinstance(
+                        auth_store["providers"].get(_wt_provider_id), dict
+                    )
+                )
                 if self.provider == "nous":
                     state = _load_provider_state(auth_store, "nous")
                     if state is None:
@@ -855,6 +936,10 @@ class CredentialPool:
                     return
 
                 _save_auth_store(auth_store)
+                if write_through_to_root and _wt_provider_id:
+                    _write_through_provider_state_to_global_root(
+                        _wt_provider_id, state
+                    )
         except Exception as exc:
             logger.debug("Failed to sync %s pool entry back to auth store: %s", self.provider, exc)
 
@@ -1040,13 +1125,17 @@ class CredentialPool:
                         logger.debug(
                             "Failed to clear terminal xAI OAuth state: %s", clear_exc
                         )
+                    removed_ids = [
+                        item.id for item in self._entries
+                        if item.source == "loopback_pkce"
+                    ]
                     self._entries = [
                         item for item in self._entries
                         if item.source != "loopback_pkce"
                     ]
                     if self._current_id == entry.id:
                         self._current_id = None
-                    self._persist()
+                    self._persist(removed_ids=removed_ids)
                     return None
             # For openai-codex: same race as xAI/nous — another Hermes process
             # may have consumed the refresh token between our proactive sync
@@ -1106,13 +1195,17 @@ class CredentialPool:
                         logger.debug(
                             "Failed to clear terminal Codex OAuth state: %s", clear_exc
                         )
+                    removed_ids = [
+                        item.id for item in self._entries
+                        if item.source == "device_code"
+                    ]
                     self._entries = [
                         item for item in self._entries
                         if item.source != "device_code"
                     ]
                     if self._current_id == entry.id:
                         self._current_id = None
-                    self._persist()
+                    self._persist(removed_ids=removed_ids)
                     return None
             # For nous: another process may have consumed the refresh token
             # between our proactive sync and the HTTP call.  Re-sync from
@@ -1169,13 +1262,17 @@ class CredentialPool:
                         auth_mod.NOUS_DEVICE_CODE_SOURCE,
                         f"manual:{auth_mod.NOUS_DEVICE_CODE_SOURCE}",
                     }
+                    removed_ids = [
+                        item.id for item in self._entries
+                        if item.source in singleton_sources
+                    ]
                     self._entries = [
                         item for item in self._entries
                         if item.source not in singleton_sources
                     ]
                     if self._current_id == entry.id:
                         self._current_id = None
-                    self._persist()
+                    self._persist(removed_ids=removed_ids)
                     return None
             self._mark_exhausted(entry, None)
             return None
@@ -1337,7 +1434,7 @@ class CredentialPool:
             pruned_ids = set(entries_to_prune)
             self._entries = [e for e in self._entries if e.id not in pruned_ids]
         if cleared_any:
-            self._persist()
+            self._persist(removed_ids=entries_to_prune)
         return available
 
     def _select_unlocked(self) -> Optional[PooledCredential]:
@@ -1511,7 +1608,11 @@ class CredentialPool:
             replace(entry, priority=new_priority)
             for new_priority, entry in enumerate(self._entries)
         ]
-        self._persist()
+        write_credential_pool(
+            self.provider,
+            [entry.to_dict() for entry in self._entries],
+            removed_ids=[removed.id],
+        )
         if self._current_id == removed.id:
             self._current_id = None
         return removed
@@ -2173,6 +2274,11 @@ def _seed_custom_pool(pool_key: str, entries: List[PooledCredential]) -> Tuple[b
 def load_pool(provider: str) -> CredentialPool:
     provider = (provider or "").strip().lower()
     raw_entries = read_credential_pool(provider)
+    disk_ids = {
+        entry.get("id")
+        for entry in raw_entries
+        if isinstance(entry, dict) and entry.get("id")
+    }
     raw_needs_sanitization = any(
         isinstance(payload, dict)
         and sanitize_borrowed_credential_payload(payload, provider) != payload
@@ -2201,8 +2307,10 @@ def load_pool(provider: str) -> CredentialPool:
         changed |= _normalize_pool_priorities(provider, entries)
 
     if changed:
+        new_ids = {entry.id for entry in entries}
         write_credential_pool(
             provider,
             [entry.to_dict() for entry in sorted(entries, key=lambda item: item.priority)],
+            removed_ids=disk_ids - new_ids,
         )
     return CredentialPool(provider, entries)

agent/curator.py

@@ -377,8 +377,10 @@ CURATOR_REVIEW_PROMPT = (
     "bodies + `references/`, `templates/`, and `scripts/` subfiles for "
     "session-specific detail — not one-session-one-skill micro-entries.\n\n"
     "Hard rules — do not violate:\n"
-    "1. DO NOT touch bundled or hub-installed skills. The candidate list "
-    "below is already filtered to agent-created skills only.\n"
+    "1. DO NOT touch bundled, hub-installed, or external-dir skills "
+    "(`skills.external_dirs`). The candidate list below is already filtered "
+    "to local curator-managed skills only; external skills are externally "
+    "owned and read-only to this background curator.\n"
     "2. DO NOT delete any skill. Archiving (moving the skill's directory "
     "into ~/.hermes/skills/.archive/) is the maximum destructive action. "
     "Archives are recoverable; deletion is not.\n"
@@ -469,8 +471,9 @@ CURATOR_REVIEW_PROMPT = (
     "skill, or `absorbed_into=\"\"` when you're truly pruning with no "
     "forwarding target. This drives cron-job skill-reference migration — "
     "guessing from your YAML summary after the fact is fragile.\n"
-    "  - terminal                       — mv a sibling into the archive "
-    "OR move its content into a support subfile\n\n"
+    "  - terminal                       — move LOCAL candidate content into "
+    "a support subfile when package integrity requires it; never mv, cp, rm, "
+    "patch, or rewrite bundled, hub-installed, or external-dir skills\n\n"
     "'keep' is a legitimate decision ONLY when the skill is already a "
     "class-level umbrella and none of the proposed merges would improve "
     "discoverability. 'This is narrow but distinct from its siblings' "
@@ -1843,6 +1846,14 @@ def _run_llm_review(prompt: str) -> Dict[str, Any]:
         # Disable recursive nudges — the curator must never spawn its own review.
         review_agent._memory_nudge_interval = 0
         review_agent._skill_nudge_interval = 0
+        # Tag this fork as autonomous background curation so skill_manage's
+        # background-review write guard fires. Without this the fork inherits
+        # the default "assistant_tool" origin, is_background_review() is False,
+        # and the external/bundled/hub-installed skill_manage guards never
+        # trigger during the curation pass they exist to protect against.
+        # turn_context.py binds this onto the write-origin ContextVar at turn
+        # start (see agent/turn_context.py).
+        review_agent._memory_write_origin = "background_review"
 
         # Redirect the forked agent's stdout/stderr to /dev/null while it
         # runs so its tool-call chatter doesn't pollute the foreground

agent/display.py

@@ -16,6 +16,7 @@ from pathlib import Path
 from typing import Any
 
 from utils import safe_json_loads
+from agent.redact import redact_sensitive_text
 from agent.tool_result_classification import file_mutation_result_landed
 
 # ANSI escape codes for coloring tool failure indicators
@@ -339,6 +340,62 @@ def _read_file_line_label(args: dict) -> str:
     return f"L{offset}-{offset + limit - 1}"
 
 
+def redact_browser_typed_text_for_display(value: Any, typed_text: Any) -> Any:
+    """Apply secret redaction to browser_type text in display-facing payloads.
+
+    Backends sometimes echo the attempted input in error strings or fallback
+    metadata.  When the raw typed value contains a recognizable secret (API
+    key, token, JWT, etc.) the redacted form differs from the raw value, so we
+    replace every occurrence of the raw value with its redacted form before a
+    browser_type result reaches logs, callbacks, the model, or chat history.
+
+    Normal typed text (search queries, addresses, form fields) matches no
+    secret pattern, so it passes through unchanged and stays readable.
+
+    Redaction is forced here regardless of the global ``security.redact_secrets``
+    preference: a typed credential leaking into chat history is a security
+    boundary, not mere log hygiene.
+    """
+    if typed_text is None:
+        return value
+    needle = str(typed_text)
+    if needle == "":
+        return value
+    redacted = redact_sensitive_text(needle, force=True)
+    if redacted == needle:
+        # Nothing secret-looking in the typed text; leave payload untouched.
+        return value
+    if isinstance(value, str):
+        return value.replace(needle, redacted)
+    if isinstance(value, dict):
+        return {
+            key: redact_browser_typed_text_for_display(item, typed_text)
+            for key, item in value.items()
+        }
+    if isinstance(value, list):
+        return [redact_browser_typed_text_for_display(item, typed_text) for item in value]
+    if isinstance(value, tuple):
+        return tuple(redact_browser_typed_text_for_display(item, typed_text) for item in value)
+    return value
+
+
+def redact_tool_args_for_display(tool_name: str, args: dict | None) -> dict | None:
+    """Return a copy of tool args safe for logs/progress UI.
+
+    For ``browser_type`` the ``text`` argument is run through the same
+    secret-pattern redactor used for logs.  Recognizable credentials (API
+    keys, tokens) are masked before the value reaches tool progress
+    notifications; normal typed text is left intact for debuggability.
+    """
+    if not isinstance(args, dict):
+        return args
+    if tool_name == "browser_type" and isinstance(args.get("text"), str):
+        safe_args = dict(args)
+        safe_args["text"] = redact_sensitive_text(args["text"], force=True)
+        return safe_args
+    return args
+
+
 def _delegate_task_goal_parts(tasks: Any, *, per_goal_len: int) -> tuple[int, list[str]]:
     if not isinstance(tasks, list):
         return 0, []
@@ -362,13 +419,14 @@ def build_tool_preview(tool_name: str, args: dict, max_len: int | None = None) -
         max_len = _tool_preview_max_len
     if not args:
         return None
+    args = redact_tool_args_for_display(tool_name, args) or args
     primary_args = {
         "terminal": "command", "web_search": "query", "web_extract": "urls",
         "read_file": "path", "write_file": "path", "patch": "path",
         "search_files": "pattern", "browser_navigate": "url",
         "browser_click": "ref", "browser_type": "text",
         "image_generate": "prompt", "text_to_speech": "text",
-        "vision_analyze": "question", "mixture_of_agents": "user_prompt",
+        "vision_analyze": "question",
         "skill_view": "name", "skills_list": "category",
         "cronjob": "action",
         "execute_code": "code", "delegate_task": "goal",
@@ -1085,6 +1143,7 @@ def get_cute_tool_message(
     When *result* is provided the line is checked for failure indicators.
     Failed tool calls get a red prefix and an informational suffix.
     """
+    args = redact_tool_args_for_display(tool_name, args) or args
     dur = f"{duration:.1f}s"
     is_failure, failure_suffix = _detect_tool_failure(tool_name, result)
     skin_prefix = get_skin_tool_prefix()
@@ -1216,8 +1275,6 @@ def get_cute_tool_message(
         return _wrap(f"┊ 🔊 speak     {_trunc(args.get('text', ''), 30)}  {dur}")
     if tool_name == "vision_analyze":
         return _wrap(f"┊ 👁️  vision    {_trunc(args.get('question', ''), 30)}  {dur}")
-    if tool_name == "mixture_of_agents":
-        return _wrap(f"┊ 🧠 reason    {_trunc(args.get('user_prompt', ''), 30)}  {dur}")
     if tool_name == "send_message":
         return _wrap(f"┊ 📨 send      {args.get('target', '?')}: \"{_trunc(args.get('message', ''), 25)}\"  {dur}")
     if tool_name == "cronjob":

agent/error_classifier.py

@@ -133,6 +133,31 @@ _RATE_LIMIT_PATTERNS = [
     "servicequotaexceededexception",
 ]
 
+# Patterns that indicate provider-side overload, NOT a per-credential rate
+# limit or billing problem.  The credential is valid — the server is just
+# busy — so the correct recovery is "back off and retry the same key", never
+# "rotate the credential" (rotating exhausts the pool while the endpoint is
+# still busy; a single-key user has nothing to rotate to).  Some providers
+# (notably Z.AI / Zhipu) reuse HTTP 429 for server-wide overload, so the 429
+# status path matches the body against this list before falling through to
+# the rate_limit default.  Phrases are kept narrow and overload-flavoured so a
+# normal rate-limit message ("you have been rate-limited") doesn't hit this
+# bucket. (#14038, #15297)
+_OVERLOADED_PATTERNS = [
+    "overloaded",
+    "temporarily overloaded",
+    "service is temporarily overloaded",
+    "service may be temporarily overloaded",
+    "server is overloaded",
+    "server overloaded",
+    "service overloaded",
+    "service is overloaded",
+    "upstream overloaded",
+    "currently overloaded",
+    "at capacity",
+    "over capacity",
+]
+
 # Usage-limit patterns that need disambiguation (could be billing OR rate_limit)
 _USAGE_LIMIT_PATTERNS = [
     "usage limit",
@@ -717,6 +742,26 @@ def classify_api_error(
 
     is_disconnect = any(p in error_msg for p in _SERVER_DISCONNECT_PATTERNS)
     if is_disconnect and not status_code:
+        # Reasoning-model override: a transport disconnect on a reasoning
+        # model is much more likely the upstream proxy idle-killing a
+        # long thinking stream than a true context overflow — even on
+        # large sessions.  The default disconnect+large-session routing
+        # below would otherwise send the user into the compression
+        # branch (should_compress=True) and silently delete
+        # conversation history on a phantom context-length error.
+        # Reasoning models have multi-minute thinking phases that
+        # routinely exceed the cloud gateway's idle window (NVIDIA
+        # NIM ~120s — first-party repro at NVIDIA/NemoClaw#4846;
+        # OpenAI worker / Anthropic stream-idle similar).  The
+        # per-reasoning-model stale-timeout floor in
+        # agent/reasoning_timeouts.py raises the stale-detector
+        # threshold to tolerate long thinking, so a true
+        # transport-layer failure here is recoverable via the retry
+        # path — not via context compression.  Reclassify as timeout.
+        # (Part 1 of Fixes #52310.)
+        from agent.reasoning_timeouts import get_reasoning_stale_timeout_floor
+        if get_reasoning_stale_timeout_floor(model) is not None:
+            return _result(FailoverReason.timeout, retryable=True)
         # Absolute token/message-count thresholds are only a proxy for smaller
         # context windows.  Large-context sessions can have hundreds of
         # messages while still being far below their actual token budget.
@@ -843,7 +888,19 @@ def _classify_by_status(
         )
 
     if status_code == 429:
-        # Already checked long_context_tier above; this is a normal rate limit
+        # Already checked long_context_tier above. Some providers (notably
+        # Z.AI / Zhipu) reuse HTTP 429 for server-wide overload — same status
+        # code as a true per-credential rate limit, but the credential is
+        # valid and the correct recovery is "back off and retry the same key",
+        # NOT "rotate the credential" (which exhausts the pool while the
+        # endpoint is still busy, and does nothing for a single-key user).
+        # Disambiguate on the error body so an overload 429 takes the
+        # transient-overload path instead of burning the pool. (#14038)
+        if any(p in error_msg for p in _OVERLOADED_PATTERNS):
+            return result_fn(
+                FailoverReason.overloaded,
+                retryable=True,
+            )
         return result_fn(
             FailoverReason.rate_limit,
             retryable=True,
@@ -1194,6 +1251,17 @@ def _classify_by_message(
             should_fallback=True,
         )
 
+    # Overloaded / server-busy patterns — must come BEFORE the rate_limit and
+    # billing checks so that a message-only "overloaded" (no 503/529 status,
+    # e.g. some Anthropic-compatible proxies) classifies as a transient
+    # overload (backoff + retry) instead of falling through to `unknown` or
+    # incorrectly triggering credential rotation.
+    if any(p in error_msg for p in _OVERLOADED_PATTERNS):
+        return result_fn(
+            FailoverReason.overloaded,
+            retryable=True,
+        )
+
     # Billing patterns
     if any(p in error_msg for p in _BILLING_PATTERNS):
         return result_fn(
@@ -1283,19 +1351,25 @@ def _extract_status_code(error: Exception) -> Optional[int]:
 
 
 def _extract_error_body(error: Exception) -> dict:
-    """Extract the structured error body from an SDK exception."""
-    body = getattr(error, "body", None)
-    if isinstance(body, dict):
-        return body
-    # Some errors have .response.json()
-    response = getattr(error, "response", None)
-    if response is not None:
-        try:
-            json_body = response.json()
-            if isinstance(json_body, dict):
-                return json_body
-        except Exception:
-            pass
+    """Extract the structured error body from an SDK exception or its cause chain."""
+    current = error
+    for _ in range(5):  # Match _extract_status_code() traversal depth.
+        body = getattr(current, "body", None)
+        if isinstance(body, dict):
+            return body
+        # Some errors have .response.json()
+        response = getattr(current, "response", None)
+        if response is not None:
+            try:
+                json_body = response.json()
+                if isinstance(json_body, dict):
+                    return json_body
+            except Exception:
+                pass
+        cause = getattr(current, "__cause__", None) or getattr(current, "__context__", None)
+        if cause is None or cause is current:
+            break
+        current = cause
     return {}
 
 

agent/file_safety.py

@@ -77,15 +77,22 @@ def build_write_denied_prefixes(home: str) -> list[str]:
     ]
 
 
-def get_safe_write_root() -> Optional[str]:
-    """Return the resolved HERMES_WRITE_SAFE_ROOT path, or None if unset."""
-    root = os.getenv("HERMES_WRITE_SAFE_ROOT", "")
-    if not root:
-        return None
-    try:
-        return os.path.realpath(os.path.expanduser(root))
-    except Exception:
-        return None
+def get_safe_write_roots() -> set[str]:
+    """Return resolved HERMES_WRITE_SAFE_ROOT paths. Supports multiple directories
+    separated by ``os.pathsep`` (``:`` on Unix, ``;`` on Windows).
+    E.g., ``/opt/data:/var/www/html`` on Unix, ``C:\\data;D:\\www`` on Windows."""
+    env = os.getenv("HERMES_WRITE_SAFE_ROOT", "")
+    if not env:
+        return set()
+    roots: set[str] = set()
+    for path in env.split(os.pathsep):
+        if path:
+            try:
+                resolved = os.path.realpath(os.path.expanduser(path))
+                roots.add(resolved)
+            except (OSError, ValueError):
+                continue
+    return roots
 
 
 def is_write_denied(path: str) -> bool:
@@ -124,9 +131,15 @@ def is_write_denied(path: str) -> bool:
         except Exception:
             pass
 
-    safe_root = get_safe_write_root()
-    if safe_root and not (resolved == safe_root or resolved.startswith(safe_root + os.sep)):
-        return True
+    safe_roots = get_safe_write_roots()
+    if safe_roots:
+        allowed = False
+        for safe_root in safe_roots:
+            if resolved == safe_root or resolved.startswith(safe_root + os.sep):
+                allowed = True
+                break
+        if not allowed:
+            return True
 
     return False
 

agent/image_routing.py

@@ -388,14 +388,98 @@ def _sniff_mime_from_bytes(raw: bytes) -> Optional[str]:
     # BMP: "BM"
     if raw.startswith(b"BM"):
         return "image/bmp"
-    # HEIC/HEIF: ftypheic / ftypheix / ftypmif1 / ftypmsf1 etc.
-    if len(raw) >= 12 and raw[4:8] == b"ftyp" and raw[8:12] in {
-        b"heic", b"heix", b"hevc", b"hevx", b"mif1", b"msf1", b"heim", b"heis",
-    }:
-        return "image/heic"
+    # ISO-BMFF family (HEIC/HEIF/AVIF): bytes 4..8 == 'ftyp', major brand at 8..12
+    if len(raw) >= 12 and raw[4:8] == b"ftyp":
+        brand = raw[8:12]
+        if brand in {b"avif", b"avis"}:
+            return "image/avif"
+        if brand in {
+            b"heic", b"heix", b"hevc", b"hevx",
+            b"mif1", b"msf1", b"heim", b"heis",
+        }:
+            return "image/heic"
+    # TIFF: II*\0 (little-endian) or MM\0* (big-endian)
+    if raw[:4] in {b"II*\x00", b"MM\x00*"}:
+        return "image/tiff"
+    # ICO: 00 00 01 00 (reserved=0, type=1=icon)
+    if raw[:4] == b"\x00\x00\x01\x00":
+        return "image/x-icon"
+    # SVG: text-based, look for an <svg tag near the start (skip BOM/whitespace)
+    head = raw[:512].lstrip().lower()
+    if head.startswith(b"<?xml") or head.startswith(b"<svg"):
+        if b"<svg" in head:
+            return "image/svg+xml"
     return None
 
 
+# Formats every major vision provider (Anthropic, OpenAI, Gemini, Bedrock)
+# accepts natively. Anything outside this set has to be transcoded to PNG
+# before we declare media_type, otherwise the provider returns HTTP 400
+# ("Could not process image" / "Unsupported image media type") and the
+# whole turn fails with no salvage path.
+#
+# Discord (and a few other chat platforms) freely accept attachments in
+# formats outside this set -- AVIF screenshots from Chromium, HEIC from
+# iPhones, TIFF from scanners, BMP from old Windows tools, ICO -- so users
+# do hit this in practice. SVG is vector and Pillow cannot rasterize it;
+# it is skipped (logged) rather than transcoded.
+_UNIVERSALLY_SUPPORTED_MIMES = frozenset({
+    "image/png", "image/jpeg", "image/gif", "image/webp",
+})
+
+
+def _transcode_to_png(raw: bytes) -> Optional[bytes]:
+    """Decode arbitrary image bytes with Pillow and re-encode as PNG.
+
+    Returns None if Pillow isn't installed or can't decode the input
+    (rare formats, corrupted bytes, missing optional decoder plugin for
+    HEIC/AVIF, or vector formats like SVG). Caller falls back to skipping
+    the image so the rest of the turn still works.
+
+    HEIC/HEIF and AVIF need optional Pillow plugins; we try to register
+    them on demand and swallow ImportError so a missing plugin just
+    looks like 'Pillow can't decode this' rather than crashing.
+    """
+    try:
+        from PIL import Image
+    except ImportError:
+        logger.info(
+            "image_routing: Pillow not installed; cannot transcode "
+            "non-standard image format to PNG. Install with `pip install Pillow` "
+            "(and `pillow-heif` / `pillow-avif-plugin` for those formats)."
+        )
+        return None
+    # Optional plugin registration. Silent on failure: an unsupported
+    # format will just fall through to Image.open raising below.
+    try:
+        import pillow_heif  # type: ignore
+
+        pillow_heif.register_heif_opener()
+    except Exception:
+        pass
+    try:
+        import pillow_avif  # type: ignore  # noqa: F401  -- registers AVIF on import
+    except Exception:
+        pass
+    try:
+        from io import BytesIO
+
+        with Image.open(BytesIO(raw)) as im:
+            # Pick an output mode PNG can serialise. Anything other than
+            # the standard set gets normalised to RGBA so transparency is
+            # preserved where the source had it.
+            if im.mode not in {"RGB", "RGBA", "L", "LA", "P"}:
+                im = im.convert("RGBA")
+            buf = BytesIO()
+            im.save(buf, format="PNG", optimize=False)
+            return buf.getvalue()
+    except Exception as exc:
+        logger.info(
+            "image_routing: Pillow could not transcode image to PNG -- %s", exc
+        )
+        return None
+
+
 def _guess_mime(path: Path, raw: Optional[bytes] = None) -> str:
     """Return image MIME type for *path*.
 
@@ -431,8 +515,18 @@ def _file_to_data_url(path: Path) -> Optional[str]:
     accept large images (OpenAI 49 MB+, Gemini 100 MB) don't pay a silent
     quality tax just because one other provider is stricter.
 
-    Returns None only if the file can't be read (missing, permission
-    denied, etc.); the caller reports those paths in ``skipped``.
+    Format compatibility IS handled here: if the sniffed MIME isn't one
+    of ``_UNIVERSALLY_SUPPORTED_MIMES`` (i.e. it's something like AVIF,
+    HEIC, BMP, TIFF, or ICO that some providers reject outright), we
+    transcode to PNG with Pillow before declaring media_type. This fixes
+    the user-visible "Could not process image" HTTP 400 from Anthropic on
+    Discord-attached AVIF/HEIC/BMP files.
+
+    Returns None if the file can't be read OR if the format isn't
+    universally supported AND Pillow can't transcode it (Pillow missing,
+    HEIC/AVIF plugin missing, vector format like SVG, corrupt bytes). The
+    caller reports those paths in ``skipped`` and the rest of the turn
+    proceeds.
     """
     try:
         raw = path.read_bytes()
@@ -440,6 +534,22 @@ def _file_to_data_url(path: Path) -> Optional[str]:
         logger.warning("image_routing: failed to read %s — %s", path, exc)
         return None
     mime = _guess_mime(path, raw=raw)
+    if mime not in _UNIVERSALLY_SUPPORTED_MIMES:
+        transcoded = _transcode_to_png(raw)
+        if transcoded is None:
+            logger.warning(
+                "image_routing: %s is %s which is not accepted by all major "
+                "vision providers and could not be transcoded to PNG; "
+                "skipping this attachment.",
+                path, mime,
+            )
+            return None
+        logger.info(
+            "image_routing: transcoded %s (%s) -> image/png for provider compatibility",
+            path.name, mime,
+        )
+        raw = transcoded
+        mime = "image/png"
     b64 = base64.b64encode(raw).decode("ascii")
     return f"data:{mime};base64,{b64}"
 

agent/moa_loop.py

@@ -0,0 +1,441 @@
+"""Mixture-of-Agents runtime helpers for /moa turns.
+
+The slash command is deliberately not a model tool. It marks one user turn as
+MoA-enabled; the normal Hermes agent loop still owns tool calling and turn
+termination, while this module gathers reference-model context before each model
+iteration.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import logging
+from concurrent.futures import ThreadPoolExecutor
+from typing import Any
+
+from agent.auxiliary_client import call_llm
+from agent.transports import get_transport
+
+logger = logging.getLogger(__name__)
+
+# Upper bound on concurrent reference-model calls. References are independent
+# advisory calls (no tools, no inter-dependence), so we fan them out the same
+# way delegate_task runs a batch: all in flight at once, results collected when
+# every reference finishes. Presets rarely list more than a handful of
+# references; this cap just protects against a pathologically large preset
+# opening dozens of sockets at once.
+_MAX_REFERENCE_WORKERS = 8
+
+
+def _slot_label(slot: dict[str, str]) -> str:
+    return f"{slot.get('provider', '').strip()}:{slot.get('model', '').strip()}"
+
+
+def _slot_runtime(slot: dict[str, str]) -> dict[str, Any]:
+    """Resolve a reference/aggregator slot to real runtime call kwargs.
+
+    A MoA slot is just a model selection — it must be called the same way any
+    model is called elsewhere, not through a bare ``call_llm(provider=...,
+    model=...)`` that leaves base_url/api_key/api_mode unresolved and lets the
+    auxiliary auto-detector guess. We route the slot's provider through
+    ``resolve_runtime_provider`` (the canonical provider→api_mode/base_url/
+    api_key resolver the CLI, gateway, and delegate_task all use), so the slot
+    gets its provider's real API surface — e.g. MiniMax → anthropic_messages,
+    GPT-5/o-series → max_completion_tokens, custom endpoints → their base_url.
+
+    Returns the kwargs to pass through to ``call_llm`` (provider/model plus the
+    resolved base_url/api_key when available). Falls back to the bare
+    provider/model on any resolution error so a misconfigured slot still
+    attempts the call rather than aborting the whole MoA turn.
+    """
+    provider = str(slot.get("provider") or "").strip()
+    model = str(slot.get("model") or "").strip()
+    out: dict[str, Any] = {"provider": provider, "model": model}
+    try:
+        from hermes_cli.runtime_provider import resolve_runtime_provider
+
+        rt = resolve_runtime_provider(requested=provider, target_model=model)
+        resolved_provider = str(rt.get("provider") or provider).strip().lower()
+        # call_llm treats an explicit base_url as a custom endpoint. That is
+        # correct for ordinary OpenAI-compatible targets, but wrong for OAuth /
+        # adapter-backed providers whose provider branch adds auth headers and
+        # request-shape adapters. Keep those providers identified by name.
+        if resolved_provider in {"openai-codex", "xai-oauth"}:
+            return out
+        # Pass the resolved endpoint through so call_llm builds the request for
+        # the provider's actual API surface instead of auto-detecting. base_url
+        # routes call_llm to the right adapter (incl. anthropic_messages mode);
+        # api_key is the resolved credential for that provider.
+        if rt.get("base_url"):
+            out["base_url"] = rt["base_url"]
+        if rt.get("api_key"):
+            out["api_key"] = rt["api_key"]
+    except Exception as exc:  # pragma: no cover - defensive
+        logger.debug("MoA slot runtime resolution failed for %s: %s", _slot_label(slot), exc)
+    return out
+
+
+def _run_reference(
+    slot: dict[str, str],
+    ref_messages: list[dict[str, Any]],
+    *,
+    temperature: float | None = None,
+    max_tokens: int | None = None,
+) -> tuple[str, str]:
+    """Call one reference model and return ``(label, text)``.
+
+    The slot is resolved to its provider's real runtime (via ``_slot_runtime``)
+    and called through the same ``call_llm`` request-building path any model
+    uses, so per-model wire-format handling (anthropic_messages,
+    max_completion_tokens, fixed/forbidden temperature) applies identically to
+    a reference as it would if that model were the acting model. MoA imposes no
+    cap of its own (``max_tokens`` defaults to ``None`` → omitted → the model's
+    real maximum); ``temperature`` is only the user's configured preset value,
+    which call_llm may still override per model.
+
+    Never raises: a failed reference becomes a labelled note so the aggregator
+    can still act with partial context. Designed to run inside a thread pool —
+    ``call_llm`` is synchronous/blocking, so threads (not asyncio) are the right
+    concurrency primitive, mirroring ``delegate_task``'s batch fan-out.
+    """
+    label = _slot_label(slot)
+    try:
+        response = call_llm(
+            task="moa_reference",
+            messages=ref_messages,
+            temperature=temperature,
+            max_tokens=max_tokens,
+            **_slot_runtime(slot),
+        )
+        return label, _extract_text(response) or "(empty response)"
+    except Exception as exc:
+        logger.warning("MoA reference model %s failed: %s", label, exc)
+        return label, f"[failed: {exc}]"
+
+
+def _run_references_parallel(
+    reference_models: list[dict[str, str]],
+    ref_messages: list[dict[str, Any]],
+    *,
+    temperature: float | None = None,
+    max_tokens: int | None = None,
+) -> list[tuple[str, str]]:
+    """Fan out all reference models in parallel, returning outputs in order.
+
+    Like ``delegate_task``'s batch mode, every reference is dispatched at once
+    and we block until all of them finish before handing the joined results to
+    the aggregator. Output order matches ``reference_models`` so the
+    ``Reference {idx}`` labelling stays stable. MoA presets that reference
+    another MoA preset are skipped here (recursion guard) with a labelled note.
+    """
+    if not reference_models:
+        return []
+
+    results: list[tuple[str, str] | None] = [None] * len(reference_models)
+    futures = {}
+    workers = min(_MAX_REFERENCE_WORKERS, len(reference_models))
+    with ThreadPoolExecutor(max_workers=workers) as executor:
+        for idx, slot in enumerate(reference_models):
+            if slot.get("provider") == "moa":
+                results[idx] = (
+                    _slot_label(slot),
+                    "[skipped: MoA presets cannot recursively reference MoA]",
+                )
+                continue
+            futures[
+                executor.submit(
+                    _run_reference,
+                    slot,
+                    ref_messages,
+                    temperature=temperature,
+                    max_tokens=max_tokens,
+                )
+            ] = idx
+        # Collect every reference before returning — the aggregator needs the
+        # complete set, so there is no early-exit / first-completed path here.
+        for future, idx in futures.items():
+            results[idx] = future.result()
+
+    return [r for r in results if r is not None]
+
+
+def _reference_messages(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    """Build an advisory-safe view of the conversation for reference models.
+
+    Reference calls are advisory: they never call tools and never emit the
+    ``tool_calls`` the main model did. Replaying the full transcript verbatim
+    (a) re-bills the ~8K-token Hermes system prompt per reference per
+    iteration and (b) risks 400s from strict providers (Mistral, Fireworks)
+    that reject orphan ``tool`` messages or ``tool_calls`` the reference never
+    produced. We keep only the user/assistant *text* turns, dropping the
+    system prompt, any ``tool``-role messages, and any ``tool_calls`` payloads.
+    """
+    trimmed: list[dict[str, Any]] = []
+    for msg in messages:
+        role = msg.get("role")
+        if role not in ("user", "assistant"):
+            # Drop system prompt and tool-result messages.
+            continue
+        content = msg.get("content")
+        if not isinstance(content, str):
+            # Skip non-text (multimodal/tool-call-only) assistant turns.
+            if not content:
+                continue
+        text = content if isinstance(content, str) else ""
+        if role == "assistant" and not text.strip():
+            # Assistant turn that was purely tool calls — nothing advisory.
+            continue
+        trimmed.append({"role": role, "content": text})
+    if not trimmed:
+        # Degenerate case (e.g. first turn was stripped): fall back to a
+        # minimal user turn so the reference still has something to answer.
+        for msg in reversed(messages):
+            if msg.get("role") == "user" and isinstance(msg.get("content"), str):
+                return [{"role": "user", "content": msg["content"]}]
+    return trimmed
+
+
+
+def _extract_text(response: Any) -> str:
+    try:
+        transport = get_transport("chat_completions")
+        if transport is None:
+            raise RuntimeError("chat_completions transport unavailable")
+        normalized = transport.normalize_response(response)
+        text = (normalized.content or "").strip()
+        if text:
+            return text
+    except Exception:
+        pass
+    try:
+        content = response.choices[0].message.content
+        return (content or "").strip()
+    except Exception:
+        return ""
+
+
+def aggregate_moa_context(
+    *,
+    user_prompt: str,
+    api_messages: list[dict[str, Any]],
+    reference_models: list[dict[str, str]],
+    aggregator: dict[str, str],
+    temperature: float = 0.6,
+    aggregator_temperature: float = 0.4,
+    max_tokens: int | None = None,
+) -> str:
+    """Run configured reference models and synthesize their advice.
+
+    Failures are returned as model-specific notes instead of aborting the normal
+    agent loop; the main model can still act with partial context.
+
+    ``max_tokens`` is ``None`` by default: MoA does not cap reference or
+    aggregator output, so each model uses its own maximum. ``call_llm`` omits
+    the parameter entirely when it is ``None`` (see its docstring), which also
+    sidesteps providers that reject ``max_tokens`` outright. A hardcoded cap
+    here previously truncated long aggregator syntheses.
+    """
+    reference_outputs: list[tuple[str, str]] = []
+    ref_messages = _reference_messages(api_messages)
+    reference_outputs = _run_references_parallel(
+        reference_models,
+        ref_messages,
+        temperature=temperature,
+        max_tokens=max_tokens,
+    )
+
+    joined = "\n\n".join(
+        f"Reference {idx} — {label}:\n{text}"
+        for idx, (label, text) in enumerate(reference_outputs, start=1)
+    )
+    synth_prompt = (
+        "You are the aggregator in a Mixture of Agents process. Synthesize the "
+        "reference responses into concise, actionable guidance for the main "
+        "Hermes agent. Focus on next steps, tool-use strategy, risks, and any "
+        "disagreements. Do not answer the user directly unless that is all that "
+        "is needed; produce context the main agent should use in its normal loop.\n\n"
+        f"Original user prompt:\n{user_prompt}\n\n"
+        f"Reference responses:\n{joined}"
+    )
+
+    agg_label = _slot_label(aggregator)
+    try:
+        response = call_llm(
+            task="moa_aggregator",
+            messages=[{"role": "user", "content": synth_prompt}],
+            temperature=aggregator_temperature,
+            max_tokens=max_tokens,
+            **_slot_runtime(aggregator),
+        )
+        synthesis = _extract_text(response)
+    except Exception as exc:
+        logger.warning("MoA aggregator model %s failed: %s", agg_label, exc)
+        synthesis = ""
+
+    if not synthesis:
+        synthesis = joined
+
+    return (
+        "[Mixture of Agents context — use this as private guidance for the "
+        "normal Hermes agent loop. You may call tools, continue reasoning, or "
+        "finish normally.]\n"
+        f"Aggregator: {agg_label}\n"
+        f"References: {', '.join(_slot_label(slot) for slot in reference_models)}\n\n"
+        f"{synthesis.strip()}"
+    )
+
+
+class MoAChatCompletions:
+    """OpenAI-chat-compatible facade where the aggregator is the acting model."""
+
+    def __init__(self, preset_name: str, reference_callback: Any = None):
+        self.preset_name = preset_name or "default"
+        # Optional display hook. Called as reference outputs become available so
+        # frontends can show each reference model's answer as a labelled block
+        # before the aggregator acts. Signature:
+        #   reference_callback(event, **kwargs)
+        # where event is one of:
+        #   "moa.reference"   kwargs: index, count, label, text
+        #   "moa.aggregating" kwargs: aggregator (label), ref_count
+        # Never raises into the model call — display is best-effort.
+        self.reference_callback = reference_callback
+        # Turn-scoped reference cache. The agent loop calls create() once per
+        # tool-loop iteration, but references are advisory for the whole turn:
+        # the advisory message view (_reference_messages) is identical across
+        # iterations (it strips tool/tool_call turns) until a new user message
+        # arrives. Re-running references every iteration would multiply their
+        # API cost by the tool-loop depth AND re-emit the same blocks to the
+        # display on every iteration. So cache outputs keyed by the advisory
+        # view's signature and reuse them — running and showing references once
+        # per user turn.
+        self._ref_cache_key: tuple | None = None
+        self._ref_cache_outputs: list[tuple[str, str]] = []
+
+    def _emit(self, event: str, **kwargs: Any) -> None:
+        cb = self.reference_callback
+        if cb is None:
+            return
+        try:
+            cb(event, **kwargs)
+        except Exception as exc:  # pragma: no cover - display must never break the turn
+            logger.debug("MoA reference_callback failed for %s: %s", event, exc)
+
+    def create(self, **api_kwargs: Any) -> Any:
+        from hermes_cli.config import load_config
+        from hermes_cli.moa_config import resolve_moa_preset
+
+        preset = resolve_moa_preset(load_config().get("moa") or {}, self.preset_name)
+        messages = list(api_kwargs.get("messages") or [])
+        reference_models = preset.get("reference_models") or []
+        aggregator = preset.get("aggregator") or {}
+        # MoA does not cap reference or aggregator output: each model uses its
+        # own maximum. Passing max_tokens=None makes call_llm omit the parameter
+        # (it never caps by default), so a long aggregator synthesis is never
+        # truncated and providers that reject max_tokens don't 400.
+        temperature = float(preset.get("reference_temperature", 0.6) or 0.6)
+        aggregator_temperature = float(preset.get("aggregator_temperature", api_kwargs.get("temperature") or 0.4) or 0.4)
+
+        # When the preset is disabled, skip the reference fan-out and let the
+        # configured aggregator act alone — it is the preset's acting model, so
+        # a disabled MoA preset is simply "use the aggregator directly."
+        if not preset.get("enabled", True):
+            reference_models = []
+
+        reference_outputs: list[tuple[str, str]] = []
+        ref_messages = _reference_messages(messages)
+
+        # Turn-scoped cache: only run + display references when the advisory
+        # view changed (i.e. a new user turn). Within one turn the agent loop
+        # calls create() once per tool iteration with the same advisory view;
+        # reuse the cached outputs and skip both the re-run and the re-emit.
+        _sig = hashlib.sha256(
+            "\u0000".join(
+                f"{m.get('role')}:{m.get('content')}" for m in ref_messages
+            ).encode("utf-8", "replace")
+        ).hexdigest()
+        _cache_key = (self.preset_name, _sig, tuple(_slot_label(s) for s in reference_models))
+        _refs_from_cache = _cache_key == self._ref_cache_key and bool(self._ref_cache_outputs)
+
+        if _refs_from_cache:
+            reference_outputs = list(self._ref_cache_outputs)
+        else:
+            reference_outputs = _run_references_parallel(
+                reference_models,
+                ref_messages,
+                temperature=temperature,
+                max_tokens=None,
+            )
+            self._ref_cache_key = _cache_key
+            self._ref_cache_outputs = list(reference_outputs)
+
+            # Surface each reference model's answer to the display BEFORE the
+            # aggregator acts — once per turn (only on the iteration that
+            # actually ran them). The user sees one labelled block per
+            # reference (rendered like a thinking block) so the MoA process is
+            # visible rather than a silent pause. Best-effort: never blocks the
+            # turn.
+            _ref_count = len(reference_outputs)
+            for _idx, (_label, _text) in enumerate(reference_outputs, start=1):
+                self._emit(
+                    "moa.reference",
+                    index=_idx,
+                    count=_ref_count,
+                    label=_label,
+                    text=_text,
+                )
+            if _ref_count:
+                self._emit(
+                    "moa.aggregating",
+                    aggregator=_slot_label(aggregator),
+                    ref_count=_ref_count,
+                )
+
+        agg_messages = [dict(m) for m in messages]
+        if reference_outputs:
+            joined = "\n\n".join(
+                f"Reference {idx} — {label}:\n{text}"
+                for idx, (label, text) in enumerate(reference_outputs, start=1)
+            )
+            guidance = (
+                "[Mixture of Agents reference context]\n"
+                f"Preset: {self.preset_name}\n"
+                f"Aggregator/acting model: {_slot_label(aggregator)}\n"
+                f"References: {', '.join(label for label, _ in reference_outputs)}\n\n"
+                "Use the reference responses below as private context. You are the aggregator and acting model: "
+                "answer the user directly or call tools as needed.\n\n"
+                f"{joined}"
+            )
+            for msg in reversed(agg_messages):
+                if msg.get("role") == "user" and isinstance(msg.get("content"), str):
+                    msg["content"] = msg["content"] + "\n\n" + guidance
+                    break
+            else:
+                agg_messages.append({"role": "user", "content": guidance})
+
+        if aggregator.get("provider") == "moa":
+            raise RuntimeError("MoA aggregator cannot be another MoA preset")
+        agg_kwargs = dict(api_kwargs)
+        agg_kwargs["messages"] = agg_messages
+        # The aggregator is the acting model. Resolve its slot to the provider's
+        # real runtime (base_url/api_key/api_mode) and call it through the same
+        # request-building path any model uses — so per-model wire-format
+        # handling (anthropic_messages, max_completion_tokens, fixed/forbidden
+        # temperature) applies identically to it. MoA imposes no output cap:
+        # max_tokens is passed through from the caller (normally None → omitted
+        # → the model's real maximum). The preset's old hardcoded 4096 default
+        # is gone — it truncated long syntheses.
+        return call_llm(
+            task="moa_aggregator",
+            messages=agg_messages,
+            temperature=aggregator_temperature,
+            max_tokens=agg_kwargs.get("max_tokens"),
+            tools=agg_kwargs.get("tools"),
+            extra_body=agg_kwargs.get("extra_body"),
+            **_slot_runtime(aggregator),
+        )
+
+
+class MoAClient:
+    def __init__(self, preset_name: str, reference_callback: Any = None):
+        self.chat = type("_MoAChat", (), {})()
+        self.chat.completions = MoAChatCompletions(preset_name, reference_callback=reference_callback)

agent/model_metadata.py

@@ -1646,6 +1646,34 @@ def get_model_context_length(
     if config_context_length is not None and isinstance(config_context_length, int) and config_context_length > 0:
         return config_context_length
 
+    # 0a. MoA virtual provider — ``model`` is a preset name, not a real model,
+    # and ``base_url`` is the local virtual endpoint, so every probe below would
+    # miss and fall through to the 256K default. The aggregator is the acting
+    # model, so resolve the context window from the aggregator slot's real
+    # provider+model instead. References are advisory-only and never bound the
+    # acting context, so they're ignored here.
+    if (provider or "").strip().lower() == "moa":
+        try:
+            from hermes_cli.config import load_config
+            from hermes_cli.moa_config import resolve_moa_preset
+            from hermes_cli.runtime_provider import resolve_runtime_provider
+
+            preset = resolve_moa_preset(load_config().get("moa") or {}, model)
+            agg = preset.get("aggregator") or {}
+            agg_provider = str(agg.get("provider") or "").strip()
+            agg_model = str(agg.get("model") or "").strip()
+            if agg_model and agg_provider and agg_provider.lower() != "moa":
+                rt = resolve_runtime_provider(requested=agg_provider, target_model=agg_model)
+                return get_model_context_length(
+                    agg_model,
+                    base_url=rt.get("base_url", "") or "",
+                    api_key=rt.get("api_key", "") or "",
+                    provider=agg_provider,
+                )
+        except Exception:
+            logger.debug("MoA aggregator context-length resolution failed", exc_info=True)
+        # Fall through to the generic default if aggregator resolution failed.
+
     # 0b. custom_providers per-model override — check before any probe.
     # This closes the gap where /model switch and display paths used to fall
     # back to 128K despite the user having a per-model context_length set.

agent/prompt_builder.py

@@ -243,7 +243,10 @@ KANBAN_GUIDANCE = (
     "- **Workspace.** `cd $HERMES_KANBAN_WORKSPACE` first. For a `worktree` kind "
     "with no `.git`, `git worktree add <path> "
     "${HERMES_KANBAN_BRANCH:-wt/$HERMES_KANBAN_TASK}` from the main repo, then "
-    "cd there.\n"
+    "cd there. For a project-linked task the workspace is a fresh "
+    "`<repo>/.worktrees/<task-id>` and `$HERMES_KANBAN_BRANCH` a deterministic "
+    "`<project-slug>/<task-id>` — the main repo is two levels up, so run "
+    "`git worktree add` from there.\n"
     "- **Deliverables.** Files a human wants go in "
     "`kanban_complete(artifacts=[<absolute paths>])` (top-level param; paths in "
     "`metadata` are NOT uploaded). Files must exist at completion.\n"
@@ -614,7 +617,12 @@ DEVELOPER_ROLE_MODELS = ("gpt-5", "codex")
 PLATFORM_HINTS = {
     "whatsapp": (
         "You are on a text messaging communication platform, WhatsApp. "
-        "Please do not use markdown as it does not render. "
+        "Standard markdown (**bold**, *italic*, ~~strike~~, # headers, "
+        "`code`, ```code blocks```, [links](url)) is auto-converted to "
+        "WhatsApp's native syntax (*bold*, _italic_, ~strike~, monospace) — "
+        "feel free to write in markdown, and use bullet lists ('- item') "
+        "freely. Tables are NOT supported — prefer bullet lists or labeled "
+        "key:value pairs. "
         "You can send media files natively: to deliver a file to the user, "
         "include MEDIA:/absolute/path/to/file in your response. The file "
         "will be sent as a native WhatsApp attachment — images (.jpg, .png, "
@@ -679,7 +687,11 @@ PLATFORM_HINTS = {
     ),
     "signal": (
         "You are on a text messaging communication platform, Signal. "
-        "Please do not use markdown as it does not render. "
+        "Standard markdown (**bold**, *italic*, ~~strike~~, # headers, "
+        "`code`, ```code blocks```) is auto-converted to Signal's native "
+        "rich formatting — feel free to write in markdown, and use bullet "
+        "lists ('- item') freely (they render as • bullets). Tables are NOT "
+        "supported — prefer bullet lists or labeled key:value pairs. "
         "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 as attachments, and other "

agent/reasoning_timeouts.py

@@ -0,0 +1,216 @@
+"""Per-reasoning-model stale-timeout floor for known reasoning models.
+
+Reasoning models (those that emit extended thinking blocks before their
+first content token) routinely exceed Hermes's default chat-model
+stale detectors:
+
+* Stream stale detector:   ``HERMES_STREAM_STALE_TIMEOUT``     default 180s
+                           ``agent/chat_completion_helpers.py:2544``
+* Non-stream stale detector: ``HERMES_API_CALL_STALE_TIMEOUT``  default 90s
+                           ``run_agent.py:1140``
+
+For NVIDIA Nemotron 3 Ultra on the hosted NIM gateway the empirical
+upstream idle kill is ~120s (first-party reproduction at
+NVIDIA/NemoClaw#4846 — TTFB ~31s, stream dies at 120s). The same
+failure mode exists on OpenAI o1/o3, Anthropic Opus 4.x thinking,
+DeepSeek R1, Qwen QwQ, xAI Grok reasoning — every cloud reasoning
+model hits upstream-proxies / load-balancers with idle timeouts
+shorter than the model's thinking phase. Result: the stale detector
+kills the connection mid-think, surfacing as
+``BrokenPipeError``/``RemoteProtocolError`` on the next read.
+
+This module provides a floor that the existing stale-detector scaling
+blocks consult via :func:`get_reasoning_stale_timeout_floor` and
+apply as ``max(default, floor)``. It is a FLOOR:
+
+* Never overrides explicit user config (``providers.<id>.models.<model>.stale_timeout_seconds``
+  or ``request_timeout_seconds`` already wins — this code never runs
+  in that branch).
+* Never lowers an existing threshold.
+* Has zero effect on non-reasoning models — they are not in the
+  allowlist and the resolver returns ``None``.
+
+Matching uses start-anchored regex on the slug-only component of
+the model name (after stripping any aggregator prefix like
+``openai/``, ``x-ai/``, ``anthropic/``).  The right-anchor matches
+end-of-string or a ``-``/``.``/``_`` slug separator, so ``qwen3-235b``
+matches the ``qwen3`` family entry (a future model slug would be
+``qwen3-235b-instruct`` and would also match) but ``some-other-qwen3``
+does NOT match ``qwen3`` (the ``-qwen3`` is not at start of slug).
+
+The ``o1`` case is the most delicate: a model named
+``llama-4-70b-o1-preview`` is a hypothetical community derivative that
+should NOT trigger the reasoning-model floor for the user (the user
+chose a non-OpenAI model, not a reasoning model).  The start-of-slug
+anchor naturally excludes this — the matched ``o1-preview`` is at
+position 11 of the slug, not at position 0.  The previous substring-
+with-trailing-hyphen design would have over-matched here, which is
+why start-of-slug anchoring is the right shape.
+
+Fixes #52217.
+"""
+
+from __future__ import annotations
+
+import re
+from typing import Optional
+
+
+# (slug, floor_seconds).  Each slug is matched as a discrete
+# word-boundary component via the wrapper regex in ``_match_any``
+# below.  Order is irrelevant — the first regex match wins.
+_REASONING_STALE_TIMEOUT_FLOORS: tuple[tuple[str, int], ...] = (
+    # NVIDIA Nemotron — reasoning models behind hosted NIM with
+    # documented 60-180s upstream idle kill (NVIDIA/NemoClaw#4846:
+    # 120s measured).
+    ("nemotron-3-ultra", 600),
+    ("nemotron-3-super", 600),
+    ("nemotron-3-nano",  300),
+    # DeepSeek — R1 reasoning model on hosted NIM / DeepSeek direct.
+    ("deepseek-r1", 600),
+    ("deepseek-reasoner", 600),
+    # Qwen — QwQ reasoning + Qwen3 thinking variants.  QwQ-32B
+    # preview is the stable slug; ``qwen3`` covers the family of
+    # thinking-mode Qwen3 models (qwen3-235b-a22b, qwen3-32b, etc.)
+    # without over-matching every Qwen3 instruct variant — the
+    # right-anchor requires the slug to be at the start of the
+    # remaining model name, so ``qwen3-235b-instruct`` (instruct is
+    # NOT a thinking variant) would still match.  Acceptable
+    # trade-off: instruct variants of qwen3 get the 180s floor
+    # even though they don't reason.  The cost is a slightly longer
+    # wait on a hung provider; the alternative (matching only
+    # ``qwen3-.*-thinking``) breaks the moment NVIDIA or Alibaba
+    # ships a slightly different naming shape.
+    ("qwq-32b", 300),
+    ("qwen3", 180),
+    # OpenAI o-series — known multi-minute TTFB.  Each variant
+    # enumerated explicitly so bare ``o1`` doesn't over-match
+    # ``olmo-1`` or hypothetical future community derivatives.
+    ("o1", 600),
+    ("o1-mini", 600),
+    ("o1-pro", 600),
+    ("o1-preview", 600),
+    ("o3", 600),
+    ("o3-pro", 600),
+    ("o3-mini", 300),
+    ("o4-mini", 300),
+    # Anthropic Claude 4.x thinking variants.  Anchored at
+    # ``claude-opus-4`` so non-thinking Claude 3.x or future
+    # non-reasoning Claude variants don't match.
+    ("claude-opus-4", 240),
+    ("claude-sonnet-4.5", 180),
+    ("claude-sonnet-4.6", 180),
+    # xAI Grok reasoning variants.  Explicit reasoning-only keys
+    # plus one for the ``non-reasoning`` variant so users picking
+    # the fast variant don't get the 300s floor.  Bare ``grok-3``,
+    # ``grok-4`` etc. don't match — only the explicit reasoning /
+    # non-reasoning pairs.
+    ("grok-4-fast-reasoning", 300),
+    ("grok-4.20-reasoning", 300),
+    ("grok-4-fast-non-reasoning", 180),
+)
+
+
+# Pre-compile each pattern.  Wrapper = start-of-slug + slug + end-or-
+# separator, where ``start-of-slug`` means start-of-string OR
+# immediately after the last ``/`` (aggregator separator) and
+# ``end-or-separator`` means end-of-string OR a ``-``/``.``/``_``.
+#
+# Why start-of-slug and not start-of-string: aggregator prefixes
+# like ``openai/`` should not affect matching — the slug identity is
+# the part after the last ``/``.  Stripping the aggregator prefix in
+# :func:`get_reasoning_stale_timeout_floor` before regex matching
+# gives the wrapper a clean start-of-string anchor.
+#
+# Why end-or-separator on the right: ``openai/o3-mini`` must match
+# the ``o3-mini`` slug (the right anchor is end-of-string).  And
+# ``openai/o3-mini-2025-01-31`` must also match ``o3-mini`` (the right
+# anchor is the ``-`` separator).  But ``openai/o3-mini-fork`` should
+# NOT match ``o3-mini`` if we wanted to exclude forks — though the
+# pattern ``o3-mini-fork`` would be matched as a derivative anyway,
+# so we accept that community forks inheriting the same prefix are
+# treated as reasoning models (a reasonable default — the upstream
+# gateway timing is the same).
+_PATTERN_CACHE: dict[str, re.Pattern[str]] = {}
+
+
+def _get_pattern(slug: str) -> re.Pattern[str]:
+    compiled = _PATTERN_CACHE.get(slug)
+    if compiled is None:
+        compiled = re.compile(
+            r"^"
+            + re.escape(slug)
+            + r"(?:$|[\-._])"
+        )
+        _PATTERN_CACHE[slug] = compiled
+    return compiled
+
+
+def _match_any(model_lower: str) -> Optional[float]:
+    """Return the floor for the first matching slug, else None.
+
+    Each table entry is matched as a start-of-slug prefix with the
+    slug-separator-or-end-of-string right-anchor.  Table iteration
+    order is irrelevant: longest slug wins (so ``o3-mini`` beats
+    ``o3`` on a model like ``openai/o3-mini``).
+    """
+    # Sort by slug length descending so longer / more-specific slugs
+    # win on shared prefixes (o3-mini beats o3).
+    sorted_floors = sorted(
+        _REASONING_STALE_TIMEOUT_FLOORS, key=lambda kv: -len(kv[0])
+    )
+    for slug, floor in sorted_floors:
+        if _get_pattern(slug).search(model_lower):
+            return float(floor)
+    return None
+
+
+def get_reasoning_stale_timeout_floor(model: object) -> Optional[float]:
+    """Return the stale-timeout floor (seconds) for a known reasoning model.
+
+    Returns ``None`` when the model is not in the allowlist or the
+    argument is empty / not a string.  Matching uses
+    word-boundary-anchored regex on the lowercased model name, so
+    ``openai/o3-mini`` matches the ``o3-mini`` slug but
+    ``olmo-1`` does NOT match ``o1`` (the ``o1`` substring is not
+    at a word boundary inside ``olmo-1``).
+
+    Aggregator prefixes (``openai/``, ``x-ai/``, ``anthropic/`` etc.)
+    are preserved through matching — the ``/`` is itself a word
+    boundary, so ``openai/o3-mini`` matches ``o3-mini`` because the
+    ``/`` before ``o3-mini`` satisfies the left-anchor alternation.
+
+    This is a FLOOR — callers must apply it as ``max(default, floor)``
+    and only when no explicit user-configured per-model
+    ``stale_timeout_seconds`` exists.
+
+    >>> get_reasoning_stale_timeout_floor("nvidia/nemotron-3-ultra-550b-a55b")
+    600.0
+    >>> get_reasoning_stale_timeout_floor("openai/o3-mini")
+    300.0
+    >>> get_reasoning_stale_timeout_floor("deepseek/deepseek-r1")
+    600.0
+    >>> get_reasoning_stale_timeout_floor("qwen/qwen3-235b-a22b-thinking")
+    180.0
+    >>> get_reasoning_stale_timeout_floor("x-ai/grok-4-fast-reasoning")
+    300.0
+    >>> get_reasoning_stale_timeout_floor("anthropic/claude-opus-4-6")
+    240.0
+    >>> get_reasoning_stale_timeout_floor("gpt-4o") is None
+    True
+    >>> get_reasoning_stale_timeout_floor("olmo-1") is None
+    True
+    >>> get_reasoning_stale_timeout_floor(None) is None
+    True
+    """
+    if not model or not isinstance(model, str):
+        return None
+    name = model.strip().lower()
+    if not name:
+        return None
+    # Strip aggregator prefix (everything before and including the
+    # last ``/``).  The wrapper regex anchors at start-of-string, so
+    # the slug identity is the bare model name.
+    if "/" in name:
+        name = name.rsplit("/", 1)[1]
+    return _match_any(name)

agent/redact.py

@@ -107,12 +107,60 @@ _PREFIX_PATTERNS = [
     r"ntn_[A-Za-z0-9]{10,}",            # Notion internal integration token
 ]
 
-# ENV assignment patterns: KEY=value where KEY contains a secret-like name
+# ENV assignment patterns: KEY=value where KEY contains a secret-like name.
+# Uppercase keys tolerate spaces around "=" (e.g. ``FOO_SECRET = bar``) because
+# an all-caps key is almost never prose/code.
 _SECRET_ENV_NAMES = r"(?:API_?KEY|TOKEN|SECRET|PASSWORD|PASSWD|CREDENTIAL|AUTH)"
 _ENV_ASSIGN_RE = re.compile(
     rf"([A-Z0-9_]{{0,50}}{_SECRET_ENV_NAMES}[A-Z0-9_]{{0,50}})\s*=\s*(['\"]?)(\S+)\2",
 )
 
+# Lowercase / dotted / hyphenated config keys from config files
+# (application.properties, .env, YAML-ish dumps): ``spring.datasource.password=secret``,
+# ``app.api.key=xyz``, ``password=secret``. The uppercase _ENV_ASSIGN_RE above
+# never matched these, so config-file passwords leaked verbatim (issue #16413).
+#
+# These run only in a config-file context, NOT in prose, code, or URLs — three
+# carve-outs preserved from the original design (#4367 + the documented
+# web-URL passthrough below):
+#   1. The value is bounded by ``[^\s&]`` (stops at whitespace AND ``&``) so
+#      form-urlencoded bodies are handled pair-by-pair (by _redact_form_body),
+#      not greedily swallowed.
+#   2. _CFG_DOTTED_RE only matches when the key is NAMESPACED (contains a dot),
+#      which is unambiguously a config key — never a prose word.
+#   3. _CFG_ANCHORED_RE matches a bare secret-word key only at line start
+#      (optionally after ``export``), so conversational ``I have password=foo``
+#      mid-sentence is left alone.
+# The colon-form URL guard (skip when ``://`` present) lives at the call site.
+_SECRET_CFG_NAMES = r"(?:api[ _.\-]?key|token|secret|passwd|password|credential|auth)"
+_CFG_VALUE = r"(['\"]?)([^\s&]+?)\2(?=[\s&]|$)"
+# Namespaced (dotted) key: the secret word may sit anywhere in a dotted path.
+_CFG_DOTTED_RE = re.compile(
+    rf"((?:[A-Za-z0-9_\-]+\.)+[A-Za-z0-9_.\-]*{_SECRET_CFG_NAMES}[A-Za-z0-9_.\-]*"
+    rf"|[A-Za-z0-9_.\-]*{_SECRET_CFG_NAMES}[A-Za-z0-9_.\-]*\.[A-Za-z0-9_.\-]+)"
+    rf"={_CFG_VALUE}",
+    re.IGNORECASE,
+)
+# Line-anchored bare key: ``password=…`` / ``export api_key=…`` at start of line.
+_CFG_ANCHORED_RE = re.compile(
+    rf"(^[ \t]*(?:export[ \t]+)?[A-Za-z0-9_\-]*{_SECRET_CFG_NAMES}[A-Za-z0-9_\-]*)={_CFG_VALUE}",
+    re.IGNORECASE | re.MULTILINE,
+)
+
+# Unquoted YAML / colon config (e.g. ``password: secret``,
+# ``spring.datasource.password: hunter2``). The secret keyword must be part of
+# the KEY (anchored to the start of the line/indent), and the value is a single
+# whitespace-free token — so prose like ``note: secret meeting`` (keyword in the
+# value) and ``error: token expired`` are left alone. Bare ``auth`` is excluded
+# from the key set so ``Authorization:`` / ``author:`` don't match (the former
+# is masked by _AUTH_HEADER_RE); ``auth_token``/``auth-token`` still match via
+# the ``token`` keyword. Quoted values defer to _JSON_FIELD_RE via the lookahead.
+_YAML_CFG_NAMES = r"(?:api[ _.\-]?key|token|secret|passwd|password|credential)"
+_YAML_ASSIGN_RE = re.compile(
+    rf"(^[ \t]*[A-Za-z0-9_.\-]*{_YAML_CFG_NAMES}[A-Za-z0-9_.\-]*)(:[ \t]*)(?!['\"])([^\s&]+)",
+    re.IGNORECASE | re.MULTILINE,
+)
+
 # JSON field patterns: "apiKey": "value", "token": "value", etc.
 _JSON_KEY_NAMES = r"(?:api_?[Kk]ey|token|secret|password|access_token|refresh_token|auth_token|bearer|secret_value|raw_secret|secret_input|key_material)"
 _JSON_FIELD_RE = re.compile(
@@ -382,6 +430,13 @@ def redact_sensitive_text(text: str, *, force: bool = False, code_file: bool = F
                 name, quote, value = m.group(1), m.group(2), m.group(3)
                 return f"{name}={quote}{_mask_token(value)}{quote}"
             text = _ENV_ASSIGN_RE.sub(_redact_env, text)
+            # Lowercase/dotted config keys (issue #16413). Skip URLs entirely —
+            # web-URL query params are intentionally passed through (see note
+            # near the bottom of this function); _DB_CONNSTR_RE still guards
+            # connection-string passwords.
+            if "://" not in text:
+                text = _CFG_DOTTED_RE.sub(_redact_env, text)
+                text = _CFG_ANCHORED_RE.sub(_redact_env, text)
 
         # JSON fields: "apiKey": "***"  (skip for code files — false positives)
         if ":" in text and '"' in text:
@@ -390,6 +445,15 @@ def redact_sensitive_text(text: str, *, force: bool = False, code_file: bool = F
                 return f'{key}: "{_mask_token(value)}"'
             text = _JSON_FIELD_RE.sub(_redact_json, text)
 
+        # Unquoted YAML / colon config: password: ***  (after JSON so quoted
+        # values are handled there; the lookahead in _YAML_ASSIGN_RE skips
+        # quotes). Skip URLs — web-URL query params pass through by design.
+        if ":" in text and "://" not in text:
+            def _redact_yaml(m):
+                key, sep, value = m.group(1), m.group(2), m.group(3)
+                return f"{key}{sep}{_mask_token(value)}"
+            text = _YAML_ASSIGN_RE.sub(_redact_yaml, text)
+
     # Authorization headers — _AUTH_HEADER_RE matches any scheme after
     # "[Proxy-]Authorization:" case-insensitively, so "uthorization" is the
     # cheapest substring gate that covers every casing without a casefold().

agent/skill_utils.py

@@ -507,6 +507,34 @@ def get_all_skills_dirs() -> List[Path]:
     return dirs
 
 
+def _resolve_for_skill_ownership(path) -> Path:
+    path_obj = path if isinstance(path, Path) else Path(str(path))
+    try:
+        return path_obj.expanduser().resolve()
+    except (OSError, RuntimeError):
+        return path_obj.expanduser().absolute()
+
+
+def is_external_skill_path(path) -> bool:
+    """Return True when ``path`` lives under a configured external skills dir.
+
+    ``skills.external_dirs`` are externally owned: Hermes can discover and view
+    their skills, and foreground user-directed tool calls may still edit them,
+    but autonomous lifecycle maintenance must treat them as read-only. This
+    helper centralizes the ownership boundary so curator/reporting/tool paths do
+    not each need to re-interpret the config.
+    """
+    candidate = _resolve_for_skill_ownership(path)
+    for root in get_external_skills_dirs():
+        resolved_root = _resolve_for_skill_ownership(root)
+        try:
+            candidate.relative_to(resolved_root)
+            return True
+        except ValueError:
+            continue
+    return False
+
+
 # ── Condition extraction ──────────────────────────────────────────────────
 
 

agent/thinking_timeout_guidance.py

@@ -0,0 +1,136 @@
+"""Thinking-timeout detection and user-facing guidance for reasoning models.
+
+When a known reasoning model (NVIDIA Nemotron 3 Ultra, OpenAI o1/o3,
+Anthropic Opus 4.x thinking, DeepSeek R1, Qwen QwQ, xAI Grok reasoning)
+hits a transport-layer error before the first content token arrives, the
+upstream proxy has almost certainly idle-killed a long thinking stream —
+not a true context overflow or a configuration error.  The user needs
+distinct guidance for this case:
+
+    "The model's thinking phase exceeded the upstream proxy's idle
+     timeout before the first content token arrived.  This is a known
+     issue with reasoning models behind cloud gateways (NVIDIA NIM,
+     OpenAI, Anthropic, DeepSeek).  Workarounds in priority order:
+     1. Set `providers.<provider>.models.<model>.stale_timeout_seconds: 900`
+        in `~/.hermes/config.yaml` to extend the per-call timeout...
+     2. Lower `reasoning_budget` or set `reasoning_effort: medium`...
+     3. Use a smaller / faster reasoning model..."
+
+The existing `_is_stream_drop` guidance at
+``agent/conversation_loop.py:3464-3486`` fires for large-file-write
+stream drops ("try execute_code with Python's open() for large files")
+which is the WRONG advice for the thinking-timeout case.  This module
+provides the detection and the message as standalone helpers so the
+detection logic is unit-testable without driving the full retry loop,
+and the message text can be regression-tested for spelling and accuracy.
+
+Part 2 of Fixes #52310.
+"""
+
+from __future__ import annotations
+
+from typing import Optional
+
+
+# Substring set that identifies a transport-layer failure on the
+# response stream.  Same shape as the existing
+# ``_SERVER_DISCONNECT_PATTERNS`` in ``agent/error_classifier.py:394``
+# but extended to also catch the OSS-level error signature
+# (``broken pipe`` / ``errno 32``) that the upstream kill surfaces
+# to the OpenAI SDK wrapper.
+_THINKING_TIMEOUT_SUBSTRINGS: tuple[str, ...] = (
+    "broken pipe",
+    "errno 32",
+    "remote protocol",
+    "connection reset",
+    "connection lost",
+    "peer closed",
+    "server disconnected",
+)
+
+
+def is_thinking_timeout(classified: object, model: str, error_msg: str) -> bool:
+    """Return True when a reasoning model's thinking phase hit a transport kill.
+
+    Args:
+        classified: a :class:`agent.error_classifier.ClassifiedError` instance
+            (duck-typed here to avoid an import cycle in unit tests).
+        model: the model slug at failure time (e.g.
+            ``"nvidia/nemotron-3-ultra-550b-a55b"``).
+        error_msg: lowercased string representation of the underlying
+            exception (typically ``str(api_error).lower()``).
+
+    Returns True when ALL conditions hold:
+        1. ``classified.reason == FailoverReason.timeout`` (the classifier
+           override at ``agent/error_classifier.py:720-738`` ensures this
+           is the case for reasoning models even on large sessions).
+        2. ``api_error`` has no ``.status_code`` attribute set (transport
+           disconnect, not an HTTP error).
+        3. ``model`` is in the reasoning-model allowlist (reuses
+           ``agent.reasoning_timeouts.get_reasoning_stale_timeout_floor``).
+        4. ``error_msg`` contains one of the transport-kill substrings.
+
+    Non-reasoning models always return False.  Non-transport errors
+    (billing / rate_limit / auth / context_overflow / format_error)
+    always return False.  HTTP-status errors always return False.
+    """
+    # Import here (not at module top) to keep this helper cheap to
+    # import even from callers that don't need it.  ``agent.reasoning_timeouts``
+    # is small and dependency-free.
+    from agent.reasoning_timeouts import get_reasoning_stale_timeout_floor
+
+    # Condition 1: classifier says timeout.  Use a string/value check
+    # rather than importing FailoverReason so this module has zero
+    # import cycles from the error_classifier package.
+    reason = getattr(classified, "reason", None)
+    reason_value = getattr(reason, "value", None)
+    if reason_value != "timeout":
+        return False
+
+    # Condition 2: no HTTP status code (transport, not API error).
+    # Caller is expected to gate on ``getattr(api_error, "status_code", None) is None``
+    # before calling this helper; the surface here is just the post-gate
+    # boolean so the caller can pass an already-prepped error_msg.
+
+    # Condition 3: reasoning model allowlist.
+    if get_reasoning_stale_timeout_floor(model) is None:
+        return False
+
+    # Condition 4: transport-kill substring in the error message.
+    error_msg_lower = (error_msg or "").lower()
+    return any(p in error_msg_lower for p in _THINKING_TIMEOUT_SUBSTRINGS)
+
+
+def build_thinking_timeout_guidance(
+    provider: str, model: str, model_label: Optional[str] = None,
+) -> str:
+    """Return the user-facing guidance string appended to ``_final_response``.
+
+    Args:
+        provider: provider slug (e.g. ``"nvidia"``, ``"openai"``).
+        model: bare model slug the user would put in their config
+            (e.g. ``"nemotron-3-ultra-550b-a55b"`` if the user uses
+            NVIDIA direct, or the full ``"nvidia/nemotron-3-ultra-550b-a55b"``
+            if they go through an aggregator).  Used verbatim in the
+            config snippet so the user can copy-paste.
+        model_label: optional short label for the model name in the
+            prose (e.g. ``"Nemotron 3 Ultra"``).  Falls back to the
+            slug if not provided.
+    """
+    label = model_label or model
+    return (
+        "\n\nThe model's thinking phase exceeded the upstream proxy's "
+        "idle timeout before the first content token arrived. This is a "
+        f"known issue with reasoning models (like {label}) behind cloud "
+        "gateways (NVIDIA NIM, OpenAI, Anthropic, DeepSeek). Workarounds "
+        "in priority order:\n"
+        f"1. Set `providers.{provider}.models.{model}.stale_timeout_seconds: 900` "
+        "in `~/.hermes/config.yaml` to extend the per-call timeout. "
+        "(Hermes's built-in floor is 600s for known reasoning models — "
+        "if you still see this after raising, the upstream cap is even "
+        "shorter.)\n"
+        "2. Lower `reasoning_budget` or set `reasoning_effort: medium` on this "
+        "model if the provider supports it.\n"
+        "3. Use a smaller / faster reasoning model if the task doesn't "
+        "require deep thinking."
+    )

agent/tool_executor.py

@@ -26,6 +26,7 @@ from agent.display import (
     build_tool_preview as _build_tool_preview,
     get_cute_tool_message as _get_cute_tool_message_impl,
     get_tool_emoji as _get_tool_emoji,
+    redact_tool_args_for_display as _redact_tool_args_for_display,
     _detect_tool_failure,
 )
 from agent.tool_guardrails import ToolGuardrailDecision
@@ -469,10 +470,11 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
     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)
+            display_args = _redact_tool_args_for_display(name, args) or args
+            args_str = json.dumps(display_args, ensure_ascii=False)
             if agent.verbose_logging:
-                print(f"  📞 Tool {i}: {name}({list(args.keys())})")
-                print(agent._wrap_verbose("Args: ", json.dumps(args, indent=2, ensure_ascii=False)))
+                print(f"  📞 Tool {i}: {name}({list(display_args.keys())})")
+                print(agent._wrap_verbose("Args: ", json.dumps(display_args, indent=2, ensure_ascii=False)))
             else:
                 args_preview = args_str[:agent.log_prefix_chars] + "..." if len(args_str) > agent.log_prefix_chars else args_str
                 print(f"  📞 Tool {i}: {name}({list(args.keys())}) - {args_preview}")
@@ -482,8 +484,9 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
             continue
         if agent.tool_progress_callback:
             try:
-                preview = _build_tool_preview(name, args)
-                agent.tool_progress_callback("tool.started", name, preview, args)
+                display_args = _redact_tool_args_for_display(name, args) or args
+                preview = _build_tool_preview(name, display_args)
+                agent.tool_progress_callback("tool.started", name, preview, display_args)
             except Exception as cb_err:
                 logging.debug(f"Tool progress callback error: {cb_err}")
 
@@ -492,7 +495,8 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
             continue
         if agent.tool_start_callback:
             try:
-                agent.tool_start_callback(tc.id, name, args)
+                display_args = _redact_tool_args_for_display(name, args) or args
+                agent.tool_start_callback(tc.id, name, display_args)
             except Exception as cb_err:
                 logging.debug(f"Tool start callback error: {cb_err}")
 
@@ -792,7 +796,8 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
 
         if not blocked and agent.tool_complete_callback:
             try:
-                agent.tool_complete_callback(tc.id, name, args, function_result)
+                display_args = _redact_tool_args_for_display(name, args) or args
+                agent.tool_complete_callback(tc.id, name, display_args, function_result)
             except Exception as cb_err:
                 logging.debug(f"Tool complete callback error: {cb_err}")
 
@@ -954,10 +959,11 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
             agent._iters_since_skill = 0
 
         if not agent.quiet_mode and getattr(agent, "tool_progress_mode", "all") != "off":
-            args_str = json.dumps(function_args, ensure_ascii=False)
+            display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+            args_str = json.dumps(display_args, ensure_ascii=False)
             if agent.verbose_logging:
-                print(f"  📞 Tool {i}: {function_name}({list(function_args.keys())})")
-                print(agent._wrap_verbose("Args: ", json.dumps(function_args, indent=2, ensure_ascii=False)))
+                print(f"  📞 Tool {i}: {function_name}({list(display_args.keys())})")
+                print(agent._wrap_verbose("Args: ", json.dumps(display_args, indent=2, ensure_ascii=False)))
             else:
                 args_preview = args_str[:agent.log_prefix_chars] + "..." if len(args_str) > agent.log_prefix_chars else args_str
                 print(f"  📞 Tool {i}: {function_name}({list(function_args.keys())}) - {args_preview}")
@@ -978,14 +984,16 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
 
         if not _execution_blocked and agent.tool_progress_callback:
             try:
-                preview = _build_tool_preview(function_name, function_args)
-                agent.tool_progress_callback("tool.started", function_name, preview, function_args)
+                display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+                preview = _build_tool_preview(function_name, display_args)
+                agent.tool_progress_callback("tool.started", function_name, preview, display_args)
             except Exception as cb_err:
                 logging.debug(f"Tool progress callback error: {cb_err}")
 
         if not _execution_blocked and agent.tool_start_callback:
             try:
-                agent.tool_start_callback(tool_call.id, function_name, function_args)
+                display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+                agent.tool_start_callback(tool_call.id, function_name, display_args)
             except Exception as cb_err:
                 logging.debug(f"Tool start callback error: {cb_err}")
 
@@ -1215,7 +1223,8 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
             if agent._should_emit_quiet_tool_messages():
                 face = random.choice(KawaiiSpinner.get_waiting_faces())
                 emoji = _get_tool_emoji(function_name)
-                preview = _build_tool_preview(function_name, function_args) or function_name
+                display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+                preview = _build_tool_preview(function_name, display_args) or function_name
                 spinner = KawaiiSpinner(f"{face} {emoji} {preview}", spinner_type='dots', print_fn=agent._print_fn)
                 spinner.start()
             _ce_result = None
@@ -1248,7 +1257,8 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
             if agent._should_emit_quiet_tool_messages() and agent._should_start_quiet_spinner():
                 face = random.choice(KawaiiSpinner.get_waiting_faces())
                 emoji = _get_tool_emoji(function_name)
-                preview = _build_tool_preview(function_name, function_args) or function_name
+                display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+                preview = _build_tool_preview(function_name, display_args) or function_name
                 spinner = KawaiiSpinner(f"{face} {emoji} {preview}", spinner_type='dots', print_fn=agent._print_fn)
                 spinner.start()
             _mem_result = None
@@ -1279,7 +1289,8 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
             if agent._should_emit_quiet_tool_messages() and agent._should_start_quiet_spinner():
                 face = random.choice(KawaiiSpinner.get_waiting_faces())
                 emoji = _get_tool_emoji(function_name)
-                preview = _build_tool_preview(function_name, function_args) or function_name
+                display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+                preview = _build_tool_preview(function_name, display_args) or function_name
                 spinner = KawaiiSpinner(f"{face} {emoji} {preview}", spinner_type='dots', print_fn=agent._print_fn)
                 spinner.start()
             _spinner_result = None
@@ -1441,7 +1452,8 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
 
         if not _execution_blocked and agent.tool_complete_callback:
             try:
-                agent.tool_complete_callback(tool_call.id, function_name, function_args, function_result)
+                display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+                agent.tool_complete_callback(tool_call.id, function_name, display_args, function_result)
             except Exception as cb_err:
                 logging.debug(f"Tool complete callback error: {cb_err}")
 

agent/turn_context.py

@@ -28,6 +28,7 @@ import uuid
 from dataclasses import dataclass
 from typing import Any, Dict, List, Optional
 
+from agent.conversation_compression import conversation_history_after_compression
 from agent.iteration_budget import IterationBudget
 from agent.model_metadata import (
     estimate_messages_tokens_rough,
@@ -400,7 +401,9 @@ def build_turn_context(
                     _orig_len, len(messages), _orig_tokens, _preflight_tokens
                 ):
                     break  # Cannot compress further: neither rows nor tokens moved
-                conversation_history = None
+                conversation_history = conversation_history_after_compression(
+                    agent, messages
+                )
                 agent._empty_content_retries = 0
                 agent._thinking_prefill_retries = 0
                 agent._last_content_with_tools = None

agent/verification_stop.py

@@ -15,9 +15,135 @@ from typing import Any, Iterable
 
 _MAX_CHANGED_PATHS_IN_NUDGE = 8
 
+# Non-code file extensions whose edits carry no verifiable runtime behavior:
+# documentation, prose, and data/markup that no test/build exercises. When a
+# turn touches ONLY these, verify-on-stop has nothing to check, so the nudge is
+# suppressed (this is fix "C" for the doc/markdown/skill false-positive — a
+# SKILL.md or README edit must never demand a /tmp verification script). A turn
+# that edits any non-listed path (a real source/code/config file) still nudges.
+_NON_CODE_VERIFY_EXTENSIONS = frozenset(
+    {
+        ".md",
+        ".markdown",
+        ".mdx",
+        ".rst",
+        ".txt",
+        ".text",
+        ".adoc",
+        ".asciidoc",
+        ".org",
+        ".log",
+        ".csv",
+        ".tsv",
+    }
+)
+
+# Filenames (case-insensitive, extension-less or otherwise) that are pure prose
+# even without a recognized doc extension.
+_NON_CODE_VERIFY_FILENAMES = frozenset(
+    {
+        "license",
+        "licence",
+        "notice",
+        "authors",
+        "contributors",
+        "changelog",
+        "codeowners",
+    }
+)
+
+
+def _is_non_code_path(raw: str) -> bool:
+    """Return True when a changed path is documentation/prose with nothing to verify."""
+    try:
+        p = Path(str(raw))
+    except Exception:
+        return False
+    suffix = p.suffix.lower()
+    if suffix in _NON_CODE_VERIFY_EXTENSIONS:
+        return True
+    if not suffix and p.name.lower() in _NON_CODE_VERIFY_FILENAMES:
+        return True
+    return False
+
+
+def _filter_verifiable_paths(paths: Iterable[str]) -> list[str]:
+    """Drop documentation/prose paths; keep paths that could have verifiable behavior."""
+    return [p for p in paths if p and not _is_non_code_path(p)]
+
+
+# Session identities (platform or source) that are NOT human conversational
+# messaging surfaces: interactive coding surfaces (CLI, TUI, desktop, codex,
+# local, gateway) and programmatic callers (API server, webhooks, tools).
+# Verify-on-stop stays ON by default for these. Any other resolved gateway
+# platform is a conversational messaging surface (Telegram, Discord, WhatsApp,
+# Signal, Slack, etc.) where the verification narrative would reach a human as
+# chat noise, so it defaults OFF. Mirrors LOCAL_SESSION_SOURCE_IDS in
+# apps/desktop/src/lib/session-source.ts; keep roughly in sync when adding a
+# local or programmatic surface. Default-deny by design: an unrecognized
+# identity is treated as messaging (OFF) so a new chat platform never leaks the
+# verification receipt before this set is updated.
+_NON_MESSAGING_SESSION_SURFACES = frozenset(
+    {
+        "",
+        "cli",
+        "codex",
+        "desktop",
+        "gateway",
+        "local",
+        "tui",
+        "tool",
+        "api_server",
+        "webhook",
+        "msgraph_webhook",
+    }
+)
+
+
+def _session_is_messaging_surface() -> bool:
+    """Return whether this turn is delivered over a human messaging channel.
+
+    The gateway binds the platform value (e.g. ``telegram``) to
+    ``HERMES_SESSION_PLATFORM``; the CLI and TUI set ``HERMES_SESSION_SOURCE``
+    (e.g. ``cli``, ``tui``) instead. Both are consulted via the session-context
+    helper (with an ``os.environ`` fallback), alongside the ``HERMES_PLATFORM``
+    override, matching the sibling platform resolution in
+    ``agent/skill_commands.py`` and ``agent/prompt_builder.py``. A turn is a
+    messaging surface when a resolved identity is present and is not a known
+    non-messaging surface.
+    """
+    try:
+        from gateway.session_context import get_session_env
+
+        platform = (
+            os.getenv("HERMES_PLATFORM")
+            or get_session_env("HERMES_SESSION_PLATFORM", "")
+        )
+        source = get_session_env("HERMES_SESSION_SOURCE", "")
+    except Exception:
+        platform = os.getenv("HERMES_PLATFORM", "") or os.environ.get(
+            "HERMES_SESSION_PLATFORM", ""
+        )
+        source = os.environ.get("HERMES_SESSION_SOURCE", "")
+    for identity in (platform, source):
+        identity = str(identity or "").strip().lower()
+        if identity and identity not in _NON_MESSAGING_SESSION_SURFACES:
+            return True
+    return False
+
 
 def verify_on_stop_enabled(config: dict[str, Any] | None = None) -> bool:
-    """Return whether edit -> verify-before-finish behavior is enabled."""
+    """Return whether edit -> verify-before-finish behavior is enabled.
+
+    Precedence: an explicit ``HERMES_VERIFY_ON_STOP`` env var wins, then an
+    explicit ``agent.verify_on_stop`` config value. The config default is
+    ``False`` (see ``DEFAULT_CONFIG``) — verify-on-stop is OFF unless the user
+    opts in. The legacy ``"auto"`` sentinel is still honored for anyone who
+    sets it explicitly: it resolves to ON for interactive coding surfaces
+    (CLI, TUI, desktop) and programmatic callers, and OFF for conversational
+    messaging surfaces (Telegram, Discord, etc.). A missing/unknown value
+    falls back to OFF.
+    """
     env = os.environ.get("HERMES_VERIFY_ON_STOP")
     if env is not None:
         return env.strip().lower() not in {"0", "false", "no", "off"}
@@ -29,9 +155,20 @@ def verify_on_stop_enabled(config: dict[str, Any] | None = None) -> bool:
         except Exception:
             config = {}
     agent_cfg = (config or {}).get("agent") if isinstance(config, dict) else None
-    if isinstance(agent_cfg, dict) and "verify_on_stop" in agent_cfg:
-        return bool(agent_cfg.get("verify_on_stop"))
-    return True
+    cfg_val = agent_cfg.get("verify_on_stop") if isinstance(agent_cfg, dict) else None
+    if isinstance(cfg_val, bool):
+        return cfg_val
+    if isinstance(cfg_val, str):
+        token = cfg_val.strip().lower()
+        if token in {"1", "true", "yes", "on"}:
+            return True
+        if token in {"0", "false", "no", "off"}:
+            return False
+        if token == "auto":
+            # Explicit opt-in to the legacy surface-aware behavior.
+            return not _session_is_messaging_surface()
+    # Missing or unknown value -> OFF (the new default).
+    return False
 
 
 def _candidate_cwds(paths: Iterable[str]) -> list[Path]:
@@ -114,7 +251,10 @@ def build_verify_on_stop_nudge(
     max_attempts: int = 2,
 ) -> str | None:
     """Return a synthetic follow-up when edited code lacks fresh verification."""
-    paths = sorted({str(p) for p in changed_paths if p})
+    # Drop documentation/prose paths (markdown, skills, README, LICENSE, ...) —
+    # they carry no verifiable behavior, so a turn that touched only those has
+    # nothing to verify and must not nudge.
+    paths = sorted({str(p) for p in _filter_verifiable_paths(changed_paths)})
     if not paths or attempts >= max_attempts:
         return None
 

apps/desktop/components.json

@@ -17,5 +17,5 @@
     "lib": "@/lib",
     "hooks": "@/hooks"
   },
-  "iconLibrary": "lucide"
+  "iconLibrary": "tabler"
 }

apps/desktop/electron/backend-env.cjs

@@ -61,10 +61,7 @@ function buildDesktopBackendPath({
   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 }
-  )
+  return appendUniquePathEntries([hermesNodeBin, venvBin, currentPath, saneEntries], { delimiter })
 }
 
 function normalizeHermesHomeRoot(hermesHome, { pathModule = pathModuleForPlatform(process.platform) } = {}) {

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

@@ -76,10 +76,7 @@ test('normalizeHermesHomeRoot maps profile homes back to the global Hermes root'
     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'
-  )
+  assert.equal(normalizeHermesHomeRoot('/Users/test/.hermes', { pathModule: path.posix }), '/Users/test/.hermes')
 })
 
 test('Windows PATH casing and delimiter are preserved without POSIX sane entries', () => {
@@ -104,8 +101,5 @@ test('Windows PATH casing and delimiter are preserved without POSIX sane entries
 })
 
 test('appendUniquePathEntries drops empty entries and keeps first occurrence', () => {
-  assert.equal(
-    appendUniquePathEntries([':/a::/b', ['/a', '/c']], { delimiter: ':' }),
-    '/a:/b:/c'
-  )
+  assert.equal(appendUniquePathEntries([':/a::/b', ['/a', '/c']], { delimiter: ':' }), '/a:/b:/c')
 })

apps/desktop/electron/backend-ready.cjs

@@ -1,3 +1,5 @@
+const fs = require('node:fs')
+
 const _READY_RE = /^HERMES_DASHBOARD_READY port=(\d+)/m
 
 // The announcement clock starts the instant the backend process is spawned —
@@ -94,9 +96,76 @@ function waitForDashboardPort(child, timeoutMs = resolvePortAnnounceTimeoutMs())
   })
 }
 
+function readDashboardReadyFile(readyFile) {
+  if (!readyFile) return null
+  try {
+    const parsed = JSON.parse(fs.readFileSync(readyFile, 'utf8'))
+    const port = Number(parsed?.port)
+    return Number.isInteger(port) && port > 0 ? port : null
+  } catch {
+    return null
+  }
+}
+
+function waitForDashboardReadyFile(readyFile, child, timeoutMs = resolvePortAnnounceTimeoutMs()) {
+  return new Promise((resolve, reject) => {
+    let done = false
+    let interval = null
+
+    function cleanup() {
+      if (done) return
+      done = true
+      clearTimeout(timer)
+      if (interval) clearInterval(interval)
+      child.off('exit', onExit)
+      child.off('error', onError)
+    }
+
+    function check() {
+      const port = readDashboardReadyFile(readyFile)
+      if (port) {
+        cleanup()
+        resolve(port)
+      }
+    }
+
+    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.on('exit', onExit)
+    child.on('error', onError)
+    interval = setInterval(check, 50)
+    if (typeof interval.unref === 'function') interval.unref()
+    check()
+  })
+}
+
+function waitForDashboardPortAnnouncement(child, options = {}) {
+  const timeoutMs = options.timeoutMs ?? resolvePortAnnounceTimeoutMs()
+  if (options.readyFile) {
+    return waitForDashboardReadyFile(options.readyFile, child, timeoutMs)
+  }
+  return waitForDashboardPort(child, timeoutMs)
+}
+
 module.exports = {
   waitForDashboardPort,
+  waitForDashboardPortAnnouncement,
+  waitForDashboardReadyFile,
+  readDashboardReadyFile,
   resolvePortAnnounceTimeoutMs,
   DEFAULT_PORT_ANNOUNCE_TIMEOUT_MS,
-  MIN_PORT_ANNOUNCE_TIMEOUT_MS,
+  MIN_PORT_ANNOUNCE_TIMEOUT_MS
 }

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

@@ -14,12 +14,18 @@
 const test = require('node:test')
 const assert = require('node:assert/strict')
 const { EventEmitter } = require('node:events')
+const fs = require('node:fs')
+const os = require('node:os')
+const path = require('node:path')
 
 const {
+  readDashboardReadyFile,
   waitForDashboardPort,
+  waitForDashboardPortAnnouncement,
+  waitForDashboardReadyFile,
   resolvePortAnnounceTimeoutMs,
   DEFAULT_PORT_ANNOUNCE_TIMEOUT_MS,
-  MIN_PORT_ANNOUNCE_TIMEOUT_MS,
+  MIN_PORT_ANNOUNCE_TIMEOUT_MS
 } = require('./backend-ready.cjs')
 
 // A minimal stand-in for a spawned child process: an EventEmitter with a
@@ -119,3 +125,75 @@ test('a late announcement after timeout does not throw (listeners torn down)', a
     child.stdout.emit('data', 'HERMES_DASHBOARD_READY port=9999\n')
   })
 })
+
+// ---------------------------------------------------------------------------
+// ready-file port announcement
+// ---------------------------------------------------------------------------
+
+function mkTmpReadyFile() {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-ready-test-'))
+  return {
+    dir,
+    file: path.join(dir, 'ready.json'),
+    cleanup: () => fs.rmSync(dir, { recursive: true, force: true })
+  }
+}
+
+test('readDashboardReadyFile returns a valid port from JSON', () => {
+  const tmp = mkTmpReadyFile()
+  try {
+    fs.writeFileSync(tmp.file, JSON.stringify({ port: 4567 }))
+    assert.equal(readDashboardReadyFile(tmp.file), 4567)
+  } finally {
+    tmp.cleanup()
+  }
+})
+
+test('readDashboardReadyFile ignores missing, malformed, or invalid files', () => {
+  const tmp = mkTmpReadyFile()
+  try {
+    assert.equal(readDashboardReadyFile(tmp.file), null)
+    fs.writeFileSync(tmp.file, '{')
+    assert.equal(readDashboardReadyFile(tmp.file), null)
+    fs.writeFileSync(tmp.file, JSON.stringify({ port: 0 }))
+    assert.equal(readDashboardReadyFile(tmp.file), null)
+  } finally {
+    tmp.cleanup()
+  }
+})
+
+test('waitForDashboardReadyFile resolves when the ready file appears', async () => {
+  const tmp = mkTmpReadyFile()
+  const child = makeFakeChild()
+  try {
+    const p = waitForDashboardReadyFile(tmp.file, child, 1000)
+    setTimeout(() => fs.writeFileSync(tmp.file, JSON.stringify({ port: 8765 })), 20)
+    assert.equal(await p, 8765)
+  } finally {
+    tmp.cleanup()
+  }
+})
+
+test('waitForDashboardPortAnnouncement uses ready file when provided', async () => {
+  const tmp = mkTmpReadyFile()
+  const child = makeFakeChild()
+  try {
+    const p = waitForDashboardPortAnnouncement(child, { readyFile: tmp.file, timeoutMs: 1000 })
+    setTimeout(() => fs.writeFileSync(tmp.file, JSON.stringify({ port: 9876 })), 20)
+    assert.equal(await p, 9876)
+  } finally {
+    tmp.cleanup()
+  }
+})
+
+test('waitForDashboardReadyFile rejects when the child exits before file readiness', async () => {
+  const tmp = mkTmpReadyFile()
+  const child = makeFakeChild()
+  try {
+    const p = waitForDashboardReadyFile(tmp.file, child, 1000)
+    child.emit('exit', 1, null)
+    await assert.rejects(p, /exited before port announcement/)
+  } finally {
+    tmp.cleanup()
+  }
+})

apps/desktop/electron/bootstrap-runner.cjs

@@ -179,7 +179,13 @@ function downloadInstallScript(commit, destPath) {
   })
 }
 
-async function resolveInstallScript({ installStamp, sourceRepoRoot, hermesHome, emit, _download = downloadInstallScript }) {
+async function resolveInstallScript({
+  installStamp,
+  sourceRepoRoot,
+  hermesHome,
+  emit,
+  _download = downloadInstallScript
+}) {
   // 1. Dev shortcut: prefer a local checkout's installer so we can iterate
   //    without pushing. SOURCE_REPO_ROOT comes from main.cjs (path.resolve
   //    of APP_ROOT/../..).
@@ -293,15 +299,19 @@ 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, hiddenWindowsChildOptions({
-      stdio: ['ignore', 'pipe', 'pipe'],
-      env: {
-        ...process.env,
-        // Pass HERMES_HOME through so install.ps1 respects the caller's
-        // choice rather than re-computing the default.
-        HERMES_HOME: hermesHome || process.env.HERMES_HOME || ''
-      }
-    }))
+    const child = spawn(
+      ps,
+      fullArgs,
+      hiddenWindowsChildOptions({
+        stdio: ['ignore', 'pipe', 'pipe'],
+        env: {
+          ...process.env,
+          // Pass HERMES_HOME through so install.ps1 respects the caller's
+          // choice rather than re-computing the default.
+          HERMES_HOME: hermesHome || process.env.HERMES_HOME || ''
+        }
+      })
+    )
 
     let stdout = ''
     let stderr = ''

apps/desktop/electron/connection-config.cjs

@@ -261,12 +261,7 @@ function cookiesHaveSession(cookies) {
  */
 function cookiesHaveLiveSession(cookies) {
   if (!Array.isArray(cookies)) return false
-  return cookies.some(
-    c =>
-      c &&
-      c.value &&
-      (AT_COOKIE_VARIANTS.includes(c.name) || RT_COOKIE_VARIANTS.includes(c.name))
-  )
+  return cookies.some(c => c && c.value && (AT_COOKIE_VARIANTS.includes(c.name) || RT_COOKIE_VARIANTS.includes(c.name)))
 }
 
 module.exports = {

apps/desktop/electron/desktop-uninstall.cjs

@@ -138,10 +138,7 @@ function buildPosixCleanupScript({ desktopPid, pythonExe, pythonPath, agentRoot,
   if (pythonPath) {
     lines.push(`export PYTHONPATH=${q(pythonPath)}\${PYTHONPATH:+:$PYTHONPATH}`)
   }
-  lines.push(
-    `cd ${q(agentRoot)} 2>/dev/null || true`,
-    `${q(pythonExe)} ${uninstallArgs.map(q).join(' ')} || true`
-  )
+  lines.push(`cd ${q(agentRoot)} 2>/dev/null || true`, `${q(pythonExe)} ${uninstallArgs.map(q).join(' ')} || true`)
   if (appPath) {
     lines.push(`rm -rf ${q(appPath)} || true`)
   }
@@ -169,7 +166,15 @@ function buildPosixCleanupScript({ desktopPid, pythonExe, pythonPath, agentRoot,
  * Removal: even after the desktop PID is gone, Windows releases directory
  * handles lazily, so a single `rmdir /s /q` can half-fail — retry up to 10x.
  */
-function buildWindowsCleanupScript({ desktopPid, pythonExe, pythonPath, agentRoot, uninstallArgs, appPath, hermesHome }) {
+function buildWindowsCleanupScript({
+  desktopPid,
+  pythonExe,
+  pythonPath,
+  agentRoot,
+  uninstallArgs,
+  appPath,
+  hermesHome
+}) {
   const pid = Number(desktopPid) || 0
   // cmd.exe has no string escaping inside quotes; strip embedded quotes (paths
   // under %LOCALAPPDATA% never contain them). `&`/`^` in a path would still be

apps/desktop/electron/desktop-uninstall.test.cjs

@@ -101,10 +101,7 @@ test('resolveRemovableAppPath uses APPIMAGE on Linux when set', () => {
 })
 
 test('resolveRemovableAppPath finds the unpacked dir on Linux', () => {
-  assert.equal(
-    resolveRemovableAppPath('/opt/hermes/linux-unpacked/hermes', 'linux', {}),
-    '/opt/hermes/linux-unpacked'
-  )
+  assert.equal(resolveRemovableAppPath('/opt/hermes/linux-unpacked/hermes', 'linux', {}), '/opt/hermes/linux-unpacked')
   // A system-package install (/usr/bin) → null, left to apt/dnf.
   assert.equal(resolveRemovableAppPath('/usr/bin/hermes', 'linux', {}), null)
 })

apps/desktop/electron/embed-referer.cjs

@@ -0,0 +1,48 @@
+'use strict'
+
+const { session } = require('electron')
+
+const EMBED_SESSION_PARTITION = 'persist:hermes-embed'
+const EMBED_REFERER = 'https://www.youtube.com/'
+const YOUTUBE_REFERER_HOST_RE =
+  /(^|\.)(youtube\.com|youtube-nocookie\.com|googlevideo\.com|ytimg\.com|youtubei\.googleapis\.com)$/i
+
+function installEmbedRefererForSession(embedSession) {
+  if (!embedSession) {
+    return
+  }
+
+  embedSession.webRequest.onBeforeSendHeaders((details, callback) => {
+    let host = ''
+
+    try {
+      host = new URL(details.url).hostname
+    } catch {
+      host = ''
+    }
+
+    if (!YOUTUBE_REFERER_HOST_RE.test(host)) {
+      callback({ requestHeaders: details.requestHeaders })
+      return
+    }
+
+    const headers = { ...details.requestHeaders }
+
+    if (!headers.Referer && !headers.referer) {
+      headers.Referer = EMBED_REFERER
+    }
+
+    callback({ requestHeaders: headers })
+  })
+}
+
+/** Stamp Referer on YouTube requests in the embed webview partition only. */
+function installEmbedReferer() {
+  try {
+    installEmbedRefererForSession(session.fromPartition(EMBED_SESSION_PARTITION))
+  } catch {
+    // Non-fatal: embeds still render; YouTube may show referer errors.
+  }
+}
+
+module.exports = { installEmbedReferer }

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

@@ -92,9 +92,7 @@ async function readDirForIpc(dirPath, options = {}) {
   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)
-    )
+    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))
 

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

@@ -349,7 +349,10 @@ test('readDirForIpc bounds concurrent stats while preserving complete sorted out
   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(`${path.sep}node_modules`)), false)
+  assert.equal(
+    statCalls.some(fullPath => fullPath.endsWith(`${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(
@@ -357,8 +360,5 @@ test('readDirForIpc bounds concurrent stats while preserving complete sorted out
     expectedNames
   )
   assert.equal(result.entries.find(entry => entry.name === failedName)?.isDirectory, false)
-  assert.equal(
-    result.entries.filter(entry => entry.isDirectory).length,
-    successfulDirectoryNames.size
-  )
+  assert.equal(result.entries.filter(entry => entry.isDirectory).length, successfulDirectoryNames.size)
 })

apps/desktop/electron/git-repo-scan.cjs

@@ -0,0 +1,96 @@
+'use strict'
+
+// Repo-first discovery: walk bounded roots for git repos using only Node's `fs`
+// — no native addon, so it just works for anyone who pulls main (no
+// electron-rebuild). Mirrors how GitHub Desktop scans: stop at the first `.git`
+// (don't descend into a repo), cap depth, and skip heavy non-repo trees so the
+// first scan stays fast. Results are cached by the backend after the first run.
+
+const fs = require('node:fs')
+const os = require('node:os')
+const path = require('node:path')
+
+const fsp = fs.promises
+
+// Shallow on purpose: real projects live a few levels under home
+// (`~/www/repo`, `~/code/org/repo`); deeper `.git` dirs are almost always
+// fixtures/vendored/eval checkouts (e.g. `~/www/ha-evals/tasks/*/repo`). Repos
+// you actually use but keep deeper still surface via session-derived discovery,
+// so this only prunes noise, never repos with history.
+const DEFAULT_MAX_DEPTH = 3
+const MAX_CONCURRENCY = 32
+
+// Big trees that are never themselves repos and would waste the walk. Anything
+// hidden (dotdirs like .cache/.Trash/.npm) is skipped wholesale below, so this
+// only needs the non-hidden heavyweights.
+const JUNK_DIRS = new Set(['Applications', 'Library', 'node_modules', 'site-packages', 'vendor', 'venv'])
+
+async function mapLimit(items, limit, fn) {
+  let cursor = 0
+
+  async function worker() {
+    while (cursor < items.length) {
+      const index = cursor
+      cursor += 1
+      await fn(items[index])
+    }
+  }
+
+  await Promise.all(Array.from({ length: Math.min(limit, items.length) }, worker))
+}
+
+/**
+ * Scan `roots` (default: the home dir) for git repositories. Returns deduped
+ * `{ root, label }` entries. `options.maxDepth` caps recursion (default 3).
+ */
+async function scanGitRepos(roots, options = {}) {
+  const maxDepth = Number(options.maxDepth) || DEFAULT_MAX_DEPTH
+  const searchRoots = Array.isArray(roots) && roots.length > 0 ? roots : [os.homedir()]
+  const found = new Map()
+
+  async function walk(dir, depth) {
+    if (depth > maxDepth) {
+      return
+    }
+
+    let entries
+    try {
+      entries = await fsp.readdir(dir, { withFileTypes: true })
+    } catch {
+      return // unreadable / permission denied
+    }
+
+    // A `.git` DIRECTORY marks a real repo root (a main checkout). A `.git`
+    // FILE is a linked worktree or submodule — those belong to their parent
+    // repo as lanes, not as separate projects, so we don't list them (and we
+    // keep descending in case a real repo sits deeper). This is what kills the
+    // worktree/eval-repo duplicate explosion.
+    if (entries.some(entry => entry.name === '.git' && entry.isDirectory())) {
+      const root = dir.replace(/[/\\]+$/, '')
+      found.set(root, path.basename(root) || root)
+
+      return
+    }
+
+    const subdirs = []
+    for (const entry of entries) {
+      // Real directories only (skip symlinks to avoid loops), no hidden dirs, no
+      // known heavy trees.
+      if (!entry.isDirectory() || entry.name.startsWith('.') || JUNK_DIRS.has(entry.name)) {
+        continue
+      }
+
+      subdirs.push(path.join(dir, entry.name))
+    }
+
+    await mapLimit(subdirs, MAX_CONCURRENCY, sub => walk(sub, depth + 1))
+  }
+
+  await mapLimit(searchRoots.map(root => String(root || '').trim()).filter(Boolean), MAX_CONCURRENCY, root =>
+    walk(root, 0)
+  )
+
+  return [...found.entries()].map(([root, label]) => ({ label, root }))
+}
+
+module.exports = { scanGitRepos }

apps/desktop/electron/git-review-ops.cjs

@@ -0,0 +1,684 @@
+'use strict'
+
+// Git ops backing the coding rail + Codex-style review pane. Built on `simple-git`
+// (a maintained wrapper around the system git binary — same git the rest of the
+// app shells to, no native build) so we read structured status()/diffSummary()
+// results instead of hand-parsing porcelain. Reads degrade to null/empty on a
+// non-repo / remote backend; mutations reject so the renderer can toast.
+
+const { execFile } = require('node:child_process')
+const fs = require('node:fs/promises')
+const path = require('node:path')
+
+const simpleGit = require('simple-git')
+
+const { resolveRequestedPathForIpc } = require('./hardening.cjs')
+
+const COMMIT_CONTEXT_DIFF_MAX_CHARS = 120_000
+const COMMIT_CONTEXT_UNTRACKED_MAX = 80
+const UNTRACKED_LINE_COUNT_CONCURRENCY = 16
+const UNTRACKED_LINE_COUNT_MAX_BYTES = 1024 * 1024
+
+// GUI-launched Electron apps on macOS inherit only a minimal PATH (no
+// /opt/homebrew/bin or /usr/local/bin), so `gh` — and the `git` gh shells out
+// to — aren't found. Augment PATH with the resolved gh dir + the common
+// package-manager bins so gh runs the same way it does in a terminal.
+function ghEnv(ghBin) {
+  const extra = [ghBin ? path.dirname(ghBin) : '', '/opt/homebrew/bin', '/usr/local/bin', '/usr/bin'].filter(
+    dir => dir && dir !== '.'
+  )
+
+  return { ...process.env, PATH: [...extra, process.env.PATH].filter(Boolean).join(path.delimiter) }
+}
+
+// Run the `gh` CLI in a repo. Resolves { ok, stdout } so callers branch on
+// availability/auth without a throw. gh missing/unauthed → ok:false.
+function runGh(args, cwd, ghBin) {
+  return new Promise(resolve => {
+    execFile(
+      ghBin || 'gh',
+      args,
+      { cwd, env: ghEnv(ghBin), windowsHide: true, timeout: 30_000, maxBuffer: 8 * 1024 * 1024 },
+      (err, stdout) => resolve({ ok: !err, stdout: String(stdout || '') })
+    )
+  })
+}
+
+function gitFor(cwd, gitBin) {
+  return simpleGit({ baseDir: cwd, binary: gitBin || 'git', maxConcurrentProcesses: 4, trimmed: false })
+}
+
+// simple-git reports renames as `old => new` (and `dir/{old => new}/f`); resolve
+// to the NEW path so the row addresses the real file for diff/stage.
+function resolveRenamePath(raw) {
+  const path = String(raw || '').trim()
+
+  if (!path.includes(' => ')) {
+    return path
+  }
+
+  const brace = path.match(/^(.*)\{(.*) => (.*)\}(.*)$/)
+
+  if (brace) {
+    const [, prefix, , to, suffix] = brace
+
+    return `${prefix}${to}${suffix}`.replace(/\/{2,}/g, '/')
+  }
+
+  return path.split(' => ').pop().trim()
+}
+
+// DiffResult.files → Map<path, {added, removed}> (binary files carry no line
+// delta).
+function countsByPath(summary) {
+  const map = new Map()
+
+  for (const file of summary.files) {
+    map.set(resolveRenamePath(file.file), {
+      added: file.binary ? 0 : file.insertions,
+      removed: file.binary ? 0 : file.deletions
+    })
+  }
+
+  return map
+}
+
+// Untracked files don't appear in diffSummary(); count insertions from disk so
+// the review tree can show +N for new files (matches an all-add diff view).
+// Insertions = line count: newline bytes, plus one for a final unterminated
+// line. Binary (NUL byte) → 0, mirroring git numstat's "-".
+async function untrackedInsertions(cwd, relPath) {
+  try {
+    const fullPath = path.join(cwd, relPath)
+    const stat = await fs.stat(fullPath)
+
+    if (!stat.isFile() || stat.size > UNTRACKED_LINE_COUNT_MAX_BYTES) {
+      return 0
+    }
+
+    const buf = await fs.readFile(fullPath)
+
+    if (buf.includes(0)) {
+      return 0
+    }
+
+    let lines = 0
+
+    for (const byte of buf) {
+      if (byte === 10) {
+        lines++
+      }
+    }
+
+    return buf.length > 0 && buf[buf.length - 1] !== 10 ? lines + 1 : lines
+  } catch {
+    return 0
+  }
+}
+
+function capText(text, maxChars, label = 'truncated') {
+  const value = String(text || '')
+
+  if (value.length <= maxChars) {
+    return value
+  }
+
+  return `${value.slice(0, maxChars)}\n# ${label}: ${value.length - maxChars} chars omitted\n`
+}
+
+async function fillUntrackedCounts(cwd, files) {
+  const pending = files.filter(file => file.status === '?' && file.added === 0 && file.removed === 0)
+
+  for (let i = 0; i < pending.length; i += UNTRACKED_LINE_COUNT_CONCURRENCY) {
+    await Promise.all(
+      pending.slice(i, i + UNTRACKED_LINE_COUNT_CONCURRENCY).map(async file => {
+        file.added = await untrackedInsertions(cwd, file.path)
+      })
+    )
+  }
+}
+
+// Resolve the base ref for "all branch changes": merge-base with the remote
+// default branch (origin/HEAD), falling back to common trunk names.
+async function branchBase(git) {
+  const candidates = []
+
+  try {
+    const head = (await git.revparse(['--abbrev-ref', 'origin/HEAD'])).trim()
+
+    if (head) {
+      candidates.push(head)
+    }
+  } catch {
+    // No origin/HEAD configured.
+  }
+
+  candidates.push('origin/main', 'origin/master', 'main', 'master')
+
+  for (const ref of candidates) {
+    try {
+      const base = (await git.raw(['merge-base', 'HEAD', ref])).trim()
+
+      if (base) {
+        return base
+      }
+    } catch {
+      // Ref doesn't exist; try the next candidate.
+    }
+  }
+
+  return null
+}
+
+// Resolve the repo's default branch NAME ("main" / "master" / …), preferring
+// the remote's HEAD, then common local trunk names. Null when none is found
+// (e.g. a fresh repo with only a feature branch). Used to offer "branch off the
+// trunk" regardless of which branch you're currently on.
+async function defaultBranchName(git) {
+  try {
+    const head = (await git.revparse(['--abbrev-ref', 'origin/HEAD'])).trim()
+
+    // "origin/main" → "main"; skip the bare "origin/HEAD" placeholder.
+    if (head && head !== 'origin/HEAD') {
+      return head.replace(/^origin\//, '')
+    }
+  } catch {
+    // No origin/HEAD configured.
+  }
+
+  // Prefer a local trunk, then a remote-only one (returns the clean name either
+  // way) so "branch off main" works even before main is checked out locally.
+  for (const ref of [
+    'refs/heads/main',
+    'refs/heads/master',
+    'refs/remotes/origin/main',
+    'refs/remotes/origin/master'
+  ]) {
+    try {
+      await git.raw(['rev-parse', '--verify', '--quiet', ref])
+
+      return ref.replace(/^refs\/(?:heads|remotes\/origin)\//, '')
+    } catch {
+      // Ref doesn't exist; try the next candidate.
+    }
+  }
+
+  return null
+}
+
+// A status file's single-letter classification, preferring the staged (index)
+// code over the worktree code; untracked wins (simple-git marks both '?').
+function statusLetter(file) {
+  if (file.index === '?' || file.working_dir === '?') {
+    return '?'
+  }
+
+  const code = file.index && file.index !== ' ' ? file.index : file.working_dir
+
+  return (code || 'M').toUpperCase()
+}
+
+const isStaged = file => Boolean(file.index && file.index !== ' ' && file.index !== '?')
+
+async function reviewList(repoPath, scope, baseRef, gitBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review list' })
+  } catch {
+    return { files: [], base: null }
+  }
+
+  const git = gitFor(cwd, gitBin)
+
+  try {
+    if (scope === 'branch' || scope === 'lastTurn') {
+      const base = scope === 'branch' ? await branchBase(git) : baseRef
+
+      if (!base) {
+        return { files: [], base: null }
+      }
+
+      const range = scope === 'branch' ? `${base}...HEAD` : base
+      const summary = await git.diffSummary([range])
+      const files = summary.files.map(file => ({
+        path: resolveRenamePath(file.file),
+        added: file.binary ? 0 : file.insertions,
+        removed: file.binary ? 0 : file.deletions,
+        status: 'M',
+        staged: false
+      }))
+
+      // "Last turn" also surfaces files created since the baseline (untracked).
+      if (scope === 'lastTurn') {
+        const status = await git.status()
+
+        for (const path of status.not_added) {
+          if (!files.some(f => f.path === path)) {
+            files.push({ path, added: 0, removed: 0, status: '?', staged: false })
+          }
+        }
+      }
+
+      files.sort((a, b) => a.path.localeCompare(b.path))
+      await fillUntrackedCounts(cwd, files)
+
+      return { files, base }
+    }
+
+    // Default: uncommitted (staged + unstaged + untracked), one row per path.
+    const [status, staged, unstaged] = await Promise.all([
+      git.status(),
+      git.diffSummary(['--cached']),
+      git.diffSummary([])
+    ])
+    const stagedCounts = countsByPath(staged)
+    const unstagedCounts = countsByPath(unstaged)
+
+    const files = status.files.map(file => {
+      const filePath = resolveRenamePath(file.path)
+      const sc = stagedCounts.get(filePath) || { added: 0, removed: 0 }
+      const uc = unstagedCounts.get(filePath) || { added: 0, removed: 0 }
+
+      return {
+        path: filePath,
+        added: sc.added + uc.added,
+        removed: sc.removed + uc.removed,
+        status: statusLetter(file),
+        staged: isStaged(file)
+      }
+    })
+
+    files.sort((a, b) => a.path.localeCompare(b.path))
+    await fillUntrackedCounts(cwd, files)
+
+    return { files, base: null }
+  } catch {
+    return { files: [], base: null }
+  }
+}
+
+async function reviewDiff(repoPath, filePath, scope, baseRef, staged, gitBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review diff' })
+  } catch {
+    return ''
+  }
+
+  const git = gitFor(cwd, gitBin)
+  const safe = args => git.diff(args).catch(() => '')
+
+  if (scope === 'branch') {
+    const base = await branchBase(git)
+
+    return base ? safe([`${base}...HEAD`, '--', filePath]) : ''
+  }
+
+  if (scope === 'lastTurn') {
+    return baseRef ? safe([baseRef, '--', filePath]) : ''
+  }
+
+  if (staged) {
+    return safe(['--cached', '--', filePath])
+  }
+
+  const worktree = await safe(['--', filePath])
+
+  if (worktree.trim()) {
+    return worktree
+  }
+
+  // Untracked file: no worktree diff exists, so synthesize an all-add diff via
+  // --no-index (exits non-zero by design when files differ, so go around
+  // simple-git's reject-on-nonzero with a raw execFile).
+  return new Promise(resolve => {
+    execFile(
+      gitBin || 'git',
+      ['diff', '--no-index', '--', '/dev/null', filePath],
+      { cwd, windowsHide: true, timeout: 30_000, maxBuffer: 32 * 1024 * 1024 },
+      (_err, stdout) => resolve(String(stdout || ''))
+    )
+  })
+}
+
+// Working-tree-vs-HEAD diff for ONE file — the "what changed since the last
+// commit" view used by the file preview. Unlike reviewDiff this never synthesizes
+// a full-add for a clean tracked file (so a pristine file shows no diff); it only
+// all-adds a genuinely untracked file.
+async function fileDiffVsHead(repoPath, filePath, gitBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'File diff' })
+  } catch {
+    return ''
+  }
+
+  const git = gitFor(cwd, gitBin)
+  const head = await git.diff(['HEAD', '--', filePath]).catch(() => '')
+
+  if (head.trim()) {
+    return head
+  }
+
+  // No tracked changes vs HEAD. Only synthesize an all-add diff for a file git
+  // doesn't know yet; a clean tracked file must return empty.
+  const status = await git.raw(['status', '--porcelain', '--', filePath]).catch(() => '')
+
+  if (!status.trim().startsWith('??')) {
+    return ''
+  }
+
+  return new Promise(resolve => {
+    execFile(
+      gitBin || 'git',
+      ['diff', '--no-index', '--', '/dev/null', filePath],
+      { cwd, windowsHide: true, timeout: 30_000, maxBuffer: 32 * 1024 * 1024 },
+      (_err, stdout) => resolve(String(stdout || ''))
+    )
+  })
+}
+
+async function reviewStage(repoPath, filePath, gitBin) {
+  const cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review stage' })
+
+  await gitFor(cwd, gitBin).raw(filePath ? ['add', '--', filePath] : ['add', '-A'])
+
+  return { ok: true }
+}
+
+async function reviewUnstage(repoPath, filePath, gitBin) {
+  const cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review unstage' })
+
+  await gitFor(cwd, gitBin).raw(filePath ? ['reset', '-q', 'HEAD', '--', filePath] : ['reset', '-q', 'HEAD'])
+
+  return { ok: true }
+}
+
+// Discard changes back to the committed state. Destructive — the renderer
+// confirms first. Restores tracked files and removes untracked ones.
+async function reviewRevert(repoPath, filePath, gitBin) {
+  const cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review revert' })
+  const git = gitFor(cwd, gitBin)
+
+  if (filePath) {
+    await git.raw(['checkout', 'HEAD', '--', filePath]).catch(() => undefined)
+    await git.raw(['clean', '-fd', '--', filePath]).catch(() => undefined)
+  } else {
+    await git.raw(['checkout', 'HEAD', '--', '.']).catch(() => undefined)
+    await git.raw(['clean', '-fd']).catch(() => undefined)
+  }
+
+  return { ok: true }
+}
+
+// Resolve a ref to a commit sha (captures the turn baseline for "Last turn").
+async function reviewRevParse(repoPath, ref, gitBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review rev-parse' })
+  } catch {
+    return null
+  }
+
+  try {
+    return (await gitFor(cwd, gitBin).revparse([ref || 'HEAD'])).trim() || null
+  } catch {
+    return null
+  }
+}
+
+// Commit the working tree. Mirrors VS Code: if nothing is staged, stage
+// everything first ("commit all"), then commit. Optionally push afterward,
+// setting upstream on the first push.
+async function reviewCommit(repoPath, message, push, gitBin) {
+  const cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review commit' })
+  const git = gitFor(cwd, gitBin)
+  const status = await git.status()
+
+  if (status.staged.length === 0) {
+    await git.raw(['add', '-A'])
+  }
+
+  await git.commit(message)
+
+  if (push) {
+    const fresh = await git.status()
+
+    if (fresh.tracking) {
+      await git.push()
+    } else if (fresh.current) {
+      await git.raw(['push', '-u', 'origin', fresh.current])
+    }
+  }
+
+  return { ok: true }
+}
+
+// Gather the context the model needs to draft a commit message: the diff of
+// what *will* be committed (staged when anything is staged, else everything
+// vs HEAD — mirroring reviewCommit's "stage all when nothing staged" rule),
+// the names of untracked files (which carry no diff), and recent commit
+// subjects for style. Diff is capped so the payload stays bounded. Reads only.
+async function reviewCommitContext(repoPath, gitBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review commit context' })
+  } catch {
+    return { diff: '', recent: '' }
+  }
+
+  const git = gitFor(cwd, gitBin)
+  const safe = args => git.diff(args).catch(() => '')
+
+  let status
+  try {
+    status = await git.status()
+  } catch {
+    return { diff: '', recent: '' }
+  }
+
+  // What will land: staged changes if any, otherwise all tracked changes vs HEAD.
+  let diff = capText(
+    status.staged.length > 0 ? await safe(['--cached']) : await safe(['HEAD']),
+    COMMIT_CONTEXT_DIFF_MAX_CHARS,
+    'diff truncated for commit-message generation'
+  )
+
+  // Untracked files have no diff — list them so new files aren't invisible.
+  const untracked = status.not_added || []
+  if (untracked.length > 0) {
+    const visible = untracked.slice(0, COMMIT_CONTEXT_UNTRACKED_MAX)
+    const omitted = untracked.length - visible.length
+    const note =
+      `\n# New (untracked) files:\n${visible.map(p => `#   ${p}`).join('\n')}\n` +
+      (omitted > 0 ? `#   ... ${omitted} more omitted\n` : '')
+
+    diff = diff ? `${diff}${note}` : note
+  }
+
+  const recent = await git.raw(['log', '-n', '10', '--pretty=format:%s']).catch(() => '')
+
+  return { diff: diff || '', recent: String(recent || '').trim() }
+}
+
+async function reviewPush(repoPath, gitBin) {
+  const cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review push' })
+  const git = gitFor(cwd, gitBin)
+  const status = await git.status()
+
+  if (status.tracking) {
+    await git.push()
+  } else if (status.current) {
+    await git.raw(['push', '-u', 'origin', status.current])
+  }
+
+  return { ok: true }
+}
+
+// gh availability + auth + whether this branch already has a PR. Reads only;
+// drives the PR button's enabled/label state. `ghReady` is false when gh is
+// missing OR not authenticated — either way the PR action can't run.
+async function reviewShipInfo(repoPath, ghBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review ship info' })
+  } catch {
+    return { ghReady: false, pr: null }
+  }
+
+  const auth = await runGh(['auth', 'status'], cwd, ghBin)
+
+  if (!auth.ok) {
+    return { ghReady: false, pr: null }
+  }
+
+  const view = await runGh(['pr', 'view', '--json', 'url,state,number'], cwd, ghBin)
+
+  if (!view.ok) {
+    // gh exits non-zero when no PR exists for the branch — that's not an error.
+    return { ghReady: true, pr: null }
+  }
+
+  try {
+    const pr = JSON.parse(view.stdout)
+
+    return { ghReady: true, pr: pr && pr.url ? { url: pr.url, state: pr.state, number: pr.number } : null }
+  } catch {
+    return { ghReady: true, pr: null }
+  }
+}
+
+// Create a PR for the current branch (pushing first so gh has a remote ref),
+// letting gh fill title/body from the commits. Returns the new PR url.
+async function reviewCreatePr(repoPath, gitBin, ghBin) {
+  const cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review create PR' })
+
+  await reviewPush(repoPath, gitBin).catch(() => undefined)
+
+  const created = await runGh(['pr', 'create', '--fill'], cwd, ghBin)
+
+  if (!created.ok) {
+    throw new Error('gh pr create failed (is gh installed and authenticated?)')
+  }
+
+  const url = created.stdout.trim().split('\n').filter(Boolean).pop() || ''
+
+  return { url }
+}
+
+// Compact working-tree status for the composer coding rail: branch, ahead/behind,
+// per-state change counts, +/- vs HEAD, and a capped changed-file list.
+async function repoStatus(repoPath, gitBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Repo status' })
+  } catch {
+    return null
+  }
+
+  // Session cwds can point at a deleted worktree for a moment (or forever in a
+  // stale row). simple-git throws at construction time on a missing baseDir, so
+  // fail soft and hide the coding rail instead of spamming IPC handler errors.
+  try {
+    const stat = await fs.stat(cwd)
+    if (!stat.isDirectory()) {
+      return null
+    }
+  } catch {
+    return null
+  }
+
+  let git
+  try {
+    git = gitFor(cwd, gitBin)
+  } catch {
+    return null
+  }
+  let status
+
+  try {
+    status = await git.status()
+  } catch {
+    // Not a repo / git unavailable / remote backend.
+    return null
+  }
+
+  const detached = typeof status.detached === 'boolean' ? status.detached : !status.current
+  const files = status.files.map(file => ({
+    path: file.path,
+    staged: isStaged(file),
+    unstaged: Boolean(file.working_dir && file.working_dir !== ' ' && file.working_dir !== '?'),
+    untracked: file.index === '?' || file.working_dir === '?',
+    conflicted: file.index === 'U' || file.working_dir === 'U'
+  }))
+
+  const result = {
+    branch: detached ? null : status.current || null,
+    defaultBranch: await defaultBranchName(git),
+    detached,
+    ahead: status.ahead || 0,
+    behind: status.behind || 0,
+    staged: files.filter(f => f.staged).length,
+    unstaged: files.filter(f => f.unstaged).length,
+    untracked: status.not_added.length,
+    conflicted: status.conflicted.length,
+    changed: files.length,
+    added: 0,
+    removed: 0,
+    files: files.slice(0, 200)
+  }
+
+  // +/- vs HEAD (staged + unstaged tracked changes). No HEAD yet → leave 0.
+  try {
+    const summary = await git.diffSummary(['HEAD'])
+    result.added = summary.insertions
+    result.removed = summary.deletions
+  } catch {
+    // No commits yet.
+  }
+
+  // `git diff HEAD` ignores untracked files, so a turn that only creates new
+  // files (the common case — a fresh module, a demo dir) showed +0 in the rail
+  // while the review pane counted them. Fold untracked insertions into `added`
+  // so the rail matches reality. Bounded (size cap + concurrency) like the
+  // review tree; only the capped file slice is counted so a huge untracked tree
+  // can't stall the probe.
+  try {
+    const untracked = status.not_added.slice(0, 500)
+    for (let i = 0; i < untracked.length; i += UNTRACKED_LINE_COUNT_CONCURRENCY) {
+      const batch = await Promise.all(
+        untracked.slice(i, i + UNTRACKED_LINE_COUNT_CONCURRENCY).map(path => untrackedInsertions(cwd, path))
+      )
+      result.added += batch.reduce((sum, n) => sum + n, 0)
+    }
+  } catch {
+    // Best-effort: a probe failure just leaves untracked lines uncounted.
+  }
+
+  return result
+}
+
+module.exports = {
+  branchBase,
+  fileDiffVsHead,
+  repoStatus,
+  resolveRenamePath,
+  reviewCommit,
+  reviewCommitContext,
+  reviewCreatePr,
+  reviewDiff,
+  reviewList,
+  reviewPush,
+  reviewRevParse,
+  reviewRevert,
+  reviewShipInfo,
+  reviewStage,
+  reviewUnstage
+}

apps/desktop/electron/git-review-ops.test.cjs

@@ -0,0 +1,22 @@
+'use strict'
+
+const assert = require('node:assert/strict')
+const test = require('node:test')
+
+const { resolveRenamePath } = require('./git-review-ops.cjs')
+
+test('resolveRenamePath: plain path is unchanged', () => {
+  assert.equal(resolveRenamePath('src/a.ts'), 'src/a.ts')
+})
+
+test('resolveRenamePath: simple rename resolves to the new path', () => {
+  assert.equal(resolveRenamePath('old.ts => new.ts'), 'new.ts')
+})
+
+test('resolveRenamePath: brace rename resolves to the new path', () => {
+  assert.equal(resolveRenamePath('src/{old => new}/file.ts'), 'src/new/file.ts')
+})
+
+test('resolveRenamePath: brace rename collapsing a segment', () => {
+  assert.equal(resolveRenamePath('src/{lib => }/file.ts'), 'src/file.ts')
+})

apps/desktop/electron/git-worktree-ops.cjs

@@ -0,0 +1,350 @@
+'use strict'
+
+// Git-driven worktree operations for the desktop "Start work" flow: spin up a
+// fresh worktree the lightest way (`git worktree add -b`), list real worktrees,
+// and remove them. Git is the source of truth; the renderer just drives these.
+
+const path = require('node:path')
+const fs = require('node:fs')
+const { execFile } = require('node:child_process')
+
+const { resolveRequestedPathForIpc } = require('./hardening.cjs')
+
+function runGit(gitBin, args, cwd) {
+  return new Promise((resolve, reject) => {
+    execFile(
+      gitBin,
+      args,
+      { cwd, windowsHide: true, timeout: 30_000, maxBuffer: 8 * 1024 * 1024 },
+      (err, stdout, stderr) => {
+        if (err) {
+          err.stderr = String(stderr || '')
+          reject(err)
+
+          return
+        }
+
+        resolve(String(stdout || ''))
+      }
+    )
+  })
+}
+
+// Parse `git worktree list --porcelain`. The first record is the main worktree.
+function parseWorktrees(out) {
+  const trees = []
+  let cur = null
+
+  for (const line of out.split('\n')) {
+    if (line.startsWith('worktree ')) {
+      if (cur) {
+        trees.push(cur)
+      }
+
+      cur = { path: line.slice(9).trim(), branch: null, detached: false, bare: false, locked: false }
+    } else if (!cur) {
+      continue
+    } else if (line.startsWith('branch ')) {
+      cur.branch = line
+        .slice(7)
+        .trim()
+        .replace(/^refs\/heads\//, '')
+    } else if (line === 'detached') {
+      cur.detached = true
+    } else if (line === 'bare') {
+      cur.bare = true
+    } else if (line.startsWith('locked')) {
+      cur.locked = true
+    }
+  }
+
+  if (cur) {
+    trees.push(cur)
+  }
+
+  return trees
+}
+
+async function listWorktrees(repoPath, gitBin) {
+  let resolved
+
+  try {
+    resolved = resolveRequestedPathForIpc(repoPath, { purpose: 'Worktree list' })
+  } catch {
+    return []
+  }
+
+  try {
+    const out = await runGit(gitBin, ['worktree', 'list', '--porcelain'], resolved)
+
+    return parseWorktrees(out).map((tree, index) => ({
+      path: tree.path,
+      branch: tree.branch,
+      isMain: index === 0,
+      detached: tree.detached,
+      locked: tree.locked
+    }))
+  } catch {
+    return []
+  }
+}
+
+// A git-ref-safe branch name (spaces → "-", drop forbidden chars, trim edges),
+// or "" when nothing usable remains. Mirrors the renderer's `gitRef`, so a bad
+// value can't reach `git` no matter the caller (the GUI also enforces live).
+function sanitizeBranch(name) {
+  return String(name || '')
+    .replace(/\s+/g, '-')
+    .replace(/[^\w./-]/g, '')
+    .replace(/-{2,}/g, '-')
+    .replace(/\/{2,}/g, '/')
+    .replace(/\.{2,}/g, '.')
+    .replace(/^[-./]+|[-./]+$/g, '')
+}
+
+function slugify(name) {
+  const slug = String(name || '')
+    .trim()
+    .toLowerCase()
+    .replace(/[^a-z0-9]+/g, '-')
+    .replace(/^-+|-+$/g, '')
+    .slice(0, 40)
+    .replace(/-+$/g, '')
+
+  return slug || 'work'
+}
+
+const TRUNK_BRANCHES = ['main', 'master']
+
+async function gitLine(gitBin, args, cwd) {
+  try {
+    return (await runGit(gitBin, args, cwd)).trim()
+  } catch {
+    return ''
+  }
+}
+
+async function defaultBranch(gitBin, cwd) {
+  const remote = (
+    await gitLine(gitBin, ['symbolic-ref', '--quiet', '--short', 'refs/remotes/origin/HEAD'], cwd)
+  ).replace(/^origin\//, '')
+
+  if (remote) {
+    return remote
+  }
+
+  const configured = await gitLine(gitBin, ['config', '--get', 'init.defaultBranch'], cwd)
+
+  if (configured) {
+    return configured
+  }
+
+  for (const branch of TRUNK_BRANCHES) {
+    if (await gitLine(gitBin, ['show-ref', '--verify', `refs/heads/${branch}`], cwd)) {
+      return branch
+    }
+  }
+
+  return ''
+}
+
+// A brand-new project folder isn't a git repo — and a freshly-init'd one has no
+// commit to branch from — so `git worktree add` would fail. Make the dir a repo
+// with a root commit on the user's behalf so worktrees "just work". No-op for a
+// repo that already has commits; never touches the user's files (the seed commit
+// is `--allow-empty`), and never inits a dir that already lives inside a repo.
+async function ensureGitRepo(gitBin, dir) {
+  let needsRoot = false
+
+  try {
+    const inside = (await runGit(gitBin, ['rev-parse', '--is-inside-work-tree'], dir)).trim()
+
+    if (inside !== 'true') {
+      await runGit(gitBin, ['init'], dir)
+      needsRoot = true
+    } else {
+      // Repo exists; a worktree still needs a HEAD to branch from.
+      try {
+        await runGit(gitBin, ['rev-parse', '--verify', 'HEAD'], dir)
+      } catch {
+        needsRoot = true
+      }
+    }
+  } catch {
+    await runGit(gitBin, ['init'], dir)
+    needsRoot = true
+  }
+
+  if (needsRoot) {
+    // Inline identity so the seed commit lands even with no global git config.
+    await runGit(
+      gitBin,
+      [
+        '-c',
+        'user.email=hermes@localhost',
+        '-c',
+        'user.name=Hermes',
+        'commit',
+        '--allow-empty',
+        '-m',
+        'Initial commit'
+      ],
+      dir
+    )
+  }
+}
+
+// Resolve the repo's MAIN worktree root, so `.worktrees/` always nests under the
+// primary checkout even when called from a linked worktree.
+async function mainRoot(gitBin, cwd) {
+  const list = await listWorktrees(cwd, gitBin)
+  const main = list.find(tree => tree.isMain)
+
+  return main ? main.path : cwd
+}
+
+function uniqueDir(base) {
+  let dir = base
+  let n = 1
+
+  while (fs.existsSync(dir)) {
+    n += 1
+    dir = `${base}-${n}`
+  }
+
+  return dir
+}
+
+async function addExistingBranchWorktree(gitBin, root, name) {
+  const branch = sanitizeBranch(name)
+
+  if (!branch) {
+    throw new Error('Branch name is required.')
+  }
+
+  if (branch === (await defaultBranch(gitBin, root))) {
+    await runGit(gitBin, ['switch', branch], root)
+
+    return { path: root, branch, repoRoot: root }
+  }
+
+  const dir = uniqueDir(path.join(root, '.worktrees', slugify(branch)))
+  await runGit(gitBin, ['worktree', 'add', dir, branch], root)
+
+  return { path: dir, branch, repoRoot: root }
+}
+
+async function addWorktree(repoPath, options, gitBin) {
+  const resolved = resolveRequestedPathForIpc(repoPath, { purpose: 'Worktree add' })
+  // A new project's folder may not be a git repo yet — init it (with a root
+  // commit) so the worktree has something to branch from.
+  await ensureGitRepo(gitBin, resolved)
+  const root = await mainRoot(gitBin, resolved)
+  const opts = options || {}
+
+  if (opts.existingBranch) {
+    return addExistingBranchWorktree(gitBin, root, opts.existingBranch)
+  }
+
+  const slug = slugify(opts.name || `work-${Date.now().toString(36)}`)
+  const branch = sanitizeBranch(opts.branch) || `hermes/${slug}`
+  const dir = uniqueDir(path.join(root, '.worktrees', slug))
+
+  const args = ['worktree', 'add', '-b', branch, dir]
+
+  if (opts.base) {
+    args.push(String(opts.base))
+  }
+
+  try {
+    await runGit(gitBin, args, root)
+  } catch (err) {
+    // Branch name may already exist — retry checking out the existing branch
+    // into a fresh worktree dir instead of failing the whole flow.
+    if (/already exists/i.test(err.stderr || '')) {
+      await runGit(gitBin, ['worktree', 'add', dir, branch], root)
+    } else {
+      throw err
+    }
+  }
+
+  return { path: dir, branch, repoRoot: root }
+}
+
+async function removeWorktree(repoPath, worktreePath, options, gitBin) {
+  const resolvedRepo = resolveRequestedPathForIpc(repoPath, { purpose: 'Worktree remove (repo)' })
+  const resolvedTree = resolveRequestedPathForIpc(worktreePath, { purpose: 'Worktree remove (tree)' })
+  const root = await mainRoot(gitBin, resolvedRepo)
+  const args = ['worktree', 'remove']
+
+  if (options && options.force) {
+    args.push('--force')
+  }
+
+  args.push(resolvedTree)
+  await runGit(gitBin, args, root)
+
+  return { removed: resolvedTree }
+}
+
+// List local branches for the "convert a branch into a worktree" picker, most
+// recently committed first. Each carries whether it's already checked out in a
+// worktree and, when checked out, that worktree's path. Empty on a non-repo /
+// remote backend where the probe can't run.
+async function listBranches(repoPath, gitBin) {
+  let resolved
+
+  try {
+    resolved = resolveRequestedPathForIpc(repoPath, { purpose: 'Branch list' })
+  } catch {
+    return []
+  }
+
+  try {
+    const out = await runGit(
+      gitBin,
+      ['for-each-ref', '--format=%(refname:short)', '--sort=-committerdate', 'refs/heads'],
+      resolved
+    )
+    const trees = await listWorktrees(resolved, gitBin)
+    const pathByBranch = new Map(trees.filter(tree => tree.branch).map(tree => [tree.branch, tree.path]))
+    const trunk = await defaultBranch(gitBin, resolved)
+
+    return out
+      .split('\n')
+      .map(line => line.trim())
+      .filter(Boolean)
+      .map(name => ({
+        name,
+        checkedOut: pathByBranch.has(name),
+        isDefault: Boolean(trunk && name === trunk),
+        worktreePath: pathByBranch.get(name) || null
+      }))
+  } catch {
+    return []
+  }
+}
+
+async function switchBranch(repoPath, branch, gitBin) {
+  const resolved = resolveRequestedPathForIpc(repoPath, { purpose: 'Branch switch' })
+  const target = sanitizeBranch(branch)
+
+  if (!target) {
+    throw new Error('Branch name is required.')
+  }
+
+  await runGit(gitBin, ['switch', target], resolved)
+
+  return { branch: target }
+}
+
+module.exports = {
+  addWorktree,
+  ensureGitRepo,
+  listBranches,
+  listWorktrees,
+  parseWorktrees,
+  removeWorktree,
+  sanitizeBranch,
+  switchBranch
+}

apps/desktop/electron/git-worktree-ops.test.cjs

@@ -0,0 +1,214 @@
+'use strict'
+
+const assert = require('node:assert/strict')
+const { execFileSync } = require('node:child_process')
+const fs = require('node:fs')
+const os = require('node:os')
+const path = require('node:path')
+const test = require('node:test')
+
+const {
+  addWorktree,
+  ensureGitRepo,
+  listBranches,
+  parseWorktrees,
+  sanitizeBranch,
+  switchBranch
+} = require('./git-worktree-ops.cjs')
+
+test('sanitizeBranch: spaces → hyphens, forbidden chars dropped, edges trimmed', () => {
+  assert.equal(sanitizeBranch('beach vibes'), 'beach-vibes')
+  assert.equal(sanitizeBranch('feat/cool thing'), 'feat/cool-thing')
+  assert.equal(sanitizeBranch('  wip~^:? '), 'wip')
+  assert.equal(sanitizeBranch('///'), '')
+})
+
+test('parseWorktrees: main checkout + linked worktree', () => {
+  const out = [
+    'worktree /repo',
+    'HEAD abc123',
+    'branch refs/heads/main',
+    '',
+    'worktree /repo/.worktrees/feat',
+    'HEAD def456',
+    'branch refs/heads/hermes/feat',
+    ''
+  ].join('\n')
+
+  const trees = parseWorktrees(out)
+
+  assert.equal(trees.length, 2)
+  assert.equal(trees[0].path, '/repo')
+  assert.equal(trees[0].branch, 'main')
+  assert.equal(trees[1].path, '/repo/.worktrees/feat')
+  assert.equal(trees[1].branch, 'hermes/feat')
+})
+
+test('parseWorktrees: detached + locked flags', () => {
+  const out = ['worktree /repo/wt', 'HEAD abc', 'detached', 'locked reason', ''].join('\n')
+  const trees = parseWorktrees(out)
+
+  assert.equal(trees.length, 1)
+  assert.equal(trees[0].detached, true)
+  assert.equal(trees[0].locked, true)
+  assert.equal(trees[0].branch, null)
+})
+
+test('parseWorktrees: empty input', () => {
+  assert.deepEqual(parseWorktrees(''), [])
+})
+
+test('ensureGitRepo: inits a plain dir with a root commit so worktrees branch', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-wt-'))
+  const git = (...args) => execFileSync('git', args, { cwd: dir }).toString().trim()
+
+  try {
+    await ensureGitRepo('git', dir)
+    assert.match(git('rev-parse', '--verify', 'HEAD'), /^[0-9a-f]{7,}$/)
+
+    // The whole point: a worktree can now branch off the seeded root commit.
+    execFileSync('git', ['worktree', 'add', '-b', 'wt', path.join(dir, '.worktrees', 'wt')], { cwd: dir })
+    assert.ok(fs.existsSync(path.join(dir, '.worktrees', 'wt')))
+
+    // Idempotent: an already-committed repo gets no extra commit.
+    await ensureGitRepo('git', dir)
+    assert.equal(git('rev-list', '--count', 'HEAD'), '1')
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('switchBranch: switches a normal checkout branch', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-switch-'))
+  const git = (...args) => execFileSync('git', args, { cwd: dir }).toString().trim()
+
+  try {
+    await ensureGitRepo('git', dir)
+    execFileSync('git', ['branch', 'feature'], { cwd: dir })
+
+    await switchBranch(dir, 'feature', 'git')
+
+    assert.equal(git('branch', '--show-current'), 'feature')
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('listBranches: lists locals and flags the checked-out branch', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-branches-'))
+
+  try {
+    await ensureGitRepo('git', dir)
+    const current = execFileSync('git', ['branch', '--show-current'], { cwd: dir }).toString().trim()
+    execFileSync('git', ['branch', 'feature'], { cwd: dir })
+
+    const branches = await listBranches(dir, 'git')
+    const names = branches.map(b => b.name).sort()
+
+    assert.deepEqual(names, [current, 'feature'].sort())
+    // The repo's own checkout is flagged; the unused branch is convertible.
+    assert.equal(branches.find(b => b.name === current).checkedOut, true)
+    assert.equal(branches.find(b => b.name === current).isDefault, true)
+    assert.equal(fs.realpathSync(branches.find(b => b.name === current).worktreePath), fs.realpathSync(dir))
+    assert.equal(branches.find(b => b.name === 'feature').checkedOut, false)
+    assert.equal(branches.find(b => b.name === 'feature').isDefault, false)
+    assert.equal(branches.find(b => b.name === 'feature').worktreePath, null)
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('listBranches: flags a free default branch as default, not checked out', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-branches-default-'))
+  const git = (...args) => execFileSync('git', args, { cwd: dir }).toString().trim()
+
+  try {
+    await ensureGitRepo('git', dir)
+    const trunk = git('branch', '--show-current')
+    execFileSync('git', ['switch', '-c', 'rawr'], { cwd: dir })
+
+    const branches = await listBranches(dir, 'git')
+    const defaultBranch = branches.find(b => b.name === trunk)
+
+    assert.equal(defaultBranch.checkedOut, false)
+    assert.equal(defaultBranch.isDefault, true)
+    assert.equal(defaultBranch.worktreePath, null)
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('listBranches: a branch claimed by a worktree is flagged checked out', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-branches-wt-'))
+
+  try {
+    await ensureGitRepo('git', dir)
+    execFileSync('git', ['branch', 'feature'], { cwd: dir })
+    // addWorktree converts the existing "feature" branch into a worktree.
+    const result = await addWorktree(dir, { existingBranch: 'feature' }, 'git')
+
+    assert.equal(result.branch, 'feature')
+    assert.ok(fs.existsSync(result.path))
+
+    const branches = await listBranches(dir, 'git')
+
+    assert.equal(branches.find(b => b.name === 'feature').checkedOut, true)
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('listBranches: empty on a non-repo path', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-nonrepo-'))
+
+  try {
+    assert.deepEqual(await listBranches(dir, 'git'), [])
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('addWorktree: existingBranch checks the branch out without a new branch', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-convert-'))
+  const git = (...args) => execFileSync('git', args, { cwd: dir }).toString().trim()
+
+  try {
+    await ensureGitRepo('git', dir)
+    execFileSync('git', ['branch', 'cool/feature'], { cwd: dir })
+
+    const before = git('branch', '--list').split('\n').length
+    const result = await addWorktree(dir, { existingBranch: 'cool/feature' }, 'git')
+
+    // No new branch was created — only the existing one is checked out.
+    assert.equal(git('branch', '--list').split('\n').length, before)
+    assert.equal(result.branch, 'cool/feature')
+    // Dir is named off the branch slug, nested under the main repo's .worktrees.
+    assert.match(result.path, /[/\\]\.worktrees[/\\]cool-feature/)
+    assert.equal(
+      execFileSync('git', ['branch', '--show-current'], { cwd: result.path }).toString().trim(),
+      'cool/feature'
+    )
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('addWorktree: existing default branch switches the main checkout, not .worktrees/main', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-convert-default-'))
+  const git = (...args) => execFileSync('git', args, { cwd: dir }).toString().trim()
+
+  try {
+    await ensureGitRepo('git', dir)
+    const trunk = git('branch', '--show-current')
+    execFileSync('git', ['switch', '-c', 'rawr'], { cwd: dir })
+
+    const result = await addWorktree(dir, { existingBranch: trunk }, 'git')
+
+    assert.equal(result.branch, trunk)
+    assert.equal(fs.realpathSync(result.path), fs.realpathSync(dir))
+    assert.equal(git('branch', '--show-current'), trunk)
+    assert.equal(fs.existsSync(path.join(dir, '.worktrees', trunk)), false)
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})

apps/desktop/electron/git-worktrees.cjs

@@ -1,174 +0,0 @@
-'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
-}

apps/desktop/electron/hardening.cjs

@@ -186,7 +186,10 @@ async function statForIpc(fsImpl, resolvedPath, purpose, typeLabel) {
     if (code === 'ENOENT' || code === 'ENOTDIR') {
       throw ipcPathError(code || 'ENOENT', `${purpose} failed: ${typeLabel} does not exist.`)
     }
-    throw ipcPathError(code || 'read-error', `${purpose} failed: ${error instanceof Error ? error.message : String(error)}`)
+    throw ipcPathError(
+      code || 'read-error',
+      `${purpose} failed: ${error instanceof Error ? error.message : String(error)}`
+    )
   }
 }
 
@@ -201,7 +204,10 @@ async function realpathForIpc(fsImpl, resolvedPath, purpose) {
     return realPath
   } catch (error) {
     const code = error && typeof error === 'object' ? error.code : ''
-    throw ipcPathError(code || 'read-error', `${purpose} failed: ${error instanceof Error ? error.message : String(error)}`)
+    throw ipcPathError(
+      code || 'read-error',
+      `${purpose} failed: ${error instanceof Error ? error.message : String(error)}`
+    )
   }
 }
 

apps/desktop/electron/main.cjs

@@ -21,10 +21,10 @@ const crypto = require('node:crypto')
 const fs = require('node:fs')
 const http = require('node:http')
 const https = require('node:https')
-const net = require('node:net')
 const path = require('node:path')
 const { pathToFileURL } = require('node:url')
 const { execFileSync, spawn } = require('node:child_process')
+const { installEmbedReferer } = require('./embed-referer.cjs')
 const { detectRemoteDisplay, isWindowsBinaryPathInWsl, isWslEnvironment } = require('./bootstrap-platform.cjs')
 const { runBootstrap } = require('./bootstrap-runner.cjs')
 const {
@@ -38,11 +38,13 @@ const { canImportHermesCli, verifyHermesCli } = require('./backend-probes.cjs')
 const { createLinkTitleWindow } = require('./link-title-window.cjs')
 const { probeGatewayWebSocket } = require('./gateway-ws-probe.cjs')
 const { adoptServedDashboardToken } = require('./dashboard-token.cjs')
-const { waitForDashboardPort } = require('./backend-ready.cjs')
+const { waitForDashboardPortAnnouncement } = require('./backend-ready.cjs')
 const { serializeJsonBody, setJsonRequestHeaders } = require('./oauth-net-request.cjs')
 const { fetchMarketplaceThemes, searchMarketplaceThemes } = require('./vscode-marketplace.cjs')
 const { buildDesktopBackendEnv, normalizeHermesHomeRoot } = require('./backend-env.cjs')
 const { readWindowsUserEnvVar } = require('./windows-user-env.cjs')
+const { readWslWindowsClipboardImage } = require('./wsl-clipboard-image.cjs')
+const { nativeOverlayWidth: computeNativeOverlayWidth } = require('./titlebar-overlay-width.cjs')
 const { readDirForIpc } = require('./fs-read-dir.cjs')
 const { readLiveUpdateMarker } = require('./update-marker.cjs')
 const {
@@ -55,7 +57,23 @@ const {
   buildRelaunchScript
 } = require('./update-relaunch.cjs')
 const { gitRootForIpc } = require('./git-root.cjs')
-const { worktreesForIpc } = require('./git-worktrees.cjs')
+const { addWorktree, listBranches, listWorktrees, removeWorktree, switchBranch } = require('./git-worktree-ops.cjs')
+const {
+  fileDiffVsHead,
+  repoStatus,
+  reviewCommit,
+  reviewCommitContext,
+  reviewCreatePr,
+  reviewDiff,
+  reviewList,
+  reviewPush,
+  reviewRevParse,
+  reviewRevert,
+  reviewShipInfo,
+  reviewStage,
+  reviewUnstage
+} = require('./git-review-ops.cjs')
+const { scanGitRepos } = require('./git-repo-scan.cjs')
 const { OFFICIAL_REPO_HTTPS_URL, isOfficialSshRemote } = require('./update-remote.cjs')
 const { resolveBehindCount, shouldCountCommits } = require('./update-count.cjs')
 const { runRebuildWithRetry } = require('./update-rebuild.cjs')
@@ -170,6 +188,16 @@ if (REMOTE_DISPLAY_REASON) {
   )
 }
 
+// WSLg: Chromium blocklists the Mesa vGPU → software compositing → typing lag.
+// /dev/dxg means a real GPU is available; un-blocklist it. Skipped when a remote
+// display already forced software (SSH'd-into-WSL).
+if (IS_WSL && !REMOTE_DISPLAY_REASON && fs.existsSync('/dev/dxg')) {
+  app.commandLine.appendSwitch('ignore-gpu-blocklist')
+  app.commandLine.appendSwitch('enable-gpu-rasterization')
+  app.commandLine.appendSwitch('enable-zero-copy')
+  console.log('[hermes] WSL GPU passthrough (/dev/dxg) detected; enabling GPU acceleration')
+}
+
 ipcMain.handle('hermes:get-remote-display-reason', () => REMOTE_DISPLAY_REASON)
 
 // Keep the renderer running at full speed while the window is in the background
@@ -302,9 +330,7 @@ function hermesManagedNodePathEntries() {
 }
 
 function pathWithHermesManagedNode(...entries) {
-  return [...hermesManagedNodePathEntries(), ...entries, process.env.PATH]
-    .filter(Boolean)
-    .join(path.delimiter)
+  return [...hermesManagedNodePathEntries(), ...entries, process.env.PATH].filter(Boolean).join(path.delimiter)
 }
 
 // ACTIVE_HERMES_ROOT — the canonical mutable Hermes install. Same path
@@ -382,14 +408,10 @@ const WINDOW_BUTTON_POSITION = {
   x: 24,
   y: TITLEBAR_HEIGHT / 2 - MACOS_TRAFFIC_LIGHTS_HEIGHT / 2
 }
-// Width Electron reserves for the Windows/Linux native min/max/close cluster
-// when `titleBarOverlay` is enabled. The OS paints these buttons in the
-// top-right corner of the renderer; we have to leave that much room on the
-// right edge so our system tools (file browser, haptics, settings) don't sit
-// underneath them. macOS uses left-side traffic lights instead and reports a
-// position via getWindowButtonPosition(), so this width is non-zero only on
-// non-macOS platforms.
-const NATIVE_OVERLAY_BUTTON_WIDTH = 144
+// Right-edge window-control reservation lives in titlebar-overlay-width.cjs
+// (pure + unit-testable); computeNativeOverlayWidth() applies it per platform.
+// It's only the pre-layout fallback — the renderer measures the exact overlay
+// width live via the Window Controls Overlay API.
 const APP_ICON_PATHS = [
   path.join(APP_ROOT, 'public', 'apple-touch-icon.png'),
   path.join(APP_ROOT, 'dist', 'apple-touch-icon.png'),
@@ -503,25 +525,48 @@ function getWindowBackgroundColor() {
   return nativeTheme.shouldUseDarkColors ? '#111111' : '#f7f7f7'
 }
 
+// Transparent WCO — renderer chrome shows through. rgba(0,0,0,0) can fall back
+// to GetFrameColor() on some Electron builds; rgba(1,0,0,0) is the escape hatch.
+const TITLEBAR_OVERLAY_COLOR = 'rgba(1, 0, 0, 0)'
+
 function getTitleBarOverlayOptions() {
   if (IS_MAC) {
     return { height: TITLEBAR_HEIGHT }
   }
 
-  if (rendererTitleBarTheme) {
-    return {
-      color: rendererTitleBarTheme.background,
-      height: TITLEBAR_HEIGHT,
-      symbolColor: rendererTitleBarTheme.foreground
-    }
+  // Windows + WSLg paint WCO natively; plain Linux disables it (frameless hidden
+  // titlebar still applies).
+  if (!IS_WINDOWS && !IS_WSL) {
+    return false
   }
 
-  const useDarkColors = nativeTheme.shouldUseDarkColors
-
   return {
-    color: useDarkColors ? '#111111' : '#f7f7f7',
+    color: TITLEBAR_OVERLAY_COLOR,
     height: TITLEBAR_HEIGHT,
-    symbolColor: useDarkColors ? '#f7f7f7' : '#242424'
+    symbolColor:
+      rendererTitleBarTheme && isHexColor(rendererTitleBarTheme.foreground)
+        ? rendererTitleBarTheme.foreground
+        : nativeTheme.shouldUseDarkColors
+          ? '#f7f7f7'
+          : '#242424'
+  }
+}
+
+// Push refreshed overlay options to a live window after a theme/appearance
+// change. No-op only on plain (non-WSL) Linux, where getTitleBarOverlayOptions()
+// returns false; the try/catch additionally guards builds where
+// setTitleBarOverlay isn't supported.
+function applyTitleBarOverlay(win) {
+  const options = getTitleBarOverlayOptions()
+  if (!options || typeof options !== 'object') {
+    return
+  }
+
+  try {
+    win?.setTitleBarOverlay?.(options)
+  } catch {
+    // Overlay not supported on this platform/build — leave the frameless
+    // titlebar as-is.
   }
 }
 
@@ -744,6 +789,9 @@ let rendererReloadTimes = []
 // instead of re-running install.ps1 in a hot loop. Cleared explicitly by
 // the renderer's "Reload and retry" path or by quitting the app.
 let bootstrapFailure = null
+// Latched non-bootstrap backend spawn failure — stops getConnection() from
+// respawning hermes dashboard children in a tight loop while boot is broken.
+let backendStartFailure = null
 // Active first-launch install, so the renderer's Cancel button (and app quit)
 // can abort the in-flight install.sh/ps1 instead of leaving it running.
 let bootstrapAbortController = null
@@ -1254,6 +1302,36 @@ function isCommandScript(command) {
   return IS_WINDOWS && /\.(cmd|bat)$/i.test(command || '')
 }
 
+function unwrapWindowsVenvHermesCommand(command, dashboardArgs) {
+  if (!IS_WINDOWS || !command || isCommandScript(command)) return null
+
+  const resolved = path.resolve(String(command))
+  if (!/^hermes(?:\.exe)?$/i.test(path.basename(resolved))) return null
+
+  const scriptsDir = path.dirname(resolved)
+  if (path.basename(scriptsDir).toLowerCase() !== 'scripts') return null
+
+  const venvRoot = path.dirname(scriptsDir)
+  const python = getNoConsoleVenvPython(venvRoot)
+  if (!fileExists(python)) return null
+
+  const root = path.dirname(venvRoot)
+  return {
+    label: `existing Hermes no-console Python at ${python}`,
+    command: python,
+    args: ['-m', 'hermes_cli.main', ...dashboardArgs],
+    bootstrap: false,
+    env: buildDesktopBackendEnv({
+      hermesHome: HERMES_HOME,
+      pythonPathEntries: [...(directoryExists(root) ? [root] : []), ...getVenvSitePackagesEntries(venvRoot)],
+      venvRoot
+    }),
+    kind: 'python',
+    readyFile: true,
+    shell: false
+  }
+}
+
 function normalizeExecutablePathForCompare(commandPath) {
   if (!commandPath) return null
 
@@ -1474,6 +1552,97 @@ function getVenvPython(venvRoot) {
   return path.join(venvRoot, IS_WINDOWS ? path.join('Scripts', 'python.exe') : path.join('bin', 'python'))
 }
 
+function readVenvHome(venvRoot) {
+  try {
+    const cfg = fs.readFileSync(path.join(venvRoot, 'pyvenv.cfg'), 'utf8')
+    const match = cfg.match(/^home\s*=\s*(.+?)\s*$/im)
+    return match ? match[1].trim() : null
+  } catch {
+    return null
+  }
+}
+
+function getNoConsoleVenvPython(venvRoot) {
+  if (!IS_WINDOWS) return getVenvPython(venvRoot)
+
+  // Prefer the venv's own pythonw shim — it carries pyvenv.cfg / site-packages
+  // wiring. Falling back to the base uv/python.org pythonw.exe skips the venv
+  // and breaks imports (yaml, hermes_cli, …) even when PYTHONPATH is patched.
+  const venvPythonw = path.join(venvRoot, 'Scripts', 'pythonw.exe')
+  if (fileExists(venvPythonw)) return venvPythonw
+
+  const baseHome = readVenvHome(venvRoot)
+  if (baseHome) {
+    const basePythonw = path.join(baseHome, 'pythonw.exe')
+    if (fileExists(basePythonw)) return basePythonw
+  }
+
+  return venvPythonw
+}
+
+function toNoConsolePython(pythonPath) {
+  if (!IS_WINDOWS || !pythonPath) return pythonPath
+
+  const resolved = String(pythonPath)
+  if (/pythonw\.exe$/i.test(resolved)) return resolved
+
+  if (/python\.exe$/i.test(resolved)) {
+    const pythonw = path.join(path.dirname(resolved), 'pythonw.exe')
+    if (fileExists(pythonw)) return pythonw
+  }
+
+  return pythonPath
+}
+
+function applyWindowsNoConsoleSpawnHints(backend) {
+  if (!IS_WINDOWS || !backend?.command) return backend
+
+  const usesHermesModule =
+    backend.kind === 'python' ||
+    (Array.isArray(backend.args) && backend.args[0] === '-m' && backend.args[1] === 'hermes_cli.main')
+
+  if (!usesHermesModule) return backend
+
+  backend.command = toNoConsolePython(backend.command)
+  if (/pythonw\.exe$/i.test(path.basename(String(backend.command || '')))) {
+    backend.readyFile = true
+  }
+
+  return backend
+}
+
+function getVenvSitePackagesEntries(venvRoot) {
+  const entries = []
+  if (!venvRoot) return entries
+
+  if (IS_WINDOWS) {
+    const sitePackages = path.join(venvRoot, 'Lib', 'site-packages')
+    if (directoryExists(sitePackages)) entries.push(sitePackages)
+    return entries
+  }
+
+  const version = (() => {
+    try {
+      const cfg = fs.readFileSync(path.join(venvRoot, 'pyvenv.cfg'), 'utf8')
+      const match = cfg.match(/^version_info\s*=\s*(\d+\.\d+)/im)
+      return match ? match[1].trim() : null
+    } catch {
+      return null
+    }
+  })()
+  if (version) {
+    const sitePackages = path.join(venvRoot, 'lib', `python${version}`, 'site-packages')
+    if (directoryExists(sitePackages)) entries.push(sitePackages)
+  }
+  return entries
+}
+
+function makeDashboardReadyFile() {
+  const dir = path.join(app.getPath('userData'), 'backend-ready')
+  fs.mkdirSync(dir, { recursive: true })
+  return path.join(dir, `dashboard-${process.pid}-${Date.now()}-${crypto.randomBytes(6).toString('hex')}.json`)
+}
+
 // resolveGitBinary — locate git.exe on Windows. A fresh installer-driven
 // install only has PortableGit under %LOCALAPPDATA%\hermes\git (never on
 // PATH), so a bare spawn('git') ENOENTs and self-update checks fail with
@@ -1503,6 +1672,30 @@ function resolveGitBinary() {
   return _gitBinaryCache
 }
 
+// resolveGhBinary — locate the GitHub CLI. GUI-launched apps get a minimal PATH
+// that omits Homebrew (/opt/homebrew/bin, /usr/local/bin) where `gh` usually
+// lives, so a bare spawn('gh') ENOENTs even though `gh` works in the user's
+// terminal. Check the common install locations first, then PATH. Cached.
+let _ghBinaryCache = null
+function resolveGhBinary() {
+  if (_ghBinaryCache) return _ghBinaryCache
+
+  const candidates = []
+
+  if (IS_WINDOWS) {
+    candidates.push(path.join(process.env['ProgramFiles'] || 'C:\\Program Files', 'GitHub CLI', 'gh.exe'))
+    if (process.env.LOCALAPPDATA) {
+      candidates.push(path.join(process.env.LOCALAPPDATA, 'Microsoft', 'WinGet', 'Links', 'gh.exe'))
+    }
+  } else {
+    const home = app.getPath('home')
+    candidates.push('/opt/homebrew/bin/gh', '/usr/local/bin/gh', '/usr/bin/gh', path.join(home, '.local', 'bin', 'gh'))
+  }
+
+  _ghBinaryCache = candidates.find(fileExists) || findOnPath('gh') || 'gh'
+  return _ghBinaryCache
+}
+
 function recentHermesLog() {
   return hermesLog.slice(-20).join('\n')
 }
@@ -1982,7 +2175,8 @@ async function applyUpdates(opts = {}) {
 
     emitUpdateProgress({
       stage: 'restart',
-      message: 'Updating Hermes — this window will close and the updater will open. Don’t reopen Hermes yourself; it restarts automatically when the update finishes.',
+      message:
+        'Updating Hermes — this window will close and the updater will open. Don’t reopen Hermes yourself; it restarts automatically when the update finishes.',
       percent: 100
     })
     repairMacUpdaterHelper(updater)
@@ -2065,7 +2259,9 @@ async function handOffWindowsBootstrapRecovery(reason) {
   })
   child.unref()
 
-  rememberLog(`[bootstrap] handed off ${reason} recovery to updater: ${updater} ${updaterArgs.join(' ')}; exiting desktop to release app.asar`)
+  rememberLog(
+    `[bootstrap] handed off ${reason} recovery to updater: ${updater} ${updaterArgs.join(' ')}; exiting desktop to release app.asar`
+  )
   // Same dwell as the in-app update hand-off (#50419): give the updater's
   // window time to appear before we vanish, so the recovery doesn't look like
   // a crash and provoke a mid-recovery relaunch.
@@ -2590,20 +2786,24 @@ function createPythonBackend(root, label, dashboardArgs, options = {}) {
   const python = findPythonForRoot(root)
   if (!python) return null
 
-  return {
+  const venvRoot = path.join(root, 'venv')
+  const venvPython = getVenvPython(venvRoot)
+  const command = IS_WINDOWS && fileExists(venvPython) ? getNoConsoleVenvPython(venvRoot) : toNoConsolePython(python)
+
+  return applyWindowsNoConsoleSpawnHints({
     kind: 'python',
     label,
-    command: python,
+    command,
     args: ['-m', 'hermes_cli.main', ...dashboardArgs],
     env: buildDesktopBackendEnv({
       hermesHome: HERMES_HOME,
       pythonPathEntries: [root],
-      venvRoot: path.join(root, 'venv')
+      venvRoot
     }),
     root,
     bootstrap: Boolean(options.bootstrap),
     shell: false
-  }
+  })
 }
 
 // createActiveBackend — build a backend pointing at ACTIVE_HERMES_ROOT, the
@@ -2612,11 +2812,12 @@ function createPythonBackend(root, label, dashboardArgs, options = {}) {
 // ensureRuntime() to create / refresh it before launch.
 function createActiveBackend(dashboardArgs) {
   const venvPython = getVenvPython(VENV_ROOT)
+  const command = fileExists(venvPython) ? getNoConsoleVenvPython(VENV_ROOT) : toNoConsolePython(findSystemPython())
 
-  return {
+  return applyWindowsNoConsoleSpawnHints({
     kind: 'python',
     label: `Hermes at ${ACTIVE_HERMES_ROOT}`,
-    command: fileExists(venvPython) ? venvPython : findSystemPython(),
+    command,
     args: ['-m', 'hermes_cli.main', ...dashboardArgs],
     env: buildDesktopBackendEnv({
       hermesHome: HERMES_HOME,
@@ -2626,7 +2827,7 @@ function createActiveBackend(dashboardArgs) {
     root: ACTIVE_HERMES_ROOT,
     bootstrap: true,
     shell: false
-  }
+  })
 }
 
 function resolveHermesBackend(dashboardArgs) {
@@ -2687,6 +2888,11 @@ function resolveHermesBackend(dashboardArgs) {
     }
 
     if (hermesCommand) {
+      const unwrapped = unwrapWindowsVenvHermesCommand(hermesCommand, dashboardArgs)
+      if (unwrapped) {
+        return unwrapped
+      }
+
       // Smoke-test the candidate before trusting it. A `hermes` shim
       // left behind by a half-uninstalled pip install (or a venv
       // entry-point pointing at a deleted interpreter) still resolves
@@ -2696,15 +2902,17 @@ function resolveHermesBackend(dashboardArgs) {
       // and lets the resolver fall through to step 6 / bootstrap.
       const shellForProbe = isCommandScript(hermesCommand)
       if (verifyHermesCli(hermesCommand, { shell: shellForProbe })) {
-        return {
-          label: `existing Hermes CLI at ${hermesCommand}`,
-          command: hermesCommand,
-          args: dashboardArgs,
-          bootstrap: false,
-          env: {},
-          kind: 'command',
-          shell: shellForProbe
-        }
+        return (
+          unwrapWindowsVenvHermesCommand(hermesCommand, dashboardArgs) || {
+            label: `existing Hermes CLI at ${hermesCommand}`,
+            command: hermesCommand,
+            args: dashboardArgs,
+            bootstrap: false,
+            env: {},
+            kind: 'command',
+            shell: shellForProbe
+          }
+        )
       }
       rememberLog(
         `Ignoring existing Hermes CLI at ${hermesCommand}: --version probe failed; falling through to bootstrap.`
@@ -2726,15 +2934,15 @@ function resolveHermesBackend(dashboardArgs) {
     // failure, fall through to step 6 so the bootstrap runner pulls
     // a uv-managed 3.11 into %LOCALAPPDATA%\hermes\hermes-agent\venv.
     if (canImportHermesCli(python)) {
-      return {
+      return applyWindowsNoConsoleSpawnHints({
         kind: 'python',
         label: `installed hermes_cli module via ${python}`,
-        command: python,
+        command: toNoConsolePython(python),
         args: ['-m', 'hermes_cli.main', ...dashboardArgs],
         bootstrap: false,
         env: {},
         shell: false
-      }
+      })
     }
     rememberLog(`Ignoring system Python ${python}: hermes_cli is not importable; falling through to bootstrap.`)
   }
@@ -2768,7 +2976,7 @@ function resolveHermesBackend(dashboardArgs) {
 async function ensureRuntime(backend) {
   if (!backend.bootstrap) {
     await advanceBootProgress('runtime.external', `Using ${backend.label}`, 32)
-    return backend
+    return applyWindowsNoConsoleSpawnHints(backend)
   }
 
   // backend.kind === 'bootstrap-needed' means resolveHermesBackend couldn't
@@ -2784,7 +2992,9 @@ async function ensureRuntime(backend) {
     rememberLog('[bootstrap] no Hermes install found; starting first-launch bootstrap')
 
     if (await handOffWindowsBootstrapRecovery('bootstrap-needed')) {
-      const handoffError = new Error('Hermes recovery was handed off to Hermes Setup. The desktop will restart when recovery completes.')
+      const handoffError = new Error(
+        'Hermes recovery was handed off to Hermes Setup. The desktop will restart when recovery completes.'
+      )
       handoffError.isBootstrapFailure = true
       handoffError.bootstrapHandedOff = true
       bootstrapFailure = handoffError
@@ -2908,7 +3118,7 @@ async function ensureRuntime(backend) {
     )
   }
 
-  backend.command = venvPython
+  backend.command = getNoConsoleVenvPython(VENV_ROOT)
   backend.label = `Hermes at ${ACTIVE_HERMES_ROOT} (venv: ${VENV_ROOT})`
   updateBootProgress({
     phase: 'runtime.ready',
@@ -2917,10 +3127,9 @@ async function ensureRuntime(backend) {
     running: true,
     error: null
   })
-  return backend
+  return applyWindowsNoConsoleSpawnHints(backend)
 }
 
-
 function fetchJson(url, token, options = {}) {
   return new Promise((resolve, reject) => {
     const body = options.body === undefined ? undefined : Buffer.from(JSON.stringify(options.body))
@@ -3579,11 +3788,7 @@ function getWindowButtonPosition() {
 }
 
 function getNativeOverlayWidth() {
-  // macOS reports traffic-light coords via windowButtonPosition; the
-  // titlebarOverlay there doesn't reserve right-edge space. Windows/Linux
-  // render the native window-controls overlay on the right, so the renderer
-  // needs to inset its right cluster by this much to clear them.
-  return IS_MAC ? 0 : NATIVE_OVERLAY_BUTTON_WIDTH
+  return computeNativeOverlayWidth({ isWindows: IS_WINDOWS, isWsl: IS_WSL })
 }
 
 function getWindowState() {
@@ -4831,6 +5036,7 @@ function resetBootProgressForReconnect() {
 
 function resetHermesConnection() {
   connectionPromise = null
+  backendStartFailure = null
 
   if (hermesProcess && !hermesProcess.killed) {
     hermesProcess.kill('SIGTERM')
@@ -4992,6 +5198,7 @@ async function spawnPoolBackend(profile, entry) {
   const backend = await ensureRuntime(resolveHermesBackend(dashboardArgs))
   const hermesCwd = resolveHermesCwd()
   const webDist = resolveWebDist()
+  const readyFile = backend.readyFile ? makeDashboardReadyFile() : null
 
   rememberLog(`Starting Hermes backend for profile "${profile}" via ${backend.label}`)
 
@@ -5012,7 +5219,8 @@ async function spawnPoolBackend(profile, entry) {
         // Marks this dashboard backend as desktop-spawned so it runs the cron
         // scheduler tick loop (the gateway isn't running under the app).
         HERMES_DESKTOP: '1',
-        HERMES_WEB_DIST: webDist
+        HERMES_WEB_DIST: webDist,
+        ...(readyFile ? { HERMES_DESKTOP_READY_FILE: readyFile } : {})
       },
       shell: backend.shell,
       stdio: ['ignore', 'pipe', 'pipe']
@@ -5045,7 +5253,10 @@ async function spawnPoolBackend(profile, entry) {
   })
 
   // Discover the ephemeral port the child bound to
-  const port = await Promise.race([waitForDashboardPort(child), startFailed])
+  const port = await Promise.race([waitForDashboardPortAnnouncement(child, { readyFile }), startFailed])
+  if (readyFile) {
+    fs.unlink(readyFile, () => {})
+  }
   entry.port = port
 
   const baseUrl = `http://127.0.0.1:${port}`
@@ -5158,6 +5369,9 @@ async function startHermes() {
   if (bootstrapFailure) {
     throw bootstrapFailure
   }
+  if (backendStartFailure) {
+    throw backendStartFailure
+  }
   if (connectionPromise) return connectionPromise
 
   connectionPromise = (async () => {
@@ -5211,6 +5425,7 @@ async function startHermes() {
     const backend = await ensureRuntime(resolveHermesBackend(dashboardArgs))
     const hermesCwd = resolveHermesCwd()
     const webDist = resolveWebDist()
+    const readyFile = backend.readyFile ? makeDashboardReadyFile() : null
 
     await advanceBootProgress('backend.spawn', `Starting Hermes backend via ${backend.label}`, 84)
     rememberLog(`Starting Hermes backend via ${backend.label}`)
@@ -5237,7 +5452,8 @@ async function startHermes() {
           // Marks this dashboard backend as desktop-spawned so it runs the cron
           // scheduler tick loop (the gateway isn't running under the app).
           HERMES_DESKTOP: '1',
-          HERMES_WEB_DIST: webDist
+          HERMES_WEB_DIST: webDist,
+          ...(readyFile ? { HERMES_DESKTOP_READY_FILE: readyFile } : {})
         },
         shell: backend.shell,
         stdio: ['ignore', 'pipe', 'pipe']
@@ -5293,12 +5509,19 @@ async function startHermes() {
 
     await advanceBootProgress('backend.port', 'Waiting for Hermes backend to launch', 86)
     // Discover the ephemeral port the child bound to
-    const port = await Promise.race([waitForDashboardPort(hermesProcess), backendStartFailed])
+    const port = await Promise.race([
+      waitForDashboardPortAnnouncement(hermesProcess, { readyFile }),
+      backendStartFailed
+    ])
+    if (readyFile) {
+      fs.unlink(readyFile, () => {})
+    }
 
     const baseUrl = `http://127.0.0.1:${port}`
     await advanceBootProgress('backend.wait', 'Waiting for Hermes backend to become ready', 90)
     await Promise.race([waitForHermes(baseUrl, token), backendStartFailed])
     backendReady = true
+    backendStartFailure = null
     const authToken = await adoptServedDashboardToken(baseUrl, token, {
       // The exit/error handlers null hermesProcess when the child dies.
       childAlive: () => hermesProcess !== null && hermesProcess.exitCode === null && !hermesProcess.killed,
@@ -5324,6 +5547,7 @@ async function startHermes() {
     }
   })().catch(error => {
     const message = error instanceof Error ? error.message : String(error)
+    backendStartFailure = error instanceof Error ? error : new Error(message)
     updateBootProgress(
       {
         error: message,
@@ -5623,7 +5847,7 @@ function createWindow() {
     if (!nativeThemeListenerInstalled) {
       nativeThemeListenerInstalled = true
       nativeTheme.on('updated', () => {
-        mainWindow?.setTitleBarOverlay?.(getTitleBarOverlayOptions())
+        applyTitleBarOverlay(mainWindow)
       })
     }
   }
@@ -5807,19 +6031,32 @@ ipcMain.handle('hermes:pet-overlay:close', async () => {
 
   return { ok: true }
 })
-// Drag: the overlay reports a new absolute screen position (it already knows the
-// pointer's screen coords), we just move the window.
+// Drag/resize: the overlay reports new absolute screen bounds (it already knows
+// the pointer's screen coords). Drag keeps the size constant; the wheel-to-scale
+// gesture grows/shrinks it so the sprite is never cropped by the window edge.
+// The window is created non-resizable (no stray edge-drag on the transparent
+// frameless panel), which on Windows/Linux also blocks programmatic setBounds
+// sizing — so briefly flip resizable on whenever the size actually changes.
 ipcMain.on('hermes:pet-overlay:set-bounds', (_event, bounds) => {
   if (!petOverlayWindow || petOverlayWindow.isDestroyed() || !bounds) {
     return
   }
 
-  petOverlayWindow.setBounds({
-    x: Math.round(bounds.x),
-    y: Math.round(bounds.y),
-    width: Math.max(80, Math.round(bounds.width)),
-    height: Math.max(80, Math.round(bounds.height))
-  })
+  const win = petOverlayWindow
+  const width = Math.max(80, Math.round(bounds.width))
+  const height = Math.max(80, Math.round(bounds.height))
+  const [curW, curH] = win.getSize()
+  const resizing = width !== curW || height !== curH
+
+  if (resizing && !win.isResizable()) {
+    win.setResizable(true)
+  }
+
+  win.setBounds({ x: Math.round(bounds.x), y: Math.round(bounds.y), width, height })
+
+  if (resizing) {
+    win.setResizable(false)
+  }
 })
 // Click-through: the overlay window is a full rectangle but only the pet pixels
 // should be interactive. The renderer toggles this as the cursor enters/leaves
@@ -5889,6 +6126,7 @@ ipcMain.handle('hermes:bootstrap:reset', async () => {
   rememberLog('[bootstrap] reset requested by renderer; clearing latched failure')
   await teardownPrimaryBackendAndWait()
   bootstrapFailure = null
+  backendStartFailure = null
   bootstrapState = {
     active: false,
     manifest: null,
@@ -5915,6 +6153,7 @@ ipcMain.handle('hermes:bootstrap:repair', async () => {
     rememberLog(`[bootstrap] failed to remove marker during repair: ${error.message}`)
   }
   bootstrapFailure = null
+  backendStartFailure = null
   resetHermesConnection()
   return { ok: true }
 })
@@ -6283,11 +6522,21 @@ ipcMain.handle('hermes:saveImageBuffer', async (_event, payload) => {
 
 ipcMain.handle('hermes:saveClipboardImage', async () => {
   const image = clipboard.readImage()
-  if (!image || image.isEmpty()) {
-    return ''
+  if (image && !image.isEmpty()) {
+    return writeComposerImage(image.toPNG(), '.png')
   }
 
-  return writeComposerImage(image.toPNG(), '.png')
+  // WSL2/WSLg doesn't bridge clipboard *images* from the Windows host to the
+  // Linux clipboard Electron reads, so a host screenshot looks empty above.
+  // Pull it straight off the Windows clipboard via PowerShell as a fallback.
+  if (IS_WSL) {
+    const png = readWslWindowsClipboardImage()
+    if (png) {
+      return writeComposerImage(png, '.png')
+    }
+  }
+
+  return ''
 })
 
 ipcMain.handle('hermes:normalizePreviewTarget', (_event, target, baseDir) =>
@@ -6307,7 +6556,7 @@ ipcMain.on('hermes:titlebar-theme', (_event, payload) => {
     background: payload.background,
     foreground: payload.foreground
   }
-  mainWindow?.setTitleBarOverlay?.(getTitleBarOverlayOptions())
+  applyTitleBarOverlay(mainWindow)
 })
 
 // Pin the native appearance to the app theme (see NATIVE_THEME_CONFIG_PATH).
@@ -6596,7 +6845,160 @@ ipcMain.handle('hermes:fs:readDir', async (_event, dirPath) => readDirForIpc(dir
 
 ipcMain.handle('hermes:fs:gitRoot', async (_event, startPath) => gitRootForIpc(startPath))
 
-ipcMain.handle('hermes:fs:worktrees', async (_event, cwds) => worktreesForIpc(cwds))
+// Reveal a path in the OS file manager (Finder / Explorer / Files).
+ipcMain.handle('hermes:fs:reveal', async (_event, targetPath) => {
+  const target = String(targetPath || '').trim()
+
+  if (!target) {
+    return false
+  }
+
+  try {
+    shell.showItemInFolder(target)
+
+    return true
+  } catch {
+    return false
+  }
+})
+
+// Rename a file/folder in place. The renderer passes the existing path + a new
+// base name; the destination is resolved in the SAME parent dir so a rename can
+// never move the item elsewhere or traverse out. Rejects on a name collision.
+ipcMain.handle('hermes:fs:rename', async (_event, targetPath, newName) => {
+  const src = String(targetPath || '').trim()
+  const name = String(newName || '').trim()
+
+  if (!src || !name || name === '.' || name === '..' || name.includes('/') || name.includes('\\')) {
+    throw new Error('Invalid rename')
+  }
+
+  const dst = path.join(path.dirname(src), name)
+
+  if (dst === src) {
+    return { path: dst }
+  }
+
+  if (fs.existsSync(dst)) {
+    throw new Error(`"${name}" already exists`)
+  }
+
+  await fs.promises.rename(src, dst)
+
+  return { path: dst }
+})
+
+// Write a small UTF-8 text file (e.g. a project's IDEA.md at creation). The path
+// is hardened (resolveRequestedPathForIpc) and the parent must already exist —
+// this never creates directory trees or escapes the allowed roots, and content
+// is size-capped so it can't be abused as a bulk-write primitive.
+ipcMain.handle('hermes:fs:writeText', async (_event, filePath, content) => {
+  const raw = String(filePath || '').trim()
+
+  if (!raw) {
+    throw new Error('Invalid path')
+  }
+
+  const text = String(content ?? '')
+
+  if (text.length > 1_000_000) {
+    throw new Error('Content too large')
+  }
+
+  const resolved = resolveRequestedPathForIpc(expandUserPath(raw), { purpose: 'Write text file' })
+
+  if (!directoryExists(path.dirname(resolved))) {
+    throw new Error('Parent directory does not exist')
+  }
+
+  await fs.promises.writeFile(resolved, text, 'utf8')
+
+  return { path: resolved }
+})
+
+// Move a file/folder to the OS trash (recoverable) — the VS Code "Delete"
+// default. `shell.trashItem` routes to Finder/Explorer/Files trash per platform.
+ipcMain.handle('hermes:fs:trash', async (_event, targetPath) => {
+  const target = String(targetPath || '').trim()
+
+  if (!target) {
+    throw new Error('Invalid delete')
+  }
+
+  await shell.trashItem(target)
+
+  return true
+})
+
+// Git-driven worktree management ("Start work" flow). Errors surface to the
+// renderer as rejected promises so it can toast a friendly message.
+ipcMain.handle('hermes:git:worktreeList', async (_event, repoPath) => listWorktrees(repoPath, resolveGitBinary()))
+
+ipcMain.handle('hermes:git:worktreeAdd', async (_event, repoPath, options) =>
+  addWorktree(repoPath, options || {}, resolveGitBinary())
+)
+
+ipcMain.handle('hermes:git:worktreeRemove', async (_event, repoPath, worktreePath, options) =>
+  removeWorktree(repoPath, worktreePath, options || {}, resolveGitBinary())
+)
+
+ipcMain.handle('hermes:git:branchSwitch', async (_event, repoPath, branch) =>
+  switchBranch(repoPath, branch, resolveGitBinary())
+)
+
+ipcMain.handle('hermes:git:branchList', async (_event, repoPath) => listBranches(repoPath, resolveGitBinary()))
+
+// Compact repo status (branch, ahead/behind, change counts + files) for the
+// composer coding rail. Returns null on a non-repo / remote backend so the rail
+// hides cleanly rather than erroring.
+ipcMain.handle('hermes:git:repoStatus', async (_event, repoPath) => repoStatus(repoPath, resolveGitBinary()))
+
+// Codex-style review pane: list changed files for a scope, fetch one file's
+// unified diff, and stage / unstage / revert. Reads return empty on failure;
+// mutations reject so the renderer can toast.
+ipcMain.handle('hermes:git:review:list', async (_event, repoPath, scope, baseRef) =>
+  reviewList(repoPath, scope, baseRef, resolveGitBinary())
+)
+ipcMain.handle('hermes:git:review:diff', async (_event, repoPath, filePath, scope, baseRef, staged) =>
+  reviewDiff(repoPath, filePath, scope, baseRef, staged, resolveGitBinary())
+)
+// Working-tree-vs-HEAD diff for one file (the preview's "show the diff" view).
+ipcMain.handle('hermes:git:fileDiff', async (_event, repoPath, filePath) =>
+  fileDiffVsHead(repoPath, filePath, resolveGitBinary())
+)
+ipcMain.handle('hermes:git:review:stage', async (_event, repoPath, filePath) =>
+  reviewStage(repoPath, filePath ?? null, resolveGitBinary())
+)
+ipcMain.handle('hermes:git:review:unstage', async (_event, repoPath, filePath) =>
+  reviewUnstage(repoPath, filePath ?? null, resolveGitBinary())
+)
+ipcMain.handle('hermes:git:review:revert', async (_event, repoPath, filePath) =>
+  reviewRevert(repoPath, filePath ?? null, resolveGitBinary())
+)
+ipcMain.handle('hermes:git:review:revParse', async (_event, repoPath, ref) =>
+  reviewRevParse(repoPath, ref, resolveGitBinary())
+)
+ipcMain.handle('hermes:git:review:commit', async (_event, repoPath, message, push) =>
+  reviewCommit(repoPath, message, Boolean(push), resolveGitBinary())
+)
+ipcMain.handle('hermes:git:review:commitContext', async (_event, repoPath) =>
+  reviewCommitContext(repoPath, resolveGitBinary())
+)
+ipcMain.handle('hermes:git:review:push', async (_event, repoPath) => reviewPush(repoPath, resolveGitBinary()))
+ipcMain.handle('hermes:git:review:shipInfo', async (_event, repoPath) => reviewShipInfo(repoPath, resolveGhBinary()))
+ipcMain.handle('hermes:git:review:createPr', async (_event, repoPath) =>
+  reviewCreatePr(repoPath, resolveGitBinary(), resolveGhBinary())
+)
+
+// Repo-first project discovery: scan bounded roots for git repos (pure fs walk,
+// no native addon). Never throws to the renderer — failures yield an empty list.
+ipcMain.handle('hermes:git:scanRepos', async (_event, roots, options) => {
+  try {
+    return await scanGitRepos(roots || [], options || {})
+  } catch {
+    return []
+  }
+})
 
 ipcMain.handle('hermes:terminal:start', async (event, payload = {}) => {
   if (!nodePty) {
@@ -7046,6 +7448,7 @@ app.whenReady().then(() => {
   }
   installMediaPermissions()
   registerMediaProtocol()
+  installEmbedReferer()
   registerDeepLinkProtocol()
   ensureWslWindowsFonts()
   configureSpellChecker()

apps/desktop/electron/oauth-net-request.test.cjs

@@ -30,5 +30,8 @@ test('setJsonRequestHeaders does not set Electron-restricted Content-Length', ()
   setJsonRequestHeaders(request)
 
   assert.deepEqual(headers, [['Content-Type', 'application/json']])
-  assert.equal(headers.some(([name]) => name.toLowerCase() === 'content-length'), false)
+  assert.equal(
+    headers.some(([name]) => name.toLowerCase() === 'content-length'),
+    false
+  )
 })

apps/desktop/electron/preload.cjs

@@ -82,7 +82,35 @@ contextBridge.exposeInMainWorld('hermesDesktop', {
   getRecentLogs: () => ipcRenderer.invoke('hermes:logs:recent'),
   readDir: dirPath => ipcRenderer.invoke('hermes:fs:readDir', dirPath),
   gitRoot: startPath => ipcRenderer.invoke('hermes:fs:gitRoot', startPath),
-  worktrees: cwds => ipcRenderer.invoke('hermes:fs:worktrees', cwds),
+  revealPath: targetPath => ipcRenderer.invoke('hermes:fs:reveal', targetPath),
+  renamePath: (targetPath, newName) => ipcRenderer.invoke('hermes:fs:rename', targetPath, newName),
+  writeTextFile: (filePath, content) => ipcRenderer.invoke('hermes:fs:writeText', filePath, content),
+  trashPath: targetPath => ipcRenderer.invoke('hermes:fs:trash', targetPath),
+  git: {
+    worktreeList: repoPath => ipcRenderer.invoke('hermes:git:worktreeList', repoPath),
+    worktreeAdd: (repoPath, options) => ipcRenderer.invoke('hermes:git:worktreeAdd', repoPath, options),
+    worktreeRemove: (repoPath, worktreePath, options) =>
+      ipcRenderer.invoke('hermes:git:worktreeRemove', repoPath, worktreePath, options),
+    branchSwitch: (repoPath, branch) => ipcRenderer.invoke('hermes:git:branchSwitch', repoPath, branch),
+    branchList: repoPath => ipcRenderer.invoke('hermes:git:branchList', repoPath),
+    repoStatus: repoPath => ipcRenderer.invoke('hermes:git:repoStatus', repoPath),
+    fileDiff: (repoPath, filePath) => ipcRenderer.invoke('hermes:git:fileDiff', repoPath, filePath),
+    scanRepos: (roots, options) => ipcRenderer.invoke('hermes:git:scanRepos', roots, options),
+    review: {
+      list: (repoPath, scope, baseRef) => ipcRenderer.invoke('hermes:git:review:list', repoPath, scope, baseRef),
+      diff: (repoPath, filePath, scope, baseRef, staged) =>
+        ipcRenderer.invoke('hermes:git:review:diff', repoPath, filePath, scope, baseRef, staged),
+      stage: (repoPath, filePath) => ipcRenderer.invoke('hermes:git:review:stage', repoPath, filePath),
+      unstage: (repoPath, filePath) => ipcRenderer.invoke('hermes:git:review:unstage', repoPath, filePath),
+      revert: (repoPath, filePath) => ipcRenderer.invoke('hermes:git:review:revert', repoPath, filePath),
+      revParse: (repoPath, ref) => ipcRenderer.invoke('hermes:git:review:revParse', repoPath, ref),
+      commit: (repoPath, message, push) => ipcRenderer.invoke('hermes:git:review:commit', repoPath, message, push),
+      commitContext: repoPath => ipcRenderer.invoke('hermes:git:review:commitContext', repoPath),
+      push: repoPath => ipcRenderer.invoke('hermes:git:review:push', repoPath),
+      shipInfo: repoPath => ipcRenderer.invoke('hermes:git:review:shipInfo', repoPath),
+      createPr: repoPath => ipcRenderer.invoke('hermes:git:review:createPr', repoPath)
+    }
+  },
   terminal: {
     dispose: id => ipcRenderer.invoke('hermes:terminal:dispose', id),
     resize: (id, size) => ipcRenderer.invoke('hermes:terminal:resize', id, size),

apps/desktop/electron/titlebar-overlay-width.cjs

@@ -0,0 +1,11 @@
+// Pre-layout fallback for WCO right-edge reservation (--titlebar-tools-right).
+// Live width comes from navigator.windowControlsOverlay in the renderer.
+
+const OVERLAY_FALLBACK_WIDTH = 144
+
+/** @param {{ isWindows?: boolean, isWsl?: boolean }} opts */
+function nativeOverlayWidth({ isWindows = false, isWsl = false } = {}) {
+  return isWindows || isWsl ? OVERLAY_FALLBACK_WIDTH : 0
+}
+
+module.exports = { OVERLAY_FALLBACK_WIDTH, nativeOverlayWidth }

apps/desktop/electron/titlebar-overlay-width.test.cjs

@@ -0,0 +1,29 @@
+const assert = require('node:assert/strict')
+const test = require('node:test')
+
+const { OVERLAY_FALLBACK_WIDTH, nativeOverlayWidth } = require('./titlebar-overlay-width.cjs')
+
+// This static reservation is only the pre-layout FALLBACK. Once laid out the
+// renderer reads the exact width from navigator.windowControlsOverlay
+// (use-window-controls-overlay-width.ts) and uses these values only when the WCO
+// API is unavailable.
+
+test('Windows reserves the overlay fallback width', () => {
+  assert.equal(nativeOverlayWidth({ isWindows: true }), OVERLAY_FALLBACK_WIDTH)
+})
+
+test('WSLg paints the same WCO, so it reserves the same fallback width', () => {
+  // The original bug: WSL fell through to 0, so the right tools sat under the
+  // controls and the title overran into them.
+  assert.equal(nativeOverlayWidth({ isWsl: true }), OVERLAY_FALLBACK_WIDTH)
+})
+
+test('plain Linux and macOS reserve nothing', () => {
+  assert.equal(nativeOverlayWidth({ isWindows: false, isWsl: false }), 0)
+  assert.equal(nativeOverlayWidth(), 0)
+  assert.equal(nativeOverlayWidth({}), 0)
+})
+
+test('the fallback width is a sane positive pixel value', () => {
+  assert.ok(Number.isInteger(OVERLAY_FALLBACK_WIDTH) && OVERLAY_FALLBACK_WIDTH > 0)
+})

apps/desktop/electron/update-count.test.cjs

@@ -7,45 +7,81 @@ const { resolveBehindCount, shouldCountCommits } = require('./update-count.cjs')
 // unconditionally, so a shallow checkout with no merge-base surfaced the bogus
 // rev-list count (e.g. 12104). This asserts the new shallow/no-merge-base branch.
 test('shallow checkout with no merge-base does NOT trust the bogus rev-list count', () => {
-  assert.equal(resolveBehindCount({
-    countStr: '12104', currentSha: 'aaa', targetSha: 'bbb',
-    isShallow: true, hasMergeBase: false,
-  }), 1)
+  assert.equal(
+    resolveBehindCount({
+      countStr: '12104',
+      currentSha: 'aaa',
+      targetSha: 'bbb',
+      isShallow: true,
+      hasMergeBase: false
+    }),
+    1
+  )
 })
 
 test('shallow checkout with no merge-base but identical SHA reports up-to-date', () => {
-  assert.equal(resolveBehindCount({
-    countStr: '12104', currentSha: 'abc', targetSha: 'abc',
-    isShallow: true, hasMergeBase: false,
-  }), 0)
+  assert.equal(
+    resolveBehindCount({
+      countStr: '12104',
+      currentSha: 'abc',
+      targetSha: 'abc',
+      isShallow: true,
+      hasMergeBase: false
+    }),
+    0
+  )
 })
 
 test('shallow checkout WITH a merge-base keeps the exact count (reliable)', () => {
-  assert.equal(resolveBehindCount({
-    countStr: '3', currentSha: 'aaa', targetSha: 'bbb',
-    isShallow: true, hasMergeBase: true,
-  }), 3)
+  assert.equal(
+    resolveBehindCount({
+      countStr: '3',
+      currentSha: 'aaa',
+      targetSha: 'bbb',
+      isShallow: true,
+      hasMergeBase: true
+    }),
+    3
+  )
 })
 
 test('full (non-shallow) clone keeps the exact count path unchanged', () => {
-  assert.equal(resolveBehindCount({
-    countStr: '7', currentSha: 'aaa', targetSha: 'bbb',
-    isShallow: false, hasMergeBase: true,
-  }), 7)
+  assert.equal(
+    resolveBehindCount({
+      countStr: '7',
+      currentSha: 'aaa',
+      targetSha: 'bbb',
+      isShallow: false,
+      hasMergeBase: true
+    }),
+    7
+  )
 })
 
 test('up-to-date full clone reports 0', () => {
-  assert.equal(resolveBehindCount({
-    countStr: '0', currentSha: 'x', targetSha: 'x',
-    isShallow: false, hasMergeBase: true,
-  }), 0)
+  assert.equal(
+    resolveBehindCount({
+      countStr: '0',
+      currentSha: 'x',
+      targetSha: 'x',
+      isShallow: false,
+      hasMergeBase: true
+    }),
+    0
+  )
 })
 
 test('non-numeric count falls back to 0 (defensive, unchanged behaviour)', () => {
-  assert.equal(resolveBehindCount({
-    countStr: '', currentSha: 'aaa', targetSha: 'bbb',
-    isShallow: false, hasMergeBase: true,
-  }), 0)
+  assert.equal(
+    resolveBehindCount({
+      countStr: '',
+      currentSha: 'aaa',
+      targetSha: 'bbb',
+      isShallow: false,
+      hasMergeBase: true
+    }),
+    0
+  )
 })
 
 // shouldCountCommits gates the expensive `rev-list --count` in checkUpdates().
@@ -68,12 +104,24 @@ test('full (non-shallow) clone always runs the count', () => {
 // The skip path produces an empty countStr; resolveBehindCount must NOT trust
 // it and must fall through to the SHA compare (mirrors the live call site).
 test('skipped-count path resolves via SHA compare, never via empty countStr', () => {
-  assert.equal(resolveBehindCount({
-    countStr: '', currentSha: 'aaa', targetSha: 'bbb',
-    isShallow: true, hasMergeBase: false,
-  }), 1)
-  assert.equal(resolveBehindCount({
-    countStr: '', currentSha: 'same', targetSha: 'same',
-    isShallow: true, hasMergeBase: false,
-  }), 0)
+  assert.equal(
+    resolveBehindCount({
+      countStr: '',
+      currentSha: 'aaa',
+      targetSha: 'bbb',
+      isShallow: true,
+      hasMergeBase: false
+    }),
+    1
+  )
+  assert.equal(
+    resolveBehindCount({
+      countStr: '',
+      currentSha: 'same',
+      targetSha: 'same',
+      isShallow: true,
+      hasMergeBase: false
+    }),
+    0
+  )
 })

apps/desktop/electron/update-relaunch.test.cjs

@@ -62,7 +62,10 @@ test('resolveUnpackedRelease is null for AppImage / .deb / .rpm / dev / unresolv
   assert.equal(resolveUnpackedRelease('/usr/lib/hermes/hermes', ROOT, 'linux'), null)
   assert.equal(resolveUnpackedRelease('/opt/Hermes/hermes', ROOT, 'linux'), null)
   // dev electron
-  assert.equal(resolveUnpackedRelease('/home/u/.hermes/hermes-agent/node_modules/electron/dist/electron', ROOT, 'linux'), null)
+  assert.equal(
+    resolveUnpackedRelease('/home/u/.hermes/hermes-agent/node_modules/electron/dist/electron', ROOT, 'linux'),
+    null
+  )
   // empty / missing
   assert.equal(resolveUnpackedRelease('', ROOT, 'linux'), null)
   assert.equal(resolveUnpackedRelease(path.join(UNPACKED, 'hermes'), '', 'linux'), null)

apps/desktop/electron/update-remote.cjs

@@ -39,7 +39,9 @@ function canonicalGitHubRemote(url) {
 }
 
 function isSshRemote(url) {
-  const value = String(url || '').trim().toLowerCase()
+  const value = String(url || '')
+    .trim()
+    .toLowerCase()
   return value.startsWith('git@') || value.startsWith('ssh://')
 }
 

apps/desktop/electron/vscode-marketplace.cjs

@@ -26,7 +26,11 @@ const REQUEST_TIMEOUT_MS = 20_000
 const ID_RE = /^[\w-]+\.[\w-]+$/
 
 /** Minimal HTTPS helper with redirect-following, timeout, and a size cap. */
-function request(url, { method = 'GET', headers = {}, body = null, maxBytes = MAX_VSIX_BYTES } = {}, redirectsLeft = MAX_REDIRECTS) {
+function request(
+  url,
+  { method = 'GET', headers = {}, body = null, maxBytes = MAX_VSIX_BYTES } = {},
+  redirectsLeft = MAX_REDIRECTS
+) {
   return new Promise((resolve, reject) => {
     const req = https.request(url, { method, headers }, res => {
       const status = res.statusCode ?? 0
@@ -42,7 +46,13 @@ function request(url, { method = 'GET', headers = {}, body = null, maxBytes = MA
         const next = new URL(res.headers.location, url).toString()
         res.resume()
         // Redirects to the CDN are plain GETs (drop the POST body).
-        resolve(request(next, { method: 'GET', headers: { 'User-Agent': headers['User-Agent'] }, maxBytes }, redirectsLeft - 1))
+        resolve(
+          request(
+            next,
+            { method: 'GET', headers: { 'User-Agent': headers['User-Agent'] }, maxBytes },
+            redirectsLeft - 1
+          )
+        )
 
         return
       }

apps/desktop/electron/window-state.test.cjs

@@ -26,7 +26,16 @@ const LAPTOP = [{ workArea: { x: 0, y: 0, width: 1366, height: 728 } }]
 // ─── sanitizeWindowState ───────────────────────────────────────────────────
 
 test('sanitizeWindowState rejects missing/garbage input', () => {
-  for (const bad of [null, undefined, 'nope', 42, {}, { width: 'x', height: 800 }, { width: NaN, height: 800 }, { width: 1000 }]) {
+  for (const bad of [
+    null,
+    undefined,
+    'nope',
+    42,
+    {},
+    { width: 'x', height: 800 },
+    { width: NaN, height: 800 },
+    { width: 1000 }
+  ]) {
     assert.equal(sanitizeWindowState(bad), null)
   }
 })
@@ -112,9 +121,13 @@ test('computeWindowOptions does not clamp when displays are unknown', () => {
 test('debounce coalesces a burst into one trailing run', t => {
   t.mock.timers.enable({ apis: ['setTimeout'] })
   let calls = 0
-  const d = debounce(() => { calls += 1 }, 250)
+  const d = debounce(() => {
+    calls += 1
+  }, 250)
 
-  d(); d(); d()
+  d()
+  d()
+  d()
   assert.equal(calls, 0)
   t.mock.timers.tick(249)
   assert.equal(calls, 0)
@@ -125,7 +138,9 @@ test('debounce coalesces a burst into one trailing run', t => {
 test('debounce.flush runs now and cancels the pending timer', t => {
   t.mock.timers.enable({ apis: ['setTimeout'] })
   let calls = 0
-  const d = debounce(() => { calls += 1 }, 250)
+  const d = debounce(() => {
+    calls += 1
+  }, 250)
 
   d()
   d.flush()

apps/desktop/electron/windows-child-process.test.cjs

@@ -12,7 +12,8 @@ function readElectronFile(name) {
 }
 
 function requireHiddenChildOptions(source, needle) {
-  const index = source.indexOf(needle)
+  const match = needle instanceof RegExp ? needle.exec(source) : null
+  const index = needle instanceof RegExp ? (match?.index ?? -1) : source.indexOf(needle)
   assert.notEqual(index, -1, `missing call site: ${needle}`)
   const snippet = source.slice(index, index + 700)
   assert.match(
@@ -28,14 +29,28 @@ test('desktop background child processes opt into hidden Windows consoles', () =
   assert.match(source, /function hiddenWindowsChildOptions\(options = \{\}\)/)
 
   requireHiddenChildOptions(source, "execFileSync(\n          'reg'")
-  requireHiddenChildOptions(source, 'execFileSync(pyExe')
-  requireHiddenChildOptions(source, 'spawn(resolveGitBinary()')
+  requireHiddenChildOptions(source, /execFileSync\(\s*pyExe/)
+  requireHiddenChildOptions(source, /spawn\(\s*resolveGitBinary\(\)/)
   requireHiddenChildOptions(source, "execFileSync('taskkill'")
-  requireHiddenChildOptions(source, 'spawn(command, args')
+  requireHiddenChildOptions(source, /spawn\(\s*command,\s*args/)
   requireHiddenChildOptions(source, "spawn('curl'")
-  requireHiddenChildOptions(source, 'spawn(backend.command, backend.args')
-  requireHiddenChildOptions(source, 'hermesProcess = spawn(backend.command, backend.args')
-  requireHiddenChildOptions(source, "spawn(py, ['-m', 'hermes_cli.main', 'uninstall', '--gui-summary']")
+  requireHiddenChildOptions(source, /spawn\(\s*backend\.command,\s*backend\.args/)
+  requireHiddenChildOptions(source, /hermesProcess = spawn\(\s*backend\.command,\s*backend\.args/)
+  requireHiddenChildOptions(source, /spawn\(\s*py,\s*\['-m', 'hermes_cli\.main', 'uninstall', '--gui-summary'\]/)
+
+  assert.match(source, /function unwrapWindowsVenvHermesCommand\(command, dashboardArgs\)/)
+  assert.match(source, /existing Hermes no-console Python at/)
+  assert.match(source, /function getNoConsoleVenvPython\(venvRoot\)/)
+  assert.match(source, /function toNoConsolePython\(pythonPath\)/)
+  assert.match(source, /function applyWindowsNoConsoleSpawnHints\(backend\)/)
+  assert.match(source, /function readVenvHome\(venvRoot\)/)
+  assert.match(source, /path\.join\(venvRoot, 'Scripts', 'pythonw\.exe'\)/)
+  assert.match(source, /backendStartFailure/)
+  assert.match(source, /HERMES_DESKTOP_READY_FILE/)
+  assert.match(source, /readyFile: true/)
+  assert.match(source, /function getVenvSitePackagesEntries\(venvRoot\)/)
+  assert.match(source, /path\.join\(venvRoot, 'Lib', 'site-packages'\)/)
+  assert.match(source, /args: \['-m', 'hermes_cli\.main', \.\.\.dashboardArgs\]/)
 })
 
 test('intentional or interactive desktop child processes stay documented', () => {

apps/desktop/electron/windows-user-env.cjs

@@ -21,8 +21,7 @@ const { execFileSync } = require('node:child_process')
 // the requested value line isn't present.
 function parseRegQueryValue(stdout, name) {
   if (!stdout || !name) return null
-  const typePattern =
-    /^(\S+)\s+(?:REG_SZ|REG_EXPAND_SZ|REG_MULTI_SZ|REG_DWORD|REG_QWORD|REG_BINARY|REG_NONE)\s+(.*)$/
+  const typePattern = /^(\S+)\s+(?:REG_SZ|REG_EXPAND_SZ|REG_MULTI_SZ|REG_DWORD|REG_QWORD|REG_BINARY|REG_NONE)\s+(.*)$/
   for (const rawLine of String(stdout).split(/\r?\n/)) {
     const line = rawLine.trim()
     const match = line.match(typePattern)
@@ -47,10 +46,7 @@ function expandWindowsEnvRefs(value, env = process.env) {
 // Read a User-scoped env var from HKCU\Environment. Windows-only: returns null
 // off-Windows (without spawning), on any spawn error, when `reg` exits non-zero
 // (the value doesn't exist), or when the value is empty.
-function readWindowsUserEnvVar(
-  name,
-  { platform = process.platform, env = process.env, exec = execFileSync } = {}
-) {
+function readWindowsUserEnvVar(name, { platform = process.platform, env = process.env, exec = execFileSync } = {}) {
   if (platform !== 'win32' || !name) return null
   let stdout
   try {

apps/desktop/electron/windows-user-env.test.cjs

@@ -1,21 +1,12 @@
 const assert = require('node:assert/strict')
 const { test } = require('node:test')
 
-const {
-  expandWindowsEnvRefs,
-  parseRegQueryValue,
-  readWindowsUserEnvVar
-} = require('./windows-user-env.cjs')
+const { expandWindowsEnvRefs, parseRegQueryValue, readWindowsUserEnvVar } = require('./windows-user-env.cjs')
 
 // ── parseRegQueryValue ─────────────────────────────────────────────────────
 
 test('parseRegQueryValue extracts a REG_SZ value', () => {
-  const out = [
-    '',
-    'HKEY_CURRENT_USER\\Environment',
-    '    HERMES_HOME    REG_SZ    F:\\Hermes\\data',
-    ''
-  ].join('\r\n')
+  const out = ['', 'HKEY_CURRENT_USER\\Environment', '    HERMES_HOME    REG_SZ    F:\\Hermes\\data', ''].join('\r\n')
   assert.equal(parseRegQueryValue(out, 'HERMES_HOME'), 'F:\\Hermes\\data')
 })
 
@@ -39,10 +30,7 @@ test('parseRegQueryValue returns null when the value line is absent', () => {
 // ── expandWindowsEnvRefs ───────────────────────────────────────────────────
 
 test('expandWindowsEnvRefs expands %VAR% case-insensitively', () => {
-  assert.equal(
-    expandWindowsEnvRefs('%UserProfile%\\h', { USERPROFILE: 'C:\\Users\\jeff' }),
-    'C:\\Users\\jeff\\h'
-  )
+  assert.equal(expandWindowsEnvRefs('%UserProfile%\\h', { USERPROFILE: 'C:\\Users\\jeff' }), 'C:\\Users\\jeff\\h')
 })
 
 test('expandWindowsEnvRefs leaves literal paths and unknown refs intact', () => {

apps/desktop/electron/workspace-cwd.cjs

@@ -14,11 +14,7 @@ function isPackagedInstallPath(dir, { installRoots, isPackaged }) {
     return false
   }
 
-  const roots = new Set(
-    (installRoots ?? [])
-      .filter(Boolean)
-      .map(candidate => path.resolve(String(candidate)))
-  )
+  const roots = new Set((installRoots ?? []).filter(Boolean).map(candidate => path.resolve(String(candidate))))
 
   for (const root of roots) {
     if (resolved === root) {

apps/desktop/electron/workspace-cwd.test.cjs

@@ -13,33 +13,21 @@ const { isPackagedInstallPath } = require('./workspace-cwd.cjs')
 const installRoot = path.resolve('/opt/Hermes')
 
 test('isPackagedInstallPath returns false when not packaged', () => {
-  assert.equal(
-    isPackagedInstallPath(installRoot, { isPackaged: false, installRoots: [installRoot] }),
-    false
-  )
+  assert.equal(isPackagedInstallPath(installRoot, { isPackaged: false, installRoots: [installRoot] }), false)
 })
 
 test('isPackagedInstallPath flags the install root itself', () => {
-  assert.equal(
-    isPackagedInstallPath(installRoot, { isPackaged: true, installRoots: [installRoot] }),
-    true
-  )
+  assert.equal(isPackagedInstallPath(installRoot, { isPackaged: true, installRoots: [installRoot] }), true)
 })
 
 test('isPackagedInstallPath flags paths nested under the install root', () => {
   const nested = path.join(installRoot, 'resources', 'app.asar')
 
-  assert.equal(
-    isPackagedInstallPath(nested, { isPackaged: true, installRoots: [installRoot] }),
-    true
-  )
+  assert.equal(isPackagedInstallPath(nested, { isPackaged: true, installRoots: [installRoot] }), true)
 })
 
 test('isPackagedInstallPath ignores paths outside the install root', () => {
   const homeProject = path.resolve('/home/user/projects/demo')
 
-  assert.equal(
-    isPackagedInstallPath(homeProject, { isPackaged: true, installRoots: [installRoot] }),
-    false
-  )
+  assert.equal(isPackagedInstallPath(homeProject, { isPackaged: true, installRoots: [installRoot] }), false)
 })

apps/desktop/electron/wsl-clipboard-image.cjs

@@ -0,0 +1,92 @@
+// Pull a Windows-host clipboard image from inside WSL2 via PowerShell (WSLg
+// bridges text but not images). Returns PNG bytes or null; exec injectable.
+
+const { execFileSync } = require('node:child_process')
+
+// STA is mandatory: System.Windows.Forms.Clipboard throws ThreadStateException
+// off a single-threaded apartment. We emit base64 (not raw bytes) so the PNG
+// survives stdout's text decoding intact, and write with [Console]::Out.Write
+// to avoid a trailing newline.
+const PS_SCRIPT = [
+  'Add-Type -AssemblyName System.Windows.Forms,System.Drawing',
+  '$img = [System.Windows.Forms.Clipboard]::GetImage()',
+  'if ($null -eq $img) { exit 0 }',
+  '$ms = New-Object System.IO.MemoryStream',
+  '$img.Save($ms, [System.Drawing.Imaging.ImageFormat]::Png)',
+  '[Console]::Out.Write([System.Convert]::ToBase64String($ms.ToArray()))'
+].join('\n')
+
+// PowerShell's -EncodedCommand takes UTF-16LE base64. Encoding the whole script
+// this way sidesteps every layer of WSL→Windows quoting (spaces, quotes,
+// brackets, newlines) that plain -Command arguments would mangle.
+function encodePowerShellCommand(script) {
+  return Buffer.from(String(script), 'utf16le').toString('base64')
+}
+
+// Locate powershell.exe. The bare name resolves through WSL's Windows-interop
+// PATH on every standard WSL2 setup; the absolute fallback covers a stripped
+// PATH. Returns the first candidate — execFile surfaces ENOENT if it's wrong
+// and we fall back to null.
+function powershellCandidates() {
+  return ['powershell.exe', '/mnt/c/Windows/System32/WindowsPowerShell/v1.0/powershell.exe']
+}
+
+function decodeClipboardImageBase64(stdout) {
+  const b64 = String(stdout || '').trim()
+  if (!b64) return null
+
+  let buffer
+  try {
+    buffer = Buffer.from(b64, 'base64')
+  } catch {
+    return null
+  }
+
+  // Guard against partial / garbage output: require a real PNG signature.
+  const PNG_SIGNATURE = Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a])
+  if (buffer.length < PNG_SIGNATURE.length || !buffer.subarray(0, PNG_SIGNATURE.length).equals(PNG_SIGNATURE)) {
+    return null
+  }
+
+  return buffer
+}
+
+// Read the Windows clipboard image from inside WSL. Returns a PNG Buffer, or
+// null when there's no image, PowerShell is unreachable, or output is invalid.
+// Linux-only by contract (caller gates on IS_WSL); never throws.
+function readWslWindowsClipboardImage({ exec = execFileSync, candidates = powershellCandidates() } = {}) {
+  const encoded = encodePowerShellCommand(PS_SCRIPT)
+
+  for (const ps of candidates) {
+    try {
+      const stdout = exec(
+        ps,
+        ['-NoProfile', '-NonInteractive', '-STA', '-ExecutionPolicy', 'Bypass', '-EncodedCommand', encoded],
+        {
+          encoding: 'utf8',
+          windowsHide: true,
+          timeout: 8000,
+          // A 4K screenshot base64s to a few MB; give stdout generous headroom.
+          maxBuffer: 64 * 1024 * 1024,
+          // PowerShell writes progress/CLIXML noise to stderr — ignore it.
+          stdio: ['ignore', 'pipe', 'ignore']
+        }
+      )
+      const decoded = decodeClipboardImageBase64(stdout)
+      if (decoded) return decoded
+      // Empty stdout = no image on the clipboard; stop, don't try fallbacks.
+      if (String(stdout || '').trim() === '') return null
+    } catch {
+      // This powershell.exe candidate is missing/failed — try the next one.
+    }
+  }
+
+  return null
+}
+
+module.exports = {
+  decodeClipboardImageBase64,
+  encodePowerShellCommand,
+  powershellCandidates,
+  readWslWindowsClipboardImage
+}

apps/desktop/electron/wsl-clipboard-image.test.cjs

@@ -0,0 +1,114 @@
+const assert = require('node:assert/strict')
+const test = require('node:test')
+
+const {
+  decodeClipboardImageBase64,
+  encodePowerShellCommand,
+  powershellCandidates,
+  readWslWindowsClipboardImage
+} = require('./wsl-clipboard-image.cjs')
+
+const PNG_SIGNATURE = Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a])
+
+function fakePngBuffer(extraBytes = 16) {
+  return Buffer.concat([PNG_SIGNATURE, Buffer.alloc(extraBytes, 0x42)])
+}
+
+test('encodePowerShellCommand produces UTF-16LE base64 PowerShell can decode', () => {
+  const encoded = encodePowerShellCommand('Write-Output "hi"')
+  const roundTripped = Buffer.from(encoded, 'base64').toString('utf16le')
+  assert.equal(roundTripped, 'Write-Output "hi"')
+})
+
+test('decodeClipboardImageBase64 returns a Buffer for valid PNG base64', () => {
+  const png = fakePngBuffer()
+  const decoded = decodeClipboardImageBase64(png.toString('base64'))
+  assert.ok(Buffer.isBuffer(decoded))
+  assert.ok(decoded.equals(png))
+})
+
+test('decodeClipboardImageBase64 trims surrounding whitespace before decoding', () => {
+  const png = fakePngBuffer()
+  const decoded = decodeClipboardImageBase64(`\n  ${png.toString('base64')}  \r\n`)
+  assert.ok(decoded && decoded.equals(png))
+})
+
+test('decodeClipboardImageBase64 returns null for empty / whitespace input', () => {
+  assert.equal(decodeClipboardImageBase64(''), null)
+  assert.equal(decodeClipboardImageBase64('   \n  '), null)
+  assert.equal(decodeClipboardImageBase64(null), null)
+  assert.equal(decodeClipboardImageBase64(undefined), null)
+})
+
+test('decodeClipboardImageBase64 rejects base64 without a PNG signature', () => {
+  // Valid base64, but the decoded bytes are not a PNG.
+  const notPng = Buffer.from('this is not a png at all').toString('base64')
+  assert.equal(decodeClipboardImageBase64(notPng), null)
+})
+
+test('readWslWindowsClipboardImage decodes the first candidate that returns a PNG', () => {
+  const png = fakePngBuffer()
+  const calls = []
+  const exec = (cmd, args) => {
+    calls.push({ cmd, args })
+    return png.toString('base64')
+  }
+
+  const result = readWslWindowsClipboardImage({ exec, candidates: ['powershell.exe'] })
+  assert.ok(result && result.equals(png))
+  assert.equal(calls.length, 1)
+  assert.equal(calls[0].cmd, 'powershell.exe')
+  // -STA is mandatory for System.Windows.Forms.Clipboard.
+  assert.ok(calls[0].args.includes('-STA'))
+  assert.ok(calls[0].args.includes('-EncodedCommand'))
+})
+
+test('readWslWindowsClipboardImage returns null and stops when stdout is empty (no image)', () => {
+  let count = 0
+  const exec = () => {
+    count += 1
+    return ''
+  }
+
+  const result = readWslWindowsClipboardImage({
+    exec,
+    candidates: ['powershell.exe', '/mnt/c/Windows/System32/WindowsPowerShell/v1.0/powershell.exe']
+  })
+  assert.equal(result, null)
+  // Empty stdout means "no image on the clipboard" — don't probe further candidates.
+  assert.equal(count, 1)
+})
+
+test('readWslWindowsClipboardImage falls through to the next candidate when one throws', () => {
+  const png = fakePngBuffer()
+  const seen = []
+  const exec = cmd => {
+    seen.push(cmd)
+    if (cmd === 'powershell.exe') {
+      throw Object.assign(new Error('not found'), { code: 'ENOENT' })
+    }
+    return png.toString('base64')
+  }
+
+  const result = readWslWindowsClipboardImage({
+    exec,
+    candidates: ['powershell.exe', '/mnt/c/Windows/System32/WindowsPowerShell/v1.0/powershell.exe']
+  })
+  assert.ok(result && result.equals(png))
+  assert.deepEqual(seen, ['powershell.exe', '/mnt/c/Windows/System32/WindowsPowerShell/v1.0/powershell.exe'])
+})
+
+test('readWslWindowsClipboardImage returns null when every candidate throws', () => {
+  const exec = () => {
+    throw new Error('boom')
+  }
+
+  const result = readWslWindowsClipboardImage({ exec, candidates: ['a', 'b'] })
+  assert.equal(result, null)
+})
+
+test('powershellCandidates lists the bare name first, then the absolute fallback', () => {
+  const candidates = powershellCandidates()
+  assert.equal(candidates[0], 'powershell.exe')
+  assert.ok(candidates.some(c => c.endsWith('WindowsPowerShell/v1.0/powershell.exe')))
+})

apps/desktop/package.json

@@ -18,7 +18,7 @@
     "profile:main": "wait-on http://127.0.0.1:5174 && cross-env XCURSOR_SIZE=24 HERMES_DESKTOP_DEV_SERVER=http://127.0.0.1:5174 electron --inspect=9229 .",
     "profile:main:cpu": "wait-on http://127.0.0.1:5174 && cross-env XCURSOR_SIZE=24 NODE_OPTIONS=--cpu-prof HERMES_DESKTOP_DEV_SERVER=http://127.0.0.1:5174 electron .",
     "start": "npm run build && electron .",
-    "build": "node scripts/assert-root-install.cjs && node scripts/write-build-stamp.cjs && node scripts/stage-native-deps.cjs && tsc -b && vite build && npm run postbuild",
+    "build": "node scripts/assert-root-install.cjs && node scripts/write-build-stamp.cjs && node scripts/stage-native-deps.cjs && tsc -b && vite build &&  node scripts/bundle-electron-main.mjs && npm run postbuild",
     "postbuild": "node scripts/assert-dist-built.cjs",
     "prebuilder": "node scripts/patch-electron-builder-mac-binary.cjs",
     "builder": "cross-env NODE_OPTIONS=--max-old-space-size=16384 node scripts/run-electron-builder.cjs",
@@ -37,7 +37,7 @@
     "test:desktop:nsis": "node scripts/test-desktop.mjs nsis",
     "test:desktop:existing": "node scripts/test-desktop.mjs existing",
     "test:desktop:fresh": "node scripts/test-desktop.mjs fresh",
-    "test:desktop:platforms": "node --test electron/bootstrap-platform.test.cjs electron/hardening.test.cjs electron/backend-env.test.cjs electron/backend-probes.test.cjs electron/backend-ready.test.cjs electron/bootstrap-runner.test.cjs electron/connection-config.test.cjs electron/dashboard-token.test.cjs electron/gateway-ws-probe.test.cjs electron/oauth-net-request.test.cjs electron/desktop-uninstall.test.cjs electron/session-windows.test.cjs electron/link-title-window.test.cjs electron/workspace-cwd.test.cjs electron/fs-read-dir.test.cjs electron/git-root.test.cjs electron/windows-child-process.test.cjs electron/update-remote.test.cjs electron/update-count.test.cjs electron/update-rebuild.test.cjs electron/update-marker.test.cjs electron/update-relaunch.test.cjs electron/windows-user-env.test.cjs electron/window-state.test.cjs",
+    "test:desktop:platforms": "node --test electron/bootstrap-platform.test.cjs electron/hardening.test.cjs electron/backend-env.test.cjs electron/backend-probes.test.cjs electron/backend-ready.test.cjs electron/bootstrap-runner.test.cjs electron/connection-config.test.cjs electron/dashboard-token.test.cjs electron/gateway-ws-probe.test.cjs electron/oauth-net-request.test.cjs electron/desktop-uninstall.test.cjs electron/session-windows.test.cjs electron/link-title-window.test.cjs electron/workspace-cwd.test.cjs electron/fs-read-dir.test.cjs electron/git-root.test.cjs electron/git-worktree-ops.test.cjs electron/windows-child-process.test.cjs electron/update-remote.test.cjs electron/update-count.test.cjs electron/update-rebuild.test.cjs electron/update-marker.test.cjs electron/update-relaunch.test.cjs electron/windows-user-env.test.cjs electron/wsl-clipboard-image.test.cjs electron/titlebar-overlay-width.test.cjs electron/window-state.test.cjs",
     "typecheck": "tsc -p . --noEmit",
     "lint": "eslint src/ electron/",
     "lint:fix": "eslint src/ electron/ --fix",
@@ -51,11 +51,17 @@
     "@assistant-ui/react-streamdown": "^0.1.11",
     "@audiowave/react": "^0.6.2",
     "@chenglou/pretext": "^0.0.6",
+    "@codemirror/commands": "^6.10.4",
+    "@codemirror/language": "^6.12.4",
+    "@codemirror/language-data": "^6.5.2",
+    "@codemirror/state": "^6.7.0",
+    "@codemirror/view": "^6.43.3",
     "@dnd-kit/core": "^6.3.1",
     "@dnd-kit/sortable": "^10.0.0",
     "@dnd-kit/utilities": "^3.2.2",
     "@hermes/shared": "file:../shared",
     "@icons-pack/react-simple-icons": "=13.11.1",
+    "@lezer/highlight": "^1.2.3",
     "@nanostores/react": "^1.1.0",
     "@nous-research/ui": "^0.13.0",
     "@radix-ui/react-slot": "^1.2.4",
@@ -75,11 +81,13 @@
     "clsx": "^2.1.1",
     "cmdk": "^1.1.1",
     "dnd-core": "^14.0.1",
+    "dompurify": "^3.4.11",
     "hast-util-from-html-isomorphic": "^2.0.0",
     "hast-util-to-text": "^4.0.2",
     "ignore": "^7.0.5",
     "katex": "^0.16.45",
     "leva": "^0.10.1",
+    "mermaid": "^11.15.0",
     "motion": "^12.38.0",
     "nanostores": "^1.3.0",
     "node-pty": "1.1.0",
@@ -93,6 +101,7 @@
     "remark-math": "^6.0.0",
     "remend": "^1.3.0",
     "shiki": "^4.0.2",
+    "simple-git": "^3.36.0",
     "streamdown": "^2.5.0",
     "tailwind-merge": "^3.5.0",
     "tailwindcss": "^4.2.4",

apps/desktop/scripts/bundle-electron-main.mjs

@@ -0,0 +1,33 @@
+#!/usr/bin/env node
+// bundle-electron-main.mjs — bundles electron/main.cjs into a single
+// self-contained file so the nix build doesn't need to ship node_modules/.
+//
+// `electron` is provided by the runtime; `node-pty` is staged separately
+// via stage-native-deps.cjs.  `preload.cjs` is NOT require()'d by main —
+// Electron loads it via path.join(__dirname, 'preload.cjs') — so it stays
+// as a separate file and doesn't need bundling.
+import { build } from 'esbuild'
+import { resolve, dirname } from 'node:path'
+import { fileURLToPath } from 'node:url'
+import { renameSync } from 'node:fs'
+
+const here = dirname(fileURLToPath(import.meta.url))
+const root = resolve(here, '..')
+const entry = resolve(root, 'electron/main.cjs')
+const tmp = resolve(root, 'electron/main.bundled.cjs')
+
+await build({
+  entryPoints: [entry],
+  bundle: true,
+  platform: 'node',
+  format: 'cjs',
+  target: 'node20',
+  outfile: tmp,
+  external: ['electron', 'node-pty'],
+  logLevel: 'info'
+})
+
+// Overwrite the original with the bundled version.
+renameSync(tmp, entry)
+
+console.log(`bundled ${entry}`)

apps/desktop/src/app/agents/index.tsx

@@ -3,10 +3,11 @@ import { type ReactNode, useEffect, useMemo, useState } from 'react'
 
 import { useElapsedSeconds } from '@/components/chat/activity-timer'
 import { ActivityTimerText } from '@/components/chat/activity-timer-text'
+import { Codicon } from '@/components/ui/codicon'
 import { FadeText } from '@/components/ui/fade-text'
 import { GlyphSpinner } from '@/components/ui/glyph-spinner'
 import { type Translations, useI18n } from '@/i18n'
-import { AlertCircle, CheckCircle2, Sparkles } from '@/lib/icons'
+import { AlertCircle, CheckCircle2 } from '@/lib/icons'
 import { useEnterAnimation } from '@/lib/use-enter-animation'
 import { cn } from '@/lib/utils'
 import {
@@ -209,7 +210,7 @@ function SubagentTree({ tree }: { tree: SubagentNode[] }) {
   if (tree.length === 0) {
     return (
       <div className="grid place-items-center gap-3 py-12 text-center">
-        <Sparkles className="size-6 text-muted-foreground/60" />
+        <Codicon className="text-muted-foreground/60" name="hubot" size="1.5rem" />
         <p className="text-sm font-medium text-foreground/90">{t.agents.emptyTitle}</p>
         <p className="max-w-md text-xs leading-relaxed text-muted-foreground/75">{t.agents.emptyDesc}</p>
       </div>

apps/desktop/src/app/artifacts/index.tsx

@@ -477,17 +477,20 @@ export function ArtifactsView({ setStatusbarItemGroup: _setStatusbarItemGroup, .
     }
   }, [artifacts])
 
-  const openArtifact = useCallback(async (href: string) => {
-    try {
-      if (window.hermesDesktop?.openExternal) {
-        await window.hermesDesktop.openExternal(href)
-      } else {
-        window.open(href, '_blank', 'noopener,noreferrer')
+  const openArtifact = useCallback(
+    async (href: string) => {
+      try {
+        if (window.hermesDesktop?.openExternal) {
+          await window.hermesDesktop.openExternal(href)
+        } else {
+          window.open(href, '_blank', 'noopener,noreferrer')
+        }
+      } catch (err) {
+        notifyError(err, a.openFailed)
       }
-    } catch (err) {
-      notifyError(err, a.openFailed)
-    }
-  }, [a])
+    },
+    [a]
+  )
 
   const markImageFailed = useCallback((id: string) => {
     setFailedImageIds(current => {
@@ -839,7 +842,8 @@ const ARTIFACT_COLUMNS: readonly ArtifactColumn[] = [
   {
     Cell: PrimaryCell,
     bodyClassName: 'p-0',
-    header: (filter, a) => (filter === 'link' ? a.colTitleLink : filter === 'file' ? a.colTitleFile : a.colTitleDefault),
+    header: (filter, a) =>
+      filter === 'link' ? a.colTitleLink : filter === 'file' ? a.colTitleFile : a.colTitleDefault,
     id: 'primary',
     width: filter => (filter === 'link' ? 'w-[50%]' : 'w-[35%]')
   },

apps/desktop/src/app/chat/composer/attachments.test.tsx

@@ -2,9 +2,9 @@ import { cleanup, render, screen } from '@testing-library/react'
 import { afterEach, describe, expect, it } from 'vitest'
 
 import { I18nProvider } from '@/i18n/context'
+import type { ComposerAttachment } from '@/store/composer'
 
 import { AttachmentList } from './attachments'
-import type { ComposerAttachment } from '@/store/composer'
 
 function makeAttachment(id: string, label = 'test.pdf'): ComposerAttachment {
   return { id, kind: 'file', label }
@@ -32,7 +32,10 @@ describe('AttachmentList', () => {
 
   it('renders empty list without error', () => {
     renderWithI18n(<AttachmentList attachments={[]} />)
-    const container = screen.getByTestId?.('composer-attachments') ?? document.querySelector('[data-slot="composer-attachments"]')
+
+    const container =
+      screen.getByTestId?.('composer-attachments') ?? document.querySelector('[data-slot="composer-attachments"]')
+
     expect(container).toBeDefined()
   })
 
@@ -55,10 +58,7 @@ describe('AttachmentList', () => {
   })
 
   it('does not crash when attachments array contains null entries', () => {
-    const attachments = [
-      null as unknown as ComposerAttachment,
-      makeAttachment('a', 'valid.txt')
-    ]
+    const attachments = [null as unknown as ComposerAttachment, makeAttachment('a', 'valid.txt')]
 
     expect(() => {
       renderWithI18n(<AttachmentList attachments={attachments} />)

apps/desktop/src/app/chat/composer/context-menu.tsx

@@ -73,7 +73,11 @@ export function ContextMenu({
           <ContextMenuItem disabled={!onPickImages} icon={ImageIcon} onSelect={onPickImages}>
             {c.images}
           </ContextMenuItem>
-          <ContextMenuItem disabled={!onPasteClipboardImage} icon={Clipboard} onSelect={onPasteClipboardImage}>
+          <ContextMenuItem
+            disabled={!onPasteClipboardImage}
+            icon={Clipboard}
+            onSelect={onPasteClipboardImage ? () => void onPasteClipboardImage() : undefined}
+          >
             {c.pasteImage}
           </ContextMenuItem>
           <ContextMenuItem icon={Link} onSelect={onOpenUrlDialog}>
@@ -167,7 +171,7 @@ interface ContextMenuItemProps {
 interface ContextMenuProps {
   onInsertText: (text: string) => void
   onOpenUrlDialog: () => void
-  onPasteClipboardImage?: () => void
+  onPasteClipboardImage?: (opts?: { silent?: boolean }) => Promise<boolean> | void
   onPickFiles?: () => void
   onPickFolders?: () => void
   onPickImages?: () => void

apps/desktop/src/app/chat/composer/enter-submit-dom-race.test.tsx

@@ -59,8 +59,10 @@ function Harness({
     }
 
     const editor = editorRef.current
+
     if (editor) {
       const domText = composerPlainText(editor)
+
       if (domText !== draftRef.current) {
         draftRef.current = domText
         setDraft(domText)
@@ -127,9 +129,11 @@ function Harness({
 describe('composer Enter submit — live DOM vs stale composer state (#39630)', () => {
   it('sends the just-typed text on Enter even when composer state has not synced', async () => {
     const onSubmit = vi.fn()
+
     const { getByTestId } = render(
       <Harness onCancel={vi.fn()} onDrain={vi.fn()} onQueue={vi.fn()} onSubmit={onSubmit} />
     )
+
     const editor = getByTestId('editor')
 
     // Fast typing: the DOM has the text but NO input event fired, so `draft`
@@ -146,9 +150,11 @@ describe('composer Enter submit — live DOM vs stale composer state (#39630)',
     const onQueue = vi.fn()
     const onDrain = vi.fn()
     const onCancel = vi.fn()
+
     const { getByTestId } = render(
       <Harness busy onCancel={onCancel} onDrain={onDrain} onQueue={onQueue} onSubmit={vi.fn()} queued={['queued-1']} />
     )
+
     const editor = getByTestId('editor')
 
     await act(async () => {
@@ -165,9 +171,11 @@ describe('composer Enter submit — live DOM vs stale composer state (#39630)',
     const onCancel = vi.fn()
     const onSubmit = vi.fn()
     const onQueue = vi.fn()
+
     const { getByTestId } = render(
       <Harness busy onCancel={onCancel} onDrain={vi.fn()} onQueue={onQueue} onSubmit={onSubmit} />
     )
+
     const editor = getByTestId('editor')
 
     await act(async () => {
@@ -183,9 +191,11 @@ describe('composer Enter submit — live DOM vs stale composer state (#39630)',
   it('drains the next queued prompt on Enter when idle with a truly empty editor', async () => {
     const onDrain = vi.fn()
     const onSubmit = vi.fn()
+
     const { getByTestId } = render(
       <Harness onCancel={vi.fn()} onDrain={onDrain} onQueue={vi.fn()} onSubmit={onSubmit} queued={['queued-1']} />
     )
+
     const editor = getByTestId('editor')
 
     await act(async () => {
@@ -200,9 +210,18 @@ describe('composer Enter submit — live DOM vs stale composer state (#39630)',
   it('keeps reconnect drafts editable but blocks Enter submit until the gateway returns', async () => {
     const onSubmit = vi.fn()
     const onDrain = vi.fn()
+
     const { getByTestId } = render(
-      <Harness disabled onCancel={vi.fn()} onDrain={onDrain} onQueue={vi.fn()} onSubmit={onSubmit} queued={['queued-1']} />
+      <Harness
+        disabled
+        onCancel={vi.fn()}
+        onDrain={onDrain}
+        onQueue={vi.fn()}
+        onSubmit={onSubmit}
+        queued={['queued-1']}
+      />
     )
+
     const editor = getByTestId('editor')
 
     await act(async () => {

apps/desktop/src/app/chat/composer/focus.ts

@@ -10,8 +10,8 @@
  * steal focus from the composer effect.
  */
 
-import { RICH_INPUT_SLOT } from './rich-editor'
 import type { InlineRefInput } from './inline-refs'
+import { RICH_INPUT_SLOT } from './rich-editor'
 
 export type ComposerTarget = 'edit' | 'main'
 export type ComposerInsertMode = 'block' | 'inline'
@@ -34,8 +34,14 @@ interface InsertRefsDetail {
 const FOCUS_EVENT = 'hermes:composer-focus'
 const INSERT_EVENT = 'hermes:composer-insert'
 const INSERT_REFS_EVENT = 'hermes:composer-insert-refs'
+const SUBMIT_EVENT = 'hermes:composer-submit'
 const VOICE_TOGGLE_EVENT = 'hermes:composer-voice-toggle'
 
+interface SubmitDetail {
+  target: ComposerTarget
+  text: string
+}
+
 let activeTarget: ComposerTarget = 'main'
 
 const resolve = (target: ComposerTarget | 'active') => (target === 'active' ? activeTarget : target)
@@ -106,6 +112,23 @@ export const requestComposerInsertRefs = (
 export const onComposerInsertRefsRequest = (handler: (detail: InsertRefsDetail) => void) =>
   subscribe<InsertRefsDetail>(INSERT_REFS_EVENT, handler)
 
+/** Submit a prompt through a composer as if the user typed + sent it. Lets
+ * external panels (e.g. the review pane's "let the agent ship it" button) hand
+ * the agent a task without the user round-tripping through the input. */
+export const requestComposerSubmit = (
+  text: string,
+  { target = 'active' }: { target?: ComposerTarget | 'active' } = {}
+) => {
+  const trimmed = text.trim()
+
+  if (trimmed) {
+    dispatch<SubmitDetail>(SUBMIT_EVENT, { target: resolve(target), text: trimmed })
+  }
+}
+
+export const onComposerSubmitRequest = (handler: (detail: SubmitDetail) => void) =>
+  subscribe<SubmitDetail>(SUBMIT_EVENT, handler)
+
 /** Toggle the active composer's voice conversation — the `composer.voice`
  *  hotkey (Ctrl+B) reaching into the composer that owns the voice state. */
 export const requestVoiceToggle = () => dispatch<{ at: number }>(VOICE_TOGGLE_EVENT, { at: Date.now() })

apps/desktop/src/app/chat/composer/help-hint.tsx

@@ -33,7 +33,7 @@ export function HelpHint() {
 
       <Section title={c.hotkeys}>
         {COMPOSER_HOTKEY_ROWS.map(row => (
-          <HotkeyRow description={c.hotkeyDescs[row.id] ?? ''} combos={[...row.combos]} key={row.id} />
+          <HotkeyRow combos={[...row.combos]} description={c.hotkeyDescs[row.id] ?? ''} key={row.id} />
         ))}
       </Section>
 

apps/desktop/src/app/chat/composer/hooks/use-mic-recorder.ts

@@ -59,7 +59,11 @@ function micError(error: unknown, copy: MicRecorderErrorCopy): Error {
   return new Error(copy.microphoneStartFailed)
 }
 
-export function useMicRecorder(copy: MicRecorderErrorCopy): { handle: MicRecorderHandle; level: number; recording: boolean } {
+export function useMicRecorder(copy: MicRecorderErrorCopy): {
+  handle: MicRecorderHandle
+  level: number
+  recording: boolean
+} {
   const [level, setLevel] = useState(0)
   const [recording, setRecording] = useState(false)
 

apps/desktop/src/app/chat/composer/hooks/use-popout-drag.ts

@@ -1,19 +1,12 @@
-import {
-  type PointerEvent as ReactPointerEvent,
-  type RefObject,
-  useCallback,
-  useEffect,
-  useRef,
-  useState
-} from 'react'
+import { type PointerEvent as ReactPointerEvent, type RefObject, useCallback, useEffect, useRef, useState } from 'react'
 
 import {
   POPOUT_ESTIMATED_HEIGHT,
   POPOUT_WIDTH_REM,
-  readPopoutBounds,
-  setComposerPopoutPosition,
   type PopoutPosition,
-  type PopoutSize
+  type PopoutSize,
+  readPopoutBounds,
+  setComposerPopoutPosition
 } from '@/store/composer-popout'
 
 // Floating surface long-press before it becomes draggable (the 5px platform drags
@@ -80,6 +73,7 @@ function dockProximityOf(rect: DOMRect) {
   const verticalGap = window.innerHeight - DOCK_ZONE_BOTTOM_PX - rect.bottom
 
   const v = verticalGap <= 0 ? 1 : Math.max(0, 1 - verticalGap / DOCK_VERTICAL_FALLOFF_PX)
+
   const h =
     horizontalDist <= DOCK_ZONE_CENTER_TOLERANCE_PX
       ? 1

apps/desktop/src/app/chat/composer/hooks/use-slash-completions.ts

@@ -98,12 +98,14 @@ export function useSlashCompletions(options: {
 
         const matches = (
           needle
-            ? $sessions.get().filter(
-                session =>
-                  sessionTitle(session).toLowerCase().includes(needle) ||
-                  (session.preview ?? '').toLowerCase().includes(needle) ||
-                  session.id.toLowerCase().includes(needle)
-              )
+            ? $sessions
+                .get()
+                .filter(
+                  session =>
+                    sessionTitle(session).toLowerCase().includes(needle) ||
+                    (session.preview ?? '').toLowerCase().includes(needle) ||
+                    session.id.toLowerCase().includes(needle)
+                )
             : $sessions.get()
         ).slice(0, SESSION_INLINE_LIMIT)
 
@@ -135,9 +137,7 @@ export function useSlashCompletions(options: {
           // Prefer the categorized layout so the popover renders section headers
           // (Session, Tools & Skills, ...). Fall back to the flat list when the
           // backend didn't categorize.
-          const sections = catalog.categories?.length
-            ? catalog.categories
-            : [{ name: '', pairs: catalog.pairs ?? [] }]
+          const sections = catalog.categories?.length ? catalog.categories : [{ name: '', pairs: catalog.pairs ?? [] }]
 
           const items = sections.flatMap(section =>
             section.pairs.map(([command, meta]) => ({
@@ -151,10 +151,9 @@ export function useSlashCompletions(options: {
           return { items, query }
         }
 
-        const result = await gateway.request<{ items?: CompletionEntry[]; replace_from?: number }>(
-          'complete.slash',
-          { text }
-        )
+        const result = await gateway.request<{ items?: CompletionEntry[]; replace_from?: number }>('complete.slash', {
+          text
+        })
 
         // Arg-completion items (replace_from > 1) carry just the arg stub —
         // e.g. complete.slash returns `{text: "alice"}` for `/personality alic`

apps/desktop/src/app/chat/composer/hooks/use-voice-conversation.ts

@@ -220,22 +220,25 @@ export function useVoiceConversation({
     }
   }, [handle, handleTurn, onFatalError, voiceCopy.couldNotStartSession, voiceCopy.microphoneFailed])
 
-  const speak = useCallback(async (text: string) => {
-    setStatus('speaking')
+  const speak = useCallback(
+    async (text: string) => {
+      setStatus('speaking')
 
-    try {
-      await playSpeechText(text, { source: 'voice-conversation' })
-    } catch (error) {
-      notifyError(error, voiceCopy.playbackFailed)
-    } finally {
-      if (enabledRef.current) {
-        pendingStartRef.current = true
-        setStatus('idle')
-      } else {
-        setStatus('idle')
+      try {
+        await playSpeechText(text, { source: 'voice-conversation' })
+      } catch (error) {
+        notifyError(error, voiceCopy.playbackFailed)
+      } finally {
+        if (enabledRef.current) {
+          pendingStartRef.current = true
+          setStatus('idle')
+        } else {
+          setStatus('idle')
+        }
       }
-    }
-  }, [voiceCopy.playbackFailed])
+    },
+    [voiceCopy.playbackFailed]
+  )
 
   const start = useCallback(async () => {
     if (!onTranscribeAudio) {
@@ -255,7 +258,14 @@ export function useVoiceConversation({
     consumePendingResponse()
     pendingStartRef.current = true
     await startListening()
-  }, [consumePendingResponse, onFatalError, onTranscribeAudio, startListening, voiceCopy.configureSpeechToText, voiceCopy.unavailable])
+  }, [
+    consumePendingResponse,
+    onFatalError,
+    onTranscribeAudio,
+    startListening,
+    voiceCopy.configureSpeechToText,
+    voiceCopy.unavailable
+  ])
 
   const end = useCallback(async () => {
     pendingStartRef.current = false

apps/desktop/src/app/chat/composer/index.tsx

@@ -45,8 +45,8 @@ import {
   $composerPoppedOut,
   POPOUT_WIDTH_REM,
   readPopoutBounds,
-  setComposerPoppedOut,
-  setComposerPopoutPosition
+  setComposerPopoutPosition,
+  setComposerPoppedOut
 } from '@/store/composer-popout'
 import {
   $queuedPromptsBySession,
@@ -60,8 +60,11 @@ import {
   updateQueuedPrompt
 } from '@/store/composer-queue'
 import { $statusItemsBySession } from '@/store/composer-status'
-import { $previewStatusBySession } from '@/store/preview-status'
 import { notify } from '@/store/notifications'
+import { $previewStatusBySession } from '@/store/preview-status'
+import { listRepoBranches, requestStartWorkSession, startWorkInRepo, switchBranchInRepo } from '@/store/projects'
+import { $activeSessionAwaitingInput } from '@/store/prompts'
+import { toggleReview } from '@/store/review'
 import { $gatewayState, $messages, setSessionPickerOpen } from '@/store/session'
 import { $threadScrolledUp } from '@/store/thread-scroll'
 import { isSecondaryWindow } from '@/store/windows'
@@ -80,6 +83,7 @@ import {
   onComposerFocusRequest,
   onComposerInsertRefsRequest,
   onComposerInsertRequest,
+  onComposerSubmitRequest,
   onComposerVoiceToggleRequest
 } from './focus'
 import { HelpHint } from './help-hint'
@@ -108,6 +112,7 @@ import {
   slashChipElement
 } from './rich-editor'
 import { ComposerStatusStack } from './status-stack'
+import { CodingStatusRow } from './status-stack/coding-row'
 import { detectTrigger, extractClipboardImageBlobs, textBeforeCaret, type TriggerState } from './text-utils'
 import { ComposerTriggerPopover } from './trigger-popover'
 import type { ChatBarProps } from './types'
@@ -225,6 +230,11 @@ export function ChatBar({
   const statusItemsBySession = useStore($statusItemsBySession)
   const previewStatusBySession = useStore($previewStatusBySession)
   const scrolledUp = useStore($threadScrolledUp)
+  // The turn is parked on the user (clarify / approval / sudo / secret). Esc must
+  // not interrupt it — there's nothing actively running to stop, and stopping
+  // would discard a question the user may want to come back to. The blocking
+  // prompt owns its own dismissal (Skip, Reject, dialog close).
+  const awaitingInput = useStore($activeSessionAwaitingInput)
   // Pop-out is a shared, persisted state — but secondary windows (the Ctrl+Shift+N
   // tiny window, subagent watch windows) always start docked and can't pop out:
   // a floating composer makes no sense in a single-session side window, and it
@@ -274,14 +284,17 @@ export function ChatBar({
     poppedOut ? handleComposerDock() : handleComposerPopOut()
   }, [handleComposerDock, handleComposerPopOut, poppedOut])
 
-  const { dockProximity, dragging, onPointerDown: onComposerGesturePointerDown } =
-    useComposerPopoutGestures({
-      composerRef,
-      onDock: handleComposerDock,
-      onPopOut: handleComposerPopOut,
-      poppedOut,
-      position: popoutPosition
-    })
+  const {
+    dockProximity,
+    dragging,
+    onPointerDown: onComposerGesturePointerDown
+  } = useComposerPopoutGestures({
+    composerRef,
+    onDock: handleComposerDock,
+    onPopOut: handleComposerPopOut,
+    poppedOut,
+    position: popoutPosition
+  })
 
   const draftRef = useRef(draft)
   const pendingDraftPersistRef = useRef<{ scope: string | null; text: string } | null>(null)
@@ -780,6 +793,16 @@ export function ChatBar({
     if (!pastedText) {
       event.preventDefault()
 
+      // Under WSL2/WSLg the Windows host clipboard doesn't bridge *images* to
+      // the Linux clipboard the DOM paste event reads, so a host screenshot
+      // arrives as an empty paste (no blobs, no text). Fall back to the main
+      // process, which pulls the image straight off the Windows clipboard.
+      // Silent so a genuinely-empty paste doesn't pop a "no image" warning.
+      if (onPasteClipboardImage) {
+        triggerHaptic('selection')
+        void onPasteClipboardImage({ silent: true })
+      }
+
       return
     }
 
@@ -812,8 +835,7 @@ export function ChatBar({
   // Suppress the "No matches" empty state once a slash command is past its name:
   // a no-arg command has nothing to offer, and a fully-typed arg commits on
   // Space/Tab — neither should dead-end on a popover.
-  const argStageEmpty =
-    trigger?.kind === '/' && slashArgStage(trigger.query) && !triggerLoading && !triggerItems.length
+  const argStageEmpty = trigger?.kind === '/' && slashArgStage(trigger.query) && !triggerLoading && !triggerItems.length
 
   const closeTrigger = () => {
     setTrigger(null)
@@ -840,7 +862,14 @@ export function ChatBar({
       id: text,
       type: 'slash',
       label: text.slice(1),
-      metadata: { command: slashCommandToken(trigger.query), display: text, meta: '', group: '', action: '', rawText: text }
+      metadata: {
+        command: slashCommandToken(trigger.query),
+        display: text,
+        meta: '',
+        group: '',
+        action: '',
+        rawText: text
+      }
     })
   }
 
@@ -980,10 +1009,7 @@ export function ChatBar({
 
     // Non-collapsed Backspace/Delete: native selection-delete is ~O(n²) on large
     // drafts (Ctrl+A → Delete froze ~1.3s). Collapsed carets fall through.
-    if (
-      (event.key === 'Backspace' || event.key === 'Delete') &&
-      deleteSelectionInEditor(event.currentTarget)
-    ) {
+    if ((event.key === 'Backspace' || event.key === 'Delete') && deleteSelectionInEditor(event.currentTarget)) {
       event.preventDefault()
       flushEditorToDraft(event.currentTarget)
 
@@ -1194,8 +1220,10 @@ export function ChatBar({
         return
       }
 
-      // Otherwise Esc interrupts the running turn (Stop-button parity).
-      if (busy) {
+      // Otherwise Esc interrupts the running turn (Stop-button parity) — unless
+      // the turn is parked waiting on the user, where Esc must not discard the
+      // pending prompt.
+      if (busy && !awaitingInput) {
         event.preventDefault()
         triggerHaptic('cancel')
         void Promise.resolve(onCancel())
@@ -1351,6 +1379,80 @@ export function ChatBar({
     }
   }, [setComposerText])
 
+  // Hand a worktree off to the controller: open a fresh session anchored there,
+  // carrying the composer draft as its first turn. Clearing here means the draft
+  // travels to the new session instead of getting stashed under this one.
+  const openInWorktree = useCallback(
+    (path: string) => {
+      const text = draftRef.current
+      clearDraft()
+      clearComposerAttachments()
+      requestStartWorkSession(path, text)
+    },
+    [clearDraft]
+  )
+
+  // Branch off into a NEW worktree (base = branch name, or current HEAD). A
+  // create failure throws back to the row (which toasts) before we touch the
+  // draft; a missing cwd / remote backend no-ops (the row hides the affordance).
+  const handleBranchOff = useCallback(
+    async (branch: string, base?: string) => {
+      const repoPath = cwd?.trim()
+      const result = repoPath && (await startWorkInRepo(repoPath, { base, branch, name: branch }))
+
+      if (result) {
+        openInWorktree(result.path)
+      }
+    },
+    [cwd, openInWorktree]
+  )
+
+  // Convert an EXISTING branch into a fresh worktree + session (no new branch).
+  // Mirrors handleBranchOff's hand-off: create the worktree, then open a session
+  // anchored there carrying the draft.
+  const handleConvertBranch = useCallback(
+    async (branch: string, path?: null | string, isDefault?: boolean) => {
+      if (path?.trim()) {
+        openInWorktree(path)
+
+        return
+      }
+
+      const repoPath = cwd?.trim()
+
+      if (repoPath && isDefault) {
+        await switchBranchInRepo(repoPath, branch)
+        openInWorktree(repoPath)
+
+        return
+      }
+
+      const result = repoPath && (await startWorkInRepo(repoPath, { existingBranch: branch }))
+
+      if (result) {
+        openInWorktree(result.path)
+      }
+    },
+    [cwd, openInWorktree]
+  )
+
+  const handleListBranches = useCallback(async () => {
+    const repoPath = cwd?.trim()
+
+    return repoPath ? listRepoBranches(repoPath) : []
+  }, [cwd])
+
+  const handleSwitchBranch = useCallback(
+    async (branch: string) => {
+      const repoPath = cwd?.trim()
+
+      if (repoPath) {
+        await switchBranchInRepo(repoPath, branch)
+      }
+    },
+    [cwd]
+  )
+
   const loadIntoComposer = (text: string, attachments: ComposerAttachment[]) => {
     draftRef.current = text
     setComposerText(text)
@@ -1674,6 +1776,46 @@ export function ChatBar({
     }
   }, [autoDrainNext, busy, queuedPrompts.length])
 
+  // Esc cancels the in-flight turn when the CHAT has focus — not just the
+  // composer input (which has its own handler above). Clicking into the
+  // transcript and hitting Esc now stops the run, matching the Stop button.
+  // Intentional only: we bail if (a) the composer/another field already
+  // handled Esc (defaultPrevented), (b) focus is in any input/textarea/
+  // contenteditable (you're typing, not stopping), or (c) a dialog/popover is
+  // open — Esc must close that overlay, never double as canceling the stream
+  // behind it. A latest-handler ref keeps the listener registered once.
+  const escCancelRef = useRef<(event: globalThis.KeyboardEvent) => void>(() => {})
+
+  escCancelRef.current = (event: globalThis.KeyboardEvent) => {
+    // `awaitingInput`: the turn is parked on a clarify / approval / sudo / secret
+    // prompt, which owns Esc (or is meant to persist) — never cancel the stream
+    // out from under it.
+    if (event.key !== 'Escape' || event.defaultPrevented || !busy || awaitingInput) {
+      return
+    }
+
+    const active = document.activeElement as HTMLElement | null
+
+    if (active && (active.tagName === 'INPUT' || active.tagName === 'TEXTAREA' || active.isContentEditable)) {
+      return
+    }
+
+    if (document.querySelector('[role="dialog"],[role="alertdialog"],[data-radix-popper-content-wrapper]')) {
+      return
+    }
+
+    event.preventDefault()
+    triggerHaptic('cancel')
+    void Promise.resolve(onCancel())
+  }
+
+  useEffect(() => {
+    const onKeyDown = (event: globalThis.KeyboardEvent) => escCancelRef.current(event)
+    window.addEventListener('keydown', onKeyDown)
+
+    return () => window.removeEventListener('keydown', onKeyDown)
+  }, [])
+
   // Queue-edit cleanup: on session swap the scope effect already stashed the
   // edit snapshot; only restore into the composer when still on the same scope.
   useEffect(() => {
@@ -1706,6 +1848,22 @@ export function ChatBar({
       .catch(restore)
   }
 
+  // External "submit this prompt" requests (e.g. the review pane's agent-ship
+  // button) route through the same send path. A ref keeps the listener stable
+  // while always calling the latest dispatchSubmit closure.
+  const dispatchSubmitRef = useRef(dispatchSubmit)
+  dispatchSubmitRef.current = dispatchSubmit
+
+  useEffect(
+    () =>
+      onComposerSubmitRequest(({ target, text }) => {
+        if (target === 'main' && !inputDisabled) {
+          dispatchSubmitRef.current(text)
+        }
+      }),
+    [inputDisabled]
+  )
+
   const submitDraft = () => {
     if (disabled) {
       return
@@ -2099,7 +2257,7 @@ export function ChatBar({
           <div className="relative w-full rounded-[inherit]">
             <div
               className={cn(
-                'group/composer-surface relative z-4 isolate rounded-[inherit] border border-[color-mix(in_srgb,var(--dt-composer-ring)_calc(18%*var(--composer-ring-strength)),var(--dt-input))] transition-[border-color] duration-200 ease-out focus-within:border-[color-mix(in_srgb,var(--dt-composer-ring)_calc(45%*var(--composer-ring-strength)),transparent)]',
+                'group/composer-surface relative z-4 isolate grid grid-rows-[auto_1fr] overflow-hidden rounded-[inherit] border border-[color-mix(in_srgb,var(--dt-composer-ring)_calc(18%*var(--composer-ring-strength)),var(--dt-input))]',
                 COMPOSER_DROP_FADE_CLASS,
                 dragActive && COMPOSER_DROP_ACTIVE_CLASS
               )}
@@ -2114,10 +2272,20 @@ export function ChatBar({
                   composerSurfaceGlass
                 )}
               />
+              <CodingStatusRow
+                onBranchOff={handleBranchOff}
+                onConvertBranch={handleConvertBranch}
+                onListBranches={handleListBranches}
+                onOpen={toggleReview}
+                onOpenWorktree={openInWorktree}
+                onSwitchBranch={handleSwitchBranch}
+              />
               <div
                 className={cn(
                   'relative z-1 flex min-h-0 w-full flex-col gap-(--composer-row-gap) overflow-hidden rounded-[inherit] px-(--composer-surface-pad-x) py-(--composer-surface-pad-y) transition-opacity duration-200 ease-out',
-                  scrolledUp ? 'opacity-30 group-hover/composer:opacity-100 group-focus-within/composer-surface:opacity-100' : 'opacity-100'
+                  scrolledUp
+                    ? 'opacity-30 group-hover/composer:opacity-100 group-focus-within/composer-surface:opacity-100'
+                    : 'opacity-100'
                 )}
                 data-slot="composer-fade"
               >

apps/desktop/src/app/chat/composer/inline-refs.ts

@@ -3,12 +3,7 @@ import { contextPath } from '@/lib/chat-runtime'
 
 import type { DroppedFile } from '../hooks/use-composer-actions'
 
-import {
-  composerPlainText,
-  normalizeComposerEditorDom,
-  placeCaretEnd,
-  refChipElement
-} from './rich-editor'
+import { composerPlainText, normalizeComposerEditorDom, placeCaretEnd, refChipElement } from './rich-editor'
 
 /** A chip to insert: a raw `@kind:value` string, or a typed value + display label. */
 export type InlineRefInput = string | { kind: string; label?: string; value: string }
@@ -159,6 +154,7 @@ export function insertInlineRefsIntoEditor(editor: HTMLDivElement, refs: readonl
   editor.focus({ preventScroll: true })
 
   const selection = window.getSelection()
+
   const range =
     selection?.rangeCount && editor.contains(selection.getRangeAt(0).commonAncestorContainer)
       ? selection.getRangeAt(0)

apps/desktop/src/app/chat/composer/model-pill.tsx

@@ -94,13 +94,7 @@ export function ModelPill({
     <DropdownMenu onOpenChange={setOpen} open={open}>
       <Tip label={title} side="top">
         <DropdownMenuTrigger asChild>
-          <Button
-            aria-label={title}
-            className={pillClass}
-            disabled={disabled}
-            type="button"
-            variant="ghost"
-          >
+          <Button aria-label={title} className={pillClass} disabled={disabled} type="button" variant="ghost">
             {label}
           </Button>
         </DropdownMenuTrigger>

apps/desktop/src/app/chat/composer/status-stack/coding-row.tsx

@@ -0,0 +1,469 @@
+import { useStore } from '@nanostores/react'
+import { memo, useCallback, useEffect, useRef, useState } from 'react'
+
+import { StatusRow } from '@/components/chat/status-row'
+import { Button } from '@/components/ui/button'
+import { Codicon } from '@/components/ui/codicon'
+import { Command, CommandEmpty, CommandGroup, CommandInput, CommandItem, CommandList } from '@/components/ui/command'
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle
+} from '@/components/ui/dialog'
+import { DiffCount } from '@/components/ui/diff-count'
+import {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuLabel,
+  DropdownMenuSeparator,
+  DropdownMenuTrigger
+} from '@/components/ui/dropdown-menu'
+import { SanitizedInput } from '@/components/ui/sanitized-input'
+import type { HermesGitBranch } from '@/global'
+import { useI18n } from '@/i18n'
+import { gitRef } from '@/lib/sanitize'
+import { $repoStatus, $repoWorktrees } from '@/store/coding-status'
+import { notifyError } from '@/store/notifications'
+import { $newWorktreeRequest } from '@/store/projects'
+
+// Tiny uppercase section header, matching the composer "+" menu's labels.
+const MENU_SECTION = 'text-[0.625rem] font-semibold uppercase tracking-wider text-(--ui-text-tertiary)'
+
+interface BranchActionCopy {
+  branchCreateWorktree: string
+  branchOpenExisting: string
+  branchSwitchHome: string
+}
+
+const branchActionLabel = (branch: HermesGitBranch, copy: BranchActionCopy) => {
+  if (branch.checkedOut) {
+    return copy.branchOpenExisting
+  }
+
+  return branch.isDefault ? copy.branchSwitchHome : copy.branchCreateWorktree
+}
+
+interface CodingStatusRowProps {
+  /** Branch the current draft off into a fresh worktree + session, based on
+   *  `base` (a branch name; omitted = current HEAD). The composer owns the
+   *  draft, so it supplies the orchestration; the row just collects the new
+   *  branch name + base. Omitted (e.g. remote backend) hides the affordance. */
+  onBranchOff?: (branch: string, base?: string) => Promise<void>
+  /** Check an existing branch out into a fresh worktree + session (no new
+   *  branch). Drives the dialog's "convert a branch" picker. */
+  onConvertBranch?: (branch: string, path?: null | string, isDefault?: boolean) => Promise<void>
+  /** List the repo's local branches for the "convert a branch" picker. */
+  onListBranches?: () => Promise<HermesGitBranch[]>
+  /** Open the review pane (changed files + diffs). */
+  onOpen?: () => void
+  /** Jump into an existing worktree (open a fresh session anchored there). */
+  onOpenWorktree?: (path: string) => void
+  /** Switch the current repo checkout to another branch. */
+  onSwitchBranch?: (branch: string) => Promise<void>
+}
+
+/**
+ * The always-on coding-context row, the BASE of the composer status stack:
+ * current branch, dirty summary (+/-), and ahead/behind. A touch more prominent
+ * than the per-turn rows above it (larger branch label, accent glyph), and the
+ * entry point to the review pane. Hidden when the active session isn't in a
+ * local git repo (the probe returns null).
+ */
+export const CodingStatusRow = memo(function CodingStatusRow({
+  onBranchOff,
+  onConvertBranch,
+  onListBranches,
+  onOpen,
+  onOpenWorktree,
+  onSwitchBranch
+}: CodingStatusRowProps) {
+  const { t } = useI18n()
+  const s = t.statusStack.coding
+  const p = t.sidebar.projects
+  const status = useStore($repoStatus)
+  const worktrees = useStore($repoWorktrees)
+
+  const [branchOpen, setBranchOpen] = useState(false)
+  const [branchName, setBranchName] = useState('')
+  const [branchBase, setBranchBase] = useState<string | undefined>(undefined)
+  const [branchPending, setBranchPending] = useState(false)
+  const [convertMode, setConvertMode] = useState(false)
+  const [branches, setBranches] = useState<HermesGitBranch[]>([])
+  const [branchesLoading, setBranchesLoading] = useState(false)
+
+  const loadBranches = useCallback(async () => {
+    if (!onListBranches) {
+      return
+    }
+
+    setBranchesLoading(true)
+
+    try {
+      setBranches(await onListBranches())
+    } catch {
+      setBranches([])
+    } finally {
+      setBranchesLoading(false)
+    }
+  }, [onListBranches])
+
+  // Open the name dialog for a chosen base. Deferred so the dropdown finishes
+  // closing before the dialog grabs focus (Radix focus-trap handoff races
+  // otherwise).
+  const startBranch = (base: string | undefined) => {
+    setBranchBase(base)
+    setBranchName('')
+    setConvertMode(false)
+    setTimeout(() => setBranchOpen(true), 0)
+  }
+
+  const startConvert = () => {
+    setBranchBase(undefined)
+    setBranchName('')
+    setConvertMode(true)
+    void loadBranches()
+    setTimeout(() => setBranchOpen(true), 0)
+  }
+
+  const enterConvert = () => {
+    setConvertMode(true)
+    void loadBranches()
+  }
+
+  const convertBranch = async (branch: HermesGitBranch) => {
+    if (branchPending || !branch || !onConvertBranch) {
+      return
+    }
+
+    setBranchPending(true)
+
+    try {
+      await onConvertBranch(branch.name, branch.worktreePath, branch.isDefault)
+      setBranchOpen(false)
+    } catch (err) {
+      notifyError(err, p.startWorkFailed)
+    } finally {
+      setBranchPending(false)
+    }
+  }
+
+  // Global ⌘⇧B (workspace.newWorktree): open the name dialog for a worktree off
+  // current HEAD. The rail only renders inside a repo, so the hotkey naturally
+  // no-ops elsewhere. Guarded by a token ref so it fires on the keypress, not on
+  // mount or unrelated re-renders.
+  const worktreeReq = useStore($newWorktreeRequest)
+  const lastWorktreeReqRef = useRef(worktreeReq)
+
+  useEffect(() => {
+    if (worktreeReq === lastWorktreeReqRef.current) {
+      return
+    }
+
+    lastWorktreeReqRef.current = worktreeReq
+
+    if (!onBranchOff) {
+      return
+    }
+
+    setBranchBase(undefined)
+    setBranchName('')
+    setConvertMode(false)
+    setBranchOpen(true)
+  }, [onBranchOff, worktreeReq])
+
+  const submitBranch = async () => {
+    const branch = branchName.trim()
+
+    if (branchPending || !branch || !onBranchOff) {
+      return
+    }
+
+    setBranchPending(true)
+
+    try {
+      await onBranchOff(branch, branchBase)
+      setBranchOpen(false)
+      setBranchName('')
+    } catch (err) {
+      notifyError(err, p.startWorkFailed)
+    } finally {
+      setBranchPending(false)
+    }
+  }
+
+  const switchToBranch = async (branch: string) => {
+    if (!onSwitchBranch) {
+      return
+    }
+
+    try {
+      await onSwitchBranch(branch)
+    } catch (err) {
+      notifyError(err, s.switchFailed(branch))
+    }
+  }
+
+  if (!status) {
+    return null
+  }
+
+  const branchLabel = status.detached ? s.detached : status.branch || s.noBranch
+  // The kebab offers branching off the trunk and/or the current branch. The
+  // worktree-add bases the new branch on `base` (a branch name; undefined =
+  // current HEAD). We dedupe so "on main" shows a single trunk entry, and fall
+  // back to a plain off-HEAD branch when no trunk is detected.
+  const current = status.detached ? null : status.branch
+  const branchTargets: { base: string | undefined; label: string }[] = []
+
+  // Current branch first (the 99% "branch off where I am"), then the trunk just
+  // below it ("New branch from main"), deduped when they're the same.
+  if (current) {
+    branchTargets.push({ base: current, label: s.branchOffFrom(current) })
+  }
+
+  if (status.defaultBranch && status.defaultBranch !== current) {
+    branchTargets.push({ base: status.defaultBranch, label: s.branchOffFrom(status.defaultBranch) })
+  }
+
+  if (branchTargets.length === 0) {
+    branchTargets.push({ base: undefined, label: s.newBranch })
+  }
+
+  const switchTarget =
+    onSwitchBranch && current && status.defaultBranch && status.defaultBranch !== current ? status.defaultBranch : null
+
+  // Other worktrees to jump into — everything except the one we're already in
+  // (matched by its checked-out branch) and the bare/main placeholder entry.
+  const otherWorktrees = onOpenWorktree
+    ? worktrees.filter(w => w.path && !w.detached && w.branch && w.branch !== current)
+    : []
+
+  const hasLineDelta = status.added > 0 || status.removed > 0
+  // Untracked files carry no line delta vs HEAD, so surface them as a count when
+  // they're the only change (otherwise +/- tells the story).
+  const untrackedOnly = !hasLineDelta && status.untracked > 0
+
+  return (
+    <>
+      <StatusRow
+        // The base "where am I working" strip is part of the composer surface
+        // itself, so it inherits the composer's width and clipped top radius.
+        className="coding-status-bar min-h-7 rounded-t-[inherit] rounded-b-none border-b border-(--ui-stroke-tertiary) px-3.5 py-1.5 hover:bg-transparent"
+        // Static branch glyph — never the loading spinner. This row only renders
+        // once `status` exists, so a spinner here only ever fired on *refreshes*
+        // of an already-loaded repo (window focus, turn settle), reading as an
+        // annoying icon "blip" with no first-load value. Refreshes are silent.
+        leading={<Codicon className="text-(--ui-green)" name="git-branch" size="0.8rem" />}
+        onActivate={onOpen}
+      >
+        <div className="flex min-w-0 flex-1 items-center gap-1">
+          <span
+            className="min-w-0 truncate text-xs font-normal text-muted-foreground/92 transition-colors group-hover/status-row:text-foreground/90"
+            title={branchLabel}
+          >
+            {branchLabel}
+          </span>
+
+          {/* Branch actions kebab — same pattern as the session/worktree rows.
+              ALWAYS laid out; only its opacity flips on hover/focus/open, so
+              revealing it never reflows the row (no layout shift). pointer-events
+              follow opacity so the invisible trigger isn't clickable at rest. */}
+          {onBranchOff && (
+            <DropdownMenu>
+              <DropdownMenuTrigger asChild>
+                <Button
+                  aria-label={s.newBranch}
+                  className="pointer-events-none size-4 shrink-0 text-muted-foreground/60 opacity-0 transition hover:text-foreground group-hover/status-row:pointer-events-auto group-hover/status-row:opacity-100 group-focus-within/status-row:pointer-events-auto group-focus-within/status-row:opacity-100 data-[state=open]:pointer-events-auto data-[state=open]:opacity-100"
+                  onClick={event => event.stopPropagation()}
+                  onKeyDown={event => {
+                    // The row's onActivate also fires on Enter/Space; keep it from
+                    // opening the review pane when the kebab is the focus target.
+                    if (event.key === 'Enter' || event.key === ' ') {
+                      event.stopPropagation()
+                    }
+                  }}
+                  size="icon-xs"
+                  variant="ghost"
+                >
+                  <Codicon name="kebab-vertical" size="0.8rem" />
+                </Button>
+              </DropdownMenuTrigger>
+              {/* The row sits at the bottom of the screen (above the composer),
+                  so the menu opens upward. */}
+              <DropdownMenuContent align="end" className="w-60" side="top" sideOffset={6}>
+                <DropdownMenuLabel className={MENU_SECTION}>{s.newBranch}</DropdownMenuLabel>
+                {branchTargets.map(target => (
+                  <DropdownMenuItem key={target.base ?? '__head__'} onSelect={() => startBranch(target.base)}>
+                    <span className="truncate">{target.label}</span>
+                  </DropdownMenuItem>
+                ))}
+
+                {switchTarget && (
+                  <DropdownMenuItem onSelect={() => void switchToBranch(switchTarget)}>
+                    <span className="truncate">{s.switchTo(switchTarget)}</span>
+                  </DropdownMenuItem>
+                )}
+
+                <DropdownMenuSeparator />
+                <DropdownMenuLabel className={MENU_SECTION}>{s.worktrees}</DropdownMenuLabel>
+                {otherWorktrees.map(worktree => (
+                  <DropdownMenuItem key={worktree.path} onSelect={() => onOpenWorktree?.(worktree.path)}>
+                    <span className="truncate">{worktree.branch}</span>
+                  </DropdownMenuItem>
+                ))}
+                {/* Create a fresh worktree off the current HEAD (the generic
+                    "spin up a worktree here", mirroring the sidebar's + button). */}
+                <DropdownMenuItem onSelect={() => startBranch(undefined)}>
+                  <span className="truncate">{p.startWork}</span>
+                </DropdownMenuItem>
+                {/* Check an EXISTING branch out into a worktree (no new branch). */}
+                {onConvertBranch && (
+                  <DropdownMenuItem onSelect={() => startConvert()}>
+                    <span className="truncate">{p.convertBranch}</span>
+                  </DropdownMenuItem>
+                )}
+              </DropdownMenuContent>
+            </DropdownMenu>
+          )}
+        </div>
+
+        {(status.ahead > 0 || status.behind > 0) && (
+          <span className="ml-auto flex shrink-0 items-center gap-1.5 text-[0.68rem] leading-4 text-muted-foreground/75 tabular-nums">
+            {status.ahead > 0 && (
+              <span className="flex items-center gap-0.5" title={s.ahead(status.ahead)}>
+                <span aria-hidden>↑</span>
+                {status.ahead}
+              </span>
+            )}
+            {status.behind > 0 && (
+              <span className="flex items-center gap-0.5" title={s.behind(status.behind)}>
+                <span aria-hidden>↓</span>
+                {status.behind}
+              </span>
+            )}
+          </span>
+        )}
+
+        {hasLineDelta ? (
+          <DiffCount
+            added={status.added}
+            className={`text-[0.72rem] leading-4 ${status.ahead === 0 && status.behind === 0 ? 'ml-auto' : ''}`}
+            removed={status.removed}
+          />
+        ) : untrackedOnly ? (
+          <span
+            className={`shrink-0 text-[0.72rem] leading-4 text-amber-500/90 ${status.ahead === 0 && status.behind === 0 ? 'ml-auto' : ''}`}
+          >
+            {s.changed(status.untracked)}
+          </span>
+        ) : null}
+      </StatusRow>
+
+      <Dialog onOpenChange={open => !branchPending && setBranchOpen(open)} open={branchOpen}>
+        <DialogContent className="max-w-md">
+          <DialogHeader>
+            <DialogTitle>{convertMode ? p.convertBranchTitle : p.newWorktreeTitle}</DialogTitle>
+            <DialogDescription>
+              {convertMode ? p.convertBranchDesc : p.newWorktreeDesc}
+              {!convertMode && branchBase && (
+                <span className="mt-1 block text-(--ui-text-secondary)">{s.branchOffFrom(branchBase)}</span>
+              )}
+            </DialogDescription>
+          </DialogHeader>
+
+          {convertMode ? (
+            <Command
+              className="rounded-md border border-(--ui-stroke-tertiary)"
+              // The branch name is the authoritative key; filter on it directly.
+              filter={(value, search) => (value.toLowerCase().includes(search.toLowerCase()) ? 1 : 0)}
+            >
+              <CommandInput autoFocus disabled={branchPending} placeholder={p.convertBranchPlaceholder} />
+              <CommandList className="max-h-64">
+                <CommandEmpty>{branchesLoading ? p.branchesLoading : p.noBranches}</CommandEmpty>
+                <CommandGroup>
+                  {branches.map(branch => (
+                    <CommandItem
+                      disabled={branchPending}
+                      key={branch.name}
+                      onSelect={() => void convertBranch(branch)}
+                      value={branch.name}
+                    >
+                      <Codicon className="shrink-0 text-(--ui-text-tertiary)" name="git-branch" size="0.8rem" />
+                      <span className="truncate">{branch.name}</span>
+                      <span className="ml-auto shrink-0 text-[0.625rem] text-(--ui-text-tertiary)">
+                        {branchActionLabel(branch, p)}
+                      </span>
+                    </CommandItem>
+                  ))}
+                </CommandGroup>
+              </CommandList>
+            </Command>
+          ) : (
+            <SanitizedInput
+              autoFocus
+              disabled={branchPending}
+              onKeyDown={event => {
+                if (event.key === 'Enter') {
+                  event.preventDefault()
+                  void submitBranch()
+                } else if (event.key === 'Escape') {
+                  setBranchOpen(false)
+                }
+              }}
+              onValueChange={setBranchName}
+              placeholder={p.branchPlaceholder}
+              sanitize={gitRef}
+              value={branchName}
+            />
+          )}
+
+          {convertMode ? (
+            <DialogFooter className="sm:justify-start">
+              <Button
+                className="px-0 text-(--ui-text-secondary) hover:text-foreground"
+                disabled={branchPending}
+                onClick={() => setConvertMode(false)}
+                type="button"
+                variant="link"
+              >
+                {t.common.cancel}
+              </Button>
+            </DialogFooter>
+          ) : (
+            <DialogFooter className="sm:justify-between">
+              {onConvertBranch ? (
+                <Button
+                  className="px-0 text-(--ui-text-secondary) hover:text-foreground"
+                  disabled={branchPending}
+                  onClick={enterConvert}
+                  type="button"
+                  variant="link"
+                >
+                  {p.convertBranchInstead}
+                </Button>
+              ) : (
+                <span />
+              )}
+              <div className="flex items-center gap-2">
+                <Button disabled={branchPending} onClick={() => setBranchOpen(false)} type="button" variant="ghost">
+                  {t.common.cancel}
+                </Button>
+                <Button
+                  disabled={branchPending || !branchName.trim()}
+                  onClick={() => void submitBranch()}
+                  type="button"
+                >
+                  {p.startWork}
+                </Button>
+              </div>
+            </DialogFooter>
+          )}
+        </DialogContent>
+      </Dialog>
+    </>
+  )
+})

apps/desktop/src/app/chat/composer/status-stack/index.tsx

@@ -30,6 +30,19 @@ import { StatusItemRow } from './status-row'
 // emit no event when they die). Only armed while a running row is on screen.
 const BACKGROUND_POLL_MS = 5_000
 
+// A localhost/loopback preview is only meaningful while its dev server is up, so
+// we tie it to a live background process rather than persisting dismissals or
+// letting dead URLs pile up. File previews (a real on-disk artifact) stand alone.
+const isLocalhostPreview = (target: string): boolean => /\b(?:localhost|127\.0\.0\.1|0\.0\.0\.0)\b/i.test(target)
+
+// Real codicons per group (no sparkles): a checklist for todos, a bot for
+// subagents, a background process glyph for background tasks.
+const GROUP_ICON: Record<StatusGroup['type'], string> = {
+  todo: 'checklist',
+  subagent: 'hubot',
+  background: 'server-process'
+}
+
 const groupLabel = (group: StatusGroup, s: Translations['statusStack']) => {
   if (group.type === 'todo') {
     return s.todos(group.items.filter(i => i.todoStatus === 'completed').length, group.items.length)
@@ -74,6 +87,10 @@ export function ComposerStatusStack({ queue, sessionId }: ComposerStatusStackPro
 
   const hasRunningBackground = groups.some(g => g.type === 'background' && g.items.some(i => i.state === 'running'))
 
+  // Drop localhost previews once no dev server is left running — that's what made
+  // dead `localhost:5174` chips stick around. On-disk file previews are kept.
+  const visiblePreviews = previews.filter(item => hasRunningBackground || !isLocalhostPreview(item.target))
+
   useEffect(() => {
     if (!sessionId || !hasRunningBackground) {
       return
@@ -89,6 +106,18 @@ export function ComposerStatusStack({ queue, sessionId }: ComposerStatusStackPro
   const openSubagent = (item: ComposerStatusItem) =>
     item.sessionId ? void openSessionInNewWindow(item.sessionId, { watch: true }) : openAgents()
 
+  // Preview links live as child rows of the background group — a localhost dev
+  // server and its preview are the same thing — so they no longer float as an
+  // odd, differently-indented standalone block under the stack.
+  const previewRows =
+    visiblePreviews.length > 0 && sessionId
+      ? visiblePreviews.map(item => (
+          <PreviewStatusRow item={item} key={item.id} onDismiss={id => dismissPreviewArtifact(sessionId, id)} />
+        ))
+      : []
+
+  const hasBackgroundGroup = groups.some(g => g.type === 'background')
+
   const sections: { key: string; node: ReactNode }[] = groups.map(group => ({
     key: group.type,
     node: (
@@ -107,11 +136,7 @@ export function ComposerStatusStack({ queue, sessionId }: ComposerStatusStackPro
           ) : undefined
         }
         defaultCollapsed={group.type !== 'todo'}
-        icon={
-          group.type === 'todo' ? (
-            <Codicon className="text-muted-foreground/70" name="checklist" size="0.8rem" />
-          ) : undefined
-        }
+        icon={<Codicon className="text-muted-foreground/70" name={GROUP_ICON[group.type]} size="0.8rem" />}
         label={groupLabel(group, t.statusStack)}
       >
         {group.items.map(item => (
@@ -120,25 +145,20 @@ export function ComposerStatusStack({ queue, sessionId }: ComposerStatusStackPro
             key={item.id}
             onDismiss={sessionId ? id => dismissBackgroundProcess(sessionId, id) : undefined}
             onOpen={() => openSubagent(item)}
-            onStop={sessionId ? id => stopBackgroundProcess(sessionId, id) : undefined}
+            onStop={sessionId ? id => void stopBackgroundProcess(sessionId, id) : undefined}
           />
         ))}
+        {group.type === 'background' && previewRows}
       </StatusSection>
     )
   }))
 
-  if (previews.length > 0 && sessionId) {
+  // No background group to host them (e.g. a standalone on-disk file preview):
+  // keep the previews as their own row block so they don't disappear.
+  if (previewRows.length > 0 && !hasBackgroundGroup) {
     sections.push({
       key: 'preview',
-      // Not a collapsible group — preview links just sit there, one line each,
-      // each individually closeable.
-      node: (
-        <div className="px-1 py-0.5">
-          {previews.map(item => (
-            <PreviewStatusRow item={item} key={item.id} onDismiss={id => dismissPreviewArtifact(sessionId, id)} />
-          ))}
-        </div>
-      )
+      node: <div className="px-1 py-0.5">{previewRows}</div>
     })
   }
 
@@ -190,12 +210,10 @@ export function ComposerStatusStack({ queue, sessionId }: ComposerStatusStackPro
 
   return (
     <div
-      // Sits above the composer (bottom-full), nudged down by the shell's 0.5rem
-      // top pad (pt-2 on composer-root) plus 1px so its bottom edge overlaps the
-      // composer surface's top border. z BELOW the surface (z-4) so the surface's
-      // top border paints over our transparent bottom border — one seam, no
-      // double line.
-      className="absolute inset-x-0 bottom-full z-3 max-h-[40vh] translate-y-[calc(0.5rem+1px)] overflow-y-auto"
+      // Sits in the overlay lane above the composer. The composer root has pt-2
+      // before the actual surface; translate by that amount so the stack returns
+      // to its original attachment point without intruding into the repo strip.
+      className="absolute inset-x-0 bottom-full z-3 max-h-[40vh] translate-y-2 overflow-y-auto"
       onPointerDownCapture={() => blurComposerInput()}
       ref={stackRef}
     >
@@ -205,17 +223,19 @@ export function ComposerStatusStack({ queue, sessionId }: ComposerStatusStackPro
           Rounded top, square bottom; the bottom border is TRANSPARENT — the
           composer surface's visible top border (which sits at a higher z) is the
           single shared seam, so the two read as one fused capsule. */}
-      <div className={cn(composerDockCard('top'), 'mx-2 rounded-b-none border-b border-b-transparent pt-0.5 pb-1')}>
-        <div
-          className={cn(
-            'transition-opacity duration-200 ease-out',
-            scrolledUp ? 'opacity-30 group-hover/composer:opacity-100' : 'opacity-100'
-          )}
-        >
-          {sections.map(section => (
-            <div key={section.key}>{section.node}</div>
-          ))}
-        </div>
+      <div
+        className={cn(
+          composerDockCard('top'),
+          // Inset (mx-2) so the stack reads slightly narrower than the composer
+          // surface below it — the original look.
+          'mx-2 overflow-hidden rounded-b-none border-b border-b-transparent pt-0.5',
+          'transition-opacity duration-200 ease-out',
+          scrolledUp ? 'opacity-30 group-hover/composer:opacity-100' : 'opacity-100'
+        )}
+      >
+        {sections.map(section => (
+          <div key={section.key}>{section.node}</div>
+        ))}
       </div>
     </div>
   )

apps/desktop/src/app/chat/composer/status-stack/preview-row.tsx

@@ -6,7 +6,6 @@ import { Button } from '@/components/ui/button'
 import { Codicon } from '@/components/ui/codicon'
 import { Tip } from '@/components/ui/tooltip'
 import { useI18n } from '@/i18n'
-import { ChevronRight, X } from '@/lib/icons'
 import { normalizeOrLocalPreviewTarget } from '@/lib/local-preview'
 import { cn } from '@/lib/utils'
 import { PREVIEW_PANE_ID } from '@/store/layout'
@@ -76,50 +75,52 @@ export const PreviewStatusRow = memo(function PreviewStatusRow({ item, onDismiss
 
   return (
     <StatusRow
-      leading={<ChevronRight aria-hidden className="size-3 text-muted-foreground/80" />}
-      onActivate={() => void togglePreview()}
+      leading={
+        <Codicon
+          aria-hidden
+          className={cn('text-muted-foreground/70', opening && 'animate-pulse')}
+          name="globe"
+          size="0.8rem"
+        />
+      }
+      // Plain click opens the link in the browser; ⌘/Ctrl-click opens it in the
+      // in-app preview pane instead. (isOpen still toggles the pane closed.)
+      onActivate={event => {
+        if (event.metaKey || event.ctrlKey) {
+          void togglePreview()
+        } else {
+          void openInBrowser()
+        }
+      }}
       trailing={
-        <span className="-my-1 flex items-center gap-0.5">
-          <Tip label={t.preview.openInBrowser}>
-            <Button
-              aria-label={t.preview.openInBrowser}
-              className="size-4 rounded-md text-muted-foreground/60 hover:text-foreground/90"
-              onClick={event => {
-                event.stopPropagation()
-                void openInBrowser()
-              }}
-              size="icon-xs"
-              type="button"
-              variant="ghost"
-            >
-              <Codicon name="link-external" size="0.75rem" />
-            </Button>
-          </Tip>
-          <Tip label={t.statusStack.dismiss}>
-            <Button
-              aria-label={t.statusStack.dismiss}
-              className="size-4 rounded-md text-muted-foreground/60 hover:text-foreground/90"
-              onClick={event => {
-                event.stopPropagation()
-                onDismiss(item.id)
-              }}
-              size="icon-xs"
-              type="button"
-              variant="ghost"
-            >
-              <X size={12} />
-            </Button>
-          </Tip>
-        </span>
+        <Tip label={t.statusStack.dismiss}>
+          <Button
+            aria-label={t.statusStack.dismiss}
+            className="-my-1 size-4 rounded-md text-muted-foreground/60 hover:text-foreground/90"
+            onClick={event => {
+              event.stopPropagation()
+              onDismiss(item.id)
+            }}
+            size="icon-xs"
+            type="button"
+            variant="ghost"
+          >
+            <Codicon name="close" size="0.75rem" />
+          </Button>
+        </Tip>
       }
       trailingVisible
     >
-      <span className="min-w-0 max-w-[18rem] truncate text-[0.73rem] leading-4 text-foreground/92" title={item.target}>
-        {item.label}
-      </span>
-      <span className={cn('shrink-0 text-[0.62rem] leading-4 text-muted-foreground/70', opening && 'animate-pulse')}>
-        {opening ? t.preview.opening : isOpen ? t.preview.hide : t.preview.openPreview}
-      </span>
+      <Tip
+        label={
+          <span className="flex flex-col gap-0.5">
+            <span>{item.target}</span>
+            <span className="opacity-70">{t.preview.linkHint}</span>
+          </span>
+        }
+      >
+        <span className="min-w-0 max-w-[18rem] truncate text-[0.73rem] leading-4 text-foreground/92">{item.label}</span>
+      </Tip>
     </StatusRow>
   )
 })

apps/desktop/src/app/chat/composer/status-stack/status-row.tsx

@@ -8,7 +8,6 @@ import { DisclosureCaret } from '@/components/ui/disclosure-caret'
 import { GlyphSpinner } from '@/components/ui/glyph-spinner'
 import { Tip } from '@/components/ui/tooltip'
 import { type Translations, useI18n } from '@/i18n'
-import { ArrowUpRight, X } from '@/lib/icons'
 import type { TodoStatus } from '@/lib/todos'
 import { cn } from '@/lib/utils'
 import type { ComposerStatusItem } from '@/store/composer-status'
@@ -50,7 +49,7 @@ function leadingGlyph(item: ComposerStatusItem, s: Translations['statusStack']):
     return (
       <GlyphSpinner
         ariaLabel={s.running}
-        className="text-[0.9rem] leading-none text-muted-foreground/80"
+        className="text-[0.85rem] leading-none text-muted-foreground/80"
         spinner="braille"
       />
     )
@@ -117,11 +116,11 @@ export const StatusItemRow = memo(function StatusItemRow({ item, onDismiss, onOp
                 type="button"
                 variant="ghost"
               >
-                <X size={12} />
+                <Codicon name="close" size="0.75rem" />
               </Button>
             </Tip>
           ) : canOpen ? (
-            <ArrowUpRight aria-hidden className="size-3.5 text-muted-foreground/55" />
+            <Codicon aria-hidden className="text-muted-foreground/55" name="link-external" size="0.85rem" />
           ) : undefined
         }
       >

apps/desktop/src/app/chat/composer/trigger-popover.test.tsx

@@ -11,7 +11,14 @@ function renderPopover(kind: '@' | '/', loading = false) {
 
   const rendered = render(
     <I18nProvider configClient={null} initialLocale="zh">
-      <ComposerTriggerPopover activeIndex={0} items={[]} kind={kind} loading={loading} onHover={onHover} onPick={onPick} />
+      <ComposerTriggerPopover
+        activeIndex={0}
+        items={[]}
+        kind={kind}
+        loading={loading}
+        onHover={onHover}
+        onPick={onPick}
+      />
     </I18nProvider>
   )
 

apps/desktop/src/app/chat/composer/types.ts

@@ -46,7 +46,7 @@ export interface ChatBarProps {
   onAddUrl?: (url: string) => void
   onAttachImageBlob?: (blob: Blob) => Promise<boolean | void> | boolean | void
   onAttachDroppedItems?: (candidates: DroppedFile[]) => Promise<boolean | void> | boolean | void
-  onPasteClipboardImage?: () => void
+  onPasteClipboardImage?: (opts?: { silent?: boolean }) => Promise<boolean> | void
   onPickFiles?: () => void
   onPickFolders?: () => void
   onPickImages?: () => void

apps/desktop/src/app/chat/hooks/use-composer-actions.ts

@@ -226,9 +226,10 @@ const attachToMain = (attachment: ComposerAttachment) => {
 export function useComposerActions({ activeSessionId, currentCwd, requestGateway }: ComposerActionsOptions) {
   const { t } = useI18n()
   const copy = t.desktop
+
   const addTextToDraft = useCallback((text: string) => {
     requestComposerInsert(text, { mode: 'block' })
-  }, [copy.imagePreviewFailed])
+  }, [])
 
   const addTerminalSelectionAttachment = useCallback((text: string, label = 'selection') => {
     const trimmed = text.trim()
@@ -329,35 +330,38 @@ export function useComposerActions({ activeSessionId, currentCwd, requestGateway
     [currentCwd]
   )
 
-  const attachImagePath = useCallback(async (filePath: string) => {
-    if (!filePath) {
-      return false
-    }
-
-    const baseAttachment: ComposerAttachment = {
-      id: attachmentId('image', filePath),
-      kind: 'image',
-      label: pathLabel(filePath),
-      detail: filePath,
-      path: filePath
-    }
-
-    attachToMain(baseAttachment)
-
-    try {
-      const previewUrl = await window.hermesDesktop?.readFileDataUrl(filePath)
-
-      if (previewUrl) {
-        addComposerAttachment({ ...baseAttachment, previewUrl })
+  const attachImagePath = useCallback(
+    async (filePath: string) => {
+      if (!filePath) {
+        return false
       }
 
-      return true
-    } catch (err) {
-      notifyError(err, copy.imagePreviewFailed)
+      const baseAttachment: ComposerAttachment = {
+        id: attachmentId('image', filePath),
+        kind: 'image',
+        label: pathLabel(filePath),
+        detail: filePath,
+        path: filePath
+      }
 
-      return true
-    }
-  }, [])
+      attachToMain(baseAttachment)
+
+      try {
+        const previewUrl = await window.hermesDesktop?.readFileDataUrl(filePath)
+
+        if (previewUrl) {
+          addComposerAttachment({ ...baseAttachment, previewUrl })
+        }
+
+        return true
+      } catch (err) {
+        notifyError(err, copy.imagePreviewFailed)
+
+        return true
+      }
+    },
+    [copy.imagePreviewFailed]
+  )
 
   const attachImageBlob = useCallback(
     async (blob: Blob) => {
@@ -411,25 +415,36 @@ export function useComposerActions({ activeSessionId, currentCwd, requestGateway
     }
   }, [attachImagePath, copy.attachImages, currentCwd, t.composer.images])
 
-  const pasteClipboardImage = useCallback(async () => {
-    try {
-      const path = await window.hermesDesktop?.saveClipboardImage()
+  const pasteClipboardImage = useCallback(
+    async ({ silent = false }: { silent?: boolean } = {}) => {
+      try {
+        const path = await window.hermesDesktop?.saveClipboardImage()
 
-      if (!path) {
-        notify({
-          kind: 'warning',
-          title: copy.clipboard,
-          message: copy.noClipboardImage
-        })
+        if (!path) {
+          if (!silent) {
+            notify({
+              kind: 'warning',
+              title: copy.clipboard,
+              message: copy.noClipboardImage
+            })
+          }
 
-        return
+          return false
+        }
+
+        await attachImagePath(path)
+
+        return true
+      } catch (err) {
+        if (!silent) {
+          notifyError(err, copy.clipboardPasteFailed)
+        }
+
+        return false
       }
-
-      await attachImagePath(path)
-    } catch (err) {
-      notifyError(err, copy.clipboardPasteFailed)
-    }
-  }, [attachImagePath, copy.clipboard, copy.clipboardPasteFailed, copy.noClipboardImage])
+    },
+    [attachImagePath, copy.clipboard, copy.clipboardPasteFailed, copy.noClipboardImage]
+  )
 
   const attachContextFolderPath = useCallback(
     (folderPath: string) => {

apps/desktop/src/app/chat/index.tsx

@@ -75,7 +75,7 @@ interface ChatViewProps extends Omit<React.ComponentProps<'div'>, 'onSubmit'> {
   maxVoiceRecordingSeconds?: number
   onAttachImageBlob: (blob: Blob) => Promise<boolean | void> | boolean | void
   onAttachDroppedItems: (candidates: DroppedFile[]) => Promise<boolean | void> | boolean | void
-  onPasteClipboardImage: () => void
+  onPasteClipboardImage: (opts?: { silent?: boolean }) => Promise<boolean> | void
   onPickFiles: () => void
   onPickFolders: () => void
   onPickImages: () => void
@@ -88,7 +88,7 @@ interface ChatViewProps extends Omit<React.ComponentProps<'div'>, 'onSubmit'> {
   onThreadMessagesChange: (messages: readonly ThreadMessage[]) => void
   onEdit: (message: AppendMessage) => Promise<void>
   onReload: (parentId: string | null) => Promise<void>
-  onRestoreToMessage?: (messageId: string) => Promise<void>
+  onRestoreToMessage?: (messageId: string, target?: { text?: string; userOrdinal?: number | null }) => Promise<void>
   onRetryResume: (sessionId: string) => void
   onTranscribeAudio?: (audio: Blob) => Promise<string>
   onDismissError?: (messageId: string) => void
@@ -317,7 +317,12 @@ export function ChatView({
   // The compact new-session pop-out skips the wordmark/tagline intro — it's a
   // scratch window, not the full-height empty state.
   const showIntro =
-    !isSecondaryWindow() && freshDraftReady && !isRoutedSessionView && !selectedSessionId && !activeSessionId && messagesEmpty
+    !isSecondaryWindow() &&
+    freshDraftReady &&
+    !isRoutedSessionView &&
+    !selectedSessionId &&
+    !activeSessionId &&
+    messagesEmpty
 
   // Session is still loading if the route references a session we haven't
   // resumed yet. Once `activeSessionId` is set (runtime has resumed), the

apps/desktop/src/app/chat/right-rail/preview-file.tsx

@@ -6,7 +6,7 @@ import type {
   MouseEvent as ReactMouseEvent,
   ReactNode
 } from 'react'
-import { useEffect, useMemo, useState } from 'react'
+import { Fragment, useCallback, useEffect, useMemo, useRef, useState } from 'react'
 import ShikiHighlighter from 'react-shiki'
 import { Streamdown } from 'streamdown'
 
@@ -14,15 +14,31 @@ import { requestComposerFocus, requestComposerInsertRefs } from '@/app/chat/comp
 import { droppedFileInlineRef } from '@/app/chat/composer/inline-refs'
 import { HERMES_PATHS_MIME } from '@/app/chat/hooks/use-composer-actions'
 import { isAddSelectionShortcut } from '@/app/right-sidebar/terminal/selection'
+import { CodeEditor } from '@/components/chat/code-editor'
+import { FileDiffPanel } from '@/components/chat/diff-lines'
+import { chunkTextLines, useFixedRowWindow } from '@/components/chat/fixed-row-window'
 import { PageLoader } from '@/components/page-loader'
 import { translateNow, useI18n } from '@/i18n'
-import { readDesktopFileDataUrl, readDesktopFileText } from '@/lib/desktop-fs'
+import {
+  desktopFileDiff,
+  desktopGitRoot,
+  readDesktopFileDataUrl,
+  readDesktopFileText,
+  writeDesktopFileText
+} from '@/lib/desktop-fs'
+import { Check, Pencil, X } from '@/lib/icons'
+import { shikiLanguageForFilename } from '@/lib/markdown-code'
 import { cn } from '@/lib/utils'
 import type { PreviewTarget } from '@/store/preview'
+import { setPreviewDirty } from '@/store/preview-edit'
 import { $currentCwd } from '@/store/session'
+import { notifyWorkspaceChanged } from '@/store/workspace-events'
 
 const SHIKI_THEME = { dark: 'github-dark-default', light: 'github-light-default' } as const
 const TEXT_PREVIEW_MAX_BYTES = 512 * 1024
+const SOURCE_CHUNK_LINES = 200
+const SOURCE_LINE_PX = 20
+const SOURCE_OVERSCAN_LINES = 400
 
 type EmptyStateTone = 'neutral' | 'warning'
 
@@ -126,6 +142,8 @@ interface LocalPreviewState {
   binary?: boolean
   byteSize?: number
   dataUrl?: string
+  /** Working-tree-vs-HEAD unified diff, when the file has uncommitted changes. */
+  diff?: string
   error?: string
   language?: string
   loading: boolean
@@ -133,6 +151,19 @@ interface LocalPreviewState {
   truncated?: boolean
 }
 
+// True when focus is in a field that should swallow plain keystrokes (so the
+// bare-`e` edit shortcut never fires while the user is typing in the composer,
+// a search box, or the editor itself).
+function isTypableElement(el: Element | null): boolean {
+  if (!el) {
+    return false
+  }
+
+  const tag = el.tagName
+
+  return tag === 'INPUT' || tag === 'TEXTAREA' || tag === 'SELECT' || (el as HTMLElement).isContentEditable
+}
+
 function filePathForTarget(target: PreviewTarget) {
   if (target.path) {
     return target.path
@@ -299,27 +330,92 @@ function MarkdownPreview({ text }: { text: string }) {
   )
 }
 
-function PreviewToggle({ asSource, onToggle }: { asSource: boolean; onToggle: () => void }) {
+function PreviewModeSwitcher({
+  active,
+  modes,
+  onSelect,
+  trailing
+}: {
+  active: PreviewViewMode
+  modes: PreviewViewMode[]
+  onSelect: (mode: PreviewViewMode) => void
+  trailing?: ReactNode
+}) {
   const { t } = useI18n()
+  const showModes = modes.length > 1
+
+  if (!showModes && !trailing) {
+    return null
+  }
+
+  const label: Record<PreviewViewMode, string> = {
+    diff: t.preview.diff,
+    rendered: t.preview.renderedPreview,
+    source: t.preview.source
+  }
 
   return (
-    <div className="sticky top-0 z-10 flex justify-end border-b border-border/40 bg-transparent px-3 py-1 backdrop-blur">
-      <button
-        className="text-[0.625rem] font-bold text-muted-foreground underline decoration-current/20 underline-offset-4 transition-colors hover:text-foreground"
-        onClick={onToggle}
-        type="button"
-      >
-        {asSource ? t.preview.renderedPreview : t.preview.source}
-      </button>
+    // Fixed height so the header is byte-identical between read and edit modes —
+    // swapping the trailing controls must never move the body below it.
+    <div className="flex h-7 shrink-0 items-center justify-end gap-3 border-b border-border/40 px-3">
+      {showModes &&
+        modes.map(mode => (
+          <button
+            className={cn(
+              'text-[0.625rem] font-bold underline-offset-4 transition-colors',
+              mode === active
+                ? 'text-foreground underline decoration-current/30'
+                : 'text-muted-foreground hover:text-foreground'
+            )}
+            key={mode}
+            onClick={() => onSelect(mode)}
+            type="button"
+          >
+            {label[mode]}
+          </button>
+        ))}
+      {trailing && <div className="flex items-center gap-1.5">{trailing}</div>}
     </div>
   )
 }
 
-// Gutter and Shiki output share `font-mono text-xs leading-relaxed py-3` so
-// each line aligns vertically. The selection overlay relies on the same
-// `text-xs * leading-relaxed = 1.21875rem` line-height to position itself.
-const SOURCE_LINE_HEIGHT_REM = 1.21875
-const SOURCE_PAD_Y_REM = 0.75
+// Cancel / Save controls rendered as the header's trailing slot (not a bar of
+// their own) so edit mode reuses the read-mode header row verbatim.
+function EditControls({
+  dirty,
+  onCancel,
+  onSave,
+  saving
+}: {
+  dirty: boolean
+  onCancel: () => void
+  onSave: () => void
+  saving: boolean
+}) {
+  const { t } = useI18n()
+
+  return (
+    <>
+      <button
+        className="flex items-center gap-1 rounded-md px-1.5 text-[0.625rem] font-bold text-muted-foreground transition-colors hover:bg-accent hover:text-foreground"
+        onClick={onCancel}
+        type="button"
+      >
+        <X className="size-3" />
+        {t.common.cancel}
+      </button>
+      <button
+        className="flex items-center gap-1 rounded-md bg-primary px-2 py-0.5 text-[0.625rem] font-bold text-primary-foreground shadow-xs transition-opacity hover:opacity-90 disabled:opacity-50"
+        disabled={!dirty || saving}
+        onClick={onSave}
+        type="button"
+      >
+        <Check className="size-3" />
+        {saving ? t.common.saving : t.common.save}
+      </button>
+    </>
+  )
+}
 
 interface LineSelection {
   end: number
@@ -337,7 +433,18 @@ function startLineDrag(event: ReactDragEvent<HTMLElement>, filePath: string, { e
 
 function SourceView({ filePath, language, text }: { filePath: string; language: string; text: string }) {
   const { t } = useI18n()
-  const lineCount = useMemo(() => Math.max(1, text.split('\n').length), [text])
+  const chunks = useMemo(() => chunkTextLines(text, SOURCE_CHUNK_LINES), [text])
+  const lastChunk = chunks.at(-1)
+  const totalLines = lastChunk ? lastChunk.start + lastChunk.lines.length : 0
+
+  const { afterRows, beforeRows, endChunk, onScroll, scrollerRef, startChunk } = useFixedRowWindow({
+    overscanRows: SOURCE_OVERSCAN_LINES,
+    rowPx: SOURCE_LINE_PX,
+    rowsPerChunk: SOURCE_CHUNK_LINES,
+    totalRows: totalLines
+  })
+
+  const visibleChunks = chunks.slice(startChunk, endChunk + 1)
   const [selection, setSelection] = useState<LineSelection | null>(null)
   const inSelection = (line: number) => selection != null && line >= selection.start && line <= selection.end
 
@@ -394,69 +501,97 @@ function SourceView({ filePath, language, text }: { filePath: string; language:
   }, [filePath, selection])
 
   return (
-    <div className="grid min-w-max grid-cols-[auto_minmax(0,1fr)] font-mono text-xs leading-relaxed">
-      <div className="select-none py-3 text-right text-muted-foreground/55">
-        {Array.from({ length: lineCount }, (_, index) => {
-          const line = index + 1
-          const selected = inSelection(line)
+    <div className="h-full overflow-auto" onScroll={onScroll} ref={scrollerRef}>
+      <div className="grid min-w-max grid-cols-[auto_minmax(0,1fr)] font-mono text-[0.7rem] leading-relaxed">
+        {beforeRows > 0 && <div aria-hidden className="col-span-2" style={{ height: beforeRows * SOURCE_LINE_PX }} />}
+        {visibleChunks.map(chunk => (
+          <Fragment key={chunk.start}>
+            <div className="select-none text-right text-muted-foreground/55">
+              {chunk.lines.map((_lineText, offset) => {
+                const line = chunk.start + offset + 1
+                const selected = inSelection(line)
 
-          return (
-            <div
-              className={cn(
-                'cursor-pointer px-3 tabular-nums transition-colors',
-                selected
-                  ? 'bg-amber-200/45 text-amber-900 dark:bg-amber-300/20 dark:text-amber-100'
-                  : 'hover:text-foreground'
-              )}
-              draggable
-              key={line}
-              onClick={event => handleLineClick(event, line)}
-              onDragStart={event => handleDragStart(event, line)}
-              title={t.preview.sourceLineTitle}
-            >
-              {line}
+                return (
+                  <div
+                    className={cn(
+                      'h-5 w-9 cursor-pointer pr-2 leading-5 tabular-nums transition-colors',
+                      selected
+                        ? 'bg-amber-200/45 text-amber-900 dark:bg-amber-300/20 dark:text-amber-100'
+                        : 'hover:text-foreground'
+                    )}
+                    draggable
+                    key={line}
+                    onClick={event => handleLineClick(event, line)}
+                    onDragStart={event => handleDragStart(event, line)}
+                    title={t.preview.sourceLineTitle}
+                  >
+                    {line}
+                  </div>
+                )
+              })}
             </div>
-          )
-        })}
-      </div>
-      <div
-        className="relative [&_pre]:m-0 [&_pre]:px-3 [&_pre]:py-3 [&_pre]:bg-transparent!"
-        data-selectable-text="true"
-      >
-        {selection && (
-          <div
-            aria-hidden
-            className="pointer-events-none absolute inset-x-0 bg-amber-200/35 dark:bg-amber-300/10"
-            style={{
-              top: `calc(${SOURCE_PAD_Y_REM}rem + ${selection.start - 1} * ${SOURCE_LINE_HEIGHT_REM}rem)`,
-              height: `calc(${selection.end - selection.start + 1} * ${SOURCE_LINE_HEIGHT_REM}rem)`
-            }}
-          />
-        )}
-        <ShikiHighlighter
-          addDefaultStyles={false}
-          as="div"
-          defaultColor="light-dark()"
-          delay={80}
-          language={language || 'text'}
-          showLanguage={false}
-          theme={SHIKI_THEME}
-        >
-          {text}
-        </ShikiHighlighter>
+            <div className="preview-source-code min-w-0 [&_pre]:m-0" data-selectable-text="true">
+              <ShikiHighlighter
+                addDefaultStyles={false}
+                as="div"
+                defaultColor="light-dark()"
+                delay={80}
+                language={language || 'text'}
+                showLanguage={false}
+                theme={SHIKI_THEME}
+              >
+                {chunk.text}
+              </ShikiHighlighter>
+            </div>
+          </Fragment>
+        ))}
+        {afterRows > 0 && <div aria-hidden className="col-span-2" style={{ height: afterRows * SOURCE_LINE_PX }} />}
       </div>
     </div>
   )
 }
 
+type PreviewViewMode = 'diff' | 'rendered' | 'source'
+
 export function LocalFilePreview({ reloadKey, target }: { reloadKey: number; target: PreviewTarget }) {
   const { t } = useI18n()
   const [state, setState] = useState<LocalPreviewState>({ loading: true })
   const [forcePreview, setForcePreview] = useState(false)
-  const [renderMarkdownAsSource, setRenderMarkdownAsSource] = useState(false)
+  // User-picked view; null = auto (diff when changed, else rendered markdown,
+  // else source). Reset when the previewed file changes.
+  const [userMode, setUserMode] = useState<null | PreviewViewMode>(null)
+  // Spot-editor state. The editor owns its buffer (keyed by `editorKey`); the
+  // live draft + the snapshot the user started from live in refs so typing
+  // never re-renders this (large) component — `dirty` is the only render-worthy
+  // signal and it flips just once when crossing the clean↔dirty boundary.
+  // `selfReload` re-runs the load after a save without the parent.
+  const [editing, setEditing] = useState(false)
+  const draftRef = useRef('')
+  const baselineRef = useRef('')
+  const [dirty, setDirty] = useState(false)
+  const [editorKey, setEditorKey] = useState(0)
+  const [saving, setSaving] = useState(false)
+  const [saveError, setSaveError] = useState<null | string>(null)
+  const [conflict, setConflict] = useState(false)
+  const [selfReload, setSelfReload] = useState(0)
+  // For the bare-`e` shortcut: the read-view root (to detect focus-within) and a
+  // hover flag (no state — only the keydown handler reads it).
+  const readViewRef = useRef<HTMLDivElement>(null)
+  const hoverRef = useRef(false)
   const filePath = filePathForTarget(target)
   const isImage = target.previewKind === 'image'
 
+  useEffect(() => {
+    setUserMode(null)
+    setEditing(false)
+    setDirty(false)
+    setSaving(false)
+    setSaveError(null)
+    setConflict(false)
+    draftRef.current = ''
+    baselineRef.current = ''
+  }, [filePath, reloadKey])
+
   // HTML files are rendered as source code, not in a webview - so they take
   // the same path as plain text files. `previewKind === 'binary'` arrives
   // when the file is forcibly previewed past the binary refusal screen.
@@ -508,6 +643,22 @@ export function LocalFilePreview({ reloadKey, target }: { reloadKey: number; tar
             text: shouldBlock ? undefined : result.text,
             truncated: result.truncated
           })
+
+          // Best-effort: fetch the file's working-tree-vs-HEAD diff so the
+          // preview can offer a DIFF view when there are uncommitted changes.
+          // Empty (clean file / not a repo / remote) just hides the option.
+          if (!shouldBlock) {
+            try {
+              const root = await desktopGitRoot(filePath)
+              const diff = root ? await desktopFileDiff(root, filePath) : ''
+
+              if (active && diff.trim()) {
+                setState(prev => (prev.text === result.text ? { ...prev, diff } : prev))
+              }
+            } catch {
+              // No diff available; the preview just shows source.
+            }
+          }
         }
       } catch (error) {
         if (active) {
@@ -524,7 +675,188 @@ export function LocalFilePreview({ reloadKey, target }: { reloadKey: number; tar
     return () => {
       active = false
     }
-  }, [blockedByTarget, filePath, forcePreview, isImage, isText, reloadKey, target.dataUrl, target.language])
+  }, [blockedByTarget, filePath, forcePreview, isImage, isText, reloadKey, selfReload, target.dataUrl, target.language])
+
+  // Editing is only offered for whole, readable text — never images, binaries,
+  // or files we only loaded the first 512 KB of (saving would drop the tail).
+  const canEdit =
+    isText && !isImage && !blockedByTarget && state.text !== undefined && !state.truncated && !state.binary
+
+  // Per-keystroke: update the draft ref (no render) and only set `dirty` when it
+  // actually changes — React bails on an identical value, so a long typing run
+  // triggers a single re-render at most.
+  const handleEditorChange = useCallback((value: string) => {
+    draftRef.current = value
+    const next = value !== baselineRef.current
+    setDirty(prev => (prev === next ? prev : next))
+  }, [])
+
+  // Publish the unsaved state to the rail so the tab can show a modified dot.
+  // Keyed by url; cleared on unmount/tab-change so a stale dot never lingers.
+  useEffect(() => {
+    setPreviewDirty(target.url, editing && dirty)
+
+    return () => setPreviewDirty(target.url, false)
+  }, [target.url, editing, dirty])
+
+  const beginEdit = () => {
+    const text = state.text ?? ''
+    baselineRef.current = text
+    draftRef.current = text
+    setDirty(false)
+    setEditorKey(key => key + 1)
+    setSaving(false)
+    setSaveError(null)
+    setConflict(false)
+    setEditing(true)
+  }
+
+  // Latest `beginEdit` for the keydown listener, so the listener can stay
+  // subscribed across renders without recreating itself or going stale.
+  const beginEditRef = useRef(beginEdit)
+  beginEditRef.current = beginEdit
+
+  // Bare `e` enters edit mode when the file pane is hovered or focused and no
+  // typable field has focus — a fast, button-free path (double-click felt laggy
+  // because of the browser's click-disambiguation delay).
+  useEffect(() => {
+    if (!canEdit || editing) {
+      return
+    }
+
+    const onKeyDown = (event: KeyboardEvent) => {
+      if (event.key !== 'e' || event.metaKey || event.ctrlKey || event.altKey) {
+        return
+      }
+
+      if (isTypableElement(document.activeElement)) {
+        return
+      }
+
+      const root = readViewRef.current
+      const focusWithin = Boolean(root && document.activeElement && root.contains(document.activeElement))
+
+      if (!hoverRef.current && !focusWithin) {
+        return
+      }
+
+      event.preventDefault()
+      beginEditRef.current()
+    }
+
+    window.addEventListener('keydown', onKeyDown)
+
+    return () => window.removeEventListener('keydown', onKeyDown)
+  }, [canEdit, editing])
+
+  const cancelEdit = () => {
+    setEditing(false)
+    setSaveError(null)
+    setConflict(false)
+  }
+
+  const discardAndReload = () => {
+    setEditing(false)
+    setConflict(false)
+    setSaveError(null)
+    setSelfReload(n => n + 1)
+  }
+
+  const saveEdit = async (force = false) => {
+    if (saving) {
+      return
+    }
+
+    setSaving(true)
+    setSaveError(null)
+
+    try {
+      // Stale-on-disk guard: re-read what's on disk now and compare to the
+      // snapshot the user started from. If something changed underneath (an
+      // agent edit, an external save), don't clobber it silently — surface the
+      // choice. `force` is the user picking "overwrite" from that banner.
+      if (!force) {
+        try {
+          const current = await readTextPreview(filePath)
+
+          if (!current.binary && (current.text ?? '') !== baselineRef.current) {
+            setConflict(true)
+            setSaving(false)
+
+            return
+          }
+        } catch {
+          // Couldn't re-read for the check — fall through and attempt the write.
+        }
+      }
+
+      await writeDesktopFileText(filePath, draftRef.current)
+      baselineRef.current = draftRef.current
+      setDirty(false)
+      setConflict(false)
+      setEditing(false)
+      notifyWorkspaceChanged()
+      setSelfReload(n => n + 1)
+    } catch (error) {
+      setSaveError(error instanceof Error ? error.message : String(error))
+    } finally {
+      setSaving(false)
+    }
+  }
+
+  // Rendered before the loading/error branches so a background re-read (file
+  // watcher, workspace tick) can't unmount the editor and drop the draft. Uses
+  // the SAME container + fixed-height header as the read view so entering edit
+  // never shifts the body — only the trailing controls and the body swap.
+  if (editing) {
+    return (
+      <div className="flex h-full flex-col overflow-hidden bg-transparent">
+        <PreviewModeSwitcher
+          active="source"
+          modes={[]}
+          onSelect={() => {}}
+          trailing={<EditControls dirty={dirty} onCancel={cancelEdit} onSave={() => void saveEdit()} saving={saving} />}
+        />
+        {conflict && (
+          <div className="shrink-0 border-b border-amber-400/40 bg-amber-50 px-3 py-2 text-[0.7rem] text-amber-900 dark:border-amber-300/30 dark:bg-amber-300/10 dark:text-amber-100">
+            <div className="font-semibold">{t.preview.diskChangedTitle}</div>
+            <div className="mt-0.5 leading-relaxed">{t.preview.diskChangedBody}</div>
+            <div className="mt-1.5 flex gap-3">
+              <button
+                className="font-bold underline underline-offset-4 transition-opacity hover:opacity-80"
+                onClick={() => void saveEdit(true)}
+                type="button"
+              >
+                {t.preview.overwrite}
+              </button>
+              <button
+                className="font-bold underline underline-offset-4 transition-opacity hover:opacity-80"
+                onClick={discardAndReload}
+                type="button"
+              >
+                {t.preview.discardReload}
+              </button>
+            </div>
+          </div>
+        )}
+        {saveError && (
+          <div className="shrink-0 border-b border-destructive/40 bg-destructive/10 px-3 py-1.5 text-[0.7rem] text-destructive">
+            {t.preview.saveFailed(saveError)}
+          </div>
+        )}
+        <div className="min-h-0 flex-1 overflow-hidden">
+          <CodeEditor
+            filePath={filePath}
+            initialValue={baselineRef.current}
+            key={editorKey}
+            onCancel={cancelEdit}
+            onChange={handleEditorChange}
+            onSave={() => void saveEdit()}
+          />
+        </div>
+      </div>
+    )
+  }
 
   if (state.loading) {
     return <PageLoader label={t.preview.loading} />
@@ -544,11 +876,7 @@ export function LocalFilePreview({ reloadKey, target }: { reloadKey: number; tar
 
     return (
       <PreviewEmptyState
-        body={
-          binary
-            ? t.preview.binaryBody(target.label)
-            : t.preview.largeBody(target.label, formatBytes(size))
-        }
+        body={binary ? t.preview.binaryBody(target.label) : t.preview.largeBody(target.label, formatBytes(size))}
         primaryAction={{ label: t.preview.previewAnyway, onClick: () => setForcePreview(true) }}
         title={binary ? t.preview.binaryTitle : t.preview.largeTitle}
         tone="warning"
@@ -571,29 +899,79 @@ export function LocalFilePreview({ reloadKey, target }: { reloadKey: number; tar
 
   if (isText && state.text !== undefined) {
     const isMarkdown = (state.language || target.language) === 'markdown'
-    const showRendered = isMarkdown && !renderMarkdownAsSource
+    const hasDiff = Boolean(state.diff && state.diff.trim())
+    // Order the toggle reads left→right; default lands on the most useful view.
+    const modes: PreviewViewMode[] = []
+
+    if (isMarkdown) {
+      modes.push('rendered')
+    }
+
+    modes.push('source')
+
+    if (hasDiff) {
+      modes.push('diff')
+    }
+
+    const autoMode: PreviewViewMode = hasDiff ? 'diff' : isMarkdown ? 'rendered' : 'source'
+    const mode = userMode && modes.includes(userMode) ? userMode : autoMode
 
     return (
-      <div className="h-full overflow-auto bg-transparent">
+      <div
+        className="flex h-full flex-col overflow-hidden bg-transparent"
+        onMouseEnter={() => {
+          hoverRef.current = true
+        }}
+        onMouseLeave={() => {
+          hoverRef.current = false
+        }}
+        ref={readViewRef}
+      >
         {state.truncated && (
           <div className="border-b border-border/60 bg-muted/35 px-3 py-1.5 text-[0.68rem] text-muted-foreground">
             {t.preview.truncated}
           </div>
         )}
-        {isMarkdown && <PreviewToggle asSource={!showRendered} onToggle={() => setRenderMarkdownAsSource(s => !s)} />}
-        {showRendered ? (
-          <MarkdownPreview text={state.text} />
-        ) : (
-          <SourceView filePath={filePath} language={state.language || 'text'} text={state.text} />
-        )}
+        <PreviewModeSwitcher
+          active={mode}
+          modes={modes}
+          onSelect={setUserMode}
+          trailing={
+            canEdit ? (
+              <button
+                className="flex items-center gap-1 text-[0.625rem] font-bold text-muted-foreground underline-offset-4 transition-colors hover:text-foreground"
+                onClick={beginEdit}
+                title={`${t.preview.edit} (e)`}
+                type="button"
+              >
+                <Pencil className="size-3" />
+                {t.preview.edit}
+              </button>
+            ) : null
+          }
+        />
+        <div className="min-h-0 flex-1 overflow-auto">
+          {mode === 'rendered' ? (
+            <MarkdownPreview text={state.text} />
+          ) : mode === 'diff' ? (
+            <FileDiffPanel
+              className="mx-0 mb-0 h-full max-h-none"
+              diff={state.diff ?? ''}
+              fullText={state.text}
+              path={filePath}
+              showLineNumbers
+            />
+          ) : (
+            <SourceView
+              filePath={filePath}
+              language={shikiLanguageForFilename(filePath) || state.language || 'text'}
+              text={state.text}
+            />
+          )}
+        </div>
       </div>
     )
   }
 
-  return (
-    <PreviewEmptyState
-      body={t.preview.noInlineBody(target.mimeType || '')}
-      title={t.preview.noInlineTitle}
-    />
-  )
+  return <PreviewEmptyState body={t.preview.noInlineBody(target.mimeType || '')} title={t.preview.noInlineTitle} />
 }

apps/desktop/src/app/chat/right-rail/preview-pane.test.tsx

@@ -7,7 +7,9 @@ import { PreviewPane } from './preview-pane'
 
 describe('PreviewPane console state', () => {
   beforeEach(() => {
-    vi.stubGlobal('requestAnimationFrame', (callback: FrameRequestCallback) => window.setTimeout(() => callback(Date.now()), 0))
+    vi.stubGlobal('requestAnimationFrame', (callback: FrameRequestCallback) =>
+      window.setTimeout(() => callback(Date.now()), 0)
+    )
     vi.stubGlobal('cancelAnimationFrame', (id: number) => window.clearTimeout(id))
   })
 

apps/desktop/src/app/chat/right-rail/preview.tsx

@@ -3,10 +3,19 @@ import { useEffect, useMemo } from 'react'
 
 import type { SetTitlebarToolGroup } from '@/app/shell/titlebar-controls'
 import { Codicon } from '@/components/ui/codicon'
+import {
+  ContextMenu,
+  ContextMenuContent,
+  ContextMenuItem,
+  ContextMenuSeparator,
+  ContextMenuTrigger
+} from '@/components/ui/context-menu'
 import { Tip } from '@/components/ui/tooltip'
 import { translateNow, useI18n } from '@/i18n'
+import { formatCombo } from '@/lib/keybinds/combo'
 import { cn } from '@/lib/utils'
 import {
+  $panesFlipped,
   $rightRailActiveTabId,
   RIGHT_RAIL_PREVIEW_TAB_ID,
   type RightRailTabId,
@@ -16,10 +25,13 @@ import {
   $filePreviewTabs,
   $previewReloadRequest,
   $previewTarget,
+  closeOtherRightRailTabs,
   closeRightRail,
   closeRightRailTab,
+  closeRightRailTabsToRight,
   type PreviewTarget
 } from '@/store/preview'
+import { $dirtyPreviewUrls } from '@/store/preview-edit'
 
 import { PreviewPane } from './preview-pane'
 
@@ -56,12 +68,16 @@ export function ChatPreviewRail({ onRestartServer, setTitlebarToolGroup }: ChatP
   const { t } = useI18n()
   const previewReloadRequest = useStore($previewReloadRequest)
   const activeTabId = useStore($rightRailActiveTabId)
+  const panesFlipped = useStore($panesFlipped)
   const filePreviewTabs = useStore($filePreviewTabs)
   const previewTarget = useStore($previewTarget)
+  const dirtyPreviewUrls = useStore($dirtyPreviewUrls)
 
   const tabs = useMemo<readonly RailTab[]>(
     () => [
-      ...(previewTarget ? [{ id: RIGHT_RAIL_PREVIEW_TAB_ID, label: t.preview.tab, target: previewTarget } as RailTab] : []),
+      ...(previewTarget
+        ? [{ id: RIGHT_RAIL_PREVIEW_TAB_ID, label: t.preview.tab, target: previewTarget } as RailTab]
+        : []),
       ...filePreviewTabs.map(({ id, target }) => ({ id, label: tabLabelFor(target), target }) as RailTab)
     ],
     [filePreviewTabs, previewTarget, t.preview.tab]
@@ -82,68 +98,109 @@ export function ChatPreviewRail({ onRestartServer, setTitlebarToolGroup }: ChatP
   const isPreview = activeTab.id === RIGHT_RAIL_PREVIEW_TAB_ID
 
   return (
-    <aside className="relative flex h-full w-full min-w-0 flex-col overflow-hidden border-l border-(--ui-stroke-tertiary) bg-(--ui-editor-surface-background) text-(--ui-text-tertiary)">
+    <aside
+      className={cn(
+        'relative flex h-full w-full min-w-0 flex-col overflow-hidden border-(--ui-stroke-tertiary) bg-(--ui-editor-surface-background) text-(--ui-text-tertiary)',
+        panesFlipped ? 'border-r' : 'border-l'
+      )}
+      // Windows/WSLg paint Electron's Window Controls Overlay across our
+      // titlebar band, so the editor-style tab strip (which normally sits IN that
+      // band) would land under the fixed titlebar tools. --right-rail-top-inset
+      // (set by AppShell only when the overlay is present) drops the rail one
+      // titlebar-height so it opens below the band. 0px elsewhere → unchanged.
+      style={{ paddingTop: 'var(--right-rail-top-inset, 0px)' }}
+    >
       <div className="group/rail-tabs flex h-(--titlebar-height) shrink-0 border-b border-(--ui-stroke-tertiary) bg-(--ui-sidebar-surface-background)">
         <div
           className="flex min-w-0 flex-1 overflow-x-auto overflow-y-hidden overscroll-x-contain [-ms-overflow-style:none] [scrollbar-width:none] [&::-webkit-scrollbar]:hidden"
           role="tablist"
         >
-          {tabs.map(tab => {
+          {tabs.map((tab, index) => {
             const active = tab.id === activeTab.id
+            const hasOthers = tabs.length > 1
+            const hasTabsToRight = index < tabs.length - 1
+            const dirty = Boolean(dirtyPreviewUrls[tab.target.url])
 
             return (
-              <div
-                className={cn(
-                  'group/tab relative flex h-full min-w-0 max-w-48 shrink-0 items-center text-[0.6875rem] font-medium [-webkit-app-region:no-drag] last:border-r last:border-(--ui-stroke-quaternary)',
-                  active
-                    ? 'bg-(--ui-editor-surface-background) text-foreground [--tab-bg:var(--ui-editor-surface-background)]'
-                    : 'border-r border-(--ui-stroke-quaternary) text-(--ui-text-tertiary) [--tab-bg:var(--ui-sidebar-surface-background)] hover:bg-(--chrome-action-hover) hover:text-foreground'
-                )}
-                key={tab.id}
-                // Middle-click closes the tab, matching browser/IDE muscle
-                // memory. `onMouseDown` swallows the middle-button press so
-                // Chromium doesn't switch into autoscroll mode.
-                onAuxClick={event => {
-                  if (event.button !== 1) {
-                    return
-                  }
+              <ContextMenu key={tab.id}>
+                <ContextMenuTrigger asChild>
+                  <div
+                    className={cn(
+                      'group/tab relative flex h-full min-w-0 max-w-48 shrink-0 items-center text-[0.6875rem] font-medium [-webkit-app-region:no-drag] last:border-r last:border-(--ui-stroke-quaternary)',
+                      active
+                        ? 'bg-(--ui-editor-surface-background) text-foreground [--tab-bg:var(--ui-editor-surface-background)]'
+                        : 'border-r border-(--ui-stroke-quaternary) text-(--ui-text-tertiary) [--tab-bg:var(--ui-sidebar-surface-background)] hover:bg-(--chrome-action-hover) hover:text-foreground'
+                    )}
+                    // Middle-click closes the tab, matching browser/IDE muscle
+                    // memory. `onMouseDown` swallows the middle-button press so
+                    // Chromium doesn't switch into autoscroll mode.
+                    onAuxClick={event => {
+                      if (event.button !== 1) {
+                        return
+                      }
 
-                  event.preventDefault()
-                  closeRightRailTab(tab.id)
-                }}
-                onMouseDown={event => {
-                  if (event.button === 1) {
-                    event.preventDefault()
-                  }
-                }}
-              >
-                {active && (
-                  <span aria-hidden="true" className="absolute inset-x-0 top-0 h-px bg-(--ui-stroke-primary)" />
-                )}
-                <Tip label={tab.label}>
-                  <button
-                    aria-selected={active}
-                    className="flex h-full min-w-0 max-w-full items-center overflow-hidden pl-3 pr-2 text-left outline-none"
-                    onClick={() => selectRightRailTab(tab.id)}
-                    role="tab"
-                    type="button"
+                      event.preventDefault()
+                      closeRightRailTab(tab.id)
+                    }}
+                    onMouseDown={event => {
+                      if (event.button === 1) {
+                        event.preventDefault()
+                      }
+                    }}
                   >
-                    <span className="block min-w-0 truncate">{tab.label}</span>
-                  </button>
-                </Tip>
-                <span
-                  aria-hidden="true"
-                  className="pointer-events-none absolute inset-y-0 right-0 w-9 bg-[linear-gradient(to_right,transparent,var(--tab-bg)_55%)] opacity-0 transition-opacity group-hover/tab:opacity-100 group-focus-within/tab:opacity-100"
-                />
-                <button
-                  aria-label={t.preview.closeTab(tab.label)}
-                  className="pointer-events-none absolute right-1.5 top-1/2 grid size-4 -translate-y-1/2 place-items-center rounded-sm text-(--ui-text-tertiary) opacity-0 transition-[background-color,color,opacity] hover:bg-(--ui-bg-secondary) hover:text-foreground focus-visible:pointer-events-auto focus-visible:opacity-100 group-hover/tab:pointer-events-auto group-hover/tab:opacity-100 group-focus-within/tab:pointer-events-auto group-focus-within/tab:opacity-100"
-                  onClick={() => closeRightRailTab(tab.id)}
-                  type="button"
-                >
-                  <Codicon name="close" size="0.75rem" />
-                </button>
-              </div>
+                    {active && (
+                      <span aria-hidden="true" className="absolute inset-x-0 top-0 h-px bg-(--ui-stroke-primary)" />
+                    )}
+                    <Tip label={tab.target.path || tab.target.url || tab.label}>
+                      <button
+                        aria-selected={active}
+                        className="flex h-full min-w-0 max-w-full items-center overflow-hidden pl-3 pr-2 text-left outline-none"
+                        onClick={() => selectRightRailTab(tab.id)}
+                        role="tab"
+                        type="button"
+                      >
+                        <span className="block min-w-0 truncate">{tab.label}</span>
+                      </button>
+                    </Tip>
+                    <span
+                      aria-hidden="true"
+                      className="pointer-events-none absolute inset-y-0 right-0 w-9 bg-[linear-gradient(to_right,transparent,var(--tab-bg)_55%)] opacity-0 transition-opacity group-hover/tab:opacity-100 group-focus-within/tab:opacity-100"
+                    />
+                    {dirty && (
+                      <span
+                        aria-hidden="true"
+                        className="pointer-events-none absolute right-1.5 top-1/2 grid size-4 -translate-y-1/2 place-items-center opacity-100 transition-opacity group-hover/tab:opacity-0 group-focus-within/tab:opacity-0"
+                      >
+                        {/* Amber (our warn color); a tab-bg ring + soft drop keeps it
+                            legible where it overlaps the filename. */}
+                        <span className="size-2 rounded-full bg-amber-500 shadow-[0_0_0_2px_var(--tab-bg),0_1px_2px_rgba(0,0,0,0.45)] dark:bg-amber-400" />
+                      </span>
+                    )}
+                    <button
+                      aria-label={t.preview.closeTab(tab.label)}
+                      className="pointer-events-none absolute right-1.5 top-1/2 grid size-4 -translate-y-1/2 place-items-center rounded-sm text-(--ui-text-tertiary) opacity-0 transition-[background-color,color,opacity] hover:bg-(--ui-bg-secondary) hover:text-foreground focus-visible:pointer-events-auto focus-visible:opacity-100 group-hover/tab:pointer-events-auto group-hover/tab:opacity-100 group-focus-within/tab:pointer-events-auto group-focus-within/tab:opacity-100"
+                      onClick={() => closeRightRailTab(tab.id)}
+                      type="button"
+                    >
+                      <Codicon name="close" size="0.75rem" />
+                    </button>
+                  </div>
+                </ContextMenuTrigger>
+                <ContextMenuContent>
+                  <ContextMenuItem onSelect={() => closeRightRailTab(tab.id)}>
+                    {t.common.close}
+                    <span className="ml-auto pl-4 text-(--ui-text-tertiary)">{formatCombo('mod+w')}</span>
+                  </ContextMenuItem>
+                  <ContextMenuItem disabled={!hasOthers} onSelect={() => closeOtherRightRailTabs(tab.id)}>
+                    {t.preview.closeOthers}
+                  </ContextMenuItem>
+                  <ContextMenuItem disabled={!hasTabsToRight} onSelect={() => closeRightRailTabsToRight(tab.id)}>
+                    {t.preview.closeToRight}
+                  </ContextMenuItem>
+                  <ContextMenuSeparator />
+                  <ContextMenuItem onSelect={closeRightRail}>{t.preview.closeAll}</ContextMenuItem>
+                </ContextMenuContent>
+              </ContextMenu>
             )
           })}
         </div>