release: v0.9.5

The first message sent from Discord was silently dropped because the
cog delegated the initial message to the bridge via CreateWorkstream-
Message, but the bridge published response events to the per-workstream
Redis pub/sub channel before the Discord bot had subscribed to it.
Redis pub/sub is fire-and-forget — events with no subscribers are lost.

Fix: create the workstream with initial_message="" (no delegation),
subscribe to the per-workstream event channel, then send the message
through router.send_message() — the same path the second message
already uses successfully.

Applied to both @mention handler and /ask slash command.

.github/renovate.json

@@ -5,55 +5,40 @@
     "helpers:pinGitHubActionDigests",
     ":separateMajorReleases"
   ],
-  "labels": [
-    "dependencies"
+  "gitIgnoredAuthors": [
+    "41898282+github-actions[bot]@users.noreply.github.com"
   ],
+  "labels": ["dependencies"],
   "prConcurrentLimit": 5,
   "prHourlyLimit": 2,
-  "schedule": [
-    "before 9am on Monday"
-  ],
+  "schedule": ["before 9am on Monday"],
   "timezone": "America/New_York",
   "lockFileMaintenance": {
     "enabled": true,
-    "schedule": [
-      "before 9am on Monday"
-    ]
+    "schedule": ["before 9am on Monday"]
   },
   "customManagers": [
     {
       "customType": "regex",
       "description": "Track vendored KaTeX version",
-      "managerFilePatterns": [
-        "/pyproject\\.toml$/"
-      ],
-      "matchStrings": [
-        "katex-(?<currentValue>[\\d.]+)/"
-      ],
+      "managerFilePatterns": ["/pyproject\\.toml$/"],
+      "matchStrings": ["katex-(?<currentValue>[\\d.]+)/"],
       "depNameTemplate": "katex",
       "datasourceTemplate": "npm"
     },
     {
       "customType": "regex",
       "description": "Track vendored Highlight.js version",
-      "managerFilePatterns": [
-        "/pyproject\\.toml$/"
-      ],
-      "matchStrings": [
-        "hljs-(?<currentValue>[\\d.]+)/"
-      ],
+      "managerFilePatterns": ["/pyproject\\.toml$/"],
+      "matchStrings": ["hljs-(?<currentValue>[\\d.]+)/"],
       "depNameTemplate": "highlight.js",
       "datasourceTemplate": "npm"
     },
     {
       "customType": "regex",
       "description": "Track vendored Mermaid version",
-      "managerFilePatterns": [
-        "/pyproject\\.toml$/"
-      ],
-      "matchStrings": [
-        "mermaid-(?<currentValue>[\\d.]+)/"
-      ],
+      "managerFilePatterns": ["/pyproject\\.toml$/"],
+      "matchStrings": ["mermaid-(?<currentValue>[\\d.]+)/"],
       "depNameTemplate": "mermaid",
       "datasourceTemplate": "npm"
     }
@@ -62,14 +47,8 @@
     {
       "description": "LLM SDKs — always review manually",
       "groupName": "LLM SDKs",
-      "matchPackageNames": [
-        "openai",
-        "anthropic",
-        "mcp"
-      ],
-      "schedule": [
-        "before 9am on Monday"
-      ],
+      "matchPackageNames": ["openai", "anthropic", "mcp"],
+      "schedule": ["before 9am on Monday"],
       "automerge": false
     },
     {
@@ -83,73 +62,38 @@
         "httpx-sse",
         "pydantic"
       ],
-      "schedule": [
-        "before 9am on Wednesday"
-      ],
+      "schedule": ["before 9am on Wednesday"],
       "automerge": true,
-      "matchUpdateTypes": [
-        "patch"
-      ]
+      "matchUpdateTypes": ["patch"]
     },
     {
       "description": "Database layer",
       "groupName": "Database",
-      "matchPackageNames": [
-        "sqlalchemy",
-        "alembic",
-        "psycopg"
-      ],
-      "schedule": [
-        "before 9am on Wednesday"
-      ],
+      "matchPackageNames": ["sqlalchemy", "alembic", "psycopg"],
+      "schedule": ["before 9am on Wednesday"],
       "automerge": true,
-      "matchUpdateTypes": [
-        "patch"
-      ]
+      "matchUpdateTypes": ["patch"]
     },
     {
       "description": "Security-critical — always review manually",
       "groupName": "Security",
-      "matchPackageNames": [
-        "PyJWT",
-        "pyjwt",
-        "bcrypt"
-      ],
+      "matchPackageNames": ["PyJWT", "pyjwt", "bcrypt"],
       "automerge": false
     },
     {
       "description": "Infrastructure dependencies",
       "groupName": "Infrastructure",
-      "matchPackageNames": [
-        "structlog",
-        "redis",
-        "croniter",
-        "discord.py"
-      ],
-      "schedule": [
-        "before 9am on the first day of the month"
-      ],
+      "matchPackageNames": ["structlog", "redis", "croniter", "discord.py"],
+      "schedule": ["before 9am on the first day of the month"],
       "automerge": true,
-      "matchUpdateTypes": [
-        "patch"
-      ]
+      "matchUpdateTypes": ["patch"]
     },
     {
-      "description": "Vendored JS — requires manual file download after merge",
+      "description": "Vendored JS — CI workflow downloads files automatically",
       "groupName": "Vendored JS",
-      "matchPackageNames": [
-        "katex",
-        "highlight.js",
-        "mermaid"
-      ],
-      "schedule": [
-        "before 9am on the first day of the month"
-      ],
-      "automerge": false,
-      "prBodyNotes": [
-        "This PR updates version references only.",
-        "After merging, run `scripts/update-vendored-js.sh <lib> <version>` to download the actual files."
-      ]
+      "matchPackageNames": ["katex", "highlight.js", "mermaid"],
+      "schedule": ["before 9am on the first day of the month"],
+      "automerge": false
     },
     {
       "description": "Dev/test tooling",
@@ -162,46 +106,29 @@
         "pytest-cov",
         "pre-commit"
       ],
-      "schedule": [
-        "before 9am on the first day of the month"
-      ],
+      "schedule": ["before 9am on the first day of the month"],
       "automerge": true,
-      "matchUpdateTypes": [
-        "patch"
-      ]
+      "matchUpdateTypes": ["patch"]
     },
     {
       "description": "Docker base images",
       "groupName": "Docker Images",
-      "matchManagers": [
-        "dockerfile",
-        "docker-compose"
-      ],
-      "schedule": [
-        "before 9am on the first day of the month"
-      ],
+      "matchManagers": ["dockerfile", "docker-compose"],
+      "schedule": ["before 9am on the first day of the month"],
       "automerge": false
     },
     {
       "description": "TypeScript SDK dev dependencies",
       "groupName": "TypeScript SDK",
-      "matchFileNames": [
-        "sdk/typescript/**"
-      ],
-      "schedule": [
-        "before 9am on the first day of the month"
-      ],
+      "matchFileNames": ["sdk/typescript/**"],
+      "schedule": ["before 9am on the first day of the month"],
       "automerge": true,
-      "matchUpdateTypes": [
-        "patch"
-      ]
+      "matchUpdateTypes": ["patch"]
     },
     {
       "description": "GitHub Actions — group all action updates",
       "groupName": "GitHub Actions",
-      "matchManagers": [
-        "github-actions"
-      ],
+      "matchManagers": ["github-actions"],
       "automerge": false
     }
   ]

.github/workflows/ci.yml

@@ -47,11 +47,37 @@ jobs:
           name: coverage-${{ matrix.python-version }}
           path: coverage.xml
 
+  test-postgres:
+    runs-on: ubuntu-latest
+    services:
+      postgres:
+        image: postgres:18
+        env:
+          POSTGRES_USER: postgres
+          POSTGRES_PASSWORD: postgres
+          POSTGRES_DB: turnstone_test
+        ports:
+          - 5432:5432
+        options: >-
+          --health-cmd="pg_isready -U postgres"
+          --health-interval=10s
+          --health-timeout=5s
+          --health-retries=5
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6
+        with:
+          python-version: "3.14"
+      - run: pip install -e ".[test,mq,postgres]"
+      - run: pytest tests/ -m "not live" --storage-backend=postgresql -q
+        env:
+          TURNSTONE_TEST_PG_URL: postgresql+psycopg://postgres:postgres@localhost:5432/turnstone_test
+
   lock-check:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6
-      - uses: astral-sh/setup-uv@e06108dd0aef18192324c70427afc47652e63a82 # v7
+      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7
         with:
           uv-version: "0.9.18"
       - run: uv lock --check
@@ -60,7 +86,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6
-      - uses: astral-sh/setup-uv@e06108dd0aef18192324c70427afc47652e63a82 # v7
+      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7
         with:
           uv-version: "0.9.18"
       - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6

.github/workflows/publish.yml

@@ -19,7 +19,7 @@ jobs:
           python-version: "3.14"
       - run: pip install build
       - run: python -m build
-      - uses: pypa/gh-action-pypi-publish@release/v1
+      - uses: pypa/gh-action-pypi-publish@ed0c53931b1dc9bd32cbe73a98c7f6766f8a527e # release/v1
 
       - name: Create GitHub Release
         uses: softprops/action-gh-release@153bb8e04406b158c6c84fc1615b65b24149a1fe # v2

.github/workflows/vendor-js.yml

@@ -0,0 +1,86 @@
+name: Complete Vendored JS Updates
+
+# When Renovate bumps a vendored JS version in pyproject.toml, this
+# workflow downloads the actual files and commits them to the PR branch
+# so the PR is merge-ready without manual intervention.
+#
+# Note: the commit is made with GITHUB_TOKEN, so it won't re-trigger CI
+# automatically.  The reviewer should re-run CI once this workflow passes,
+# or Renovate's next rebase will trigger it.
+
+on:
+  pull_request:
+    paths:
+      - pyproject.toml
+  workflow_dispatch:
+    inputs:
+      pr_number:
+        description: "PR number to update"
+        required: true
+        type: number
+
+permissions:
+  contents: write
+  pull-requests: read
+
+jobs:
+  vendor-js:
+    if: github.actor == 'renovate[bot]' || github.event_name == 'workflow_dispatch'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Resolve PR head ref
+        id: ref
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          if [[ "${{ github.event_name }}" == "workflow_dispatch" ]]; then
+            ref=$(gh pr view "${{ inputs.pr_number }}" --repo "${{ github.repository }}" --json headRefName -q .headRefName)
+          else
+            ref="${{ github.head_ref }}"
+          fi
+          echo "head_ref=${ref}" >> "$GITHUB_OUTPUT"
+
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6
+        with:
+          ref: ${{ steps.ref.outputs.head_ref }}
+
+      - name: Detect vendored JS changes
+        id: detect
+        run: |
+          updates=()
+          for lib in katex hljs mermaid; do
+            version=$(grep -oE "${lib}-[0-9.]+" pyproject.toml | head -1 | sed "s/${lib}-//")
+            [[ -z "$version" ]] && continue
+            [[ -d "turnstone/shared_static/${lib}-${version}" ]] && continue
+            updates+=("${lib}:${version}")
+          done
+          if [[ ${#updates[@]} -eq 0 ]]; then
+            echo "found=false" >> "$GITHUB_OUTPUT"
+          else
+            echo "found=true" >> "$GITHUB_OUTPUT"
+            printf '%s\n' "${updates[@]}" > /tmp/updates.txt
+            echo "Libs to update:"
+            cat /tmp/updates.txt
+          fi
+
+      - name: Download vendored files
+        if: steps.detect.outputs.found == 'true'
+        run: |
+          while IFS=: read -r lib version; do
+            echo "::group::Updating ${lib} to ${version}"
+            bash scripts/update-vendored-js.sh "$lib" "$version"
+            echo "::endgroup::"
+          done < /tmp/updates.txt
+
+      - name: Commit and push
+        if: steps.detect.outputs.found == 'true'
+        run: |
+          git config user.name "github-actions[bot]"
+          git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
+          git add -A
+          if git diff --cached --quiet; then
+            echo "No changes to commit"
+            exit 0
+          fi
+          git commit -m "chore: download vendored JS files"
+          git push

.trivyignore

@@ -0,0 +1,4 @@
+# libexpat integer overflow — no fix available in Debian repos yet
+# https://avd.aquasec.com/nvd/cve-2026-25210
+# Review: remove this entry once a patched libexpat1 is published
+CVE-2026-25210

Dockerfile

@@ -8,10 +8,14 @@ FROM python:3.14-slim
 LABEL org.opencontainers.image.title="turnstone" \
       org.opencontainers.image.description="Multi-node AI orchestration platform"
 
-COPY --from=ghcr.io/astral-sh/uv:0.10.10 /uv /usr/local/bin/uv
+COPY --from=ghcr.io/astral-sh/uv:0.11.2 /uv /usr/local/bin/uv
 
-# System dependencies for psycopg (PostgreSQL client library)
-RUN apt-get update && apt-get upgrade -y && apt-get install -y --no-install-recommends libpq5 \
+# Remove the slim image's man page exclusion so man-db has actual content
+RUN rm -f /etc/dpkg/dpkg.cfg.d/docker
+
+# System dependencies: psycopg (libpq5), developer tooling for agent workflows
+RUN apt-get update && apt-get upgrade -y && apt-get install -y --no-install-recommends \
+    libpq5 git curl jq man-db manpages \
     && rm -rf /var/lib/apt/lists/*
 
 # Non-root user
@@ -25,12 +29,12 @@ ENV UV_COMPILE_BYTECODE=1
 # Install dependencies first (cached layer — only re-runs when deps change)
 COPY pyproject.toml uv.lock README.md LICENSE ./
 RUN uv sync --frozen --no-install-project --no-dev \
-    --extra mq --extra console --extra sim --extra postgres --extra discord --extra anthropic
+    --extra all
 
 # Install the project itself
 COPY turnstone/ turnstone/
 RUN uv sync --frozen --no-dev \
-    --extra mq --extra console --extra sim --extra postgres --extra discord --extra anthropic
+    --extra all
 
 # Add venv to PATH so entry points are found
 ENV PATH="/app/.venv/bin:$PATH"

README.md

@@ -5,9 +5,11 @@
 [![Python](https://img.shields.io/pypi/pyversions/turnstone)](https://pypi.org/project/turnstone/)
 [![License](https://img.shields.io/badge/license-BSL--1.1-blue)](LICENSE)
 
-Multi-node AI orchestration platform. Deploy tool-using AI agents across a cluster of servers, driven by message queues or interactive interfaces.
+Experimental multi-node AI orchestration platform. Deploy tool-using AI agents across a cluster of servers, driven by message queues or interactive interfaces.
 
-Named after the [Ruddy Turnstone](https://en.wikipedia.org/wiki/Ruddy_turnstone) — a bird that flips rocks to expose what's hiding underneath.
+> **Beta — Use at your own risk.** Turnstone is under active development and has not reached a stable release. APIs, configuration formats, and database schemas may change between versions without migration paths. We make no guarantees of determinism, reliability, or backward compatibility. Evaluate thoroughly before deploying to any environment where these properties matter.
+
+Named after the [Ruddy Turnstone](https://en.wikipedia.org/wiki/Ruddy_turnstone) (*Arenaria interpres*) — a shorebird that flips stones to discover what's hiding underneath.
 
 ## What it does
 
@@ -145,7 +147,7 @@ Turnstone includes a built-in governance layer for enterprise deployments — ma
 - **RBAC** — 15 granular permissions, 3 built-in roles (admin / operator / viewer), custom roles, privilege escalation prevention
 - **OIDC SSO** — single sign-on via any OpenID Connect provider (Okta, Azure AD, Google, Keycloak); Authorization Code Flow with PKCE, auto-provisioning, claim-based role mapping with demotion propagation; see [docs/oidc.md](docs/oidc.md)
 - **Tool policies** — glob-pattern rules (`allow` / `deny` / `ask`) with priority ordering; automate approvals or lock down dangerous tools
-- **Skills** — reusable behavioral profiles with system prompts, `{{variable}}` substitution, session config (model, temperature, token budget), install-time security scanning, version history, external discovery (skills.sh / GitHub), and runtime `load_skill` tool for model-driven skill activation
+- **Skills** — reusable behavioral profiles with system prompts, `{{variable}}` substitution, session config (model, temperature, token budget), install-time security scanning, version history, external discovery (skills.sh / GitHub), and runtime `skill` tool for model-driven skill activation
 - **Usage tracking** — per-request token and tool metrics, aggregation by day / model / user, automatic 90-day pruning
 - **Audit logging** — append-only event trail for all admin mutations, IP-aware, 365-day retention
 
@@ -288,7 +290,7 @@ All entry points read `~/.config/turnstone/config.toml`. CLI flags override conf
 [api]
 base_url = "http://localhost:8000/v1"
 api_key = ""
-tavily_key = ""        # only needed for local/vLLM models without native search
+# tavily_key = ""      # only needed for local/vLLM models without native search
 
 [model]
 name = ""              # empty = auto-detect
@@ -308,7 +310,7 @@ search_max_results = 5     # max tools returned per search query
 [server]
 host = "0.0.0.0"
 port = 8080
-max_workstreams = 10       # auto-evicts oldest idle when full
+max_workstreams = 50       # auto-evicts oldest idle when full
 
 [redis]
 host = "localhost"
@@ -340,7 +342,7 @@ burst = 20
 backend = "sqlite"     # "sqlite" (default) or "postgresql"
 path = ".turnstone.db" # SQLite file path (relative to working directory)
 # url = "postgresql+psycopg://user:pass@host:5432/turnstone"  # PostgreSQL
-# pool_size = 5        # PostgreSQL connection pool size
+# pool_size = 2        # PostgreSQL connection pool size (per process)
 
 [judge]
 enabled = true         # intent validation for tool approvals (--no-judge to disable)
@@ -394,7 +396,7 @@ Idle workstreams are automatically cleaned up after 2 hours (configurable). In m
 - `turnstone_judge_llm_latency_seconds` — LLM judge evaluation latency histogram
 - `turnstone_judge_enabled` — whether the intent validation judge is active (0/1)
 
-Per-workstream metrics are labeled by `ws_id` (bounded to 10 max workstreams).
+Per-workstream metrics are labeled by `ws_id` (bounded by `[server].max_workstreams`).
 
 ### Health & Rate Limiting
 
@@ -404,7 +406,7 @@ Per-workstream metrics are labeled by `ws_id` (bounded to 10 max workstreams).
 
 **Per-IP rate limiting.** When `[ratelimit].enabled` is true, each client IP is tracked with a token-bucket limiter (`requests_per_second` / `burst`). Rate limiting is applied in `do_GET`/`do_POST` after authentication but before route dispatch. `/health` and `/metrics` are exempt. Requests that exceed the limit receive HTTP 429 with a `Retry-After` header.
 
-**Workstream eviction.** When `WorkstreamManager.create()` would exceed `max_workstreams`, the oldest IDLE workstream is automatically evicted and the `turnstone_workstreams_evicted_total` counter is incremented. Configure via `[server].max_workstreams` (default 10).
+**Workstream eviction.** When `WorkstreamManager.create()` would exceed `max_workstreams`, the oldest IDLE workstream is automatically evicted and the `turnstone_workstreams_evicted_total` counter is incremented. Configure via `[server].max_workstreams` (default 50).
 
 ## Requirements
 
@@ -413,6 +415,7 @@ Per-workstream metrics are labeled by `ws_id` (bounded to 10 max workstreams).
 - Redis (for message queue bridge — `pip install turnstone[mq]`)
 - Anthropic provider (optional — `pip install turnstone[anthropic]`)
 - PostgreSQL (optional, for production — `pip install turnstone[postgres]`)
+- Math sandbox packages (optional — `pip install turnstone[sandbox]` for sympy, numpy, scipy, pytest)
 - [Git LFS](https://git-lfs.com/) (for cloning — diagram PNGs are stored in LFS)
 
 ## License

deploy/docker-compose.tls.yml

@@ -0,0 +1,146 @@
+# TLS overlay — enables mTLS across the turnstone cluster.
+#
+# Usage (requires base compose.yaml with production profile):
+#   docker compose -f compose.yaml -f deploy/docker-compose.tls.yml --profile production up
+#
+# The tls-init service bootstraps a CA and issues a cert for Redis.
+# All turnstone services auto-provision their own certs via the
+# console's ACME endpoint.
+
+services:
+  # Bootstrap: create CA + Redis cert before anything starts.
+  # Runs as root to create directories in the volume, then chowns
+  # to turnstone:turnstone with restrictive perms (keys 0600).
+  tls-init:
+    build: .
+    user: root
+    command:
+      - sh
+      - -c
+      - |
+        set -e
+        turnstone-admin tls-bootstrap --out /certs --issue redis
+        chown -R turnstone:turnstone /certs
+        find /certs -type d -exec chmod 750 {} +
+        find /certs -type f -name '*key.pem' -exec chmod 600 {} +
+        find /certs -type f ! -name '*key.pem' -exec chmod 640 {} +
+    volumes:
+      - tls-certs:/certs
+    networks:
+      - turnstone-net
+    restart: "no"
+
+  # Console: runs the internal CA + ACME server
+  console:
+    depends_on:
+      tls-init:
+        condition: service_completed_successfully
+    volumes:
+      - tls-certs:/certs:ro
+    environment:
+      TURNSTONE_TLS_ENABLED: "true"
+      TURNSTONE_TLS_SANS: "console"
+      TURNSTONE_CONSOLE_URL: "http://console:8090"
+    command:
+      - turnstone-console
+      - --host=0.0.0.0
+      - --port=8090
+      - --redis-host=redis
+      - --redis-port=6379
+      - --poll-interval=${CONSOLE_POLL_INTERVAL:-10}
+      - --redis-tls
+      - --redis-tls-ca=/certs/ca.pem
+
+  # Server: auto-provisions certs via console ACME, serves HTTPS
+  server:
+    depends_on:
+      console:
+        condition: service_healthy
+    volumes:
+      - tls-certs:/certs:ro
+    environment:
+      TURNSTONE_TLS_ENABLED: "true"
+      TURNSTONE_TLS_SANS: "server"
+    # Disable healthcheck — server serves HTTPS with mTLS which the
+    # stdlib healthcheck script can't satisfy. The base compose
+    # healthcheck uses plain HTTP which won't work on an HTTPS listener.
+    # TODO: wire healthcheck with client cert from /certs volume
+    healthcheck:
+      disable: true
+
+  # Bridge: mTLS to server + Redis TLS
+  bridge:
+    depends_on:
+      console:
+        condition: service_healthy
+      server:
+        condition: service_started
+      redis:
+        condition: service_healthy
+    volumes:
+      - tls-certs:/certs:ro
+    environment:
+      TURNSTONE_TLS_ENABLED: "true"
+      TURNSTONE_TLS_SANS: "bridge"
+    command:
+      - turnstone-bridge
+      - --server-url=http://server:8080
+      - --redis-host=redis
+      - --redis-port=6379
+      - --heartbeat-ttl=${HEARTBEAT_TTL:-60}
+      - --approval-timeout=${APPROVAL_TIMEOUT:-3600}
+      - --redis-tls
+      - --redis-tls-ca=/certs/ca.pem
+
+  # Channel: Redis TLS
+  channel:
+    depends_on:
+      console:
+        condition: service_healthy
+      redis:
+        condition: service_healthy
+    volumes:
+      - tls-certs:/certs:ro
+    environment:
+      TURNSTONE_TLS_ENABLED: "true"
+      TURNSTONE_TLS_SANS: "channel"
+    command:
+      - sh
+      - -c
+      - >-
+        turnstone-channel
+        --redis-host=redis
+        --redis-port=6379
+        --redis-tls
+        --redis-tls-ca=/certs/ca.pem
+        --http-host=0.0.0.0
+        $${TURNSTONE_DISCORD_GUILD:+--discord-guild $$TURNSTONE_DISCORD_GUILD}
+
+  # Redis: TLS with certs from bootstrap
+  redis:
+    depends_on:
+      tls-init:
+        condition: service_completed_successfully
+    volumes:
+      - tls-certs:/certs:ro
+    command:
+      - sh
+      - -c
+      - |
+        ARGS="--tls-port 6379 --port 0 \
+          --tls-cert-file /certs/certs/redis/cert.pem \
+          --tls-key-file /certs/certs/redis/key.pem \
+          --tls-ca-cert-file /certs/ca.pem \
+          --tls-auth-clients no"
+        if [ -n "$$REDIS_PASSWORD" ]; then
+          ARGS="$$ARGS --requirepass $$REDIS_PASSWORD"
+        fi
+        exec redis-server $$ARGS
+    healthcheck:
+      test: ["CMD-SHELL", "if [ -n \"$$REDIS_PASSWORD\" ]; then redis-cli --tls --cacert /certs/ca.pem -a $$REDIS_PASSWORD ping; else redis-cli --tls --cacert /certs/ca.pem ping; fi"]
+      interval: 5s
+      timeout: 3s
+      retries: 5
+
+volumes:
+  tls-certs:

deploy/openshell/routes.yaml

@@ -0,0 +1,49 @@
+# OpenShell inference routing for Turnstone.
+#
+# When using inference routing, the sandbox process connects to
+# https://inference.local instead of the real LLM API. The OpenShell
+# proxy intercepts, rewrites credentials, and forwards to the backend.
+#
+# This keeps real API keys out of the sandbox entirely — the process
+# only sees opaque placeholder tokens in its environment.
+#
+# Usage:
+#   openshell sandbox run \
+#     --inference-routes deploy/openshell/routes.yaml \
+#     ...
+#
+# Then start turnstone with:
+#   python3 -m turnstone.server --base-url https://inference.local
+#
+# CUSTOMIZE: uncomment one of the provider blocks below.
+
+routes:
+
+  # --- OpenAI ---
+  # - name: inference.local
+  #   endpoint: https://api.openai.com/v1
+  #   model: gpt-5
+  #   provider_type: openai
+  #   protocols:
+  #     - openai_chat_completions
+  #     - model_discovery
+  #   api_key_env: OPENAI_API_KEY
+
+  # --- Anthropic ---
+  # - name: inference.local
+  #   endpoint: https://api.anthropic.com
+  #   model: claude-sonnet-4-6
+  #   provider_type: anthropic
+  #   protocols:
+  #     - anthropic_messages
+  #   api_key_env: ANTHROPIC_API_KEY
+
+  # --- Local model server (vLLM / llama.cpp) ---
+  # No secret resolution needed — local servers typically have no auth.
+  # Omit both api_key and api_key_env to skip credential injection.
+  # - name: inference.local
+  #   endpoint: http://localhost:8000/v1
+  #   model: meta-llama/Llama-3.1-70B-Instruct
+  #   protocols:
+  #     - openai_chat_completions
+  #     - model_discovery

deploy/openshell/turnstone-policy.yaml

@@ -0,0 +1,333 @@
+# OpenShell sandbox policy for Turnstone AI orchestration platform.
+#
+# This policy wraps a turnstone-server process (the primary sandbox target).
+# The bridge, console, and channel gateway are separate processes that would
+# each need their own sandbox with a tailored policy variant.
+#
+# Usage:
+#   openshell sandbox run \
+#     --policy deploy/openshell/turnstone-policy.yaml \
+#     --workdir /project \
+#     -- python3 -m turnstone.server --host 0.0.0.0 --port 8080
+#
+# For inference routing (keeps real API keys out of the sandbox):
+#   openshell sandbox run \
+#     --policy deploy/openshell/turnstone-policy.yaml \
+#     --inference-routes deploy/openshell/routes.yaml \
+#     --workdir /project \
+#     -- python3 -m turnstone.server --host 0.0.0.0 --port 8080 \
+#         --base-url https://inference.local
+#
+# Note: inference.local is intercepted by the OpenShell proxy before
+# network policy evaluation — no network_policies entry is needed for it.
+#
+# Customization points (search for "CUSTOMIZE"):
+#   - OIDC issuer endpoint
+#   - Redis host/port (if not localhost)
+#   - MCP HTTP server endpoints
+#   - Additional tool binaries
+#   - web_fetch domain allowlist
+
+version: 1
+
+# ---------------------------------------------------------------------------
+# Filesystem: Landlock kernel enforcement
+# ---------------------------------------------------------------------------
+# Static — cannot be changed after sandbox creation.
+# include_workdir adds the --workdir path to read_write automatically.
+
+filesystem_policy:
+  include_workdir: true
+
+  read_only:
+    # Python runtime + installed packages (includes turnstone package)
+    - /usr
+    - /lib
+    - /lib64
+    # System essentials
+    - /etc
+    - /proc
+    - /dev/urandom
+    # Turnstone config (read-only — writes go to database)
+    # CUSTOMIZE: adjust if config lives elsewhere
+    - /home/sandbox/.config/turnstone
+
+  read_write:
+    # Working directory is added via include_workdir
+    # Temp files (bash tool scripts, eval workdirs)
+    - /tmp
+    # Shell redirections (2>/dev/null)
+    - /dev/null
+    # SQLite database (default location is workdir, covered by include_workdir)
+    # Logs
+    - /var/log
+
+landlock:
+  # best_effort: degrade gracefully on kernels without Landlock (< 5.13)
+  # Change to hard_requirement for production hardened deployments
+  compatibility: best_effort
+
+# ---------------------------------------------------------------------------
+# Process: privilege separation
+# ---------------------------------------------------------------------------
+
+process:
+  run_as_user: sandbox
+  run_as_group: sandbox
+
+# ---------------------------------------------------------------------------
+# Network: per-endpoint, per-binary allowlisting
+# ---------------------------------------------------------------------------
+# Default-deny. Only listed host:port pairs are reachable.
+# Child processes (MCP servers, bash subcommands) inherit the network
+# namespace — they cannot bypass the proxy.
+
+network_policies:
+
+  # --- LLM API providers ---
+
+  openai_api:
+    name: openai-api
+    endpoints:
+      - host: api.openai.com
+        port: 443
+    binaries:
+      - path: /usr/bin/python3*
+      - path: /usr/local/bin/python3*
+
+  anthropic_api:
+    name: anthropic-api
+    endpoints:
+      - host: api.anthropic.com
+        port: 443
+    binaries:
+      - path: /usr/bin/python3*
+      - path: /usr/local/bin/python3*
+
+  # --- Web search fallback (Tavily) ---
+
+  tavily_api:
+    name: tavily-search
+    endpoints:
+      - host: api.tavily.com
+        port: 443
+    binaries:
+      - path: /usr/bin/python3*
+      - path: /usr/local/bin/python3*
+
+  # --- Skill discovery ---
+
+  skills_registry:
+    name: skills-registry
+    endpoints:
+      - host: skills.sh
+        port: 443
+    binaries:
+      - path: /usr/bin/python3*
+      - path: /usr/local/bin/python3*
+
+  github_api:
+    name: github-api
+    endpoints:
+      - host: api.github.com
+        port: 443
+        protocol: rest
+        tls: terminate
+        enforcement: enforce
+        access: read-only
+      - host: raw.githubusercontent.com
+        port: 443
+    binaries:
+      - path: /usr/bin/python3*
+      - path: /usr/local/bin/python3*
+
+  mcp_registry:
+    name: mcp-registry
+    endpoints:
+      - host: registry.modelcontextprotocol.io
+        port: 443
+        protocol: rest
+        tls: terminate
+        enforcement: enforce
+        access: read-only
+    binaries:
+      - path: /usr/bin/python3*
+      - path: /usr/local/bin/python3*
+
+  # --- OIDC SSO ---
+  # CUSTOMIZE: replace with your identity provider's hostname
+
+  # oidc_provider:
+  #   name: oidc-provider
+  #   endpoints:
+  #     - host: login.example.com
+  #       port: 443
+  #   binaries:
+  #     - path: /usr/bin/python3*
+  #     - path: /usr/local/bin/python3*
+
+  # --- Redis (MQ) ---
+  # CUSTOMIZE: if Redis is not on localhost, add host + allowed_ips.
+  # localhost is blocked by default SSRF protection, so we need allowed_ips.
+
+  redis:
+    name: redis-mq
+    endpoints:
+      - port: 6379
+        allowed_ips:
+          - "127.0.0.1"
+    binaries:
+      - path: /usr/bin/python3*
+      - path: /usr/local/bin/python3*
+
+  # --- Discord (channel integration) ---
+  # Uncomment if using turnstone-channel with Discord adapter.
+
+  # discord:
+  #   name: discord
+  #   endpoints:
+  #     - host: discord.com
+  #       port: 443
+  #     - host: gateway.discord.gg
+  #       port: 443
+  #     - host: cdn.discordapp.com
+  #       port: 443
+  #   binaries:
+  #     - path: /usr/bin/python3*
+  #     - path: /usr/local/bin/python3*
+
+  # --- web_fetch tool: curated domain allowlist ---
+  #
+  # This is the hard tradeoff. Turnstone's web_fetch tool lets the LLM
+  # fetch arbitrary public URLs. OpenShell cannot allow "all HTTPS" —
+  # every domain must be enumerated.
+  #
+  # Strategy: allowlist the domains your workloads actually need.
+  # The web_fetch tool will return a connection error for unlisted domains,
+  # which the LLM handles gracefully (it tells the user it can't reach
+  # that site).
+  #
+  # CUSTOMIZE: add domains your workstreams need to fetch from.
+
+  web_fetch_common:
+    name: web-fetch-common
+    endpoints:
+      # Documentation sites
+      - host: "**.readthedocs.io"
+        port: 443
+      - host: docs.python.org
+        port: 443
+      - host: "**.github.io"
+        port: 443
+      # Package registries (metadata lookups)
+      - host: pypi.org
+        port: 443
+      - host: www.npmjs.com
+        port: 443
+      # Stack Overflow / reference
+      - host: stackoverflow.com
+        port: 443
+      - host: "**.stackexchange.com"
+        port: 443
+      # Wikipedia
+      - host: "**.wikipedia.org"
+        port: 443
+    binaries:
+      - path: /usr/bin/python3*
+      - path: /usr/local/bin/python3*
+
+  # --- MCP HTTP servers ---
+  # CUSTOMIZE: add endpoints for any MCP servers using streamable-http
+  # transport. stdio-transport MCP servers need no network entry (they
+  # communicate via stdin/stdout pipes within the sandbox).
+
+  # mcp_http_servers:
+  #   name: mcp-http
+  #   endpoints:
+  #     - host: mcp.internal.example.com
+  #       port: 443
+  #   binaries:
+  #     - path: /usr/bin/python3*
+  #     - path: /usr/local/bin/python3*
+
+  # --- Bash tool: curl/wget ---
+  # The bash tool can run curl/wget. These inherit the network namespace
+  # so they can only reach allowed endpoints. But they need binary entries
+  # to pass the proxy's identity check.
+
+  bash_network_tools:
+    name: bash-network-tools
+    endpoints:
+      # Mirrors web_fetch_common — curl/wget should have the same reach.
+      - host: "**.readthedocs.io"
+        port: 443
+      - host: docs.python.org
+        port: 443
+      - host: "**.github.io"
+        port: 443
+      - host: pypi.org
+        port: 443
+      - host: www.npmjs.com
+        port: 443
+      - host: stackoverflow.com
+        port: 443
+      - host: "**.stackexchange.com"
+        port: 443
+      - host: "**.wikipedia.org"
+        port: 443
+    binaries:
+      - path: /usr/bin/curl
+      - path: /usr/bin/wget
+
+  # --- Package installation ---
+  # pip install / uv add from the bash tool.
+
+  package_registries:
+    name: package-install
+    endpoints:
+      - host: pypi.org
+        port: 443
+      - host: files.pythonhosted.org
+        port: 443
+      - host: "**.pypi.org"
+        port: 443
+    binaries:
+      - path: /usr/bin/pip*
+      - path: /usr/local/bin/pip*
+      - path: /usr/bin/uv
+      - path: /usr/local/bin/uv
+      - path: /usr/bin/python3*
+      - path: /usr/local/bin/python3*
+
+  # --- Git operations ---
+  # read-only: clone, fetch, pull. No push (L7 enforcement).
+
+  git_operations:
+    name: git-read-only
+    endpoints:
+      - host: github.com
+        port: 443
+        protocol: rest
+        tls: terminate
+        enforcement: enforce
+        rules:
+          - allow:
+              method: GET
+              path: "/**/info/refs*"
+          - allow:
+              method: POST
+              path: "/**/git-upload-pack"
+      - host: gitlab.com
+        port: 443
+        protocol: rest
+        tls: terminate
+        enforcement: enforce
+        rules:
+          - allow:
+              method: GET
+              path: "/**/info/refs*"
+          - allow:
+              method: POST
+              path: "/**/git-upload-pack"
+    binaries:
+      - path: /usr/bin/git

docs/api-reference.md

@@ -386,10 +386,10 @@ Each item in `items` (shared by `tool_info` and `approve_request`):
 {"type": "tool_output_chunk", "call_id": "call_abc123", "chunk": "Building project...\n"}
 ```
 
-**`tool_result`** -- final output from a completed tool execution. The `call_id` matches the corresponding `tool_info`/`approve_request` item and any preceding `tool_output_chunk` events. For bash tools, this arrives after all streaming chunks and includes both stdout and stderr.
+**`tool_result`** -- final output from a completed tool execution. The `call_id` matches the corresponding `tool_info`/`approve_request` item and any preceding `tool_output_chunk` events. For bash tools, this arrives after all streaming chunks and includes both stdout and stderr. The `is_error` field is `true` when the tool execution failed (e.g. bash exit code >= 2 or signal, file not found, timeout). Exit code 1 is ambiguous (e.g. `grep` no-match) and is not flagged. User denials are tracked separately via a `denied` flag. Clients should use `is_error` instead of text-prefix heuristics.
 
 ```json
-{"type": "tool_result", "call_id": "call_abc123", "name": "bash", "output": "file1.py\nfile2.py\n"}
+{"type": "tool_result", "call_id": "call_abc123", "name": "bash", "output": "file1.py\nfile2.py\n", "is_error": false}
 ```
 
 **`status`** -- token usage statistics, sent after each model turn.
@@ -452,9 +452,12 @@ after `/clear` or `/new` commands).
 {"type": "clear_ui"}
 ```
 
-**`cancelled`** -- the generation was cancelled by the user (via the Stop
-button or `POST /v1/api/cancel`). The client should finalize any in-progress
-assistant message with whatever partial content was streamed.
+**`cancelled`** -- a cancel request was acknowledged (via the Stop button or
+`POST /v1/api/cancel`). This signals that cancellation is in progress, not
+that it is complete. The worker thread may still be finishing — wait for
+`stream_end` before transitioning to a ready state. The client should clear
+any in-progress assistant rendering but not re-enable the send button until
+`stream_end` arrives.
 
 ```json
 {"type": "cancelled"}
@@ -554,7 +557,7 @@ Possible `state` values:
 | `error`     | An error occurred                               |
 
 **Fan-out pattern:** Each connected client receives its own bounded queue
-(`maxsize=500`). A dedicated fan-out thread reads from the shared global queue
+(`maxsize=1000`). A dedicated fan-out thread reads from the shared global queue
 and copies each event to every client queue. If a client queue is full, the
 event is silently dropped for that client.
 
@@ -793,23 +796,35 @@ containing the resumed session's messages.
 
 Cancels the active generation in a workstream. Sets a cooperative cancellation
 flag that is checked at multiple points in the generation loop (per streaming
-chunk, before tool execution, inside bash commands). The session transitions to
-`idle` state and preserves any partial content already streamed.
+chunk, before tool execution, inside bash commands). Also closes the underlying
+HTTP stream to the LLM provider, unblocking any pending read immediately.
+The session transitions to `idle` state and preserves any partial content
+already streamed.
 
 If the workstream is waiting for tool approval or plan review, the pending
 prompt is automatically denied/rejected to unblock the worker thread.
 
 Calling this endpoint when the workstream is already idle is a harmless no-op.
 
+**Force cancel:** When `force` is `true`, the server abandons the stuck worker
+thread immediately and transitions the workstream to `idle`. The abandoned
+thread continues to wind down in the background (killing any running
+subprocesses and exiting at the next cancellation checkpoint). During this
+wind-down it may emit a final `stream_end` event which the server suppresses
+for the orphaned thread. Use force cancel when cooperative cancel has not
+resolved within a few seconds — the web UI offers this as a "Force Stop"
+button automatically.
+
 **Request body:**
 
 ```json
-{"ws_id": "abc123"}
+{"ws_id": "abc123", "force": false}
 ```
 
 | Field  | Type   | Required | Description          |
 |--------|--------|----------|----------------------|
 | `ws_id`| string | yes      | Target workstream ID |
+| `force`| bool   | no       | Abandon stuck worker immediately (default: `false`) |
 
 **Response:**
 

docs/architecture.md

@@ -3,7 +3,7 @@
 Turnstone is an AI orchestration platform with tool use, parallel workstreams, and persistent
 memory. It connects to any OpenAI-compatible API (local vLLM, OpenAI, etc.) or
 Anthropic's native Messages API via pluggable provider adapters, and gives the
-model 17 built-in tools plus external tools via MCP (Model Context Protocol) for
+model 19 built-in tools plus external tools via MCP (Model Context Protocol) for
 reading, writing, searching, planning, and executing code.
 
 The core design principle is a **UI-agnostic engine with pluggable frontends**.
@@ -91,7 +91,7 @@ turnstone/
     _config.py        Base ChannelConfig dataclass
     discord/          Discord adapter (bot, cog, views, streaming, config)
   shared_static/      Shared design system (base.css, auth.js, theme.js, toast.js, utils.js, kb.js)
-    katex-0.16.38/    Vendored KaTeX math rendering library (MIT, woff2 fonts)
+    katex-0.16.44/    Vendored KaTeX math rendering library (MIT, woff2 fonts)
   ui/
     colors.py         ANSI color constants with NO_COLOR support
     markdown.py       Streaming terminal markdown renderer (line-buffered)
@@ -100,7 +100,7 @@ turnstone/
       index.html      Single-page app shell (links to CSS and JS)
       style.css       Page-specific UI styles (dashboard, markdown elements, approval blocks)
       renderer.js     Markdown + LaTeX renderer (tables, nested lists, blockquotes, KaTeX math)
-      app.js          Page-specific client-side JavaScript (SSE, workstreams, tool approval)
+      app.js          Split-pane UI (Pane class, binary layout tree, SSE, tool approval)
   tools/
     *.json            15 tool schemas (OpenAI function-calling format + turnstone metadata)
 ```
@@ -242,7 +242,7 @@ class SessionUI(Protocol):
     def on_content_token(self, text: str) -> None: ...
     def on_stream_end(self) -> None: ...
     def approve_tools(self, items: list[dict]) -> tuple[bool, str | None]: ...
-    def on_tool_result(self, call_id: str, name: str, output: str) -> None: ...
+    def on_tool_result(self, call_id: str, name: str, output: str, *, is_error: bool = False) -> None: ...
     def on_tool_output_chunk(self, call_id: str, chunk: str) -> None: ...
     def on_status(self, usage: dict, context_window: int, effort: str) -> None: ...
     def on_plan_review(self, content: str) -> str: ...
@@ -259,7 +259,7 @@ class SessionUI(Protocol):
 | Class | Module | Notes |
 |-------|--------|-------|
 | `TerminalUI` | `turnstone.cli` | ANSI colors, `MarkdownRenderer`, `Spinner`, readline-based `input()` for approval |
-| `WebUI` | `turnstone.server` | SSE event queue per workstream, `threading.Event` for blocking on approval/plan |
+| `WebUI` | `turnstone.server` | SSE event queue per workstream + global broadcast, `threading.Event` for blocking on approval/plan. `on_state_change` sends to both per-workstream and global SSE (the browser UI uses per-workstream `state_change` events to manage busy/idle transitions; `stream_end` only finalizes markdown rendering). |
 | `NullUI` | `turnstone.eval` | Discards all output; `approve_tools` always returns `(True, None)` |
 
 ### WorkstreamTerminalUI
@@ -353,7 +353,7 @@ remove the tab immediately. Controlled by `--workstream-idle-timeout` (default:
 
 **Workstream eviction at capacity:** When `WorkstreamManager.create()` would
 exceed `max_workstreams` (configurable via `[server].max_workstreams`, default
-10), the oldest IDLE workstream is automatically evicted to make room. The
+50), the oldest IDLE workstream is automatically evicted to make room. The
 `turnstone_workstreams_evicted_total` counter is incremented on each eviction.
 If no IDLE workstream is available the create request fails as before.
 
@@ -375,12 +375,19 @@ non-idle background workstreams above the input prompt.
 ### Web Workstreams
 
 - **Tab bar**: Each workstream renders as a tab with a colored state indicator
-  (CSS `@keyframes pulse` animation per state).
-- **Per-tab SSE**: `connectContentSSE(wsId)` opens
-  `/v1/api/events?ws_id=<id>` for the active tab's event stream.
+  (CSS `@keyframes pulse` animation per state). Clicking a tab switches the
+  focused pane's workstream (or focuses an existing pane showing that ws).
+- **Split panes**: The UI supports tiling multiple workstreams side-by-side or
+  stacked via a binary layout tree. Each `Pane` instance encapsulates its own
+  SSE connection, message area, input, and state (busy, approval, streaming).
+  Split via right-click context menu, pane header buttons, or keyboard
+  (`Ctrl+\`, `Ctrl+Shift+\`). Max 6 panes; no duplicate workstreams across panes.
+  Layout persisted to `localStorage`.
+- **Per-pane SSE**: `Pane.connectSSE(wsId)` opens
+  `/v1/api/events?ws_id=<id>` for each pane's event stream independently.
 - **Global SSE**: `connectGlobalSSE()` opens `/v1/api/events/global` which
   receives `ws_state` broadcasts from all workstreams, used to update tab
-  indicators without switching.
+  indicators and pane headers without switching.
 - **New tab / close**: POST `/v1/api/workstreams/new`, POST `/v1/api/workstreams/close`.
 
 ### Thread Safety
@@ -828,10 +835,16 @@ and are the single source of truth for both backends and Alembic migrations.
 backend = "sqlite"                  # "sqlite" | "postgresql"
 path = ".turnstone.db"              # SQLite file path
 url = ""                            # PostgreSQL connection URL
-pool_size = 5                       # PostgreSQL connection pool size
+pool_size = 2                       # PostgreSQL connection pool size (per process)
 ```
 
-Environment variables: `TURNSTONE_DB_BACKEND`, `TURNSTONE_DB_URL`, `TURNSTONE_DB_PATH`.
+Environment variables: `TURNSTONE_DB_BACKEND`, `TURNSTONE_DB_URL`, `TURNSTONE_DB_PATH`,
+`TURNSTONE_DB_POOL_SIZE`.
+
+The default pool is intentionally small (2 base + 3 overflow = 5 per process)
+because all database operations are short-burst queries that hold connections for
+milliseconds. For clusters with many nodes sharing a PostgreSQL instance, use
+[PgBouncer](pgbouncer.md) in transaction pooling mode.
 
 ### Persistence and Resume
 
@@ -953,6 +966,9 @@ warns if the summary was truncated.
   seeded with `history.replaceState({turnstone: 'dashboard'})` on load. The
   `popstate` listener restores the correct tab or shows the dashboard,
   guarded by `_historyNavigation = true` to prevent re-entrant pushState.
+- **Pane focus**: `mousedown` and `focusin` events on pane containers update
+  `focusedPaneId`. Approval shortcuts (y/n/a) apply to the focused pane.
+  `Ctrl+Alt+Arrow` cycles focus between panes.
 
 ### Eval Resilience
 

docs/console.md

@@ -69,10 +69,11 @@ All reads and writes to the node/workstream map are protected by a single `threa
 
 ### Scale Considerations
 
-- **10,000 workstreams** at ~500 bytes each = ~5 MB in memory
-- **1,000 nodes** polled in parallel with 50 threads at ~100ms each = ~2 second poll cycle
+- **50,000 workstreams** (1,000 nodes × 50 per node) at ~500 bytes each = ~25 MB in memory
+- **1,000 nodes** polled in parallel — fan-out concurrency is configurable via `cluster.node_fan_out_limit` (default 200), yielding 5 batches at ~100ms each = ~0.5 second poll cycle
 - **Filtering and pagination** run in-memory on the full workstream list — sub-millisecond at this scale
-- **SSE fan-out** uses the same per-client queue pattern as the per-node server — backed-up clients get events dropped, not blocking
+- **SSE fan-out** uses per-client queues (2,000 events) — backed-up clients get events dropped, not blocking
+- **Database** — for clusters sharing PostgreSQL, use [PgBouncer](pgbouncer.md) in transaction pooling mode
 
 ---
 
@@ -364,7 +365,7 @@ SSE streams (`/v1/api/events`, `/v1/api/events/global`) are proxied as raw byte
 
 ### Authentication
 
-The proxy forwards the user's JWT to upstream server nodes — it extracts the token from the incoming request's cookie (or `Authorization` header) and adds it as a `Bearer` header on the proxied request.  Since all services share the same `TURNSTONE_JWT_SECRET`, the user's JWT is valid on every node without re-authentication.  The console's own auth middleware also checks proxy routes — `POST` requests to proxy write endpoints (`/v1/api/send`, `/v1/api/approve`, etc.) require `write` scope, preventing read-only tokens from escalating via proxy.  The static `--auth-token` / `proxy_auth_token` is used as a fallback when no user JWT is present.
+The proxy mints a short-lived (5-minute) JWT per request carrying the real user's `user_id`, `scopes`, and `permissions` with `aud: turnstone-server`.  The user's console JWT (`aud: turnstone-console`) cannot be forwarded directly — it would be rejected by the server's audience validation — so the console re-signs a new server-audience JWT from the validated `AuthResult`.  This preserves audit attribution (the upstream server sees the real user, not a service identity) and enforces scope narrowing as defense in depth (a read-only console user's proxied request carries only `read` scope).  The JWT `src` claim is set to `"console-proxy"` for audit traceability.  When no user context is available (auth disabled), the proxy falls back to a `ServiceTokenManager` with service identity `console-proxy`.  The static `--auth-token` / `proxy_auth_token` is used as a final fallback.
 
 ---
 

docs/diagrams/02-package-structure.puml

@@ -97,7 +97,7 @@ package "turnstone/sdk/" <<Rectangle>> {
 
 ' Tool schemas
 package "turnstone/tools/" <<Rectangle>> {
-    component [*.json\n18 tool schemas] as schemas <<artifact>>
+    component [*.json\n19 tool schemas] as schemas <<artifact>>
 }
 
 ' Entry point dependencies

docs/diagrams/03-core-engine-classes.puml

@@ -12,7 +12,7 @@ interface "SessionUI" as SessionUI <<Protocol>> {
     + on_content_token(text: str)
     + on_stream_end()
     + approve_tools(items: list) → (bool, str|None)
-    + on_tool_result(call_id: str, name: str, output: str)
+    + on_tool_result(call_id: str, name: str, output: str, *, is_error: bool = False)
     + on_tool_output_chunk(call_id: str, chunk: str)
     + on_status(usage: dict, ctx_window: int, effort: str)
     + on_plan_review(content: str) → str
@@ -250,13 +250,11 @@ class "MCPClientManager" as MCPMgr {
 
 ' ToolSearchManager
 class "ToolSearchManager" as ToolSearchMgr {
-    - _all_tools: list[dict]
     - _always_on: list[dict]
     - _deferred: list[dict]
     - _expanded: dict[str, None]
     - _index: BM25Index
     --
-    + should_activate() → bool
     + get_visible_tools() → list[dict]
     + get_deferred_tools() → list[dict]
     + get_expanded_names() → list[str]

docs/diagrams/04-conversation-turn.puml

@@ -133,7 +133,8 @@ group loop [while tool_calls present]
         note right of TP
           bash: on_tool_output_chunk(call_id, line)
           called per stdout line,
-          then on_tool_result(call_id, name, output).
+          then on_tool_result(call_id, name, output, is_error).
+          is_error=True when execution failed.
           call_id routes chunks/results to correct
           tool div during parallel execution.
           Other tools: on_tool_result() only.

docs/diagrams/05-tool-pipeline.puml

@@ -24,7 +24,7 @@ partition "Phase 1: Prepare" #E8F5E9 {
         :Dispatch to _prepare_{func_name}();
 
         note right
-            **Dispatch table (17 tools):**
+            **Dispatch table (19 built-in + tool_search):**
             ┌───────────────┬──────────────────┐
             │ Tool          │ Needs Approval?  │
             ├───────────────┼──────────────────┤
@@ -33,16 +33,19 @@ partition "Phase 1: Prepare" #E8F5E9 {
             │ write_file    │ ✓ Yes            │
             │ edit_file     │ ✓ Yes            │
             │ search        │ ✗ Auto-approve   │
+            │ diff_file     │ ✗ Auto-approve   │
             │ math          │ ✗ Auto-approve   │
             │ man           │ ✗ Auto-approve   │
             │ web_fetch     │ ✗ Auto-approve   │
             │ web_search    │ ✗ Auto-approve   │
             │ tool_search   │ ✗ Auto-approve   │
-            │ task          │ ✓ Yes            │
-            │ plan          │ ✓ Yes            │
+            │ task_agent    │ ✓ Yes            │
+            │ plan_agent    │ ✓ Yes            │
             │ memory        │ ✗ Auto-approve   │
             │ recall        │ ✗ Auto-approve   │
             │ notify        │ ✗ Auto-approve   │
+            │ watch         │ ✓ create only    │
+            │ skill         │ ✓ load only      │
             │ read_resource │ ✓ Yes            │
             │ use_prompt    │ ✓ Yes            │
             ├───────────────┼──────────────────┤
@@ -127,7 +130,7 @@ partition "Phase 3: Execute" #E3F2FD {
     :_truncate_output() on each result\n(max context_window × chars_per_token × 0.5 chars\ndefault: ~context_window × 2 chars);
 
     :bash: ui.on_tool_output_chunk(call_id, line) per stdout line;
-    :ui.on_tool_result(call_id, name, output) for each;
+    :ui.on_tool_result(call_id, name, output, is_error) for each;
 
     if (plan tool was executed?) then (yes)
         :ui.on_plan_review(output);

docs/diagrams/06-mq-protocol.puml

@@ -61,6 +61,7 @@ package "Inbound Messages (Client → Bridge)" as InPkg #FFF3E0 {
         + target_node: str = ""
         + initial_message: str = ""
         + skill: str = ""
+        + user_id: str = ""
     }
 
     class CloseWorkstreamMessage {
@@ -146,6 +147,7 @@ package "Outbound Events (Bridge → Client)" as OutPkg #E3F2FD {
             + call_id: str
             + name: str
             + output: str
+            + is_error: bool
         }
         class PlanReviewEvent {
             type = "plan_review"

docs/diagrams/09-workstream-states.puml

@@ -40,12 +40,23 @@ running --> error : Exception during\ntool execution
 
 error --> thinking : New send() call\n_emit_state("thinking")
 
-thinking --> idle : cancel() called\n_emit_state("idle")
+thinking --> idle : cancel() called\nstream aborted\n_emit_state("idle")
 
 running --> idle : cancel() called\n_emit_state("idle")
 
 attention --> idle : cancel() unblocks\napproval/plan wait\n_emit_state("idle")
 
+note left of idle
+  **Cancel escalation:**
+  1. **Cooperative**: cancel() sets event + closes
+     SDK stream → worker exits at next checkpoint
+  2. **Force**: force=true abandons the worker
+     thread, emits stream_end immediately.
+     Orphaned thread still kills subprocesses
+     but skips message mutations (generation
+     counter prevents stale writes).
+end note
+
 note right of thinking
   **Emitted via:**
   session._emit_state(state)

docs/diagrams/11-console-data-flow.puml

@@ -105,7 +105,7 @@ Server -> CC : get_snapshot()
 CC --> Server : ClusterSnapshot\n(full current state)
 
 Server -> CC : register_listener(queue)
-note right : Per-client queue.Queue(maxsize=500)\nSSE via EventSourceResponse + run_in_executor()
+note right : Per-client queue.Queue(maxsize=2000)\nSSE via EventSourceResponse + run_in_executor()
 
 Server -> Browser : data: {"type":"snapshot",...}\n(full state as first SSE event)
 
@@ -154,7 +154,7 @@ activate Server #FFECB3
 Server -> CC : _pick_best_node() or\nget_node_detail(node_id)
 CC --> Server : node validated
 
-Server -> Server : Build CreateWorkstreamMessage\n{target_node:"nodeA", name:"new-task"}
+Server -> Server : Build CreateWorkstreamMessage\n{target_node:"nodeA", name:"new-task",\nuser_id: from auth_result}
 
 Server -> Redis : RPUSH turnstone:inbound:nodeA\n(directed queue)
 Server --> Browser : {status:"ok", correlation_id:"abc",\ntarget_node:"nodeA"}
@@ -163,7 +163,8 @@ deactivate Server
 note right of Redis
   Bridge on Node-A picks up the
   message from its directed queue,
-  POSTs to /v1/api/workstreams/new,
+  POSTs to /v1/api/workstreams/new
+  (forwarding user_id in payload),
   registers ownership, publishes
   ws_created to cluster channel.
 end note

docs/diagrams/12-deployment.puml

@@ -56,6 +56,26 @@ node "Docker Host" as host {
             end note
         }
 
+        node "postgres (profile: production)" <<pgautoupgrade>> as pg_node {
+            component [PostgreSQL\nport 5432] as postgres
+            note bottom of postgres
+                Healthcheck: pg_isready
+                Volume: postgres-data
+                Required for cluster
+                  and production profiles
+            end note
+        }
+
+        node "pgbouncer (optional)" <<bitnami/pgbouncer>> as pgb_node {
+            component [PgBouncer\nport 6432] as pgbouncer
+            note bottom of pgbouncer
+                pool_mode: transaction
+                Recommended for clusters
+                  > 50 nodes
+                See docs/pgbouncer.md
+            end note
+        }
+
         node "sim (profile: sim)" <<turnstone image>> as sim_node {
             component [turnstone-sim] as sim
             note bottom of sim
@@ -92,6 +112,11 @@ console --> server : HTTP polling + proxy\n(GET /v1/api/dashboard,\nproxy /node/
 
 sim --> redis : Redis protocol\n(queues + pubsub + keys)
 
+' Database connections (production/cluster profiles)
+server ..> pgbouncer : PostgreSQL\n(pool_size=2)
+console ..> pgbouncer : PostgreSQL\n(auth/admin)
+pgbouncer --> postgres : transaction\npooling
+
 ' Environment variables
 note right of host
     **Environment Variables:**
@@ -99,13 +124,17 @@ note right of host
     • OPENAI_API_KEY — API key
     • REDIS_PASSWORD — Redis auth
     • TURNSTONE_AUTH_TOKEN — API auth
+    • TURNSTONE_DB_URL — PostgreSQL URL
+    • POSTGRES_PASSWORD — DB password
 end note
 
 ' Volumes
 database "redis-data" as rv
 database "turnstone-data" as tv
+database "postgres-data" as pv
 
 redis_node --> rv
 server_node --> tv
+pg_node --> pv
 
 @enduml

docs/diagrams/14-storage-architecture.puml

@@ -53,10 +53,10 @@ class "SQLiteBackend" as SQLite <<sqlite>> {
 
 class "PostgreSQLBackend" as PG <<postgres>> {
     -_engine: sa.Engine
-    +__init__(url: str, pool_size: int)
+    +__init__(url: str, pool_size: int = 2,\n    max_overflow: int = 3)
     --
     tsvector + ILIKE search
-    Connection pooling
+    Connection pooling (5 max per process)
 }
 
 ' -- Schema --
@@ -151,7 +151,7 @@ note right of Registry
   backend = "sqlite" | "postgresql"
   url = "postgresql+psycopg://..."
   path = ".turnstone.db"
-  pool_size = 5
+  pool_size = 2 (+ 3 overflow)
 end note
 
 note bottom of SQLite
@@ -162,8 +162,9 @@ end note
 
 note bottom of PG
   Production backend.
-  Multi-node / Docker
-  default.
+  Multi-node / Docker default.
+  Use PgBouncer (transaction mode)
+  for clusters > 50 nodes.
 end note
 
 @enduml

docs/diagrams/15-auth-architecture.puml

@@ -176,4 +176,15 @@ note bottom of SH
   Both share JWT signing secret
 end note
 
+note left of JWT
+  **Console Proxy Token Minting**
+  When proxying requests to server nodes:
+  1. Console AuthMiddleware validates user JWT (aud: turnstone-console)
+  2. Proxy mints new JWT (aud: turnstone-server)
+     with real user_id, scopes, permissions
+  3. src: "console-proxy" for audit traceability
+  4. 5-minute expiry (fresh per request)
+  5. Fallback: ServiceTokenManager if no user context
+end note
+
 @enduml

docs/diagrams/26-skills-discovery-architecture.puml

@@ -33,7 +33,7 @@ package "Core Modules" as core #181825 {
 }
 
 package "Session Runtime" as runtime #181825 {
-  rectangle "load_skill tool\nsession.py" as loadtool
+  rectangle "skill tool\nsession.py" as loadtool
   rectangle "set_skill()\nsession.py" as setskill
   rectangle "_load_skills()\nsession.py" as loadskills
 }
@@ -80,6 +80,7 @@ importui --> install : POST (github source)
 ' Annotations
 note right of parser
   YAML frontmatter -> ParsedSkill
+  allowed-tools (standard) -> allowed_tools (internal)
   Anthropic + Hermes tag formats
   Name validation (lowercase+hyphens)
 end note

docs/diagrams/png/02-package-structure.png

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:efcc7cbe8161a54b5ec24bdfd47e8a142f70029e6e66c707e811b99369f85ebf
-size 310079
+oid sha256:2b3ea69f852e93dc1bc7943db0c71d0cdd1afcbcf470a8737189ae56e85b3206
+size 310075

docs/diagrams/png/03-core-engine-classes.png

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0e605963c649574c7bf987b2d338257c78bc1fc68b524ac8276bb25365035e06
-size 594096
+oid sha256:6471e611beebf647f3a191eb16588571a404cc52a43067883a2b6f06dd936376
+size 594676

docs/diagrams/png/04-conversation-turn.png

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:da9d32000e3d92d92ce621661ced60f276f9b5be652f5ed6123b400505415f4a
-size 319702
+oid sha256:3aa8d972bba40d78152f9f0c762b9f5ec616d8052c45fa52b7dd1c679ed81d61
+size 325245

docs/diagrams/png/05-tool-pipeline.png

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:43844b07d36beb04db871f6795a3f3be17852a6a484fdc0ea207403bd7f512a6
-size 274286
+oid sha256:674712a0563f51837383184652efeb28b7bec13378be636e89d2959bfba39d1e
+size 281519

docs/diagrams/png/06-mq-protocol.png

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2636e2d4d2f84f26f93de6e984b780f6ad5bf8d3618a0202ea0a2ee80859c5b5
-size 312409
+oid sha256:d831c5e10266f0232262b6a29b0ea8e45b3cc63df75f716a80f18462b4f85e66
+size 319125

docs/diagrams/png/09-workstream-states.png

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7896c6e041b6dbb89d034468fa980c8fe645df5eb969d45ef966ccc6399edac2
-size 200083
+oid sha256:7bf27afa267d5b8d6da38e83213ed1b8e87639d5105a0a1ccc2e5a4bf4d3b67e
+size 185282

docs/diagrams/png/11-console-data-flow.png

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a74b4b8b5dbfb1a51a01100b731477968942b01218bad9451a3d5a9cb3003294
-size 411665
+oid sha256:e3f1ad0fcd55eaca3b8ad9c5abc07432803641c54ede9fc93c79df144cf77d1c
+size 407761

docs/diagrams/png/12-deployment.png

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:84524f4bc900708ac8adf081591d336f862830188eb8505e71a0f071b339d923
-size 252599
+oid sha256:09065fef028d05e6df425fd8abefaf5a2ca04b66802f2e3975f289fa597f63ed
+size 309656

docs/diagrams/png/14-storage-architecture.png

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c94556889abb382cd5b818639fc0a4706beef3d9c7a0b4cbedc763943d657dd0
-size 244998
+oid sha256:b047cdc318c505f0f0895a65e14c5cc7552716053055cca57fa0a77db150e618
+size 255458

docs/docker.md

@@ -109,9 +109,12 @@ All configuration is via environment variables in `.env` (copy from `.env.exampl
 |----------|---------|-------------|
 | `TURNSTONE_DB_BACKEND` | `sqlite` | Storage backend: `sqlite` or `postgresql` |
 | `TURNSTONE_DB_URL` | — | Database URL (e.g. `postgresql://user:pass@db:5432/turnstone`). For SQLite, defaults to `/data/.turnstone.db` |
+| `TURNSTONE_DB_POOL_SIZE` | `2` | PostgreSQL connection pool size per process (default: 2 base + 3 overflow = 5 max) |
 
 The database stores workstream history, user accounts, and API tokens. When using JWT auth, a database backend is required for user storage.
 
+> **Large clusters:** Each turnstone process maintains a small connection pool (5 max). At hundreds of nodes this adds up — use [PgBouncer](pgbouncer.md) in transaction pooling mode between turnstone and PostgreSQL.
+
 > **First-time setup:** After deploying with auth enabled, create an initial admin user by running `turnstone-admin create-user` inside the container:
 >
 > ```bash
@@ -151,6 +154,8 @@ POSTGRES_PASSWORD=secret docker compose --profile cluster up
 
 The default `server` and `bridge` also run alongside the cluster nodes (11 total). All nodes are accessible via the console dashboard at `:8090`.
 
+For production clusters beyond ~50 nodes, add PgBouncer between turnstone services and PostgreSQL. See [PgBouncer Connection Pooling](pgbouncer.md) for Docker Compose and Helm configuration.
+
 ## Volumes
 
 | Volume | Mount | Purpose |

docs/eval.md

@@ -2,7 +2,8 @@
 
 `turnstone-eval` is the evaluation and prompt optimization system for turnstone. It
 runs test cases against the LLM, scores tool call sequences against expected
-actions, and optionally uses the model to self-optimize the developer prompt.
+actions, and optionally uses a multi-agent pipeline to optimize the developer
+prompt and tool descriptions.
 
 Source: `turnstone/eval.py`
 
@@ -10,15 +11,24 @@ Source: `turnstone/eval.py`
 
 ## Overview
 
-The system works in an iterative loop:
+The system uses UCB tree search to explore prompt variants:
 
-1. Run each test case N times against the current developer prompt.
-2. Score each run by comparing the actual tool call sequence to expected actions.
-3. If not all tests pass, use the model to rewrite the prompt based on failures.
-4. Repeat until all tests pass or max iterations are reached.
+1. Maintain an **evolution tree** of prompt variants, starting from the initial prompt.
+2. Each iteration, **UCB1 selects** the most promising node to evaluate.
+3. Run each test case N times against the selected prompt.
+4. Score each run by comparing the actual tool call sequence to expected actions.
+5. If not all tests pass, run a **three-phase optimization pipeline**:
+   - Phase 1: Analyst diagnoses semantic failure patterns
+   - Phase 2: Tool optimizer adjusts tool descriptions (when `--optimize-tools`)
+   - Phase 3: Prompt optimizer proposes a child variant (when not `--optimize-tools`)
+6. Add the child to the tree and repeat until all tests pass or max iterations reached.
 
-When optimization is disabled (`--no-optimize`), only step 1 and 2 execute
-(a single iteration).
+This approach (inspired by [Learning to Self-Evolve](https://arxiv.org/abs/2603.18620))
+prevents irrecoverable collapse from bad edits — UCB naturally backtracks to
+high-scoring ancestors instead of following a linear chain.
+
+When optimization is disabled (`--no-optimize`), only steps 2-4 execute
+(a single iteration evaluating the root node).
 
 ---
 
@@ -65,6 +75,7 @@ Test suites are JSON files with this structure:
 | `match_mode`       | no       | `"ordered_subset"` | How to match actual vs expected actions (see Scoring). |
 | `max_turns`        | no       | `10`               | Maximum conversation turns before stopping. |
 | `n_runs`           | no       | suite default or 3 | Per-case override for number of runs. |
+| `holdout`          | no       | `false`            | If `true`, this case is evaluated but excluded from optimizer feedback. Used to measure progress without overfitting. |
 
 ### Expected Action Specs
 
@@ -135,6 +146,7 @@ deterministic, non-interactive execution suitable for automated testing.
 | Stdout          | Normal                  | Suppressed during execution |
 | Tool logging    | Display only            | Structured `tool_call_log`  |
 | System prompt   | Built-in developer prompt | Overridable via constructor |
+| Cancellation    | N/A                     | `_cancelled` event for timeout cleanup |
 
 ### NullUI
 
@@ -156,20 +168,34 @@ def send_headless(
 Runs a complete multi-turn conversation:
 
 1. Appends the user message.
-2. Calls the model API (non-streaming).
-3. If tool calls are returned, executes them (with stdout suppressed) and
+2. Checks `_cancelled` event — stops if set (timeout cleanup).
+3. Calls the model API (non-streaming).
+4. If tool calls are returned, executes them (with stdout suppressed) and
    logs each call to `self.tool_call_log`.
-4. Repeats up to `max_turns` or until the model responds without tool calls.
-5. Returns the tool call log: list of dicts with keys `tool`, `args`,
+5. Repeats up to `max_turns` or until the model responds without tool calls.
+6. Returns the tool call log: list of dicts with keys `tool`, `args`,
    `result` (truncated to 500 chars), and `turn`.
 
 Parallel tool calls are capped at 10 per turn to prevent degenerate repetition.
 
+### Timeout and Cancellation
+
+Each test runs in a `ThreadPoolExecutor(max_workers=1)` with a per-test
+timeout (`--test-timeout`). Each attempt gets its own `OpenAI` client with
+a matching httpx read timeout. On timeout, three layers of defense prevent
+zombie connections:
+
+1. **httpx timeout**: Per-request read timeout aborts the HTTP call and
+   releases the server slot.
+2. **`_cancelled` event**: Prevents the orphan thread from starting new turns.
+3. **`run_client.close()`**: Closes the connection pool to abort any
+   in-flight request.
+
 ### Retry Logic
 
 `send_headless()` is called inside `_run_single_test()` with retry logic:
 3 attempts with exponential backoff (sleep `2^attempt` seconds) on any
-exception. This prevents transient API errors from poisoning eval scores.
+exception. `TimeoutError` is re-raised immediately (no retry).
 
 ---
 
@@ -180,68 +206,175 @@ Each test case runs in isolation:
 1. A fresh temp directory is created.
 2. Setup files are written to the temp directory.
 3. The working directory is changed to the temp directory.
-4. A new `HeadlessSession` is created with the current developer prompt.
-5. `send_headless()` runs the user prompt through the conversation loop.
-6. The tool log is scored against expected actions.
-7. The temp directory is cleaned up.
+4. A per-attempt `OpenAI` client is created with httpx timeout matching `--test-timeout`.
+5. A new `HeadlessSession` is created with the current developer prompt.
+6. `send_headless()` runs the user prompt through the conversation loop.
+7. The tool log is scored against expected actions.
+8. The temp directory is cleaned up.
 
 The memory database is also isolated per test (an ephemeral SQLite database
 in the temp directory) so tests do not pollute each other or the user's
 real memory store.
 
+### Parallel Execution
+
+With `--parallel N` (N > 1), tests run in a `ProcessPoolExecutor` with N
+workers. Each subprocess creates its own `OpenAI` client. This is suitable
+for remote API endpoints but will overwhelm local inference servers. The
+default (`--parallel 1`) runs tests serially.
+
 ---
 
-## Optimization Loop
+## Model Roles
 
-`run_optimization()` is the main entry point for iterative prompt optimization.
+The eval pipeline uses up to five separate model roles, each independently
+configurable. All roles inherit from the test model by default, with a
+cascade chain:
+
+```
+test model (--base-url, --model)
+  └─ optimizer (--optimizer-*)
+       ├─ observer (--observer-*)
+       ├─ analyst (--analyst-*)
+       ├─ diversifier (--diversifier-*)
+       └─ tool optimizer (--tool-optimizer-*)
+```
+
+| Role | Purpose | When it runs |
+|------|---------|--------------|
+| **Test** | The model being evaluated | Every iteration |
+| **Analyst** | Diagnoses semantic failure patterns with tool use | When pass rate < 100% |
+| **Optimizer** | Rewrites the developer prompt | Every iteration (unless `--optimize-tools`) |
+| **Tool optimizer** | Rewrites tool descriptions | When `--optimize-tools` is set |
+| **Observer** | Tunes the optimizer's strategy | Every 3 iterations |
+| **Diversifier** | Generates prompt paraphrases | Once before the loop (when `--diversify N`) |
+
+Typical setup: local model for test, Opus for analyst, Sonnet for
+optimizer/observer/diversifier.
+
+---
+
+## Optimization Pipeline
 
 ### Flow
 
 ```
 for iteration in 0..max_iterations:
-    1. Run all test cases n_runs times with current prompt
-    2. Score and aggregate results
-    3. Save intermediate results to JSON
-    4. If all tests pass -> stop
-    5. Every 3 iterations (at iteration 2, 5, 8, ...):
-       -> Observer reviews optimizer strategy
-       -> Reset prompt to best-performing iteration
-    6. Propose new prompt via optimizer model call
-    7. If prompt unchanged -> stop
-    8. Continue with new prompt
+    1. UCB select → pick the most promising tree node
+    2. Run all test cases n_runs times with selected node's prompt
+    3. Update node score (rolling mean) and visit count
+    4. Save intermediate results + tree state to JSON
+    5. If all tests pass → stop
+    6. Phase 1: Analyst diagnoses semantic failure patterns
+    7. Phase 2 (--optimize-tools only): Tool optimizer adjusts descriptions
+    8. Phase 3 (default only): Prompt optimizer proposes new prompt
+    9. Every 3 iterations: Observer tunes the optimizer's strategy
+   10. Add child node to tree (if prompt or tools changed)
 ```
 
-### Prompt Proposal (`_propose_prompt_modification`)
+### Phase 1: Analyst (`_run_analyst`)
 
-Uses the model to rewrite the developer prompt based on test results:
+A multi-turn agent with `math` (Python) and `bash` tools for computing
+statistics. It receives per-case results with failure classifications and
+produces a structured diagnosis:
 
-- **Input**: Current prompt, test case definitions, per-case results with
-  actual vs expected tool sequences, and a history of the last 3 iterations.
-- **Optimizer system prompt** (`OPTIMIZER_SYSTEM`): Instructs the model to
-  act as a text rewriter. Key guidance includes:
-  - Address critical failure modes (text-only responses, write_file vs edit_file,
-    unnecessary search before create, missing plan calls).
-  - Preserve phrasing that drives 100% pass rate on passing tests.
-  - Use direct imperative style with concrete tool call examples.
-  - Stay within 130% of original prompt length.
-- **Output**: The rewritten prompt text (stripped of reasoning tags and code fences).
+- **Failure patterns**: Shared root causes across failing cases
+- **Success/failure contrast**: What distinguishes passing from failing cases
+- **Consistency signals**: Systematic (0%), flaky (1-79%), marginal (80-99%)
+- **Recommended fixes**: Priority-ordered patterns/examples to add or adjust
 
-### Observer System (`_observe_and_update_optimizer`)
+The analyst is instructed to frame fixes as patterns and examples, not
+imperative rules — this feeds cleaner signal to the optimizer.
 
-Every 3 iterations, a meta-level "observer" reviews the optimizer's strategy:
+In `--optimize-tools` mode, the analyst receives the current tool descriptions
+(with any overrides applied) and focuses on tool confusion and description
+issues rather than system prompt patterns.
 
-- Analyzes the iteration history: score trends, regressions, prompt length changes,
+### Phase 2: Tool Optimizer (`_propose_tool_overrides`)
+
+Runs when `--optimize-tools` is set. Receives the current tool descriptions,
+confusion failures (where the model picked the wrong tool), and the analyst's
+diagnosis. Returns a JSON override dict that modifies tool descriptions.
+
+Overrides are validated against known tool names — only `description` and
+`parameters` changes are accepted (no tool renaming at eval time).
+
+After each iteration, changed descriptions are logged as old → new diffs
+for easy visual inspection.
+
+### Phase 3: Prompt Optimizer (`_propose_prompt_modification`)
+
+Skipped in `--optimize-tools` mode. Receives the current prompt, test
+results with per-case pass rates and deltas from the parent node, and the
+analyst's diagnosis. Returns a rewritten prompt.
+
+The optimizer is instructed to prefer patterns over rules — concrete tool
+chain examples teach better than imperative directives like "ALWAYS" or
+"NEVER." If the current prompt contains rule-heavy language, the optimizer
+is guided to replace it with examples.
+
+### Two Optimization Surfaces
+
+The system supports alternating between two optimization surfaces:
+
+1. **System prompt optimization** (default): Freeze tool descriptions,
+   optimize the developer prompt. Run until scores plateau.
+2. **Tool description optimization** (`--optimize-tools`): Freeze the system
+   prompt, optimize tool descriptions only. Run until scores plateau.
+
+Each surface lifts the floor for the other — tool description improvements
+may unlock system prompt gains that weren't reachable before, and vice versa.
+
+### Observer (`_observe_and_update_optimizer`)
+
+Every 3 iterations, a meta-level observer reviews the optimizer's strategy:
+
+- Analyzes iteration history: score trends, regressions, prompt length changes,
   and diffs between iterations.
-- Summarizes the optimizer's behavioral patterns (list style, header usage, length).
-- Uses `OBSERVER_SYSTEM` to rewrite the optimizer's own system prompt.
+- Detects whether the optimizer is producing rule-heavy or pattern-based output.
+- Rewrites the optimizer's own system prompt to correct course.
 - Rejects degenerate outputs (over 200% of input length).
-- After updating the optimizer prompt, resets the developer prompt to the
-  best-performing iteration so far.
 
-This two-level optimization (optimizer + observer) helps the system escape
-local minima and adjust its rewriting strategy.
+### Prompt Diversification
 
-### Result Persistence
+When `--diversify N` is set, the diversifier generates N paraphrased variants
+of each test case's user prompt before the optimization loop. Each run cycles
+through variants (round-robin), testing robustness across phrasings.
+
+Variants can be cached back to the test suite JSON with `--save-variants`,
+and auto-loaded on subsequent runs even without `--diversify`.
+
+---
+
+## Evolution Tree
+
+The optimization maintains a tree of prompt variants (`EvolutionNode`), where
+each node stores its prompt text, tool overrides, aggregated score, and visit
+count. The root node (ID 0) contains the initial prompt.
+
+**UCB1 selection**: Each iteration picks the node with the highest Upper
+Confidence Bound score: `R_bar + C * sqrt(ln(N) / v)`, where `R_bar` is the
+node's mean score, `N` is total visits across all nodes, `v` is the node's
+visit count, and `C` is the exploration constant (`--explore-constant`,
+default sqrt(2)). Unvisited nodes are always selected first.
+
+### Holdout Cases
+
+Test cases with `"holdout": true` are evaluated every iteration but excluded
+from the optimizer's feedback. This prevents the optimizer from overfitting
+to specific test cases. Node scores are computed from holdout cases only
+(when present). If fewer than 2 non-holdout cases remain, holdout is disabled.
+
+### Improvement-Based Feedback
+
+The optimizer sees delta scores (`delta=+20%`) alongside absolute pass rates,
+showing how each case improved relative to the parent node's evaluation. This
+provides a cleaner signal than absolute scores alone — the optimizer can
+distinguish beneficial edits from harmful ones regardless of starting point.
+
+---
+
+## Result Persistence
 
 After each iteration, results are written to the output JSON file. The
 structure is:
@@ -251,9 +384,15 @@ structure is:
   "meta": {
     "model": "model-name",
     "base_url": "http://localhost:8000/v1",
+    "optimizer_model": "claude-opus-4-6",
+    "observer_model": "claude-opus-4-6",
     "started": "2025-01-01T00:00:00",
     "test_suite": "tests.json",
-    "n_runs_default": 3
+    "n_runs_default": 3,
+    "explore_constant": 1.414,
+    "holdout_ids": [],
+    "diversify": 10,
+    "prompt_variants": {"case_id": ["variant1", "variant2"]}
   },
   "iterations": [
     {
@@ -261,7 +400,11 @@ structure is:
       "prompt": "the developer prompt used",
       "prompt_diff": null,
       "optimizer_system": "the optimizer system prompt",
+      "analyst": "analyst diagnosis output",
+      "tool_overrides": {"bash": {"description": "..."}},
       "timestamp": "2025-01-01T00:01:00",
+      "tree_node_id": 0,
+      "tree_child_id": 1,
       "cases": {
         "test_name": {
           "runs": [
@@ -287,9 +430,21 @@ structure is:
         "overall_pass_rate": 0.8,
         "overall_avg_score": 0.87,
         "json_dumps": 0,
-        "per_case_pass_rates": {"test_name": 1.0, ...}
+        "per_case_pass_rates": {"test_name": 1.0}
       }
     }
+  ],
+  "tree": [
+    {
+      "node_id": 0,
+      "parent_id": null,
+      "prompt": "initial prompt",
+      "tool_overrides": {},
+      "score": 0.85,
+      "visit_count": 3,
+      "children": [1, 2],
+      "iteration": 0
+    }
   ]
 }
 ```
@@ -306,26 +461,57 @@ turnstone-eval tests.json                          # evaluate + optimize
 turnstone-eval tests.json --no-optimize            # evaluate only (single iteration)
 turnstone-eval tests.json --n-runs 5 --max-iter 10 # more thorough evaluation
 turnstone-eval tests.json --prompt custom.txt      # start from a custom prompt
+turnstone-eval tests.json --optimize-tools         # optimize tool descriptions only
+turnstone-eval tests.json --diversify 10           # test with prompt variants
 turnstone-eval tests.json -v                       # verbose per-turn logging
 ```
 
+### Multi-model setup (local test model, cloud optimizer)
+
+```
+turnstone-eval tests.json \
+  --base-url http://localhost:8000/v1 \
+  --optimizer-base-url https://api.anthropic.com \
+  --optimizer-model claude-sonnet-4-6 \
+  --analyst-model claude-opus-4-6
+```
+
 ### All Options
 
-| Flag                | Default                       | Description |
-|---------------------|-------------------------------|-------------|
-| `test_file`         | (positional, required)        | Path to test cases JSON file. |
-| `--base-url`        | `http://localhost:8000/v1`    | API base URL. |
-| `--model`           | auto-detect                   | Model name. Auto-detected from the API if not specified. |
-| `--prompt`          | turnstone built-in prompt         | Path to initial prompt text file. |
-| `--n-runs`          | from tests.json or 3          | Number of runs per test case. |
-| `--max-iter`        | 5                             | Maximum optimization iterations. |
-| `--no-optimize`     | false                         | Run evaluation only (sets max-iter to 1). |
-| `--temperature`     | 0.7                           | Sampling temperature. |
-| `--max-tokens`      | 32768                         | Max completion tokens. |
-| `--reasoning-effort` | `medium`                     | Reasoning effort: `low`, `medium`, or `high`. |
-| `--context-window`  | 131072                        | Context window size. |
-| `--output`          | `eval_results.json`           | Output results file path. |
-| `-v`, `--verbose`   | false                         | Show detailed per-turn logging (API calls, tool args, results). |
+| Flag                    | Default                    | Description |
+|-------------------------|----------------------------|-------------|
+| `test_file`             | (positional, required)     | Path to test cases JSON file. |
+| `--base-url`            | `http://localhost:8000/v1` | API base URL for the test model. |
+| `--model`               | auto-detect                | Model name. Auto-detected from the API if not specified. |
+| `--prompt`              | turnstone built-in prompt  | Path to initial prompt text file. |
+| `--n-runs`              | from tests.json or 3       | Number of runs per test case. |
+| `--max-iter`            | 5                          | Maximum optimization iterations. |
+| `--no-optimize`         | false                      | Run evaluation only (sets max-iter to 1). |
+| `--temperature`         | 0.7                        | Sampling temperature. |
+| `--max-tokens`          | 32768                      | Max completion tokens. |
+| `--reasoning-effort`    | `medium`                   | Reasoning effort: `low`, `medium`, or `high`. |
+| `--context-window`      | 131072                     | Context window size. |
+| `--output`              | `eval_results.json`        | Output results file path. |
+| `-v`, `--verbose`       | false                      | Show detailed per-turn logging. |
+| `--explore-constant`    | 1.414 (sqrt(2))            | UCB exploration constant C. |
+| `--test-timeout`        | 300                        | Per-test timeout in seconds. |
+| `--suite-timeout`       | 0 (unlimited)              | Total suite timeout in seconds. |
+| `--no-fast-fail`        | false                      | Disable early termination on all-zero initial runs. |
+| `--parallel`            | 1 (serial)                 | Parallel workers (0=auto, N=use N workers). |
+| `--optimizer-model`     | same as `--model`          | Model for prompt optimization. |
+| `--optimizer-base-url`  | same as `--base-url`       | Base URL for optimizer model. |
+| `--observer-model`      | same as optimizer           | Model for meta-optimization (observer). |
+| `--observer-base-url`   | same as optimizer           | Base URL for observer model. |
+| `--analyst-model`       | same as optimizer           | Model for failure analysis. |
+| `--analyst-base-url`    | same as optimizer           | Base URL for analyst model. |
+| `--diversify`           | 0 (disabled)               | Generate N prompt variants per test case. |
+| `--diversifier-model`   | same as optimizer           | Model for prompt diversification. |
+| `--diversifier-base-url`| same as optimizer           | Base URL for diversifier model. |
+| `--save-variants`       | false                      | Save generated variants back to test suite JSON. |
+| `--optimize-tools`      | false                      | Optimize tool descriptions only (freeze system prompt). |
+| `--tool-optimizer-model` | same as optimizer          | Model for tool description optimization. |
+| `--tool-optimizer-base-url` | same as optimizer      | Base URL for tool optimizer model. |
+| `--save-tools`          | false                      | Write optimized tool descriptions back to `turnstone/tools/*.json`. |
 
 ### Precedence for n_runs
 

docs/governance.md

@@ -70,7 +70,7 @@ etc.) since workstream templates were merged into the skills system in v0.8.0.
   `{{node_id}}` (server node ID). Unrecognized placeholders are kept as-is.
 - **Runtime switching**: `/template <name>` to switch, `/template clear` to revert
   to defaults, `/template` to show current. Persisted across resume.
-- **Model-driven loading**: The `load_skill` built-in tool lets the model
+- **Model-driven loading**: The `skill` built-in tool lets the model
   discover and activate skills mid-conversation. `search` action finds skills
   by query (auto-approved); `load` action activates by name (requires user
   approval since it changes session behavior). Main session only.
@@ -83,11 +83,15 @@ etc.) since workstream templates were merged into the skills system in v0.8.0.
   precedence on name collision. MCP-synced content updates reset `is_default` to
   prevent compromised servers from injecting defaults. Admin UI shows origin badge
   and disables edit/delete for MCP-sourced skills.
+- **Spec fields**: Skills support the full Agent Skills standard frontmatter:
+  `name`, `description`, `license`, `compatibility`, `metadata` (author, version),
+  `allowed-tools`. The `license` and `compatibility` fields are preserved on import
+  and editable in the admin UI. See https://agentskills.io/specification.
 - **Security scanning**: Skills are automatically scanned at creation and update
   time. The scanner evaluates four risk axes: content risk (command execution,
   data exfiltration), supply chain risk (pipe-to-shell, transitive installs),
   vulnerability risk (prompt injection, insecure credentials), and declared
-  capability risk (from `allowed_tools`). Results populate the `scan_status`
+  capability risk (from `allowed-tools` in SKILL.md). Results populate the `scan_status`
   (safe/low/medium/high/critical) and `scan_report` (JSON breakdown) columns.
   These fields are system-managed and cannot be overwritten via the admin API.
 - **Discovery**: External skills can be discovered and installed from registries:
@@ -100,6 +104,20 @@ etc.) since workstream templates were merged into the skills system in v0.8.0.
     Discovery view has search bar, result cards, and "Import from GitHub" modal.
   - SDK: `discover_skills(q)` and `install_skill(source, skill_id=..., url=...)`
     on both Python and TypeScript console clients.
+- **Runtime config on installed skills**: Installed (readonly) skills can have
+  their runtime configuration edited — model, temperature, reasoning effort,
+  token budget, max tokens, agent max turns, auto-approve, allowed tools,
+  and enabled flag. The server restricts updates to these fields only via
+  `_SKILL_RUNTIME_CONFIG_FIELDS` filtering; spec/content fields (name,
+  description, tags, license, compatibility, content, activation) remain
+  immutable. The admin UI shows "Save Config" instead of "Save" for these
+  skills. Audit action: `skill.update.config`.
+- **Admin UI**: Create/Edit skill modals use a two-column spec manifest layout
+  (left: Identity / Manifest / Deployment; right: Skill Content editor with
+  monospace font). Runtime Config is a collapsible 3-column grid below.
+  License uses an SPDX identifier dropdown (MIT, Apache-2.0, GPL-3.0, etc.).
+  Installed skills show a cyan origin badge with source URL, spec fields are
+  disabled, and all collapsible sections auto-expand in view mode.
 
 ### Usage Tracking
 

docs/judge.md

@@ -299,7 +299,7 @@ four independent risk axes:
    obfuscation, download-execute chains, executable URLs from untrusted domains
 3. **Vulnerability risk** — prompt injection patterns, insecure credential
    handling, third-party content exposure (indirect prompt injection surface)
-4. **Declared capability risk** — parsed from the skill's `allowed_tools` field.
+4. **Declared capability risk** — parsed from `allowed-tools` in the skill's SKILL.md.
    `Bash(*)` (unrestricted shell) is high risk. `Bash(git:*)` is low.
    Read-only tools are safe.
 
@@ -338,7 +338,7 @@ from the output before it enters the conversation.
 | Priority | Category | Risk | Examples |
 |----------|----------|------|----------|
 | 1 | Prompt injection | high | Override phrases, role injection (`{"role":"system"}`), instruction override markers |
-| 2 | Credential leakage | high | API keys, private key blocks, connection strings, `.env` format secrets |
+| 2 | Credential leakage | high | API keys, private key blocks, connection strings, `.env` format secrets, JSON secrets (`"api_key": "..."`, `"password": "..."`, etc.) |
 | 3 | Encoded payloads | medium | Script data URIs, hex shellcode sequences |
 | 4 | Adversarial URLs | medium | Cloud metadata endpoints, credential-bearing query parameters |
 | 5 | System info disclosure | low | Private IP addresses, sensitive file paths |

docs/openshell.md

@@ -0,0 +1,288 @@
+# OpenShell Sandbox Integration
+
+Turnstone can run inside an [OpenShell](https://github.com/NVIDIA/OpenShell)
+sandbox for kernel-enforced security boundaries around tool execution. OpenShell
+provides four layers of defense that Turnstone's application-level safety model
+does not cover:
+
+| Layer | Mechanism | What it prevents |
+|-------|-----------|------------------|
+| Filesystem | Landlock | Writes to `/etc`, `~/.ssh`, system paths |
+| Network | Network namespace + seccomp + HTTP CONNECT proxy | Connections to unlisted hosts |
+| Process | `setuid` drop + verification | Privilege escalation to root |
+| Credentials | Proxy-level secret resolution | API keys in sandbox memory |
+
+Turnstone's own safety layers (human approval, intent judge, tool policies,
+output guard) remain active inside the sandbox and handle threats at the semantic
+level -- what the LLM *means* to do with its legitimate access.
+
+> See also: [Security and Authentication](security.md),
+> [Intent Validation](judge.md), [Governance](governance.md)
+
+---
+
+## Quick Start
+
+```bash
+# Run turnstone-server in an OpenShell sandbox
+openshell sandbox run \
+  --policy deploy/openshell/turnstone-policy.yaml \
+  --workdir /path/to/project \
+  -- python3 -m turnstone.server --host 0.0.0.0 --port 8080
+```
+
+With inference routing (API keys never enter the sandbox):
+
+```bash
+openshell sandbox run \
+  --policy deploy/openshell/turnstone-policy.yaml \
+  --inference-routes deploy/openshell/routes.yaml \
+  --workdir /path/to/project \
+  -- python3 -m turnstone.server --host 0.0.0.0 --port 8080 \
+      --base-url https://inference.local
+```
+
+The `inference.local` hostname is intercepted by the OpenShell proxy before
+network policy evaluation -- no network policy entry is needed for it.
+
+---
+
+## Policy Files
+
+### `deploy/openshell/turnstone-policy.yaml`
+
+The main sandbox policy. Covers filesystem, process, and network rules.
+
+### `deploy/openshell/routes.yaml`
+
+Inference routing configuration. Maps `inference.local` to real LLM API
+backends. Uncomment and configure the provider(s) you use.
+
+---
+
+## Filesystem Policy
+
+The policy uses Landlock (Linux 5.13+) for kernel-enforced filesystem access
+control. Paths are locked at sandbox creation and cannot be changed at runtime.
+
+| Path | Access | Purpose |
+|------|--------|---------|
+| `--workdir` | read-write | Project files (auto-added via `include_workdir`) |
+| `/tmp` | read-write | Bash tool temp scripts, eval workdirs |
+| `/dev/null` | read-write | Shell redirections (`2>/dev/null`) |
+| `/var/log` | read-write | Log files |
+| `/usr`, `/lib`, `/lib64` | read-only | Python runtime, installed packages |
+| `/etc` | read-only | System config, SSL certificates |
+| `/proc`, `/dev/urandom` | read-only | Process info, entropy |
+| `~/.config/turnstone` | read-only | Config file (writes go to database) |
+
+Landlock runs in `best_effort` mode by default -- degrades gracefully on kernels
+without Landlock support. Set `compatibility: hard_requirement` for production
+hardened deployments.
+
+---
+
+## Network Policy
+
+Default-deny. Only explicitly listed host:port pairs are reachable. All child
+processes (MCP servers, bash commands, grep) inherit the network namespace and
+cannot bypass the proxy.
+
+### Included endpoints
+
+| Policy | Hosts | Purpose |
+|--------|-------|---------|
+| `openai_api` | `api.openai.com` | OpenAI LLM API |
+| `anthropic_api` | `api.anthropic.com` | Anthropic LLM API |
+| `tavily_api` | `api.tavily.com` | Web search fallback |
+| `skills_registry` | `skills.sh` | Skill discovery |
+| `github_api` | `api.github.com` (read-only L7), `raw.githubusercontent.com` | Skill fetch, GitHub API |
+| `mcp_registry` | `registry.modelcontextprotocol.io` (read-only L7) | MCP server discovery |
+| `redis` | `127.0.0.1:6379` | Message queue |
+| `web_fetch_common` | readthedocs, python docs, GitHub Pages, PyPI, npm, Stack Overflow, Wikipedia | Curated web_fetch domains |
+| `bash_network_tools` | Same as `web_fetch_common` | curl/wget from bash tool |
+| `package_registries` | `pypi.org`, `files.pythonhosted.org` | pip/uv package installs |
+| `git_operations` | `github.com`, `gitlab.com` (L7: clone/fetch only, no push) | Git read-only operations |
+
+### L7 enforcement
+
+Endpoints marked with `protocol: rest` and `tls: terminate` get HTTP-level
+inspection. The proxy TLS-terminates using an ephemeral per-sandbox CA, parses
+each request, and evaluates method + path against the rules.
+
+The `github_api`, `mcp_registry`, and `git_operations` policies use L7
+enforcement:
+
+- **GitHub API / MCP Registry**: `access: read-only` -- only GET, HEAD, OPTIONS
+  allowed
+- **Git operations**: explicit rules allowing only `info/refs` (GET) and
+  `git-upload-pack` (POST) -- clone and fetch work, push is blocked
+
+### Commented-out sections
+
+The policy includes commented blocks for optional integrations. Uncomment and
+configure as needed:
+
+- **OIDC** -- add your identity provider's hostname
+- **Discord** -- `discord.com`, `gateway.discord.gg`, `cdn.discordapp.com`
+- **MCP HTTP servers** -- any MCP servers using streamable-http transport
+
+---
+
+## Customizing the Domain Allowlist
+
+The `web_fetch` tool lets the LLM fetch arbitrary public URLs, but OpenShell
+cannot allow "all HTTPS" -- bare wildcard hosts are rejected by policy
+validation. Instead, the policy ships with a curated set of common reference
+domains.
+
+To add domains your workloads need:
+
+```yaml
+# In turnstone-policy.yaml, under web_fetch_common.endpoints:
+      - host: docs.example.com
+        port: 443
+
+# Also add to bash_network_tools.endpoints if curl/wget should reach it:
+      - host: docs.example.com
+        port: 443
+```
+
+Wildcard patterns are supported:
+
+- `*.example.com` -- matches one subdomain level (e.g. `api.example.com`)
+- `**.example.com` -- matches any depth (e.g. `deep.sub.example.com`)
+
+Unlisted domains return connection errors, which the LLM handles gracefully by
+telling the user it cannot reach that site.
+
+---
+
+## Inference Routing
+
+Inference routing keeps real API keys completely outside the sandbox. The
+sandbox process only sees opaque placeholder tokens in its environment
+(`openshell:resolve:env:ANTHROPIC_API_KEY`). The proxy rewrites these to real
+credentials on the wire before forwarding to the upstream API.
+
+### Setup
+
+1. Edit `deploy/openshell/routes.yaml` -- uncomment your provider:
+
+```yaml
+routes:
+  # OpenAI
+  - name: inference.local
+    endpoint: https://api.openai.com/v1
+    model: gpt-5
+    provider_type: openai
+    protocols:
+      - openai_chat_completions
+      - model_discovery
+    api_key_env: OPENAI_API_KEY
+
+  # Or Anthropic
+  - name: inference.local
+    endpoint: https://api.anthropic.com
+    model: claude-sonnet-4-6
+    provider_type: anthropic
+    protocols:
+      - anthropic_messages
+    api_key_env: ANTHROPIC_API_KEY
+```
+
+2. Start with `--inference-routes` and point turnstone at `inference.local`:
+
+```bash
+openshell sandbox run \
+  --inference-routes deploy/openshell/routes.yaml \
+  --base-url https://inference.local \
+  ...
+```
+
+3. When inference routing is active, the `openai_api` and `anthropic_api`
+   network policies can be removed from the sandbox policy -- the proxy handles
+   LLM traffic on a separate code path that bypasses OPA entirely.
+
+### Local model servers
+
+For local servers (vLLM, llama.cpp) with no authentication, omit both
+`api_key` and `api_key_env` from the route config. No credential resolution
+is needed.
+
+---
+
+## MCP Server Subprocesses
+
+MCP servers using stdio transport are spawned as child processes of turnstone.
+They automatically inherit all sandbox constraints:
+
+- **Network namespace** -- kernel-level, cannot be bypassed
+- **Landlock filesystem** -- kernel-level, cannot be relaxed
+- **Seccomp socket filter** -- kernel-level, inherited on fork
+
+No per-subprocess policy entries are needed for these constraints. However, if
+an MCP server makes outbound network requests (through the proxy), its binary
+must appear in a `binaries[]` entry for the relevant network policy. The proxy
+identifies the requesting process via `/proc/<pid>/exe` (not `argv[0]`, which
+is spoofable).
+
+Example for a Python-based MCP server that calls an external API:
+
+```yaml
+  mcp_external_api:
+    name: mcp-external
+    endpoints:
+      - host: api.example.com
+        port: 443
+    binaries:
+      - path: /usr/bin/python3*
+      - path: /usr/local/bin/python3*
+```
+
+MCP servers using streamable-http transport are remote -- they need a network
+policy entry for their host:port but no binary entry (the Python process making
+the HTTP call is already covered by the standard `python3*` binary entries).
+
+---
+
+## Security Model: Which Layer Enforces What
+
+```
+                      OpenShell (infrastructure)          Turnstone (application)
+                     ─────────────────────────────     ──────────────────────────────
+Filesystem access     Landlock kernel enforcement       (no enforcement)
+Network egress        Netns + seccomp + proxy + OPA     SSRF check on web_fetch
+Credentials           Placeholder injection + proxy     Output guard redaction
+Privilege level       setuid drop + verification        (no enforcement)
+Tool semantics        (no visibility)                   Heuristic + LLM judge
+Tool policies         (no visibility)                   fnmatch admin policies
+Prompt injection      (no visibility)                   Output guard detection
+Human approval        (no visibility)                   Approval gate + "always"
+```
+
+OpenShell constrains what the process can physically reach. Turnstone constrains
+what the LLM does with its legitimate access. Neither layer is sufficient alone:
+
+- Without OpenShell: a bash command can `curl` secrets to any endpoint, write to
+  `/etc/crontab`, or read `~/.ssh/id_rsa` -- all gated only by human approval
+- Without Turnstone: the LLM can `rm -rf` the entire workdir, run destructive
+  commands, or consume prompt injection payloads -- all within the sandbox's
+  allowed scope
+
+---
+
+## Hardening Checklist
+
+For production deployments:
+
+- [ ] Set `landlock.compatibility: hard_requirement`
+- [ ] Enable inference routing (removes API keys from sandbox)
+- [ ] Remove `openai_api`/`anthropic_api` network policies when using inference
+      routing (traffic goes through the router, not direct)
+- [ ] Review and trim `web_fetch_common` domains to your actual needs
+- [ ] Remove `package_registries` policy if pip/uv installs are not needed
+- [ ] Add your OIDC provider endpoint if using SSO
+- [ ] Set Redis `allowed_ips` to your actual Redis host if not localhost
+- [ ] Consider removing `bash_network_tools` entirely if bash should not have
+      network access

docs/pgbouncer.md

@@ -0,0 +1,200 @@
+# PgBouncer Connection Pooling
+
+Turnstone cluster deployments share a single PostgreSQL instance across
+all server nodes, bridge processes, and the console. Each process
+maintains a small connection pool (2 base + 3 overflow = 5 max). At
+scale this adds up — a 100-node cluster opens up to 500 connections,
+and a 1000-node cluster up to 5,000.
+
+PostgreSQL's default `max_connections` is 100, and each real connection
+allocates ~5–10 MB of backend memory. PgBouncer sits between turnstone
+and PostgreSQL, multiplexing thousands of lightweight client connections
+down to a small number of real database connections.
+
+---
+
+## Why PgBouncer works well with turnstone
+
+All turnstone database operations are short-burst queries: acquire a
+connection, execute 1–3 statements, commit, release. No operation holds
+a connection for more than a few milliseconds. This makes **transaction
+pooling mode** ideal — PgBouncer assigns a real connection only for the
+duration of each transaction, then returns it to the pool.
+
+| Cluster size | Client connections (max) | PgBouncer server connections needed |
+|--------------|------------------------|-------------------------------------|
+| 10 nodes     | 50                     | 10–20                               |
+| 100 nodes    | 500                    | 20–40                               |
+| 500 nodes    | 2,500                  | 30–60                               |
+| 1,000 nodes  | 5,000                  | 40–80                               |
+
+The server connection count stays low because most client connections
+are idle at any given moment.
+
+---
+
+## Docker Compose
+
+Add PgBouncer between turnstone services and PostgreSQL:
+
+```yaml
+services:
+  pgbouncer:
+    image: bitnami/pgbouncer:latest
+    environment:
+      POSTGRESQL_HOST: postgres
+      POSTGRESQL_PORT: "5432"
+      POSTGRESQL_DATABASE: turnstone
+      POSTGRESQL_USERNAME: ${POSTGRES_USER:-turnstone}
+      POSTGRESQL_PASSWORD: ${POSTGRES_PASSWORD:?}
+      PGBOUNCER_POOL_MODE: transaction
+      PGBOUNCER_DEFAULT_POOL_SIZE: "40"
+      PGBOUNCER_MAX_CLIENT_CONN: "5000"
+      PGBOUNCER_MAX_DB_CONNECTIONS: "80"
+      PGBOUNCER_SERVER_IDLE_TIMEOUT: "300"
+    ports:
+      - "6432:6432"
+    networks:
+      - turnstone-net
+    depends_on:
+      postgres:
+        condition: service_healthy
+    healthcheck:
+      test: ["CMD", "pg_isready", "-h", "127.0.0.1", "-p", "6432"]
+      interval: 5s
+      timeout: 3s
+      retries: 5
+```
+
+Then point turnstone services at PgBouncer instead of PostgreSQL
+directly by changing the `DATABASE_URL` (or `TURNSTONE_DB_URL`):
+
+```bash
+# Before (direct)
+TURNSTONE_DB_URL=postgresql://turnstone:secret@postgres:5432/turnstone
+
+# After (via PgBouncer)
+TURNSTONE_DB_URL=postgresql://turnstone:secret@pgbouncer:6432/turnstone
+```
+
+---
+
+## Helm / Kubernetes
+
+Add a PgBouncer deployment or use a Helm chart like
+[bitnami/pgbouncer](https://github.com/bitnami/charts/tree/main/bitnami/pgbouncer).
+
+In `values.yaml`, point the database at PgBouncer:
+
+```yaml
+database:
+  backend: postgresql
+  external:
+    host: pgbouncer
+    port: 6432
+    database: turnstone
+    username: turnstone
+    existingSecret: turnstone-db-secret
+```
+
+PgBouncer configuration:
+
+```yaml
+pgbouncer:
+  poolMode: transaction
+  defaultPoolSize: 40
+  maxClientConn: 5000
+  maxDbConnections: 80
+```
+
+---
+
+## Configuration reference
+
+| PgBouncer setting | Recommended | Notes |
+|-------------------|-------------|-------|
+| `pool_mode` | `transaction` | Required — turnstone uses short-burst queries with no session state |
+| `default_pool_size` | 40 | Real PostgreSQL connections per database. Start here, increase if you see `no more connections allowed` |
+| `max_client_conn` | 5000 | Upper bound on client connections. Set to `cluster_nodes × 5` |
+| `max_db_connections` | 80 | Hard cap on real connections to PostgreSQL. Keep below PG `max_connections` minus headroom for admin/monitoring |
+| `server_idle_timeout` | 300 | Close idle server connections after 5 minutes |
+| `server_lifetime` | 3600 | Recycle server connections after 1 hour |
+
+On the PostgreSQL side:
+
+| PostgreSQL setting | Recommended | Notes |
+|--------------------|-------------|-------|
+| `max_connections` | 100 | Default is fine — PgBouncer is the only client. Set higher than `max_db_connections` to leave room for admin connections |
+| `shared_buffers` | 25% of RAM | Standard PostgreSQL tuning |
+
+---
+
+## Turnstone pool settings
+
+Each turnstone process maintains its own SQLAlchemy connection pool to
+PgBouncer (which then multiplexes to PostgreSQL):
+
+| Environment variable | Default | Description |
+|---------------------|---------|-------------|
+| `TURNSTONE_DB_POOL_SIZE` | 2 | Base pool size per process |
+| `TURNSTONE_DB_BACKEND` | sqlite | Set to `postgresql` for cluster deployments |
+| `TURNSTONE_DB_URL` | — | Connection URL (point at PgBouncer, not PostgreSQL directly) |
+
+The default pool of 2 + 3 overflow = 5 connections per process is
+intentionally small to support large clusters. You should not need to
+increase this — turnstone's database operations are all short-burst
+context-managed queries that hold connections for milliseconds.
+
+SQLAlchemy `pool_pre_ping` is enabled, so stale connections (e.g. after
+PgBouncer restarts) are automatically detected and replaced.
+
+---
+
+## Monitoring
+
+PgBouncer exposes stats via its admin console (connect to
+PgBouncer port with user `pgbouncer`):
+
+```sql
+-- Active and waiting clients
+SHOW POOLS;
+
+-- Per-database stats
+SHOW STATS;
+
+-- Current client connections
+SHOW CLIENTS;
+```
+
+Key metrics to watch:
+
+- **`cl_active`** — clients with a server connection assigned. Should be
+  well below `max_db_connections`.
+- **`cl_waiting`** — clients waiting for a server connection. Sustained
+  non-zero values mean you need more `default_pool_size`.
+- **`sv_active`** — active server (PostgreSQL) connections. Should stay
+  below PostgreSQL `max_connections`.
+
+---
+
+## Troubleshooting
+
+**"no more connections allowed (max_client_conn)"** — PgBouncer is
+rejecting new client connections. Increase `max_client_conn` to match
+your cluster size × 5.
+
+**"no more connections allowed (max_db_connections)"** — PgBouncer
+cannot open more connections to PostgreSQL. Increase
+`max_db_connections` and ensure PostgreSQL `max_connections` is higher.
+
+**Connections timing out on startup** — If all nodes start
+simultaneously, the burst of initial connections (migrations, health
+checks) can temporarily exceed the pool. PgBouncer queues excess
+clients by default — this resolves itself within seconds.
+
+**Prepared statements not supported** — PgBouncer in `transaction` mode
+does not support prepared statements. Turnstone's SQLAlchemy layer does
+not use server-side prepared statements by default, so this is not an
+issue.
+
+See also: [Docker deployment](docker.md) · [Security](security.md)

docs/sdk.md

@@ -75,7 +75,7 @@ Both `TurnstoneServer` (sync) and `AsyncTurnstoneServer` (async) expose:
 | | `approve(*, ws_id, approved, feedback, always)` | `StatusResponse` |
 | | `plan_feedback(*, ws_id, feedback)` | `StatusResponse` |
 | | `command(*, ws_id, command)` | `StatusResponse` |
-| | `cancel(ws_id)` | `StatusResponse` |
+| | `cancel(ws_id, *, force=False)` | `StatusResponse` |
 | **Streaming** | `stream_events(ws_id)` | `Iterator[ServerEvent]` |
 | | `stream_global_events()` | `Iterator[ServerEvent]` |
 | **High-level** | `send_and_wait(message, ws_id, *, timeout, on_event)` | `TurnResult` |
@@ -127,7 +127,7 @@ SSE events are deserialized into typed dataclasses. Use `event.type` to discrimi
 | `reasoning` | `ReasoningEvent` | `text` |
 | `tool_info` | `ToolInfoEvent` | `items` |
 | `approve_request` | `ApproveRequestEvent` | `items` |
-| `tool_result` | `ToolResultEvent` | `call_id`, `name`, `output` |
+| `tool_result` | `ToolResultEvent` | `call_id`, `name`, `output`, `is_error` |
 | `tool_output_chunk` | `ToolOutputChunkEvent` | `call_id`, `chunk` |
 | `status` | `StatusEvent` | `prompt_tokens`, `total_tokens`, `pct`, `effort`, `cache_creation_tokens`, `cache_read_tokens` |
 | `plan_review` | `PlanReviewEvent` | `content` |

docs/security.md

@@ -457,14 +457,30 @@ without any database.
 ### Proxy auth forwarding
 
 When the console proxies requests to server nodes (via `/node/{id}/...`
-routes), it uses a dedicated **service proxy token** with
-`aud: turnstone-server` and `write` scope.  The user's console JWT
-(which has `aud: turnstone-console`) is **not** forwarded — it would be
-rejected by the server's audience validation.
+routes), it mints a **short-lived user-scoped JWT** with
+`aud: turnstone-server` carrying the real user's `user_id`, `scopes`,
+and `permissions`.  The user's console JWT (which has
+`aud: turnstone-console`) is **not** forwarded directly — it would be
+rejected by the server's audience validation.  Instead, the console
+re-signs a new JWT targeted at the server audience.
 
-The proxy token is managed by a `ServiceTokenManager` that auto-rotates
-1-hour JWTs, refreshing at 80% of lifetime.  If `--auth-token` is
-provided, that static token is used instead.
+Each proxied request gets a fresh JWT (5-minute expiry).  This ensures:
+
+- **Audit attribution** — the upstream server records the real user in
+  `ctx_user_id` and audit events, not a generic service identity.
+- **Scope narrowing** — a read-only console user's proxied request
+  carries only `read` scope, not the full `{read, write, approve}` set.
+  The server enforces this as defense in depth.
+- **Permission forwarding** — granular RBAC permissions from the
+  console JWT are carried through to the server.
+
+The JWT `src` claim is set to `"console-proxy"`, allowing servers to
+distinguish proxied requests from direct logins in audit logs.
+
+When no user context is available (auth disabled, or internal requests),
+the proxy falls back to a `ServiceTokenManager` with service identity
+`console-proxy` and full scopes.  If `--auth-token` is provided, that
+static token is used as a final fallback.
 
 ### Service-to-service authentication
 
@@ -475,7 +491,7 @@ auto-rotating JWTs when communicating with server nodes:
 |---------|----------|-------|----------|---------|
 | Bridge | `bridge` | `approve` | `turnstone-server` | Tool approval proxy, message relay |
 | Console collector | `console-collector` | `read` | `turnstone-server` | Node health polling |
-| Console proxy | `console-proxy` | `write` | `turnstone-server` | Proxied API calls |
+| Console proxy (fallback) | `console-proxy` | `approve` | `turnstone-server` | Proxied API calls when no user context |
 | Channel notify | `system` | `write` | `turnstone-channel` | Notification delivery to channel gateway |
 
 Service tokens use 1-hour expiry with automatic refresh via
@@ -483,6 +499,17 @@ Service tokens use 1-hour expiry with automatic refresh via
 httpx event hooks to ensure rotated tokens are picked up on SSE
 reconnects.
 
+### User identity in MQ-dispatched workstreams
+
+When the console creates a workstream via MQ (the normal path), the
+authenticated user's `user_id` is embedded in the
+`CreateWorkstreamMessage`.  The bridge forwards this `user_id` in the
+HTTP payload when calling the server's `POST /v1/api/workstreams/new`.
+The server accepts a `user_id` from the request body **only when the
+caller is a trusted service** — identified by `token_source` matching
+`bridge`, `console-proxy`, or `console`.  Regular API callers cannot
+override `user_id`; the server always uses their JWT identity.
+
 Note that the channel gateway uses a distinct JWT audience
 (`turnstone-channel`) from the server (`turnstone-server`) and console
 (`turnstone-console`).  A server-scoped JWT cannot authenticate to the

docs/settings.md

@@ -51,7 +51,7 @@ connection, Redis, auth secrets, server bind address). These stay in
 | Bridge identity | `[bridge]` | config.toml / env |
 | Console bind | `[console]` | config.toml / env |
 
-**ConfigStore settings** (~40 settings) are loaded from the database after
+**ConfigStore settings** (48 settings) are loaded from the database after
 storage initialization:
 
 | Section | Settings |
@@ -60,10 +60,12 @@ storage initialization:
 | `session` | instructions, retention_days, compact_max_tokens, auto_compact_pct |
 | `tools` | timeout, truncation, agent_max_turns, skip_permissions, search, search_threshold, search_max_results |
 | `server` | workstream_idle_timeout, max_workstreams |
+| `cluster` | node_fan_out_limit, mcp_max_servers |
 | `mcp` | config_path, refresh_interval, registry_url |
-| `ratelimit` | enabled, requests_per_second, burst |
+| `ratelimit` | enabled, requests_per_second, burst, trusted_proxies |
 | `health` | backend_probe_interval, backend_probe_timeout, circuit_breaker_threshold, circuit_breaker_cooldown |
-| `judge` | enabled, model, provider, base_url, api_key, confidence_threshold, max_context_ratio, timeout, read_only_tools |
+| `judge` | enabled, model, provider, base_url, api_key, confidence_threshold, max_context_ratio, timeout, read_only_tools, output_guard, redact_secrets |
+| `skills` | discovery_url |
 | `memory` | relevance_k, fetch_limit, max_content, nudge_cooldown, nudges |
 
 Settings are addressed by dotted key (e.g. `memory.relevance_k`). Each has a

docs/tls.md

@@ -0,0 +1,221 @@
+# TLS / mTLS
+
+Turnstone supports end-to-end transport encryption with mutual TLS (mTLS) for
+inter-service communication, powered by [lacme](https://pypi.org/project/lacme/).
+
+---
+
+## Quick Start (Docker Compose)
+
+```bash
+docker compose -f compose.yaml -f deploy/docker-compose.tls.yml up
+```
+
+This:
+1. Bootstraps an internal CA and issues certs for Redis/PostgreSQL
+2. Starts the console with TLS enabled (internal CA + ACME server)
+3. Server nodes auto-provision certs via the console's ACME endpoint
+4. All inter-service communication uses mTLS
+
+---
+
+## Architecture
+
+```
+Console (CA + ACME Server)
+  +-- CertificateAuthority (owns root key, signs certs)
+  +-- ACMEResponder (mounted at /acme, RFC 8555)
+  +-- GET /acme/ca.pem (root cert for node bootstrapping)
+                  |
+                  | ACME protocol (auto-approve, no challenge validation)
+      +-----------+-----------+
+      |           |           |
+  Server(s)    Bridge    Channel GW
+  (auto-cert    (mTLS     (mTLS
+   + renewal)   client)    client)
+```
+
+**Two cert paths on the console:**
+- **Internal cert** (mTLS): Always from the internal CA. Used for cluster
+  service mesh communication.
+- **Frontend cert** (HTTPS): From an external ACME CA (e.g. Let's Encrypt)
+  if `tls.acme_directory` is set, otherwise self-issued from the internal CA.
+
+---
+
+## Configuration
+
+### Settings (ConfigStore / Admin Settings tab)
+
+| Setting | Default | Description |
+|---------|---------|-------------|
+| `tls.enabled` | `false` | Master switch for internal mTLS |
+| `tls.acme_directory` | `""` | External ACME CA URL for console frontend cert |
+
+### Bootstrap Config (config.toml)
+
+These are needed before storage is available:
+
+```toml
+[redis]
+tls = false
+tls_ca = ""      # path to CA cert
+tls_cert = ""    # path to client cert
+tls_key = ""     # path to client key
+
+[database]
+sslmode = "prefer"   # disable, allow, prefer, require, verify-full
+sslrootcert = ""     # path to CA cert
+sslcert = ""         # path to client cert
+sslkey = ""          # path to client key
+```
+
+### Hardcoded Defaults
+
+| Parameter | Value | Notes |
+|-----------|-------|-------|
+| CA common name | "Turnstone CA" | |
+| CA validity | 10 years | |
+| Cert validity | 48 hours | Short-lived, auto-renewed |
+| Renewal interval | 24 hours | Half of validity |
+| ACME auto-approve | true | Internal network, no challenge validation |
+
+---
+
+## CLI
+
+### Offline Bootstrap
+
+Create a CA and infrastructure certs without a running console:
+
+```bash
+# Bootstrap CA + Redis + PostgreSQL certs
+turnstone-admin tls-bootstrap --out /certs --issue redis --issue postgres
+
+# Output:
+#   /certs/ca.pem              (CA root certificate)
+#   /certs/certs/redis/        (Redis cert + key)
+#   /certs/certs/postgres/     (PostgreSQL cert + key)
+```
+
+The output directory is chmod 0700 (contains the CA private key).
+
+### Online Cert Issuance
+
+Request certs from a running console's ACME endpoint:
+
+```bash
+# Download CA root cert (TOFU — verify fingerprint)
+turnstone-admin tls-ca-cert --out ca.pem --console-url http://console:8080
+
+# Request a cert for a domain
+turnstone-admin tls-issue worker-1.internal --out /certs --console-url http://console:8080
+
+# List issued certs
+turnstone-admin tls-list --console-url http://console:8080 --auth-token $TOKEN
+```
+
+### Console URL Discovery
+
+If `--console-url` is not provided, the CLI discovers it from the `services`
+table in the shared database. The console registers itself on startup.
+
+---
+
+## Admin UI
+
+The **TLS** tab in the console admin panel (System group) shows:
+- CA status (common name, certificate count)
+- Certificate table (domain, SANs, issued, expires)
+- Force-renew and delete actions per certificate
+
+---
+
+## SDK
+
+### Python
+
+```python
+from turnstone.sdk import TurnstoneServer
+
+client = TurnstoneServer(
+    base_url="https://server:8080",
+    token="tok_xxx",
+    ca_cert="/path/to/ca.pem",
+    client_cert="/path/to/cert.pem",
+    client_key="/path/to/key.pem",
+)
+```
+
+### TypeScript
+
+```typescript
+import { TurnstoneServer } from "@turnstone/sdk";
+import { Agent } from "undici";
+import * as fs from "fs";
+
+const agent = new Agent({
+  connect: {
+    ca: fs.readFileSync("/path/to/ca.pem"),
+    cert: fs.readFileSync("/path/to/cert.pem"),
+    key: fs.readFileSync("/path/to/key.pem"),
+  },
+});
+
+const client = new TurnstoneServer({
+  baseUrl: "https://server:8080",
+  token: "tok_xxx",
+  // Node.js 18+ uses undici under the hood
+  fetch: (url, init) =>
+    fetch(url, { ...init, dispatcher: agent } as RequestInit),
+});
+```
+
+---
+
+## How It Works
+
+### Node Bootstrap Flow
+
+1. Node starts, connects to shared database (plain connection)
+2. Discovers console URL from `services` table
+3. Fetches CA root cert from `http://console/acme/ca.pem` (plain HTTP, TOFU)
+4. Requests service cert via ACME protocol (plain HTTP, JWS-signed)
+5. Starts auto-renewal (24h interval, re-issues before expiry)
+6. All subsequent inter-service communication uses mTLS
+
+### Console Startup Flow
+
+1. Read `tls.enabled` from ConfigStore
+2. Initialize CA (load from DB or generate new root key)
+3. Mount ACME responder at `/acme` (serves `/ca.pem` natively)
+4. Issue console certs (internal + optional frontend)
+5. Start CA-direct auto-renewal (no network, signs directly)
+6. Register console URL in services table with heartbeat
+
+---
+
+## Troubleshooting
+
+### Cert expired / mTLS connection refused
+
+Certs are valid for 48 hours. If auto-renewal stopped (e.g. console was down),
+restart the service to re-request a cert.
+
+### "No console service found"
+
+The console registers itself in the `services` table on startup. If the console
+hasn't started or the registration expired (1 hour TTL), nodes can't discover
+it. Use `--console-url` explicitly.
+
+### Let's Encrypt for console frontend
+
+Set `tls.acme_directory` to `https://acme-v02.api.letsencrypt.org/directory`
+in the admin Settings tab. The console will request a publicly trusted cert
+for its HTTPS endpoint. Internal mTLS still uses the private CA.
+
+### Verifying the cert chain
+
+```bash
+openssl s_client -connect server:8080 -CAfile ca.pem
+```

docs/tools.md

@@ -1,6 +1,6 @@
 # Tools Reference
 
-turnstone exposes 18 built-in tools plus any number of external MCP tools to the
+turnstone exposes 19 built-in tools plus any number of external MCP tools to the
 LLM via the OpenAI function-calling interface. Built-in tools are defined as JSON
 files under `turnstone/tools/` and loaded at startup by `turnstone/core/tools.py`.
 MCP tools are discovered from configured MCP servers at startup by
@@ -46,12 +46,12 @@ schema plus turnstone-specific metadata keys:
 
 | Name                | Description |
 |---------------------|-------------|
-| `TOOLS`             | All 17 tool definitions (sent to the model). |
+| `TOOLS`             | All 19 tool definitions (sent to the model). |
 | `AGENT_TOOLS`       | Tools with `agent: true` -- available to plan sub-agents. Read-only tools. |
 | `TASK_AGENT_TOOLS`  | Tools with `task_agent: true` -- available to task sub-agents. Includes write operations. |
 | `AGENT_AUTO_TOOLS`  | Set of tool names with `auto_approve: true` -- no user confirmation needed. |
 | `TASK_AUTO_TOOLS`   | Same as `AGENT_AUTO_TOOLS` (identical filter). |
-| `BUILTIN_TOOL_NAMES`| Frozenset of all 17 built-in tool names. Used by tool search to distinguish always-on tools from deferrable MCP tools. |
+| `BUILTIN_TOOL_NAMES`| Frozenset of all 19 built-in tool names. Used by tool search to distinguish always-on tools from deferrable MCP tools. |
 | `PRIMARY_KEY_MAP`   | Dict mapping tool name to its `primary_key` parameter name. |
 
 ---
@@ -69,7 +69,7 @@ Tool execution follows a three-phase pipeline inside `ChatSession._execute_tools
 - Parses the JSON arguments (with fallback for malformed JSON).
 - If JSON parsing fails entirely, uses `PRIMARY_KEY_MAP` to map a bare string
   to the correct parameter.
-- Dispatches to the matching `_prepare_{func_name}()` handler. There are 17
+- Dispatches to the matching `_prepare_{func_name}()` handler. There are 19
   built-in tools plus `tool_search` (synthetic, client-side BM25 fallback) and
   the generic `_prepare_mcp_tool()` handler for MCP tools.
 - Validates arguments and builds a preview dict containing:
@@ -102,10 +102,15 @@ Each item's `execute` callable is invoked:
 - Errored or denied items return their error/denial message without executing.
 - The `bash` tool streams stdout incrementally: each line calls
   `ui.on_tool_output_chunk(call_id, line)` as it is produced, then the final
-  combined output (stdout + stderr) is delivered via `ui.on_tool_result(call_id, name, output)`.
+  combined output (stdout + stderr) is delivered via
+  `ui.on_tool_result(call_id, name, output, is_error=...)`.
   The `call_id` links `tool_info`/`approve_request` items to their streaming chunks and
   final result, enabling correct routing when multiple bash tools run in parallel.
-  Other tools deliver results atomically via `ui.on_tool_result(call_id, name, output)` only.
+  The `is_error` flag is `True` when the tool execution failed (e.g. bash exit code >= 2
+  or signal, file not found, timeout). Exit code 1 is ambiguous and not flagged; user
+  denials are tracked separately. This removes the need for text-prefix heuristics.
+  Other tools deliver results atomically via
+  `ui.on_tool_result(call_id, name, output, is_error=...)` only.
 - Special post-execution gate for `plan`: the plan output is shown to the user
   for review, and the user can reject or annotate it.
 
@@ -183,8 +188,11 @@ Execute a bash command and return stdout + stderr.
 | Parameter | Type   | Required | Description |
 |-----------|--------|----------|-------------|
 | `command` | string | yes      | The bash command to execute. |
+| `timeout` | integer | no      | Timeout in seconds (1-600). Omit to use the global `tools.timeout` setting (typically 120s). |
+| `stop_on_error` | boolean | no | Enable `set -e` so the script exits on the first command failure. Default false. |
 
-- **What it does**: Runs the command in a subprocess with a configurable timeout. Commands are sanitized and checked against a blocklist (e.g. `rm -rf /`).
+- **What it does**: Runs the command in a subprocess with a configurable timeout. Commands are sanitized and checked against a blocklist (e.g. `rm -rf /`). Environment variables containing secrets are scrubbed (`*_KEY`, `*_SECRET`, `*_TOKEN`, etc.).
+- **Output format**: Stdout is returned directly. Stderr lines are prefixed with `[stderr]` so the model can distinguish them. When the command itself redirects stderr to stdout (`2>&1`), no prefix is added. Output exceeding 256KB is truncated (head + tail preserved, middle replaced with a truncation notice).
 - **Auto-approve**: No -- requires user confirmation.
 - **Agent availability**: `task_agent` only (not available to plan sub-agents).
 
@@ -216,8 +224,9 @@ Write content to a file, creating it if needed.
 |-----------|--------|----------|-------------|
 | `path`    | string | yes      | Absolute or relative file path. |
 | `content` | string | yes      | The full file content to write. |
+| `mode`    | string | no       | `"overwrite"` (default) replaces the file. `"append"` adds content to the end. |
 
-- **What it does**: Creates or overwrites the file at the given path. Parent directories are created as needed.
+- **What it does**: Creates or overwrites (or appends to) the file at the given path. Parent directories are created as needed.
 - **Auto-approve**: No -- requires user confirmation.
 - **Agent availability**: `task_agent` only.
 
@@ -225,21 +234,44 @@ Write content to a file, creating it if needed.
 
 ### edit_file
 
-Replace an exact string in a file with new content.
+Replace exact strings in a file, or apply multiple replacements atomically.
 
 | Parameter    | Type    | Required | Description |
 |--------------|---------|----------|-------------|
 | `path`       | string  | yes      | Absolute or relative file path. |
-| `old_string` | string  | yes      | The exact text to find and replace. |
-| `new_string` | string  | yes      | The replacement text. |
+| `old_string` | string  | no*      | The exact text to find and replace. |
+| `new_string` | string  | no*      | The replacement text. |
 | `near_line`  | integer | no       | Disambiguate when `old_string` matches multiple locations. |
+| `edits`       | array   | no*      | Multiple replacements to apply atomically (see below). |
+| `replace_all` | boolean | no       | Replace ALL occurrences of `old_string`. Cannot combine with `near_line` or `edits`. |
 
-- **What it does**: Finds `old_string` in the file and replaces it with `new_string`. Fails if the string is not found or matches multiple locations (unless `near_line` is provided to pick the nearest match). Requires a prior `read_file` call on the same path.
+\* Provide either `old_string`+`new_string` (single edit) or `edits` array (batch), not both.
+
+- **What it does**: Finds `old_string` in the file and replaces it with `new_string`. Fails if the string is not found or matches multiple locations (unless `near_line` or `replace_all` is provided). Requires a prior `read_file` or `diff_file` call on the same path.
+- **Batch mode**: The `edits` array accepts multiple `{old_string, new_string, near_line?}` entries applied atomically. All edits are validated before any are applied. Overlapping edits (two entries targeting the same text region) are rejected. Edits are applied in reverse file-position order so character offsets stay stable.
+- **Replace-all mode**: When `replace_all` is true, all occurrences are replaced via `str.replace()`. The approval preview shows the occurrence count.
 - **Auto-approve**: No -- requires user confirmation.
 - **Agent availability**: `task_agent` only.
 
 ---
 
+### diff_file
+
+Show a unified diff between two files, or between a file and a provided string.
+
+| Parameter       | Type    | Required | Description |
+|-----------------|---------|----------|-------------|
+| `path_a`        | string  | yes      | Path to the first file. |
+| `path_b`        | string  | no       | Path to the second file. Mutually exclusive with `content_b`. |
+| `content_b`     | string  | no       | String content to compare against `path_a`. Mutually exclusive with `path_b`. |
+| `context_lines` | integer | no       | Number of context lines around changes (default 3, max 20). |
+
+- **What it does**: Returns unified diff output using Python's `difflib`. Binary files (containing null bytes) are rejected with a clear error. Files read through `diff_file` satisfy `edit_file`'s read guard — you can diff then edit without a separate `read_file` call. Large diffs are streamed with early cutoff at the tool truncation limit.
+- **Auto-approve**: Yes (read-only).
+- **Agent availability**: `agent` and `task_agent`.
+
+---
+
 ### search
 
 Search file contents for a regex pattern.
@@ -265,8 +297,9 @@ Execute Python code for math and computation in a sandbox.
 |-----------|--------|----------|-------------|
 | `code`    | string | yes      | Python code to execute. Must use `print()` for output. |
 
-- **What it does**: Runs Python code in a sandboxed environment with pre-imported libraries: `sympy`, `numpy`, `scipy`, `math`, `fractions`, `itertools`, `functools`, `collections`, `decimal`, `operator`, `random`, `re`, `string`. Common sympy names (`symbols`, `solve`, `simplify`, `sqrt`, `Matrix`, etc.) are pre-imported.
-- **Auto-approve**: No -- requires user confirmation.
+- **What it does**: Runs Python code in a sandboxed environment with pre-imported libraries: `sympy`, `numpy`, `scipy`, `math`, `fractions`, `itertools`, `functools`, `collections`, `decimal`, `operator`, `random`, `re`, `string`. Common sympy names (`symbols`, `solve`, `simplify`, `sqrt`, `Matrix`, etc.) are pre-imported. `pytest` is also available for import.
+- **Installation**: `sympy`, `numpy`, `scipy`, and `pytest` require the `[sandbox]` extras group: `pip install turnstone[sandbox]` (included in `[all]`).
+- **Auto-approve**: Yes.
 - **Agent availability**: `agent` and `task_agent`.
 
 ---
@@ -492,7 +525,7 @@ data.get("mergedAt") is not None
 
 ---
 
-### load_skill
+### skill
 
 Discover and activate skills at runtime during a conversation. The model can
 search for available skills and load one by name, replacing the current active
@@ -543,7 +576,7 @@ pre-configure skills at workstream creation.
 | `watch`      | Monitor    | No (create)  | No    | No         | `command`   |
 | `read_resource`| MCP      | No           | Yes   | Yes        | `uri`       |
 | `use_prompt` | MCP        | No           | Yes   | Yes        | `name`      |
-| `load_skill` | Skills     | No (load)    | No    | No         | `name`      |
+| `skill`      | Skills     | No (load)    | No    | No         | `name`      |
 | `tool_search`| Search     | Yes          | No    | No         | `query`     |
 
 ---
@@ -591,12 +624,12 @@ CLI flags override the config file:
 
 ### How it works
 
-1. **Threshold check**: At session startup, `ToolSearchManager.should_activate()`
-   counts total tools (built-in + MCP). If the count is below the threshold, tool
-   search stays off and all tools are sent to the model directly.
+1. **Threshold check**: At session startup, if the total tool count (built-in + MCP)
+   is below the threshold, tool search stays off and all tools are sent to the model
+   directly.
 
 2. **Partitioning**: When active, tools are split into two sets:
-   - **Always-on** -- the 17 built-in tools (members of `BUILTIN_TOOL_NAMES`).
+   - **Always-on** -- the 19 built-in tools (members of `BUILTIN_TOOL_NAMES`).
      These are always visible to the model.
    - **Deferred** -- all MCP tools. These are not sent in the tool list unless
      the model searches for them.
@@ -639,7 +672,7 @@ MCP-compatible service.
 3. **Schema conversion**: Each MCP tool's `inputSchema` is converted to OpenAI
    function-calling format. The tool name is prefixed: `mcp__{server}__{tool}`.
 
-4. **Merging**: MCP tools are appended after the 17 built-in tools via
+4. **Merging**: MCP tools are appended after the 19 built-in tools via
    `merge_mcp_tools()`. Built-in tools appear first, giving them natural LLM priority.
    When dynamic tool search is active, MCP tools are deferred rather than directly
    visible -- the model discovers them via search as needed (see

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "turnstone"
-version = "0.8.2"
+version = "0.9.5"
 description = "Multi-node AI orchestration platform with tool use, agent routing, and cluster simulation."
 readme = "README.md"
 license = "BUSL-1.1"
@@ -51,8 +51,11 @@ console = ["redis>=7.2", "croniter>=3.0"]
 sim = ["redis>=7.2"]
 anthropic = ["anthropic>=0.39"]
 postgres = ["psycopg[binary]>=3.2"]
+ddg = ["ddgs>=9.0"]
 discord = ["discord.py>=2.4", "redis>=7.2"]
-all = ["turnstone[mq,console,sim,anthropic,postgres,discord]"]
+tls = ["lacme>=1.0.4"]
+sandbox = ["sympy>=1.13", "numpy>=2.0", "scipy>=1.14", "pytest>=9.0"]
+all = ["turnstone[mq,console,sim,anthropic,postgres,discord,ddg,tls,sandbox]"]
 
 [project.scripts]
 turnstone = "turnstone.cli:main"
@@ -77,7 +80,7 @@ include = [
   "turnstone/console/static/*.js",
   "turnstone/shared_static/*.css",
   "turnstone/shared_static/*.js",
-  "turnstone/shared_static/katex-0.16.38/**/*",
+  "turnstone/shared_static/katex-0.16.44/**/*",
   "turnstone/shared_static/hljs-11.11.1/**/*",
   "turnstone/shared_static/mermaid-11.13.0/**/*",
   "turnstone/sdk/py.typed",
@@ -164,6 +167,14 @@ ignore_missing_imports = true
 module = ["frontmatter", "frontmatter.*"]
 ignore_missing_imports = true
 
+[[tool.mypy.overrides]]
+module = ["ddgs", "ddgs.*"]
+ignore_missing_imports = true
+
+[[tool.mypy.overrides]]
+module = ["lacme", "lacme.*"]
+ignore_missing_imports = true
+
 [[tool.mypy.overrides]]
 module = ["turnstone.channels.discord.*"]
 disallow_subclassing_any = false

scripts/update-vendored-js.sh

@@ -28,9 +28,19 @@ usage() {
 LIB="$1"
 VERSION="$2"
 
-# Detect current version from pyproject.toml
+# Detect current version from the filesystem (not pyproject.toml, which
+# Renovate may have already updated).  Falls back to pyproject.toml if
+# no directory is found.
 detect_old_version() {
     local pattern="$1"
+    # Look for existing directory: e.g. turnstone/shared_static/katex-0.16.38
+    local dir
+    dir=$(find "${STATIC_DIR}" -maxdepth 1 -type d -name "${pattern}-*" | head -1)
+    if [[ -n "$dir" ]]; then
+        basename "$dir" | sed "s/${pattern}-//"
+        return
+    fi
+    # Fallback to pyproject.toml
     grep -oE "${pattern}-[0-9.]+" pyproject.toml | head -1 | sed "s/${pattern}-//"
 }
 
@@ -51,9 +61,19 @@ update_refs() {
     done
 }
 
+check_same_version() {
+    if [[ "$1" == "$2" ]]; then
+        echo "ERROR: Old version ($1) == new version ($2). Nothing to update."
+        echo "If the old directory was already removed, re-download with:"
+        echo "  rm -rf ${STATIC_DIR}/${3}-${1} && $0 $3 $2"
+        exit 1
+    fi
+}
+
 case "$LIB" in
     katex)
         OLD_VERSION=$(detect_old_version "katex")
+        check_same_version "$OLD_VERSION" "$VERSION" "katex"
         OLD_DIR="${STATIC_DIR}/katex-${OLD_VERSION}"
         NEW_DIR="${STATIC_DIR}/katex-${VERSION}"
 
@@ -87,6 +107,7 @@ case "$LIB" in
 
     hljs)
         OLD_VERSION=$(detect_old_version "hljs")
+        check_same_version "$OLD_VERSION" "$VERSION" "hljs"
         OLD_DIR="${STATIC_DIR}/hljs-${OLD_VERSION}"
         NEW_DIR="${STATIC_DIR}/hljs-${VERSION}"
 
@@ -107,6 +128,7 @@ case "$LIB" in
 
     mermaid)
         OLD_VERSION=$(detect_old_version "mermaid")
+        check_same_version "$OLD_VERSION" "$VERSION" "mermaid"
         OLD_DIR="${STATIC_DIR}/mermaid-${OLD_VERSION}"
         NEW_DIR="${STATIC_DIR}/mermaid-${VERSION}"
 

sdk/typescript/openapi-server.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "turnstone Server API",
-    "version": "0.7.0",
+    "version": "0.9.2",
     "description": "Single-node workstream management, chat interaction, and real-time streaming."
   },
   "paths": {
@@ -458,6 +458,27 @@
         }
       }
     },
+    "/v1/api/models": {
+      "get": {
+        "summary": "List available model aliases",
+        "operationId": "v1_api_models_get",
+        "tags": [
+          "Models"
+        ],
+        "responses": {
+          "200": {
+            "description": "Success",
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/ListAvailableModelsResponse"
+                }
+              }
+            }
+          }
+        }
+      }
+    },
     "/v1/api/auth/login": {
       "post": {
         "summary": "Authenticate with a token",
@@ -1223,6 +1244,12 @@
             "description": "Target workstream ID",
             "title": "Ws Id",
             "type": "string"
+          },
+          "force": {
+            "default": false,
+            "description": "Force cancel: abandon the stuck worker thread immediately. Use when cooperative cancel has not resolved within a few seconds.",
+            "title": "Force",
+            "type": "boolean"
           }
         },
         "required": [
@@ -1559,6 +1586,11 @@
             "title": "Version",
             "type": "string"
           },
+          "node_id": {
+            "default": "",
+            "title": "Node Id",
+            "type": "string"
+          },
           "uptime_seconds": {
             "default": 0.0,
             "title": "Uptime Seconds",
@@ -1569,6 +1601,12 @@
             "title": "Model",
             "type": "string"
           },
+          "max_ws": {
+            "default": 10,
+            "description": "Maximum concurrent workstreams",
+            "title": "Max Ws",
+            "type": "integer"
+          },
           "workstreams": {
             "$ref": "#/components/schemas/WorkstreamCounts",
             "default": {
@@ -1966,7 +2004,43 @@
         ],
         "title": "ListSkillSummaryResponse",
         "type": "object"
+      },
+      "AvailableModelInfo": {
+        "properties": {
+          "alias": {
+            "title": "Alias",
+            "type": "string"
+          },
+          "model": {
+            "title": "Model",
+            "type": "string"
+          },
+          "provider": {
+            "title": "Provider",
+            "type": "string"
+          }
+        },
+        "required": [
+          "alias",
+          "model",
+          "provider"
+        ],
+        "title": "AvailableModelInfo",
+        "type": "object"
+      },
+      "ListAvailableModelsResponse": {
+        "properties": {
+          "models": {
+            "items": {
+              "$ref": "#/components/schemas/AvailableModelInfo"
+            },
+            "title": "Models",
+            "type": "array"
+          }
+        },
+        "title": "ListAvailableModelsResponse",
+        "type": "object"
       }
     }
   }
-}
+}

sdk/typescript/package-lock.json

@@ -9,29 +9,31 @@
       "version": "0.3.0",
       "license": "BUSL-1.1",
       "devDependencies": {
-        "typescript": "^5.4",
+        "typescript": "^6.0.0",
         "vitest": "^4.1"
       }
     },
     "node_modules/@emnapi/core": {
-      "version": "1.9.0",
-      "resolved": "https://registry.npmjs.org/@emnapi/core/-/core-1.9.0.tgz",
-      "integrity": "sha512-0DQ98G9ZQZOxfUcQn1waV2yS8aWdZ6kJMbYCJB3oUBecjWYO1fqJ+a1DRfPF3O5JEkwqwP1A9QEN/9mYm2Yd0w==",
+      "version": "1.9.1",
+      "resolved": "https://registry.npmjs.org/@emnapi/core/-/core-1.9.1.tgz",
+      "integrity": "sha512-mukuNALVsoix/w1BJwFzwXBN/dHeejQtuVzcDsfOEsdpCumXb/E9j8w11h5S54tT1xhifGfbbSm/ICrObRb3KA==",
       "dev": true,
       "license": "MIT",
       "optional": true,
+      "peer": true,
       "dependencies": {
         "@emnapi/wasi-threads": "1.2.0",
         "tslib": "^2.4.0"
       }
     },
     "node_modules/@emnapi/runtime": {
-      "version": "1.9.0",
-      "resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.9.0.tgz",
-      "integrity": "sha512-QN75eB0IH2ywSpRpNddCRfQIhmJYBCJ1x5Lb3IscKAL8bMnVAKnRg8dCoXbHzVLLH7P38N2Z3mtulB7W0J0FKw==",
+      "version": "1.9.1",
+      "resolved": "https://registry.npmjs.org/@emnapi/runtime/-/runtime-1.9.1.tgz",
+      "integrity": "sha512-VYi5+ZVLhpgK4hQ0TAjiQiZ6ol0oe4mBx7mVv7IflsiEp0OWoVsp/+f9Vc1hOhE0TtkORVrI1GvzyreqpgWtkA==",
       "dev": true,
       "license": "MIT",
       "optional": true,
+      "peer": true,
       "dependencies": {
         "tslib": "^2.4.0"
       }
@@ -43,6 +45,7 @@
       "dev": true,
       "license": "MIT",
       "optional": true,
+      "peer": true,
       "dependencies": {
         "tslib": "^2.4.0"
       }
@@ -55,36 +58,28 @@
       "license": "MIT"
     },
     "node_modules/@napi-rs/wasm-runtime": {
-      "version": "1.1.1",
-      "resolved": "https://registry.npmjs.org/@napi-rs/wasm-runtime/-/wasm-runtime-1.1.1.tgz",
-      "integrity": "sha512-p64ah1M1ld8xjWv3qbvFwHiFVWrq1yFvV4f7w+mzaqiR4IlSgkqhcRdHwsGgomwzBH51sRY4NEowLxnaBjcW/A==",
+      "version": "1.1.2",
+      "resolved": "https://registry.npmjs.org/@napi-rs/wasm-runtime/-/wasm-runtime-1.1.2.tgz",
+      "integrity": "sha512-sNXv5oLJ7ob93xkZ1XnxisYhGYXfaG9f65/ZgYuAu3qt7b3NadcOEhLvx28hv31PgX8SZJRYrAIPQilQmFpLVw==",
       "dev": true,
       "license": "MIT",
       "optional": true,
       "dependencies": {
-        "@emnapi/core": "^1.7.1",
-        "@emnapi/runtime": "^1.7.1",
         "@tybys/wasm-util": "^0.10.1"
       },
       "funding": {
         "type": "github",
         "url": "https://github.com/sponsors/Brooooooklyn"
-      }
-    },
-    "node_modules/@oxc-project/runtime": {
-      "version": "0.115.0",
-      "resolved": "https://registry.npmjs.org/@oxc-project/runtime/-/runtime-0.115.0.tgz",
-      "integrity": "sha512-Rg8Wlt5dCbXhQnsXPrkOjL1DTSvXLgb2R/KYfnf1/K+R0k6UMLEmbQXPM+kwrWqSmWA2t0B1EtHy2/3zikQpvQ==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": "^20.19.0 || >=22.12.0"
+      },
+      "peerDependencies": {
+        "@emnapi/core": "^1.7.1",
+        "@emnapi/runtime": "^1.7.1"
       }
     },
     "node_modules/@oxc-project/types": {
-      "version": "0.115.0",
-      "resolved": "https://registry.npmjs.org/@oxc-project/types/-/types-0.115.0.tgz",
-      "integrity": "sha512-4n91DKnebUS4yjUHl2g3/b2T+IUdCfmoZGhmwsovZCDaJSs+QkVAM+0AqqTxHSsHfeiMuueT75cZaZcT/m0pSw==",
+      "version": "0.122.0",
+      "resolved": "https://registry.npmjs.org/@oxc-project/types/-/types-0.122.0.tgz",
+      "integrity": "sha512-oLAl5kBpV4w69UtFZ9xqcmTi+GENWOcPF7FCrczTiBbmC0ibXxCwyvZGbO39rCVEuLGAZM84DH0pUIyyv/YJzA==",
       "dev": true,
       "license": "MIT",
       "funding": {
@@ -92,9 +87,9 @@
       }
     },
     "node_modules/@rolldown/binding-android-arm64": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-android-arm64/-/binding-android-arm64-1.0.0-rc.9.tgz",
-      "integrity": "sha512-lcJL0bN5hpgJfSIz/8PIf02irmyL43P+j1pTCfbD1DbLkmGRuFIA4DD3B3ZOvGqG0XiVvRznbKtN0COQVaKUTg==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-android-arm64/-/binding-android-arm64-1.0.0-rc.12.tgz",
+      "integrity": "sha512-pv1y2Fv0JybcykuiiD3qBOBdz6RteYojRFY1d+b95WVuzx211CRh+ytI/+9iVyWQ6koTh5dawe4S/yRfOFjgaA==",
       "cpu": [
         "arm64"
       ],
@@ -109,9 +104,9 @@
       }
     },
     "node_modules/@rolldown/binding-darwin-arm64": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-arm64/-/binding-darwin-arm64-1.0.0-rc.9.tgz",
-      "integrity": "sha512-J7Zk3kLYFsLtuH6U+F4pS2sYVzac0qkjcO5QxHS7OS7yZu2LRs+IXo+uvJ/mvpyUljDJ3LROZPoQfgBIpCMhdQ==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-arm64/-/binding-darwin-arm64-1.0.0-rc.12.tgz",
+      "integrity": "sha512-cFYr6zTG/3PXXF3pUO+umXxt1wkRK/0AYT8lDwuqvRC+LuKYWSAQAQZjCWDQpAH172ZV6ieYrNnFzVVcnSflAg==",
       "cpu": [
         "arm64"
       ],
@@ -126,9 +121,9 @@
       }
     },
     "node_modules/@rolldown/binding-darwin-x64": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-x64/-/binding-darwin-x64-1.0.0-rc.9.tgz",
-      "integrity": "sha512-iwtmmghy8nhfRGeNAIltcNXzD0QMNaaA5U/NyZc1Ia4bxrzFByNMDoppoC+hl7cDiUq5/1CnFthpT9n+UtfFyg==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-x64/-/binding-darwin-x64-1.0.0-rc.12.tgz",
+      "integrity": "sha512-ZCsYknnHzeXYps0lGBz8JrF37GpE9bFVefrlmDrAQhOEi4IOIlcoU1+FwHEtyXGx2VkYAvhu7dyBf75EJQffBw==",
       "cpu": [
         "x64"
       ],
@@ -143,9 +138,9 @@
       }
     },
     "node_modules/@rolldown/binding-freebsd-x64": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-freebsd-x64/-/binding-freebsd-x64-1.0.0-rc.9.tgz",
-      "integrity": "sha512-DLFYI78SCiZr5VvdEplsVC2Vx53lnA4/Ga5C65iyldMVaErr86aiqCoNBLl92PXPfDtUYjUh+xFFor40ueNs4Q==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-freebsd-x64/-/binding-freebsd-x64-1.0.0-rc.12.tgz",
+      "integrity": "sha512-dMLeprcVsyJsKolRXyoTH3NL6qtsT0Y2xeuEA8WQJquWFXkEC4bcu1rLZZSnZRMtAqwtrF/Ib9Ddtpa/Gkge9Q==",
       "cpu": [
         "x64"
       ],
@@ -160,9 +155,9 @@
       }
     },
     "node_modules/@rolldown/binding-linux-arm-gnueabihf": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm-gnueabihf/-/binding-linux-arm-gnueabihf-1.0.0-rc.9.tgz",
-      "integrity": "sha512-CsjTmTwd0Hri6iTw/DRMK7kOZ7FwAkrO4h8YWKoX/kcj833e4coqo2wzIFywtch/8Eb5enQ/lwLM7w6JX1W5RQ==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm-gnueabihf/-/binding-linux-arm-gnueabihf-1.0.0-rc.12.tgz",
+      "integrity": "sha512-YqWjAgGC/9M1lz3GR1r1rP79nMgo3mQiiA+Hfo+pvKFK1fAJ1bCi0ZQVh8noOqNacuY1qIcfyVfP6HoyBRZ85Q==",
       "cpu": [
         "arm"
       ],
@@ -177,13 +172,16 @@
       }
     },
     "node_modules/@rolldown/binding-linux-arm64-gnu": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-gnu/-/binding-linux-arm64-gnu-1.0.0-rc.9.tgz",
-      "integrity": "sha512-2x9O2JbSPxpxMDhP9Z74mahAStibTlrBMW0520+epJH5sac7/LwZW5Bmg/E6CXuEF53JJFW509uP+lSedaUNxg==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-gnu/-/binding-linux-arm64-gnu-1.0.0-rc.12.tgz",
+      "integrity": "sha512-/I5AS4cIroLpslsmzXfwbe5OmWvSsrFuEw3mwvbQ1kDxJ822hFHIx+vsN/TAzNVyepI/j/GSzrtCIwQPeKCLIg==",
       "cpu": [
         "arm64"
       ],
       "dev": true,
+      "libc": [
+        "glibc"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -194,13 +192,16 @@
       }
     },
     "node_modules/@rolldown/binding-linux-arm64-musl": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-musl/-/binding-linux-arm64-musl-1.0.0-rc.9.tgz",
-      "integrity": "sha512-JA1QRW31ogheAIRhIg9tjMfsYbglXXYGNPLdPEYrwFxdbkQCAzvpSCSHCDWNl4hTtrol8WeboCSEpjdZK8qrCg==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-musl/-/binding-linux-arm64-musl-1.0.0-rc.12.tgz",
+      "integrity": "sha512-V6/wZztnBqlx5hJQqNWwFdxIKN0m38p8Jas+VoSfgH54HSj9tKTt1dZvG6JRHcjh6D7TvrJPWFGaY9UBVOaWPw==",
       "cpu": [
         "arm64"
       ],
       "dev": true,
+      "libc": [
+        "musl"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -211,13 +212,16 @@
       }
     },
     "node_modules/@rolldown/binding-linux-ppc64-gnu": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-ppc64-gnu/-/binding-linux-ppc64-gnu-1.0.0-rc.9.tgz",
-      "integrity": "sha512-aOKU9dJheda8Kj8Y3w9gnt9QFOO+qKPAl8SWd7JPHP+Cu0EuDAE5wokQubLzIDQWg2myXq2XhTpOVS07qqvT+w==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-ppc64-gnu/-/binding-linux-ppc64-gnu-1.0.0-rc.12.tgz",
+      "integrity": "sha512-AP3E9BpcUYliZCxa3w5Kwj9OtEVDYK6sVoUzy4vTOJsjPOgdaJZKFmN4oOlX0Wp0RPV2ETfmIra9x1xuayFB7g==",
       "cpu": [
         "ppc64"
       ],
       "dev": true,
+      "libc": [
+        "glibc"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -228,13 +232,16 @@
       }
     },
     "node_modules/@rolldown/binding-linux-s390x-gnu": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-s390x-gnu/-/binding-linux-s390x-gnu-1.0.0-rc.9.tgz",
-      "integrity": "sha512-OalO94fqj7IWRn3VdXWty75jC5dk4C197AWEuMhIpvVv2lw9fiPhud0+bW2ctCxb3YoBZor71QHbY+9/WToadA==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-s390x-gnu/-/binding-linux-s390x-gnu-1.0.0-rc.12.tgz",
+      "integrity": "sha512-nWwpvUSPkoFmZo0kQazZYOrT7J5DGOJ/+QHHzjvNlooDZED8oH82Yg67HvehPPLAg5fUff7TfWFHQS8IV1n3og==",
       "cpu": [
         "s390x"
       ],
       "dev": true,
+      "libc": [
+        "glibc"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -245,13 +252,16 @@
       }
     },
     "node_modules/@rolldown/binding-linux-x64-gnu": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-gnu/-/binding-linux-x64-gnu-1.0.0-rc.9.tgz",
-      "integrity": "sha512-cVEl1vZtBsBZna3YMjGXNvnYYrOJ7RzuWvZU0ffvJUexWkukMaDuGhUXn0rjnV0ptzGVkvc+vW9Yqy6h8YX4pg==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-gnu/-/binding-linux-x64-gnu-1.0.0-rc.12.tgz",
+      "integrity": "sha512-RNrafz5bcwRy+O9e6P8Z/OCAJW/A+qtBczIqVYwTs14pf4iV1/+eKEjdOUta93q2TsT/FI0XYDP3TCky38LMAg==",
       "cpu": [
         "x64"
       ],
       "dev": true,
+      "libc": [
+        "glibc"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -262,13 +272,16 @@
       }
     },
     "node_modules/@rolldown/binding-linux-x64-musl": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-musl/-/binding-linux-x64-musl-1.0.0-rc.9.tgz",
-      "integrity": "sha512-UzYnKCIIc4heAKgI4PZ3dfBGUZefGCJ1TPDuLHoCzgrMYPb5Rv6TLFuYtyM4rWyHM7hymNdsg5ik2C+UD9VDbA==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-musl/-/binding-linux-x64-musl-1.0.0-rc.12.tgz",
+      "integrity": "sha512-Jpw/0iwoKWx3LJ2rc1yjFrj+T7iHZn2JDg1Yny1ma0luviFS4mhAIcd1LFNxK3EYu3DHWCps0ydXQ5i/rrJ2ig==",
       "cpu": [
         "x64"
       ],
       "dev": true,
+      "libc": [
+        "musl"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -279,9 +292,9 @@
       }
     },
     "node_modules/@rolldown/binding-openharmony-arm64": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-openharmony-arm64/-/binding-openharmony-arm64-1.0.0-rc.9.tgz",
-      "integrity": "sha512-+6zoiF+RRyf5cdlFQP7nm58mq7+/2PFaY2DNQeD4B87N36JzfF/l9mdBkkmTvSYcYPE8tMh/o3cRlsx1ldLfog==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-openharmony-arm64/-/binding-openharmony-arm64-1.0.0-rc.12.tgz",
+      "integrity": "sha512-vRugONE4yMfVn0+7lUKdKvN4D5YusEiPilaoO2sgUWpCvrncvWgPMzK00ZFFJuiPgLwgFNP5eSiUlv2tfc+lpA==",
       "cpu": [
         "arm64"
       ],
@@ -296,9 +309,9 @@
       }
     },
     "node_modules/@rolldown/binding-wasm32-wasi": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-wasm32-wasi/-/binding-wasm32-wasi-1.0.0-rc.9.tgz",
-      "integrity": "sha512-rgFN6sA/dyebil3YTlL2evvi/M+ivhfnyxec7AccTpRPccno/rPoNlqybEZQBkcbZu8Hy+eqNJCqfBR8P7Pg8g==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-wasm32-wasi/-/binding-wasm32-wasi-1.0.0-rc.12.tgz",
+      "integrity": "sha512-ykGiLr/6kkiHc0XnBfmFJuCjr5ZYKKofkx+chJWDjitX+KsJuAmrzWhwyOMSHzPhzOHOy7u9HlFoa5MoAOJ/Zg==",
       "cpu": [
         "wasm32"
       ],
@@ -313,9 +326,9 @@
       }
     },
     "node_modules/@rolldown/binding-win32-arm64-msvc": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.0.0-rc.9.tgz",
-      "integrity": "sha512-lHVNUG/8nlF1IQk1C0Ci574qKYyty2goMiPlRqkC5R+3LkXDkL5Dhx8ytbxq35m+pkHVIvIxviD+TWLdfeuadA==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.0.0-rc.12.tgz",
+      "integrity": "sha512-5eOND4duWkwx1AzCxadcOrNeighiLwMInEADT0YM7xeEOOFcovWZCq8dadXgcRHSf3Ulh1kFo/qvzoFiCLOL1Q==",
       "cpu": [
         "arm64"
       ],
@@ -330,9 +343,9 @@
       }
     },
     "node_modules/@rolldown/binding-win32-x64-msvc": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-x64-msvc/-/binding-win32-x64-msvc-1.0.0-rc.9.tgz",
-      "integrity": "sha512-G0oA4+w1iY5AGi5HcDTxWsoxF509hrFIPB2rduV5aDqS9FtDg1CAfa7V34qImbjfhIcA8C+RekocJZA96EarwQ==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-x64-msvc/-/binding-win32-x64-msvc-1.0.0-rc.12.tgz",
+      "integrity": "sha512-PyqoipaswDLAZtot351MLhrlrh6lcZPo2LSYE+VDxbVk24LVKAGOuE4hb8xZQmrPAuEtTZW8E6D2zc5EUZX4Lw==",
       "cpu": [
         "x64"
       ],
@@ -347,9 +360,9 @@
       }
     },
     "node_modules/@rolldown/pluginutils": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/@rolldown/pluginutils/-/pluginutils-1.0.0-rc.9.tgz",
-      "integrity": "sha512-w6oiRWgEBl04QkFZgmW+jnU1EC9b57Oihi2ot3HNWIQRqgHp5PnYDia5iZ5FF7rpa4EQdiqMDXjlqKGXBhsoXw==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/@rolldown/pluginutils/-/pluginutils-1.0.0-rc.12.tgz",
+      "integrity": "sha512-HHMwmarRKvoFsJorqYlFeFRzXZqCt2ETQlEDOb9aqssrnVBB1/+xgTGtuTrIk5vzLNX1MjMtTf7W9z3tsSbrxw==",
       "dev": true,
       "license": "MIT"
     },
@@ -397,31 +410,31 @@
       "license": "MIT"
     },
     "node_modules/@vitest/expect": {
-      "version": "4.1.0",
-      "resolved": "https://registry.npmjs.org/@vitest/expect/-/expect-4.1.0.tgz",
-      "integrity": "sha512-EIxG7k4wlWweuCLG9Y5InKFwpMEOyrMb6ZJ1ihYu02LVj/bzUwn2VMU+13PinsjRW75XnITeFrQBMH5+dLvCDA==",
+      "version": "4.1.2",
+      "resolved": "https://registry.npmjs.org/@vitest/expect/-/expect-4.1.2.tgz",
+      "integrity": "sha512-gbu+7B0YgUJ2nkdsRJrFFW6X7NTP44WlhiclHniUhxADQJH5Szt9mZ9hWnJPJ8YwOK5zUOSSlSvyzRf0u1DSBQ==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
         "@standard-schema/spec": "^1.1.0",
         "@types/chai": "^5.2.2",
-        "@vitest/spy": "4.1.0",
-        "@vitest/utils": "4.1.0",
+        "@vitest/spy": "4.1.2",
+        "@vitest/utils": "4.1.2",
         "chai": "^6.2.2",
-        "tinyrainbow": "^3.0.3"
+        "tinyrainbow": "^3.1.0"
       },
       "funding": {
         "url": "https://opencollective.com/vitest"
       }
     },
     "node_modules/@vitest/mocker": {
-      "version": "4.1.0",
-      "resolved": "https://registry.npmjs.org/@vitest/mocker/-/mocker-4.1.0.tgz",
-      "integrity": "sha512-evxREh+Hork43+Y4IOhTo+h5lGmVRyjqI739Rz4RlUPqwrkFFDF6EMvOOYjTx4E8Tl6gyCLRL8Mu7Ry12a13Tw==",
+      "version": "4.1.2",
+      "resolved": "https://registry.npmjs.org/@vitest/mocker/-/mocker-4.1.2.tgz",
+      "integrity": "sha512-Ize4iQtEALHDttPRCmN+FKqOl2vxTiNUhzobQFFt/BM1lRUTG7zRCLOykG/6Vo4E4hnUdfVLo5/eqKPukcWW7Q==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@vitest/spy": "4.1.0",
+        "@vitest/spy": "4.1.2",
         "estree-walker": "^3.0.3",
         "magic-string": "^0.30.21"
       },
@@ -430,7 +443,7 @@
       },
       "peerDependencies": {
         "msw": "^2.4.9",
-        "vite": "^6.0.0 || ^7.0.0 || ^8.0.0-0"
+        "vite": "^6.0.0 || ^7.0.0 || ^8.0.0"
       },
       "peerDependenciesMeta": {
         "msw": {
@@ -442,26 +455,26 @@
       }
     },
     "node_modules/@vitest/pretty-format": {
-      "version": "4.1.0",
-      "resolved": "https://registry.npmjs.org/@vitest/pretty-format/-/pretty-format-4.1.0.tgz",
-      "integrity": "sha512-3RZLZlh88Ib0J7NQTRATfc/3ZPOnSUn2uDBUoGNn5T36+bALixmzphN26OUD3LRXWkJu4H0s5vvUeqBiw+kS0A==",
+      "version": "4.1.2",
+      "resolved": "https://registry.npmjs.org/@vitest/pretty-format/-/pretty-format-4.1.2.tgz",
+      "integrity": "sha512-dwQga8aejqeuB+TvXCMzSQemvV9hNEtDDpgUKDzOmNQayl2OG241PSWeJwKRH3CiC+sESrmoFd49rfnq7T4RnA==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "tinyrainbow": "^3.0.3"
+        "tinyrainbow": "^3.1.0"
       },
       "funding": {
         "url": "https://opencollective.com/vitest"
       }
     },
     "node_modules/@vitest/runner": {
-      "version": "4.1.0",
-      "resolved": "https://registry.npmjs.org/@vitest/runner/-/runner-4.1.0.tgz",
-      "integrity": "sha512-Duvx2OzQ7d6OjchL+trw+aSrb9idh7pnNfxrklo14p3zmNL4qPCDeIJAK+eBKYjkIwG96Bc6vYuxhqDXQOWpoQ==",
+      "version": "4.1.2",
+      "resolved": "https://registry.npmjs.org/@vitest/runner/-/runner-4.1.2.tgz",
+      "integrity": "sha512-Gr+FQan34CdiYAwpGJmQG8PgkyFVmARK8/xSijia3eTFgVfpcpztWLuP6FttGNfPLJhaZVP/euvujeNYar36OQ==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@vitest/utils": "4.1.0",
+        "@vitest/utils": "4.1.2",
         "pathe": "^2.0.3"
       },
       "funding": {
@@ -469,14 +482,14 @@
       }
     },
     "node_modules/@vitest/snapshot": {
-      "version": "4.1.0",
-      "resolved": "https://registry.npmjs.org/@vitest/snapshot/-/snapshot-4.1.0.tgz",
-      "integrity": "sha512-0Vy9euT1kgsnj1CHttwi9i9o+4rRLEaPRSOJ5gyv579GJkNpgJK+B4HSv/rAWixx2wdAFci1X4CEPjiu2bXIMg==",
+      "version": "4.1.2",
+      "resolved": "https://registry.npmjs.org/@vitest/snapshot/-/snapshot-4.1.2.tgz",
+      "integrity": "sha512-g7yfUmxYS4mNxk31qbOYsSt2F4m1E02LFqO53Xpzg3zKMhLAPZAjjfyl9e6z7HrW6LvUdTwAQR3HHfLjpko16A==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@vitest/pretty-format": "4.1.0",
-        "@vitest/utils": "4.1.0",
+        "@vitest/pretty-format": "4.1.2",
+        "@vitest/utils": "4.1.2",
         "magic-string": "^0.30.21",
         "pathe": "^2.0.3"
       },
@@ -485,9 +498,9 @@
       }
     },
     "node_modules/@vitest/spy": {
-      "version": "4.1.0",
-      "resolved": "https://registry.npmjs.org/@vitest/spy/-/spy-4.1.0.tgz",
-      "integrity": "sha512-pz77k+PgNpyMDv2FV6qmk5ZVau6c3R8HC8v342T2xlFxQKTrSeYw9waIJG8KgV9fFwAtTu4ceRzMivPTH6wSxw==",
+      "version": "4.1.2",
+      "resolved": "https://registry.npmjs.org/@vitest/spy/-/spy-4.1.2.tgz",
+      "integrity": "sha512-DU4fBnbVCJGNBwVA6xSToNXrkZNSiw59H8tcuUspVMsBDBST4nfvsPsEHDHGtWRRnqBERBQu7TrTKskmjqTXKA==",
       "dev": true,
       "license": "MIT",
       "funding": {
@@ -495,15 +508,15 @@
       }
     },
     "node_modules/@vitest/utils": {
-      "version": "4.1.0",
-      "resolved": "https://registry.npmjs.org/@vitest/utils/-/utils-4.1.0.tgz",
-      "integrity": "sha512-XfPXT6a8TZY3dcGY8EdwsBulFCIw+BeeX0RZn2x/BtiY/75YGh8FeWGG8QISN/WhaqSrE2OrlDgtF8q5uhOTmw==",
+      "version": "4.1.2",
+      "resolved": "https://registry.npmjs.org/@vitest/utils/-/utils-4.1.2.tgz",
+      "integrity": "sha512-xw2/TiX82lQHA06cgbqRKFb5lCAy3axQ4H4SoUFhUsg+wztiet+co86IAMDtF6Vm1hc7J6j09oh/rgDn+JdKIQ==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@vitest/pretty-format": "4.1.0",
+        "@vitest/pretty-format": "4.1.2",
         "convert-source-map": "^2.0.0",
-        "tinyrainbow": "^3.0.3"
+        "tinyrainbow": "^3.1.0"
       },
       "funding": {
         "url": "https://opencollective.com/vitest"
@@ -749,6 +762,9 @@
         "arm64"
       ],
       "dev": true,
+      "libc": [
+        "glibc"
+      ],
       "license": "MPL-2.0",
       "optional": true,
       "os": [
@@ -770,6 +786,9 @@
         "arm64"
       ],
       "dev": true,
+      "libc": [
+        "musl"
+      ],
       "license": "MPL-2.0",
       "optional": true,
       "os": [
@@ -791,6 +810,9 @@
         "x64"
       ],
       "dev": true,
+      "libc": [
+        "glibc"
+      ],
       "license": "MPL-2.0",
       "optional": true,
       "os": [
@@ -812,6 +834,9 @@
         "x64"
       ],
       "dev": true,
+      "libc": [
+        "musl"
+      ],
       "license": "MPL-2.0",
       "optional": true,
       "os": [
@@ -922,9 +947,9 @@
       "license": "ISC"
     },
     "node_modules/picomatch": {
-      "version": "4.0.3",
-      "resolved": "https://registry.npmjs.org/picomatch/-/picomatch-4.0.3.tgz",
-      "integrity": "sha512-5gTmgEY/sqK6gFXLIsQNH19lWb4ebPDLA4SdLP7dsWkIXHWlG66oPuVvXSGFPppYZz8ZDZq0dYYrbHfBCVUb1Q==",
+      "version": "4.0.4",
+      "resolved": "https://registry.npmjs.org/picomatch/-/picomatch-4.0.4.tgz",
+      "integrity": "sha512-QP88BAKvMam/3NxH6vj2o21R6MjxZUAd6nlwAS/pnGvN9IVLocLHxGYIzFhg6fUQ+5th6P4dv4eW9jX3DSIj7A==",
       "dev": true,
       "license": "MIT",
       "engines": {
@@ -964,14 +989,14 @@
       }
     },
     "node_modules/rolldown": {
-      "version": "1.0.0-rc.9",
-      "resolved": "https://registry.npmjs.org/rolldown/-/rolldown-1.0.0-rc.9.tgz",
-      "integrity": "sha512-9EbgWge7ZH+yqb4d2EnELAntgPTWbfL8ajiTW+SyhJEC4qhBbkCKbqFV4Ge4zmu5ziQuVbWxb/XwLZ+RIO7E8Q==",
+      "version": "1.0.0-rc.12",
+      "resolved": "https://registry.npmjs.org/rolldown/-/rolldown-1.0.0-rc.12.tgz",
+      "integrity": "sha512-yP4USLIMYrwpPHEFB5JGH1uxhcslv6/hL0OyvTuY+3qlOSJvZ7ntYnoWpehBxufkgN0cvXxppuTu5hHa/zPh+A==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@oxc-project/types": "=0.115.0",
-        "@rolldown/pluginutils": "1.0.0-rc.9"
+        "@oxc-project/types": "=0.122.0",
+        "@rolldown/pluginutils": "1.0.0-rc.12"
       },
       "bin": {
         "rolldown": "bin/cli.mjs"
@@ -980,21 +1005,21 @@
         "node": "^20.19.0 || >=22.12.0"
       },
       "optionalDependencies": {
-        "@rolldown/binding-android-arm64": "1.0.0-rc.9",
-        "@rolldown/binding-darwin-arm64": "1.0.0-rc.9",
-        "@rolldown/binding-darwin-x64": "1.0.0-rc.9",
-        "@rolldown/binding-freebsd-x64": "1.0.0-rc.9",
-        "@rolldown/binding-linux-arm-gnueabihf": "1.0.0-rc.9",
-        "@rolldown/binding-linux-arm64-gnu": "1.0.0-rc.9",
-        "@rolldown/binding-linux-arm64-musl": "1.0.0-rc.9",
-        "@rolldown/binding-linux-ppc64-gnu": "1.0.0-rc.9",
-        "@rolldown/binding-linux-s390x-gnu": "1.0.0-rc.9",
-        "@rolldown/binding-linux-x64-gnu": "1.0.0-rc.9",
-        "@rolldown/binding-linux-x64-musl": "1.0.0-rc.9",
-        "@rolldown/binding-openharmony-arm64": "1.0.0-rc.9",
-        "@rolldown/binding-wasm32-wasi": "1.0.0-rc.9",
-        "@rolldown/binding-win32-arm64-msvc": "1.0.0-rc.9",
-        "@rolldown/binding-win32-x64-msvc": "1.0.0-rc.9"
+        "@rolldown/binding-android-arm64": "1.0.0-rc.12",
+        "@rolldown/binding-darwin-arm64": "1.0.0-rc.12",
+        "@rolldown/binding-darwin-x64": "1.0.0-rc.12",
+        "@rolldown/binding-freebsd-x64": "1.0.0-rc.12",
+        "@rolldown/binding-linux-arm-gnueabihf": "1.0.0-rc.12",
+        "@rolldown/binding-linux-arm64-gnu": "1.0.0-rc.12",
+        "@rolldown/binding-linux-arm64-musl": "1.0.0-rc.12",
+        "@rolldown/binding-linux-ppc64-gnu": "1.0.0-rc.12",
+        "@rolldown/binding-linux-s390x-gnu": "1.0.0-rc.12",
+        "@rolldown/binding-linux-x64-gnu": "1.0.0-rc.12",
+        "@rolldown/binding-linux-x64-musl": "1.0.0-rc.12",
+        "@rolldown/binding-openharmony-arm64": "1.0.0-rc.12",
+        "@rolldown/binding-wasm32-wasi": "1.0.0-rc.12",
+        "@rolldown/binding-win32-arm64-msvc": "1.0.0-rc.12",
+        "@rolldown/binding-win32-x64-msvc": "1.0.0-rc.12"
       }
     },
     "node_modules/siginfo": {
@@ -1081,9 +1106,9 @@
       "optional": true
     },
     "node_modules/typescript": {
-      "version": "5.9.3",
-      "resolved": "https://registry.npmjs.org/typescript/-/typescript-5.9.3.tgz",
-      "integrity": "sha512-jl1vZzPDinLr9eUt3J/t7V6FgNEw9QjvBPdysz9KfQDD41fQrC2Y4vKQdiaUpFT4bXlb1RHhLpp8wtm6M5TgSw==",
+      "version": "6.0.2",
+      "resolved": "https://registry.npmjs.org/typescript/-/typescript-6.0.2.tgz",
+      "integrity": "sha512-bGdAIrZ0wiGDo5l8c++HWtbaNCWTS4UTv7RaTH/ThVIgjkveJt83m74bBHMJkuCbslY8ixgLBVZJIOiQlQTjfQ==",
       "dev": true,
       "license": "Apache-2.0",
       "bin": {
@@ -1095,17 +1120,16 @@
       }
     },
     "node_modules/vite": {
-      "version": "8.0.0",
-      "resolved": "https://registry.npmjs.org/vite/-/vite-8.0.0.tgz",
-      "integrity": "sha512-fPGaRNj9Zytaf8LEiBhY7Z6ijnFKdzU/+mL8EFBaKr7Vw1/FWcTBAMW0wLPJAGMPX38ZPVCVgLceWiEqeoqL2Q==",
+      "version": "8.0.3",
+      "resolved": "https://registry.npmjs.org/vite/-/vite-8.0.3.tgz",
+      "integrity": "sha512-B9ifbFudT1TFhfltfaIPgjo9Z3mDynBTJSUYxTjOQruf/zHH+ezCQKcoqO+h7a9Pw9Nm/OtlXAiGT1axBgwqrQ==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@oxc-project/runtime": "0.115.0",
         "lightningcss": "^1.32.0",
-        "picomatch": "^4.0.3",
+        "picomatch": "^4.0.4",
         "postcss": "^8.5.8",
-        "rolldown": "1.0.0-rc.9",
+        "rolldown": "1.0.0-rc.12",
         "tinyglobby": "^0.2.15"
       },
       "bin": {
@@ -1122,7 +1146,7 @@
       },
       "peerDependencies": {
         "@types/node": "^20.19.0 || >=22.12.0",
-        "@vitejs/devtools": "^0.0.0-alpha.31",
+        "@vitejs/devtools": "^0.1.0",
         "esbuild": "^0.27.0",
         "jiti": ">=1.21.0",
         "less": "^4.0.0",
@@ -1174,19 +1198,19 @@
       }
     },
     "node_modules/vitest": {
-      "version": "4.1.0",
-      "resolved": "https://registry.npmjs.org/vitest/-/vitest-4.1.0.tgz",
-      "integrity": "sha512-YbDrMF9jM2Lqc++2530UourxZHmkKLxrs4+mYhEwqWS97WJ7wOYEkcr+QfRgJ3PW9wz3odRijLZjHEaRLTNbqw==",
+      "version": "4.1.2",
+      "resolved": "https://registry.npmjs.org/vitest/-/vitest-4.1.2.tgz",
+      "integrity": "sha512-xjR1dMTVHlFLh98JE3i/f/WePqJsah4A0FK9cc8Ehp9Udk0AZk6ccpIZhh1qJ/yxVWRZ+Q54ocnD8TXmkhspGg==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@vitest/expect": "4.1.0",
-        "@vitest/mocker": "4.1.0",
-        "@vitest/pretty-format": "4.1.0",
-        "@vitest/runner": "4.1.0",
-        "@vitest/snapshot": "4.1.0",
-        "@vitest/spy": "4.1.0",
-        "@vitest/utils": "4.1.0",
+        "@vitest/expect": "4.1.2",
+        "@vitest/mocker": "4.1.2",
+        "@vitest/pretty-format": "4.1.2",
+        "@vitest/runner": "4.1.2",
+        "@vitest/snapshot": "4.1.2",
+        "@vitest/spy": "4.1.2",
+        "@vitest/utils": "4.1.2",
         "es-module-lexer": "^2.0.0",
         "expect-type": "^1.3.0",
         "magic-string": "^0.30.21",
@@ -1197,8 +1221,8 @@
         "tinybench": "^2.9.0",
         "tinyexec": "^1.0.2",
         "tinyglobby": "^0.2.15",
-        "tinyrainbow": "^3.0.3",
-        "vite": "^6.0.0 || ^7.0.0 || ^8.0.0-0",
+        "tinyrainbow": "^3.1.0",
+        "vite": "^6.0.0 || ^7.0.0 || ^8.0.0",
         "why-is-node-running": "^2.3.0"
       },
       "bin": {
@@ -1214,13 +1238,13 @@
         "@edge-runtime/vm": "*",
         "@opentelemetry/api": "^1.9.0",
         "@types/node": "^20.0.0 || ^22.0.0 || >=24.0.0",
-        "@vitest/browser-playwright": "4.1.0",
-        "@vitest/browser-preview": "4.1.0",
-        "@vitest/browser-webdriverio": "4.1.0",
-        "@vitest/ui": "4.1.0",
+        "@vitest/browser-playwright": "4.1.2",
+        "@vitest/browser-preview": "4.1.2",
+        "@vitest/browser-webdriverio": "4.1.2",
+        "@vitest/ui": "4.1.2",
         "happy-dom": "*",
         "jsdom": "*",
-        "vite": "^6.0.0 || ^7.0.0 || ^8.0.0-0"
+        "vite": "^6.0.0 || ^7.0.0 || ^8.0.0"
       },
       "peerDependenciesMeta": {
         "@edge-runtime/vm": {

sdk/typescript/package.json

@@ -32,7 +32,7 @@
   ],
   "license": "BUSL-1.1",
   "devDependencies": {
-    "typescript": "^5.4",
+    "typescript": "^6.0.0",
     "vitest": "^4.1"
   }
 }

sdk/typescript/src/base.ts

@@ -1,6 +1,15 @@
 import { TurnstoneAPIError } from "./errors.js";
 import { parseSSEStream } from "./sse.js";
 
+export interface TlsOptions {
+  /** Path to CA certificate PEM file (Node.js only). */
+  caCert?: string;
+  /** Path to client certificate PEM file for mTLS (Node.js only). */
+  clientCert?: string;
+  /** Path to client key PEM file for mTLS (Node.js only). */
+  clientKey?: string;
+}
+
 export interface ClientOptions {
   /** Server base URL (e.g. "http://localhost:8080"). */
   baseUrl: string;
@@ -8,6 +17,13 @@ export interface ClientOptions {
   token?: string;
   /** Custom fetch implementation (defaults to globalThis.fetch). */
   fetch?: typeof globalThis.fetch;
+  /**
+   * TLS certificate paths for documentation and tooling.
+   * The SDK does not read these directly — pass a custom `fetch`
+   * configured with your runtime's TLS agent (e.g. Node.js https.Agent).
+   * See docs/tls.md for examples.
+   */
+  tls?: TlsOptions;
 }
 
 export interface RequestOptions {

sdk/typescript/src/console.ts

@@ -44,6 +44,7 @@ import type {
   OrgInfo,
   RoleInfo,
   ScheduleInfo,
+  DeleteSettingResponse,
   SettingInfo,
   StatusResponse,
   ToolPolicyInfo,
@@ -394,7 +395,10 @@ export class TurnstoneConsole extends BaseClient {
     });
   }
 
-  async deleteSetting(key: string, nodeId?: string): Promise<StatusResponse> {
+  async deleteSetting(
+    key: string,
+    nodeId?: string,
+  ): Promise<DeleteSettingResponse> {
     const params: Record<string, string> = {};
     if (nodeId) params.node_id = nodeId;
     return this.request("DELETE", `/v1/api/admin/settings/${key}`, {

sdk/typescript/src/events.ts

@@ -38,6 +38,11 @@ export interface StreamEndEvent {
   type: "stream_end";
 }
 
+export interface StateChangeEvent {
+  type: "state_change";
+  state: "idle" | "thinking" | "running" | "attention" | "error";
+}
+
 export interface ToolInfoEvent {
   type: "tool_info";
   items: Array<Record<string, unknown>>;
@@ -59,6 +64,7 @@ export interface ToolResultEvent {
   call_id: string;
   name: string;
   output: string;
+  is_error?: boolean;
 }
 
 export interface ToolOutputChunkEvent {
@@ -77,6 +83,8 @@ export interface StatusEvent {
   effort: string;
   cache_creation_tokens?: number;
   cache_read_tokens?: number;
+  tool_calls_this_turn?: number;
+  turn_count?: number;
 }
 
 export interface PlanReviewEvent {
@@ -149,6 +157,7 @@ export type ServerEvent =
   | ContentEvent
   | ReasoningEvent
   | StreamEndEvent
+  | StateChangeEvent
   | ToolInfoEvent
   | ApproveRequestEvent
   | ApprovalResolvedEvent
@@ -246,6 +255,10 @@ export function isStreamEndEvent(e: ServerEvent): e is StreamEndEvent {
   return e.type === "stream_end";
 }
 
+export function isStateChangeEvent(e: ServerEvent): e is StateChangeEvent {
+  return e.type === "state_change";
+}
+
 export function isToolResultEvent(e: ServerEvent): e is ToolResultEvent {
   return e.type === "tool_result";
 }

sdk/typescript/src/index.ts

@@ -19,7 +19,7 @@
 // Clients
 export { TurnstoneServer } from "./server.js";
 export { TurnstoneConsole } from "./console.js";
-export type { ClientOptions } from "./base.js";
+export type { ClientOptions, TlsOptions } from "./base.js";
 
 // Errors
 export { TurnstoneAPIError } from "./errors.js";
@@ -35,6 +35,7 @@ export type {
   ContentEvent,
   ReasoningEvent,
   StreamEndEvent,
+  StateChangeEvent,
   ToolInfoEvent,
   ApproveRequestEvent,
   ApprovalResolvedEvent,
@@ -65,6 +66,7 @@ export {
   isReasoningEvent,
   isErrorEvent,
   isStreamEndEvent,
+  isStateChangeEvent,
   isToolResultEvent,
   isWsStateEvent,
   isApproveRequestEvent,
@@ -98,6 +100,7 @@ export type {
   AuthLoginResponse,
   AuthStatusResponse,
   AuthSetupResponse,
+  DeleteSettingResponse,
   StatusResponse,
   ErrorResponse,
   ClusterOverviewResponse,

sdk/typescript/src/server.ts

@@ -93,10 +93,13 @@ export class TurnstoneServer extends BaseClient {
     });
   }
 
-  async cancel(wsId: string): Promise<StatusResponse> {
-    return this.request("POST", "/v1/api/cancel", {
-      json: { ws_id: wsId },
-    });
+  async cancel(
+    wsId: string,
+    opts?: { force?: boolean },
+  ): Promise<StatusResponse> {
+    const body: Record<string, unknown> = { ws_id: wsId };
+    if (opts?.force) body.force = true;
+    return this.request("POST", "/v1/api/cancel", { json: body });
   }
 
   // -- Streaming ------------------------------------------------------------

sdk/typescript/src/types.ts

@@ -10,6 +10,12 @@ export interface StatusResponse {
   status: string;
 }
 
+export interface DeleteSettingResponse {
+  status: string;
+  key: string;
+  default: unknown;
+}
+
 export interface AuthLoginRequest {
   token: string;
 }
@@ -186,7 +192,10 @@ export interface SkillInfo {
   agent_max_turns: number | null;
   notify_on_complete: string;
   enabled: boolean;
+  priority: number;
   allowed_tools: string;
+  license: string;
+  compatibility: string;
   resource_count: number;
   created: string;
   updated: string;
@@ -213,7 +222,10 @@ export interface CreateSkillRequest {
   agent_max_turns?: number | null;
   notify_on_complete?: string;
   enabled?: boolean;
+  priority?: number;
   allowed_tools?: string;
+  license?: string;
+  compatibility?: string;
 }
 
 export interface UpdateSkillRequest {
@@ -236,7 +248,10 @@ export interface UpdateSkillRequest {
   agent_max_turns?: number | null;
   notify_on_complete?: string;
   enabled?: boolean;
+  priority?: number;
   allowed_tools?: string;
+  license?: string;
+  compatibility?: string;
 }
 
 export interface ListSkillsResponse {
@@ -396,6 +411,7 @@ export interface ConsoleCreateWsRequest {
   model?: string;
   initial_message?: string;
   skill?: string;
+  resume_ws?: string;
 }
 
 export interface ConsoleCreateWsResponse {

tests.json

@@ -1,5 +1,5 @@
 {
-  "description": "turnstone behavior tests — tool selection, sequencing, and multi-step reasoning",
+  "description": "turnstone behavior tests \u2014 tool selection, sequencing, and multi-step reasoning",
   "defaults": {
     "n_runs": 5,
     "max_turns": 15
@@ -8,35 +8,91 @@
     {
       "id": "read-before-edit",
       "description": "Must read_file before edit_file on the same path",
-      "user_prompt": "Fix the typo in config.py — change 'recieve' to 'receive'",
+      "user_prompt": "Fix the typo in config.py \u2014 change 'recieve' to 'receive'",
       "setup": {
         "files": {
           "config.py": "# Config module\ndef recieve_data(source):\n    \"\"\"Recieve data from source.\"\"\"\n    return source.read()\n"
         }
       },
       "expected_actions": [
-        { "tool": "read_file", "args": { "path": "config.py" } },
-        { "tool": "edit_file", "args": { "path": "config.py" } }
+        {
+          "tool": "read_file",
+          "args": {
+            "path": "config.py"
+          }
+        },
+        {
+          "tool": "edit_file",
+          "args": {
+            "path": "config.py"
+          }
+        }
       ],
-      "match_mode": "ordered_subset"
+      "match_mode": "ordered_subset",
+      "user_prompts": [
+        "Fix the typo in config.py \u2014 change 'recieve' to 'receive'",
+        "In config.py, correct the misspelling of 'recieve' to 'receive'",
+        "Please update config.py by replacing 'recieve' with the correct spelling 'receive'",
+        "There's a typo in config.py: 'recieve' should be 'receive'. Please fix it.",
+        "Could you change 'recieve' to 'receive' in config.py?",
+        "Go ahead and fix 'recieve' \u2192 'receive' in config.py",
+        "I need the word 'recieve' corrected to 'receive' in the file config.py",
+        "config.py has a spelling error \u2014 'recieve' needs to be changed to 'receive'",
+        "Kindly rectify the typographical error in config.py, replacing 'recieve' with 'receive'",
+        "Hey, swap 'recieve' for 'receive' in config.py"
+      ]
     },
     {
       "id": "write-file-not-bash",
       "description": "Use write_file for file creation, not bash echo/cat",
       "user_prompt": "Create a file called hello.py that prints hello world",
       "expected_actions": [
-        { "tool": "write_file", "args_pattern": { "path": "hello\\.py" } }
+        {
+          "tool": "write_file",
+          "args_pattern": {
+            "path": "hello\\.py"
+          }
+        }
       ],
-      "match_mode": "subset"
+      "match_mode": "subset",
+      "user_prompts": [
+        "Create a file called hello.py that prints hello world",
+        "Make a hello.py file that outputs hello world",
+        "Please write a Python file named hello.py which prints hello world",
+        "I need a file called hello.py that prints hello world",
+        "Could you create hello.py with code that prints hello world?",
+        "Write hello.py \u2014 it should print hello world",
+        "Generate a hello.py file that outputs \"hello world\"",
+        "I'd like you to create a file named hello.py that prints hello world",
+        "Set up a file called hello.py to print hello world",
+        "Kindly produce a hello.py file whose purpose is to print hello world"
+      ]
     },
     {
       "id": "bash-for-commands",
       "description": "Use bash for running system commands",
       "user_prompt": "What Python version is installed?",
       "expected_actions": [
-        { "tool": "bash", "args_pattern": { "command": "python" } }
+        {
+          "tool": "bash",
+          "args_pattern": {
+            "command": "python"
+          }
+        }
       ],
-      "match_mode": "subset"
+      "match_mode": "subset",
+      "user_prompts": [
+        "What Python version is installed?",
+        "Check which version of Python is currently installed",
+        "Can you tell me the installed Python version?",
+        "python --version please",
+        "I need to know what version of Python is on this system",
+        "Which Python version do we have?",
+        "Could you look up the Python version that's installed here?",
+        "Determine the currently installed Python version",
+        "What's the Python version on this machine?",
+        "Please check the Python version"
+      ]
     },
     {
       "id": "search-for-patterns",
@@ -49,13 +105,30 @@
         }
       },
       "expected_actions": [
-        { "tool": "search", "args_pattern": { "query": "test_" } }
+        {
+          "tool": "search",
+          "args_pattern": {
+            "query": "test_"
+          }
+        }
       ],
-      "match_mode": "subset"
+      "match_mode": "subset",
+      "user_prompts": [
+        "Find all functions that start with 'test_' in the project",
+        "List every function in the project whose name begins with 'test_'",
+        "I need to locate all functions prefixed with 'test_' across the project",
+        "Could you search the project for any functions starting with 'test_'?",
+        "Show me all the test_ prefixed functions in this project",
+        "Hunt down every function that has a 'test_' prefix in the codebase",
+        "I'm looking for all functions named test_* throughout the project",
+        "Search the entire project for functions whose names start with test_",
+        "What functions beginning with 'test_' exist in this project?",
+        "Please identify all functions with the 'test_' prefix in the project files"
+      ]
     },
     {
       "id": "multi-file-edit",
-      "description": "Read and edit multiple files — must read before editing each, and edit both",
+      "description": "Read and edit multiple files \u2014 must read before editing each, and edit both",
       "user_prompt": "Change the default port from 8000 to 9000 in both server.py and config.py",
       "setup": {
         "files": {
@@ -64,12 +137,32 @@
         }
       },
       "expected_actions": [
-        { "tool": "read_file" },
-        { "tool": "read_file" },
-        { "tool": "edit_file" },
-        { "tool": "edit_file" }
+        {
+          "tool": "read_file"
+        },
+        {
+          "tool": "read_file"
+        },
+        {
+          "tool": "edit_file"
+        },
+        {
+          "tool": "edit_file"
+        }
       ],
-      "match_mode": "ordered_subset"
+      "match_mode": "ordered_subset",
+      "user_prompts": [
+        "Change the default port from 8000 to 9000 in both server.py and config.py",
+        "Update the default port to 9000 instead of 8000 in server.py and config.py",
+        "Could you modify the port number from 8000 to 9000 in both config.py and server.py?",
+        "Please replace port 8000 with 9000 in server.py and config.py",
+        "I need the default port switched from 8000 to 9000 in both server.py and config.py",
+        "In server.py and config.py, the default port should be changed from 8000 to 9000",
+        "Swap out port 8000 for 9000 in config.py and server.py",
+        "Would you mind updating the default port value from 8000 to 9000 across both server.py and config.py?",
+        "The default port in server.py and config.py needs to be 9000 instead of 8000 \u2014 please make that change",
+        "Go ahead and change 8000 to 9000 for the default port in both server.py and config.py"
+      ]
     },
     {
       "id": "search-then-edit",
@@ -83,51 +176,147 @@
         }
       },
       "expected_actions": [
-        { "tool": "search", "args_pattern": { "query": "MAX_RETRIES" } },
-        { "tool": "read_file" },
-        { "tool": "edit_file", "args_pattern": { "old_string": "3" } }
+        {
+          "tool": "search",
+          "args_pattern": {
+            "query": "MAX_RETRIES"
+          }
+        },
+        {
+          "tool": "read_file"
+        },
+        {
+          "tool": "edit_file",
+          "args_pattern": {
+            "old_string": "3"
+          }
+        }
       ],
-      "match_mode": "ordered_subset"
+      "match_mode": "ordered_subset",
+      "user_prompts": [
+        "Find where MAX_RETRIES is defined and change it from 3 to 5",
+        "Locate the definition of MAX_RETRIES and update its value from 3 to 5",
+        "Could you search for where MAX_RETRIES is defined and modify it from 3 to 5?",
+        "I need MAX_RETRIES changed from 3 to 5 \u2014 find where it's defined and update it",
+        "Please find the MAX_RETRIES definition and bump it from 3 to 5",
+        "Hunt down MAX_RETRIES in the codebase and change its value from 3 to 5",
+        "Where is MAX_RETRIES set to 3? Change it to 5.",
+        "Search the code for the MAX_RETRIES definition and alter it from 3 to 5",
+        "I'd like you to locate MAX_RETRIES (currently 3) and set it to 5 instead",
+        "Go find MAX_RETRIES and switch it from 3 to 5"
+      ]
     },
     {
       "id": "bash-git-log",
       "description": "Use bash for git commands, not other tools",
       "user_prompt": "Show me the git log for the last 5 commits",
       "expected_actions": [
-        { "tool": "bash", "args_pattern": { "command": "git\\s+log" } }
+        {
+          "tool": "bash",
+          "args_pattern": {
+            "command": "git\\s+log"
+          }
+        }
       ],
-      "match_mode": "subset"
+      "match_mode": "subset",
+      "user_prompts": [
+        "Show me the git log for the last 5 commits",
+        "Display the 5 most recent git commits",
+        "Can you pull up the git log limited to the last five commits?",
+        "I need to see the git log showing only the previous 5 commits",
+        "git log for the 5 latest commits, please",
+        "Would you mind showing me the last five entries in the git log?",
+        "Print out the most recent 5 commits from the git log",
+        "I'd like to review the git log \u2014 just the last 5 commits",
+        "Show the recent 5 commit history using git log",
+        "Could you display the git commit history for the past five commits?"
+      ]
     },
     {
       "id": "write-then-run",
       "description": "Create a script and run it to verify it works",
       "user_prompt": "Create a Python script called fib.py that prints the first 10 Fibonacci numbers, then run it to verify",
       "expected_actions": [
-        { "tool": "write_file", "args_pattern": { "path": "fib\\.py" } },
-        { "tool": "bash", "args_pattern": { "command": "python" } }
+        {
+          "tool": "write_file",
+          "args_pattern": {
+            "path": "fib\\.py"
+          }
+        },
+        {
+          "tool": "bash",
+          "args_pattern": {
+            "command": "python"
+          }
+        }
       ],
-      "match_mode": "ordered_subset"
+      "match_mode": "ordered_subset",
+      "user_prompts": [
+        "Create a Python script called fib.py that prints the first 10 Fibonacci numbers, then run it to verify",
+        "Write a Python file named fib.py that outputs the first 10 Fibonacci numbers, and execute it to confirm it works",
+        "Please make fib.py \u2013 a Python script printing the first ten Fibonacci numbers \u2013 then run it to check the output",
+        "I need a Python script fib.py that prints the first 10 Fibonacci numbers. Execute it afterwards to verify correctness.",
+        "Could you create fib.py to display the first 10 Fibonacci numbers in Python, and then run it to make sure it works?",
+        "Draft a script called fib.py in Python that outputs the first ten Fibonacci numbers, then execute it to validate",
+        "Hey, write me a fib.py that prints the first 10 Fibonacci numbers and run it so we can see it works",
+        "Generate a Python program fib.py which prints the initial 10 Fibonacci numbers, and verify by running it",
+        "Kindly produce a Python script named fib.py to print the first 10 Fibonacci numbers, then execute the script to confirm its output",
+        "Make a file fib.py containing Python code to print the first 10 Fibonacci numbers. Then run it to verify."
+      ]
     },
     {
       "id": "no-bash-for-file-write",
-      "description": "Should NOT use bash (echo/cat/heredoc) to create files — only write_file",
+      "description": "Should NOT use bash (echo/cat/heredoc) to create files \u2014 only write_file",
       "user_prompt": "Create a new file called README.md with a title and description of this project",
       "expected_actions": [
-        { "tool": "write_file", "args_pattern": { "path": "README" } }
+        {
+          "tool": "write_file",
+          "args_pattern": {
+            "path": "README"
+          }
+        }
       ],
-      "match_mode": "subset"
+      "match_mode": "subset",
+      "user_prompts": [
+        "Create a new file called README.md with a title and description of this project",
+        "Make a README.md file that includes a project title and description",
+        "I need a README.md created with a title and a brief description of the project",
+        "Please generate a README.md file containing the project's title and description",
+        "Could you set up a README.md with a title and project description?",
+        "Write a README.md that has a title and describes this project",
+        "Go ahead and create README.md \u2014 it should have a title and a description of the project",
+        "I'd like you to produce a new README.md file featuring a project title and description",
+        "Kindly establish a README.md file incorporating both a title and a description for this project",
+        "Spin up a README.md with a project title and description in it"
+      ]
     },
     {
-      "id": "plan-before-refactor",
-      "description": "Use the plan tool before a large refactoring task",
-      "user_prompt": "I need to refactor this codebase to separate the database layer from the API layer. Use the plan tool to think through the approach before making any changes.",
+      "id": "plan-when-asked",
+      "description": "Call the plan tool when the user asks to plan",
+      "user_prompt": "Plan how to add user authentication to this app.",
       "setup": {
         "files": {
-          "app.py": "import sqlite3\nfrom flask import Flask, jsonify\n\napp = Flask(__name__)\nDB = 'data.db'\n\ndef get_db():\n    return sqlite3.connect(DB)\n\n@app.route('/users')\ndef list_users():\n    db = get_db()\n    users = db.execute('SELECT * FROM users').fetchall()\n    db.close()\n    return jsonify(users)\n\n@app.route('/users/<int:uid>')\ndef get_user(uid):\n    db = get_db()\n    user = db.execute('SELECT * FROM users WHERE id=?', (uid,)).fetchone()\n    db.close()\n    return jsonify(user)\n\nif __name__ == '__main__':\n    app.run(port=8000)\n"
+          "app.py": "from flask import Flask, jsonify\n\napp = Flask(__name__)\n\n@app.route('/users')\ndef list_users():\n    return jsonify([{'id': 1, 'name': 'Alice'}])\n\nif __name__ == '__main__':\n    app.run(port=8000)\n"
         }
       },
-      "expected_actions": [{ "tool": "create_plan" }],
-      "match_mode": "subset"
+      "expected_actions": [
+        {
+          "tool": "plan_agent"
+        }
+      ],
+      "match_mode": "subset",
+      "user_prompts": [
+        "Plan how to add user authentication to this app.",
+        "Make a plan for adding pagination to the API endpoints.",
+        "Plan out how to add error handling to this application.",
+        "I need a plan for adding logging to this codebase.",
+        "Plan the approach for adding unit tests to this app.",
+        "How would you approach adding user authentication to this app? Lay out a plan.",
+        "I'd like you to outline a strategy for implementing user authentication in this application.",
+        "Could you come up with a plan for integrating user authentication into this app?",
+        "Think through the steps needed to add user auth to this app and present a plan.",
+        "Draft a plan for incorporating user authentication functionality into this application."
+      ]
     },
     {
       "id": "edit-not-rewrite",
@@ -139,10 +328,32 @@
         }
       },
       "expected_actions": [
-        { "tool": "read_file", "args": { "path": "utils.py" } },
-        { "tool": "edit_file", "args": { "path": "utils.py" } }
+        {
+          "tool": "read_file",
+          "args": {
+            "path": "utils.py"
+          }
+        },
+        {
+          "tool": "edit_file",
+          "args": {
+            "path": "utils.py"
+          }
+        }
       ],
-      "match_mode": "ordered_subset"
+      "match_mode": "ordered_subset",
+      "user_prompts": [
+        "Add a docstring to the process_data function in utils.py",
+        "Please add a docstring to the process_data function in utils.py",
+        "Could you write a docstring for process_data in utils.py?",
+        "Insert a docstring into the process_data function found in utils.py",
+        "I need a docstring added to process_data in utils.py",
+        "Put a docstring on the process_data function in utils.py",
+        "The process_data function in utils.py is missing a docstring \u2014 please add one",
+        "Would you mind adding a docstring to process_data in utils.py?",
+        "In utils.py, the process_data function needs a docstring",
+        "Add documentation via a docstring to the process_data function within utils.py"
+      ]
     },
     {
       "id": "bash-run-tests",
@@ -154,45 +365,130 @@
         }
       },
       "expected_actions": [
-        { "tool": "bash", "args_pattern": { "command": "pytest|python.*test" } }
+        {
+          "tool": "bash",
+          "args_pattern": {
+            "command": "pytest|python.*test"
+          }
+        }
       ],
-      "match_mode": "subset"
+      "match_mode": "subset",
+      "user_prompts": [
+        "Run the tests",
+        "Execute the test suite",
+        "Please go ahead and run the tests",
+        "Could you run the tests for me?",
+        "I need the tests to be run",
+        "Kick off the tests",
+        "Let's run the tests",
+        "Fire up the tests",
+        "Go ahead and execute the tests",
+        "I'd like you to run the tests"
+      ]
     },
     {
       "id": "web-fetch-url",
       "description": "Use web_fetch when asked to retrieve content from a URL",
       "user_prompt": "Fetch the contents of https://example.com and summarize what's on the page",
       "expected_actions": [
-        { "tool": "web_fetch", "args_pattern": { "url": "example\\.com" } }
+        {
+          "tool": "web_fetch",
+          "args_pattern": {
+            "url": "example\\.com"
+          }
+        }
       ],
-      "match_mode": "subset"
+      "match_mode": "subset",
+      "user_prompts": [
+        "Fetch the contents of https://example.com and summarize what's on the page",
+        "Go to https://example.com and give me a summary of what you find there",
+        "Could you pull up https://example.com and tell me what the page is about?",
+        "Retrieve the content from https://example.com, then provide a summary of it",
+        "I need you to grab https://example.com and summarize its contents for me",
+        "Please access https://example.com and give me an overview of the page",
+        "What's on https://example.com? Fetch it and summarize for me.",
+        "Download the page at https://example.com and provide a brief summary",
+        "I'd like a summary of whatever is at https://example.com \u2014 please fetch it first",
+        "Hit https://example.com and let me know what's there in summary form"
+      ]
     },
     {
       "id": "man-page-lookup",
       "description": "Use man tool to look up command documentation",
       "user_prompt": "Look up the man page for tar and tell me what the --xattrs flag does",
       "expected_actions": [
-        { "tool": "man", "args_pattern": { "page": "tar" } }
+        {
+          "tool": "man",
+          "args_pattern": {
+            "page": "tar"
+          }
+        }
       ],
-      "match_mode": "subset"
+      "match_mode": "subset",
+      "user_prompts": [
+        "Look up the man page for tar and tell me what the --xattrs flag does",
+        "What does the --xattrs flag do in tar? Check the man page for me.",
+        "Could you pull up the man page for tar and explain the --xattrs option?",
+        "I need to know what --xattrs does in tar \u2014 can you check the man page?",
+        "Check tar's man page and let me know the purpose of the --xattrs flag.",
+        "Please consult the tar man page and describe what the --xattrs flag is for.",
+        "Hey, look at the tar man page real quick \u2014 what's --xattrs do?",
+        "I'd like you to read the tar man page and summarize the --xattrs option for me.",
+        "Would you mind checking the man page for tar to find out what --xattrs means?",
+        "Look into the tar manual and explain the --xattrs flag to me."
+      ]
     },
     {
       "id": "math-calculation",
       "description": "Use the math tool for precise calculations, not bash or mental math",
       "user_prompt": "What is 2^64 - 1? Use the math tool to calculate it precisely.",
       "expected_actions": [
-        { "tool": "math", "args_pattern": { "code": "2.*64" } }
+        {
+          "tool": "math",
+          "args_pattern": {
+            "code": "2.*64"
+          }
+        }
       ],
-      "match_mode": "subset"
+      "match_mode": "subset",
+      "user_prompts": [
+        "What is 2^64 - 1? Use the math tool to calculate it precisely.",
+        "Calculate 2^64 - 1 for me using the math tool, please.",
+        "I need the exact value of 2^64 - 1. Please use the math tool.",
+        "Could you use the math tool to compute 2^64 minus 1 precisely?",
+        "Use the math tool to tell me what 2^64 - 1 equals.",
+        "I'm curious: what's 2^64 - 1? Compute it with the math tool.",
+        "Please precisely determine 2^64 - 1 via the math tool.",
+        "Mind using the math tool to figure out 2^64 - 1 exactly?",
+        "I'd like to know the precise result of 2^64 - 1 \u2014 use the math tool for this.",
+        "Leverage the math tool to give me an exact answer for 2^64 - 1."
+      ]
     },
     {
       "id": "web-search-query",
       "description": "Use web_search for general knowledge lookups, not web_fetch",
       "user_prompt": "Search the web for the current population of Tokyo",
       "expected_actions": [
-        { "tool": "web_search", "args_pattern": { "query": "Tokyo" } }
+        {
+          "tool": "web_search",
+          "args_pattern": {
+            "query": "Tokyo"
+          }
+        }
       ],
-      "match_mode": "subset"
+      "match_mode": "subset",
+      "user_prompts": [
+        "Search the web for the current population of Tokyo",
+        "What's Tokyo's current population? Look it up on the web.",
+        "Could you do a web search to find out how many people currently live in Tokyo?",
+        "Please search online for Tokyo's present-day population.",
+        "I need you to look up the current population of Tokyo on the web.",
+        "Find me Tokyo's current population via a web search.",
+        "Web search: what is the current population of Tokyo?",
+        "I'd like to know Tokyo's current population\u2014can you search the web for that?",
+        "Look up how many people live in Tokyo right now using a web search.",
+        "Do a web search for the population of Tokyo as of now."
+      ]
     }
   ]
 }

tests/conftest.py

@@ -1,11 +1,23 @@
+from __future__ import annotations
+
+import os
 from unittest.mock import MagicMock
 
 import pytest
 
 
+def pytest_addoption(parser: pytest.Parser) -> None:
+    parser.addoption(
+        "--storage-backend",
+        default="sqlite",
+        choices=["sqlite", "postgresql"],
+        help="Storage backend for integration tests (default: sqlite)",
+    )
+
+
 @pytest.fixture
 def tmp_db(tmp_path):
-    """Provide a temporary SQLite storage backend."""
+    """Provide a temporary SQLite storage backend (singleton registry)."""
     from turnstone.core.storage import init_storage, reset_storage
 
     db_path = str(tmp_path / "test.db")
@@ -15,6 +27,68 @@ def tmp_db(tmp_path):
     reset_storage()
 
 
+@pytest.fixture
+def storage_backend(request, tmp_path):
+    """Shared storage backend fixture — respects --storage-backend flag.
+
+    Returns a StorageBackend instance (SQLite or PostgreSQL).
+    Tests that use this fixture run against whichever backend CI selects.
+    """
+    from turnstone.core.storage import init_storage, reset_storage
+
+    backend_type = request.config.getoption("--storage-backend")
+    reset_storage()
+
+    if backend_type == "postgresql":
+        pg_url = os.environ.get(
+            "TURNSTONE_TEST_PG_URL",
+            "postgresql+psycopg://postgres:postgres@localhost:5432/turnstone_test",
+        )
+        backend = init_storage("postgresql", url=pg_url, run_migrations=False)
+        yield backend
+        # Truncate all tables between tests — faster than DELETE and resets
+        # autoincrement sequences.  CASCADE handles any future FK constraints.
+        # NOTE: accesses backend._engine (SQLAlchemy internal) — both SQLite
+        # and PostgreSQL backends expose this.  If a non-SQLAlchemy backend is
+        # ever added, this cleanup will need a protocol-level hook.
+        try:
+            import sqlalchemy as sa
+
+            from turnstone.core.storage._schema import metadata as db_metadata
+
+            with backend._engine.connect() as conn:
+                table_names = ", ".join(t.name for t in reversed(db_metadata.sorted_tables))
+                conn.execute(sa.text(f"TRUNCATE {table_names} RESTART IDENTITY CASCADE"))
+                conn.commit()
+        except Exception:
+            pass  # best-effort cleanup; reset_storage disposes engine
+        finally:
+            reset_storage()
+    else:
+        db_path = str(tmp_path / "test.db")
+        backend = init_storage("sqlite", path=db_path, run_migrations=False)
+        yield backend
+        reset_storage()
+
+
+@pytest.fixture
+def backend(storage_backend):
+    """Alias for storage_backend — used by test_storage_sqlite.py etc."""
+    return storage_backend
+
+
+@pytest.fixture
+def db(storage_backend):
+    """Alias for storage_backend — used by domain-specific storage tests."""
+    return storage_backend
+
+
+@pytest.fixture
+def storage(storage_backend):
+    """Alias for storage_backend — used by services/skill resource tests."""
+    return storage_backend
+
+
 @pytest.fixture
 def mock_openai_client():
     """Return a minimal mock OpenAI client."""

tests/test_api_versioning.py

@@ -19,6 +19,7 @@ class TestServerVersioning:
 
         mock_mgr = MagicMock()
         mock_mgr.list_all.return_value = []
+        mock_mgr.max_workstreams = 10
         app = create_app(
             workstreams=mock_mgr,
             global_queue=queue.Queue(),

tests/test_auth.py

@@ -783,6 +783,7 @@ class TestServerAuth:
         mock_ws.session = mock_session
         mock_mgr = MagicMock()
         mock_mgr.list_all.return_value = [mock_ws]
+        mock_mgr.max_workstreams = 10
 
         app = srv_mod.create_app(
             workstreams=mock_mgr,
@@ -1001,6 +1002,7 @@ class TestServerLogin:
         mock_ws.session = mock_session
         mock_mgr = MagicMock()
         mock_mgr.list_all.return_value = [mock_ws]
+        mock_mgr.max_workstreams = 10
 
         app = srv_mod.create_app(
             workstreams=mock_mgr,
@@ -1235,6 +1237,52 @@ class TestJWTAudienceIssuer:
         result = validate_jwt(token, self.SECRET, audience="")
         assert result is not None
 
+    def test_create_jwt_expiry_seconds(self):
+        import jwt as pyjwt
+
+        from turnstone.core.auth import create_jwt
+
+        token = create_jwt("user1", frozenset({"read"}), "test", self.SECRET, expiry_seconds=300)
+        payload = pyjwt.decode(
+            token, self.SECRET, algorithms=["HS256"], options={"verify_aud": False}
+        )
+        assert payload["exp"] - payload["iat"] == 300
+
+    def test_create_jwt_expiry_seconds_overrides_hours(self):
+        import jwt as pyjwt
+
+        from turnstone.core.auth import create_jwt
+
+        token = create_jwt(
+            "user1",
+            frozenset({"read"}),
+            "test",
+            self.SECRET,
+            expiry_hours=24,
+            expiry_seconds=60,
+        )
+        payload = pyjwt.decode(
+            token, self.SECRET, algorithms=["HS256"], options={"verify_aud": False}
+        )
+        # expiry_seconds takes precedence over expiry_hours
+        assert payload["exp"] - payload["iat"] == 60
+
+    def test_create_jwt_expiry_seconds_rejects_zero(self):
+        import pytest
+
+        from turnstone.core.auth import create_jwt
+
+        with pytest.raises(ValueError, match="expiry_seconds must be positive"):
+            create_jwt("user1", frozenset({"read"}), "test", self.SECRET, expiry_seconds=0)
+
+    def test_create_jwt_expiry_seconds_rejects_negative(self):
+        import pytest
+
+        from turnstone.core.auth import create_jwt
+
+        with pytest.raises(ValueError, match="expiry_seconds must be positive"):
+            create_jwt("user1", frozenset({"read"}), "test", self.SECRET, expiry_seconds=-1)
+
 
 class TestServiceTokenManager:
     SECRET = "test-secret-that-is-at-least-32-chars"
@@ -1369,8 +1417,11 @@ class TestCorsConfigurable:
 
         import turnstone.server as srv_mod
 
+        mgr = MagicMock()
+        mgr.list_all.return_value = []
+        mgr.max_workstreams = 10
         app = srv_mod.create_app(
-            workstreams=MagicMock(),
+            workstreams=mgr,
             global_queue=queue.Queue(),
             global_listeners=[],
             global_listeners_lock=threading.Lock(),
@@ -1388,8 +1439,11 @@ class TestCorsConfigurable:
 
         import turnstone.server as srv_mod
 
+        mgr = MagicMock()
+        mgr.list_all.return_value = []
+        mgr.max_workstreams = 10
         app = srv_mod.create_app(
-            workstreams=MagicMock(),
+            workstreams=mgr,
             global_queue=queue.Queue(),
             global_listeners=[],
             global_listeners_lock=threading.Lock(),

tests/test_auth_identity.py

@@ -130,7 +130,7 @@ class TestParseScopes:
 
 
 class TestJWT:
-    SECRET = "test-secret-key-for-jwt"
+    SECRET = "test-secret-key-for-jwt-min-32b!"
 
     def test_round_trip(self):
         scopes = frozenset({"read", "write"})
@@ -155,7 +155,7 @@ class TestJWT:
 
     def test_invalid_signature(self):
         token = create_jwt("user1", frozenset({"read"}), "db", self.SECRET)
-        assert validate_jwt(token, "wrong-secret") is None
+        assert validate_jwt(token, "wrong-secret-key-for-jwt-min-32b") is None
 
     def test_malformed_token(self):
         assert validate_jwt("not.a.jwt", self.SECRET) is None
@@ -217,7 +217,7 @@ class TestAuthenticateToken:
         assert result.scopes == frozenset({"read", "write", "approve"})
 
     def test_jwt_token(self):
-        secret = "test-secret"
+        secret = "test-secret-key-for-jwt-min-32b!"
         jwt_tok = create_jwt("user1", frozenset({"read", "write"}), "db", secret)
         cfg = AuthConfig(enabled=True)
         result = _authenticate_token(jwt_tok, cfg, jwt_secret=secret)
@@ -304,7 +304,7 @@ class TestCheckRequestScopes:
         assert result.has_scope("approve")
 
     def test_jwt_with_scopes(self):
-        secret = "test"
+        secret = "test-secret-key-for-jwt-min-32b!"
         jwt_tok = create_jwt("u1", frozenset({"read", "write"}), "db", secret)
         cfg = AuthConfig(enabled=True)
         allowed, status, msg, result = check_request(
@@ -319,7 +319,7 @@ class TestCheckRequestScopes:
         assert result.user_id == "u1"
 
     def test_jwt_insufficient_scope(self):
-        secret = "test"
+        secret = "test-secret-key-for-jwt-min-32b!"
         jwt_tok = create_jwt("u1", frozenset({"read"}), "db", secret)
         cfg = AuthConfig(enabled=True)
         allowed, status, msg, _ = check_request(

tests/test_bridge_stress.py

@@ -0,0 +1,357 @@
+"""Stress tests for bridge.py threading — race conditions in approval,
+plan review, and workstream lifecycle.
+
+Each scenario is run many times (ITERATIONS) with threading.Barrier to
+maximize timing overlap.  Uses mock broker (no Redis) and no HTTP calls.
+
+Races tested:
+1. Duplicate approval on SSE reconnect (TOCTOU in _pending_approvals)
+2. Duplicate plan review on SSE reconnect (TOCTOU in _pending_plan_reviews)
+3. approve_set stale reference escape during concurrent update
+4. _running flag visibility across threads on shutdown
+5. Approval thread exits within bounded time after timeout
+6. Concurrent approval + workstream close leaves no orphaned state
+"""
+
+from __future__ import annotations
+
+import threading
+import time
+from collections import Counter
+from unittest.mock import MagicMock, patch
+
+from turnstone.mq.bridge import Bridge
+
+ITERATIONS = 100
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _make_bridge(**overrides) -> Bridge:
+    """Create a Bridge with a mock broker (no Redis or HTTP)."""
+    broker = MagicMock()
+    defaults = dict(
+        server_url="http://localhost:8080",
+        broker=broker,
+        node_id="test-node",
+        approval_timeout=1,
+    )
+    defaults.update(overrides)
+    bridge = Bridge(**defaults)
+    # Replace real httpx client with a mock so daemon threads spawned by
+    # _handle_approval / _handle_plan_review don't make real HTTP calls
+    # after the test's patch context exits.
+    bridge._http.close()
+    bridge._http = MagicMock()
+    return bridge
+
+
+def _approval_items(tool_name: str = "bash") -> list[dict]:
+    return [{"func_name": tool_name, "needs_approval": True, "approval_label": tool_name}]
+
+
+def _wait_pending_resolved(bridge: Bridge, key: str, attr: str, deadline_s: float = 3.0) -> bool:
+    """Poll until the pending entry is resolved (tombstone) or absent."""
+    deadline = time.monotonic() + deadline_s
+    while time.monotonic() < deadline:
+        with bridge._lock:
+            entries = getattr(bridge, attr)
+            if key not in entries:
+                return True
+            _, resolved_at = entries[key]
+            if resolved_at > 0:
+                return True
+        time.sleep(0.01)
+    return False
+
+
+# ---------------------------------------------------------------------------
+# Race 1: Duplicate approval on SSE reconnect
+# ---------------------------------------------------------------------------
+
+
+class TestDuplicateApproval:
+    """Two threads call _handle_approval for the same ws_id simultaneously.
+    Only one should create a pending entry; the other should be skipped."""
+
+    def test_no_duplicate_approvals(self):
+        sent_count = Counter()
+
+        for _ in range(ITERATIONS):
+            bridge = _make_bridge()
+            bridge._broker.pop_response.return_value = '{"type": "approve", "approved": true}'
+            barrier = threading.Barrier(2, timeout=5)
+
+            def _call_approval(bridge=bridge, barrier=barrier):
+                barrier.wait()
+                bridge._handle_approval("ws-1", {"items": _approval_items()})
+
+            t1 = threading.Thread(target=_call_approval)
+            t2 = threading.Thread(target=_call_approval)
+            with (
+                patch.object(bridge, "_api_approve") as mock_approve,
+                patch.object(bridge, "_publish_ws"),
+            ):
+                t1.start()
+                t2.start()
+                t1.join(timeout=5)
+                t2.join(timeout=5)
+                assert not t1.is_alive(), "Thread 1 hung"
+                assert not t2.is_alive(), "Thread 2 hung"
+
+                # Wait for spawned _wait_approval threads to resolve
+                _wait_pending_resolved(bridge, "ws-1", "_pending_approvals")
+
+                sent_count[mock_approve.call_count] += 1
+
+        # At most 1 approval should be forwarded per iteration
+        assert sent_count.get(2, 0) == 0, (
+            f"Duplicate approvals sent in {sent_count[2]}/{ITERATIONS} iterations"
+        )
+
+
+# ---------------------------------------------------------------------------
+# Race 2: Duplicate plan review on SSE reconnect
+# ---------------------------------------------------------------------------
+
+
+class TestDuplicatePlanReview:
+    """Two threads call _handle_plan_review simultaneously.
+    Only one should create a pending entry."""
+
+    def test_no_duplicate_plan_reviews(self):
+        sent_count = Counter()
+
+        for _ in range(ITERATIONS):
+            bridge = _make_bridge()
+            bridge._broker.pop_response.return_value = (
+                '{"type": "plan_feedback", "feedback": "looks good"}'
+            )
+            barrier = threading.Barrier(2, timeout=5)
+
+            def _call_plan(bridge=bridge, barrier=barrier):
+                barrier.wait()
+                bridge._handle_plan_review("ws-1", {"content": "plan text"})
+
+            t1 = threading.Thread(target=_call_plan)
+            t2 = threading.Thread(target=_call_plan)
+            with patch.object(bridge, "_publish_ws"), patch.object(bridge._http, "post"):
+                t1.start()
+                t2.start()
+                t1.join(timeout=5)
+                t2.join(timeout=5)
+                assert not t1.is_alive(), "Thread 1 hung"
+                assert not t2.is_alive(), "Thread 2 hung"
+
+                # Wait for spawned _wait_plan threads to resolve
+                _wait_pending_resolved(bridge, "ws-1", "_pending_plan_reviews")
+
+                sent_count[bridge._http.post.call_count] += 1
+
+        assert sent_count.get(2, 0) == 0, (
+            f"Duplicate plan reviews sent in {sent_count[2]}/{ITERATIONS} iterations"
+        )
+
+
+# ---------------------------------------------------------------------------
+# Race 3: approve_set stale reference during concurrent update
+# ---------------------------------------------------------------------------
+
+
+class TestApproveSetConsistency:
+    """One thread reads approve_set for auto-approve check while another
+    updates it via _wait_approval 'always' path.  The auto-approve
+    decision should be consistent (either all-approved or not)."""
+
+    def test_approve_set_never_partially_visible(self):
+        for _ in range(ITERATIONS):
+            bridge = _make_bridge()
+            with bridge._lock:
+                bridge._ws_approve_tools["ws-1"] = {"read_file", "search"}
+
+            barrier = threading.Barrier(2, timeout=5)
+            results = []
+
+            def _reader(bridge=bridge, barrier=barrier, results=results):
+                barrier.wait()
+                with bridge._lock:
+                    snap = bridge._ws_approve_tools.get("ws-1", set()).copy()
+                results.append(snap)
+
+            def _writer(bridge=bridge, barrier=barrier):
+                barrier.wait()
+                with bridge._lock:
+                    existing = bridge._ws_approve_tools.get("ws-1", set())
+                    bridge._ws_approve_tools["ws-1"] = existing | {"bash", "write_file"}
+
+            t1 = threading.Thread(target=_reader)
+            t2 = threading.Thread(target=_writer)
+            t1.start()
+            t2.start()
+            t1.join(timeout=5)
+            t2.join(timeout=5)
+            assert not t1.is_alive(), "Reader hung"
+            assert not t2.is_alive(), "Writer hung"
+
+            snap = results[0]
+            assert snap in (
+                {"read_file", "search"},
+                {"read_file", "search", "bash", "write_file"},
+            ), f"Partial set observed: {snap}"
+
+
+# ---------------------------------------------------------------------------
+# Race 4: _running flag visibility across threads
+# ---------------------------------------------------------------------------
+
+
+class TestRunningFlagVisibility:
+    """All threads reading _running should see False within a bounded time
+    after the main thread sets it."""
+
+    def test_all_threads_observe_shutdown(self):
+        bridge = _make_bridge()
+        observed_false = threading.Event()
+        threads_running = []
+
+        def _spin_checker():
+            while bridge._running:
+                time.sleep(0.001)
+            observed_false.set()
+
+        for _ in range(5):
+            t = threading.Thread(target=_spin_checker, daemon=True)
+            threads_running.append(t)
+            t.start()
+
+        time.sleep(0.01)
+        bridge._running = False
+
+        for t in threads_running:
+            t.join(timeout=1)
+            assert not t.is_alive(), "Thread did not observe _running=False"
+
+        assert observed_false.is_set()
+
+
+# ---------------------------------------------------------------------------
+# Race 5: Approval thread exits within bounded time
+# ---------------------------------------------------------------------------
+
+
+class TestApprovalThreadTimeout:
+    """An approval thread blocked on pop_response should exit within the
+    configured approval_timeout, not hang indefinitely."""
+
+    def test_approval_thread_exits_within_timeout(self):
+        for _ in range(10):
+            bridge = _make_bridge(approval_timeout=0.5)
+
+            def _slow_pop(queue_name, timeout=300):
+                time.sleep(min(timeout, 0.5))
+                return None
+
+            bridge._broker.pop_response.side_effect = _slow_pop
+
+            with patch.object(bridge, "_publish_ws"), patch.object(bridge, "_api_approve"):
+                bridge._handle_approval("ws-1", {"items": _approval_items()})
+
+            # The pending entry should be resolved within the timeout
+            resolved = _wait_pending_resolved(bridge, "ws-1", "_pending_approvals", deadline_s=3.0)
+            assert resolved, "Approval thread did not exit within expected timeout"
+
+
+# ---------------------------------------------------------------------------
+# Race 6: Concurrent approval + workstream close
+# ---------------------------------------------------------------------------
+
+
+class TestApprovalDuringClose:
+    """An approval arriving at the exact same time as a ws_closed event
+    should not leave orphaned state."""
+
+    def test_no_orphaned_pending_after_close(self):
+        for _ in range(ITERATIONS):
+            bridge = _make_bridge(approval_timeout=0.1)
+            bridge._broker.pop_response.return_value = None  # timeout
+
+            barrier = threading.Barrier(2, timeout=5)
+
+            def _send_approval(bridge=bridge, barrier=barrier):
+                barrier.wait()
+                with patch.object(bridge, "_publish_ws"), patch.object(bridge, "_api_approve"):
+                    bridge._handle_approval("ws-1", {"items": _approval_items()})
+
+            def _close_ws(bridge=bridge, barrier=barrier):
+                barrier.wait()
+                with (
+                    patch.object(bridge, "_publish_global"),
+                    patch.object(bridge, "_publish_cluster"),
+                ):
+                    bridge._handle_global_event({"type": "ws_closed", "ws_id": "ws-1"})
+
+            t1 = threading.Thread(target=_send_approval)
+            t2 = threading.Thread(target=_close_ws)
+            t1.start()
+            t2.start()
+            t1.join(timeout=5)
+            t2.join(timeout=5)
+            assert not t1.is_alive(), "Approval thread hung"
+            assert not t2.is_alive(), "Close thread hung"
+
+            # Wait for spawned _wait_approval thread to resolve (if close
+            # didn't remove the entry first)
+            resolved = _wait_pending_resolved(bridge, "ws-1", "_pending_approvals")
+            assert resolved, "Orphaned pending approval"
+
+
+# ---------------------------------------------------------------------------
+# Race 7: Plan review refinement loop (tombstone → cleanup → re-entry)
+# ---------------------------------------------------------------------------
+
+
+class TestPlanReviewRefinementLoop:
+    """After a plan review is resolved, a ws_state event should clean up the
+    tombstone so the refinement-loop plan_review event is handled correctly."""
+
+    def test_refinement_loop_allows_reentry(self):
+        for _ in range(ITERATIONS):
+            bridge = _make_bridge()
+            bridge._broker.pop_response.return_value = (
+                '{"type": "plan_feedback", "feedback": "refine this"}'
+            )
+
+            # Step 1: first plan review — creates pending entry, resolves it
+            with patch.object(bridge, "_publish_ws"), patch.object(bridge._http, "post"):
+                bridge._handle_plan_review("ws-1", {"content": "plan v1"})
+
+            _wait_pending_resolved(bridge, "ws-1", "_pending_plan_reviews")
+
+            # Verify tombstone is present (resolved_at > 0)
+            with bridge._lock:
+                assert "ws-1" in bridge._pending_plan_reviews
+                assert bridge._pending_plan_reviews["ws-1"][1] > 0
+
+            # Step 2: ws_state event cleans up the resolved tombstone
+            with (
+                patch.object(bridge, "_publish_ws"),
+                patch.object(bridge, "_publish_global"),
+                patch.object(bridge, "_publish_cluster"),
+            ):
+                bridge._handle_global_event(
+                    {"type": "ws_state", "ws_id": "ws-1", "state": "working"}
+                )
+
+            with bridge._lock:
+                assert "ws-1" not in bridge._pending_plan_reviews
+
+            # Step 3: refinement plan_review arrives — should create new entry
+            with patch.object(bridge, "_publish_ws"), patch.object(bridge._http, "post"):
+                bridge._handle_plan_review("ws-1", {"content": "plan v2"})
+
+            _wait_pending_resolved(bridge, "ws-1", "_pending_plan_reviews")
+
+            with bridge._lock:
+                assert "ws-1" in bridge._pending_plan_reviews

tests/test_cancel.py

@@ -1,5 +1,6 @@
 """Tests for generation cancellation (cooperative cancel via threading.Event)."""
 
+import contextlib
 import threading
 import time
 from dataclasses import dataclass, field
@@ -7,7 +8,7 @@ from unittest.mock import MagicMock, patch
 
 import pytest
 
-from turnstone.core.session import ChatSession, GenerationCancelled
+from turnstone.core.session import ChatSession, GenerationCancelled, _CancelRef
 
 
 class NullUI:
@@ -36,7 +37,7 @@ class NullUI:
     def approve_tools(self, items):
         return True, None
 
-    def on_tool_result(self, call_id, name, output):
+    def on_tool_result(self, call_id, name, output, **kwargs):
         pass
 
     def on_tool_output_chunk(self, call_id, chunk):
@@ -179,7 +180,7 @@ class TestCancelDuringToolExecution:
     """Cancel while tools are being executed."""
 
     def test_rollback_incomplete_tool_results(self, tmp_db):
-        """When cancelled during tool execution, incomplete results are rolled back."""
+        """When cancelled during tool execution, synthesized results replace missing tool outputs."""
         ui = NullUI()
         session = _make_session(ui=ui)
 
@@ -235,13 +236,15 @@ class TestCancelDuringToolExecution:
 
         # Session should be idle
         assert ui.states[-1] == "idle"
-        # No tool result messages should remain (rolled back)
-        roles = [m["role"] for m in session.messages]
-        assert "tool" not in roles
-        # The assistant message with tool_calls should also be rolled back
-        for m in session.messages:
-            if m["role"] == "assistant":
-                assert "tool_calls" not in m or not m["tool_calls"]
+        # Cancelled tool calls should have synthesized results
+        tool_msgs = [m for m in session.messages if m["role"] == "tool"]
+        assert len(tool_msgs) == 1
+        assert tool_msgs[0]["tool_call_id"] == "tc_1"
+        assert "Cancelled by user" in tool_msgs[0]["content"]
+        assert tool_msgs[0].get("is_error") is True
+        # The assistant message with tool_calls should still be present
+        assistant_msgs = [m for m in session.messages if m.get("tool_calls")]
+        assert len(assistant_msgs) == 1
 
 
 class TestCancelWhenIdle:
@@ -407,3 +410,402 @@ class TestStreamFlushBeforeToolCalls:
         stream_end_idx = next(i for i, e in enumerate(events) if e[0] == "stream_end")
         late_content = [e for e in events[stream_end_idx + 1 :] if e[0] == "content"]
         assert late_content == [], f"Content after stream_end: {late_content}"
+
+
+class TestStreamAbort:
+    """Tests for cancel() closing the underlying SDK stream."""
+
+    def test_cancel_closes_cancel_stream(self, tmp_db):
+        """cancel() calls .close() on the stored SDK stream handle."""
+        session = _make_session()
+        mock_stream = MagicMock()
+        session._cancel_stream = mock_stream
+        session.cancel()
+        mock_stream.close.assert_called_once()
+        assert session._cancel_event.is_set()
+
+    def test_cancel_without_stream_is_safe(self, tmp_db):
+        """cancel() with no active stream just sets the event."""
+        session = _make_session()
+        assert session._cancel_stream is None
+        session.cancel()  # Should not raise
+        assert session._cancel_event.is_set()
+
+    def test_cancel_stream_close_error_suppressed(self, tmp_db):
+        """Errors from stream.close() are suppressed."""
+        session = _make_session()
+        mock_stream = MagicMock()
+        mock_stream.close.side_effect = RuntimeError("already closed")
+        session._cancel_stream = mock_stream
+        session.cancel()  # Should not raise
+        assert session._cancel_event.is_set()
+
+    def test_cancel_ref_populated_after_first_chunk(self, tmp_db):
+        """_cancel_ref is populated by the provider after the first chunk
+        arrives (lazy generator evaluation)."""
+        ui = NullUI()
+        session = _make_session(ui=ui)
+
+        @dataclass
+        class FakeChunk:
+            content_delta: str = ""
+            reasoning_delta: str = ""
+            tool_call_deltas: list = field(default_factory=list)
+            usage: None = None
+            finish_reason: str = "stop"
+            info_delta: str = ""
+            provider_blocks: list = field(default_factory=list)
+
+        sdk_stream = MagicMock()
+
+        def fake_provider_stream():
+            # Simulate provider appending to cancel_ref before first yield
+            session._cancel_ref.append(sdk_stream)
+            yield FakeChunk(content_delta="hi", finish_reason="stop")
+
+        with (
+            patch.object(
+                session,
+                "_create_stream_with_retry",
+                return_value=fake_provider_stream(),
+            ),
+            patch.object(session, "_full_messages", return_value=[]),
+        ):
+            session.send("test")
+
+        # After stream completes, cancel_stream should be cleared
+        assert session._cancel_stream is None
+        assert len(session._cancel_ref) == 0
+
+    def test_transport_error_during_cancel_becomes_generation_cancelled(self, tmp_db):
+        """When cancel() closes the stream, the resulting transport error
+        is converted to GenerationCancelled."""
+        ui = NullUI()
+        session = _make_session(ui=ui)
+
+        @dataclass
+        class FakeChunk:
+            content_delta: str = ""
+            reasoning_delta: str = ""
+            tool_call_deltas: list = field(default_factory=list)
+            usage: None = None
+            finish_reason: str = ""
+            info_delta: str = ""
+            provider_blocks: list = field(default_factory=list)
+
+        def stream_that_errors():
+            yield FakeChunk(content_delta="Hello")
+            session._cancel_event.set()
+            raise ConnectionError("stream closed")
+
+        with (
+            patch.object(
+                session,
+                "_create_stream_with_retry",
+                return_value=stream_that_errors(),
+            ),
+            patch.object(session, "_full_messages", return_value=[]),
+        ):
+            session.send("test")
+
+        # Should complete as cancelled, not error
+        assert "idle" in ui.states
+        assert any("cancelled" in i.lower() for i in ui.infos)
+        # Partial content preserved
+        assistant_msgs = [m for m in session.messages if m["role"] == "assistant"]
+        assert len(assistant_msgs) == 1
+        assert assistant_msgs[0]["content"] == "Hello"
+
+    def test_non_cancel_exception_not_swallowed(self, tmp_db):
+        """Exceptions during streaming that aren't caused by cancel
+        should propagate normally."""
+        ui = NullUI()
+        session = _make_session(ui=ui)
+
+        @dataclass
+        class FakeChunk:
+            content_delta: str = ""
+            reasoning_delta: str = ""
+            tool_call_deltas: list = field(default_factory=list)
+            usage: None = None
+            finish_reason: str = ""
+            info_delta: str = ""
+            provider_blocks: list = field(default_factory=list)
+
+        def stream_that_errors():
+            yield FakeChunk(content_delta="Hello")
+            raise ValueError("unexpected error")
+
+        with (
+            patch.object(
+                session,
+                "_create_stream_with_retry",
+                return_value=stream_that_errors(),
+            ),
+            patch.object(session, "_full_messages", return_value=[]),
+            pytest.raises(ValueError, match="unexpected error"),
+        ):
+            session.send("test")
+
+    def test_check_cancelled_between_retries(self, tmp_db):
+        """_try_stream checks for cancellation between retry attempts."""
+        session = _make_session()
+        session.cancel()
+
+        with pytest.raises(GenerationCancelled):
+            session._try_stream(
+                client=MagicMock(),
+                model="test",
+                msgs=[],
+            )
+
+
+class TestCancelRef:
+    """Tests for the _CancelRef list proxy."""
+
+    def test_append_sets_cancel_stream(self, tmp_db):
+        """Appending a stream handle to _CancelRef sets _cancel_stream eagerly."""
+        session = _make_session()
+        mock_stream = MagicMock()
+        assert session._cancel_stream is None
+
+        session._cancel_ref.append(mock_stream)
+
+        assert session._cancel_stream is mock_stream
+
+    def test_append_closes_stream_when_already_cancelled(self, tmp_db):
+        """If cancel is already set when a stream is appended, it is closed immediately."""
+        session = _make_session()
+        session.cancel()  # Set cancel event before stream is created
+
+        mock_stream = MagicMock()
+        session._cancel_ref.append(mock_stream)
+
+        mock_stream.close.assert_called_once()
+
+    def test_append_does_not_close_stream_when_not_cancelled(self, tmp_db):
+        """Stream is not closed if cancel hasn't been requested."""
+        session = _make_session()
+        mock_stream = MagicMock()
+
+        session._cancel_ref.append(mock_stream)
+
+        mock_stream.close.assert_not_called()
+        assert session._cancel_stream is mock_stream
+
+    def test_append_close_error_suppressed(self, tmp_db):
+        """Errors from stream.close() during eager close are suppressed."""
+        session = _make_session()
+        session.cancel()
+
+        mock_stream = MagicMock()
+        mock_stream.close.side_effect = RuntimeError("already closed")
+
+        session._cancel_ref.append(mock_stream)  # Should not raise
+
+    def test_cancel_ref_is_cancel_ref_instance(self, tmp_db):
+        """ChatSession._cancel_ref is a _CancelRef instance."""
+        session = _make_session()
+        assert isinstance(session._cancel_ref, _CancelRef)
+
+    def test_cancel_ref_cleared_after_stream_ends(self, tmp_db):
+        """_cancel_ref is cleared in the send() finally block after streaming."""
+        ui = NullUI()
+        session = _make_session(ui=ui)
+        mock_stream = MagicMock()
+        session._cancel_ref.append(mock_stream)
+        assert len(session._cancel_ref) == 1
+
+        @dataclass
+        class FakeChunk:
+            content_delta: str = ""
+            reasoning_delta: str = ""
+            tool_call_deltas: list = field(default_factory=list)
+            usage: None = None
+            finish_reason: str = "stop"
+            info_delta: str = ""
+            provider_blocks: list = field(default_factory=list)
+
+        with (
+            patch.object(
+                session,
+                "_create_stream_with_retry",
+                return_value=iter([FakeChunk(content_delta="hi", finish_reason="stop")]),
+            ),
+            patch.object(session, "_full_messages", return_value=[]),
+        ):
+            session.send("test")
+
+        # After send() completes, _cancel_ref is cleared in the finally block
+        assert len(session._cancel_ref) == 0
+
+
+class TestForceCancelGeneration:
+    """Tests for per-generation tracking that prevents orphaned-thread side-effects."""
+
+    def test_check_cancelled_raises_for_orphaned_generation(self, tmp_db):
+        """_check_cancelled raises GenerationCancelled when my_generation is stale."""
+        session = _make_session()
+        session._generation = 2  # Simulate two generations having run
+
+        with pytest.raises(GenerationCancelled):
+            session._check_cancelled(my_generation=1)  # Generation 1 is orphaned
+
+    def test_check_cancelled_ok_for_current_generation(self, tmp_db):
+        """_check_cancelled does not raise when my_generation matches current."""
+        session = _make_session()
+        session._generation = 3
+        session._check_cancelled(my_generation=3)  # Should not raise
+
+    def test_force_cancel_orphaned_thread_does_not_mutate_messages(self, tmp_db):
+        """An abandoned generation (force-cancel) cannot append to session.messages."""
+        ui = NullUI()
+        session = _make_session(ui=ui)
+
+        # We can't trivially test the full threading scenario in a unit test,
+        # so directly verify that _check_cancelled raises when my_generation
+        # is stale, which is what guards _stream_response against orphaned
+        # (force-cancelled) threads continuing to mutate messages.
+        session._generation = 5
+        with pytest.raises(GenerationCancelled):
+            session._check_cancelled(my_generation=4)  # orphaned generation
+
+    def test_new_cancel_event_per_generation_in_send(self, tmp_db):
+        """send() replaces _cancel_event with a fresh Event each generation."""
+        ui = NullUI()
+        session = _make_session(ui=ui)
+
+        @dataclass
+        class FakeChunk:
+            content_delta: str = ""
+            reasoning_delta: str = ""
+            tool_call_deltas: list = field(default_factory=list)
+            usage: None = None
+            finish_reason: str = "stop"
+            info_delta: str = ""
+            provider_blocks: list = field(default_factory=list)
+
+        original_event = session._cancel_event
+
+        with (
+            patch.object(
+                session,
+                "_create_stream_with_retry",
+                return_value=iter([FakeChunk(content_delta="hi", finish_reason="stop")]),
+            ),
+            patch.object(session, "_full_messages", return_value=[]),
+        ):
+            session.send("test")
+
+        # After send() completes, _cancel_event should be a NEW Event
+        # (not the same object as before the call).
+        assert session._cancel_event is not original_event
+        assert not session._cancel_event.is_set()
+
+
+class TestForceCancelThreaded:
+    """Force cancel with actual threads — verifies orphaned thread behavior."""
+
+    def test_force_cancel_orphan_does_not_mutate_messages(self, tmp_db):
+        """After force cancel + new send(), the orphaned thread must not
+        append stale content to session.messages."""
+        ui = NullUI()
+        session = _make_session(ui=ui)
+
+        @dataclass
+        class FakeChunk:
+            content_delta: str = ""
+            reasoning_delta: str = ""
+            tool_call_deltas: list = field(default_factory=list)
+            usage: None = None
+            finish_reason: str = ""
+            info_delta: str = ""
+            provider_blocks: list = field(default_factory=list)
+
+        barrier = threading.Event()
+        old_done = threading.Event()
+
+        def slow_stream():
+            yield FakeChunk(content_delta="Old content")
+            barrier.set()  # signal: first chunk delivered
+            time.sleep(2)  # simulate stuck stream
+            yield FakeChunk(content_delta=" more", finish_reason="stop")
+
+        # Start generation 1 (will get stuck)
+        with (
+            patch.object(session, "_create_stream_with_retry", return_value=slow_stream()),
+            patch.object(session, "_full_messages", return_value=[]),
+        ):
+
+            def run_old():
+                with contextlib.suppress(Exception):
+                    session.send("old message")
+                old_done.set()
+
+            t1 = threading.Thread(target=run_old, daemon=True)
+            t1.start()
+            assert barrier.wait(timeout=5), "stream did not start"
+
+        # Force cancel: simulate what the server does
+        session.cancel()
+        # Increment generation as new send() would
+        session._generation += 1
+        session._cancel_event = threading.Event()
+
+        # Wait for old thread to notice generation mismatch and exit
+        assert old_done.wait(timeout=10), "orphaned thread did not exit"
+
+        # The orphaned thread should NOT have appended its content
+        assistant_msgs = [m for m in session.messages if m["role"] == "assistant"]
+        # May have partial content from before cancel, but NOT the full
+        # "Old content more" that would appear without the generation guard
+        for msg in assistant_msgs:
+            assert "more" not in msg.get("content", "")
+
+    def test_force_cancel_then_new_send_succeeds(self, tmp_db):
+        """A new send() after force cancel works cleanly."""
+        ui = NullUI()
+        session = _make_session(ui=ui)
+
+        @dataclass
+        class FakeChunk:
+            content_delta: str = ""
+            reasoning_delta: str = ""
+            tool_call_deltas: list = field(default_factory=list)
+            usage: None = None
+            finish_reason: str = "stop"
+            info_delta: str = ""
+            provider_blocks: list = field(default_factory=list)
+
+        barrier = threading.Event()
+
+        def stuck_stream():
+            yield FakeChunk(content_delta="stuck")
+            barrier.set()
+            time.sleep(2)
+            yield FakeChunk(content_delta=" end", finish_reason="stop")
+
+        # Start stuck generation
+        with (
+            patch.object(session, "_create_stream_with_retry", return_value=stuck_stream()),
+            patch.object(session, "_full_messages", return_value=[]),
+        ):
+            t = threading.Thread(target=lambda: session.send("old"), daemon=True)
+            t.start()
+            assert barrier.wait(timeout=5), "stream did not start"
+
+        # Force cancel
+        session.cancel()
+
+        # New generation should work
+        fresh_stream = iter([FakeChunk(content_delta="Fresh response")])
+        with (
+            patch.object(session, "_create_stream_with_retry", return_value=fresh_stream),
+            patch.object(session, "_full_messages", return_value=[]),
+        ):
+            session.send("new message")
+
+        # The new generation should have completed successfully
+        assert "idle" in ui.states
+        assistant_msgs = [m for m in session.messages if m["role"] == "assistant"]
+        assert any("Fresh response" in m.get("content", "") for m in assistant_msgs)

tests/test_channel_discord.py

@@ -390,6 +390,7 @@ class TestApprovalVerdictDisplay:
         bot.config.streaming_edit_interval = 1.5
         bot.config.auto_approve = False
         bot.config.auto_approve_tools = []
+        bot.storage = None
         bot._streaming = {}
         bot._pending_approval_msgs = {}
         bot._notify_reply_channels = {}

tests/test_channel_storage.py

@@ -2,17 +2,6 @@
 
 from __future__ import annotations
 
-import pytest
-
-from turnstone.core.storage._sqlite import SQLiteBackend
-
-
-@pytest.fixture
-def db(tmp_path):
-    """Fresh SQLite backend for each test."""
-    backend = SQLiteBackend(str(tmp_path / "test.db"))
-    return backend
-
 
 class TestChannelUserCRUD:
     """Tests for channel_users table operations."""

tests/test_config.py

@@ -3,54 +3,55 @@
 import argparse
 
 import turnstone.core.config as config_mod
-from turnstone.core.config import apply_config, load_config
+from turnstone.core.config import apply_config, load_config, set_config_path
 
 
 def _reset_cache():
     """Clear the module-level config cache between tests."""
     config_mod._cache = None
+    config_mod._config_path = None
 
 
-def test_load_config_missing_file(tmp_path, monkeypatch):
+def test_load_config_missing_file(tmp_path):
     _reset_cache()
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", tmp_path / "nope.toml")
+    set_config_path(str(tmp_path / "nope.toml"))
     assert load_config() == {}
 
 
-def test_load_config_valid_toml(tmp_path, monkeypatch):
+def test_load_config_valid_toml(tmp_path):
     _reset_cache()
     cfg = tmp_path / "config.toml"
     cfg.write_text('[redis]\nhost = "10.0.0.1"\nport = 6380\npassword = "secret"\n')
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", cfg)
+    set_config_path(str(cfg))
     result = load_config()
     assert result["redis"]["host"] == "10.0.0.1"
     assert result["redis"]["port"] == 6380
     assert result["redis"]["password"] == "secret"
 
 
-def test_load_config_section(tmp_path, monkeypatch):
+def test_load_config_section(tmp_path):
     _reset_cache()
     cfg = tmp_path / "config.toml"
     cfg.write_text('[api]\nbase_url = "http://x:8000/v1"\n[redis]\nhost = "y"\n')
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", cfg)
+    set_config_path(str(cfg))
     assert load_config("redis") == {"host": "y"}
     assert load_config("api") == {"base_url": "http://x:8000/v1"}
     assert load_config("nonexistent") == {}
 
 
-def test_load_config_invalid_toml(tmp_path, monkeypatch):
+def test_load_config_invalid_toml(tmp_path):
     _reset_cache()
     cfg = tmp_path / "config.toml"
     cfg.write_text("this is not valid toml [[[")
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", cfg)
+    set_config_path(str(cfg))
     assert load_config() == {}
 
 
-def test_load_config_caches(tmp_path, monkeypatch):
+def test_load_config_caches(tmp_path):
     _reset_cache()
     cfg = tmp_path / "config.toml"
     cfg.write_text('[api]\nbase_url = "http://first"\n')
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", cfg)
+    set_config_path(str(cfg))
     first = load_config()
     assert first["api"]["base_url"] == "http://first"
 
@@ -60,14 +61,14 @@ def test_load_config_caches(tmp_path, monkeypatch):
     assert second["api"]["base_url"] == "http://first"
 
 
-def test_apply_config_sets_defaults(tmp_path, monkeypatch):
+def test_apply_config_sets_defaults(tmp_path):
     _reset_cache()
     cfg = tmp_path / "config.toml"
     cfg.write_text(
         '[redis]\nhost = "redis.local"\nport = 7777\npassword = "pw"\n'
         '[bridge]\nserver_url = "http://bridge:9090"\n'
     )
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", cfg)
+    set_config_path(str(cfg))
 
     parser = argparse.ArgumentParser()
     parser.add_argument("--redis-host", default="localhost")
@@ -84,11 +85,11 @@ def test_apply_config_sets_defaults(tmp_path, monkeypatch):
     assert args.server_url == "http://bridge:9090"
 
 
-def test_apply_config_cli_overrides(tmp_path, monkeypatch):
+def test_apply_config_cli_overrides(tmp_path):
     _reset_cache()
     cfg = tmp_path / "config.toml"
     cfg.write_text('[redis]\nhost = "config-host"\nport = 7777\n')
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", cfg)
+    set_config_path(str(cfg))
 
     parser = argparse.ArgumentParser()
     parser.add_argument("--redis-host", default="localhost")
@@ -102,11 +103,11 @@ def test_apply_config_cli_overrides(tmp_path, monkeypatch):
     assert args.redis_port == 7777  # config wins (no CLI override)
 
 
-def test_apply_config_missing_keys_keep_defaults(tmp_path, monkeypatch):
+def test_apply_config_missing_keys_keep_defaults(tmp_path):
     _reset_cache()
     cfg = tmp_path / "config.toml"
     cfg.write_text('[redis]\nhost = "only-host"\n')  # no port, no password
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", cfg)
+    set_config_path(str(cfg))
 
     parser = argparse.ArgumentParser()
     parser.add_argument("--redis-host", default="localhost")
@@ -121,9 +122,9 @@ def test_apply_config_missing_keys_keep_defaults(tmp_path, monkeypatch):
     assert args.redis_password is None  # original default kept
 
 
-def test_apply_config_no_file(tmp_path, monkeypatch):
+def test_apply_config_no_file(tmp_path):
     _reset_cache()
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", tmp_path / "nope.toml")
+    set_config_path(str(tmp_path / "nope.toml"))
 
     parser = argparse.ArgumentParser()
     parser.add_argument("--redis-host", default="localhost")
@@ -133,11 +134,11 @@ def test_apply_config_no_file(tmp_path, monkeypatch):
     assert args.redis_host == "localhost"
 
 
-def test_apply_config_model_section(tmp_path, monkeypatch):
+def test_apply_config_model_section(tmp_path):
     _reset_cache()
     cfg = tmp_path / "config.toml"
     cfg.write_text('[model]\nname = "qwen-72b"\ntemperature = 0.3\n')
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", cfg)
+    set_config_path(str(cfg))
 
     parser = argparse.ArgumentParser()
     parser.add_argument("--model", default=None)
@@ -158,7 +159,7 @@ def test_tavily_key_from_config(tmp_path, monkeypatch):
 
     cfg = tmp_path / "config.toml"
     cfg.write_text('[api]\ntavily_key = "tvly-from-config"\n')
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", cfg)
+    set_config_path(str(cfg))
     monkeypatch.delenv("TAVILY_API_KEY", raising=False)
 
     key = config_mod.get_tavily_key()
@@ -174,14 +175,14 @@ def test_tavily_key_fallback_to_env(tmp_path, monkeypatch):
     # Config exists but no tavily_key in it
     cfg = tmp_path / "config.toml"
     cfg.write_text("[api]\n")
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", cfg)
+    set_config_path(str(cfg))
     monkeypatch.setenv("TAVILY_API_KEY", "tvly-from-env")
 
     key = config_mod.get_tavily_key()
     assert key == "tvly-from-env"
 
 
-def test_apply_config_judge_section(tmp_path, monkeypatch):
+def test_apply_config_judge_section(tmp_path):
     """apply_config() loads [judge] section and maps to argparse dests."""
     _reset_cache()
     cfg = tmp_path / "config.toml"
@@ -193,7 +194,7 @@ def test_apply_config_judge_section(tmp_path, monkeypatch):
         "timeout = 30.0\n"
         "read_only_tools = false\n"
     )
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", cfg)
+    set_config_path(str(cfg))
 
     parser = argparse.ArgumentParser()
     parser.add_argument("--judge", dest="judge_enabled", action="store_true", default=False)
@@ -212,12 +213,12 @@ def test_apply_config_judge_section(tmp_path, monkeypatch):
     assert args.judge_read_only_tools is False
 
 
-def test_apply_config_judge_cli_overrides(tmp_path, monkeypatch):
+def test_apply_config_judge_cli_overrides(tmp_path):
     """CLI flags override config.toml [judge] values."""
     _reset_cache()
     cfg = tmp_path / "config.toml"
     cfg.write_text("[judge]\nenabled = true\nconfidence_threshold = 0.85\n")
-    monkeypatch.setattr(config_mod, "CONFIG_PATH", cfg)
+    set_config_path(str(cfg))
 
     parser = argparse.ArgumentParser()
     parser.add_argument("--judge", dest="judge_enabled", action="store_true", default=False)
@@ -229,3 +230,35 @@ def test_apply_config_judge_cli_overrides(tmp_path, monkeypatch):
 
     assert args.judge_enabled is False  # CLI wins
     assert args.judge_confidence == 0.85  # config wins (no CLI override)
+
+
+def test_set_config_path_overrides_default(tmp_path):
+    """set_config_path() overrides the default config location."""
+    _reset_cache()
+    cfg = tmp_path / "custom.toml"
+    cfg.write_text('[api]\nbase_url = "http://custom:9999"\n')
+    set_config_path(str(cfg))
+    assert load_config("api") == {"base_url": "http://custom:9999"}
+
+
+def test_env_var_overrides_default(tmp_path, monkeypatch):
+    """$TURNSTONE_CONFIG env var overrides the default config location."""
+    _reset_cache()
+    cfg = tmp_path / "env.toml"
+    cfg.write_text('[api]\nbase_url = "http://env:7777"\n')
+    monkeypatch.setenv("TURNSTONE_CONFIG", str(cfg))
+    assert load_config("api") == {"base_url": "http://env:7777"}
+
+
+def test_set_config_path_overrides_env_var(tmp_path, monkeypatch):
+    """set_config_path() takes precedence over $TURNSTONE_CONFIG."""
+    _reset_cache()
+    env_cfg = tmp_path / "env.toml"
+    env_cfg.write_text('[api]\nbase_url = "http://env"\n')
+    monkeypatch.setenv("TURNSTONE_CONFIG", str(env_cfg))
+
+    explicit_cfg = tmp_path / "explicit.toml"
+    explicit_cfg.write_text('[api]\nbase_url = "http://explicit"\n')
+    set_config_path(str(explicit_cfg))
+
+    assert load_config("api") == {"base_url": "http://explicit"}

tests/test_console.py

@@ -3,7 +3,7 @@
 import asyncio
 import json
 import queue
-from unittest.mock import MagicMock
+from unittest.mock import MagicMock, patch
 
 import pytest
 
@@ -47,8 +47,8 @@ class MockBroker:
 # ---------------------------------------------------------------------------
 
 
-def _make_collector(broker=None, poll_interval=999, discovery_interval=999):
-    """Create a collector with long intervals so threads don't auto-fire."""
+def _make_collector(broker=None, poll_interval=0, discovery_interval=999):
+    """Create a collector with zero poll interval (no jitter delay in tests)."""
     b = broker or MockBroker()
     return ClusterCollector(
         broker=b,
@@ -264,6 +264,57 @@ class TestCollectorPolling:
         assert q.empty()
         assert len(c._nodes["node-a"].workstreams) == 0
 
+    def test_poll_401_preserves_workstreams_and_marks_unreachable(self):
+        """A 401 from the server must NOT wipe workstream data."""
+        c = _make_collector()
+        c._nodes["node-a"] = NodeSnapshot(
+            node_id="node-a",
+            server_url="http://a:8080",
+            reachable=True,
+            workstreams={"ws1": {"id": "ws1", "name": "existing", "state": "idle"}},
+        )
+
+        # Mock httpx to return 401
+        import httpx as _httpx
+
+        mock_response = _httpx.Response(
+            401,
+            json={"error": "Unauthorized"},
+            request=_httpx.Request("GET", "http://a:8080/v1/api/dashboard"),
+        )
+
+        with patch.object(c._http_client, "get", return_value=mock_response):
+            c._poll_all_nodes()
+
+        # Workstream data must be preserved, node marked unreachable
+        assert c._nodes["node-a"].reachable is False
+        assert "ws1" in c._nodes["node-a"].workstreams
+        assert c._nodes["node-a"].workstreams["ws1"]["name"] == "existing"
+
+    def test_poll_403_preserves_workstreams(self):
+        """A 403 should also preserve state and mark unreachable."""
+        c = _make_collector()
+        c._nodes["node-a"] = NodeSnapshot(
+            node_id="node-a",
+            server_url="http://a:8080",
+            reachable=True,
+            workstreams={"ws1": {"id": "ws1", "name": "keep-me", "state": "running"}},
+        )
+
+        import httpx as _httpx
+
+        mock_response = _httpx.Response(
+            403,
+            json={"error": "Forbidden"},
+            request=_httpx.Request("GET", "http://a:8080/v1/api/dashboard"),
+        )
+
+        with patch.object(c._http_client, "get", return_value=mock_response):
+            c._poll_all_nodes()
+
+        assert c._nodes["node-a"].reachable is False
+        assert "ws1" in c._nodes["node-a"].workstreams
+
 
 class TestCollectorEvents:
     """Real-time event handling from cluster channel."""
@@ -902,6 +953,8 @@ class TestConsoleWorkstreamCreation:
             ],
             2,
         )
+        # get_all_nodes delegates to get_nodes (mirrors real implementation)
+        collector.get_all_nodes.side_effect = lambda: collector.get_nodes.return_value[0]
         return collector
 
     @pytest.fixture()
@@ -1050,6 +1103,39 @@ class TestConsoleWorkstreamCreation:
         assert resp.status_code == 200
         assert resp.json()["target_node"] == "pool"
 
+    def test_create_with_resume_ws_directed(self, client_and_broker, mock_collector):
+        """resume_ws is forwarded in directed dispatch."""
+        client, broker = client_and_broker
+        resp = client.post(
+            "/v1/api/cluster/workstreams/new",
+            json={"node_id": "node-a", "resume_ws": "old-ws-id-123"},
+        )
+        assert resp.status_code == 200
+        msg = json.loads(broker.push_inbound.call_args[0][0])
+        assert msg["resume_ws"] == "old-ws-id-123"
+
+    def test_create_with_resume_ws_pool(self, client_and_broker, mock_collector):
+        """resume_ws is forwarded in pool dispatch."""
+        client, broker = client_and_broker
+        resp = client.post(
+            "/v1/api/cluster/workstreams/new",
+            json={"node_id": "pool", "resume_ws": "old-ws-id-456"},
+        )
+        assert resp.status_code == 200
+        msg = json.loads(broker.push_inbound.call_args[0][0])
+        assert msg["resume_ws"] == "old-ws-id-456"
+
+    def test_create_with_resume_ws_auto(self, client_and_broker, mock_collector):
+        """resume_ws is forwarded in auto-select dispatch."""
+        client, broker = client_and_broker
+        resp = client.post(
+            "/v1/api/cluster/workstreams/new",
+            json={"resume_ws": "old-ws-id-789"},
+        )
+        assert resp.status_code == 200
+        msg = json.loads(broker.push_inbound.call_args[0][0])
+        assert msg["resume_ws"] == "old-ws-id-789"
+
 
 # ---------------------------------------------------------------------------
 # Proxy tests
@@ -1182,47 +1268,49 @@ class TestProxyRewriting:
 class TestPickBestNode:
     """Test the _pick_best_node helper."""
 
+    @staticmethod
+    def _mock_collector(nodes: list) -> MagicMock:
+        collector = MagicMock(spec=ClusterCollector)
+        collector.get_nodes.return_value = (nodes, len(nodes))
+        collector.get_all_nodes.side_effect = lambda: collector.get_nodes.return_value[0]
+        return collector
+
     def test_picks_node_with_most_headroom(self):
         from turnstone.console.server import _pick_best_node
 
-        collector = MagicMock(spec=ClusterCollector)
-        collector.get_nodes.return_value = (
+        collector = self._mock_collector(
             [
                 {"node_id": "busy", "reachable": True, "max_ws": 10, "ws_total": 9},
                 {"node_id": "free", "reachable": True, "max_ws": 10, "ws_total": 2},
                 {"node_id": "mid", "reachable": True, "max_ws": 10, "ws_total": 5},
-            ],
-            3,
+            ]
         )
         assert _pick_best_node(collector) == "free"
 
     def test_skips_unreachable_nodes(self):
         from turnstone.console.server import _pick_best_node
 
-        collector = MagicMock(spec=ClusterCollector)
-        collector.get_nodes.return_value = (
+        collector = self._mock_collector(
             [
                 {"node_id": "down", "reachable": False, "max_ws": 10, "ws_total": 0},
                 {"node_id": "up", "reachable": True, "max_ws": 10, "ws_total": 5},
-            ],
-            2,
+            ]
         )
         assert _pick_best_node(collector) == "up"
 
     def test_returns_empty_when_no_nodes(self):
         from turnstone.console.server import _pick_best_node
 
-        collector = MagicMock(spec=ClusterCollector)
-        collector.get_nodes.return_value = ([], 0)
+        collector = self._mock_collector([])
         assert _pick_best_node(collector) == ""
 
     def test_returns_empty_when_all_unreachable(self):
         from turnstone.console.server import _pick_best_node
 
-        collector = MagicMock(spec=ClusterCollector)
-        collector.get_nodes.return_value = (
-            [{"node_id": "down", "reachable": False, "max_ws": 10, "ws_total": 0}],
-            1,
+        collector = self._mock_collector(
+            [
+                {"node_id": "down", "reachable": False, "max_ws": 10, "ws_total": 0},
+            ]
         )
         assert _pick_best_node(collector) == ""
 
@@ -1611,6 +1699,236 @@ class TestSSEProxy:
         asyncio.run(_run())
 
 
+# ---------------------------------------------------------------------------
+# Proxy auth header propagation
+# ---------------------------------------------------------------------------
+
+
+class TestProxyAuthHeaders:
+    """Verify _proxy_auth_headers mints user-scoped JWTs for proxy requests."""
+
+    SECRET = "test-secret-that-is-at-least-32-chars"
+
+    def _make_request(
+        self, *, auth_result=None, jwt_secret="", proxy_token_mgr=None, proxy_auth_token=""
+    ):
+        """Build a minimal fake request for _proxy_auth_headers."""
+
+        class _State:
+            pass
+
+        class _AppState:
+            pass
+
+        class _App:
+            state = _AppState()
+
+        class _Request:
+            state = _State()
+            app = _App()
+
+        req = _Request()
+        req.state.auth_result = auth_result
+        req.app.state.jwt_secret = jwt_secret
+        req.app.state.proxy_token_mgr = proxy_token_mgr
+        req.app.state.proxy_auth_token = proxy_auth_token
+        return req
+
+    def test_mints_user_jwt(self):
+        """Real user auth_result → JWT with correct sub, scopes, src, aud, permissions."""
+        import jwt as pyjwt
+
+        from turnstone.console.server import _proxy_auth_headers
+        from turnstone.core.auth import JWT_AUD_SERVER, AuthResult
+
+        auth = AuthResult(
+            user_id="alice",
+            scopes=frozenset({"read", "write"}),
+            token_source="jwt",
+            permissions=frozenset({"admin.users"}),
+        )
+        req = self._make_request(auth_result=auth, jwt_secret=self.SECRET)
+        headers = _proxy_auth_headers(req)
+
+        assert "Authorization" in headers
+        token = headers["Authorization"].removeprefix("Bearer ")
+        payload = pyjwt.decode(token, self.SECRET, algorithms=["HS256"], audience=JWT_AUD_SERVER)
+        assert payload["sub"] == "alice"
+        assert set(payload["scopes"].split(",")) == {"read", "write"}
+        assert payload["src"] == "console-proxy"
+        assert payload["aud"] == JWT_AUD_SERVER
+        assert payload["permissions"] == "admin.users"
+
+    def test_narrows_scopes(self):
+        """Read-only user → JWT carries only read scope, not full {read,write,approve}."""
+        import jwt as pyjwt
+
+        from turnstone.console.server import _proxy_auth_headers
+        from turnstone.core.auth import JWT_AUD_SERVER, AuthResult
+
+        auth = AuthResult(
+            user_id="viewer",
+            scopes=frozenset({"read"}),
+            token_source="jwt",
+        )
+        req = self._make_request(auth_result=auth, jwt_secret=self.SECRET)
+        headers = _proxy_auth_headers(req)
+
+        token = headers["Authorization"].removeprefix("Bearer ")
+        payload = pyjwt.decode(token, self.SECRET, algorithms=["HS256"], audience=JWT_AUD_SERVER)
+        assert payload["scopes"] == "read"
+
+    def test_short_expiry(self):
+        """Minted JWT expires in 300 seconds, not hours."""
+        import jwt as pyjwt
+
+        from turnstone.console.server import _proxy_auth_headers
+        from turnstone.core.auth import AuthResult
+
+        auth = AuthResult(
+            user_id="alice",
+            scopes=frozenset({"read"}),
+            token_source="jwt",
+        )
+        req = self._make_request(auth_result=auth, jwt_secret=self.SECRET)
+        headers = _proxy_auth_headers(req)
+
+        token = headers["Authorization"].removeprefix("Bearer ")
+        payload = pyjwt.decode(
+            token, self.SECRET, algorithms=["HS256"], options={"verify_aud": False}
+        )
+        assert payload["exp"] - payload["iat"] == 300
+
+    def test_fallback_no_user(self):
+        """No auth_result → falls back to ServiceTokenManager."""
+        from turnstone.console.server import _proxy_auth_headers
+        from turnstone.core.auth import ServiceTokenManager
+
+        mgr = ServiceTokenManager(
+            user_id="console-proxy",
+            scopes=frozenset({"read", "write", "approve"}),
+            source="console",
+            secret=self.SECRET,
+        )
+        req = self._make_request(proxy_token_mgr=mgr)
+        headers = _proxy_auth_headers(req)
+
+        assert "Authorization" in headers
+        assert headers["Authorization"] == f"Bearer {mgr.token}"
+
+    def test_fallback_no_secret(self):
+        """auth_result present but empty jwt_secret → falls back to ServiceTokenManager."""
+        from turnstone.console.server import _proxy_auth_headers
+        from turnstone.core.auth import AuthResult, ServiceTokenManager
+
+        auth = AuthResult(
+            user_id="alice",
+            scopes=frozenset({"read"}),
+            token_source="jwt",
+        )
+        mgr = ServiceTokenManager(
+            user_id="console-proxy",
+            scopes=frozenset({"read", "write", "approve"}),
+            source="console",
+            secret=self.SECRET,
+        )
+        req = self._make_request(auth_result=auth, jwt_secret="", proxy_token_mgr=mgr)
+        headers = _proxy_auth_headers(req)
+
+        # Should use ServiceTokenManager, not mint a user JWT
+        assert headers["Authorization"] == f"Bearer {mgr.token}"
+
+    def test_fallback_static_token(self):
+        """No auth_result, no ServiceTokenManager → uses static proxy_auth_token."""
+        from turnstone.console.server import _proxy_auth_headers
+
+        req = self._make_request(proxy_auth_token="static-tok-123")
+        headers = _proxy_auth_headers(req)
+
+        assert headers == {"Authorization": "Bearer static-tok-123"}
+
+
+# ---------------------------------------------------------------------------
+# Server: trusted user_id forwarding on create_workstream
+# ---------------------------------------------------------------------------
+
+
+class TestCreateWorkstreamUserIdTrust:
+    """Verify that only trusted service tokens can forward user_id in create_workstream."""
+
+    def _extract_uid(self, body: dict, auth_result) -> str:
+        """Replicate the trust check from server.py:create_workstream."""
+        auth = auth_result
+        uid: str = getattr(auth, "user_id", "") or ""
+        trusted_sources = {"bridge", "console"}
+        if (
+            body.get("user_id")
+            and isinstance(body["user_id"], str)
+            and auth is not None
+            and auth.token_source in trusted_sources
+        ):
+            uid = body["user_id"]
+        return uid
+
+    def test_bridge_can_forward_user_id(self):
+        from turnstone.core.auth import AuthResult
+
+        auth = AuthResult(
+            user_id="bridge",
+            scopes=frozenset({"approve"}),
+            token_source="bridge",
+        )
+        uid = self._extract_uid({"user_id": "real-user-abc"}, auth)
+        assert uid == "real-user-abc"
+
+    def test_console_service_can_forward_user_id(self):
+        from turnstone.core.auth import AuthResult
+
+        auth = AuthResult(
+            user_id="console",
+            scopes=frozenset({"approve"}),
+            token_source="console",
+        )
+        uid = self._extract_uid({"user_id": "real-user-abc"}, auth)
+        assert uid == "real-user-abc"
+
+    def test_console_proxy_user_cannot_override_user_id(self):
+        """End-user tokens via console-proxy must NOT override user_id."""
+        from turnstone.core.auth import AuthResult
+
+        auth = AuthResult(
+            user_id="real-user-abc",
+            scopes=frozenset({"read", "write"}),
+            token_source="console-proxy",
+        )
+        uid = self._extract_uid({"user_id": "impersonated-user"}, auth)
+        # Should use JWT identity, NOT the body override
+        assert uid == "real-user-abc"
+
+    def test_direct_user_cannot_override_user_id(self):
+        """Direct JWT login must NOT override user_id."""
+        from turnstone.core.auth import AuthResult
+
+        auth = AuthResult(
+            user_id="real-user-abc",
+            scopes=frozenset({"read", "write"}),
+            token_source="password",
+        )
+        uid = self._extract_uid({"user_id": "impersonated-user"}, auth)
+        assert uid == "real-user-abc"
+
+    def test_no_body_user_id_uses_jwt(self):
+        from turnstone.core.auth import AuthResult
+
+        auth = AuthResult(
+            user_id="bridge",
+            scopes=frozenset({"approve"}),
+            token_source="bridge",
+        )
+        uid = self._extract_uid({"name": "test-ws"}, auth)
+        assert uid == "bridge"
+
+
 # ---------------------------------------------------------------------------
 # Collector — MCP aggregation in get_overview()
 # ---------------------------------------------------------------------------

tests/test_edit_file.py

@@ -0,0 +1,460 @@
+"""Tests for edit_file tool — single edit and batch edit modes."""
+
+from __future__ import annotations
+
+import os
+from unittest.mock import MagicMock
+
+import pytest
+
+from turnstone.core.session import ChatSession
+
+
+@pytest.fixture
+def session(tmp_db, mock_openai_client):
+    """Create a ChatSession wired to a temp database."""
+    return ChatSession(
+        client=mock_openai_client,
+        model="test-model",
+        ui=MagicMock(),
+        instructions=None,
+        temperature=0.5,
+        max_tokens=1000,
+        tool_timeout=10,
+    )
+
+
+@pytest.fixture
+def sample_file(tmp_path):
+    """Create a sample file and return its path."""
+    p = tmp_path / "test.py"
+    p.write_text("line1\nline2\nline3\nline4\nline5\n")
+    return str(p)
+
+
+def _mark_read(session: ChatSession, path: str) -> None:
+    """Simulate a prior read_file so the edit guard passes."""
+    resolved = os.path.realpath(os.path.expanduser(path))
+    session._read_files.add(resolved)
+
+
+# ── Single edit (backward compat) ────────────────────────────────────
+
+
+class TestSingleEdit:
+    def test_basic_replace(self, session, sample_file):
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "old_string": "line2",
+                "new_string": "replaced",
+            },
+        )
+        assert result["needs_approval"]
+        assert result["func_name"] == "edit_file"
+
+        call_id, msg = session._exec_edit_file(result)
+        assert call_id == "c1"
+        assert "applied 1 edit" in msg
+        with open(sample_file) as f:
+            assert f.read() == "line1\nreplaced\nline3\nline4\nline5\n"
+
+    def test_missing_path(self, session):
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "old_string": "a",
+                "new_string": "b",
+            },
+        )
+        assert result.get("error")
+        assert "missing path" in result["error"]
+
+    def test_missing_old_string(self, session, sample_file):
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "new_string": "b",
+            },
+        )
+        assert result.get("error")
+        assert "old_string" in result["error"]
+
+    def test_identical_strings(self, session, sample_file):
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "old_string": "line1",
+                "new_string": "line1",
+            },
+        )
+        assert result.get("error")
+        assert "identical" in result["error"]
+
+    def test_old_string_not_found(self, session, sample_file):
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "old_string": "nonexistent",
+                "new_string": "replaced",
+            },
+        )
+        assert result.get("error")
+        assert "not found" in result["error"]
+
+    def test_must_read_first(self, session, sample_file):
+        # Don't call _mark_read — should fail
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "old_string": "line1",
+                "new_string": "replaced",
+            },
+        )
+        assert result.get("error")
+        assert "must read_file" in result["error"]
+
+    def test_multiple_occurrences_without_near_line(self, session, tmp_path):
+        p = tmp_path / "dup.txt"
+        p.write_text("foo\nbar\nfoo\n")
+        path = str(p)
+        _mark_read(session, path)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": path,
+                "old_string": "foo",
+                "new_string": "baz",
+            },
+        )
+        assert result.get("error")
+        assert "found 2 times" in result["error"]
+
+    def test_near_line_disambiguates(self, session, tmp_path):
+        p = tmp_path / "dup.txt"
+        p.write_text("foo\nbar\nfoo\n")
+        path = str(p)
+        _mark_read(session, path)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": path,
+                "old_string": "foo",
+                "new_string": "baz",
+                "near_line": 3,
+            },
+        )
+        assert result["needs_approval"]
+
+        call_id, msg = session._exec_edit_file(result)
+        assert "applied 1 edit" in msg
+        with open(path) as f:
+            assert f.read() == "foo\nbar\nbaz\n"
+
+    def test_deletion(self, session, sample_file):
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "old_string": "line3\n",
+                "new_string": "",
+            },
+        )
+        assert result["needs_approval"]
+        assert "deletion" in result["preview"]
+
+        call_id, msg = session._exec_edit_file(result)
+        with open(sample_file) as f:
+            assert f.read() == "line1\nline2\nline4\nline5\n"
+
+
+# ── Batch edits ──────────────────────────────────────────────────────
+
+
+class TestBatchEdit:
+    def test_two_edits_applied_atomically(self, session, sample_file):
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "edits": [
+                    {"old_string": "line1", "new_string": "first"},
+                    {"old_string": "line5", "new_string": "last"},
+                ],
+            },
+        )
+        assert result["needs_approval"]
+        assert "2 edits" in result["header"]
+
+        call_id, msg = session._exec_edit_file(result)
+        assert "applied 2 edits" in msg
+        with open(sample_file) as f:
+            assert f.read() == "first\nline2\nline3\nline4\nlast\n"
+
+    def test_three_edits_middle_of_file(self, session, sample_file):
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "edits": [
+                    {"old_string": "line2", "new_string": "second"},
+                    {"old_string": "line3", "new_string": "third"},
+                    {"old_string": "line4", "new_string": "fourth"},
+                ],
+            },
+        )
+        assert result["needs_approval"]
+
+        call_id, msg = session._exec_edit_file(result)
+        assert "applied 3 edits" in msg
+        with open(sample_file) as f:
+            assert f.read() == "line1\nsecond\nthird\nfourth\nline5\n"
+
+    def test_overlapping_edits_rejected(self, session, tmp_path):
+        p = tmp_path / "overlap.txt"
+        p.write_text("abcdefgh\n")
+        path = str(p)
+        _mark_read(session, path)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": path,
+                "edits": [
+                    {"old_string": "abcdef", "new_string": "XXX"},
+                    {"old_string": "defgh", "new_string": "YYY"},
+                ],
+            },
+        )
+        assert result["needs_approval"]
+
+        call_id, msg = session._exec_edit_file(result)
+        assert "overlap" in msg.lower()
+        # File should be untouched
+        with open(path) as f:
+            assert f.read() == "abcdefgh\n"
+
+    def test_batch_edit_not_found(self, session, sample_file):
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "edits": [
+                    {"old_string": "line1", "new_string": "first"},
+                    {"old_string": "nonexistent", "new_string": "oops"},
+                ],
+            },
+        )
+        assert result.get("error")
+        assert "edits[1]" in result["error"]
+        assert "not found" in result["error"]
+
+    def test_batch_edit_missing_old_string(self, session, sample_file):
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "edits": [
+                    {"old_string": "line1", "new_string": "first"},
+                    {"new_string": "oops"},
+                ],
+            },
+        )
+        assert result.get("error")
+        assert "edits[1]" in result["error"]
+        assert "old_string" in result["error"]
+
+    def test_batch_edit_identical_strings(self, session, sample_file):
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "edits": [
+                    {"old_string": "line1", "new_string": "line1"},
+                ],
+            },
+        )
+        assert result.get("error")
+        assert "identical" in result["error"]
+
+    def test_batch_with_near_line(self, session, tmp_path):
+        p = tmp_path / "dup.txt"
+        p.write_text("foo\nbar\nfoo\nbaz\n")
+        path = str(p)
+        _mark_read(session, path)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": path,
+                "edits": [
+                    {"old_string": "foo", "new_string": "first_foo", "near_line": 1},
+                    {"old_string": "foo", "new_string": "second_foo", "near_line": 3},
+                ],
+            },
+        )
+        assert result["needs_approval"]
+
+        call_id, msg = session._exec_edit_file(result)
+        assert "applied 2 edits" in msg
+        with open(path) as f:
+            assert f.read() == "first_foo\nbar\nsecond_foo\nbaz\n"
+
+    def test_batch_with_deletion(self, session, sample_file):
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "edits": [
+                    {"old_string": "line2\n", "new_string": ""},
+                    {"old_string": "line4\n", "new_string": ""},
+                ],
+            },
+        )
+        assert result["needs_approval"]
+
+        call_id, msg = session._exec_edit_file(result)
+        with open(sample_file) as f:
+            assert f.read() == "line1\nline3\nline5\n"
+
+    def test_single_item_edits_array(self, session, sample_file):
+        """An edits array with one item should work like a single edit."""
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "edits": [
+                    {"old_string": "line3", "new_string": "middle"},
+                ],
+            },
+        )
+        assert result["needs_approval"]
+        # Single edit — no "(N edits)" count in header
+        assert "edits)" not in result["header"]
+
+        call_id, msg = session._exec_edit_file(result)
+        assert "applied 1 edit" in msg
+        with open(sample_file) as f:
+            assert f.read() == "line1\nline2\nmiddle\nline4\nline5\n"
+
+
+# ── Mutual exclusivity ──────────────────────────────────────────────
+
+
+class TestMutualExclusivity:
+    def test_both_single_and_batch_rejected(self, session, sample_file):
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "old_string": "line1",
+                "new_string": "replaced",
+                "edits": [
+                    {"old_string": "line2", "new_string": "also_replaced"},
+                ],
+            },
+        )
+        assert result.get("error")
+        assert "not both" in result["error"]
+
+    def test_neither_single_nor_batch(self, session, sample_file):
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+            },
+        )
+        assert result.get("error")
+        assert "old_string" in result["error"]
+
+    def test_empty_edits_array_falls_through_to_single(self, session, sample_file):
+        """An empty edits array should be treated as no batch."""
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "edits": [],
+            },
+        )
+        # Falls through to single-edit path, which requires old_string
+        assert result.get("error")
+        assert "old_string" in result["error"]
+
+
+# ── TOCTOU edge cases ───────────────────────────────────────────────
+
+
+class TestExecEdgeCases:
+    def test_file_changed_between_prepare_and_exec(self, session, sample_file):
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "old_string": "line2",
+                "new_string": "replaced",
+            },
+        )
+        assert result["needs_approval"]
+
+        # Modify the file after prepare
+        with open(sample_file, "w") as f:
+            f.write("completely different content\n")
+
+        call_id, msg = session._exec_edit_file(result)
+        assert "no longer found" in msg
+
+    def test_file_deleted_between_prepare_and_exec(self, session, sample_file):
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "old_string": "line2",
+                "new_string": "replaced",
+            },
+        )
+        assert result["needs_approval"]
+
+        os.unlink(sample_file)
+
+        call_id, msg = session._exec_edit_file(result)
+        assert "Error" in msg
+
+    def test_batch_file_changed_partial_match(self, session, sample_file):
+        """If file changes so one edit fails, none should be applied."""
+        _mark_read(session, sample_file)
+        result = session._prepare_edit_file(
+            "c1",
+            {
+                "path": sample_file,
+                "edits": [
+                    {"old_string": "line1", "new_string": "first"},
+                    {"old_string": "line5", "new_string": "last"},
+                ],
+            },
+        )
+        assert result["needs_approval"]
+
+        # Remove line5 between prepare and exec
+        with open(sample_file, "w") as f:
+            f.write("line1\nline2\nline3\nline4\n")
+
+        call_id, msg = session._exec_edit_file(result)
+        assert "no longer found" in msg
+        # line1 should NOT have been edited (atomic failure)
+        with open(sample_file) as f:
+            assert "line1" in f.read()

tests/test_env_scrub.py

@@ -0,0 +1,149 @@
+"""Tests for turnstone.core.env — subprocess environment scrubbing."""
+
+from __future__ import annotations
+
+import os
+from unittest.mock import patch
+
+from turnstone.core.env import _is_safe, _is_secret, scrubbed_env
+
+
+class TestIsSecret:
+    def test_explicit_scrub_list(self):
+        assert _is_secret("OPENAI_API_KEY") is True
+        assert _is_secret("ANTHROPIC_API_KEY") is True
+        assert _is_secret("TURNSTONE_JWT_SECRET") is True
+        assert _is_secret("AWS_SECRET_ACCESS_KEY") is True
+
+    def test_suffix_matching(self):
+        assert _is_secret("MY_CUSTOM_API_KEY") is True
+        assert _is_secret("DB_PASSWORD") is True
+        assert _is_secret("AUTH_TOKEN") is True
+        assert _is_secret("SERVICE_CREDENTIAL") is True
+        assert _is_secret("GCP_CREDENTIALS") is True
+
+    def test_safe_vars_not_secret(self):
+        assert _is_secret("PATH") is False
+        assert _is_secret("HOME") is False
+        assert _is_secret("LANG") is False
+
+    def test_no_false_positives_on_substring(self):
+        """Suffix matching avoids false positives like MONKEYTYPE."""
+        assert _is_secret("MONKEYTYPE") is False
+        assert _is_secret("KEYBOARD_LAYOUT") is False
+        assert _is_secret("PYTHONPATH") is False
+        assert _is_secret("EDITOR") is False
+        assert _is_secret("GOPATH") is False
+
+
+class TestIsSafe:
+    def test_safe_names(self):
+        assert _is_safe("PATH") is True
+        assert _is_safe("HOME") is True
+        assert _is_safe("TERM") is True
+        assert _is_safe("MANWIDTH") is True
+
+    def test_safe_prefixes(self):
+        assert _is_safe("LC_ALL") is True
+        assert _is_safe("LC_CTYPE") is True
+        assert _is_safe("XDG_RUNTIME_DIR") is True
+
+    def test_non_safe_names(self):
+        assert _is_safe("OPENAI_API_KEY") is False
+        assert _is_safe("CUSTOM_VAR") is False
+
+
+class TestScrubbedEnv:
+    def test_strips_api_keys(self):
+        fake_env = {
+            "PATH": "/usr/bin",
+            "HOME": "/home/user",
+            "OPENAI_API_KEY": "sk-secret",
+            "ANTHROPIC_API_KEY": "ant-secret",
+            "CUSTOM_VAR": "safe_value",
+        }
+        with patch.dict(os.environ, fake_env, clear=True):
+            result = scrubbed_env()
+
+        assert result["PATH"] == "/usr/bin"
+        assert result["HOME"] == "/home/user"
+        assert result["CUSTOM_VAR"] == "safe_value"
+        assert "OPENAI_API_KEY" not in result
+        assert "ANTHROPIC_API_KEY" not in result
+
+    def test_strips_pattern_matched_secrets(self):
+        fake_env = {
+            "PATH": "/usr/bin",
+            "MY_SERVICE_TOKEN": "tok-123",
+            "DB_PASSWORD": "pass123",
+        }
+        with patch.dict(os.environ, fake_env, clear=True):
+            result = scrubbed_env()
+
+        assert "MY_SERVICE_TOKEN" not in result
+        assert "DB_PASSWORD" not in result
+
+    def test_extra_vars_merged(self):
+        fake_env = {"PATH": "/usr/bin"}
+        with patch.dict(os.environ, fake_env, clear=True):
+            result = scrubbed_env(extra={"MANWIDTH": "80"})
+
+        assert result["MANWIDTH"] == "80"
+        assert result["PATH"] == "/usr/bin"
+
+    def test_passthrough_overrides_scrub(self):
+        fake_env = {
+            "PATH": "/usr/bin",
+            "OPENAI_API_KEY": "sk-needed",
+        }
+        with patch.dict(os.environ, fake_env, clear=True):
+            result = scrubbed_env(passthrough=["OPENAI_API_KEY"])
+
+        assert result["OPENAI_API_KEY"] == "sk-needed"
+
+    def test_preserves_locale_vars(self):
+        fake_env = {
+            "PATH": "/usr/bin",
+            "LC_ALL": "en_US.UTF-8",
+            "LC_CTYPE": "en_US.UTF-8",
+        }
+        with patch.dict(os.environ, fake_env, clear=True):
+            result = scrubbed_env()
+
+        assert result["LC_ALL"] == "en_US.UTF-8"
+        assert result["LC_CTYPE"] == "en_US.UTF-8"
+
+    def test_preserves_unknown_non_secret_vars(self):
+        fake_env = {
+            "PATH": "/usr/bin",
+            "PYTHONPATH": "/opt/lib",
+            "GOPATH": "/home/user/go",
+        }
+        with patch.dict(os.environ, fake_env, clear=True):
+            result = scrubbed_env()
+
+        assert result["PYTHONPATH"] == "/opt/lib"
+        assert result["GOPATH"] == "/home/user/go"
+
+    def test_extra_can_reintroduce_scrubbed_var(self):
+        """extra= intentionally overrides scrubbing (operator-controlled)."""
+        fake_env = {"PATH": "/usr/bin", "OPENAI_API_KEY": "sk-original"}
+        with patch.dict(os.environ, fake_env, clear=True):
+            result = scrubbed_env(extra={"OPENAI_API_KEY": "sk-injected"})
+
+        assert result["OPENAI_API_KEY"] == "sk-injected"
+
+    def test_less_prefix_does_not_leak_secrets(self):
+        """LESS pager vars are safe but LESS_SECRET_TOKEN is not."""
+        fake_env = {
+            "PATH": "/usr/bin",
+            "LESS": "-R",
+            "LESSOPEN": "| lesspipe %s",
+            "LESS_SECRET_TOKEN": "tok-secret",
+        }
+        with patch.dict(os.environ, fake_env, clear=True):
+            result = scrubbed_env()
+
+        assert result["LESS"] == "-R"
+        assert result["LESSOPEN"] == "| lesspipe %s"
+        assert "LESS_SECRET_TOKEN" not in result

tests/test_governance_storage.py

@@ -8,18 +8,8 @@ from __future__ import annotations
 
 from datetime import UTC, datetime
 
-import pytest
 import sqlalchemy as sa
 
-from turnstone.core.storage._sqlite import SQLiteBackend
-
-
-@pytest.fixture()
-def db(tmp_path):
-    """Create a fresh SQLite backend for each test."""
-    return SQLiteBackend(str(tmp_path / "test.db"))
-
-
 # ---------------------------------------------------------------------------
 # Roles
 # ---------------------------------------------------------------------------

tests/test_healthcheck.py

@@ -221,6 +221,63 @@ class TestBackendHealthMonitor:
         mon.record_failure()
         assert mon.circuit_state == CircuitState.OPEN  # type: ignore[comparison-overlap]
 
+    def test_probe_loop_autonomous_recovery(
+        self, mock_client: MagicMock, mock_metrics: MagicMock
+    ) -> None:
+        """_probe_loop transitions OPEN → HALF_OPEN → CLOSED without user requests."""
+        # Use very short intervals so the test is fast
+        mon = BackendHealthMonitor(
+            client=mock_client,
+            probe_interval=0.05,
+            probe_timeout=1.0,
+            failure_threshold=1,
+            cooldown=0.1,
+        )
+        # Trip the circuit
+        mon.record_failure()
+        assert mon.circuit_state == CircuitState.OPEN
+
+        # Backend is healthy — probe_once will succeed
+        mock_client.with_options.return_value.models.list.return_value = MagicMock()
+
+        # Start the probe loop and wait for autonomous recovery
+        mon.start()
+        try:
+            import time
+
+            deadline = time.monotonic() + 5.0
+            while mon.circuit_state != CircuitState.CLOSED and time.monotonic() < deadline:
+                time.sleep(0.05)
+            assert mon.circuit_state == CircuitState.CLOSED
+            # User requests should flow again without anyone calling acquire_request_permit
+            assert mon.acquire_request_permit() is True
+        finally:
+            mon.stop()
+            if mon._thread:
+                mon._thread.join(timeout=2.0)
+
+    def test_probe_loop_no_user_permit_during_probe(
+        self, mock_client: MagicMock, mock_metrics: MagicMock
+    ) -> None:
+        """While background probe is in HALF_OPEN, user requests are blocked."""
+        mon = BackendHealthMonitor(
+            client=mock_client,
+            probe_interval=0.05,
+            probe_timeout=1.0,
+            failure_threshold=1,
+            cooldown=0.1,
+        )
+        mon.record_failure()
+        assert mon.circuit_state == CircuitState.OPEN
+
+        # Force into HALF_OPEN as the probe loop would
+        with mon._lock:
+            mon._state = CircuitState.HALF_OPEN
+            mon._half_open_permit = False  # probe consumes it
+
+        # User requests should be blocked — only the probe gets through
+        assert mon.acquire_request_permit() is False
+
     def test_stop_thread(self, mock_client: MagicMock) -> None:
         """stop() signals the probe loop to exit."""
         mon = _make_monitor(mock_client)

tests/test_judge.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import json
 import time
+from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
 from typing import Any
 from unittest.mock import MagicMock
@@ -179,11 +180,13 @@ class TestErrorHandling:
         provider = _make_mock_provider(side_effect=RuntimeError("API error"))
         judge = _make_judge(provider)
 
-        result = judge._evaluate_single(
-            _make_item(),
-            [{"role": "user", "content": "test"}],
-            MagicMock(),
-        )
+        with ThreadPoolExecutor(max_workers=1) as pool:
+            result = judge._evaluate_single(
+                _make_item(),
+                [{"role": "user", "content": "test"}],
+                cancel_event=None,
+                executor=pool,
+            )
         assert result is None
 
     def test_provider_error_heuristic_still_returned(self):
@@ -212,11 +215,13 @@ class TestErrorHandling:
         result_mock.content = ""
 
         judge = _make_judge(provider)
-        result = judge._evaluate_single(
-            _make_item(),
-            [{"role": "user", "content": "test"}],
-            MagicMock(),
-        )
+        with ThreadPoolExecutor(max_workers=1) as pool:
+            result = judge._evaluate_single(
+                _make_item(),
+                [{"role": "user", "content": "test"}],
+                cancel_event=None,
+                executor=pool,
+            )
         assert result is None
 
 
@@ -256,11 +261,13 @@ class TestMultiTurnToolUse:
         provider.create_completion.side_effect = [turn1, turn2]
 
         judge = _make_judge(provider)
-        verdict = judge._evaluate_single(
-            _make_item(),
-            [{"role": "user", "content": "test"}],
-            MagicMock(),
-        )
+        with ThreadPoolExecutor(max_workers=1) as pool:
+            verdict = judge._evaluate_single(
+                _make_item(),
+                [{"role": "user", "content": "test"}],
+                cancel_event=None,
+                executor=pool,
+            )
         assert verdict is not None
         assert verdict.tier == "llm"
         assert provider.create_completion.call_count == 2
@@ -302,11 +309,13 @@ class TestMultiTurnToolUse:
         ]
 
         judge = _make_judge(provider)
-        judge._evaluate_single(
-            _make_item(),
-            [{"role": "user", "content": "test"}],
-            MagicMock(),
-        )
+        with ThreadPoolExecutor(max_workers=1) as pool:
+            judge._evaluate_single(
+                _make_item(),
+                [{"role": "user", "content": "test"}],
+                cancel_event=None,
+                executor=pool,
+            )
         # Should have called create_completion exactly _JUDGE_MAX_TURNS times
         assert provider.create_completion.call_count == 5
 

tests/test_judge_storage.py

@@ -4,16 +4,6 @@ from __future__ import annotations
 
 from datetime import UTC, datetime, timedelta
 
-import pytest
-
-from turnstone.core.storage._sqlite import SQLiteBackend
-
-
-@pytest.fixture()
-def db(tmp_path):
-    """Fresh SQLite backend for each test."""
-    return SQLiteBackend(str(tmp_path / "test.db"))
-
 
 def _make_verdict_kwargs(**overrides):
     """Build default kwargs for create_intent_verdict."""

tests/test_load_skill.py

@@ -1,4 +1,4 @@
-"""Tests for the load_skill built-in tool."""
+"""Tests for the skill built-in tool."""
 
 from __future__ import annotations
 
@@ -9,25 +9,25 @@ from turnstone.core.tools import BUILTIN_TOOL_NAMES, PRIMARY_KEY_MAP
 
 
 class TestToolRegistration:
-    """Verify load_skill is registered correctly."""
+    """Verify skill is registered correctly."""
 
     def test_in_builtin_tool_names(self) -> None:
-        assert "load_skill" in BUILTIN_TOOL_NAMES
+        assert "skill" in BUILTIN_TOOL_NAMES
 
     def test_not_agent_tool(self) -> None:
         from turnstone.core.tools import AGENT_TOOLS
 
         names = {t["function"]["name"] for t in AGENT_TOOLS}
-        assert "load_skill" not in names
+        assert "skill" not in names
 
     def test_not_task_agent_tool(self) -> None:
         from turnstone.core.tools import TASK_AGENT_TOOLS
 
         names = {t["function"]["name"] for t in TASK_AGENT_TOOLS}
-        assert "load_skill" not in names
+        assert "skill" not in names
 
     def test_has_primary_key(self) -> None:
-        assert PRIMARY_KEY_MAP.get("load_skill") == "name"
+        assert PRIMARY_KEY_MAP.get("skill") == "name"
 
 
 # ---------------------------------------------------------------------------
@@ -54,6 +54,7 @@ def _make_session(skills: list[dict[str, Any]] | None = None):
     session._notify_on_complete = "{}"
     session.messages = []
     session._config = {}
+    session._tool_error_flags = {}
 
     # Stub set_skill to just record the call
     session._set_skill_called: list[str | None] = []
@@ -82,12 +83,12 @@ def _make_session(skills: list[dict[str, Any]] | None = None):
 
 
 class TestPrepareLoadSkill:
-    """Test _prepare_load_skill validation and item dict shape."""
+    """Test _prepare_skill validation and item dict shape."""
 
     def test_load_valid(self) -> None:
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "load", "name": "code-review"})
-        assert item["func_name"] == "load_skill"
+        item = session._prepare_skill("call-1", {"action": "load", "name": "code-review"})
+        assert item["func_name"] == "skill"
         assert item["action"] == "load"
         assert item["name"] == "code-review"
         assert item["needs_approval"] is True
@@ -96,19 +97,19 @@ class TestPrepareLoadSkill:
 
     def test_load_missing_name(self) -> None:
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "load"})
+        item = session._prepare_skill("call-1", {"action": "load"})
         assert "error" in item
         assert "name" in item["error"].lower()
         assert item["needs_approval"] is False
 
     def test_load_empty_name(self) -> None:
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "load", "name": ""})
+        item = session._prepare_skill("call-1", {"action": "load", "name": ""})
         assert "error" in item
 
     def test_search_with_query(self) -> None:
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "search", "query": "code review"})
+        item = session._prepare_skill("call-1", {"action": "search", "query": "code review"})
         assert item["action"] == "search"
         assert item["query"] == "code review"
         assert item["needs_approval"] is False
@@ -116,30 +117,30 @@ class TestPrepareLoadSkill:
 
     def test_search_without_query(self) -> None:
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "search"})
+        item = session._prepare_skill("call-1", {"action": "search"})
         assert item["action"] == "search"
         assert item["query"] == ""
         assert item["needs_approval"] is False
 
     def test_invalid_action(self) -> None:
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "delete"})
+        item = session._prepare_skill("call-1", {"action": "delete"})
         assert "error" in item
         assert "delete" in item["error"]
 
     def test_empty_action(self) -> None:
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": ""})
+        item = session._prepare_skill("call-1", {"action": ""})
         assert "error" in item
 
     def test_header_for_load(self) -> None:
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "load", "name": "my-skill"})
+        item = session._prepare_skill("call-1", {"action": "load", "name": "my-skill"})
         assert "my-skill" in item["header"]
 
     def test_header_for_search(self) -> None:
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "search", "query": "testing"})
+        item = session._prepare_skill("call-1", {"action": "search", "query": "testing"})
         assert "testing" in item["header"]
 
 
@@ -149,7 +150,7 @@ class TestPrepareLoadSkill:
 
 
 class TestExecLoadSkill:
-    """Test _exec_load_skill execution logic."""
+    """Test _exec_skill execution logic."""
 
     def test_load_existing_skill(self) -> None:
         skills = [
@@ -164,8 +165,8 @@ class TestExecLoadSkill:
         session, _, fake_get = _make_session(skills)
 
         with patch("turnstone.core.session.get_skill_by_name", side_effect=fake_get):
-            item = session._prepare_load_skill("call-1", {"action": "load", "name": "code-review"})
-            call_id, result = session._exec_load_skill(item)
+            item = session._prepare_skill("call-1", {"action": "load", "name": "code-review"})
+            call_id, result = session._exec_skill(item)
 
         assert call_id == "call-1"
         assert "code-review" in result
@@ -177,8 +178,8 @@ class TestExecLoadSkill:
         session, _, fake_get = _make_session([])
 
         with patch("turnstone.core.session.get_skill_by_name", side_effect=fake_get):
-            item = session._prepare_load_skill("call-1", {"action": "load", "name": "nope"})
-            call_id, result = session._exec_load_skill(item)
+            item = session._prepare_skill("call-1", {"action": "load", "name": "nope"})
+            call_id, result = session._exec_skill(item)
 
         assert "not found" in result.lower()
         assert session._set_skill_called == []
@@ -188,8 +189,8 @@ class TestExecLoadSkill:
         session, _, fake_get = _make_session(skills)
 
         with patch("turnstone.core.session.get_skill_by_name", side_effect=fake_get):
-            item = session._prepare_load_skill("call-1", {"action": "load", "name": "test"})
-            session._exec_load_skill(item)
+            item = session._prepare_skill("call-1", {"action": "load", "name": "test"})
+            session._exec_skill(item)
 
         session.ui.on_tool_result.assert_called_once()
 
@@ -216,10 +217,10 @@ class TestExecLoadSkill:
         mock_storage.list_prompt_templates.return_value = skills
 
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "search", "query": "code"})
+        item = session._prepare_skill("call-1", {"action": "search", "query": "code"})
 
         with patch("turnstone.core.storage._registry.get_storage", return_value=mock_storage):
-            call_id, result = session._exec_load_skill(item)
+            call_id, result = session._exec_skill(item)
 
         assert "code-review" in result
         # docs-writer shouldn't match "code" query
@@ -241,10 +242,10 @@ class TestExecLoadSkill:
         mock_storage.list_prompt_templates.return_value = skills
 
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "search"})
+        item = session._prepare_skill("call-1", {"action": "search"})
 
         with patch("turnstone.core.storage._registry.get_storage", return_value=mock_storage):
-            call_id, result = session._exec_load_skill(item)
+            call_id, result = session._exec_skill(item)
 
         # Should be limited to 10
         assert result.count("skill-") == 10
@@ -254,10 +255,10 @@ class TestExecLoadSkill:
         mock_storage.list_prompt_templates.return_value = []
 
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "search", "query": "nonexistent"})
+        item = session._prepare_skill("call-1", {"action": "search", "query": "nonexistent"})
 
         with patch("turnstone.core.storage._registry.get_storage", return_value=mock_storage):
-            call_id, result = session._exec_load_skill(item)
+            call_id, result = session._exec_skill(item)
 
         assert "no skills found" in result.lower()
 
@@ -276,21 +277,21 @@ class TestExecLoadSkill:
         mock_storage.list_prompt_templates.return_value = skills
 
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "search", "query": "risky"})
+        item = session._prepare_skill("call-1", {"action": "search", "query": "risky"})
 
         with patch("turnstone.core.storage._registry.get_storage", return_value=mock_storage):
-            call_id, result = session._exec_load_skill(item)
+            call_id, result = session._exec_skill(item)
 
         assert "high" in result
 
     def test_search_storage_failure_returns_empty(self) -> None:
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "search", "query": "test"})
+        item = session._prepare_skill("call-1", {"action": "search", "query": "test"})
 
         with patch(
             "turnstone.core.storage._registry.get_storage", side_effect=RuntimeError("no storage")
         ):
-            call_id, result = session._exec_load_skill(item)
+            call_id, result = session._exec_skill(item)
 
         assert "no skills found" in result.lower()
 
@@ -307,10 +308,8 @@ class TestExecLoadSkill:
         session, _, fake_get = _make_session(skills)
 
         with patch("turnstone.core.session.get_skill_by_name", side_effect=fake_get):
-            item = session._prepare_load_skill(
-                "call-1", {"action": "load", "name": "disabled-skill"}
-            )
-            call_id, result = session._exec_load_skill(item)
+            item = session._prepare_skill("call-1", {"action": "load", "name": "disabled-skill"})
+            call_id, result = session._exec_skill(item)
 
         assert "not found" in result.lower()
         assert session._set_skill_called == []
@@ -321,8 +320,8 @@ class TestExecLoadSkill:
         session._skill_name = "active"
 
         with patch("turnstone.core.session.get_skill_by_name", side_effect=fake_get):
-            item = session._prepare_load_skill("call-1", {"action": "load", "name": "active"})
-            call_id, result = session._exec_load_skill(item)
+            item = session._prepare_skill("call-1", {"action": "load", "name": "active"})
+            call_id, result = session._exec_skill(item)
 
         assert "already active" in result.lower()
         assert session._set_skill_called == []
@@ -352,10 +351,10 @@ class TestExecLoadSkill:
         mock_storage.list_prompt_templates.return_value = skills
 
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "search"})
+        item = session._prepare_skill("call-1", {"action": "search"})
 
         with patch("turnstone.core.storage._registry.get_storage", return_value=mock_storage):
-            call_id, result = session._exec_load_skill(item)
+            call_id, result = session._exec_skill(item)
 
         assert "enabled-skill" in result
         assert "disabled-skill" not in result
@@ -375,14 +374,114 @@ class TestExecLoadSkill:
         mock_storage.list_prompt_templates.return_value = skills
 
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "search", "query": "code review"})
+        item = session._prepare_skill("call-1", {"action": "search", "query": "code review"})
 
         with patch("turnstone.core.storage._registry.get_storage", return_value=mock_storage):
-            call_id, result = session._exec_load_skill(item)
+            call_id, result = session._exec_skill(item)
 
         assert "code-review" in result
 
     def test_preparer_load_has_approval_label(self) -> None:
         session, _, _ = _make_session()
-        item = session._prepare_load_skill("call-1", {"action": "load", "name": "my-skill"})
-        assert item["approval_label"] == "load_skill__my-skill"
+        item = session._prepare_skill("call-1", {"action": "load", "name": "my-skill"})
+        assert item["approval_label"] == "skill__my-skill"
+
+
+# ---------------------------------------------------------------------------
+# Tests: Skill Catalog Disclosure (Agent Skills standard compliance)
+# ---------------------------------------------------------------------------
+
+
+class TestSkillCatalogDisclosure:
+    """Verify <available-skills> catalog appears in system messages."""
+
+    def _build_session_with_system_messages(
+        self,
+        search_skills: list[dict[str, Any]] | None = None,
+    ) -> Any:
+        """Build a session and call _init_system_messages to get dev_parts."""
+        from turnstone.core.session import ChatSession
+
+        session = ChatSession.__new__(ChatSession)
+        ui = MagicMock()
+        session.ui = ui
+        session.model = "test-model"
+        session._ws_id = "ws-test"
+        session._node_id = "node-1"
+        session._skill_name = None
+        session._skill_content = None
+        session._skill_resources = {}
+        session._applied_skill_content = None
+        session.context_window = 128000
+        session.messages = []
+        session._config = {}
+        session.creative_mode = False
+        session.instructions = ""
+        session.system_messages = []
+        session._agent_system_messages = []
+        session.reasoning_effort = "medium"
+        session._pending_nudge = []
+        session._tool_search = None
+        session._mcp_client = None
+        session._notify_on_complete = "{}"
+        session._tool_error_flags = {}
+
+        # Memory stubs
+        session._memory_config = MagicMock()
+        session._memory_config.fetch_limit = 0
+        session._user_id = ""
+
+        with (
+            patch(
+                "turnstone.core.session.list_skills_by_activation",
+                return_value=search_skills or [],
+            ),
+            patch.object(session, "_get_visible_memories", return_value=[]),
+        ):
+            session._init_system_messages()
+
+        return session
+
+    def test_catalog_present_with_search_skills(self) -> None:
+        skills = [
+            {"name": "pdf-processing", "description": "Extract PDF text and forms."},
+            {"name": "data-analysis", "description": "Analyze datasets."},
+        ]
+        session = self._build_session_with_system_messages(search_skills=skills)
+        content = session.system_messages[0]["content"]
+        assert "<available-skills>" in content
+        assert "pdf-processing" in content
+        assert "data-analysis" in content
+        assert "</available-skills>" in content
+
+    def test_catalog_omitted_when_no_search_skills(self) -> None:
+        session = self._build_session_with_system_messages(search_skills=[])
+        content = session.system_messages[0]["content"]
+        assert "<available-skills>" not in content
+
+    def test_catalog_capped_at_30(self) -> None:
+        skills = [{"name": f"skill-{i:03d}", "description": f"Desc {i}"} for i in range(50)]
+        session = self._build_session_with_system_messages(search_skills=skills)
+        content = session.system_messages[0]["content"]
+        # Should include first 30, not all 50
+        assert "skill-029" in content
+        assert "skill-030" not in content
+
+    def test_catalog_escapes_html(self) -> None:
+        skills = [
+            {"name": "xss-test", "description": "Handle <script> & 'quotes'."},
+        ]
+        session = self._build_session_with_system_messages(search_skills=skills)
+        content = session.system_messages[0]["content"]
+        assert "&lt;script&gt;" in content
+        assert "<script>" not in content.replace("<available-skills>", "").replace(
+            "</available-skills>", ""
+        ).replace("<skill>", "").replace("</skill>", "").replace("<name>", "").replace(
+            "</name>", ""
+        ).replace("<description>", "").replace("</description>", "")
+
+    def test_catalog_includes_hint(self) -> None:
+        skills = [{"name": "test", "description": "Test skill."}]
+        session = self._build_session_with_system_messages(search_skills=skills)
+        content = session.system_messages[0]["content"]
+        assert "/skill" in content

tests/test_mcp_admin_api.py

@@ -26,6 +26,7 @@ from turnstone.console.server import (
     admin_get_mcp_server,
     admin_import_mcp_config,
     admin_list_mcp_servers,
+    admin_mcp_reload,
     admin_update_mcp_server,
 )
 from turnstone.core.auth import AuthResult
@@ -96,6 +97,11 @@ _ROUTES = [
                 admin_import_mcp_config,
                 methods=["POST"],
             ),
+            Route(
+                "/api/admin/mcp-servers/reload",
+                admin_mcp_reload,
+                methods=["POST"],
+            ),
             Route(
                 "/api/admin/mcp-servers/{server_id}",
                 admin_get_mcp_server,
@@ -115,6 +121,21 @@ _ROUTES = [
 ]
 
 
+def _routes_with_internal() -> list[Mount]:
+    """Routes including the node-side internal endpoint (lazy-imported)."""
+    from turnstone.server import internal_mcp_reload
+
+    return [
+        Mount(
+            "/v1",
+            routes=[
+                *_ROUTES[0].routes,  # type: ignore[union-attr]
+                Route("/api/_internal/mcp-reload", internal_mcp_reload, methods=["POST"]),
+            ],
+        ),
+    ]
+
+
 @pytest.fixture
 def storage(tmp_path):
     return SQLiteBackend(str(tmp_path / "test.db"))
@@ -550,8 +571,11 @@ def _fake_request(*nodes: dict[str, Any], proxy_client: Any = None) -> MagicMock
     """Build a minimal mock request with collector and proxy_client."""
     collector = MagicMock()
     collector.get_nodes.return_value = (list(nodes), len(nodes))
+    collector.get_all_nodes.side_effect = lambda: collector.get_nodes.return_value[0]
     req = MagicMock()
+    req.state.auth_result = None
     req.app.state.collector = collector
+    req.app.state.jwt_secret = ""
     req.app.state.proxy_client = proxy_client or AsyncMock()
     req.app.state.proxy_token_mgr = None
     req.app.state.proxy_auth_token = "tok"
@@ -693,3 +717,175 @@ class TestNotifyNodesMcpReload:
         result = await _notify_nodes_mcp_reload(req)
         assert result["n1"] == {"reloaded": 2}
         assert "error" in result["n2"]
+
+
+# ---------------------------------------------------------------------------
+# Console reload endpoint: POST /v1/api/admin/mcp-servers/reload
+# ---------------------------------------------------------------------------
+
+
+class TestAdminMcpReloadEndpoint:
+    """HTTP-level tests for the console reload endpoint."""
+
+    def test_reload_success(self, client: TestClient) -> None:
+        """Reload endpoint returns status ok and fan-out results."""
+        with patch(
+            "turnstone.console.server._notify_nodes_mcp_reload",
+            new_callable=AsyncMock,
+            return_value={"n1": {"reloaded": 3}},
+        ):
+            r = client.post("/v1/api/admin/mcp-servers/reload")
+        assert r.status_code == 200
+        data = r.json()
+        assert data["status"] == "ok"
+        assert data["results"] == {"n1": {"reloaded": 3}}
+
+    def test_reload_empty_cluster(self, client: TestClient) -> None:
+        """Reload with no nodes returns empty results."""
+        with patch(
+            "turnstone.console.server._notify_nodes_mcp_reload",
+            new_callable=AsyncMock,
+            return_value={},
+        ):
+            r = client.post("/v1/api/admin/mcp-servers/reload")
+        assert r.status_code == 200
+        data = r.json()
+        assert data["status"] == "ok"
+        assert data["results"] == {}
+
+    def test_reload_permission_denied(self, client_no_perm: TestClient) -> None:
+        """Reload without admin.mcp permission is rejected."""
+        r = client_no_perm.post("/v1/api/admin/mcp-servers/reload")
+        assert r.status_code == 403
+        assert "admin.mcp" in r.json()["error"]
+
+    def test_reload_no_storage(self) -> None:
+        """Reload returns 503 when auth_storage is not available."""
+        app = Starlette(
+            routes=_ROUTES,
+            middleware=[Middleware(_InjectAuthMiddleware)],
+        )
+        # Deliberately omit app.state.auth_storage
+        no_storage_client = TestClient(app, raise_server_exceptions=False)
+        r = no_storage_client.post("/v1/api/admin/mcp-servers/reload")
+        assert r.status_code == 503
+
+    def test_reload_mixed_node_results(self, client: TestClient) -> None:
+        """Reload propagates per-node errors in results."""
+        with patch(
+            "turnstone.console.server._notify_nodes_mcp_reload",
+            new_callable=AsyncMock,
+            return_value={
+                "n1": {"reloaded": 2},
+                "n2": {"error": "Connection refused"},
+            },
+        ):
+            r = client.post("/v1/api/admin/mcp-servers/reload")
+        assert r.status_code == 200
+        data = r.json()
+        assert data["results"]["n1"] == {"reloaded": 2}
+        assert "error" in data["results"]["n2"]
+
+
+# ---------------------------------------------------------------------------
+# Node reload endpoint: POST /v1/api/_internal/mcp-reload
+# ---------------------------------------------------------------------------
+
+
+class TestInternalMcpReloadEndpoint:
+    """HTTP-level tests for the node-side MCP reload endpoint."""
+
+    @pytest.fixture()
+    def node_client(self, storage: SQLiteBackend) -> TestClient:
+        """TestClient with an MCP client manager on app.state."""
+        app = Starlette(
+            routes=_routes_with_internal(),
+            middleware=[Middleware(_InjectAuthMiddleware)],
+        )
+        app.state.auth_storage = storage
+        mgr = MagicMock()
+        mgr.reconcile_sync.return_value = {
+            "added": ["new-srv"],
+            "removed": [],
+            "updated": [],
+        }
+        app.state.mcp_client = mgr
+        return TestClient(app, raise_server_exceptions=False)
+
+    def test_reload_calls_reconcile(self, node_client: TestClient, storage: SQLiteBackend) -> None:
+        """Reload endpoint calls reconcile_sync and returns its result."""
+        with patch("turnstone.core.storage._registry.get_storage", return_value=storage):
+            r = node_client.post("/v1/api/_internal/mcp-reload")
+        assert r.status_code == 200
+        data = r.json()
+        assert data["status"] == "ok"
+        assert data["added"] == ["new-srv"]
+        assert data["removed"] == []
+        assert data["updated"] == []
+
+    def test_reload_passes_storage_to_reconcile(
+        self,
+        storage: SQLiteBackend,
+    ) -> None:
+        """Verify reconcile_sync receives the storage backend."""
+        app = Starlette(
+            routes=_routes_with_internal(),
+            middleware=[Middleware(_InjectAuthMiddleware)],
+        )
+        app.state.auth_storage = storage
+        mgr = MagicMock()
+        mgr.reconcile_sync.return_value = {"added": [], "removed": [], "updated": []}
+        app.state.mcp_client = mgr
+        c = TestClient(app, raise_server_exceptions=False)
+        with patch("turnstone.core.storage._registry.get_storage", return_value=storage):
+            r = c.post("/v1/api/_internal/mcp-reload")
+        assert r.status_code == 200
+        mgr.reconcile_sync.assert_called_once_with(storage)
+
+    def test_reload_creates_manager_when_missing(self, storage: SQLiteBackend) -> None:
+        """When mcp_client is absent, a new MCPClientManager is created."""
+        app = Starlette(
+            routes=_routes_with_internal(),
+            middleware=[Middleware(_InjectAuthMiddleware)],
+        )
+        app.state.auth_storage = storage
+        # No mcp_client on app.state
+        c = TestClient(app, raise_server_exceptions=False)
+        with (
+            patch("turnstone.core.storage._registry.get_storage", return_value=storage),
+            patch("turnstone.core.mcp_client.MCPClientManager") as mock_cls,
+        ):
+            mock_mgr = MagicMock()
+            mock_mgr.reconcile_sync.return_value = {
+                "added": [],
+                "removed": [],
+                "updated": [],
+            }
+            mock_cls.return_value = mock_mgr
+            r = c.post("/v1/api/_internal/mcp-reload")
+        assert r.status_code == 200
+        mock_cls.assert_called_once_with({})
+        mock_mgr.start.assert_called_once()
+        mock_mgr.reconcile_sync.assert_called_once_with(storage)
+
+    def test_reload_reconcile_result_in_response(self, storage: SQLiteBackend) -> None:
+        """Full reconcile result fields (added/removed/updated) appear in JSON."""
+        app = Starlette(
+            routes=_routes_with_internal(),
+            middleware=[Middleware(_InjectAuthMiddleware)],
+        )
+        app.state.auth_storage = storage
+        mgr = MagicMock()
+        mgr.reconcile_sync.return_value = {
+            "added": ["a"],
+            "removed": ["b"],
+            "updated": ["c"],
+        }
+        app.state.mcp_client = mgr
+        c = TestClient(app, raise_server_exceptions=False)
+        with patch("turnstone.core.storage._registry.get_storage", return_value=storage):
+            r = c.post("/v1/api/_internal/mcp-reload")
+        data = r.json()
+        assert data["added"] == ["a"]
+        assert data["removed"] == ["b"]
+        assert data["updated"] == ["c"]

tests/test_mcp_client.py

@@ -401,6 +401,64 @@ class TestSessionIntegration:
         prepared = session._prepare_tool(tc)
         assert "error" in prepared
         assert "Unknown tool" in prepared["error"]
+        # Error lists available tools so the model can self-correct
+        assert "bash" in prepared["error"]
+        # Surfaces warning to user
+        session.ui.on_error.assert_called_once()
+        assert "nonexistent" in session.ui.on_error.call_args[0][0]
+
+    def test_prepare_tool_strips_whitespace_from_name(self, tmp_db):
+        """Local models may produce tool names with leading/trailing whitespace."""
+        session = self._make_session(mcp_client=None)
+        tc = {
+            "id": "call_strip",
+            "function": {"name": "  bash\n", "arguments": '{"command": "echo hi"}'},
+        }
+        prepared = session._prepare_tool(tc)
+        assert prepared["func_name"] == "bash"
+        assert "error" not in prepared
+
+    def test_prepare_tool_malformed_json_surfaces_error(self, tmp_db):
+        """Malformed JSON args should surface a warning to the user and
+        give the model a hint about expected format."""
+        session = self._make_session(mcp_client=None)
+        tc = {
+            "id": "call_bad",
+            "function": {"name": "bash", "arguments": "{command: echo hi}"},
+        }
+        prepared = session._prepare_tool(tc)
+        assert "error" in prepared
+        assert "JSON parse error" in prepared["error"]
+        assert "command" in prepared["error"]  # hint about expected key
+        assert "Please retry" in prepared["error"]
+        # User-facing warning
+        session.ui.on_error.assert_called_once()
+        assert "Malformed tool call" in session.ui.on_error.call_args[0][0]
+
+    def test_ensure_tool_call_ids_dict(self, tmp_db):
+        """_ensure_tool_call_ids fills empty IDs on streaming-style dict."""
+        from turnstone.core.session import ChatSession
+
+        tool_calls_acc = {
+            0: {"id": "", "function": {"name": "bash", "arguments": "{}"}},
+            1: {"id": "", "function": {"name": "read_file", "arguments": "{}"}},
+        }
+        ChatSession._ensure_tool_call_ids(tool_calls_acc)
+        ids = [tc["id"] for tc in tool_calls_acc.values()]
+        assert all(id_.startswith("call_") for id_ in ids)
+        assert len(set(ids)) == 2  # unique
+
+    def test_ensure_tool_call_ids_list(self, tmp_db):
+        """_ensure_tool_call_ids fills empty IDs on list (agent path)."""
+        from turnstone.core.session import ChatSession
+
+        tool_calls = [
+            {"id": None, "function": {"name": "bash", "arguments": "{}"}},
+            {"id": "call_existing", "function": {"name": "bash", "arguments": "{}"}},
+        ]
+        ChatSession._ensure_tool_call_ids(tool_calls)
+        assert tool_calls[0]["id"].startswith("call_")
+        assert tool_calls[1]["id"] == "call_existing"  # preserved
 
     def test_mcp_command_no_client(self, tmp_db):
         session = self._make_session(mcp_client=None)
@@ -1368,3 +1426,130 @@ class TestShutdownCleanup:
         assert mgr.get_prompts() == []
         assert mgr._resource_map == {}
         assert mgr._prompt_map == {}
+
+
+# ---------------------------------------------------------------------------
+# TCP probe and unreachable server handling
+# ---------------------------------------------------------------------------
+
+
+class TestTCPProbe:
+    """MCPClientManager._tcp_probe should fail fast on unreachable servers."""
+
+    def test_tcp_probe_unreachable_raises_connection_error(self):
+        """Unreachable host raises ConnectionError, not TimeoutError."""
+        mgr = MCPClientManager({})
+
+        async def _run():
+            with pytest.raises(ConnectionError, match="unreachable"):
+                await mgr._tcp_probe("test-server", "http://127.0.0.1:1")
+
+        asyncio.run(_run())
+
+    def test_tcp_probe_parses_url_correctly(self):
+        """Port and host are extracted from the URL."""
+        mgr = MCPClientManager({})
+
+        async def _run():
+            # Non-routable port — should fail with ConnectionError
+            with pytest.raises(ConnectionError):
+                await mgr._tcp_probe("srv", "https://127.0.0.1:1/mcp")
+
+        asyncio.run(_run())
+
+    def test_tcp_probe_default_port_http(self):
+        """Default port 80 used for http:// URLs without explicit port."""
+        mgr = MCPClientManager({})
+
+        async def _run():
+            # Will fail (nothing on port 80), but should not crash on parsing
+            with pytest.raises(ConnectionError):
+                await mgr._tcp_probe("srv", "http://127.0.0.1")
+
+        asyncio.run(_run())
+
+    def test_tcp_probe_dns_failure(self):
+        """Unresolvable hostname raises ConnectionError."""
+        mgr = MCPClientManager({})
+
+        async def _run():
+            with pytest.raises(ConnectionError):
+                await mgr._tcp_probe("srv", "http://this.host.does.not.exist.invalid:8080/mcp")
+
+        asyncio.run(_run())
+
+
+class TestConnectOneUnreachable:
+    """_connect_one should handle unreachable HTTP servers gracefully."""
+
+    def test_unreachable_http_server_raises_connection_error(self):
+        """Unreachable HTTP MCP server raises ConnectionError without spinning."""
+        mgr = MCPClientManager({})
+        mgr._loop = asyncio.new_event_loop()
+
+        async def _run():
+            with pytest.raises(ConnectionError, match="unreachable"):
+                await mgr._connect_one(
+                    "bad-server",
+                    {
+                        "type": "http",
+                        "url": "http://127.0.0.1:1/mcp",
+                    },
+                )
+
+        mgr._loop.run_until_complete(_run())
+        mgr._loop.close()
+
+        # Server should NOT be in sessions (connection failed)
+        assert "bad-server" not in mgr._sessions
+
+    def test_connect_all_continues_after_unreachable_server(self):
+        """_connect_all logs error and continues to next server."""
+        mgr = MCPClientManager(
+            {
+                "bad": {"type": "http", "url": "http://127.0.0.1:1/mcp"},
+            }
+        )
+
+        loop = asyncio.new_event_loop()
+        loop.run_until_complete(mgr._connect_all())
+        loop.close()
+
+        assert "bad" not in mgr._sessions
+        assert "bad" in mgr._last_error
+
+
+class TestSafeCloseStack:
+    """_safe_close_stack should suppress errors from broken anyio scopes."""
+
+    def test_suppresses_runtime_error(self):
+        """RuntimeError from broken cancel scope is suppressed."""
+
+        async def _run():
+            stack = AsyncExitStack()
+            await stack.__aenter__()
+
+            # Simulate a broken close that raises RuntimeError
+            async def _broken_close():
+                raise RuntimeError("Attempted to exit cancel scope in a different task")
+
+            stack.aclose = _broken_close
+            # Should not raise
+            await MCPClientManager._safe_close_stack(stack)
+
+        asyncio.run(_run())
+
+    def test_suppresses_cancelled_error(self):
+        """CancelledError during close is suppressed."""
+
+        async def _run():
+            stack = AsyncExitStack()
+            await stack.__aenter__()
+
+            async def _cancel_close():
+                raise asyncio.CancelledError()
+
+            stack.aclose = _cancel_close
+            await MCPClientManager._safe_close_stack(stack)
+
+        asyncio.run(_run())

tests/test_mcp_registry.py

@@ -398,6 +398,72 @@ class TestResolveInstallConfig:
         config = resolve_install_config(server, "remote", 0)
         assert config["url"] == "https://us-east.example.com/mcp"
 
+    def test_remote_variable_substitution_invalid_scheme(self) -> None:
+        server = RegistryServer(
+            name="io.example/test",
+            version="1.0.0",
+            remotes=[
+                RegistryRemote(
+                    type="streamable-http",
+                    url="{scheme}://evil.example.com/mcp",
+                    variables={
+                        "scheme": RegistryRemoteVariable(is_required=True),
+                    },
+                )
+            ],
+        )
+        with pytest.raises(MCPRegistryError, match="Invalid URL scheme"):
+            resolve_install_config(server, "remote", 0, variables={"scheme": "file"})
+
+    def test_remote_variable_substitution_preserves_valid_scheme(self) -> None:
+        server = RegistryServer(
+            name="io.example/test",
+            version="1.0.0",
+            remotes=[
+                RegistryRemote(
+                    type="streamable-http",
+                    url="https://{host}.example.com/mcp",
+                    variables={
+                        "host": RegistryRemoteVariable(is_required=True),
+                    },
+                )
+            ],
+        )
+        config = resolve_install_config(server, "remote", 0, variables={"host": "api"})
+        assert config["url"] == "https://api.example.com/mcp"
+
+    def test_remote_variable_substitution_missing_hostname(self) -> None:
+        """URL like https:///mcp has valid scheme but no hostname."""
+        server = RegistryServer(
+            name="io.example/test",
+            version="1.0.0",
+            remotes=[
+                RegistryRemote(
+                    type="streamable-http",
+                    url="https:///mcp",
+                )
+            ],
+        )
+        with pytest.raises(MCPRegistryError, match="hostname is missing"):
+            resolve_install_config(server, "remote", 0)
+
+    def test_remote_variable_substitution_embedded_credentials(self) -> None:
+        server = RegistryServer(
+            name="io.example/test",
+            version="1.0.0",
+            remotes=[
+                RegistryRemote(
+                    type="streamable-http",
+                    url="https://{creds}@example.com/mcp",
+                    variables={
+                        "creds": RegistryRemoteVariable(is_required=True),
+                    },
+                )
+            ],
+        )
+        with pytest.raises(MCPRegistryError, match="embedded credentials"):
+            resolve_install_config(server, "remote", 0, variables={"creds": "user:pass"})
+
     def test_remote_no_remotes(self) -> None:
         server = RegistryServer(name="io.example/test", version="1.0.0")
         with pytest.raises(MCPRegistryError, match="no remote"):

tests/test_mcp_registry_api.py

@@ -4,7 +4,7 @@ from __future__ import annotations
 
 import json
 from typing import TYPE_CHECKING, Any
-from unittest.mock import AsyncMock, patch
+from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 from starlette.applications import Starlette
@@ -18,11 +18,13 @@ if TYPE_CHECKING:
     from starlette.responses import Response
 
 from turnstone.console.server import (
+    _get_registry_url,
     admin_registry_install,
     admin_registry_search,
 )
 from turnstone.core.auth import AuthResult
 from turnstone.core.mcp_registry import (
+    DEFAULT_REGISTRY_URL,
     MCPRegistryError,
     RegistryPackage,
     RegistryRemote,
@@ -362,7 +364,7 @@ class TestRegistryInstall:
     def test_install_max_servers(self, client: TestClient, storage: SQLiteBackend) -> None:
         import uuid
 
-        for i in range(50):
+        for i in range(200):
             storage.create_mcp_server(
                 server_id=uuid.uuid4().hex,
                 name=f"server-{i}",
@@ -549,3 +551,73 @@ class TestRegistryInstall:
 
         assert resp.status_code == 409
         assert "custom 'name'" in resp.json()["error"]
+
+
+# ---------------------------------------------------------------------------
+# _get_registry_url fallback chain tests
+# ---------------------------------------------------------------------------
+
+
+def _mock_request(storage: Any = None, config_store: Any = None) -> MagicMock:
+    """Build a mock Request with app.state.auth_storage and app.state.config_store."""
+    request = MagicMock()
+    request.app.state.auth_storage = storage
+    request.app.state.config_store = config_store
+    return request
+
+
+class TestGetRegistryUrl:
+    """Verify three-tier URL resolution: DB setting -> config.toml -> default."""
+
+    def test_returns_db_setting_when_available(self) -> None:
+        config_store = MagicMock()
+        config_store.get.return_value = "https://custom.registry.example.com"
+        request = _mock_request(config_store=config_store)
+
+        with patch("turnstone.core.config.load_config", return_value={}):
+            url = _get_registry_url(request)
+
+        assert url == "https://custom.registry.example.com"
+        config_store.get.assert_called_once_with("mcp.registry_url")
+
+    def test_falls_back_to_config_when_config_store_returns_empty(self) -> None:
+        config_store = MagicMock()
+        config_store.get.return_value = ""
+        request = _mock_request(config_store=config_store)
+
+        with patch(
+            "turnstone.core.config.load_config",
+            return_value={"registry_url": "https://config.registry.example.com"},
+        ):
+            url = _get_registry_url(request)
+
+        assert url == "https://config.registry.example.com"
+
+    def test_falls_back_to_config_when_no_config_store(self) -> None:
+        request = _mock_request()
+
+        with patch(
+            "turnstone.core.config.load_config",
+            return_value={"registry_url": "https://config.registry.example.com"},
+        ):
+            url = _get_registry_url(request)
+
+        assert url == "https://config.registry.example.com"
+
+    def test_falls_back_to_default_when_both_unavailable(self) -> None:
+        config_store = MagicMock()
+        config_store.get.return_value = ""
+        request = _mock_request(config_store=config_store)
+
+        with patch("turnstone.core.config.load_config", return_value={}):
+            url = _get_registry_url(request)
+
+        assert url == DEFAULT_REGISTRY_URL
+
+    def test_falls_back_to_default_when_no_config_store_or_config(self) -> None:
+        request = _mock_request()
+
+        with patch("turnstone.core.config.load_config", return_value={}):
+            url = _get_registry_url(request)
+
+        assert url == DEFAULT_REGISTRY_URL

tests/test_mcp_server_storage.py

@@ -3,16 +3,10 @@
 from __future__ import annotations
 
 import uuid
+from typing import TYPE_CHECKING
 
-import pytest
-
-from turnstone.core.storage._sqlite import SQLiteBackend
-
-
-@pytest.fixture
-def db(tmp_path):
-    """Fresh SQLite backend for each test."""
-    return SQLiteBackend(str(tmp_path / "test.db"))
+if TYPE_CHECKING:
+    from turnstone.core.storage._sqlite import SQLiteBackend
 
 
 def _make_id() -> str:

tests/test_metacognition.py

@@ -4,6 +4,7 @@ from turnstone.core.metacognition import (
     NUDGE_COMPLETION,
     NUDGE_CORRECTION,
     NUDGE_DENIAL,
+    NUDGE_REPEAT,
     NUDGE_RESUME,
     NUDGE_START,
     NUDGE_TOOL_ERROR,
@@ -288,3 +289,22 @@ class TestToolErrorNudge:
     def test_not_with_zero_memories(self):
         state: dict[str, float] = {}
         assert should_nudge("tool_error", state, message_count=5, memory_count=0) is False
+
+
+class TestRepeatNudge:
+    def test_format(self):
+        assert format_nudge("repeat") == NUDGE_REPEAT
+
+    def test_fires(self):
+        state: dict[str, float] = {}
+        assert should_nudge("repeat", state, message_count=5) is True
+
+    def test_cooldown(self):
+        state: dict[str, float] = {}
+        assert should_nudge("repeat", state, message_count=5) is True
+        assert should_nudge("repeat", state, message_count=6) is False
+
+    def test_no_memory_requirement(self):
+        """Repeat nudge should fire even with zero memories."""
+        state: dict[str, float] = {}
+        assert should_nudge("repeat", state, message_count=5, memory_count=0) is True

tests/test_model_definition_storage.py

@@ -0,0 +1,163 @@
+"""Tests for model definition storage CRUD operations."""
+
+from __future__ import annotations
+
+import uuid
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from turnstone.core.storage._sqlite import SQLiteBackend
+
+
+def _make_id() -> str:
+    return uuid.uuid4().hex
+
+
+class TestModelDefinitionStorage:
+    def test_create_and_get(self, db: SQLiteBackend) -> None:
+        did = _make_id()
+        db.create_model_definition(
+            definition_id=did,
+            alias="test-model",
+            model="gpt-5",
+            provider="openai",
+            base_url="https://api.openai.com/v1",
+            api_key="sk-test",
+            context_window=128000,
+        )
+        m = db.get_model_definition(did)
+        assert m is not None
+        assert m["alias"] == "test-model"
+        assert m["model"] == "gpt-5"
+        assert m["provider"] == "openai"
+        assert m["base_url"] == "https://api.openai.com/v1"
+        assert m["api_key"] == "sk-test"
+        assert m["context_window"] == 128000
+        assert m["capabilities"] == "{}"
+        assert m["enabled"] is True
+
+    def test_get_by_alias(self, db: SQLiteBackend) -> None:
+        did = _make_id()
+        db.create_model_definition(definition_id=did, alias="by-alias", model="gpt-5")
+        m = db.get_model_definition_by_alias("by-alias")
+        assert m is not None
+        assert m["definition_id"] == did
+
+    def test_get_by_alias_not_found(self, db: SQLiteBackend) -> None:
+        assert db.get_model_definition_by_alias("nope") is None
+
+    def test_get_not_found(self, db: SQLiteBackend) -> None:
+        assert db.get_model_definition("nonexistent") is None
+
+    def test_list_empty(self, db: SQLiteBackend) -> None:
+        assert db.list_model_definitions() == []
+
+    def test_list_all(self, db: SQLiteBackend) -> None:
+        db.create_model_definition(definition_id=_make_id(), alias="alpha", model="gpt-5")
+        db.create_model_definition(
+            definition_id=_make_id(), alias="beta", model="claude-opus-4-6", provider="anthropic"
+        )
+        models = db.list_model_definitions()
+        assert len(models) == 2
+        assert models[0]["alias"] == "alpha"  # ordered by alias
+        assert models[1]["alias"] == "beta"
+
+    def test_list_enabled_only(self, db: SQLiteBackend) -> None:
+        db.create_model_definition(
+            definition_id=_make_id(), alias="enabled-model", model="gpt-5", enabled=True
+        )
+        db.create_model_definition(
+            definition_id=_make_id(), alias="disabled-model", model="gpt-5", enabled=False
+        )
+        enabled = db.list_model_definitions(enabled_only=True)
+        assert len(enabled) == 1
+        assert enabled[0]["alias"] == "enabled-model"
+
+    def test_update_basic_fields(self, db: SQLiteBackend) -> None:
+        did = _make_id()
+        db.create_model_definition(
+            definition_id=did, alias="orig", model="gpt-5", base_url="http://old"
+        )
+        ok = db.update_model_definition(did, alias="renamed", base_url="http://new")
+        assert ok is True
+        m = db.get_model_definition(did)
+        assert m is not None
+        assert m["alias"] == "renamed"
+        assert m["base_url"] == "http://new"
+
+    def test_update_boolean_conversion(self, db: SQLiteBackend) -> None:
+        did = _make_id()
+        db.create_model_definition(definition_id=did, alias="booltest", model="gpt-5")
+        db.update_model_definition(did, enabled=False)
+        m = db.get_model_definition(did)
+        assert m is not None
+        assert m["enabled"] is False
+
+    def test_update_not_found(self, db: SQLiteBackend) -> None:
+        ok = db.update_model_definition("nonexistent", alias="x")
+        assert ok is False
+
+    def test_update_ignores_disallowed_fields(self, db: SQLiteBackend) -> None:
+        did = _make_id()
+        db.create_model_definition(
+            definition_id=did, alias="guard", model="gpt-5", created_by="admin"
+        )
+        original = db.get_model_definition(did)
+        assert original is not None
+        original_created = original["created"]
+        # created_by and created are not in the mutable allowlist
+        db.update_model_definition(did, created_by="evil", created="2000-01-01T00:00:00")
+        m = db.get_model_definition(did)
+        assert m is not None
+        assert m["created_by"] == "admin"  # unchanged
+        assert m["created"] == original_created  # unchanged
+
+    def test_delete(self, db: SQLiteBackend) -> None:
+        did = _make_id()
+        db.create_model_definition(definition_id=did, alias="delme", model="gpt-5")
+        ok = db.delete_model_definition(did)
+        assert ok is True
+        assert db.get_model_definition(did) is None
+
+    def test_delete_not_found(self, db: SQLiteBackend) -> None:
+        ok = db.delete_model_definition("nonexistent")
+        assert ok is False
+
+    def test_create_duplicate_alias(self, db: SQLiteBackend) -> None:
+        db.create_model_definition(definition_id=_make_id(), alias="unique", model="gpt-5")
+        # Second create with same alias but different ID should be no-op (OR IGNORE)
+        did2 = _make_id()
+        db.create_model_definition(definition_id=did2, alias="unique", model="gpt-5")
+        assert db.get_model_definition(did2) is None
+
+    def test_create_idempotent_same_id(self, db: SQLiteBackend) -> None:
+        did = _make_id()
+        db.create_model_definition(definition_id=did, alias="idem", model="gpt-5")
+        db.create_model_definition(definition_id=did, alias="idem", model="gpt-5-mini")
+        m = db.get_model_definition(did)
+        assert m is not None
+        assert m["model"] == "gpt-5"  # original preserved
+
+    def test_capabilities_json(self, db: SQLiteBackend) -> None:
+        did = _make_id()
+        caps = '{"supports_vision": true, "supports_web_search": false}'
+        db.create_model_definition(
+            definition_id=did, alias="caps-test", model="gpt-5", capabilities=caps
+        )
+        m = db.get_model_definition(did)
+        assert m is not None
+        assert m["capabilities"] == caps
+
+    def test_defaults(self, db: SQLiteBackend) -> None:
+        """Verify default values for optional fields."""
+        did = _make_id()
+        db.create_model_definition(definition_id=did, alias="defaults", model="gpt-5")
+        m = db.get_model_definition(did)
+        assert m is not None
+        assert m["provider"] == "openai"
+        assert m["base_url"] == ""
+        assert m["api_key"] == ""
+        assert m["context_window"] == 32768
+        assert m["capabilities"] == "{}"
+        assert m["enabled"] is True
+        assert m["created_by"] == ""

tests/test_model_probe.py

@@ -0,0 +1,235 @@
+"""Tests for probe_model_endpoint() and lookup_model_capabilities()."""
+
+from __future__ import annotations
+
+from typing import Any
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from turnstone.core.model_registry import probe_model_endpoint
+from turnstone.core.providers import list_known_models, lookup_model_capabilities
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _mock_model(
+    model_id: str,
+    *,
+    owned_by: str = "test",
+    meta: dict[str, Any] | None = None,
+) -> MagicMock:
+    m = MagicMock()
+    m.id = model_id
+    dumped: dict[str, Any] = {"owned_by": owned_by}
+    if meta is not None:
+        dumped["meta"] = meta
+    m.model_dump.return_value = dumped
+    return m
+
+
+def _mock_client(*models: MagicMock) -> MagicMock:
+    fast = MagicMock()
+    fast.models.list.return_value = MagicMock(data=list(models))
+    client = MagicMock()
+    client.with_options.return_value = fast
+    return client
+
+
+# ---------------------------------------------------------------------------
+# probe_model_endpoint
+# ---------------------------------------------------------------------------
+
+
+class TestProbeModelEndpoint:
+    @patch("turnstone.core.providers.create_client")
+    def test_probe_success(self, mock_cc: MagicMock) -> None:
+        m1 = _mock_model("model-a")
+        m2 = _mock_model("model-b")
+        mock_cc.return_value = _mock_client(m1, m2)
+
+        result = probe_model_endpoint("openai", "http://localhost:8000/v1", "key")
+        assert result["reachable"] is True
+        assert result["available_models"] == ["model-a", "model-b"]
+        assert result["error"] is None
+
+    @patch("turnstone.core.providers.create_client")
+    def test_target_found(self, mock_cc: MagicMock) -> None:
+        m1 = _mock_model("gpt-5")
+        mock_cc.return_value = _mock_client(m1)
+
+        result = probe_model_endpoint(
+            "openai", "http://localhost:8000/v1", "key", target_model="gpt-5"
+        )
+        assert result["model_found"] is True
+
+    @patch("turnstone.core.providers.create_client")
+    def test_target_not_found(self, mock_cc: MagicMock) -> None:
+        m1 = _mock_model("model-a")
+        mock_cc.return_value = _mock_client(m1)
+
+        result = probe_model_endpoint(
+            "openai", "http://localhost:8000/v1", "key", target_model="gpt-5"
+        )
+        assert result["model_found"] is False
+        assert result["available_models"] == ["model-a"]
+
+    @patch("turnstone.core.providers.create_client")
+    def test_no_target_model_found_is_none(self, mock_cc: MagicMock) -> None:
+        m1 = _mock_model("model-a")
+        mock_cc.return_value = _mock_client(m1)
+
+        result = probe_model_endpoint("openai", "http://localhost:8000/v1", "key")
+        assert result["model_found"] is None
+
+    @patch("turnstone.core.providers.create_client")
+    def test_context_window_llama_cpp(self, mock_cc: MagicMock) -> None:
+        m = _mock_model("qwen-32b", meta={"n_ctx_train": 131072})
+        mock_cc.return_value = _mock_client(m)
+
+        result = probe_model_endpoint("openai", "http://localhost:8000/v1", "key")
+        assert result["context_window"] == 131072
+        assert result["server_type"] == "llama.cpp"
+
+    @patch("turnstone.core.providers.create_client")
+    def test_server_type_openai(self, mock_cc: MagicMock) -> None:
+        m = _mock_model("gpt-5")
+        mock_cc.return_value = _mock_client(m)
+
+        result = probe_model_endpoint("openai", "https://api.openai.com/v1", "sk-test")
+        assert result["server_type"] == "openai"
+
+    @patch("turnstone.core.providers.create_client")
+    def test_server_type_sglang(self, mock_cc: MagicMock) -> None:
+        m = _mock_model("meta-llama/Llama-3", owned_by="sglang")
+        mock_cc.return_value = _mock_client(m)
+
+        result = probe_model_endpoint("openai", "http://localhost:30000/v1", "key")
+        assert result["server_type"] == "sglang"
+
+    @patch("turnstone.core.providers.create_client")
+    def test_server_type_vllm(self, mock_cc: MagicMock) -> None:
+        m = _mock_model("org/model-name")
+        mock_cc.return_value = _mock_client(m)
+
+        result = probe_model_endpoint("openai", "http://localhost:8000/v1", "key")
+        assert result["server_type"] == "vllm"
+
+    @patch("turnstone.core.providers.create_client")
+    def test_server_type_generic(self, mock_cc: MagicMock) -> None:
+        m = _mock_model("my-model")
+        mock_cc.return_value = _mock_client(m)
+
+        result = probe_model_endpoint("openai", "http://localhost:8000/v1", "key")
+        assert result["server_type"] == "openai-compatible"
+
+    @patch("turnstone.core.providers.create_client")
+    def test_anthropic_provider(self, mock_cc: MagicMock) -> None:
+        m = _mock_model("claude-sonnet-4-6")
+        mock_cc.return_value = _mock_client(m)
+
+        result = probe_model_endpoint(
+            "anthropic",
+            "https://api.anthropic.com",
+            "sk-ant-test",
+            target_model="claude-sonnet-4-6",
+        )
+        assert result["reachable"] is True
+        assert result["server_type"] == "anthropic"
+        assert result["context_window"] == 1000000
+
+    @patch("turnstone.core.providers.create_client")
+    def test_connection_failure(self, mock_cc: MagicMock) -> None:
+        mock_cc.side_effect = OSError("Connection refused")
+
+        result = probe_model_endpoint("openai", "http://bad:1234/v1", "key")
+        assert result["reachable"] is False
+        assert "Connection refused" in (result["error"] or "")
+
+    @patch("turnstone.core.providers.create_client")
+    def test_empty_model_list(self, mock_cc: MagicMock) -> None:
+        mock_cc.return_value = _mock_client()  # no models
+
+        result = probe_model_endpoint("openai", "http://localhost:8000/v1", "key")
+        assert result["reachable"] is True
+        assert result["available_models"] == []
+        assert "No models found" in (result["error"] or "")
+
+    @patch("turnstone.core.providers.create_client")
+    def test_context_window_openai_static_table(self, mock_cc: MagicMock) -> None:
+        """When base_url is api.openai.com and model is known, use static table."""
+        m = _mock_model("gpt-5")
+        mock_cc.return_value = _mock_client(m)
+
+        result = probe_model_endpoint(
+            "openai", "https://api.openai.com/v1", "sk-test", target_model="gpt-5"
+        )
+        assert result["context_window"] == 400000
+
+
+# ---------------------------------------------------------------------------
+# lookup_model_capabilities
+# ---------------------------------------------------------------------------
+
+
+class TestLookupModelCapabilities:
+    def test_known_openai_model(self) -> None:
+        caps = lookup_model_capabilities("openai", "gpt-5")
+        assert caps is not None
+        assert caps["context_window"] == 400000
+        assert caps["supports_temperature"] is False
+
+    def test_known_anthropic_model(self) -> None:
+        caps = lookup_model_capabilities("anthropic", "claude-opus-4-6")
+        assert caps is not None
+        assert caps["context_window"] == 1000000
+        assert caps["thinking_mode"] == "adaptive"
+
+    def test_unknown_model_returns_none(self) -> None:
+        caps = lookup_model_capabilities("openai", "totally-unknown-model")
+        assert caps is None
+
+    def test_tuples_converted_to_lists(self) -> None:
+        caps = lookup_model_capabilities("openai", "gpt-5")
+        assert caps is not None
+        for val in caps.values():
+            assert not isinstance(val, tuple), f"Found tuple: {val}"
+
+    def test_reasoning_effort_values_are_list(self) -> None:
+        caps = lookup_model_capabilities("openai", "gpt-5")
+        assert caps is not None
+        assert isinstance(caps["reasoning_effort_values"], list)
+        assert "medium" in caps["reasoning_effort_values"]
+
+    def test_openai_compatible_returns_none(self) -> None:
+        caps = lookup_model_capabilities("openai-compatible", "my-local-model")
+        assert caps is None
+
+    def test_invalid_provider_raises(self) -> None:
+        with pytest.raises(ValueError, match="Unknown provider"):
+            lookup_model_capabilities("bad-provider", "gpt-5")
+
+
+# ---------------------------------------------------------------------------
+# list_known_models
+# ---------------------------------------------------------------------------
+
+
+class TestListKnownModels:
+    def test_openai_models(self) -> None:
+        models = list_known_models("openai")
+        assert "gpt-5" in models
+        assert isinstance(models, list)
+        assert models == sorted(models)
+
+    def test_anthropic_models(self) -> None:
+        models = list_known_models("anthropic")
+        assert "claude-opus-4-6" in models
+
+    def test_openai_compatible_returns_empty(self) -> None:
+        assert list_known_models("openai-compatible") == []
+
+    def test_unknown_provider_returns_empty(self) -> None:
+        assert list_known_models("bad-provider") == []

tests/test_model_registry.py

@@ -10,6 +10,8 @@ import pytest
 from turnstone.core.model_registry import (
     ModelConfig,
     ModelRegistry,
+    _resolve_env_vars,
+    detect_model,
     load_model_registry,
 )
 
@@ -318,6 +320,280 @@ class TestLoadModelRegistry:
         assert alt_cfg.api_key == "my-key"
 
 
+# ---------------------------------------------------------------------------
+# load_model_registry with DB storage
+# ---------------------------------------------------------------------------
+
+
+class _MockStorage:
+    """Minimal storage mock returning canned model definitions."""
+
+    def __init__(self, rows: list[dict[str, Any]] | None = None) -> None:
+        self._rows = rows or []
+        self.calls: list[str] = []
+
+    def list_model_definitions(self, enabled_only: bool = False) -> list[dict[str, Any]]:
+        self.calls.append("list_model_definitions")
+        if enabled_only:
+            return [r for r in self._rows if r.get("enabled", True)]
+        return list(self._rows)
+
+
+class TestLoadModelRegistryWithDB:
+    def test_db_models_loaded(self) -> None:
+        """DB model definitions are loaded into the registry."""
+        storage = _MockStorage(
+            [
+                {
+                    "alias": "cloud-gpt",
+                    "model": "gpt-5",
+                    "provider": "openai",
+                    "base_url": "https://api.openai.com/v1",
+                    "api_key": "sk-db",
+                    "context_window": 128000,
+                    "capabilities": "{}",
+                    "enabled": True,
+                }
+            ]
+        )
+        with patch("turnstone.core.model_registry.load_config", return_value={}):
+            reg = load_model_registry("http://x/v1", "x", "x", storage=storage)
+        assert reg.has_alias("cloud-gpt")
+        cfg = reg.get_config("cloud-gpt")
+        assert cfg.model == "gpt-5"
+        assert cfg.source == "db"
+
+    def test_config_overrides_db(self) -> None:
+        """Config.toml entry overrides DB entry with same alias."""
+        storage = _MockStorage(
+            [
+                {
+                    "alias": "shared",
+                    "model": "db-model",
+                    "provider": "openai",
+                    "base_url": "http://db/v1",
+                    "api_key": "sk-db",
+                    "context_window": 32768,
+                    "capabilities": "{}",
+                    "enabled": True,
+                }
+            ]
+        )
+        fake_cfg: dict[str, Any] = {
+            "models": {
+                "shared": {
+                    "model": "config-model",
+                    "base_url": "http://config/v1",
+                },
+            },
+        }
+        with patch("turnstone.core.model_registry.load_config", return_value=fake_cfg):
+            reg = load_model_registry("http://x/v1", "x", "x", storage=storage)
+        cfg = reg.get_config("shared")
+        assert cfg.model == "config-model"
+        assert cfg.source == "config"
+
+    def test_db_only_models_coexist(self) -> None:
+        """DB models coexist alongside config.toml models."""
+        storage = _MockStorage(
+            [
+                {
+                    "alias": "db-only",
+                    "model": "db-model",
+                    "provider": "anthropic",
+                    "base_url": "",
+                    "api_key": "sk-db",
+                    "context_window": 200000,
+                    "capabilities": "{}",
+                    "enabled": True,
+                }
+            ]
+        )
+        fake_cfg: dict[str, Any] = {
+            "models": {
+                "config-only": {"model": "config-model"},
+            },
+        }
+        with patch("turnstone.core.model_registry.load_config", return_value=fake_cfg):
+            reg = load_model_registry("http://x/v1", "x", "x", storage=storage)
+        assert reg.has_alias("db-only")
+        assert reg.has_alias("config-only")
+        assert reg.has_alias("default")
+        assert reg.get_config("db-only").source == "db"
+        assert reg.get_config("config-only").source == "config"
+
+    def test_source_field_set(self) -> None:
+        """Source field correctly distinguishes origin."""
+        storage = _MockStorage(
+            [
+                {
+                    "alias": "from-db",
+                    "model": "m",
+                    "provider": "openai",
+                    "base_url": "",
+                    "api_key": "",
+                    "context_window": 32768,
+                    "capabilities": "{}",
+                    "enabled": True,
+                }
+            ]
+        )
+        with patch("turnstone.core.model_registry.load_config", return_value={}):
+            reg = load_model_registry("http://x/v1", "x", "x", storage=storage)
+        assert reg.get_config("from-db").source == "db"
+        assert reg.get_config("default").source == ""
+
+    def test_disabled_db_models_excluded(self) -> None:
+        """Disabled DB models are not loaded."""
+        storage = _MockStorage(
+            [
+                {
+                    "alias": "disabled",
+                    "model": "m",
+                    "provider": "openai",
+                    "base_url": "",
+                    "api_key": "",
+                    "context_window": 32768,
+                    "capabilities": "{}",
+                    "enabled": False,
+                }
+            ]
+        )
+        with patch("turnstone.core.model_registry.load_config", return_value={}):
+            reg = load_model_registry("http://x/v1", "x", "x", storage=storage)
+        assert not reg.has_alias("disabled")
+
+    def test_db_capabilities_parsed(self) -> None:
+        """JSON capabilities from DB are parsed into dict."""
+        storage = _MockStorage(
+            [
+                {
+                    "alias": "caps-model",
+                    "model": "m",
+                    "provider": "openai",
+                    "base_url": "",
+                    "api_key": "",
+                    "context_window": 32768,
+                    "capabilities": '{"supports_vision": true}',
+                    "enabled": True,
+                }
+            ]
+        )
+        with patch("turnstone.core.model_registry.load_config", return_value={}):
+            reg = load_model_registry("http://x/v1", "x", "x", storage=storage)
+        assert reg.get_config("caps-model").capabilities == {"supports_vision": True}
+
+    def test_db_default_alias_not_clobbered(self) -> None:
+        """DB model with alias='default' is not overwritten by CLI args."""
+        storage = _MockStorage(
+            [
+                {
+                    "alias": "default",
+                    "model": "db-default-model",
+                    "provider": "openai",
+                    "base_url": "http://db/v1",
+                    "api_key": "sk-db",
+                    "context_window": 128000,
+                    "capabilities": "{}",
+                    "enabled": True,
+                }
+            ]
+        )
+        with patch("turnstone.core.model_registry.load_config", return_value={}):
+            reg = load_model_registry("http://cli/v1", "cli-key", "cli-model", storage=storage)
+        cfg = reg.get_config("default")
+        assert cfg.model == "db-default-model"
+        assert cfg.source == "db"
+
+    def test_no_db_writes(self) -> None:
+        """Config.toml models are NOT written to storage."""
+        storage = _MockStorage()
+        fake_cfg: dict[str, Any] = {
+            "models": {"local": {"model": "llama"}},
+        }
+        with patch("turnstone.core.model_registry.load_config", return_value=fake_cfg):
+            load_model_registry("http://x/v1", "x", "x", storage=storage)
+        # Only list_model_definitions should be called, no create
+        assert storage.calls == ["list_model_definitions"]
+
+    def test_storage_failure_graceful(self) -> None:
+        """Storage errors don't prevent registry creation."""
+        storage = MagicMock()
+        storage.list_model_definitions.side_effect = RuntimeError("db down")
+        with patch("turnstone.core.model_registry.load_config", return_value={}):
+            reg = load_model_registry("http://x/v1", "x", "x", storage=storage)
+        assert reg.has_alias("default")
+
+
+# ---------------------------------------------------------------------------
+# _resolve_env_vars
+# ---------------------------------------------------------------------------
+
+
+class TestResolveEnvVars:
+    def test_expand_single(self) -> None:
+        with patch.dict("os.environ", {"MY_KEY": "secret123"}):
+            assert _resolve_env_vars("sk-${MY_KEY}") == "sk-secret123"
+
+    def test_expand_multiple(self) -> None:
+        with patch.dict("os.environ", {"A": "1", "B": "2"}):
+            assert _resolve_env_vars("${A}-${B}") == "1-2"
+
+    def test_missing_var_empty(self) -> None:
+        with patch.dict("os.environ", {}, clear=True):
+            assert _resolve_env_vars("${MISSING}") == ""
+
+    def test_no_vars(self) -> None:
+        assert _resolve_env_vars("plain-key") == "plain-key"
+
+    def test_empty_string(self) -> None:
+        assert _resolve_env_vars("") == ""
+
+
+# ---------------------------------------------------------------------------
+# ModelRegistry.reload
+# ---------------------------------------------------------------------------
+
+
+class TestRegistryReload:
+    def test_reload_replaces_models(self) -> None:
+        models_a = {"a": ModelConfig("a", "x", "x", "m1")}
+        reg = ModelRegistry(models=models_a, default="a")
+        assert reg.has_alias("a")
+
+        models_b = {"b": ModelConfig("b", "y", "y", "m2")}
+        reg.reload(models_b, "b")
+        assert not reg.has_alias("a")
+        assert reg.has_alias("b")
+        assert reg.default == "b"
+
+    def test_reload_clears_clients(self) -> None:
+        models = {"a": ModelConfig("a", "http://x/v1", "key", "m")}
+        reg = ModelRegistry(models=models, default="a")
+        # Force client creation
+        reg.get_client("a")
+        assert "a" in reg._clients
+
+        # Reload with same models — clients should be cleared
+        reg.reload(dict(models), "a")
+        assert "a" not in reg._clients
+
+    def test_reload_validates_default(self) -> None:
+        models_a = {"a": ModelConfig("a", "x", "x", "m")}
+        reg = ModelRegistry(models=models_a, default="a")
+        with pytest.raises(ValueError, match="Default model"):
+            reg.reload(models_a, "nonexistent")
+        # Registry should be unchanged after failed reload
+        assert reg.has_alias("a")
+        assert reg.default == "a"
+
+    def test_reload_validates_empty(self) -> None:
+        models_a = {"a": ModelConfig("a", "x", "x", "m")}
+        reg = ModelRegistry(models=models_a, default="a")
+        with pytest.raises(ValueError, match="at least one"):
+            reg.reload({}, "a")
+
+
 # ---------------------------------------------------------------------------
 # Session integration
 # ---------------------------------------------------------------------------
@@ -338,7 +614,7 @@ class _FakeUI:
     def approve_tools(self, items: list[dict[str, Any]]) -> tuple[bool, str | None]:
         return True, None
 
-    def on_tool_result(self, call_id: str, name: str, output: str) -> None: ...
+    def on_tool_result(self, call_id: str, name: str, output: str, **kwargs: Any) -> None: ...
     def on_tool_output_chunk(self, call_id: str, chunk: str) -> None: ...
     def on_status(self, usage: dict[str, Any], context_window: int, effort: str) -> None: ...
     def on_plan_review(self, content: str) -> str:
@@ -590,3 +866,36 @@ class TestProtocolModel:
         assert isinstance(restored, CreateWorkstreamMessage)
         assert restored.model == "local"
         assert restored.name == "ws1"
+
+
+# ---------------------------------------------------------------------------
+# detect_model — startup timeout
+# ---------------------------------------------------------------------------
+
+
+class TestDetectModelTimeout:
+    def test_uses_short_timeout_and_no_retries(self) -> None:
+        """detect_model() uses with_options(timeout=10, max_retries=0)."""
+        mock_model = MagicMock()
+        mock_model.id = "test-model"
+        mock_model.owned_by = "test"
+
+        fast_client = MagicMock()
+        fast_client.models.list.return_value = MagicMock(data=[mock_model])
+
+        client = MagicMock()
+        client.with_options.return_value = fast_client
+
+        result = detect_model(client, provider="openai")
+        client.with_options.assert_called_once_with(timeout=10.0, max_retries=0)
+        fast_client.models.list.assert_called_once()
+        assert result[0] == "test-model"
+
+    def test_connection_error_non_fatal(self) -> None:
+        """detect_model(fatal=False) returns (None, None) on connection error."""
+        client = MagicMock()
+        client.with_options.return_value = client
+        client.models.list.side_effect = OSError("Connection refused")
+
+        result = detect_model(client, provider="openai", fatal=False)
+        assert result == (None, None)

tests/test_oidc.py

@@ -22,6 +22,7 @@ from turnstone.core.oidc import (
     load_oidc_config,
     provision_oidc_user,
     validate_id_token,
+    validate_issuer_url,
 )
 
 # ---------------------------------------------------------------------------
@@ -293,6 +294,162 @@ class TestLoadOIDCConfig:
         assert cfg.redirect_base == "http://localhost:8000"
 
 
+# ---------------------------------------------------------------------------
+# SSRF Validation
+# ---------------------------------------------------------------------------
+
+
+class TestValidateIssuerURL:
+    """Tests for ``validate_issuer_url`` SSRF protection."""
+
+    def test_valid_https_url(self):
+        """Public HTTPS issuer URL passes validation."""
+        # Should not raise -- mock DNS to return a public IP.
+        with patch(
+            "socket.getaddrinfo",
+            return_value=[
+                (2, 1, 6, "", ("93.184.216.34", 0)),
+            ],
+        ):
+            validate_issuer_url("https://idp.example.com")
+
+    def test_rejects_http_non_localhost(self):
+        """HTTP is rejected for non-localhost hosts."""
+        with pytest.raises(OIDCError, match="must use HTTPS"):
+            validate_issuer_url("http://idp.example.com")
+
+    def test_allows_http_localhost(self):
+        """HTTP is allowed for localhost (development)."""
+        with patch(
+            "socket.getaddrinfo",
+            return_value=[
+                (2, 1, 6, "", ("127.0.0.1", 0)),
+            ],
+        ):
+            validate_issuer_url("http://localhost:8080")
+
+    def test_allows_http_localhost_subdomain(self):
+        """HTTP is allowed for *.localhost subdomains."""
+        with patch(
+            "socket.getaddrinfo",
+            return_value=[
+                (2, 1, 6, "", ("127.0.0.1", 0)),
+            ],
+        ):
+            validate_issuer_url("http://keycloak.localhost:8080")
+
+    def test_rejects_embedded_credentials(self):
+        """URLs with userinfo (user:pass@host) are rejected."""
+        with pytest.raises(OIDCError, match="embedded credentials"):
+            validate_issuer_url("https://admin:secret@idp.example.com")
+
+    def test_rejects_username_only(self):
+        """URLs with just a username are rejected."""
+        with pytest.raises(OIDCError, match="embedded credentials"):
+            validate_issuer_url("https://admin@idp.example.com")
+
+    def test_rejects_no_hostname(self):
+        """URLs without a hostname are rejected."""
+        with pytest.raises(OIDCError, match="no hostname"):
+            validate_issuer_url("https://")
+
+    def test_rejects_private_10_range(self):
+        """Hostnames resolving to 10.x.x.x are rejected."""
+        with (
+            patch("socket.getaddrinfo", return_value=[(2, 1, 6, "", ("10.0.0.1", 0))]),
+            pytest.raises(OIDCError, match="non-public address.*10.0.0.1"),
+        ):
+            validate_issuer_url("https://internal.corp.example.com")
+
+    def test_rejects_private_172_range(self):
+        """Hostnames resolving to 172.16-31.x.x are rejected."""
+        with (
+            patch("socket.getaddrinfo", return_value=[(2, 1, 6, "", ("172.16.0.1", 0))]),
+            pytest.raises(OIDCError, match="non-public address.*172.16.0.1"),
+        ):
+            validate_issuer_url("https://internal.corp.example.com")
+
+    def test_rejects_private_192_168_range(self):
+        """Hostnames resolving to 192.168.x.x are rejected."""
+        with (
+            patch("socket.getaddrinfo", return_value=[(2, 1, 6, "", ("192.168.1.1", 0))]),
+            pytest.raises(OIDCError, match="non-public address.*192.168.1.1"),
+        ):
+            validate_issuer_url("https://internal.corp.example.com")
+
+    def test_rejects_loopback_127(self):
+        """Hostnames resolving to 127.x.x.x are rejected (non-localhost host)."""
+        with (
+            patch("socket.getaddrinfo", return_value=[(2, 1, 6, "", ("127.0.0.1", 0))]),
+            pytest.raises(OIDCError, match="non-public address.*127.0.0.1"),
+        ):
+            validate_issuer_url("https://evil.example.com")
+
+    def test_rejects_ipv6_loopback(self):
+        """Hostnames resolving to ::1 are rejected (non-localhost host)."""
+        with (
+            patch("socket.getaddrinfo", return_value=[(10, 1, 6, "", ("::1", 0, 0, 0))]),
+            pytest.raises(OIDCError, match="non-public address.*::1"),
+        ):
+            validate_issuer_url("https://evil.example.com")
+
+    def test_rejects_ipv6_private(self):
+        """Hostnames resolving to fc00::/7 are rejected."""
+        with (
+            patch("socket.getaddrinfo", return_value=[(10, 1, 6, "", ("fd00::1", 0, 0, 0))]),
+            pytest.raises(OIDCError, match="non-public address.*fd00::1"),
+        ):
+            validate_issuer_url("https://evil.example.com")
+
+    def test_rejects_link_local(self):
+        """Hostnames resolving to link-local addresses are rejected."""
+        with (
+            patch("socket.getaddrinfo", return_value=[(2, 1, 6, "", ("169.254.169.254", 0))]),
+            pytest.raises(OIDCError, match="non-public address.*169.254.169.254"),
+        ):
+            validate_issuer_url("https://metadata.internal")
+
+    def test_rejects_unresolvable_hostname(self):
+        """DNS resolution failure is rejected."""
+        import socket as _socket
+
+        with (
+            patch("socket.getaddrinfo", side_effect=_socket.gaierror("not found")),
+            pytest.raises(OIDCError, match="cannot be resolved"),
+        ):
+            validate_issuer_url("https://nonexistent.invalid")
+
+    def test_rejects_mixed_addresses(self):
+        """If any resolved address is private, the URL is rejected."""
+        with (
+            patch(
+                "socket.getaddrinfo",
+                return_value=[
+                    (2, 1, 6, "", ("93.184.216.34", 0)),
+                    (2, 1, 6, "", ("10.0.0.1", 0)),
+                ],
+            ),
+            pytest.raises(OIDCError, match="non-public address.*10.0.0.1"),
+        ):
+            validate_issuer_url("https://dual-homed.example.com")
+
+    def test_discover_rejects_ssrf(self):
+        """discover_oidc returns enabled=False when issuer URL fails SSRF check."""
+        config = _make_config(
+            issuer="http://10.0.0.1:8080",
+            authorization_endpoint="",
+            token_endpoint="",
+            userinfo_endpoint="",
+            jwks_uri="",
+        )
+
+        async def _run():
+            result = await discover_oidc(config)
+            assert result.enabled is False
+
+        asyncio.run(_run())
+
+
 # ---------------------------------------------------------------------------
 # Redirect URI Builder
 # ---------------------------------------------------------------------------
@@ -869,6 +1026,9 @@ class TestApplyRoleMapping:
 
 
 class TestDiscoverOIDC:
+    # Mock DNS result for a public IP — reused across discovery tests.
+    _PUBLIC_ADDR = [(2, 1, 6, "", ("93.184.216.34", 0))]
+
     def test_discover_oidc_success(self):
         """Mock httpx response, verify endpoints populated."""
         config = _make_config(
@@ -891,7 +1051,10 @@ class TestDiscoverOIDC:
 
         async def _run():
             client = _mock_async_client(lambda url: _async_return(mock_response))
-            with patch("httpx.AsyncClient", return_value=client):
+            with (
+                patch("socket.getaddrinfo", return_value=self._PUBLIC_ADDR),
+                patch("httpx.AsyncClient", return_value=client),
+            ):
                 result = await discover_oidc(config)
 
             assert result.authorization_endpoint == "https://idp.example.com/authorize"
@@ -916,7 +1079,10 @@ class TestDiscoverOIDC:
 
         async def _run():
             client = _mock_async_client(_failing_get)
-            with patch("httpx.AsyncClient", return_value=client):
+            with (
+                patch("socket.getaddrinfo", return_value=self._PUBLIC_ADDR),
+                patch("httpx.AsyncClient", return_value=client),
+            ):
                 result = await discover_oidc(config)
 
             assert result.enabled is False
@@ -954,7 +1120,10 @@ class TestDiscoverOIDC:
 
         async def _run():
             client = _mock_async_client(lambda url: _async_return(mock_response))
-            with patch("httpx.AsyncClient", return_value=client):
+            with (
+                patch("socket.getaddrinfo", return_value=self._PUBLIC_ADDR),
+                patch("httpx.AsyncClient", return_value=client),
+            ):
                 result = await discover_oidc(config)
 
             assert result.enabled is False

tests/test_oidc_handlers.py

@@ -131,7 +131,7 @@ def authorize_client(storage: SQLiteBackend, oidc_config: OIDCConfig) -> TestCli
     )
     app.state.oidc_config = oidc_config
     app.state.auth_storage = storage
-    app.state.jwt_secret = "test-jwt-secret"
+    app.state.jwt_secret = "test-jwt-secret-key-padded-32b!!"
     app.state.jwks_data = {"keys": []}
     app.state.login_limiter = None
     return TestClient(app, raise_server_exceptions=False)
@@ -468,7 +468,7 @@ class TestOIDCCallback:
         )
         app.state.oidc_config = _make_oidc_config()
         app.state.auth_storage = backend
-        app.state.jwt_secret = "secret"
+        app.state.jwt_secret = "test-jwt-secret-key-padded-32b!!"
         app.state.jwks_data = {"keys": []}
         app.state.login_limiter = None
 
@@ -492,7 +492,7 @@ class TestOIDCCallback:
         )
         app.state.oidc_config = _make_oidc_config()
         app.state.auth_storage = storage
-        app.state.jwt_secret = "secret"
+        app.state.jwt_secret = "test-jwt-secret-key-padded-32b!!"
         app.state.jwks_data = {"keys": []}
         limiter = LoginRateLimiter(max_attempts=1, window_seconds=300)
         limiter.record("ip:testclient")

tests/test_oidc_storage.py

@@ -6,15 +6,6 @@ import time
 
 import pytest
 
-from turnstone.core.storage._sqlite import SQLiteBackend
-
-
-@pytest.fixture()
-def db(tmp_path):
-    """Create a fresh SQLite backend for each test."""
-    return SQLiteBackend(str(tmp_path / "test.db"))
-
-
 # ---------------------------------------------------------------------------
 # OIDC Identity CRUD
 # ---------------------------------------------------------------------------

tests/test_output_assessment_storage.py

@@ -4,16 +4,6 @@ from __future__ import annotations
 
 from datetime import UTC, datetime, timedelta
 
-import pytest
-
-from turnstone.core.storage._sqlite import SQLiteBackend
-
-
-@pytest.fixture()
-def db(tmp_path):
-    """Fresh SQLite backend for each test."""
-    return SQLiteBackend(str(tmp_path / "test.db"))
-
 
 def _make_assessment_kwargs(**overrides):
     """Build default kwargs for record_output_assessment."""

tests/test_prompt_templates_runtime.py

@@ -29,7 +29,7 @@ class NullUI:
     def approve_tools(self, items):
         return True, None
 
-    def on_tool_result(self, call_id, name, output):
+    def on_tool_result(self, call_id, name, output, **kwargs):
         pass
 
     def on_tool_output_chunk(self, call_id, chunk):

tests/test_providers.py

@@ -94,6 +94,7 @@ def _anthropic_event(
         delta.type = kwargs.get("delta_type", "text_delta")
         delta.text = kwargs.get("text", "")
         delta.thinking = kwargs.get("thinking", "")
+        delta.signature = kwargs.get("signature", "")
         delta.partial_json = kwargs.get("partial_json", "")
         event.delta = delta
         event.index = kwargs.get("index", 0)
@@ -140,6 +141,34 @@ class TestOpenAIProvider:
     def test_provider_name(self) -> None:
         assert self.provider.provider_name == "openai"
 
+    # -- _sanitize_messages ---------------------------------------------------
+
+    def test_sanitize_messages_none_content_no_tool_calls(self) -> None:
+        msgs = [{"role": "assistant", "content": None}]
+        assert self.provider._sanitize_messages(msgs) == [{"role": "assistant", "content": ""}]
+
+    def test_sanitize_messages_none_content_with_tool_calls(self) -> None:
+        msgs = [{"role": "assistant", "content": None, "tool_calls": [{"id": "1"}]}]
+        result = self.provider._sanitize_messages(msgs)
+        assert result[0]["content"] is None
+        assert result[0]["tool_calls"] == [{"id": "1"}]
+
+    def test_sanitize_messages_empty_string_passthrough(self) -> None:
+        msgs = [{"role": "assistant", "content": ""}]
+        assert self.provider._sanitize_messages(msgs) == msgs
+
+    def test_sanitize_messages_non_assistant_unchanged(self) -> None:
+        msgs = [{"role": "user", "content": None}]
+        result = self.provider._sanitize_messages(msgs)
+        assert result[0]["content"] is None
+
+    def test_sanitize_messages_does_not_mutate_original(self) -> None:
+        original = {"role": "assistant", "content": None}
+        self.provider._sanitize_messages([original])
+        assert original["content"] is None
+
+    # -- convert_tools --------------------------------------------------------
+
     def test_convert_tools_passthrough(self) -> None:
         tools = [
             {
@@ -479,10 +508,11 @@ class TestAnthropicProvider:
                         },
                     }
                 ],
-            }
+            },
+            {"role": "tool", "tool_call_id": "call_1", "content": "file contents"},
         ]
         _, converted = self.provider._convert_messages(messages)
-        assert len(converted) == 1
+        assert len(converted) == 2
         blocks = converted[0]["content"]
         assert len(blocks) == 2
         assert blocks[0] == {"type": "text", "text": "Let me check that."}
@@ -490,6 +520,8 @@ class TestAnthropicProvider:
         assert blocks[1]["id"] == "call_1"
         assert blocks[1]["name"] == "read_file"
         assert blocks[1]["input"] == {"path": "foo.py"}
+        # Tool result in user message
+        assert converted[1]["role"] == "user"
 
     def test_message_conversion_tool_results(self) -> None:
         messages = [
@@ -598,7 +630,11 @@ class TestAnthropicProvider:
         response.usage.output_tokens = 5
 
         client = MagicMock()
-        client.messages.create.return_value = response
+        stream_ctx = MagicMock()
+        stream_ctx.__enter__ = MagicMock(return_value=stream_ctx)
+        stream_ctx.__exit__ = MagicMock(return_value=False)
+        stream_ctx.get_final_message.return_value = response
+        client.messages.stream.return_value = stream_ctx
 
         result = self.provider.create_completion(
             client=client,
@@ -630,7 +666,11 @@ class TestAnthropicProvider:
         response.usage.output_tokens = 20
 
         client = MagicMock()
-        client.messages.create.return_value = response
+        stream_ctx = MagicMock()
+        stream_ctx.__enter__ = MagicMock(return_value=stream_ctx)
+        stream_ctx.__exit__ = MagicMock(return_value=False)
+        stream_ctx.get_final_message.return_value = response
+        client.messages.stream.return_value = stream_ctx
 
         result = self.provider.create_completion(
             client=client,
@@ -661,7 +701,11 @@ class TestAnthropicProvider:
         response.usage.output_tokens = 50
 
         client = MagicMock()
-        client.messages.create.return_value = response
+        stream_ctx = MagicMock()
+        stream_ctx.__enter__ = MagicMock(return_value=stream_ctx)
+        stream_ctx.__exit__ = MagicMock(return_value=False)
+        stream_ctx.get_final_message.return_value = response
+        client.messages.stream.return_value = stream_ctx
 
         result = self.provider.create_completion(
             client=client,
@@ -913,7 +957,7 @@ class TestAnthropicHelpers:
 
         provider = AnthropicProvider()
         caps = provider.get_capabilities("claude-opus-4-6")
-        assert caps.context_window == 200000
+        assert caps.context_window == 1000000
         assert caps.max_output_tokens == 128000
         assert caps.thinking_mode == "adaptive"
         assert caps.supports_effort is True
@@ -922,11 +966,11 @@ class TestAnthropicHelpers:
         from turnstone.core.providers._anthropic import AnthropicProvider
 
         provider = AnthropicProvider()
-        # Prefix match: "claude-sonnet-4" matches dated variants
-        caps = provider.get_capabilities("claude-sonnet-4-20260101")
-        assert caps.context_window == 200000
+        # Prefix match: "claude-sonnet-4-6" matches dated variants
+        caps = provider.get_capabilities("claude-sonnet-4-6-20260101")
+        assert caps.context_window == 1000000
         assert caps.token_param == "max_tokens"
-        assert caps.thinking_mode == "manual"
+        assert caps.thinking_mode == "adaptive"
 
     def test_capabilities_lookup_unknown(self) -> None:
         from turnstone.core.providers._anthropic import AnthropicProvider
@@ -1141,6 +1185,189 @@ class TestOpenAIParameterGating:
         assert kwargs["reasoning_effort"] == "medium"  # fell back from unsupported "low"
 
 
+class TestAnthropicOrphanedToolUse:
+    """Verify _convert_messages synthesizes tool_results for orphaned tool_use."""
+
+    def setup_method(self) -> None:
+        from turnstone.core.providers._anthropic import AnthropicProvider
+
+        self.provider = AnthropicProvider()
+
+    def test_orphaned_tool_use_gets_synthetic_result(self) -> None:
+        """Assistant has tool_calls but next message is user (no tool results)."""
+        messages = [
+            {"role": "user", "content": "do something"},
+            {
+                "role": "assistant",
+                "content": "I'll run that.",
+                "tool_calls": [
+                    {
+                        "id": "call_abc",
+                        "function": {"name": "bash", "arguments": '{"command": "ls"}'},
+                    }
+                ],
+            },
+            {"role": "user", "content": "never mind, do something else"},
+        ]
+        _, converted = self.provider._convert_messages(messages)
+        # Should have: user, assistant(tool_use), user(synthetic tool_result), user
+        # After _merge_consecutive, the two user messages may merge.
+        # Find the synthetic tool_result
+        tool_results = []
+        for msg in converted:
+            if msg["role"] == "user" and isinstance(msg["content"], list):
+                for block in msg["content"]:
+                    if isinstance(block, dict) and block.get("type") == "tool_result":
+                        tool_results.append(block)
+        assert len(tool_results) == 1
+        assert tool_results[0]["tool_use_id"] == "call_abc"
+        assert tool_results[0]["is_error"] is True
+        assert "cancelled" in tool_results[0]["content"].lower()
+
+    def test_multiple_orphaned_tool_calls(self) -> None:
+        """Assistant has 3 tool_calls, none have results."""
+        messages = [
+            {"role": "user", "content": "do three things"},
+            {
+                "role": "assistant",
+                "content": "",
+                "tool_calls": [
+                    {"id": "c1", "function": {"name": "bash", "arguments": "{}"}},
+                    {"id": "c2", "function": {"name": "read_file", "arguments": "{}"}},
+                    {"id": "c3", "function": {"name": "write_file", "arguments": "{}"}},
+                ],
+            },
+            {"role": "user", "content": "skip all that"},
+        ]
+        _, converted = self.provider._convert_messages(messages)
+        tool_results = []
+        for msg in converted:
+            if msg["role"] == "user" and isinstance(msg["content"], list):
+                for block in msg["content"]:
+                    if isinstance(block, dict) and block.get("type") == "tool_result":
+                        tool_results.append(block)
+        assert len(tool_results) == 3
+        result_ids = {r["tool_use_id"] for r in tool_results}
+        assert result_ids == {"c1", "c2", "c3"}
+
+    def test_partial_results_only_orphans_synthesized(self) -> None:
+        """2 tool_calls, only 1 has a result — synthesize for the missing one."""
+        messages = [
+            {"role": "user", "content": "do two things"},
+            {
+                "role": "assistant",
+                "content": "",
+                "tool_calls": [
+                    {"id": "c1", "function": {"name": "bash", "arguments": "{}"}},
+                    {"id": "c2", "function": {"name": "write_file", "arguments": "{}"}},
+                ],
+            },
+            {"role": "tool", "tool_call_id": "c1", "content": "file1.txt"},
+            {"role": "user", "content": "skip the write"},
+        ]
+        _, converted = self.provider._convert_messages(messages)
+        # c1 should have a real result, c2 should have a synthetic one
+        tool_results = []
+        for msg in converted:
+            if msg["role"] == "user" and isinstance(msg["content"], list):
+                for block in msg["content"]:
+                    if isinstance(block, dict) and block.get("type") == "tool_result":
+                        tool_results.append(block)
+        # Real result should come before synthetic (ordering matters for Anthropic)
+        assert len(tool_results) == 2
+        assert tool_results[0]["tool_use_id"] == "c1"
+        assert tool_results[0]["content"] == "file1.txt"  # real result
+        assert tool_results[0].get("is_error") is not True
+        assert tool_results[1]["tool_use_id"] == "c2"
+        assert tool_results[1]["is_error"] is True  # synthetic
+
+    def test_complete_results_no_synthesis(self) -> None:
+        """All tool_calls have results — no synthesis needed."""
+        messages = [
+            {"role": "user", "content": "do it"},
+            {
+                "role": "assistant",
+                "content": "",
+                "tool_calls": [
+                    {"id": "c1", "function": {"name": "bash", "arguments": "{}"}},
+                ],
+            },
+            {"role": "tool", "tool_call_id": "c1", "content": "done"},
+            {"role": "user", "content": "thanks"},
+        ]
+        _, converted = self.provider._convert_messages(messages)
+        # No synthetic results — only the real one (no is_error flag)
+        tool_results = []
+        for msg in converted:
+            if msg["role"] == "user" and isinstance(msg["content"], list):
+                for block in msg["content"]:
+                    if isinstance(block, dict) and block.get("type") == "tool_result":
+                        tool_results.append(block)
+        assert len(tool_results) == 1
+        assert tool_results[0]["tool_use_id"] == "c1"
+        assert tool_results[0].get("is_error") is not True
+
+    def test_trailing_orphan(self) -> None:
+        """Orphaned tool_use at end of conversation (no following messages)."""
+        messages = [
+            {"role": "user", "content": "do it"},
+            {
+                "role": "assistant",
+                "content": "Running...",
+                "tool_calls": [
+                    {"id": "c1", "function": {"name": "bash", "arguments": "{}"}},
+                ],
+            },
+        ]
+        _, converted = self.provider._convert_messages(messages)
+        tool_results = []
+        for msg in converted:
+            if msg["role"] == "user" and isinstance(msg["content"], list):
+                for block in msg["content"]:
+                    if isinstance(block, dict) and block.get("type") == "tool_result":
+                        tool_results.append(block)
+        assert len(tool_results) == 1
+        assert tool_results[0]["tool_use_id"] == "c1"
+        assert tool_results[0]["is_error"] is True
+
+    def test_provider_content_orphan(self) -> None:
+        """Orphaned tool_use inside _provider_content (Anthropic raw blocks)."""
+        messages = [
+            {"role": "user", "content": "run something"},
+            {
+                "role": "assistant",
+                "content": "Running...",
+                "_provider_content": [
+                    {"type": "text", "text": "Running..."},
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_abc",
+                        "name": "bash",
+                        "input": {"command": "sleep 30"},
+                    },
+                ],
+                "tool_calls": [
+                    {
+                        "id": "toolu_abc",
+                        "function": {"name": "bash", "arguments": '{"command": "sleep 30"}'},
+                    },
+                ],
+            },
+            {"role": "user", "content": "never mind"},
+        ]
+        _, converted = self.provider._convert_messages(messages)
+        # Should synthesize a tool_result for the orphaned tool_use in provider_content
+        tool_results = []
+        for msg in converted:
+            if msg["role"] == "user" and isinstance(msg["content"], list):
+                for block in msg["content"]:
+                    if isinstance(block, dict) and block.get("type") == "tool_result":
+                        tool_results.append(block)
+        assert len(tool_results) == 1
+        assert tool_results[0]["tool_use_id"] == "toolu_abc"
+        assert tool_results[0]["is_error"] is True
+
+
 class TestAnthropicReasoningNone:
     """Verify 'none' effort disables thinking for manual-thinking models."""
 
@@ -1180,7 +1407,7 @@ class TestAnthropicWebSearch:
         """All Anthropic models should support native web search."""
         caps = self.provider.get_capabilities("claude-opus-4-6")
         assert caps.supports_web_search is True
-        caps = self.provider.get_capabilities("claude-sonnet-4")
+        caps = self.provider.get_capabilities("claude-sonnet-4-6")
         assert caps.supports_web_search is True
         # Unknown models use default which also has web search
         caps = self.provider.get_capabilities("claude-unknown-99")
@@ -1383,7 +1610,11 @@ class TestAnthropicWebSearch:
         response.usage.output_tokens = 50
 
         client = MagicMock()
-        client.messages.create.return_value = response
+        stream_ctx = MagicMock()
+        stream_ctx.__enter__ = MagicMock(return_value=stream_ctx)
+        stream_ctx.__exit__ = MagicMock(return_value=False)
+        stream_ctx.get_final_message.return_value = response
+        client.messages.stream.return_value = stream_ctx
 
         with patch("turnstone.core.providers._anthropic._ensure_anthropic"):
             result = self.provider.create_completion(
@@ -1934,6 +2165,104 @@ class TestAnthropicProviderBlocks:
         assert blocks[2]["type"] == "web_search_tool_result"
         assert blocks[2]["encrypted_content"] == "enc_data"
 
+    def test_streaming_thinking_block_captures_signature(self) -> None:
+        """Streaming thinking block accumulates signature from signature_delta events."""
+        thinking_block = MagicMock()
+        thinking_block.type = "thinking"
+        thinking_block.model_dump.return_value = {
+            "type": "thinking",
+            "thinking": "",
+            "signature": "",
+        }
+
+        text_block = MagicMock()
+        text_block.type = "text"
+        text_block.model_dump.return_value = {"type": "text", "text": ""}
+
+        events = [
+            MagicMock(type="content_block_start", index=0, content_block=thinking_block),
+            _anthropic_event(
+                "content_block_delta", delta_type="thinking_delta", thinking="step 1", index=0
+            ),
+            _anthropic_event(
+                "content_block_delta", delta_type="thinking_delta", thinking=" step 2", index=0
+            ),
+            _anthropic_event(
+                "content_block_delta",
+                delta_type="signature_delta",
+                signature="sig_part1",
+                index=0,
+            ),
+            _anthropic_event(
+                "content_block_delta",
+                delta_type="signature_delta",
+                signature="sig_part2",
+                index=0,
+            ),
+            _anthropic_event("content_block_stop", index=0),
+            MagicMock(type="content_block_start", index=1, content_block=text_block),
+            _anthropic_event("content_block_delta", delta_type="text_delta", text="Hello", index=1),
+            _anthropic_event("content_block_stop", index=1),
+            _anthropic_event("message_delta", stop_reason="end_turn", usage_output_tokens=50),
+        ]
+
+        chunks = list(self.provider._iter_anthropic_stream(iter(events)))
+        final_chunks = [c for c in chunks if c.provider_blocks]
+        assert len(final_chunks) == 1
+        blocks = final_chunks[0].provider_blocks
+        assert blocks[0]["type"] == "thinking"
+        assert blocks[0]["thinking"] == "step 1 step 2"
+        assert blocks[0]["signature"] == "sig_part1sig_part2"
+
+    def test_thinking_block_multiturn_roundtrip(self) -> None:
+        """Thinking block with signature survives _convert_messages round-trip."""
+        provider_content = [
+            {
+                "type": "thinking",
+                "thinking": "Let me reason...",
+                "signature": "ErUBCkYIAxgCIkD_valid_sig",
+            },
+            {"type": "text", "text": "Here is my answer."},
+        ]
+        messages = [
+            {"role": "user", "content": "Question"},
+            {
+                "role": "assistant",
+                "content": "Here is my answer.",
+                "_provider_content": provider_content,
+            },
+            {"role": "user", "content": "Follow up"},
+        ]
+        _, converted = self.provider._convert_messages(messages)
+        assistant_msg = converted[1]
+        assert assistant_msg["content"] is provider_content
+        assert assistant_msg["content"][0]["signature"] == "ErUBCkYIAxgCIkD_valid_sig"
+        assert assistant_msg["content"][0]["type"] == "thinking"
+
+    def test_block_to_dict_preserves_thinking_signature(self) -> None:
+        """_block_to_dict preserves signature on thinking blocks."""
+        from turnstone.core.providers._anthropic import _block_to_dict
+
+        class FakeThinkingBlock:
+            def model_dump(self, **kwargs: Any) -> dict[str, Any]:
+                return {
+                    "type": "thinking",
+                    "thinking": "reasoning...",
+                    "signature": "abc123sig",
+                }
+
+        result = _block_to_dict(FakeThinkingBlock())
+        assert result["signature"] == "abc123sig"
+
+        # Also test fallback path (no model_dump)
+        class FallbackBlock:
+            type = "thinking"
+            thinking = "reasoning..."
+            signature = "abc123sig"
+
+        result2 = _block_to_dict(FallbackBlock())
+        assert result2["signature"] == "abc123sig"
+
 
 # ---------------------------------------------------------------------------
 # Tool search tests
@@ -2307,7 +2636,11 @@ class TestAnthropicPromptCaching:
         response.usage = usage
 
         client = MagicMock()
-        client.messages.create.return_value = response
+        stream_ctx = MagicMock()
+        stream_ctx.__enter__ = MagicMock(return_value=stream_ctx)
+        stream_ctx.__exit__ = MagicMock(return_value=False)
+        stream_ctx.get_final_message.return_value = response
+        client.messages.stream.return_value = stream_ctx
 
         result = self.provider.create_completion(
             client=client,

tests/test_reconstruct_messages.py

@@ -9,13 +9,12 @@ def _row(
     role,
     content=None,
     tool_name=None,
-    tool_args=None,
     tc_id=None,
     pdata=None,
     tool_calls=None,
 ):
-    """Build a 7-element conversation row tuple (post-migration 013 format)."""
-    return (role, content, tool_name, tool_args, tc_id, pdata, tool_calls)
+    """Build a 6-element conversation row tuple (post-migration 027 format)."""
+    return (role, content, tool_name, tc_id, pdata, tool_calls)
 
 
 class TestAssistantWithToolCalls:
@@ -227,3 +226,91 @@ class TestEdgeCases:
         assert len(msgs) == 2
         assert msgs[0]["role"] == "user"
         assert msgs[1]["role"] == "assistant"
+
+
+class TestMidConversationOrphanRepair:
+    """Mid-conversation orphaned tool_calls get synthetic tool results."""
+
+    def test_all_orphaned_mid_conversation(self):
+        """Assistant has 2 tool_calls, no tool results, then user message."""
+        tc = json.dumps(
+            [
+                {"id": "c1", "function": {"name": "bash", "arguments": "{}"}},
+                {"id": "c2", "function": {"name": "write_file", "arguments": "{}"}},
+            ]
+        )
+        rows = [
+            _row("user", "do stuff"),
+            _row("assistant", "Running...", tool_calls=tc),
+            _row("user", "never mind"),
+        ]
+        msgs = reconstruct_messages(rows, "ws1")
+        # Should have: user, assistant, tool(c1), tool(c2), user
+        assert len(msgs) == 5
+        assert msgs[2]["role"] == "tool"
+        assert msgs[2]["tool_call_id"] == "c1"
+        assert msgs[2]["is_error"] is True
+        assert msgs[3]["role"] == "tool"
+        assert msgs[3]["tool_call_id"] == "c2"
+        assert msgs[4]["role"] == "user"
+
+    def test_partial_results_mid_conversation(self):
+        """2 tool_calls, 1 result present, 1 missing — synthesize only the missing one."""
+        tc = json.dumps(
+            [
+                {"id": "c1", "function": {"name": "bash", "arguments": "{}"}},
+                {"id": "c2", "function": {"name": "write_file", "arguments": "{}"}},
+            ]
+        )
+        rows = [
+            _row("user", "do stuff"),
+            _row("assistant", "", tool_calls=tc),
+            _row("tool", "file1.txt", tool_name="bash", tc_id="c1"),
+            _row("user", "skip the write"),
+        ]
+        msgs = reconstruct_messages(rows, "ws1")
+        # Should have: user, assistant, tool(c1 real), tool(c2 synthetic), user
+        assert len(msgs) == 5
+        assert msgs[2]["role"] == "tool"
+        assert msgs[2]["tool_call_id"] == "c1"
+        assert msgs[2]["content"] == "file1.txt"
+        assert msgs[2].get("is_error") is not True
+        assert msgs[3]["role"] == "tool"
+        assert msgs[3]["tool_call_id"] == "c2"
+        assert msgs[3]["is_error"] is True
+        assert msgs[4]["role"] == "user"
+
+    def test_complete_results_no_synthesis(self):
+        """All tool_calls have results — no synthesis needed."""
+        tc = json.dumps(
+            [
+                {"id": "c1", "function": {"name": "bash", "arguments": "{}"}},
+            ]
+        )
+        rows = [
+            _row("user", "do it"),
+            _row("assistant", "", tool_calls=tc),
+            _row("tool", "done", tool_name="bash", tc_id="c1"),
+            _row("user", "thanks"),
+        ]
+        msgs = reconstruct_messages(rows, "ws1")
+        assert len(msgs) == 4
+        tool_msgs = [m for m in msgs if m["role"] == "tool"]
+        assert len(tool_msgs) == 1
+        assert tool_msgs[0].get("is_error") is not True
+
+    def test_trailing_orphan_stripped_not_synthesized(self):
+        """Trailing orphan is handled by the existing strip repair, not synthesis."""
+        tc = json.dumps(
+            [
+                {"id": "c1", "function": {"name": "bash", "arguments": "{}"}},
+            ]
+        )
+        rows = [
+            _row("user", "do it"),
+            _row("assistant", "Running...", tool_calls=tc),
+        ]
+        msgs = reconstruct_messages(rows, "ws1")
+        # Trailing strip removes the assistant message entirely
+        assert len(msgs) == 1
+        assert msgs[0]["role"] == "user"

tests/test_rewind_retry.py

@@ -0,0 +1,387 @@
+"""Tests for conversation rewind and retry functionality."""
+
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+from turnstone.core.session import ChatSession
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+class NullUI:
+    """UI adapter that discards all output."""
+
+    def on_thinking_start(self):
+        pass
+
+    def on_thinking_stop(self):
+        pass
+
+    def on_reasoning_token(self, text):
+        pass
+
+    def on_content_token(self, text):
+        pass
+
+    def on_stream_end(self):
+        pass
+
+    def approve_tools(self, items):
+        return True, None
+
+    def on_tool_result(self, call_id, name, output, **kwargs):
+        pass
+
+    def on_tool_output_chunk(self, call_id, chunk):
+        pass
+
+    def on_status(self, usage, context_window, effort):
+        pass
+
+    def on_plan_review(self, content):
+        return ""
+
+    def on_info(self, message):
+        pass
+
+    def on_error(self, message):
+        pass
+
+    def on_state_change(self, state):
+        pass
+
+    def on_rename(self, name):
+        pass
+
+    def on_output_warning(self, call_id, assessment):
+        pass
+
+
+def _make_session(tmp_db) -> ChatSession:
+    return ChatSession(
+        client=MagicMock(),
+        model="test-model",
+        ui=NullUI(),
+        instructions="",
+        temperature=0.5,
+        max_tokens=4096,
+        tool_timeout=30,
+    )
+
+
+def _populate_simple(session: ChatSession) -> None:
+    """Populate with 2 simple turns (no tool calls)."""
+    session.messages = [
+        {"role": "user", "content": "Hello"},
+        {"role": "assistant", "content": "Hi there!"},
+        {"role": "user", "content": "How are you?"},
+        {"role": "assistant", "content": "I'm fine."},
+    ]
+    session._msg_tokens = [10, 20, 10, 20]
+
+
+def _populate_with_tools(session: ChatSession) -> None:
+    """Populate with 2 turns, first has tool calls."""
+    session.messages = [
+        {"role": "user", "content": "Write a test"},
+        {
+            "role": "assistant",
+            "content": "",
+            "tool_calls": [
+                {"id": "tc1", "function": {"name": "bash", "arguments": '{"cmd":"echo hi"}'}}
+            ],
+        },
+        {"role": "tool", "tool_call_id": "tc1", "content": "hi"},
+        {"role": "assistant", "content": "Done."},
+        {"role": "user", "content": "Fix the import"},
+        {"role": "assistant", "content": "Fixed."},
+    ]
+    session._msg_tokens = [10, 20, 10, 20, 10, 20]
+
+
+# ---------------------------------------------------------------------------
+# _find_turn_boundaries
+# ---------------------------------------------------------------------------
+
+
+class TestFindTurnBoundaries:
+    def test_empty_messages(self, tmp_db):
+        session = _make_session(tmp_db)
+        assert session._find_turn_boundaries() == []
+
+    def test_single_turn(self, tmp_db):
+        session = _make_session(tmp_db)
+        session.messages = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi!"},
+        ]
+        assert session._find_turn_boundaries() == [0]
+
+    def test_multi_turn(self, tmp_db):
+        session = _make_session(tmp_db)
+        _populate_simple(session)
+        assert session._find_turn_boundaries() == [0, 2]
+
+    def test_with_tool_calls(self, tmp_db):
+        session = _make_session(tmp_db)
+        _populate_with_tools(session)
+        assert session._find_turn_boundaries() == [0, 4]
+
+
+# ---------------------------------------------------------------------------
+# rewind
+# ---------------------------------------------------------------------------
+
+
+class TestRewind:
+    def test_rewind_zero(self, tmp_db):
+        session = _make_session(tmp_db)
+        _populate_simple(session)
+        assert session.rewind(0) == 0
+        assert len(session.messages) == 4
+
+    def test_rewind_one_turn(self, tmp_db):
+        session = _make_session(tmp_db)
+        _populate_simple(session)
+        removed = session.rewind(1)
+        assert removed == 2  # user + assistant
+        assert len(session.messages) == 2
+        assert session.messages[0]["content"] == "Hello"
+        assert session.messages[1]["content"] == "Hi there!"
+        assert len(session._msg_tokens) == 2
+
+    def test_rewind_all_turns(self, tmp_db):
+        session = _make_session(tmp_db)
+        _populate_simple(session)
+        removed = session.rewind(2)
+        assert removed == 4
+        assert len(session.messages) == 0
+        assert len(session._msg_tokens) == 0
+
+    def test_rewind_clamped(self, tmp_db):
+        """Rewinding more turns than exist should clamp to available."""
+        session = _make_session(tmp_db)
+        _populate_simple(session)
+        removed = session.rewind(999)
+        assert removed == 4
+        assert len(session.messages) == 0
+
+    def test_rewind_empty(self, tmp_db):
+        session = _make_session(tmp_db)
+        assert session.rewind(1) == 0
+
+    def test_rewind_with_tools(self, tmp_db):
+        """Rewinding 1 turn on a multi-sub-turn conversation."""
+        session = _make_session(tmp_db)
+        _populate_with_tools(session)
+        removed = session.rewind(1)
+        assert removed == 2  # user "Fix the import" + assistant "Fixed."
+        assert len(session.messages) == 4
+        assert session.messages[-1]["content"] == "Done."
+
+    def test_rewind_tokens_sync(self, tmp_db):
+        """_msg_tokens stays in sync with messages."""
+        session = _make_session(tmp_db)
+        _populate_simple(session)
+        session.rewind(1)
+        assert len(session._msg_tokens) == len(session.messages)
+
+
+# ---------------------------------------------------------------------------
+# retry
+# ---------------------------------------------------------------------------
+
+
+class TestRetry:
+    def test_retry_returns_user_message(self, tmp_db):
+        session = _make_session(tmp_db)
+        _populate_simple(session)
+        msg = session.retry()
+        assert msg == "How are you?"
+        # Only Turn 1 remains, without the second user message
+        assert len(session.messages) == 2
+        assert session.messages[-1]["content"] == "Hi there!"
+
+    def test_retry_empty(self, tmp_db):
+        session = _make_session(tmp_db)
+        assert session.retry() is None
+
+    def test_retry_with_tools(self, tmp_db):
+        session = _make_session(tmp_db)
+        _populate_with_tools(session)
+        msg = session.retry()
+        assert msg == "Fix the import"
+        # Only Turn 1 remains (user + assistant w/tools + tool result + assistant)
+        assert len(session.messages) == 4
+
+    def test_retry_sets_pending(self, tmp_db):
+        """handle_command for /retry should set _pending_retry."""
+        session = _make_session(tmp_db)
+        _populate_simple(session)
+        session.handle_command("/retry")
+        assert session._pending_retry == "How are you?"
+
+    def test_retry_tokens_sync(self, tmp_db):
+        session = _make_session(tmp_db)
+        _populate_simple(session)
+        session.retry()
+        assert len(session._msg_tokens) == len(session.messages)
+
+    def test_retry_multipart_content_returns_none(self, tmp_db):
+        """retry() should refuse multipart (vision/image) messages."""
+        session = _make_session(tmp_db)
+        session.messages = [
+            {"role": "user", "content": [{"type": "text", "text": "describe this"}]},
+            {"role": "assistant", "content": "It's an image."},
+        ]
+        session._msg_tokens = [10, 20]
+        assert session.retry() is None
+        # Messages should be unchanged
+        assert len(session.messages) == 2
+
+    def test_retry_none_content_returns_none(self, tmp_db):
+        """retry() should handle content=None gracefully."""
+        session = _make_session(tmp_db)
+        session.messages = [
+            {"role": "user", "content": None},
+            {"role": "assistant", "content": "Ok."},
+        ]
+        session._msg_tokens = [10, 20]
+        assert session.retry() is None
+
+
+# ---------------------------------------------------------------------------
+# handle_command integration
+# ---------------------------------------------------------------------------
+
+
+class TestHandleCommand:
+    def test_rewind_command(self, tmp_db):
+        session = _make_session(tmp_db)
+        _populate_simple(session)
+        session.handle_command("/rewind 1")
+        assert len(session.messages) == 2
+
+    def test_rewind_no_arg(self, tmp_db):
+        session = _make_session(tmp_db)
+        ui = session.ui
+        ui.on_info = MagicMock()
+        session.handle_command("/rewind")
+        ui.on_info.assert_called_once()
+        assert "Usage" in ui.on_info.call_args[0][0]
+
+    def test_rewind_invalid_arg(self, tmp_db):
+        session = _make_session(tmp_db)
+        ui = session.ui
+        ui.on_info = MagicMock()
+        session.handle_command("/rewind abc")
+        ui.on_info.assert_called_once()
+        assert "integer" in ui.on_info.call_args[0][0]
+
+    def test_retry_nothing_to_retry(self, tmp_db):
+        session = _make_session(tmp_db)
+        ui = session.ui
+        ui.on_info = MagicMock()
+        session.handle_command("/retry")
+        ui.on_info.assert_called_once()
+        assert "Nothing" in ui.on_info.call_args[0][0]
+
+
+# ---------------------------------------------------------------------------
+# Storage integration — delete_messages_after
+# ---------------------------------------------------------------------------
+
+
+class TestDeleteMessagesAfter:
+    def test_delete_truncates_db(self, tmp_db):
+        from turnstone.core.memory import (
+            delete_messages_after,
+            load_messages,
+            register_workstream,
+            save_message,
+        )
+
+        ws_id = "test-ws-delete"
+        register_workstream(ws_id)
+        save_message(ws_id, "user", "Hello")
+        save_message(ws_id, "assistant", "Hi!")
+        save_message(ws_id, "user", "Bye")
+        save_message(ws_id, "assistant", "Goodbye!")
+
+        deleted = delete_messages_after(ws_id, 2)
+        assert deleted == 2
+
+        msgs = load_messages(ws_id)
+        assert len(msgs) == 2
+        assert msgs[0]["content"] == "Hello"
+        assert msgs[1]["content"] == "Hi!"
+
+    def test_delete_nothing(self, tmp_db):
+        from turnstone.core.memory import (
+            delete_messages_after,
+            register_workstream,
+            save_message,
+        )
+
+        ws_id = "test-ws-noop"
+        register_workstream(ws_id)
+        save_message(ws_id, "user", "Hello")
+
+        deleted = delete_messages_after(ws_id, 10)
+        assert deleted == 0
+
+    def test_delete_all(self, tmp_db):
+        from turnstone.core.memory import (
+            delete_messages_after,
+            load_messages,
+            register_workstream,
+            save_message,
+        )
+
+        ws_id = "test-ws-all"
+        register_workstream(ws_id)
+        save_message(ws_id, "user", "Hello")
+        save_message(ws_id, "assistant", "Hi!")
+
+        deleted = delete_messages_after(ws_id, 0)
+        assert deleted == 2
+        assert load_messages(ws_id) == []
+
+
+# ---------------------------------------------------------------------------
+# End-to-end: rewind + DB sync
+# ---------------------------------------------------------------------------
+
+
+class TestRewindDBSync:
+    def test_rewind_persists_to_db(self, tmp_db):
+        from turnstone.core.memory import load_messages, register_workstream, save_message
+
+        session = _make_session(tmp_db)
+        ws_id = session.ws_id
+        register_workstream(ws_id)
+
+        # Persist messages to DB and set in-memory state
+        save_message(ws_id, "user", "Hello")
+        save_message(ws_id, "assistant", "Hi!")
+        save_message(ws_id, "user", "Bye")
+        save_message(ws_id, "assistant", "Goodbye!")
+
+        session.messages = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi!"},
+            {"role": "user", "content": "Bye"},
+            {"role": "assistant", "content": "Goodbye!"},
+        ]
+        session._msg_tokens = [5, 5, 5, 5]
+
+        session.rewind(1)
+
+        # Verify DB matches in-memory state
+        db_msgs = load_messages(ws_id)
+        assert len(db_msgs) == 2
+        assert db_msgs[0]["content"] == "Hello"
+        assert db_msgs[1]["content"] == "Hi!"

tests/test_scheduled_tasks_storage.py

@@ -4,17 +4,6 @@ from __future__ import annotations
 
 import time
 
-import pytest
-
-from turnstone.core.storage._sqlite import SQLiteBackend
-
-
-@pytest.fixture
-def db(tmp_path):
-    """Fresh SQLite backend for each test."""
-    backend = SQLiteBackend(str(tmp_path / "test.db"))
-    return backend
-
 
 def _make_task_kwargs(**overrides):
     """Build default kwargs for create_scheduled_task."""

tests/test_sdk_governance.py

@@ -0,0 +1,440 @@
+"""Integration tests for SDK governance methods against a real Starlette app.
+
+Verifies round-trip serialization: SDK -> HTTP -> Starlette handler -> storage
+-> JSON response -> Pydantic model validation in the SDK client.
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any
+
+import httpx
+import pytest
+from starlette.applications import Starlette
+from starlette.middleware import Middleware
+from starlette.middleware.base import BaseHTTPMiddleware
+from starlette.routing import Mount, Route
+
+if TYPE_CHECKING:
+    from starlette.requests import Request
+    from starlette.responses import Response
+
+from turnstone.api.console_schemas import (
+    ListOrgsResponse,
+    ListRolesResponse,
+    ListToolPoliciesResponse,
+    OrgInfo,
+    RoleInfo,
+    ToolPolicyInfo,
+)
+from turnstone.api.schemas import StatusResponse
+from turnstone.console.server import (
+    admin_create_policy,
+    admin_create_role,
+    admin_delete_policy,
+    admin_delete_role,
+    admin_get_org,
+    admin_list_orgs,
+    admin_list_policies,
+    admin_list_roles,
+    admin_update_org,
+    admin_update_policy,
+    admin_update_role,
+)
+from turnstone.core.auth import AuthResult
+from turnstone.core.storage._sqlite import SQLiteBackend
+from turnstone.sdk.console import AsyncTurnstoneConsole
+
+# ---------------------------------------------------------------------------
+# Auth bypass middleware — injects a full-access AuthResult on every request.
+# ---------------------------------------------------------------------------
+
+
+class _InjectAuthMiddleware(BaseHTTPMiddleware):
+    async def dispatch(self, request: Request, call_next: Any) -> Response:
+        request.state.auth_result = AuthResult(
+            user_id="test-admin",
+            scopes=frozenset({"approve"}),
+            token_source="config",
+            permissions=frozenset(
+                {
+                    "read",
+                    "write",
+                    "approve",
+                    "admin.roles",
+                    "admin.orgs",
+                    "admin.policies",
+                }
+            ),
+        )
+        resp: Response = await call_next(request)
+        return resp
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+def _make_app() -> Starlette:
+    return Starlette(
+        routes=[
+            Mount(
+                "/v1",
+                routes=[
+                    # Roles
+                    Route("/api/admin/roles", admin_list_roles),
+                    Route("/api/admin/roles", admin_create_role, methods=["POST"]),
+                    Route("/api/admin/roles/{role_id}", admin_update_role, methods=["PUT"]),
+                    Route("/api/admin/roles/{role_id}", admin_delete_role, methods=["DELETE"]),
+                    # Orgs
+                    Route("/api/admin/orgs", admin_list_orgs),
+                    Route("/api/admin/orgs/{org_id}", admin_get_org),
+                    Route("/api/admin/orgs/{org_id}", admin_update_org, methods=["PUT"]),
+                    # Policies
+                    Route("/api/admin/policies", admin_list_policies),
+                    Route("/api/admin/policies", admin_create_policy, methods=["POST"]),
+                    Route(
+                        "/api/admin/policies/{policy_id}",
+                        admin_update_policy,
+                        methods=["PUT"],
+                    ),
+                    Route(
+                        "/api/admin/policies/{policy_id}",
+                        admin_delete_policy,
+                        methods=["DELETE"],
+                    ),
+                ],
+            ),
+        ],
+        middleware=[Middleware(_InjectAuthMiddleware)],
+    )
+
+
+@pytest.fixture
+def storage(tmp_path: Any) -> SQLiteBackend:
+    return SQLiteBackend(str(tmp_path / "test.db"))
+
+
+@pytest.fixture
+async def sdk_client(storage: SQLiteBackend):
+    """SDK client wired to a real Starlette app via ASGITransport."""
+    app = _make_app()
+    app.state.auth_storage = storage
+    transport = httpx.ASGITransport(app=app)
+    async with httpx.AsyncClient(transport=transport, base_url="http://testserver") as hc:
+        yield AsyncTurnstoneConsole(httpx_client=hc)
+
+
+# ---------------------------------------------------------------------------
+# Tests — Roles round-trip
+# ---------------------------------------------------------------------------
+
+
+class TestRolesRoundTrip:
+    @pytest.mark.anyio
+    async def test_list_roles_empty(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        resp = await sdk_client.list_roles()
+        assert isinstance(resp, ListRolesResponse)
+        assert resp.roles == []
+
+    @pytest.mark.anyio
+    async def test_create_and_list_role(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        role = await sdk_client.create_role(
+            "analyst", display_name="Data Analyst", permissions="read,write"
+        )
+        assert isinstance(role, RoleInfo)
+        assert role.name == "analyst"
+        assert role.display_name == "Data Analyst"
+        assert role.permissions == "read,write"
+        assert role.builtin is False
+        assert role.role_id  # non-empty
+
+        # List should now contain the new role
+        resp = await sdk_client.list_roles()
+        assert len(resp.roles) == 1
+        assert resp.roles[0].role_id == role.role_id
+        assert resp.roles[0].name == "analyst"
+
+    @pytest.mark.anyio
+    async def test_create_update_role(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        role = await sdk_client.create_role("ops", permissions="read")
+        assert role.permissions == "read"
+
+        updated = await sdk_client.update_role(
+            role.role_id, display_name="Operations", permissions="read,write,approve"
+        )
+        assert isinstance(updated, RoleInfo)
+        assert updated.display_name == "Operations"
+        assert updated.permissions == "read,write,approve"
+        assert updated.role_id == role.role_id
+
+    @pytest.mark.anyio
+    async def test_create_delete_role(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        role = await sdk_client.create_role("temp-role", permissions="read")
+
+        result = await sdk_client.delete_role(role.role_id)
+        assert isinstance(result, StatusResponse)
+        assert result.status == "ok"
+
+        # Verify gone
+        resp = await sdk_client.list_roles()
+        assert resp.roles == []
+
+    @pytest.mark.anyio
+    async def test_full_lifecycle(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        """Create -> list -> update -> list -> delete -> list."""
+        # Create
+        role = await sdk_client.create_role(
+            "lifecycle", display_name="Lifecycle", permissions="read"
+        )
+        role_id = role.role_id
+
+        # List confirms creation
+        roles = (await sdk_client.list_roles()).roles
+        assert len(roles) == 1
+        assert roles[0].role_id == role_id
+
+        # Update
+        updated = await sdk_client.update_role(role_id, permissions="read,write")
+        assert updated.permissions == "read,write"
+
+        # List still has one
+        roles = (await sdk_client.list_roles()).roles
+        assert len(roles) == 1
+        assert roles[0].permissions == "read,write"
+
+        # Delete
+        await sdk_client.delete_role(role_id)
+
+        # List is empty
+        roles = (await sdk_client.list_roles()).roles
+        assert roles == []
+
+    @pytest.mark.anyio
+    async def test_delete_nonexistent_role_raises(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        from turnstone.sdk._types import TurnstoneAPIError
+
+        with pytest.raises(TurnstoneAPIError) as exc_info:
+            await sdk_client.delete_role("nonexistent")
+        assert exc_info.value.status_code == 404
+
+    @pytest.mark.anyio
+    async def test_update_nonexistent_role_raises(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        from turnstone.sdk._types import TurnstoneAPIError
+
+        with pytest.raises(TurnstoneAPIError) as exc_info:
+            await sdk_client.update_role("nonexistent", display_name="Nope")
+        assert exc_info.value.status_code == 404
+
+
+# ---------------------------------------------------------------------------
+# Tests — Policies round-trip
+# ---------------------------------------------------------------------------
+
+
+class TestPoliciesRoundTrip:
+    @pytest.mark.anyio
+    async def test_list_policies_empty(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        resp = await sdk_client.list_policies()
+        assert isinstance(resp, ListToolPoliciesResponse)
+        assert resp.policies == []
+
+    @pytest.mark.anyio
+    async def test_create_and_list_policy(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        policy = await sdk_client.create_policy("Allow bash", "bash_*", "allow", priority=10)
+        assert isinstance(policy, ToolPolicyInfo)
+        assert policy.name == "Allow bash"
+        assert policy.tool_pattern == "bash_*"
+        assert policy.action == "allow"
+        assert policy.priority == 10
+        assert policy.enabled is True
+        assert policy.policy_id  # non-empty
+
+        resp = await sdk_client.list_policies()
+        assert len(resp.policies) == 1
+        assert resp.policies[0].policy_id == policy.policy_id
+
+    @pytest.mark.anyio
+    async def test_create_update_policy(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        policy = await sdk_client.create_policy("Deny write", "write_*", "deny", priority=5)
+
+        updated = await sdk_client.update_policy(
+            policy.policy_id, name="Allow write", action="allow", priority=20
+        )
+        assert isinstance(updated, ToolPolicyInfo)
+        assert updated.name == "Allow write"
+        assert updated.action == "allow"
+        assert updated.priority == 20
+        assert updated.policy_id == policy.policy_id
+
+    @pytest.mark.anyio
+    async def test_create_delete_policy(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        policy = await sdk_client.create_policy("Temp policy", "temp_*", "ask")
+
+        result = await sdk_client.delete_policy(policy.policy_id)
+        assert isinstance(result, StatusResponse)
+        assert result.status == "ok"
+
+        resp = await sdk_client.list_policies()
+        assert resp.policies == []
+
+    @pytest.mark.anyio
+    async def test_full_lifecycle(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        """Create -> list -> update -> list -> delete -> list."""
+        policy = await sdk_client.create_policy("Lifecycle", "test_*", "deny", priority=1)
+        pid = policy.policy_id
+
+        policies = (await sdk_client.list_policies()).policies
+        assert len(policies) == 1
+
+        await sdk_client.update_policy(pid, action="allow", priority=99)
+        policies = (await sdk_client.list_policies()).policies
+        assert policies[0].action == "allow"
+        assert policies[0].priority == 99
+
+        await sdk_client.delete_policy(pid)
+        policies = (await sdk_client.list_policies()).policies
+        assert policies == []
+
+    @pytest.mark.anyio
+    async def test_delete_nonexistent_policy_raises(
+        self, sdk_client: AsyncTurnstoneConsole
+    ) -> None:
+        from turnstone.sdk._types import TurnstoneAPIError
+
+        with pytest.raises(TurnstoneAPIError) as exc_info:
+            await sdk_client.delete_policy("nonexistent")
+        assert exc_info.value.status_code == 404
+
+    @pytest.mark.anyio
+    async def test_update_nonexistent_policy_raises(
+        self, sdk_client: AsyncTurnstoneConsole
+    ) -> None:
+        from turnstone.sdk._types import TurnstoneAPIError
+
+        with pytest.raises(TurnstoneAPIError) as exc_info:
+            await sdk_client.update_policy("nonexistent", name="Nope")
+        assert exc_info.value.status_code == 404
+
+    @pytest.mark.anyio
+    async def test_create_policy_invalid_action_raises(
+        self, sdk_client: AsyncTurnstoneConsole
+    ) -> None:
+        from turnstone.sdk._types import TurnstoneAPIError
+
+        with pytest.raises(TurnstoneAPIError) as exc_info:
+            await sdk_client.create_policy("Bad", "tool_*", "yolo")
+        assert exc_info.value.status_code == 400
+
+
+# ---------------------------------------------------------------------------
+# Tests — Orgs round-trip
+# ---------------------------------------------------------------------------
+
+
+class TestOrgsRoundTrip:
+    @pytest.mark.anyio
+    async def test_list_orgs_empty(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        resp = await sdk_client.list_orgs()
+        assert isinstance(resp, ListOrgsResponse)
+        assert resp.orgs == []
+
+    @pytest.mark.anyio
+    async def test_get_org(self, sdk_client: AsyncTurnstoneConsole, storage: SQLiteBackend) -> None:
+        storage.create_org(
+            org_id="org-1", name="acme", display_name="Acme Corp", settings='{"k": "v"}'
+        )
+        org = await sdk_client.get_org("org-1")
+        assert isinstance(org, OrgInfo)
+        assert org.org_id == "org-1"
+        assert org.name == "acme"
+        assert org.display_name == "Acme Corp"
+        assert org.settings == '{"k": "v"}'
+
+    @pytest.mark.anyio
+    async def test_list_orgs_after_seed(
+        self, sdk_client: AsyncTurnstoneConsole, storage: SQLiteBackend
+    ) -> None:
+        storage.create_org(org_id="org-a", name="alpha", display_name="Alpha")
+        storage.create_org(org_id="org-b", name="beta", display_name="Beta")
+
+        resp = await sdk_client.list_orgs()
+        assert len(resp.orgs) == 2
+        names = {o.name for o in resp.orgs}
+        assert names == {"alpha", "beta"}
+
+    @pytest.mark.anyio
+    async def test_update_org(
+        self, sdk_client: AsyncTurnstoneConsole, storage: SQLiteBackend
+    ) -> None:
+        storage.create_org(org_id="org-1", name="acme", display_name="Acme Corp")
+
+        updated = await sdk_client.update_org("org-1", display_name="Acme Inc.")
+        assert isinstance(updated, OrgInfo)
+        assert updated.display_name == "Acme Inc."
+        assert updated.org_id == "org-1"
+
+    @pytest.mark.anyio
+    async def test_get_nonexistent_org_raises(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        from turnstone.sdk._types import TurnstoneAPIError
+
+        with pytest.raises(TurnstoneAPIError) as exc_info:
+            await sdk_client.get_org("nonexistent")
+        assert exc_info.value.status_code == 404
+
+    @pytest.mark.anyio
+    async def test_update_nonexistent_org_raises(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        from turnstone.sdk._types import TurnstoneAPIError
+
+        with pytest.raises(TurnstoneAPIError) as exc_info:
+            await sdk_client.update_org("nonexistent", display_name="Nope")
+        assert exc_info.value.status_code == 404
+
+
+# ---------------------------------------------------------------------------
+# Tests — Pydantic model field validation
+# ---------------------------------------------------------------------------
+
+
+class TestModelValidation:
+    """Verify that all expected fields are populated and correctly typed."""
+
+    @pytest.mark.anyio
+    async def test_role_info_fields(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        role = await sdk_client.create_role("reviewer", permissions="read")
+        assert isinstance(role.role_id, str)
+        assert isinstance(role.name, str)
+        assert isinstance(role.display_name, str)
+        assert isinstance(role.permissions, str)
+        assert isinstance(role.builtin, bool)
+        assert isinstance(role.org_id, str)
+        assert isinstance(role.created, str)
+        assert isinstance(role.updated, str)
+
+    @pytest.mark.anyio
+    async def test_policy_info_fields(self, sdk_client: AsyncTurnstoneConsole) -> None:
+        policy = await sdk_client.create_policy("Test", "read_*", "allow", priority=5)
+        assert isinstance(policy.policy_id, str)
+        assert isinstance(policy.name, str)
+        assert isinstance(policy.tool_pattern, str)
+        assert isinstance(policy.action, str)
+        assert isinstance(policy.priority, int)
+        assert isinstance(policy.org_id, str)
+        assert isinstance(policy.enabled, bool)
+        assert isinstance(policy.created_by, str)
+        assert isinstance(policy.created, str)
+        assert isinstance(policy.updated, str)
+
+    @pytest.mark.anyio
+    async def test_org_info_fields(
+        self, sdk_client: AsyncTurnstoneConsole, storage: SQLiteBackend
+    ) -> None:
+        storage.create_org(org_id="org-v", name="validate", display_name="Validate")
+        org = await sdk_client.get_org("org-v")
+        assert isinstance(org.org_id, str)
+        assert isinstance(org.name, str)
+        assert isinstance(org.display_name, str)
+        assert isinstance(org.settings, str)
+        assert isinstance(org.created, str)
+        assert isinstance(org.updated, str)

tests/test_server_live.py

@@ -88,7 +88,7 @@ class RecordingUI:
     def approve_tools(self, items):
         return True, None  # auto-approve everything
 
-    def on_tool_result(self, call_id, name, output):
+    def on_tool_result(self, call_id, name, output, **kwargs):
         self.tool_results.append((call_id, name, output))
 
     def on_tool_output_chunk(self, call_id, chunk):
@@ -623,6 +623,7 @@ class TestServerHealthMetrics:
 
         mock_mgr = MagicMock()
         mock_mgr.list_all.return_value = [mock_ws]
+        mock_mgr.max_workstreams = 10
 
         app = srv_mod.create_app(
             workstreams=mock_mgr,
@@ -799,6 +800,7 @@ class TestServerRateLimiting:
 
         mock_mgr = MagicMock()
         mock_mgr.list_all.return_value = [mock_ws]
+        mock_mgr.max_workstreams = 10
 
         app = srv_mod.create_app(
             workstreams=mock_mgr,

tests/test_services_storage.py

@@ -2,15 +2,6 @@
 
 from __future__ import annotations
 
-import pytest
-
-from turnstone.core.storage._sqlite import SQLiteBackend
-
-
-@pytest.fixture
-def storage(tmp_path):
-    return SQLiteBackend(str(tmp_path / "test.db"))
-
 
 class TestServiceRegistry:
     def test_register_and_list(self, storage):

tests/test_session.py

@@ -28,7 +28,7 @@ class NullUI:
     def approve_tools(self, items):
         return True, None
 
-    def on_tool_result(self, call_id, name, output):
+    def on_tool_result(self, call_id, name, output, **kwargs):
         pass
 
     def on_tool_output_chunk(self, call_id, chunk):
@@ -214,7 +214,7 @@ class TestPlanExec:
                         "id": tc_id,
                         "type": "function",
                         "function": {
-                            "name": "create_plan",
+                            "name": "plan_agent",
                             "arguments": json.dumps({"goal": prior_prompt}),
                         },
                     }
@@ -250,7 +250,7 @@ class TestPlanExec:
             m for m in messages if m["role"] == "assistant" and m.get("tool_calls")
         ]
         assert len(assistant_with_tc) == 1
-        assert assistant_with_tc[0]["tool_calls"][0]["function"]["name"] == "create_plan"
+        assert assistant_with_tc[0]["tool_calls"][0]["function"]["name"] == "plan_agent"
 
         # The real tool result is forwarded with its original content
         tool_msgs = [m for m in messages if m["role"] == "tool"]
@@ -463,7 +463,7 @@ class TestPlanRefinement:
         with patch.object(session, "_refine_plan", side_effect=fake_refine):
             items = [
                 {
-                    "func_name": "create_plan",
+                    "func_name": "plan_agent",
                     "call_id": "c1",
                     "prompt": "add auth",
                 }
@@ -576,7 +576,7 @@ class TestPlanRefinement:
         msgs = captured["messages"]
         assert msgs[0]["role"] == "system"
         assert msgs[1]["role"] == "assistant"
-        assert msgs[1]["tool_calls"][0]["function"]["name"] == "create_plan"
+        assert msgs[1]["tool_calls"][0]["function"]["name"] == "plan_agent"
         assert msgs[2]["role"] == "tool"
         assert msgs[2]["content"] == self.GOOD_PLAN
         assert msgs[3]["role"] == "user"
@@ -753,3 +753,268 @@ class TestGetCapabilitiesOverride:
         caps = session._get_capabilities()
         # Default OpenAI provider for unknown model → no vision
         assert caps.supports_vision is False
+
+
+class TestTitleRetry:
+    """_generate_title resets _title_generated on failure."""
+
+    def test_title_generated_reset_on_failure(self, tmp_db):
+        session = _make_session()
+        session._title_generated = True
+        session.messages = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there"},
+        ]
+        # Mock provider to raise
+        session._provider = MagicMock()
+        session._provider.create_completion.side_effect = RuntimeError("API error")
+
+        session._generate_title()
+
+        assert session._title_generated is False
+
+    def test_title_generated_stays_true_on_success(self, tmp_db):
+        session = _make_session()
+        session._title_generated = True
+        session.messages = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there"},
+        ]
+        result = MagicMock()
+        result.content = "Test Title"
+        session._provider = MagicMock()
+        session._provider.create_completion.return_value = result
+
+        with patch("turnstone.core.session.update_workstream_title"):
+            session._generate_title()
+
+        # Flag stays True after successful generation
+        assert session._title_generated is True
+
+    def test_title_skipped_after_resume_changes_ws_id(self, tmp_db):
+        """If ws_id changes (via resume) during title generation, discard the result."""
+        session = _make_session()
+        session._title_generated = True
+        session.messages = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there"},
+        ]
+        original_ws_id = session._ws_id
+        result = MagicMock()
+        result.content = "Test Title"
+        session._provider = MagicMock()
+        session._provider.create_completion.return_value = result
+
+        # Simulate resume() changing ws_id while title generation is in flight
+        def _change_ws_id(*args, **kwargs):
+            session._ws_id = "different-ws-id"
+            return result
+
+        session._provider.create_completion.side_effect = _change_ws_id
+
+        with patch("turnstone.core.session.update_workstream_title") as mock_update:
+            session._generate_title()
+
+        # Title should NOT be applied to the new workstream
+        mock_update.assert_not_called()
+        # Restore for cleanup
+        session._ws_id = original_ws_id
+
+
+class TestLiveConfigUpdate:
+    """ConfigStore-backed sessions pick up settings changes at point-of-use."""
+
+    def test_memory_config_reads_from_config_store(self, tmp_db):
+        """_mem_cfg returns live values from ConfigStore when present."""
+        from turnstone.core.config_store import ConfigStore
+        from turnstone.core.storage._sqlite import SQLiteBackend
+
+        storage = SQLiteBackend(str(tmp_db), create_tables=True)
+        cs = ConfigStore(storage)
+        session = _make_session(config_store=cs)
+
+        # Default: relevance_k=5
+        assert session._mem_cfg.relevance_k == 5
+
+        # Admin changes the setting
+        cs.set("memory.relevance_k", 10, changed_by="test")
+        assert session._mem_cfg.relevance_k == 10
+
+    def test_judge_config_reads_from_config_store(self, tmp_db):
+        """_judge_cfg returns live behavioral flags from ConfigStore."""
+        from turnstone.core.config_store import ConfigStore
+        from turnstone.core.judge import JudgeConfig
+        from turnstone.core.storage._sqlite import SQLiteBackend
+
+        storage = SQLiteBackend(str(tmp_db), create_tables=True)
+        cs = ConfigStore(storage)
+        session = _make_session(
+            judge_config=JudgeConfig(),
+            config_store=cs,
+        )
+
+        # Default: enabled=True
+        assert session._judge_cfg.enabled is True
+
+        # Admin disables the judge
+        cs.set("judge.enabled", False, changed_by="test")
+        assert session._judge_cfg.enabled is False
+
+    def test_judge_client_config_stays_frozen(self, tmp_db):
+        """LLM client fields (model, provider) are frozen from creation time."""
+        from turnstone.core.config_store import ConfigStore
+        from turnstone.core.judge import JudgeConfig
+        from turnstone.core.storage._sqlite import SQLiteBackend
+
+        storage = SQLiteBackend(str(tmp_db), create_tables=True)
+        cs = ConfigStore(storage)
+        session = _make_session(
+            judge_config=JudgeConfig(model="original-model"),
+            config_store=cs,
+        )
+
+        # Change the model in ConfigStore — should NOT affect the session
+        cs.set("judge.model", "new-model", changed_by="test")
+        assert session._judge_cfg.model == "original-model"
+
+    def test_judge_disable_after_init_stops_future_use(self, tmp_db):
+        """Disabling judge.enabled after IntentJudge is created returns None."""
+        from turnstone.core.config_store import ConfigStore
+        from turnstone.core.judge import JudgeConfig
+        from turnstone.core.storage._sqlite import SQLiteBackend
+
+        storage = SQLiteBackend(str(tmp_db), create_tables=True)
+        cs = ConfigStore(storage)
+        session = _make_session(
+            judge_config=JudgeConfig(),
+            config_store=cs,
+        )
+
+        # Force judge initialization by setting a mock
+        session._judge = MagicMock()
+        assert session._ensure_judge() is not None
+
+        # Admin disables the judge — cached instance should NOT be returned
+        cs.set("judge.enabled", False, changed_by="test")
+        assert session._ensure_judge() is None
+
+    def test_fallback_to_frozen_without_config_store(self, tmp_db):
+        """Without ConfigStore (CLI mode), frozen config is used."""
+        from turnstone.core.memory_relevance import MemoryConfig
+
+        session = _make_session(memory_config=MemoryConfig(relevance_k=3))
+        assert session._mem_cfg.relevance_k == 3
+
+
+class TestAgentOutputGuard:
+    """Output guard should evaluate tool results in _run_agent, not just the main loop."""
+
+    def test_agent_loop_calls_evaluate_output(self):
+        """_run_agent passes tool output through _evaluate_output when output_guard is enabled."""
+        from turnstone.core.judge import JudgeConfig
+
+        session = _make_session(judge_config=JudgeConfig(output_guard=True))
+
+        with patch.object(session, "_evaluate_output", wraps=lambda cid, o, fn: o) as mock_eval:
+            # Simulate _run_agent getting a tool call response then a text response
+            call_count = [0]
+
+            def fake_create(**kwargs):
+                call_count[0] += 1
+                resp = MagicMock()
+                if call_count[0] == 1:
+                    # First call: model returns a tool call
+                    choice = MagicMock()
+                    choice.finish_reason = "tool_calls"
+                    tc = MagicMock()
+                    tc.id = "call_1"
+                    tc.function.name = "read_file"
+                    tc.function.arguments = '{"path": "/tmp/test"}'
+                    choice.message.tool_calls = [tc]
+                    choice.message.content = None
+                    resp.choices = [choice]
+                    resp.usage = MagicMock(prompt_tokens=10, completion_tokens=5)
+                else:
+                    # Second call: model returns text (done)
+                    choice = MagicMock()
+                    choice.finish_reason = "stop"
+                    choice.message.tool_calls = None
+                    choice.message.content = "Done"
+                    resp.choices = [choice]
+                    resp.usage = MagicMock(prompt_tokens=10, completion_tokens=5)
+                return resp
+
+            session.client.chat.completions.create = fake_create
+
+            # Mock tool preparation to return a simple output
+            def fake_prepare(tc_dict, **kwargs):
+                return {
+                    "call_id": tc_dict["id"],
+                    "func_name": "read_file",
+                    "needs_approval": False,
+                    "execute": lambda p: ("call_1", "file contents with sk-proj-SECRET123"),
+                }
+
+            with patch.object(session, "_prepare_tool", side_effect=fake_prepare):
+                session._run_agent(
+                    [{"role": "user", "content": "test"}],
+                    tools=[{"type": "function", "function": {"name": "read_file"}}],
+                    label="test",
+                )
+
+            mock_eval.assert_called_once()
+            args = mock_eval.call_args[0]
+            assert args[0] == "call_1"  # call_id
+            assert "sk-proj-SECRET123" in args[1]  # output
+            assert args[2] == "read_file"  # func_name
+
+    def test_agent_loop_skips_guard_when_disabled(self):
+        """_run_agent does not call _evaluate_output when output_guard is disabled."""
+        from turnstone.core.judge import JudgeConfig
+
+        session = _make_session(judge_config=JudgeConfig(output_guard=False))
+
+        with patch.object(session, "_evaluate_output") as mock_eval:
+            call_count = [0]
+
+            def fake_create(**kwargs):
+                call_count[0] += 1
+                resp = MagicMock()
+                if call_count[0] == 1:
+                    choice = MagicMock()
+                    choice.finish_reason = "tool_calls"
+                    tc = MagicMock()
+                    tc.id = "call_1"
+                    tc.function.name = "read_file"
+                    tc.function.arguments = '{"path": "/tmp/test"}'
+                    choice.message.tool_calls = [tc]
+                    choice.message.content = None
+                    resp.choices = [choice]
+                    resp.usage = MagicMock(prompt_tokens=10, completion_tokens=5)
+                else:
+                    choice = MagicMock()
+                    choice.finish_reason = "stop"
+                    choice.message.tool_calls = None
+                    choice.message.content = "Done"
+                    resp.choices = [choice]
+                    resp.usage = MagicMock(prompt_tokens=10, completion_tokens=5)
+                return resp
+
+            session.client.chat.completions.create = fake_create
+
+            def fake_prepare(tc_dict, **kwargs):
+                return {
+                    "call_id": tc_dict["id"],
+                    "func_name": "read_file",
+                    "needs_approval": False,
+                    "execute": lambda p: ("call_1", "safe output"),
+                }
+
+            with patch.object(session, "_prepare_tool", side_effect=fake_prepare):
+                session._run_agent(
+                    [{"role": "user", "content": "test"}],
+                    tools=[{"type": "function", "function": {"name": "read_file"}}],
+                    label="test",
+                )
+
+            mock_eval.assert_not_called()

tests/test_sessions.py

@@ -525,6 +525,37 @@ class TestWorkstreamConfig:
         assert session.instructions == "be concise"
         assert session.creative_mode is True
 
+    def test_resume_restores_model(self, tmp_db):
+        """ChatSession.resume() should restore the model from workstream config."""
+        client = MagicMock()
+        client.models.list.return_value.data = [MagicMock(id="test-model")]
+        ui = MagicMock()
+        ui.on_info = MagicMock()
+        ui.on_error = MagicMock()
+        ui.on_state_change = MagicMock()
+        ui.on_rename = MagicMock()
+
+        # Create a workstream that was using a specific model
+        register_workstream("model_ws")
+        save_message("model_ws", "user", "hello")
+        save_message("model_ws", "assistant", "hi")
+        save_workstream_config("model_ws", {"model": "gpt-5", "model_alias": ""})
+
+        # Resume into a session that was created with a different model
+        session = ChatSession(
+            client=client,
+            model="gpt-5-nano",
+            ui=ui,
+            instructions=None,
+            temperature=0.5,
+            max_tokens=1000,
+            tool_timeout=10,
+        )
+        assert session.model == "gpt-5-nano"
+        result = session.resume("model_ws")
+        assert result is True
+        assert session.model == "gpt-5"
+
 
 # ── Prune workstreams ─────────────────────────────────────────────────
 
@@ -607,3 +638,298 @@ class TestPruneWorkstreams:
         # Config rows should be cleaned up
         assert load_workstream_config("orphan_cfg") == {}
         assert load_workstream_config("stale_cfg") == {}
+
+
+# ── Parallel tool exception isolation ────────────────────────────────
+
+
+class TestParallelToolExceptionIsolation:
+    """Bug #117: one tool raising should not kill the entire batch."""
+
+    def test_exception_in_one_tool_does_not_kill_batch(self, tmp_db, mock_openai_client):
+        from unittest.mock import patch
+
+        session = ChatSession(
+            client=mock_openai_client,
+            model="test-model",
+            ui=MagicMock(),
+            instructions=None,
+            temperature=0.5,
+            max_tokens=1000,
+            tool_timeout=10,
+        )
+
+        def succeed(item):
+            return item["call_id"], "ok"
+
+        def fail(item):
+            raise RuntimeError("boom")
+
+        items = [
+            {
+                "call_id": "c1",
+                "func_name": "bash",
+                "execute": succeed,
+                "needs_approval": False,
+                "header": "test",
+                "preview": "",
+            },
+            {
+                "call_id": "c2",
+                "func_name": "math",
+                "execute": fail,
+                "needs_approval": False,
+                "header": "test",
+                "preview": "",
+            },
+        ]
+
+        tool_calls = [
+            {"id": "c1", "function": {"name": "bash", "arguments": "{}"}},
+            {"id": "c2", "function": {"name": "math", "arguments": "{}"}},
+        ]
+
+        with (
+            patch.object(session, "_prepare_tool", side_effect=items),
+            patch.object(session, "_evaluate_intent"),
+            patch.object(session, "_emit_state"),
+            patch.object(session, "_init_system_messages"),
+            patch.object(session, "_check_cancelled"),
+        ):
+            session.ui.approve_tools.return_value = (True, None)
+            results, _ = session._execute_tools(tool_calls)
+
+        assert results[0] == ("c1", "ok")
+        assert results[1][0] == "c2"
+        assert "Error executing math" in results[1][1]
+        assert "boom" in results[1][1]
+
+
+# ── Web search tool gating ───────────────────────────────────────────
+
+
+class TestWebSearchGating:
+    """Bug #117: web_search should not be offered without a backend."""
+
+    def test_web_search_filtered_when_no_backend(self, tmp_db, mock_openai_client):
+        from unittest.mock import patch
+
+        from turnstone.core.providers._protocol import ModelCapabilities
+
+        session = ChatSession(
+            client=mock_openai_client,
+            model="local-model",
+            ui=MagicMock(),
+            instructions=None,
+            temperature=0.5,
+            max_tokens=1000,
+            tool_timeout=10,
+        )
+
+        caps = ModelCapabilities(supports_web_search=False)
+        with (
+            patch.object(session, "_get_capabilities", return_value=caps),
+            patch("turnstone.core.session.get_tavily_key", return_value=None),
+            patch("turnstone.core.web_search._ddg_available", return_value=False),
+        ):
+            tools = session._get_active_tools()
+
+        names = [t.get("function", {}).get("name") for t in tools]
+        assert "web_search" not in names
+
+    def test_web_search_kept_when_tavily_available(self, tmp_db, mock_openai_client):
+        from unittest.mock import patch
+
+        from turnstone.core.providers._protocol import ModelCapabilities
+
+        session = ChatSession(
+            client=mock_openai_client,
+            model="local-model",
+            ui=MagicMock(),
+            instructions=None,
+            temperature=0.5,
+            max_tokens=1000,
+            tool_timeout=10,
+        )
+
+        caps = ModelCapabilities(supports_web_search=False)
+        with (
+            patch.object(session, "_get_capabilities", return_value=caps),
+            patch("turnstone.core.session.get_tavily_key", return_value="tvly-test-key"),
+        ):
+            tools = session._get_active_tools()
+
+        names = [t.get("function", {}).get("name") for t in tools]
+        assert "web_search" in names
+
+    def test_web_search_kept_when_native_support(self, tmp_db, mock_openai_client):
+        from unittest.mock import patch
+
+        from turnstone.core.providers._protocol import ModelCapabilities
+
+        session = ChatSession(
+            client=mock_openai_client,
+            model="gpt-5-search-api",
+            ui=MagicMock(),
+            instructions=None,
+            temperature=0.5,
+            max_tokens=1000,
+            tool_timeout=10,
+        )
+
+        caps = ModelCapabilities(supports_web_search=True)
+        with (
+            patch.object(session, "_get_capabilities", return_value=caps),
+            patch("turnstone.core.session.get_tavily_key", return_value=None),
+        ):
+            tools = session._get_active_tools()
+
+        names = [t.get("function", {}).get("name") for t in tools]
+        assert "web_search" in names
+
+
+class TestMCPToolGating:
+    """MCP tools should not be offered when no MCP servers provide them."""
+
+    def test_mcp_tools_filtered_without_mcp_client(self, tmp_db, mock_openai_client):
+        """read_resource and use_prompt excluded when no MCP client."""
+        session = ChatSession(
+            client=mock_openai_client,
+            model="local-model",
+            ui=MagicMock(),
+            instructions=None,
+            temperature=0.5,
+            max_tokens=1000,
+            tool_timeout=10,
+        )
+        assert session._mcp_client is None
+
+        tools = session._get_active_tools()
+        names = [t.get("function", {}).get("name") for t in tools]
+        assert "read_resource" not in names
+        assert "use_prompt" not in names
+
+    def test_read_resource_filtered_when_no_resources(self, tmp_db, mock_openai_client):
+        """read_resource excluded when MCP client has no resources."""
+        mcp_client = MagicMock()
+        mcp_client.get_tools.return_value = []
+        mcp_client.resource_count = 0
+        mcp_client.prompt_count = 2
+
+        session = ChatSession(
+            client=mock_openai_client,
+            model="local-model",
+            ui=MagicMock(),
+            instructions=None,
+            temperature=0.5,
+            max_tokens=1000,
+            tool_timeout=10,
+            mcp_client=mcp_client,
+        )
+
+        tools = session._get_active_tools()
+        names = [t.get("function", {}).get("name") for t in tools]
+        assert "read_resource" not in names
+        assert "use_prompt" in names
+
+    def test_use_prompt_filtered_when_no_prompts(self, tmp_db, mock_openai_client):
+        """use_prompt excluded when MCP client has no prompts."""
+        mcp_client = MagicMock()
+        mcp_client.get_tools.return_value = []
+        mcp_client.resource_count = 3
+        mcp_client.prompt_count = 0
+
+        session = ChatSession(
+            client=mock_openai_client,
+            model="local-model",
+            ui=MagicMock(),
+            instructions=None,
+            temperature=0.5,
+            max_tokens=1000,
+            tool_timeout=10,
+            mcp_client=mcp_client,
+        )
+
+        tools = session._get_active_tools()
+        names = [t.get("function", {}).get("name") for t in tools]
+        assert "use_prompt" not in names
+        assert "read_resource" in names
+
+    def test_mcp_tools_kept_when_servers_have_both(self, tmp_db, mock_openai_client):
+        """Both tools present when MCP client has resources and prompts."""
+        mcp_client = MagicMock()
+        mcp_client.get_tools.return_value = []
+        mcp_client.resource_count = 1
+        mcp_client.prompt_count = 1
+
+        session = ChatSession(
+            client=mock_openai_client,
+            model="local-model",
+            ui=MagicMock(),
+            instructions=None,
+            temperature=0.5,
+            max_tokens=1000,
+            tool_timeout=10,
+            mcp_client=mcp_client,
+        )
+
+        tools = session._get_active_tools()
+        names = [t.get("function", {}).get("name") for t in tools]
+        assert "read_resource" in names
+        assert "use_prompt" in names
+
+    def test_mcp_tools_filtered_with_tool_search_active(self, tmp_db, mock_openai_client):
+        """Gating applies even when tool_search is active (client-side path)."""
+        mcp_client = MagicMock()
+        mcp_client.get_tools.return_value = []
+        mcp_client.resource_count = 0
+        mcp_client.prompt_count = 0
+
+        session = ChatSession(
+            client=mock_openai_client,
+            model="local-model",
+            ui=MagicMock(),
+            instructions=None,
+            temperature=0.5,
+            max_tokens=1000,
+            tool_timeout=10,
+            mcp_client=mcp_client,
+            tool_search="on",
+        )
+        assert session._tool_search is not None
+
+        tools = session._get_active_tools()
+        names = [t.get("function", {}).get("name") for t in tools]
+        assert "read_resource" not in names
+        assert "use_prompt" not in names
+
+    def test_mcp_tools_filtered_with_native_tool_search(self, tmp_db, mock_openai_client):
+        """Gating applies when provider handles tool search natively."""
+        from unittest.mock import patch
+
+        from turnstone.core.providers._protocol import ModelCapabilities
+
+        mcp_client = MagicMock()
+        mcp_client.get_tools.return_value = []
+        mcp_client.resource_count = 0
+        mcp_client.prompt_count = 0
+
+        session = ChatSession(
+            client=mock_openai_client,
+            model="local-model",
+            ui=MagicMock(),
+            instructions=None,
+            temperature=0.5,
+            max_tokens=1000,
+            tool_timeout=10,
+            mcp_client=mcp_client,
+            tool_search="on",
+        )
+
+        caps = ModelCapabilities(supports_tool_search=True)
+        with patch.object(session, "_get_capabilities", return_value=caps):
+            tools = session._get_active_tools()
+
+        names = [t.get("function", {}).get("name") for t in tools]
+        assert "read_resource" not in names
+        assert "use_prompt" not in names

tests/test_settings_api.py

@@ -179,7 +179,10 @@ class TestDeleteSetting:
         # Delete it
         r = client.delete("/v1/api/admin/settings/tools.timeout")
         assert r.status_code == 200
-        assert r.json()["status"] == "ok"
+        body = r.json()
+        assert body["status"] == "ok"
+        assert body["key"] == "tools.timeout"
+        assert body["default"] == 120  # registry default for tools.timeout
 
     def test_delete_then_list_shows_default(self, client):
         client.put(
@@ -251,20 +254,46 @@ class TestSecretMasking:
         by_key = {s["key"]: s for s in r.json()["settings"]}
         assert by_key["judge.api_key"]["value"] == "***"
 
-    def test_secret_write_blocked(self, client):
-        """Secret settings cannot be modified via API."""
+    def test_secret_writable_via_api(self, client):
+        """Secret settings can be written via API (write-only pattern)."""
         r = client.put(
             "/v1/api/admin/settings/judge.api_key",
             json={"value": "sk-secret-123"},
         )
-        assert r.status_code == 403
-        assert "config.toml" in r.json()["error"]
+        assert r.status_code == 200
+        # Response value is masked even for the write confirmation
+        assert r.json()["value"] == "***"
 
-    def test_secret_shows_managed_label(self, client):
-        """Secret settings show a label instead of a value."""
+    def test_secret_sentinel_preserves_existing(self, client):
+        """Submitting '***' for a secret setting is a no-op (preserve existing)."""
+        # First write a real value
+        r1 = client.put(
+            "/v1/api/admin/settings/judge.api_key",
+            json={"value": "sk-real-key"},
+        )
+        assert r1.status_code == 200
+        # Now submit the sentinel — should return unchanged with full response shape
+        r2 = client.put(
+            "/v1/api/admin/settings/judge.api_key",
+            json={"value": "***"},
+        )
+        assert r2.status_code == 200
+        data = r2.json()
+        assert data.get("unchanged") is True
+        assert data["key"] == "judge.api_key"
+        assert data["value"] == "***"
+        assert data["type"] == "str"
+        assert data["is_secret"] is True
+
+    def test_secret_still_masked_in_list(self, client):
+        """After writing a secret, list still shows '***'."""
+        client.put(
+            "/v1/api/admin/settings/judge.api_key",
+            json={"value": "sk-written-via-api"},
+        )
         r = client.get("/v1/api/admin/settings")
         by_key = {s["key"]: s for s in r.json()["settings"]}
-        assert "managed via" in by_key["judge.api_key"]["value"]
+        assert by_key["judge.api_key"]["value"] == "***"
 
 
 # ---------------------------------------------------------------------------

tests/test_sim.py

@@ -74,7 +74,7 @@ class TestSimEngine:
 
     def test_llm_response_returns_content(self, engine):
         async def _test():
-            content, tool_calls = await engine.simulate_llm_response(True, 1)
+            content, tool_calls = await engine.simulate_llm_response(True)
             assert isinstance(content, str)
             assert len(content) > 0
             assert isinstance(tool_calls, list)
@@ -85,8 +85,8 @@ class TestSimEngine:
         async def _test():
             e1 = SimEngine(fast_config, rng=random.Random(123))
             e2 = SimEngine(fast_config, rng=random.Random(123))
-            c1, t1 = await e1.simulate_llm_response(True, 1)
-            c2, t2 = await e2.simulate_llm_response(True, 1)
+            c1, t1 = await e1.simulate_llm_response(True)
+            c2, t2 = await e2.simulate_llm_response(True)
             assert c1 == c2
             assert len(t1) == len(t2)