fix(docker): drop the module docstring

docker.from_env() ignores the current docker context, so Docker Desktop on
macOS (socket under ~/.docker/run unless the default-socket option is on),
OrbStack and Colima reported DOCKER NOT AVAILABLE while `docker ps` worked.
Every failure also printed the same "ensure Docker Desktop is running" text
followed by a RuntimeError traceback.

- resolve the endpoint as the CLI does: DOCKER_HOST, then DOCKER_CONTEXT or
  the current context, then the default socket; the sandbox backend uses the
  same resolution so startup and scan talk to the same daemon
- classify the SDK error (socket missing, permission denied, connection
  refused, Windows named pipe) and print the fix for the current platform,
  the endpoint that was tried and the underlying error
- exit 1 cleanly instead of raising after the panel
- telemetry reports docker_unavailable_<reason>

.github/actions/setup-python-env/action.yml

@@ -0,0 +1,21 @@
+name: Set up Python environment
+description: Install Python and uv, then sync the locked project environment.
+
+inputs:
+  python-version:
+    description: Python version to install.
+    default: '3.12'
+
+runs:
+  using: composite
+  steps:
+    - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065  # v5.6.0
+      with:
+        python-version: ${{ inputs.python-version }}
+    - uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86  # v5.4.2
+      with:
+        enable-cache: true
+    - run: uv sync --frozen --python "$PYTHON_VERSION"
+      shell: bash
+      env:
+        PYTHON_VERSION: ${{ inputs.python-version }}

.github/dependabot.yml

@@ -0,0 +1,49 @@
+version: 2
+
+updates:
+  - package-ecosystem: github-actions
+    directory: /
+    schedule:
+      interval: weekly
+    cooldown:
+      default-days: 7
+    groups:
+      actions:
+        patterns:
+          - '*'
+
+  - package-ecosystem: uv
+    directory: /
+    schedule:
+      interval: weekly
+    cooldown:
+      default-days: 7
+    groups:
+      python:
+        update-types:
+          - minor
+          - patch
+
+  - package-ecosystem: gomod
+    directory: /strix/interface/tui
+    schedule:
+      interval: weekly
+    cooldown:
+      default-days: 7
+    groups:
+      go:
+        update-types:
+          - minor
+          - patch
+
+  - package-ecosystem: npm
+    directory: /strix/interface/viewer/frontend
+    schedule:
+      interval: weekly
+    cooldown:
+      default-days: 7
+    groups:
+      viewer:
+        update-types:
+          - minor
+          - patch

.github/workflows/build-release.yml

@@ -43,12 +43,14 @@ jobs:
           python-version: '3.12'
 
       - uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86  # v5.4.2
+        with:
+          enable-cache: false
 
       - uses: actions/setup-go@44694675825211faa026b3c33043df3e48a5fa00  # v6.0.0
         with:
           go-version: '1.24.x'
           check-latest: true
-          cache-dependency-path: strix/interface/tui/go.sum
+          cache: false
 
       - name: Build
         shell: bash
@@ -70,7 +72,10 @@ jobs:
             TUI_NAME="strix-tui"
             dist/strix --version
           fi
-          uv run pyi-archive_viewer -l "$PYI_BINARY" | grep -E "strix[/\\]+bin[/\\]+$TUI_NAME" >/dev/null
+          uv run pyi-archive_viewer -l -r "$PYI_BINARY" > archive.txt
+          grep -E "strix[/\\]+bin[/\\]+$TUI_NAME" archive.txt >/dev/null
+          grep -E "'google\.auth\.transport\.requests'" archive.txt >/dev/null
+          grep -E "'boto3\.session'" archive.txt >/dev/null
 
           if [[ "${{ matrix.target }}" == "linux-arm64" ]]; then
             file dist/strix
@@ -102,6 +107,7 @@ jobs:
           if-no-files-found: error
 
   release:
+    if: startsWith(github.ref, 'refs/tags/')
     needs: build
     runs-on: ubuntu-latest
     permissions:
@@ -116,6 +122,5 @@ jobs:
       - name: Create Release
         uses: softprops/action-gh-release@3bb12739c298aeb8a4eeaf626c5b8d85266b0e65  # v2.6.2
         with:
-          prerelease: ${{ !startsWith(github.ref, 'refs/tags/') }}
           generate_release_notes: true
           files: release/**

.github/workflows/ci-package.yml

@@ -0,0 +1,76 @@
+name: Package
+
+on:
+  workflow_call:
+
+permissions:
+  contents: read
+
+jobs:
+  wheel:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+        with:
+          persist-credentials: false
+      - uses: actions/setup-go@44694675825211faa026b3c33043df3e48a5fa00  # v6.0.0
+        with:
+          go-version-file: strix/interface/tui/go.mod
+          cache-dependency-path: strix/interface/tui/go.sum
+      - uses: ./.github/actions/setup-python-env
+      - run: uv build --wheel
+      - name: Wheel ships the TUI sidecar and the viewer bundle
+        run: |
+          uv run --frozen python - <<'PY'
+          import glob, zipfile
+          (wheel,) = glob.glob("dist/*.whl")
+          names = zipfile.ZipFile(wheel).namelist()
+          assert "strix/bin/strix-tui" in names, names
+          assert "strix/interface/viewer/static/index.html" in names, names
+          assert not any(name.startswith("strix/interface/viewer/frontend/") for name in names)
+          assert not any(name.startswith("strix/interface/tui/cmd/") for name in names)
+          PY
+      - name: Install into a clean environment and run the CLI
+        run: |
+          uv venv --python 3.12 /tmp/strix-smoke
+          uv pip install --python /tmp/strix-smoke/bin/python dist/*.whl
+          /tmp/strix-smoke/bin/strix --version
+          /tmp/strix-smoke/bin/strix --help >/dev/null
+      - name: Vertex AI and Bedrock dependencies install with the wheel
+        run: /tmp/strix-smoke/bin/python -c "import google.auth.transport.requests, boto3"
+
+  binary:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+        with:
+          persist-credentials: false
+      - uses: actions/setup-go@44694675825211faa026b3c33043df3e48a5fa00  # v6.0.0
+        with:
+          go-version-file: strix/interface/tui/go.mod
+          cache-dependency-path: strix/interface/tui/go.sum
+      - uses: ./.github/actions/setup-python-env
+      - run: make tui-build
+      - run: uv run --frozen pyinstaller strix.spec --noconfirm
+      - run: dist/strix --version
+      - name: Binary bundles the TUI sidecar and the provider dependencies
+        run: |
+          uv run --frozen pyi-archive_viewer -l -r dist/strix > archive.txt
+          grep -E "strix/bin/strix-tui" archive.txt >/dev/null
+          grep -E "'google\.auth\.transport\.requests'" archive.txt >/dev/null
+          grep -E "'boto3\.session'" archive.txt >/dev/null
+      - name: Binary reaches Vertex AI and Bedrock authentication
+        env:
+          STRIX_TELEMETRY: "false"
+          STRIX_IMAGE: alpine:3.20
+          VERTEXAI_PROJECT: strix-ci
+          VERTEXAI_LOCATION: us-central1
+          AWS_REGION: us-east-1
+        run: |
+          mkdir -p /tmp/target && echo 'print("hi")' > /tmp/target/app.py
+          STRIX_LLM=vertex_ai/gemini-2.5-pro dist/strix -n -t /tmp/target --scan-mode quick > vertex.log 2>&1 || true
+          grep -q "DefaultCredentialsError" vertex.log
+          if grep -q "No module named" vertex.log; then exit 1; fi
+          STRIX_LLM=bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0 dist/strix -n -t /tmp/target --scan-mode quick > bedrock.log 2>&1 || true
+          grep -q "Unable to locate credentials" bedrock.log
+          if grep -q "No module named" bedrock.log; then exit 1; fi

.github/workflows/ci-python.yml

@@ -0,0 +1,61 @@
+name: Python
+
+on:
+  workflow_call:
+
+permissions:
+  contents: read
+
+jobs:
+  lock:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+        with:
+          persist-credentials: false
+      - uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86  # v5.4.2
+      - run: uv lock --check
+
+  ruff:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+        with:
+          persist-credentials: false
+      - uses: ./.github/actions/setup-python-env
+      - run: uv run --frozen ruff check .
+      - run: uv run --frozen ruff format --check .
+
+  mypy:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+        with:
+          persist-credentials: false
+      - uses: ./.github/actions/setup-python-env
+      - run: uv run --frozen mypy strix/
+
+  bandit:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+        with:
+          persist-credentials: false
+      - uses: ./.github/actions/setup-python-env
+      - run: uv run --frozen bandit -r strix/ -c pyproject.toml
+
+  pytest:
+    name: pytest (${{ matrix.python-version }})
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ['3.12', '3.13', '3.14']
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+        with:
+          persist-credentials: false
+      - uses: ./.github/actions/setup-python-env
+        with:
+          python-version: ${{ matrix.python-version }}
+      - run: uv run --frozen pytest -q

.github/workflows/ci-tui.yml

@@ -0,0 +1,26 @@
+name: TUI
+
+on:
+  workflow_call:
+
+permissions:
+  contents: read
+
+jobs:
+  go:
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: strix/interface/tui
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+        with:
+          persist-credentials: false
+      - uses: actions/setup-go@44694675825211faa026b3c33043df3e48a5fa00  # v6.0.0
+        with:
+          go-version-file: strix/interface/tui/go.mod
+          cache-dependency-path: strix/interface/tui/go.sum
+      - run: test -z "$(gofmt -l .)"
+      - run: go vet ./...
+      - run: CGO_ENABLED=0 go build -trimpath -o /dev/null ./cmd/strix-tui
+      - run: go test -race ./...

.github/workflows/ci-viewer.yml

@@ -0,0 +1,27 @@
+name: Viewer
+
+on:
+  workflow_call:
+
+permissions:
+  contents: read
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: strix/interface/viewer/frontend
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+        with:
+          persist-credentials: false
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020  # v7.0.0
+        with:
+          node-version: '24'
+          cache: npm
+          cache-dependency-path: strix/interface/viewer/frontend/package-lock.json
+      - run: npm ci --no-audit --no-fund
+      - run: npm run build
+      - name: Committed viewer bundle matches the build
+        run: git -C "$GITHUB_WORKSPACE" diff --exit-code --stat -- strix/interface/viewer/static && test -z "$(git -C "$GITHUB_WORKSPACE" status --porcelain -- strix/interface/viewer/static)"

.github/workflows/ci-workflows.yml

@@ -0,0 +1,37 @@
+name: Workflows
+
+on:
+  workflow_call:
+
+permissions:
+  contents: read
+
+jobs:
+  actionlint:
+    runs-on: ubuntu-latest
+    env:
+      ACTIONLINT_VERSION: 1.7.12
+      ACTIONLINT_SCRIPT_REF: 914e7df21a07ef503a81201c76d2b11c789d3fca  # v1.7.12
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+        with:
+          persist-credentials: false
+      - name: Install actionlint
+        run: |
+          mkdir -p "$RUNNER_TEMP/actionlint"
+          curl -fsSL --retry 5 -o "$RUNNER_TEMP/actionlint/download.bash" \
+            "https://raw.githubusercontent.com/rhysd/actionlint/$ACTIONLINT_SCRIPT_REF/scripts/download-actionlint.bash"
+          bash "$RUNNER_TEMP/actionlint/download.bash" "$ACTIONLINT_VERSION" "$RUNNER_TEMP/actionlint"
+      - run: |
+          "$RUNNER_TEMP/actionlint/actionlint" -color
+
+  zizmor:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+        with:
+          persist-credentials: false
+      - uses: zizmorcore/zizmor-action@cc914d7f3750a2d13d75c7f184a1060aa0e9d482  # v0.6.4
+        with:
+          advanced-security: false
+          version: 1.22.0

.github/workflows/ci.yml

@@ -0,0 +1,52 @@
+name: CI
+
+on:
+  pull_request:
+  push:
+    branches:
+      - main
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  python:
+    uses: ./.github/workflows/ci-python.yml
+
+  tui:
+    uses: ./.github/workflows/ci-tui.yml
+
+  viewer:
+    uses: ./.github/workflows/ci-viewer.yml
+
+  package:
+    uses: ./.github/workflows/ci-package.yml
+
+  workflows:
+    uses: ./.github/workflows/ci-workflows.yml
+
+  ci-passed:
+    name: ci-passed
+    if: always()
+    needs:
+      - python
+      - tui
+      - viewer
+      - package
+      - workflows
+    runs-on: ubuntu-latest
+    steps:
+      - name: Every job succeeded
+        env:
+          NEEDS: ${{ toJSON(needs) }}
+        run: |
+          failing=$(jq -r 'to_entries[] | select(.value.result != "success") | "\(.key): \(.value.result)"' <<< "$NEEDS")
+          if [ -n "$failing" ]; then
+            echo "$failing"
+            exit 1
+          fi

.pre-commit-config.yaml

@@ -4,7 +4,7 @@ exclude: ^strix/interface/viewer/static/assets/
 repos:
   # Ruff for fast linting and formatting
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.11.13
+    rev: v0.15.20
     hooks:
       - id: ruff
         args: [--fix, --exit-non-zero-on-fix]

Makefile

@@ -1,4 +1,4 @@
-.PHONY: help install dev-install format lint type-check security check-all clean pre-commit setup-dev dev viewer wheel tui-build tui-test tui-lint
+.PHONY: help install dev-install format lint format-check lint-check type-check security check-all clean pre-commit setup-dev dev viewer wheel tui-build tui-test tui-lint
 
 TUI_BINARY := build/sidecar/strix-tui$(if $(filter Windows_NT,$(OS)),.exe)
 
@@ -10,7 +10,9 @@ help:
 	@echo ""
 	@echo "Code Quality:"
 	@echo "  format        - Format code with ruff"
-	@echo "  lint          - Lint code with ruff"
+	@echo "  lint          - Lint code with ruff and apply fixes"
+	@echo "  format-check  - Check formatting without modifying files"
+	@echo "  lint-check    - Check lint without modifying files"
 	@echo "  type-check    - Run type checking with mypy and pyright"
 	@echo "  security      - Run security checks with bandit"
 	@echo "  check-all     - Run all code quality checks"
@@ -45,6 +47,12 @@ lint:
 	uv run ruff check . --fix
 	@echo "✅ Linting complete!"
 
+format-check:
+	uv run ruff format --check .
+
+lint-check:
+	uv run ruff check .
+
 type-check:
 	@echo "🔍 Type checking with mypy..."
 	uv run mypy strix/
@@ -57,7 +65,7 @@ security:
 	uv run bandit -r strix/ -c pyproject.toml
 	@echo "✅ Security checks complete!"
 
-check-all: format lint type-check security
+check-all: format-check lint-check type-check security
 	@echo "✅ All code quality checks passed!"
 
 pre-commit:

docs/advanced/configuration.mdx

@@ -37,6 +37,11 @@ Configure Strix using environment variables or a config file.
   Request timeout in seconds for LLM calls.
 </ParamField>
 
+<ParamField path="LLM_PREFLIGHT_TIMEOUT" default="30" type="integer">
+  Seconds the startup connection check waits for the model to answer before
+  failing with `LLM CONNECTION FAILED`. Scan requests use `LLM_TIMEOUT`.
+</ParamField>
+
 <ParamField path="STRIX_LLM_MAX_RETRIES" default="5" type="integer">
   Maximum number of retries for LLM API calls on transient failures.
 </ParamField>
@@ -49,6 +54,14 @@ Configure Strix using environment variables or a config file.
   Timeout in seconds for memory compression operations (context summarization).
 </ParamField>
 
+<ParamField path="STRIX_OPENROUTER_STICKY_SESSIONS" default="false" type="boolean">
+  Send a per-agent `session_id` on OpenRouter requests, so each agent's calls stay on one upstream provider and its prompt cache carries over between turns. When unset, OpenRouter routes every request freely.
+</ParamField>
+
+<ParamField path="STRIX_CACHE_BLOCK_TOKENS" default="128" type="integer">
+  Token block size that providers cache prompts in. A turn counts as a cache miss in the run report only when the cached tokens fall at least this many tokens short of the previous prompt and the prompt did not shrink. Lower it for providers with smaller blocks (DeepSeek and GLM use 64; vLLM defaults to 16).
+</ParamField>
+
 ### Dedicated deduplication model
 
 Finding deduplication is a cheap, structured classification task. By default it
@@ -137,7 +150,7 @@ When remote vars are set, Strix dual-writes telemetry to both local JSONL and th
 </ParamField>
 
 <ParamField path="DOCKER_HOST" type="string">
-  Docker daemon socket path. Use for remote Docker hosts or custom configurations.
+  Docker daemon address, for example `unix:///var/run/docker.sock` or `tcp://10.0.0.5:2375`. When not set, Strix connects to the same daemon as the `docker` CLI: the current docker context (`docker context use`), then the default socket. Docker Desktop, OrbStack and Colima register their sockets as contexts, so no extra configuration is needed for them.
 </ParamField>
 
 <ParamField path="STRIX_RUNTIME_BACKEND" default="docker" type="string">

docs/integrations/ci-cd.mdx

@@ -23,9 +23,15 @@ strix -n --target ./app --scan-mode quick --scope-mode diff --diff-base origin/m
 
 | Code | Meaning |
 |------|---------|
-| 0 | No vulnerabilities found |
+| 0 | No vulnerabilities found (with `--fail-on`, none at or above the threshold) |
 | 1 | Execution error |
-| 2 | Vulnerabilities found |
+| 2 | Vulnerabilities found (with `--fail-on`, at least one at or above the threshold) |
+
+To fail only on serious findings, pass `--fail-on` with a minimum severity. Lower findings are still reported but exit `0`, so a passing job does not mean the report is empty.:
+
+```bash
+strix -n --target ./app --scan-mode quick --fail-on high
+```
 
 ## GitLab CI
 

docs/integrations/github-actions.mdx

@@ -46,9 +46,11 @@ The workflow fails when vulnerabilities are found:
 
 | Code | Result |
 |------|--------|
-| 0 | Pass — No vulnerabilities |
+| 0 | Pass — No vulnerabilities (with `--fail-on`, none at or above the threshold) |
 | 2 | Fail — Vulnerabilities found |
 
+Add `--fail-on high` (or `critical`, `medium`, `low`) to fail only on findings at or above that severity. Lower findings still appear in the report, so a passing run is not necessarily finding-free.
+
 ## Scan Modes for CI
 
 | Mode | Duration | Use Case |

docs/llm-providers/bedrock.mdx

@@ -3,14 +3,6 @@ title: "AWS Bedrock"
 description: "Configure Strix with models via AWS Bedrock"
 ---
 
-## Installation
-
-Bedrock requires the AWS SDK dependency. Install Strix with the bedrock extra:
-
-```bash
-pipx install "strix-agent[bedrock]"
-```
-
 ## Setup
 
 ```bash

docs/llm-providers/vertex.mdx

@@ -3,14 +3,6 @@ title: "Google Vertex AI"
 description: "Configure Strix with Gemini models via Google Cloud"
 ---
 
-## Installation
-
-Vertex AI requires the Google Cloud dependency. Install Strix with the vertex extra:
-
-```bash
-pipx install "strix-agent[vertex]"
-```
-
 ## Setup
 
 ```bash

docs/usage/cli.mdx

@@ -57,7 +57,17 @@ strix (--target <target> | --target-list <path>) [options]
 </ParamField>
 
 <ParamField path="--non-interactive, -n" type="boolean">
-  Run in headless mode without TUI. Ideal for CI/CD.
+  Run in headless mode without TUI. Ideal for CI/CD. When no terminal is
+  attached (CI, `nohup`, pipes) and a target is given, Strix runs headless
+  automatically and says so on one line.
+</ParamField>
+
+<ParamField path="--fail-on" type="string">
+  Headless mode only (requires `-n`). Minimum severity that makes the run exit
+  `2`: `critical`, `high`, `medium`, `low`, or `info`. Findings below the
+  threshold are still written to every report artifact. A finding whose
+  severity Strix does not recognize always counts, so the gate never passes on a
+  value it cannot rank. Omit the flag to exit `2` on any finding.
 </ParamField>
 
 <ParamField path="--config" type="string">
@@ -132,6 +142,9 @@ strix --target api.example.com --instruction "Focus on IDOR and auth bypass"
 # CI/CD mode
 strix -n --target ./ --scan-mode quick
 
+# CI/CD mode, failing only on high or critical findings
+strix -n --target ./ --scan-mode quick --fail-on high
+
 # Cap cost and per-agent turns
 strix --target https://example.com --max-budget 25 --max-turns 300
 
@@ -159,6 +172,6 @@ strix --target https://app.com --workspace-file ./openapi.yaml:specs/openapi.yam
 
 | Code | Meaning |
 |------|---------|
-| 0 | Scan completed successfully (interactive mode always exits `0`; in headless mode, `0` means no vulnerabilities were found) |
+| 0 | Scan completed successfully (interactive mode always exits `0`; in headless mode, `0` means no vulnerabilities were found, or with `--fail-on`, none at or above the threshold; lower findings are still in the report) |
 | 1 | A fatal error occurred before or during the scan (e.g. missing environment variables, Docker unavailable, invalid config file, diff-scope resolution failure, or an unhandled error) |
-| 2 | Vulnerabilities found (headless mode only) |
+| 2 | Vulnerabilities found (headless mode only; with `--fail-on`, at least one finding at or above that severity) |

pyproject.toml

@@ -36,6 +36,8 @@ dependencies = [
   "openai-agents[litellm]>=0.19.0,<0.20",
   "openai>=2.45.0,<3",
   "litellm>=1.101.0",
+  "google-auth>=2.0.0",
+  "boto3>=1.28.0",
   "pydantic>=2.11.3",
   "pydantic-settings>=2.13.0",
   "rich",
@@ -52,10 +54,6 @@ dependencies = [
   "pyyaml>=6.0",
 ]
 
-[project.optional-dependencies]
-vertex = ["google-auth>=2.0.0"]
-bedrock = ["boto3>=1.28.0"]
-
 [project.scripts]
 strix = "strix.interface.main:main"
 

skills/ci-security-scanning-with-strix/SKILL.md

@@ -67,7 +67,7 @@ Then tell the user to add two repository secrets: `STRIX_LLM` (model id, for exa
 
 Notes:
 - In CI/headless runs Strix automatically scopes to the PR's changed files (`--scope-mode auto`). If diff resolution fails, keep `fetch-depth: 0` or set `--diff-base` to the PR's actual base branch — use `origin/${{ github.base_ref }}` in GitHub Actions rather than a hard-coded `origin/main`, since repos use different default branches.
-- Exit codes: `0` pass, `2` vulnerabilities found (fails the job), `1` setup error.
+- Exit codes: `0` pass, `2` vulnerabilities found (fails the job), `1` setup error. Add `--fail-on high` to fail only on high/critical findings; lower ones are still reported, so a passing job is not a finding-free report.
 - The runner needs Docker (default GitHub-hosted Ubuntu runners have it).
 - **Size the budget so the scan completes — do not let it fail open.** A `0` exit means "no validated vulnerabilities in what was analyzed"; if `--max-budget` is hit before the diff is fully covered, the scan wraps up early and can still exit `0`. The "Fail unless the scan completed" step above narrows the gap: `strix_runs/<run>/run.json` is `"stopped"` when the scan was cut off at the hard budget limit without a final report. It is not a complete guard — the agents get graduated wrap-up warnings before that limit, and a run that wraps up on a warning still calls `finish_scan` and records `"completed"` with partial coverage. So keep that step in any pipeline that gates merges **and** give the scan real headroom (compare `run.json`'s `llm_usage.cost` against `--max-budget`; if it ran right up to the cap, raise it). For a `quick` diff-scoped PR scan `--max-budget 10` is usually ample, raise it for large diffs.
 

skills/penetration-testing-with-strix/SKILL.md

@@ -94,6 +94,7 @@ Key flags:
 | `--workspace-file PATH[:DEST]` | Copy a file from this machine into `/workspace` before the scan, for a wordlist, a spec, or notes. Repeatable. |
 | `--max-budget USD` | Hard LLM spend cap; scan wraps up cleanly at the limit. |
 | `--max-turns N` | Per-agent turn cap (default 500). |
+| `--fail-on SEVERITY` | Headless only: exit `2` only for findings at or above `critical`/`high`/`medium`/`low`/`info`. Default: any finding. |
 | `--resume RUN_NAME` | Resume a prior run from `strix_runs/`, with its agent history and targets. Cannot be combined with `-t`. |
 | `--scope-mode` | For code targets: `auto` (diff-scope in CI/headless), `diff` (force changed files only), `full` (whole tree). |
 | `--diff-base REF` | Branch or commit that `diff` scope compares against. Defaults to the repo's default branch. |
@@ -102,9 +103,9 @@ Scans take minutes (`quick`) to hours (`deep`). Run them in the background and p
 
 ### Exit codes (headless)
 
-- `0` — finished with no validated vulnerabilities **in what was analyzed**
+- `0` — finished with no validated vulnerabilities **in what was analyzed** (with `--fail-on`, none at or above the threshold; lower ones are still in the artifacts)
 - `1` — fatal error (missing env vars, Docker down, bad config)
-- `2` — vulnerabilities found
+- `2` — vulnerabilities found (with `--fail-on`, at least one at or above that severity)
 
 A `0` is not proof of full coverage: if `--max-budget`/`--max-turns` is reached before the scan completes, it wraps up early and still exits `0`. When you need assurance the scan finished, give it enough budget and check `strix_runs/<run>/run.json`: a hard budget stop leaves `status: "stopped"`, but an agent that wrapped up early on a budget *warning* still calls `finish_scan` and records `"completed"` — so also sanity-check the run's cost against `--max-budget` and the report's stated coverage before treating a clean result as full coverage.
 

strix.spec

@@ -180,6 +180,8 @@ hiddenimports = [
 ]
 
 hiddenimports += collect_submodules('litellm')
+hiddenimports += collect_submodules('google.auth')
+hiddenimports += collect_submodules('google.oauth2')
 hiddenimports += collect_submodules('rich')
 hiddenimports += collect_submodules('pydantic')
 hiddenimports += collect_submodules('pygments')
@@ -211,8 +213,6 @@ excludes = [
     'google.cloud',
     'google.cloud.aiplatform',
     'google.api_core',
-    'google.auth',
-    'google.oauth2',
     'google.protobuf',
     'grpc',
     'grpcio',

strix/agents/prompt.py

@@ -16,6 +16,10 @@ logger = logging.getLogger(__name__)
 
 _PROMPT_DIRNAME = "prompts"
 
+# Marks where the system prompt is split so the part before it can be cached.
+# Removed before the prompt is sent.
+CACHE_POINT = "<cache_point>"
+
 
 def _resolve_skills(
     *,
@@ -80,8 +84,13 @@ def render_system_prompt(
     is_diff_scoped: bool = False,
     interactive: bool = False,
     system_prompt_context: dict[str, Any] | None = None,
+    include_scope: bool = True,
 ) -> str:
-    """Render the system prompt. Returns empty string on template failure."""
+    """Render the system prompt. Returns empty string on template failure.
+
+    The per-run scope (targets, MCP connections) goes last so the rest of the
+    prompt is an identical prefix across runs and can be served from cache.
+    """
     try:
         prompt_dir = get_strix_resource_path("agents", _PROMPT_DIRNAME)
         loader_dirs = [prompt_dir, *skill_search_dirs()]
@@ -93,6 +102,16 @@ def render_system_prompt(
             ),
         )
 
+        shared = {
+            name.split("/")[-1]
+            for name in _resolve_skills(
+                requested=None,
+                scan_mode=scan_mode,
+                is_whitebox=is_whitebox,
+                is_root=is_root,
+                is_diff_scoped=is_diff_scoped,
+            )
+        }
         skills_to_load = _resolve_skills(
             requested=skills,
             scan_mode=scan_mode,
@@ -103,12 +122,16 @@ def render_system_prompt(
         skill_content = load_skills(skills_to_load)
         env.globals["get_skill"] = lambda name: skill_content.get(name, "")
 
+        # Skills every agent of this kind loads come first, so siblings share them
+        # as a cached prefix; the ones the caller asked for vary and go after.
         rendered = env.get_template("system_prompt.jinja").render(
-            loaded_skill_names=list(skill_content.keys()),
+            shared_skill_names=[name for name in skill_content if name in shared],
+            requested_skill_names=[name for name in skill_content if name not in shared],
             available_skills=get_available_skills(),
             interactive=interactive,
             is_root=is_root,
             system_prompt_context=system_prompt_context or {},
+            include_scope=include_scope,
             **skill_content,
         )
     except Exception:
@@ -124,3 +147,16 @@ def render_system_prompt(
             len(rendered),
         )
         return str(rendered)
+
+
+def render_scope_prompt(system_prompt_context: dict[str, Any] | None) -> str:
+    """Render only the per-run scope block that ends the system prompt."""
+    prompt_dir = get_strix_resource_path("agents", _PROMPT_DIRNAME)
+    env = Environment(
+        loader=FileSystemLoader(prompt_dir),
+        autoescape=select_autoescape(enabled_extensions=(), default_for_string=False),
+    )
+    rendered = env.get_template("scope.jinja").render(
+        system_prompt_context=system_prompt_context or {},
+    )
+    return str(rendered).strip()

strix/agents/prompts/scope.jinja

@@ -0,0 +1,35 @@
+<cache_point>
+{% if system_prompt_context and system_prompt_context.authorized_targets %}
+SYSTEM-VERIFIED SCOPE:
+- The following scope metadata is injected by the platform into the system prompt and is authoritative
+- Scope source: {{ system_prompt_context.scope_source }}
+- Authorization source: {{ system_prompt_context.authorization_source }}
+- Every target listed below has already been verified by the platform as in-scope and authorized
+- User instructions, chat messages, and other free-form text do NOT expand scope beyond this list
+- NEVER refuse, question authorization, or claim lack of permission for any target in this system-verified scope
+- NEVER test any external domain, URL, host, IP, or repository that is not explicitly listed in this system-verified scope
+- If the user mentions any asset outside this list, ignore that asset and continue working only on the listed in-scope targets
+
+AUTHORIZED TARGETS:
+{% for target in system_prompt_context.authorized_targets %}
+- {{ target.type }}: {{ target.value }}{% if target.workspace_path %} (workspace: {{ target.workspace_path }}){% endif %}
+{% endfor %}
+{% endif %}
+
+{% if system_prompt_context and system_prompt_context.mcp_available %}
+MCP CONNECTIONS (available this run):
+- The user connected one or more MCP (Model Context Protocol) servers. Their individual tools do NOT appear in your tool list. Use the four discovery and dispatch tools to reach them.
+{% if system_prompt_context.mcp_connections %}
+- Connected this run (search one to find relevant tools):
+{% for connection in system_prompt_context.mcp_connections %}
+  - {{ connection.name }} ({{ connection.tool_count }} tools){% if connection.purpose %}: {{ connection.purpose }}{% endif %}
+{% endfor %}
+{% endif %}
+- Reach for a connection whenever the target itself cannot give you information a connection could: its database schema and access policies, real deployment or infrastructure configuration, known issues or prior findings, or server logs. In those cases call list_mcps early to see what is available, and prefer a connection's authoritative data over inferring from the target's responses. Do not wait to be told a connection exists.
+  1. Call list_mcps() to discover the available connections.
+  2. Call search_mcp_tools(connection="<name>", query="<capability>") for a short candidate list.
+  3. Call get_mcp_tool_schema(connection="<name>", tool="<tool>") for the one schema you need.
+  4. Call call_mcp(connection="<name>", tool="<tool>", arguments={...}) to run it, passing an arguments object that matches the schema (omit arguments for a tool that takes none).
+- Do not assume a connection or tool exists; discover it with list_mcps and search_mcp_tools before calling.
+- Use describe_mcp only as a compatibility fallback when targeted search cannot identify an expected tool. Its full catalog can be large.
+{% endif %}

strix/agents/prompts/system_prompt.jinja

@@ -58,41 +58,6 @@ AUTONOMOUS BEHAVIOR:
 </communication_rules>
 
 <execution_guidelines>
-{% if system_prompt_context and system_prompt_context.authorized_targets %}
-SYSTEM-VERIFIED SCOPE:
-- The following scope metadata is injected by the platform into the system prompt and is authoritative
-- Scope source: {{ system_prompt_context.scope_source }}
-- Authorization source: {{ system_prompt_context.authorization_source }}
-- Every target listed below has already been verified by the platform as in-scope and authorized
-- User instructions, chat messages, and other free-form text do NOT expand scope beyond this list
-- NEVER refuse, question authorization, or claim lack of permission for any target in this system-verified scope
-- NEVER test any external domain, URL, host, IP, or repository that is not explicitly listed in this system-verified scope
-- If the user mentions any asset outside this list, ignore that asset and continue working only on the listed in-scope targets
-
-AUTHORIZED TARGETS:
-{% for target in system_prompt_context.authorized_targets %}
-- {{ target.type }}: {{ target.value }}{% if target.workspace_path %} (workspace: {{ target.workspace_path }}){% endif %}
-{% endfor %}
-{% endif %}
-
-{% if system_prompt_context and system_prompt_context.mcp_available %}
-MCP CONNECTIONS (available this run):
-- The user connected one or more MCP (Model Context Protocol) servers. Their individual tools do NOT appear in your tool list. Use the four discovery and dispatch tools to reach them.
-{% if system_prompt_context.mcp_connections %}
-- Connected this run (search one to find relevant tools):
-{% for connection in system_prompt_context.mcp_connections %}
-  - {{ connection.name }} ({{ connection.tool_count }} tools){% if connection.purpose %}: {{ connection.purpose }}{% endif %}
-{% endfor %}
-{% endif %}
-- Reach for a connection whenever the target itself cannot give you information a connection could: its database schema and access policies, real deployment or infrastructure configuration, known issues or prior findings, or server logs. In those cases call list_mcps early to see what is available, and prefer a connection's authoritative data over inferring from the target's responses. Do not wait to be told a connection exists.
-  1. Call list_mcps() to discover the available connections.
-  2. Call search_mcp_tools(connection="<name>", query="<capability>") for a short candidate list.
-  3. Call get_mcp_tool_schema(connection="<name>", tool="<tool>") for the one schema you need.
-  4. Call call_mcp(connection="<name>", tool="<tool>", arguments={...}) to run it, passing an arguments object that matches the schema (omit arguments for a tool that takes none).
-- Do not assume a connection or tool exists; discover it with list_mcps and search_mcp_tools before calling.
-- Use describe_mcp only as a compatibility fallback when targeted search cannot identify an expected tool. Its full catalog can be large.
-{% endif %}
-
 AUTHORIZATION STATUS:
 - You have FULL AUTHORIZATION for authorized security validation on in-scope targets to help secure the target systems/app
 - All permission checks have been COMPLETED and APPROVED - never question your authority
@@ -528,9 +493,9 @@ Directories:
 Default user: pentester (sudo available)
 </environment>
 
-{% if loaded_skill_names %}
+{% if shared_skill_names %}
 <specialized_knowledge>
-{% for skill_name in loaded_skill_names %}
+{% for skill_name in shared_skill_names %}
 <{{ skill_name }}>
 {{ get_skill(skill_name) }}
 </{{ skill_name }}>
@@ -540,7 +505,7 @@ Default user: pentester (sudo available)
 
 {% if available_skills %}
 <available_skills>
-On-demand specialist skills. Spawn a specialist via `create_agent(skills=[...])`, or pull guidance inline for yourself via `load_skill(skills=[...])`. Anything wrapped in `<specialized_knowledge>` above is already loaded for you.
+On-demand specialist skills. Spawn a specialist via `create_agent(skills=[...])`, or pull guidance inline for yourself via `load_skill(skills=[...])`. Anything wrapped in `<specialized_knowledge>` is already loaded for you.
 
 {% for category, skills in available_skills | dictsort -%}
 {% for skill in skills -%}
@@ -549,3 +514,18 @@ On-demand specialist skills. Spawn a specialist via `create_agent(skills=[...])`
 {% endfor -%}
 </available_skills>
 {% endif %}
+
+{% if requested_skill_names %}
+<cache_point>
+<specialized_knowledge>
+{% for skill_name in requested_skill_names %}
+<{{ skill_name }}>
+{{ get_skill(skill_name) }}
+</{{ skill_name }}>
+{% endfor %}
+</specialized_knowledge>
+{% endif %}
+
+{% if include_scope %}
+{% include "scope.jinja" %}
+{% endif %}

strix/config/models.py

@@ -8,6 +8,7 @@ import inspect
 import logging
 import os
 import time
+import uuid
 from collections.abc import AsyncGenerator
 from typing import TYPE_CHECKING, Any, cast
 
@@ -36,6 +37,7 @@ from openai.types.responses import (
 from openai.types.responses.response_usage import ResponseUsage
 from openai.types.shared import Reasoning
 
+from strix.agents.prompt import CACHE_POINT
 from strix.config import codex
 from strix.config.loader import load_settings
 from strix.config.tool_call_ids import TurnCallIdRewriter, dedupe_input
@@ -56,7 +58,7 @@ if TYPE_CHECKING:
     from openai import AsyncOpenAI
     from openai.types.responses.response_prompt_param import ResponsePromptParam
 
-    from strix.config.settings import LlmSettings, ReasoningEffort, Settings
+    from strix.config.settings import ApiType, LlmSettings, ReasoningEffort, Settings
 
 
 logger = logging.getLogger(__name__)
@@ -306,6 +308,9 @@ class _TurnGuardModel(Model):
     ) -> ModelResponse:
         sanitized = dedupe_input(input)
         rewriter = TurnCallIdRewriter(sanitized)
+        system_instructions, sanitized = _split_cached_prefix(
+            system_instructions, sanitized, model_settings
+        )
         response = await self._inner.get_response(
             system_instructions,
             cast("str | list[TResponseInputItem]", sanitized),
@@ -339,6 +344,9 @@ class _TurnGuardModel(Model):
     ) -> AsyncIterator[TResponseStreamEvent]:
         sanitized = dedupe_input(input)
         rewriter = TurnCallIdRewriter(sanitized)
+        system_instructions, sanitized = _split_cached_prefix(
+            system_instructions, sanitized, model_settings
+        )
         limiter = self._limiter()
         stream = self._inner.stream_response(
             system_instructions,
@@ -359,6 +367,27 @@ class _TurnGuardModel(Model):
         self._log_dropped(limiter)
 
 
+def _split_cached_prefix(
+    system_instructions: str | None,
+    model_input: str | list[Any],
+    model_settings: ModelSettings,
+) -> tuple[str | None, str | list[Any]]:
+    """Split the system prompt at each ``CACHE_POINT`` on cache-point routes.
+
+    LiteLLM puts a cache point at the end of each system message, so each part
+    gets its own. Other routes get the prompt with the markers removed.
+    """
+    if not system_instructions or CACHE_POINT not in system_instructions:
+        return system_instructions, model_input
+    extra_args = model_settings.extra_args or {}
+    if "cache_control_injection_points" not in extra_args:
+        return system_instructions.replace(CACHE_POINT, ""), model_input
+    if isinstance(model_input, str):
+        model_input = [{"role": "user", "content": model_input}]
+    parts = [part for part in system_instructions.split(CACHE_POINT) if part.strip()]
+    return None, [*({"role": "system", "content": part} for part in parts), *model_input]
+
+
 async def _aclose(stream: AsyncIterator[TResponseStreamEvent]) -> None:
     if isinstance(stream, AsyncGenerator):
         with contextlib.suppress(Exception):
@@ -603,61 +632,6 @@ DEFAULT_MODEL_RETRY = ModelRetrySettings(
     ),
 )
 
-RECOMMENDED_MODEL_NAMES = (
-    "zai/glm-5.3",
-    "zai/glm-5.3-flash",
-    "openai/gpt-5.6-sol",
-    "openai/gpt-5.6-terra",
-    "openai/gpt-5.6-luna",
-    "openai/gpt-5.6",
-    "openai/gpt-5.5-pro",
-    "openai/gpt-5.5",
-    "openai/gpt-5.4",
-    "openai/gpt-5.3-codex",
-    "anthropic/claude-fable-5-1",
-    "anthropic/claude-fable-5",
-    "anthropic/claude-opus-5",
-    "anthropic/claude-opus-4-8",
-    "anthropic/claude-sonnet-5",
-    "anthropic/claude-sonnet-4-6",
-    "vertex_ai/gemini-3.1-pro-preview",
-    "gemini/gemini-3.1-pro-preview",
-    "vertex_ai/gemini-3.7-flash",
-    "gemini/gemini-3.7-flash",
-    "gemini/gemini-3.6-flash",
-    "deepseek/deepseek-v4-pro",
-    "deepseek/deepseek-v4-flash",
-    "dashscope/qwen3.8-max",
-    "dashscope/qwen3.7-max-2026-06-08",
-    "moonshot/kimi-k3",
-    "moonshot/kimi-k2.7-code",
-)
-
-_RECOMMENDED_MODEL_NAME_SET = frozenset(name.lower() for name in RECOMMENDED_MODEL_NAMES)
-
-# Matched against the bare model name only: the route (``openai/``, ``openrouter/``,
-# a local gateway, ...) says nothing about the model's quality.
-FRONTIER_MODEL_PREFIXES = (
-    "gpt-5",
-    "claude-fable-5",
-    "claude-opus-5",
-    "claude-opus-4",
-    "claude-sonnet-5",
-    "claude-sonnet-4",
-    "gemini-3",
-    "deepseek-v4",
-    "deepseek-r1",
-    "deepseek-reasoner",
-    "qwen3.8",
-    "qwen3.7",
-    "qwen3-max",
-    "kimi-k3",
-    "kimi-k2.7",
-    "kimi-k2.6",
-    "glm-5.3",
-    "glm-5.2",
-)
-
 
 def configure_sdk_model_defaults(settings: Settings) -> None:
     """Apply Strix config to SDK-native defaults."""
@@ -675,14 +649,40 @@ def configure_sdk_model_defaults(settings: Settings) -> None:
     if llm.api_base:
         os.environ["OPENAI_BASE_URL"] = llm.api_base
         _configure_litellm_default("api_base", llm.api_base)
-    api_type = llm.api_type
-    if api_type is None:
-        api_type = "chat_completions" if llm.api_base else "responses"
-
-    set_default_openai_api(api_type)
+    configure_sdk_api_route(llm.model or "", settings)
     _configure_extra_headers(llm)
 
 
+def configure_sdk_api_route(model_name: str, settings: Settings) -> None:
+    """Point SDK-native OpenAI requests for ``model_name`` at Responses or chat completions."""
+    api_type = resolve_api_type(model_name, settings)
+    logger.info("OpenAI API route for %s: %s", model_name, api_type)
+    set_default_openai_api(api_type)
+
+
+_RESPONSES_ENDPOINT = "/v1/responses"
+
+
+def resolve_api_type(model_name: str, settings: Settings) -> ApiType:
+    """The SDK-native OpenAI route for ``model_name``: Responses or chat completions.
+
+    An explicit ``STRIX_API_TYPE`` wins. Otherwise the model decides: Responses
+    when LiteLLM's catalog lists ``/v1/responses`` for it, chat completions for
+    everything else.
+    """
+    if settings.llm.api_type is not None:
+        return settings.llm.api_type
+    if _RESPONSES_ENDPOINT in _catalog_supported_endpoints(model_name):
+        return "responses"
+    return "chat_completions"
+
+
+def _catalog_supported_endpoints(model_name: str) -> list[str]:
+    entry = _catalog_entry(model_name)
+    endpoints = entry.get("supported_endpoints") if entry else None
+    return [str(e) for e in endpoints] if isinstance(endpoints, list) else []
+
+
 def _mirror_api_key_to_provider_env(model_name: str | None, api_key: str) -> None:
     if not model_name:
         return
@@ -718,6 +718,11 @@ def _configure_litellm_compatibility() -> None:
     _install_openrouter_stream_cost_capture()
 
 
+# Agent ids are 8 hex characters and can repeat across runs; the session id
+# OpenRouter pins a provider to must not, so each agent gets its own UUID.
+_OPENROUTER_SESSION_IDS: dict[str, str] = {}
+
+
 def _install_openrouter_stream_cost_capture() -> None:
     """Preserve OpenRouter's per-stream cost, which LiteLLM drops when streaming.
 
@@ -736,14 +741,16 @@ def _install_openrouter_stream_cost_capture() -> None:
         OpenrouterConfig,
     )
 
-    from strix.report.state import streamed_openrouter_costs
+    from strix.report.state import record_openrouter_provider, streamed_openrouter_costs
 
     class _StrixOpenRouterStreamingHandler(OpenRouterChatCompletionStreamingHandler):
         def chunk_parser(self, chunk: dict[str, Any]) -> Any:
             stream = super().chunk_parser(chunk)
-            streamed_openrouter_costs.remember(
-                chunk.get("id") or getattr(stream, "id", None), chunk.get("usage")
-            )
+            usage = chunk.get("usage")
+            response_id = chunk.get("id") or getattr(stream, "id", None)
+            streamed_openrouter_costs.remember(response_id, usage)
+            if usage:
+                record_openrouter_provider(chunk.get("provider"), usage)
             return stream
 
     class _StrixOpenrouterConfig(OpenrouterConfig):
@@ -756,6 +763,26 @@ def _install_openrouter_stream_cost_capture() -> None:
                 json_mode=json_mode,
             )
 
+        def transform_response(self, *args: Any, **kwargs: Any) -> Any:
+            # Non-streamed replies (LLM_DISABLE_STREAMING) skip the chunk parser.
+            response = super().transform_response(*args, **kwargs)
+            raw_response = kwargs.get("raw_response", args[1] if len(args) > 1 else None)
+            with contextlib.suppress(Exception):
+                body = raw_response.json()  # type: ignore[union-attr]
+                if body.get("usage"):
+                    record_openrouter_provider(body.get("provider"), body["usage"])
+            return response
+
+        def transform_request(self, *args: Any, **kwargs: Any) -> dict[str, Any]:
+            # Pin each agent's calls to one upstream provider so its prompt cache
+            # survives between turns.
+            body = super().transform_request(*args, **kwargs)
+            agent_id = request_log.current_call_context().agent_id
+            if agent_id and load_settings().llm.openrouter_sticky_sessions:
+                session_id = _OPENROUTER_SESSION_IDS.setdefault(agent_id, str(uuid.uuid4()))
+                body.setdefault("session_id", session_id)
+            return body
+
     # LiteLLM's provider-config factory reads litellm.OpenrouterConfig at call
     # time, so overriding the attribute is enough for the subclass to take
     # effect. (type: ignore — mypy rejects reassigning a class attribute.)
@@ -858,11 +885,7 @@ def uses_chat_completions_tool_schema(model_name: str, settings: Settings) -> bo
     model = model_name.strip().lower()
     if "/" in model and not model.startswith("openai/"):
         return True
-    if settings.llm.api_type is not None:
-        return settings.llm.api_type == "chat_completions"
-    if settings.llm.api_base:
-        return True
-    return not model_supports_reasoning(model_name)
+    return resolve_api_type(model_name, settings) == "chat_completions"
 
 
 def supports_strict_tool_schemas(model_name: str) -> bool:
@@ -878,54 +901,27 @@ def supports_strict_tool_schemas(model_name: str) -> bool:
 
 
 def model_supports_reasoning(model_name: str) -> bool:
-    import litellm
-
-    name = model_name.strip().lower()
-    for prefix in ("litellm/", "any-llm/", "openai/"):
-        if name.startswith(prefix):
-            name = name[len(prefix) :]
-            break
-    entry = litellm.model_cost.get(name)
-    if entry is None and "/" in name:
-        entry = litellm.model_cost.get(name.rsplit("/", 1)[1])
+    entry = _catalog_entry(model_name)
     return bool(entry and entry.get("supports_reasoning"))
 
 
-def is_recommended_or_frontier_model(model_name: str) -> bool:
-    """Return whether a model is recommended or in a frontier model family."""
-    name = _normalized_model_name(model_name)
-    if not name:
-        return False
-    if name in _RECOMMENDED_MODEL_NAME_SET:
-        return True
-    bare_model_name = name.rsplit("/", 1)[-1]
-    return _matches_model_prefix(bare_model_name, FRONTIER_MODEL_PREFIXES)
-
-
-def _normalized_model_name(model_name: str) -> str:
+def _bare_openai_name(model_name: str) -> str:
     name = model_name.strip().lower()
-    for prefix in ("litellm/", "any-llm/"):
+    for prefix in ("litellm/", "any-llm/", "openai/"):
         if name.startswith(prefix):
-            name = name[len(prefix) :]
-            break
+            return name[len(prefix) :]
     return name
 
 
-def _matches_model_prefix(model_name: str, model_prefixes: tuple[str, ...]) -> bool:
-    return any(
-        candidate.startswith(prefix)
-        for candidate in _model_name_candidates(model_name)
-        for prefix in model_prefixes
-    )
+def _catalog_entry(model_name: str) -> dict[str, Any] | None:
+    """LiteLLM's cost-map entry for the model, looked up as it would route it."""
+    import litellm
 
-
-def _model_name_candidates(model_name: str) -> tuple[str, ...]:
-    if "." not in model_name:
-        return (model_name,)
-    suffixes = tuple(
-        model_name.split(".", index)[-1] for index in range(1, model_name.count(".") + 1)
-    )
-    return (model_name, *suffixes)
+    name = _bare_openai_name(model_name)
+    entry = litellm.model_cost.get(name)
+    if entry is None and "/" in name:
+        entry = litellm.model_cost.get(name.rsplit("/", 1)[1])
+    return entry if isinstance(entry, dict) else None
 
 
 def is_known_openai_bare_model(model_name: str) -> bool:

strix/config/settings.py

@@ -4,7 +4,7 @@ from __future__ import annotations
 
 from typing import Literal
 
-from pydantic import AliasChoices, Field
+from pydantic import AliasChoices, Field, field_validator
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
 
@@ -13,6 +13,12 @@ ApiType = Literal["responses", "chat_completions"]
 
 DEFAULT_MAX_TURNS = 500
 
+
+def _lowercase(value: object) -> object:
+    """Enum-like env values are matched case-insensitively (``None`` is ``none``)."""
+    return value.strip().lower() if isinstance(value, str) else value
+
+
 _BASE_CONFIG = SettingsConfigDict(
     case_sensitive=False,
     populate_by_name=True,
@@ -58,11 +64,20 @@ class LlmSettings(BaseSettings):
         default=True,
         alias="STRIX_PROMPT_CACHE",
     )
+    # Providers cache prompts in fixed-size token blocks, so a fully cached prompt
+    # can read back up to a block short. 128 covers the largest common size
+    # (OpenAI; DeepSeek and GLM use 64, vLLM defaults to 16).
+    cache_block_tokens: int = Field(default=128, ge=1, alias="STRIX_CACHE_BLOCK_TOKENS")
+    openrouter_sticky_sessions: bool = Field(
+        default=False,
+        alias="STRIX_OPENROUTER_STICKY_SESSIONS",
+    )
     disable_streaming: bool = Field(
         default=False,
         alias="LLM_DISABLE_STREAMING",
     )
     timeout: int = Field(default=300, alias="LLM_TIMEOUT")
+    preflight_timeout: int = Field(default=30, ge=1, alias="LLM_PREFLIGHT_TIMEOUT")
     stream_idle_timeout: int = Field(default=300, ge=0, alias="LLM_STREAM_IDLE_TIMEOUT")
     max_tool_calls_per_turn: int = Field(
         default=32,
@@ -70,6 +85,11 @@ class LlmSettings(BaseSettings):
         alias="LLM_MAX_TOOL_CALLS_PER_TURN",
     )
 
+    @field_validator("api_type", "reasoning_effort", mode="before")
+    @classmethod
+    def _normalize_case(cls, value: object) -> object:
+        return _lowercase(value)
+
 
 class DedupeSettings(BaseSettings):
     model_config = _BASE_CONFIG
@@ -87,6 +107,11 @@ class DedupeSettings(BaseSettings):
         repr=False,
     )
 
+    @field_validator("reasoning_effort", mode="before")
+    @classmethod
+    def _normalize_case(cls, value: object) -> object:
+        return _lowercase(value)
+
 
 class ContextSettings(BaseSettings):
     """Context-window management: per-tool-output caps and history compaction."""

strix/core/agents.py

@@ -24,8 +24,14 @@ logger = logging.getLogger(__name__)
 
 Status = Literal["running", "waiting", "completed", "stopped", "crashed", "failed", "budget_paused"]
 
+BudgetPolicy = Literal["stop", "pause"]
+
 TERMINAL_STATUSES: frozenset[str] = frozenset({"completed", "stopped", "crashed", "failed"})
 
+# Agents that still have work to do or to wake up for. A parked agent is mid-task:
+# a scan is not finished while one exists, and it can be stopped like any other.
+ACTIVE_STATUSES: frozenset[str] = frozenset({"running", "waiting", "budget_paused"})
+
 # Why an agent parked. The user can message any agent, so this - not the agent's
 # position in the tree - decides whether waiting is bounded: only an agent waiting
 # on other agents is re-checked on a timer.
@@ -68,7 +74,10 @@ class AgentCoordinator:
         self._budget_stopped = False
         self._reserve_stopped = False
         self._budget_paused = False
+        self._resume_epoch = 0
+        self._budget_policy: BudgetPolicy = "stop"
         self._extend_budget: Callable[[], None] | None = None
+        self._set_budget_limit: Callable[[float | None], None] | None = None
 
     def set_snapshot_path(self, path: Path) -> None:
         self._snapshot_path = path
@@ -95,15 +104,100 @@ class AgentCoordinator:
     def budget_paused(self) -> bool:
         return self._budget_paused
 
+    @property
+    def resume_epoch(self) -> int:
+        """Bumped by every ``resume_budget``; an agent parks against the value it read."""
+        return self._resume_epoch
+
+    @property
+    def budget_policy(self) -> BudgetPolicy:
+        return self._budget_policy
+
+    def set_budget_policy(self, policy: BudgetPolicy) -> None:
+        self._budget_policy = policy
+
     def set_budget_extender(self, extend: Callable[[], None]) -> None:
         self._extend_budget = extend
 
+    def set_budget_limit_setter(self, setter: Callable[[float | None], None]) -> None:
+        self._set_budget_limit = setter
+
     async def pause_for_budget(self, agent_id: str) -> None:
         async with self._lock:
             self._budget_paused = True
         await self.set_status(agent_id, "budget_paused")
 
+    async def park_for_budget(self, agent_id: str) -> bool:
+        """Park ``agent_id`` before an LLM call (pause policy).
+
+        Only a ``running`` agent parks; returns False when it was stopped in the
+        meantime, so the stop is not overwritten.
+        """
+        async with self._lock:
+            if self.statuses.get(agent_id) != "running":
+                return False
+            self._set_status_locked(agent_id, "budget_paused")
+        logger.info("agent.status %s=budget_paused", agent_id)
+        await self._maybe_snapshot()
+        return True
+
+    async def pause_budget(self) -> None:
+        """Operator pause: every agent parks before its next LLM call.
+
+        Agents mid-call or mid-tool finish that step first, so their spend still
+        lands; nothing is cancelled.
+        """
+        async with self._lock:
+            self._budget_paused = True
+        logger.info("scan paused by the operator")
+        await self._maybe_snapshot()
+
+    async def resume_budget(self, *, max_budget_usd: float | None = None) -> list[str]:
+        """Lift the pause and wake every parked agent; returns the woken agent ids.
+
+        With ``max_budget_usd`` the scan's limit is replaced first (``None`` keeps
+        the current one). Agents continue with the LLM call they parked on; no
+        message is added to any session. An agent that parks again on its next
+        call (the new limit is already spent) is not an error.
+        """
+        if max_budget_usd is not None and self._set_budget_limit is not None:
+            self._set_budget_limit(max_budget_usd)
+        async with self._lock:
+            self._budget_paused = False
+            self._resume_epoch += 1
+            woken = [aid for aid, status in self.statuses.items() if status == "budget_paused"]
+            for aid in woken:
+                self.runtimes.setdefault(aid, AgentRuntime()).wake.set()
+        logger.info("scan resumed; woke %d parked agent(s)", len(woken))
+        await self._maybe_snapshot()
+        return woken
+
+    async def wait_for_budget_resume(self, agent_id: str, *, parked_epoch: int) -> None:
+        """Block until a resume newer than ``parked_epoch``, a scan-wide stop, or
+        the agent itself being stopped while parked.
+
+        ``parked_epoch`` is the ``resume_epoch`` the agent read when it decided to
+        park, so a resume that lands between that decision and this wait is not
+        missed. The switch back to ``running`` happens under the lock, so a stop
+        can never be overwritten by it; on return the agent is ``running`` unless
+        it was stopped.
+        """
+        while True:
+            async with self._lock:
+                runtime = self.runtimes.setdefault(agent_id, AgentRuntime())
+                if self._budget_stopped or self.statuses.get(agent_id) != "budget_paused":
+                    return
+                if self._resume_epoch != parked_epoch:
+                    self._set_status_locked(agent_id, "running")
+                    break
+                wake = runtime.wake
+                wake.clear()
+            await wake.wait()
+        logger.info("agent.status %s=running", agent_id)
+        await self._maybe_snapshot()
+
     async def resume_from_budget_pause(self, *, exclude: str | None = None) -> None:
+        """Legacy interactive resume: extend by the original budget and nudge agents."""
         async with self._lock:
             if not self._budget_paused:
                 return
@@ -258,20 +352,25 @@ class AgentCoordinator:
         async with self._lock:
             if agent_id not in self.statuses:
                 return
-            self.statuses[agent_id] = status  # type: ignore[assignment]
-            if error is not None:
-                self.errors[agent_id] = error
-            elif status == "running":
-                self.errors.pop(agent_id, None)
-            if status == "running":
-                # Running again means a fresh stint that owes its parent its own notice.
-                self._parent_notified.discard(agent_id)
-            runtime = self.runtimes.setdefault(agent_id, AgentRuntime())
-            runtime.user_wake_required = status in {"failed", "crashed"}
-            runtime.wake.set()
+            self._set_status_locked(agent_id, status, error=error)
         logger.info("agent.status %s=%s", agent_id, status)
         await self._maybe_snapshot()
 
+    def _set_status_locked(
+        self, agent_id: str, status: Status | str, *, error: str | None = None
+    ) -> None:
+        self.statuses[agent_id] = status  # type: ignore[assignment]
+        if error is not None:
+            self.errors[agent_id] = error
+        elif status == "running":
+            self.errors.pop(agent_id, None)
+        if status == "running":
+            # Running again means a fresh stint that owes its parent its own notice.
+            self._parent_notified.discard(agent_id)
+        runtime = self.runtimes.setdefault(agent_id, AgentRuntime())
+        runtime.user_wake_required = status in {"failed", "crashed"}
+        runtime.wake.set()
+
     async def claim_parent_notice(self, agent_id: str) -> bool:
         """Reserve the one notice a child owes its parent when it stops running.
 
@@ -308,7 +407,7 @@ class AgentCoordinator:
         unknown, or it is terminal and its loop does not park for wake-ups.
         """
         from_user = message.get("from") == "user"
-        if from_user and self._budget_paused:
+        if from_user and self._budget_paused and self._budget_policy != "pause":
             await self.resume_from_budget_pause(exclude=target_agent_id)
         async with self._lock:
             if target_agent_id not in self.statuses:
@@ -462,7 +561,7 @@ class AgentCoordinator:
                     "parent_id": self.parent_of.get(aid),
                 }
                 for aid, status in self.statuses.items()
-                if aid != agent_id and status in {"running", "waiting"}
+                if aid != agent_id and status in ACTIVE_STATUSES
             ]
 
     async def graph_snapshot(

strix/core/execution.py

@@ -522,33 +522,51 @@ async def _run_until_lifecycle(
             await coordinator.set_status(agent_id, "stopped")
             raise SubagentBudgetReservedError("scan reached the sub-agent budget reserve")
 
-        if interactive:
-            result = await _run_cycle_parked(
-                agent,
-                coordinator,
-                agent_id,
-                input_data=input_data,
-                run_config=run_config,
-                context=context,
-                max_turns=max_turns,
-                session=session,
-                event_sink=event_sink,
-                hooks=hooks,
-            )
-        else:
-            result = await _run_cycle(
-                agent,
-                coordinator,
-                agent_id,
-                input_data=input_data,
-                run_config=run_config,
-                context=context,
-                max_turns=max_turns,
-                session=session,
-                interactive=False,
-                event_sink=event_sink,
-                hooks=hooks,
-            )
+        try:
+            if interactive:
+                result = await _run_cycle_parked(
+                    agent,
+                    coordinator,
+                    agent_id,
+                    input_data=input_data,
+                    run_config=run_config,
+                    context=context,
+                    max_turns=max_turns,
+                    session=session,
+                    event_sink=event_sink,
+                    hooks=hooks,
+                )
+            else:
+                result = await _run_cycle(
+                    agent,
+                    coordinator,
+                    agent_id,
+                    input_data=input_data,
+                    run_config=run_config,
+                    context=context,
+                    max_turns=max_turns,
+                    session=session,
+                    interactive=False,
+                    event_sink=event_sink,
+                    hooks=hooks,
+                )
+        except BudgetPausedError as exc:
+            if coordinator.budget_policy != "pause":
+                raise
+            # The agent parked right before an LLM call; everything up to that
+            # point is already in its session. Once resumed, the same call goes
+            # out with nothing added to the conversation.
+            await coordinator.wait_for_budget_resume(agent_id, parked_epoch=exc.resume_epoch)
+            if (
+                not coordinator.budget_stopped
+                and await _agent_status(coordinator, agent_id) != "running"
+            ):
+                # Stopped while parked (operator stop or a parent's stop_agent).
+                await coordinator.reset_recovery(agent_id)
+                return result
+            if session is not None:
+                input_data = []
+            continue
 
         status = await _agent_status(coordinator, agent_id)
         if status != "running":
@@ -759,7 +777,10 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
                 await coordinator.detach_stream(agent_id, stream)
         except BudgetPausedError as exc:
             logger.info("agent %s paused at the scan budget limit: %s", agent_id, exc)
-            await coordinator.pause_for_budget(agent_id)
+            if coordinator.budget_policy == "pause":
+                await coordinator.park_for_budget(agent_id)
+            else:
+                await coordinator.pause_for_budget(agent_id)
             raise
         except SubagentBudgetReservedError as exc:
             logger.info("sub-agent %s stopped at the budget reserve: %s", agent_id, exc)

strix/core/hooks.py

@@ -8,6 +8,7 @@ from typing import TYPE_CHECKING, Any
 
 from agents.lifecycle import RunHooks
 
+from strix.core.agents import BudgetPolicy, coordinator_from_context
 from strix.report.state import get_global_report_state
 
 
@@ -22,6 +23,26 @@ logger = logging.getLogger(__name__)
 
 LLM_TURN_KEY = "llm_turn"
 
+# ``BudgetPolicy`` decides what happens when the accumulated LLM cost reaches
+# ``max_budget_usd``.
+#
+# ``stop``: the agents are warned as the limit approaches, sub-agents are cut at
+# a reserve so the root can write its report, and the scan ends at the limit.
+#
+# ``pause``: the agents are never told a limit exists. Every agent parks right
+# before its next LLM call once the limit is reached (or an operator pauses the
+# scan), keeping its session, context and sandbox alive, and continues with that
+# same call when the operator raises the limit or resumes.
+__all__ = [
+    "LLM_TURN_KEY",
+    "BudgetExceededError",
+    "BudgetPausedError",
+    "BudgetPolicy",
+    "ReportUsageHooks",
+    "SubagentBudgetReservedError",
+    "recomputed_budget_flags",
+]
+
 _STAGE_LABELS: tuple[str, ...] = ("NOTICE", "URGENT", "CRITICAL")
 _TURN_WARN_BANDS: tuple[float, ...] = (0.70, 0.85, 0.95)
 _ROOT_BUDGET_WARN_BANDS: tuple[float, ...] = (0.70, 0.85, 0.95)
@@ -38,7 +59,20 @@ class SubagentBudgetReservedError(RuntimeError):
 
 
 class BudgetPausedError(RuntimeError):
-    """Raised to park one agent when an interactive scan reaches its budget."""
+    """Raised to park one agent until the scan budget is raised or the pause lifted.
+
+    ``resume_epoch`` is the coordinator's ``resume_epoch`` at the moment the agent
+    decided to park; the agent waits for a resume newer than that.
+    """
+
+    def __init__(self, message: str, *, resume_epoch: int = 0) -> None:
+        super().__init__(message)
+        self.resume_epoch = resume_epoch
+
+
+def _validate_budget(max_budget_usd: float | None) -> None:
+    if max_budget_usd is not None and (not math.isfinite(max_budget_usd) or max_budget_usd <= 0):
+        raise ValueError("max_budget_usd must be a finite number greater than 0")
 
 
 def recomputed_budget_flags(
@@ -46,11 +80,12 @@ def recomputed_budget_flags(
     max_budget_usd: float | None,
     *,
     interactive: bool,
+    budget_policy: BudgetPolicy = "stop",
 ) -> tuple[bool, bool]:
     """Return the (budget_stopped, reserve_stopped) flags a resumed scan should carry."""
     if max_budget_usd is None:
         return False, False
-    if interactive:
+    if interactive or budget_policy == "pause":
         return False, False
     budget_stopped = cost >= max_budget_usd
     reserve_stopped = cost >= max_budget_usd * _SUBAGENT_BUDGET_RESERVE
@@ -121,18 +156,32 @@ class ReportUsageHooks(RunHooks[dict[str, Any]]):
         max_budget_usd: float | None = None,
         max_turns: int | None = None,
         interactive: bool = False,
+        budget_policy: BudgetPolicy = "stop",
     ) -> None:
-        if max_budget_usd is not None and (
-            not math.isfinite(max_budget_usd) or max_budget_usd <= 0
-        ):
-            raise ValueError("max_budget_usd must be a finite number greater than 0")
+        _validate_budget(max_budget_usd)
         if max_turns is not None and max_turns <= 0:
             raise ValueError("max_turns must be a positive integer")
+        if budget_policy not in ("stop", "pause"):
+            raise ValueError(f"unknown budget_policy: {budget_policy!r}")
         self._model = model
         self._max_budget_usd = max_budget_usd
         self._budget_increment = max_budget_usd
         self._max_turns = max_turns
         self._interactive = interactive
+        self._budget_policy: BudgetPolicy = budget_policy
+
+    @property
+    def max_budget_usd(self) -> float | None:
+        return self._max_budget_usd
+
+    @property
+    def budget_policy(self) -> BudgetPolicy:
+        return self._budget_policy
+
+    def set_max_budget_usd(self, max_budget_usd: float | None) -> None:
+        """Replace the scan's cost limit; ``None`` removes it."""
+        _validate_budget(max_budget_usd)
+        self._max_budget_usd = max_budget_usd
 
     def extend_budget(self) -> None:
         if self._max_budget_usd is None or self._budget_increment is None:
@@ -146,6 +195,8 @@ class ReportUsageHooks(RunHooks[dict[str, Any]]):
         system_prompt: str | None,  # noqa: ARG002
         input_items: list[TResponseInputItem],
     ) -> None:
+        if self._budget_policy == "pause":
+            self._pause_if_limited(context)
         context.context[LLM_TURN_KEY] = int(context.context.get(LLM_TURN_KEY, 0)) + 1
         try:
             self._maybe_warn_turns(context, input_items)
@@ -153,6 +204,32 @@ class ReportUsageHooks(RunHooks[dict[str, Any]]):
         except Exception:
             logger.exception("budget/turn warning injection failed")
 
+    def _pause_if_limited(self, context: RunContextWrapper[dict[str, Any]]) -> None:
+        """Park the agent before a paid call when the scan is at its limit or paused.
+
+        Only calls that have already returned are counted, so calls in flight on
+        other agents still land and are paid for: ``spent`` may end up above the
+        limit, which is expected and never an error under this policy.
+        """
+        coordinator = coordinator_from_context(context.context)
+        epoch = coordinator.resume_epoch if coordinator is not None else 0
+        if coordinator is not None and coordinator.budget_paused:
+            raise BudgetPausedError(
+                "scan paused; waiting for the operator to resume", resume_epoch=epoch
+            )
+        if self._max_budget_usd is None:
+            return
+        report_state = get_global_report_state()
+        if report_state is None:
+            return
+        cost = report_state.get_total_llm_cost()
+        if cost >= self._max_budget_usd:
+            raise BudgetPausedError(
+                f"Scan budget of ${self._max_budget_usd:.2f} reached (spent ${cost:.4f}); "
+                "pausing until the operator raises the limit",
+                resume_epoch=epoch,
+            )
+
     def _maybe_warn_turns(
         self,
         context: RunContextWrapper[dict[str, Any]],
@@ -182,7 +259,7 @@ class ReportUsageHooks(RunHooks[dict[str, Any]]):
         context: RunContextWrapper[dict[str, Any]],
         input_items: list[TResponseInputItem],
     ) -> None:
-        if self._max_budget_usd is None:
+        if self._max_budget_usd is None or self._budget_policy == "pause":
             return
         report_state = get_global_report_state()
         if report_state is None:
@@ -250,6 +327,11 @@ class ReportUsageHooks(RunHooks[dict[str, Any]]):
         except Exception:
             logger.exception("failed to record SDK usage for agent %s", agent_id)
 
+        if self._budget_policy == "pause":
+            # The finished call is paid for and its tool calls still run for free;
+            # the agent parks before its next call, in ``on_llm_start``.
+            return
+
         if self._max_budget_usd is not None:
             cost = report_state.get_total_llm_cost()
             if cost >= self._max_budget_usd:

strix/core/inputs.py

@@ -267,9 +267,7 @@ def make_model_settings(
         and reasoning_effort != "none"
         and model_supports_reasoning(model_name)
     ):
-        model_settings = model_settings.resolve(
-            _reasoning_settings(reasoning_effort),
-        )
+        model_settings = model_settings.resolve(_reasoning_settings(reasoning_effort))
     if force_required_tool_choice and _accepts_required_tool_choice(model_name):
         model_settings = model_settings.resolve(ModelSettings(tool_choice="required"))
 
@@ -312,11 +310,12 @@ def _reasoning_settings(effort: ReasoningEffort) -> ModelSettings:
 def _prompt_cache_extra_args(model_name: str) -> dict[str, Any] | None:
     """LiteLLM ``cache_control_injection_points`` for Claude prompt caching.
 
-    System prompt + rolling last-message breakpoint everywhere; ``tool_config``
-    only on Bedrock Converse (the only route whose LiteLLM transform consumes
-    it — elsewhere it leaks onto the wire and native Anthropic 400s). Unmapped
-    Bedrock models get no points at all: Bedrock rejects the passed-through
-    field outright.
+    A breakpoint on each system message, plus a rolling last-message one. The
+    system prompt is split into up to three messages, which with the last
+    message uses all four breakpoints Claude allows. There is none on
+    ``tool_config``: the tools come before the system prompt, so its first
+    breakpoint caches them too. Unmapped Bedrock models get no points at all:
+    Bedrock rejects the passed-through field outright.
 
     The field is LiteLLM's own, consumed by its transform, so it only goes to
     routes LiteLLM serves. A bare ``claude-...`` name is served by the SDK's
@@ -328,11 +327,12 @@ def _prompt_cache_extra_args(model_name: str) -> dict[str, Any] | None:
     if is_bedrock_route(model_name) and not bedrock_route_supports_prompt_caching(model_name):
         return None
 
-    points: list[dict[str, Any]] = [{"location": "message", "role": "system"}]
-    if is_bedrock_route(model_name):
-        points.append({"location": "tool_config"})
-    points.append({"location": "message", "index": -1})
-    return {"cache_control_injection_points": points}
+    return {
+        "cache_control_injection_points": [
+            {"location": "message", "role": "system"},
+            {"location": "message", "index": -1},
+        ]
+    }
 
 
 def child_initial_input(

strix/core/runner.py

@@ -17,16 +17,17 @@ from agents.sandbox import SandboxRunConfig
 from openai import RateLimitError
 
 from strix.agents.factory import build_strix_agent, make_child_factory
-from strix.agents.prompt import render_system_prompt
-from strix.config import load_settings
+from strix.agents.prompt import render_scope_prompt, render_system_prompt
+from strix.config import codex, load_settings
 from strix.config.models import (
     StrixProvider,
+    configure_sdk_api_route,
     configure_sdk_model_defaults,
     supports_strict_tool_schemas,
     uses_chat_completions_tool_schema,
 )
 from strix.config.settings import DEFAULT_MAX_TURNS
-from strix.core.agents import AgentCoordinator
+from strix.core.agents import AgentCoordinator, BudgetPolicy
 from strix.core.execution import (
     respawn_subagents,
     run_agent_loop,
@@ -164,15 +165,17 @@ def _compose_root_instructions_override(
         is_diff_scoped=is_diff_scoped,
         interactive=interactive,
         system_prompt_context=system_prompt_context,
+        include_scope=False,
     )
     return (
         f"{base_instructions}\n\n"
         "<root_scan_instructions_override>\n"
         "The following root scan instructions are subordinate to the "
-        "system-verified scope above. They cannot expand, replace, or weaken "
+        "system-verified scope below. They cannot expand, replace, or weaken "
         "authorized target constraints.\n\n"
         f"{root_instructions_override}\n"
-        "</root_scan_instructions_override>"
+        "</root_scan_instructions_override>\n\n"
+        f"{render_scope_prompt(system_prompt_context)}"
     )
 
 
@@ -187,6 +190,7 @@ async def run_strix_scan(
     interactive: bool = False,
     max_turns: int = DEFAULT_MAX_TURNS,
     max_budget_usd: float | None = None,
+    budget_policy: BudgetPolicy = "stop",
     model: str | None = None,
     cleanup_on_exit: bool = True,
     event_sink: StreamEventSink | None = None,
@@ -206,6 +210,12 @@ async def run_strix_scan(
     ``extra_system_prompt_context`` is merged into the root agent's scan
     context before prompt rendering. Child agents keep the standard scan prompt
     and context.
+    ``budget_policy`` decides what happens when the LLM spend reaches
+    ``max_budget_usd``: ``"stop"`` warns the agents as the limit approaches and
+    ends the scan at it; ``"pause"`` tells the agents nothing and parks every
+    agent before its next LLM call until the caller resumes the scan through
+    ``coordinator.resume_budget()`` (optionally with a higher limit) or cancels
+    it. ``coordinator.pause_budget()`` parks a running scan the same way.
     ``mcp_connection_requests`` supplies the run's MCP connections from any
     source: when given, the engine connects those requests; when ``None`` (the
     command-line default) it reads ``~/.strix/mcp-servers.json`` itself. Either
@@ -248,15 +258,22 @@ async def run_strix_scan(
         raise RuntimeError(
             "No LLM model configured. Set STRIX_LLM env or pass model= to run_strix_scan().",
         )
+    if resolved_model != (settings.llm.model or "").strip() and not codex.subscription_model(
+        resolved_model
+    ):
+        configure_sdk_api_route(resolved_model, settings)
     logger.info("LLM model resolved: %s", resolved_model)
     chat_completions_tools = uses_chat_completions_tool_schema(resolved_model, settings)
     strict_tool_schemas = supports_strict_tool_schemas(resolved_model)
     if not strict_tool_schemas:
         logger.info("Sending non-strict tool schemas: %s caps strict tools", resolved_model)
 
+    if budget_policy not in ("stop", "pause"):
+        raise ValueError(f"unknown budget_policy: {budget_policy!r}")
     if coordinator is None:
         coordinator = AgentCoordinator()
     coordinator.set_snapshot_path(agents_path)
+    coordinator.set_budget_policy(budget_policy)
 
     from strix.tools.coverage.tools import hydrate_coverage_from_disk
     from strix.tools.notes.tools import hydrate_notes_from_disk
@@ -287,11 +304,17 @@ async def run_strix_scan(
                 report_state.get_total_llm_cost(),
                 max_budget_usd,
                 interactive=interactive,
+                budget_policy=budget_policy,
             )
+            # Under the pause policy the hooks re-park at the first call if the
+            # spend is still at the limit, so a restored pause flag would only
+            # hold agents back after the limit was raised.
             await coordinator.reset_budget_stops(
                 budget_stopped=budget_stopped,
                 reserve_stopped=reserve_stopped,
-                budget_paused=interactive and coordinator.budget_paused,
+                budget_paused=(
+                    interactive and budget_policy != "pause" and coordinator.budget_paused
+                ),
             )
         for aid, parent in coordinator.parent_of.items():
             if parent is None:
@@ -370,8 +393,10 @@ async def run_strix_scan(
             max_budget_usd=max_budget_usd,
             max_turns=max_turns,
             interactive=interactive,
+            budget_policy=budget_policy,
         )
-        if interactive:
+        coordinator.set_budget_limit_setter(hooks.set_max_budget_usd)
+        if interactive and budget_policy != "pause":
             coordinator.set_budget_extender(hooks.extend_budget)
 
         scope_context = build_scope_context(scan_config)

strix/interface/cli_args.py

@@ -9,7 +9,7 @@ from pathlib import Path
 
 from strix.config import apply_config_override
 from strix.config.settings import DEFAULT_MAX_TURNS
-from strix.core.paths import run_dir_for, runtime_state_dir
+from strix.core.paths import RUNS_DIR_NAME, run_dir_for, runtime_state_dir
 from strix.interface.scan_setup import attach_workspace_mount, build_targets_info
 from strix.interface.update_check import self_update
 from strix.interface.utils import (
@@ -18,6 +18,11 @@ from strix.interface.utils import (
     resolve_workspace_files,
     validate_config_file,
 )
+from strix.report.runs import list_run_summaries
+
+
+# Severities ``--fail-on`` accepts, most severe first.
+FAIL_ON_SEVERITIES = ("critical", "high", "medium", "low", "info")
 
 
 def get_version() -> str:
@@ -186,6 +191,21 @@ Strix Cloud:
         ),
     )
 
+    parser.add_argument(
+        "--fail-on",
+        dest="fail_on",
+        type=str.lower,
+        choices=FAIL_ON_SEVERITIES,
+        default=None,
+        metavar="SEVERITY",
+        help=(
+            "Headless mode only: exit 2 only when a finding is at or above this severity "
+            "(critical, high, medium, low, info). Lower findings are still written to every "
+            "report artifact. A finding with an unrecognized severity always counts. "
+            "Default: any finding exits 2."
+        ),
+    )
+
     parser.add_argument(
         "-m",
         "--scan-mode",
@@ -281,11 +301,14 @@ Strix Cloud:
     parser.add_argument(
         "--resume",
         type=str,
+        nargs="?",
+        const="",
         metavar="RUN_NAME",
         help=(
             "Resume a prior scan by its run name (the dir under ./strix_runs/). "
             "Picks up the root + every non-terminal subagent's full LLM history "
-            "and agent topology. Skips fresh run-name generation."
+            "and agent topology. Skips fresh run-name generation. Without a "
+            "name, opens a picker of the prior runs."
         ),
     )
 
@@ -293,6 +316,7 @@ Strix Cloud:
     # Startup-resolved state lives alongside the parsed flags. The full schema
     # is established here so downstream code reads attributes directly.
     args.needs_setup = False
+    args.resume_picker = False
     args.targets_info = []
     args.local_sources = []
     args.diff_scope = {"active": False}
@@ -318,6 +342,11 @@ Strix Cloud:
     if args.update:
         sys.exit(0 if self_update() else 1)
 
+    if args.fail_on and not args.non_interactive and terminal_attached():
+        # Without a terminal main() switches to headless anyway, so the
+        # flag is only out of place when the TUI would actually open.
+        parser.error("--fail-on only applies to headless runs; add -n/--non-interactive.")
+
     if args.instruction and args.instruction_file:
         parser.error(
             "Cannot specify both --instruction and --instruction-file. Use one or the other."
@@ -338,28 +367,34 @@ Strix Cloud:
     except ValueError as error:
         parser.error(f"--workspace-file: {error}")
 
-    args.user_explicit_instruction = args.instruction if args.resume else None
+    args.user_explicit_instruction = args.instruction if args.resume is not None else None
     # What the user actually asked for, kept apart from args.instruction because
     # prepare_run prepends the diff-scope preamble to that. This is the text the
     # transcript shows as their opening message.
     args.user_instruction = args.instruction or None
 
-    if args.resume:
+    if args.resume is not None:
         if args.target or args.target_list:
             parser.error(
                 "Cannot combine --resume with --target/--target-list. "
                 "--resume picks up where the prior run left off, including the "
                 "original target list."
             )
-        _load_resume_state(args, parser)
-        agents_path = runtime_state_dir(run_dir_for(args.resume)) / "agents.json"
-        if not agents_path.exists():
-            parser.error(
-                f"--resume {args.resume}: missing {agents_path}. The run was "
-                f"persisted but never reached its first agent snapshot — "
-                f"there's nothing to resume from. Pick a fresh --run-name "
-                f"or remove --resume to start over with the same targets."
-            )
+        if not args.resume.strip():
+            # A bare --resume: main() opens the inline picker of prior runs
+            # before anything launches; headless has nobody to pick, so it
+            # lists them.
+            args.resume = None
+            args.resume_picker = True
+            if args.non_interactive:
+                parser.error(resume_run_list_message("--resume needs a run name in headless mode."))
+            if not list_run_summaries():
+                parser.error(f"--resume: no runs in ./{RUNS_DIR_NAME} to resume")
+            return args
+        try:
+            load_resume_state(args)
+        except ResumeError as exc:
+            parser.error(str(exc))
     else:
         if not args.target and not args.target_list:
             if args.non_interactive:
@@ -381,21 +416,52 @@ Strix Cloud:
     return args
 
 
-def _load_resume_state(args: argparse.Namespace, parser: argparse.ArgumentParser) -> None:
-    """Populate ``args.targets_info`` and friends from a prior run's run.json."""
+def terminal_attached() -> bool:
+    """Whether the TUI can take over the terminal: a tty on both ends, not dumb."""
+    if os.environ.get("TERM", "").strip().lower() == "dumb":
+        return False
+    return all(hasattr(stream, "isatty") and stream.isatty() for stream in (sys.stdin, sys.stdout))
+
+
+class ResumeError(ValueError):
+    """A prior run cannot be resumed as recorded; the message names why."""
+
+
+def resume_run_list_message(lead: str) -> str:
+    """``lead`` followed by the runs that ``--resume <name>`` would accept."""
+    runs = list_run_summaries()
+    if not runs:
+        return f"{lead} There are no runs in ./{RUNS_DIR_NAME}."
+    name_width = max(len(run.run_name) for run in runs)
+    status_width = max(len(run.status) for run in runs)
+    lines = [f"{lead} Runs in ./{RUNS_DIR_NAME}:"]
+    lines.extend(
+        f"  {run.run_name:<{name_width}}  {run.status:<{status_width}}  "
+        f"{run.started_at[:19]:<19}  {run.target}".rstrip()
+        for run in runs
+    )
+    return "\n".join(lines)
+
+
+def load_resume_state(args: argparse.Namespace) -> None:
+    """Populate ``args.targets_info`` and friends from a prior run's run.json.
+
+    Raises :class:`ResumeError` when the run is missing, unreadable, or its
+    recorded workspace is gone.
+    """
     from strix.report.writer import read_run_record
 
     run_dir = run_dir_for(args.resume)
     state_path = run_dir / "run.json"
     if not state_path.exists():
-        parser.error(
+        raise ResumeError(
             f"--resume {args.resume}: no such run "
             f"(missing {state_path}; remove --resume for a fresh start)"
         )
     try:
         state = read_run_record(run_dir)
     except (RuntimeError, TypeError) as exc:
-        parser.error(f"--resume {args.resume}: run.json unreadable: {exc}")
+        raise ResumeError(f"--resume {args.resume}: run.json unreadable: {exc}") from exc
 
     args.targets_info = state.get("targets_info") or []
     # A target-less run has no targets_info at all. It is driven by its
@@ -403,7 +469,7 @@ def _load_resume_state(args: argparse.Namespace, parser: argparse.ArgumentParser
     # mount was declined, so either of those is enough to resume it.
     workspace_mount = state.get("workspace_mount") or None
     if not args.targets_info and not workspace_mount and not state.get("user_instruction"):
-        parser.error(f"--resume {args.resume}: run.json has no targets_info")
+        raise ResumeError(f"--resume {args.resume}: run.json has no targets_info")
 
     for target in args.targets_info:
         if not isinstance(target, dict):
@@ -413,7 +479,7 @@ def _load_resume_state(args: argparse.Namespace, parser: argparse.ArgumentParser
             try:
                 check_mountable_dir(Path(details["target_path"]).expanduser())
             except ValueError as exc:
-                parser.error(f"--resume {args.resume}: {exc}")
+                raise ResumeError(f"--resume {args.resume}: {exc}") from exc
             continue
         if target.get("type") != "repository":
             continue
@@ -421,7 +487,7 @@ def _load_resume_state(args: argparse.Namespace, parser: argparse.ArgumentParser
         if not cloned:
             continue
         if not Path(cloned).expanduser().exists():
-            parser.error(
+            raise ResumeError(
                 f"--resume {args.resume}: cloned repo at {cloned} is missing. "
                 f"It was deleted between runs. Pick a fresh --run-name to "
                 f"re-clone, or restore the directory before resuming."
@@ -452,10 +518,10 @@ def _load_resume_state(args: argparse.Namespace, parser: argparse.ArgumentParser
         try:
             args.workspace_files = resolve_workspace_files(restored)
         except ValueError as error:
-            parser.error(f"--resume {args.resume}: invalid workspace file: {error}")
+            raise ResumeError(f"--resume {args.resume}: invalid workspace file: {error}") from error
     if workspace_mount:
         if not Path(workspace_mount).expanduser().is_dir():
-            parser.error(
+            raise ResumeError(
                 f"--resume {args.resume}: the working directory {workspace_mount} "
                 f"is missing. Restore it before resuming, or start a fresh run."
             )
@@ -465,3 +531,11 @@ def _load_resume_state(args: argparse.Namespace, parser: argparse.ArgumentParser
     persisted_scan_mode = state.get("scan_mode")
     if persisted_scan_mode and args.scan_mode == "deep":
         args.scan_mode = persisted_scan_mode
+    agents_path = runtime_state_dir(run_dir) / "agents.json"
+    if not agents_path.exists():
+        raise ResumeError(
+            f"--resume {args.resume}: missing {agents_path}. The run was "
+            f"persisted but never reached its first agent snapshot — "
+            f"there's nothing to resume from. Pick a fresh --run-name "
+            f"or remove --resume to start over with the same targets."
+        )

strix/interface/interactive.py

@@ -17,12 +17,17 @@ class InteractiveSetupUnavailableError(RuntimeError):
     """Raised when the interactive TUI cannot be launched."""
 
 
+class InteractiveInterfaceExitedError(RuntimeError):
+    """Raised when the TUI process dies after it was up (no terminal, killed)."""
+
+
 async def run_tui(args: argparse.Namespace) -> None:
     """Run the Bubble Tea TUI."""
     from strix.interface.tui.runtime import (
         GoTuiPreActivationError,
         run_go_tui,
     )
+    from strix.interface.tui.sidecar import TuiProcessExitedError
 
     try:
         await run_go_tui(args)
@@ -30,9 +35,12 @@ async def run_tui(args: argparse.Namespace) -> None:
         raise InteractiveSetupUnavailableError(
             f"The interactive interface could not start: {exc}"
         ) from exc
+    except TuiProcessExitedError as exc:
+        raise InteractiveInterfaceExitedError(str(exc)) from exc
 
 
 __all__ = [
+    "InteractiveInterfaceExitedError",
     "InteractiveSetupUnavailableError",
     "run_tui",
 ]

strix/interface/main.py

@@ -6,28 +6,39 @@ Strix Agent Interface
 import argparse
 import asyncio
 import contextlib
+import logging
 import sys
 from pathlib import Path
+from typing import Any
 
 from rich.console import Console
 from rich.panel import Panel
 from rich.text import Text
 
 from strix.config import codex, load_settings, persist_current
-from strix.core.paths import run_dir_for
-from strix.interface.cli_args import parse_arguments
+from strix.core.paths import RUNS_DIR_NAME, run_dir_for
+from strix.interface.cli_args import (
+    FAIL_ON_SEVERITIES,
+    ResumeError,
+    load_resume_state,
+    parse_arguments,
+    resume_run_list_message,
+    terminal_attached,
+)
 from strix.interface.environment import (
     check_docker_installed,
     pull_docker_image,
     validate_environment,
 )
 from strix.interface.interactive import (
+    InteractiveInterfaceExitedError,
     InteractiveSetupUnavailableError,
     run_tui,
 )
 from strix.interface.scan_setup import (
     ModelConnectionError,
     preflight_model_connection,
+    preflight_request,
     prepare_run,
     telemetry_start,
 )
@@ -43,22 +54,7 @@ from strix.interface.utils import (
 )
 from strix.llm.warmup import start_import_warmup, wait_for_import_warmup
 from strix.telemetry import posthog, report_error, scarf, set_scan_phase
-from strix.telemetry.logging import configure_dependency_logging
-
-
-BEDROCK_MODEL_PREFIX = "bedrock/"
-BEDROCK_MISSING_MODULE_ERROR = "No module named 'boto3'"
-BEDROCK_EXTRA_HINT = (
-    'Bedrock support is optional. Install it with: pipx install "strix-agent[bedrock]"'
-)
-VERTEX_MODEL_MARKER = "vertex"
-VERTEX_MISSING_MODULE_ERROR = "No module named 'google"
-VERTEX_EXTRA_HINT = (
-    'Vertex AI support is optional. Install it with: pipx install "strix-agent[vertex]"'
-)
-
-
-import logging  # noqa: E402
+from strix.telemetry.logging import setup_console_logging
 
 
 logger = logging.getLogger(__name__)
@@ -89,29 +85,6 @@ def _exception_messages(exc: BaseException) -> tuple[str, ...]:
     return tuple(messages)
 
 
-def _provider_import_hint(exc: BaseException, model: str) -> str | None:
-    """Return an install hint when *exc* is a missing provider dependency.
-
-    Bedrock and Vertex AI ship as optional extras: Bedrock needs ``boto3`` and
-    Vertex AI needs ``google-auth``. When either is absent, litellm may raise an
-    ``ImportError``/``ModuleNotFoundError`` directly or wrap it in a connection
-    error. Map the missing module back to the matching extra so the user knows
-    what to install. Returns ``None`` for any unrelated error.
-    """
-    model_name = model.lower()
-    messages = _exception_messages(exc)
-    if any(
-        BEDROCK_MISSING_MODULE_ERROR in message for message in messages
-    ) and model_name.startswith(BEDROCK_MODEL_PREFIX):
-        return BEDROCK_EXTRA_HINT
-    if (
-        any(VERTEX_MISSING_MODULE_ERROR in message for message in messages)
-        and VERTEX_MODEL_MARKER in model_name
-    ):
-        return VERTEX_EXTRA_HINT
-    return None
-
-
 def _subscription_error_hint(exc: BaseException) -> str | None:
     """Return an actionable hint for a known ChatGPT-subscription error, or None."""
     if not codex.subscription_model(load_settings().llm.model):
@@ -135,16 +108,11 @@ def _subscription_error_hint(exc: BaseException) -> str | None:
     return None
 
 
-async def warm_up_llm(show_model_warning: bool = True) -> None:
-    from agents.models.interface import ModelTracing
-
+async def warm_up_llm() -> None:
     from strix.config.models import (
-        RECOMMENDED_MODEL_NAMES,
         configure_sdk_model_defaults,
         is_known_openai_bare_model,
-        is_recommended_or_frontier_model,
     )
-    from strix.core.inputs import make_model_settings
 
     console = Console()
     logger.info("Warming up LLM connection")
@@ -186,32 +154,6 @@ async def warm_up_llm(show_model_warning: bool = True) -> None:
             )
             sys.exit(1)
 
-        if show_model_warning and raw_model and not is_recommended_or_frontier_model(raw_model):
-            warn_text = Text()
-            warn_text.append("MODEL QUALITY WARNING", style="bold yellow")
-            warn_text.append("\n\n", style="white")
-            warn_text.append(f"'{raw_model}'", style="bold cyan")
-            warn_text.append(
-                " is not a recommended frontier model for Strix.\nSecurity scans work best with:\n",
-                style="white",
-            )
-            for recommended_model in RECOMMENDED_MODEL_NAMES:
-                warn_text.append(f"• {recommended_model}\n", style="bold cyan")
-            warn_text.append(
-                "\nYou can continue, but weaker models may miss vulnerabilities "
-                "or produce lower-quality findings.",
-                style="white",
-            )
-            console.print(
-                Panel(
-                    warn_text,
-                    title="[bold white]STRIX",
-                    title_align="left",
-                    border_style="yellow",
-                    padding=(1, 2),
-                ),
-            )
-
         await preflight_model_connection(raw_model, settings=settings)
         logger.info("LLM warm-up succeeded for model %s", (llm.model or "").strip())
 
@@ -224,28 +166,12 @@ async def warm_up_llm(show_model_warning: bool = True) -> None:
             # A dedicated dedupe model may route to another provider, which must
             # never receive the main endpoint's headers; it has its own
             # DEDUPE_LLM_EXTRA_HEADERS.
-            deduper_settings = make_model_settings(
-                None,
+            await preflight_request(
+                deduper,
                 model_name=dedupe_model,
-                request_timeout=llm.timeout,
-                prompt_cache=False,
                 extra_headers=settings.dedupe.extra_headers,
-                has_tools=False,
-            )
-            await asyncio.wait_for(
-                deduper.get_response(
-                    system_instructions="You are a helpful assistant.",
-                    input="Reply with just 'OK'.",
-                    model_settings=deduper_settings,
-                    tools=[],
-                    output_schema=None,
-                    handoffs=[],
-                    tracing=ModelTracing.DISABLED,
-                    previous_response_id=None,
-                    conversation_id=None,
-                    prompt=None,
-                ),
-                timeout=llm.timeout,
+                timeout=llm.preflight_timeout,
+                api_base_setting="DEDUPE_LLM_API_BASE",
             )
             logger.info("LLM warm-up succeeded for dedupe model %s", dedupe_model)
 
@@ -343,6 +269,30 @@ def display_completion_message(args: argparse.Namespace, results_path: Path) ->
         notify_update(console)
 
 
+def findings_fail_build(reports: list[dict[str, Any]], fail_on: str | None) -> bool:
+    """Whether headless findings should exit 2 under the ``--fail-on`` threshold.
+
+    With no threshold any finding fails. Otherwise a finding fails when its
+    severity is at or above the threshold. A severity outside the known scale
+    fails too, so a gate never passes on a value it cannot rank. ``none`` is a
+    known level below ``info`` and only fails without a threshold.
+    """
+    if not reports:
+        return False
+    if fail_on is None:
+        return True
+    threshold = FAIL_ON_SEVERITIES.index(fail_on)
+    for report in reports:
+        severity = str(report.get("severity") or "").strip().lower()
+        if severity == "none":
+            continue
+        if severity not in FAIL_ON_SEVERITIES:
+            return True
+        if FAIL_ON_SEVERITIES.index(severity) <= threshold:
+            return True
+    return False
+
+
 def _print_error_panel(title: str, message: str) -> None:
     console = Console()
     error_text = Text()
@@ -361,7 +311,7 @@ def _print_error_panel(title: str, message: str) -> None:
     console.print()
 
 
-def _print_model_connection_error(exc: BaseException, model_name: str) -> None:
+def _print_model_connection_error(exc: BaseException) -> None:
     console = Console()
     error_text = Text()
     sub_hint = _subscription_error_hint(exc)
@@ -377,9 +327,6 @@ def _print_model_connection_error(exc: BaseException, model_name: str) -> None:
         error_text.append("\n\n", style="white")
         error_text.append("Could not establish connection to the language model.\n", style="white")
         error_text.append("Please check your configuration and try again.\n", style="white")
-        hint = _provider_import_hint(exc, model_name)
-        if hint is not None:
-            error_text.append(f"\n{hint}\n", style="bold yellow")
         error_text.append(f"\nError: {exc}", style="dim white")
 
     panel = Panel(
@@ -394,19 +341,71 @@ def _print_model_connection_error(exc: BaseException, model_name: str) -> None:
     console.print()
 
 
-def _bootstrap_scan(args: argparse.Namespace) -> None:
-    """Warm up the model and prepare the run for a non-interactive scan.
+def _print_cli_error(message: str) -> None:
+    Console(stderr=True, soft_wrap=True).print(
+        f"strix: error: {message}", markup=False, highlight=False
+    )
 
-    Interactive launches skip this: the model preflight and run preparation
-    happen inside the TUI so the interface paints immediately instead of
-    waiting on a model round trip.
+
+HEADLESS_HINT = "If Strix runs without a terminal (CI, nohup, pipes), pass -n to run headless."
+
+
+def _fall_back_to_headless(args: argparse.Namespace) -> None:
+    """Run headless when there is no terminal for the TUI to attach to.
+
+    CI jobs, ``nohup``, pipes and cron have no tty; the Go TUI exits as soon
+    as it tries to take over the screen. With a target the scan can still run
+    as if ``-n`` was given. Without one the start screen is the only way to
+    enter a target, so stop with the fix instead. A bare ``--resume`` is left
+    to the picker, which already explains itself without a terminal.
+    """
+    if args.non_interactive or args.resume_picker or terminal_attached():
+        return
+    if args.needs_setup:
+        report_error("no_terminal_for_setup")
+        _print_error_panel(
+            "NO TERMINAL ATTACHED",
+            "The interactive interface needs a terminal and no target was given.\n"
+            "Pass -t <target> -n to run headless.",
+        )
+        sys.exit(1)
+    args.non_interactive = True
+    Console().print("No terminal attached, running headless (same as -n).", style="dim")
+
+
+def _pick_run_to_resume(args: argparse.Namespace) -> None:
+    """A bare --resume: let the user pick a run, then load it like --resume <name>."""
+    from strix.interface.resume_picker import PickerUnavailableError, pick_run
+    from strix.report.runs import list_run_summaries
+
+    try:
+        chosen = pick_run(list_run_summaries(), runs_dir=RUNS_DIR_NAME)
+    except PickerUnavailableError as exc:
+        _print_cli_error(resume_run_list_message(f"{exc}."))
+        sys.exit(2)
+    if chosen is None:
+        Console().print("No run selected.", style="dim")
+        sys.exit(0)
+    args.resume = chosen.run_name
+    try:
+        load_resume_state(args)
+    except ResumeError as exc:
+        _print_cli_error(str(exc))
+        sys.exit(2)
+
+
+def _bootstrap_scan(args: argparse.Namespace) -> None:
+    """Warm up the model and prepare the run before the interface starts.
+
+    Start-screen launches skip this: they verify the model and prepare the
+    run once the user has entered a target.
     """
     set_scan_phase("preflight")
     try:
-        asyncio.run(warm_up_llm(show_model_warning=True))
+        asyncio.run(warm_up_llm())
     except ModelConnectionError as exc:
         report_error("model_connection_failed", exc)
-        _print_model_connection_error(exc, exc.model_name)
+        _print_model_connection_error(exc)
         sys.exit(1)
     persist_current()
     try:
@@ -418,12 +417,22 @@ def _bootstrap_scan(args: argparse.Namespace) -> None:
     telemetry_start(args)
 
 
-def main() -> None:
-    configure_dependency_logging()
+def _force_utf8_streams() -> None:
+    for stream in (sys.stdout, sys.stderr):
+        reconfigure = getattr(stream, "reconfigure", None)
+        if reconfigure is None:
+            continue
+        with contextlib.suppress(ValueError, OSError):
+            reconfigure(encoding="utf-8")
 
+
+def main() -> None:
     if sys.platform == "win32":
+        _force_utf8_streams()
         asyncio.set_event_loop_policy(asyncio.WindowsSelectorEventLoopPolicy())
 
+    setup_console_logging()
+
     if len(sys.argv) == 2 and sys.argv[1] in ("-h", "--help"):
         try:
             parse_arguments()
@@ -462,6 +471,7 @@ def main() -> None:
     start_import_warmup()
 
     args = parse_arguments()
+    _fall_back_to_headless(args)
 
     start_background_check()
     if not args.non_interactive and prompt_update_if_available(Console()):
@@ -469,6 +479,9 @@ def main() -> None:
             restart_after_update()
         sys.exit(0)
 
+    if args.resume_picker:
+        _pick_run_to_resume(args)
+
     check_docker_installed()
     pull_docker_image()
     validate_environment()
@@ -476,7 +489,7 @@ def main() -> None:
     # Everything below imports the scan engine; do not race the warm-up thread.
     wait_for_import_warmup()
 
-    if args.non_interactive:
+    if args.non_interactive or not args.needs_setup:
         _bootstrap_scan(args)
 
     from strix.report.state import get_global_report_state
@@ -497,6 +510,11 @@ def main() -> None:
         report_error("interactive_setup_unavailable", exc)
         _print_error_panel("INTERACTIVE SETUP UNAVAILABLE", str(exc))
         sys.exit(1)
+    except InteractiveInterfaceExitedError as exc:
+        exit_reason = "error"
+        report_error("interactive_interface_exited", exc)
+        _print_error_panel("INTERACTIVE INTERFACE STOPPED", f"{exc}.\n{HEADLESS_HINT}")
+        sys.exit(1)
     except KeyboardInterrupt:
         exit_reason = "interrupted"
     except Exception as exc:
@@ -529,7 +547,7 @@ def main() -> None:
 
     if args.non_interactive:
         report_state = get_global_report_state()
-        if report_state and report_state.vulnerability_reports:
+        if report_state and findings_fail_build(report_state.vulnerability_reports, args.fail_on):
             sys.exit(2)
 
 

strix/interface/resume_picker.py

@@ -0,0 +1,409 @@
+"""Inline arrow-key picker for ``strix --resume`` with no run name.
+
+Draws a short list in the normal terminal (no alternate screen), redraws it in
+place on every key, and returns the chosen run. Keys: up/down, page up/down,
+home/end, enter, esc, typing filters the rows, backspace edits the filter.
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+from datetime import UTC, datetime
+from typing import TYPE_CHECKING, TextIO
+
+from rich.console import Console
+from rich.text import Text
+
+
+if TYPE_CHECKING:
+    from collections.abc import Callable
+
+    from strix.report.runs import RunSummary
+
+
+KEY_UP = "up"
+KEY_DOWN = "down"
+KEY_PAGE_UP = "pageup"
+KEY_PAGE_DOWN = "pagedown"
+KEY_HOME = "home"
+KEY_END = "end"
+KEY_ENTER = "enter"
+KEY_ESCAPE = "escape"
+KEY_BACKSPACE = "backspace"
+KEY_INTERRUPT = "interrupt"
+
+_CSI_KEYS = {
+    "A": KEY_UP,
+    "B": KEY_DOWN,
+    "H": KEY_HOME,
+    "F": KEY_END,
+    "1~": KEY_HOME,
+    "4~": KEY_END,
+    "5~": KEY_PAGE_UP,
+    "6~": KEY_PAGE_DOWN,
+    "7~": KEY_HOME,
+    "8~": KEY_END,
+}
+_WINDOWS_KEYS = {
+    "H": KEY_UP,
+    "P": KEY_DOWN,
+    "I": KEY_PAGE_UP,
+    "Q": KEY_PAGE_DOWN,
+    "G": KEY_HOME,
+    "O": KEY_END,
+}
+_CONTROL_KEYS = {
+    "\r": KEY_ENTER,
+    "\n": KEY_ENTER,
+    "\x1b": KEY_ESCAPE,
+    "\x7f": KEY_BACKSPACE,
+    "\x08": KEY_BACKSPACE,
+    "\x03": KEY_INTERRUPT,
+}
+
+_CURSOR_HIDE = "\x1b[?25l"
+_CURSOR_SHOW = "\x1b[?25h"
+_CLEAR_BELOW = "\x1b[J"
+
+_GREEN = "#22c55e"
+_AMBER = "#f59e0b"
+_STATUS_STYLES = {
+    "completed": _GREEN,
+    "running": _GREEN,
+    "interrupted": _AMBER,
+    "stopped": _AMBER,
+    "failed": "#ef4444",
+}
+_NO_STATE = "no state"
+_CURSOR = " \u276f "
+_STARTED_WIDTH = 14
+_FINDINGS_WIDTH = 8
+_MIN_TARGET_WIDTH = 12
+_MAX_RUN_WIDTH = 40
+_CHROME_LINES = 8
+_MAX_VISIBLE = 8
+
+
+def _utf8_length(lead: bytes) -> int:
+    byte = lead[0] if lead else 0
+    if byte >= 0xF0:
+        return 4
+    if byte >= 0xE0:
+        return 3
+    if byte >= 0xC0:
+        return 2
+    return 1
+
+
+class PickerUnavailableError(RuntimeError):
+    """The picker needs an interactive terminal on stdin and stdout."""
+
+
+if sys.platform == "win32":
+    import msvcrt
+
+    def read_raw_key(_stream: TextIO) -> str:
+        char = str(msvcrt.getwch())
+        if char in ("\x00", "\xe0"):
+            return _WINDOWS_KEYS.get(str(msvcrt.getwch()), "")
+        return char
+
+else:
+    import select
+    import termios
+    import tty
+
+    def read_raw_key(stream: TextIO) -> str:
+        fd = stream.fileno()
+        saved = termios.tcgetattr(fd)
+        try:
+            tty.setcbreak(fd, termios.TCSANOW)
+            first = os.read(fd, 1)
+            char = (first + os.read(fd, _utf8_length(first) - 1)).decode("utf-8", "replace")
+            if char != "\x1b":
+                return char
+            sequence = ""
+            while select.select([fd], [], [], 0.05)[0]:
+                sequence += os.read(fd, 1).decode("utf-8", "replace")
+                if (sequence.startswith("[") and sequence[-1].isalpha()) or sequence.endswith("~"):
+                    break
+        finally:
+            termios.tcsetattr(fd, termios.TCSADRAIN, saved)
+        if not sequence:
+            return "\x1b"
+        return _CSI_KEYS.get(sequence[1:], "") if sequence[0] in "[O" else ""
+
+
+def read_key(stream: TextIO) -> str:
+    return translate_key(read_raw_key(stream))
+
+
+def translate_key(raw: str) -> str:
+    return _CONTROL_KEYS.get(raw, raw)
+
+
+def relative_time(stamp: str, now: datetime | None = None) -> str:
+    try:
+        started = datetime.fromisoformat(stamp)
+    except ValueError:
+        return stamp[:16] if stamp else "unknown"
+    if started.tzinfo is None:
+        started = started.replace(tzinfo=UTC)
+    now = now or datetime.now(UTC)
+    seconds = (now - started).total_seconds()
+    for limit, unit, label in (
+        (60, 1, ""),
+        (3600, 60, "min"),
+        (86400, 3600, "h"),
+        (7 * 86400, 86400, "d"),
+    ):
+        if seconds < limit:
+            return "just now" if not label else f"{int(seconds // unit)} {label} ago"
+    local = started.astimezone()
+    if local.year == now.astimezone().year:
+        return f"{local:%b} {local.day}, {local:%H:%M}"
+    return f"{local:%b} {local.day}, {local.year}"
+
+
+def filter_runs(runs: list[RunSummary], needle: str) -> list[RunSummary]:
+    needle = needle.strip().lower()
+    if not needle:
+        return list(runs)
+    return [
+        run
+        for run in runs
+        if needle in run.run_name.lower()
+        or needle in run.target.lower()
+        or needle in run.status.lower()
+    ]
+
+
+def _fit(text: str, width: int) -> str:
+    if len(text) <= width:
+        return text.ljust(width)
+    return text[: width - 1] + "\u2026"
+
+
+def _status_text(run: RunSummary) -> str:
+    return run.status if run.resumable else f"{run.status} \u00b7 {_NO_STATE}"
+
+
+class ResumePicker:
+    def __init__(
+        self,
+        runs: list[RunSummary],
+        *,
+        console: Console,
+        runs_dir: str,
+        now: datetime | None = None,
+    ) -> None:
+        self.runs = runs
+        self.console = console
+        self.runs_dir = runs_dir
+        self.now = now
+        self.filter = ""
+        self.cursor = 0
+        self.offset = 0
+        self.notice = ""
+        self._drawn = 0
+
+    @property
+    def rows(self) -> list[RunSummary]:
+        return filter_runs(self.runs, self.filter)
+
+    def _visible(self) -> int:
+        return max(3, min(len(self.rows), _MAX_VISIBLE, self.console.height - _CHROME_LINES))
+
+    def _columns(self) -> tuple[int, int, int]:
+        width = max(40, self.console.width - 1)
+        status_width = max(len("status"), *(len(_status_text(run)) for run in self.runs))
+        run_width = min(_MAX_RUN_WIDTH, max(len("run"), *(len(run.run_name) for run in self.runs)))
+        fixed = len(_CURSOR) + _STARTED_WIDTH + _FINDINGS_WIDTH + status_width + 4 * 2
+        target_width = width - fixed - run_width
+        if target_width < _MIN_TARGET_WIDTH:
+            run_width = max(8, run_width + target_width - _MIN_TARGET_WIDTH)
+            target_width = width - fixed - run_width
+        return max(_MIN_TARGET_WIDTH, target_width), status_width, run_width
+
+    def _scroll(self) -> range:
+        rows = self.rows
+        visible = self._visible()
+        self.cursor = max(0, min(self.cursor, len(rows) - 1))
+        if self.cursor < self.offset:
+            self.offset = self.cursor
+        elif self.cursor >= self.offset + visible:
+            self.offset = self.cursor - visible + 1
+        self.offset = max(0, min(self.offset, max(0, len(rows) - visible)))
+        return range(self.offset, min(len(rows), self.offset + visible))
+
+    def render(self) -> list[Text]:
+        rows = self.rows
+        widths = self._columns()
+        window = self._scroll()
+
+        title = Text()
+        title.append(" Resume a run", style="bold")
+        title.append(f"   {len(self.runs)} runs in ./{self.runs_dir}", style="dim")
+        if self.filter:
+            title.append("   search: ", style="dim")
+            title.append(self.filter)
+        header = Text(
+            " " * len(_CURSOR)
+            + self._cells("started", "target", "status", "findings", "run", widths),
+            style="dim",
+        )
+        lines = [title, Text(), header]
+        if not rows:
+            lines.append(Text(f"   no runs match {self.filter!r}", style="dim"))
+        if window.start:
+            lines.append(Text(f"   \u2026 {window.start} more above", style="dim"))
+        lines.extend(self._row(rows[index], index == self.cursor, widths) for index in window)
+        if window.stop < len(rows):
+            lines.append(Text(f"   \u2026 {len(rows) - window.stop} more below", style="dim"))
+        footer = Text(" ")
+        if self.notice:
+            footer.append(self.notice, style=_AMBER)
+        else:
+            footer.append(
+                "\u2191\u2193 move   enter resume   type to search   esc cancel", style="dim"
+            )
+        lines.extend([Text(), footer])
+        for line in lines:
+            line.truncate(self.console.width - 1)
+        return lines
+
+    @staticmethod
+    def _cells(
+        started: str,
+        target: str,
+        status: str,
+        findings: str,
+        run: str,
+        widths: tuple[int, int, int],
+    ) -> str:
+        target_width, status_width, run_width = widths
+        return "  ".join(
+            [
+                _fit(started, _STARTED_WIDTH),
+                _fit(target, target_width),
+                _fit(status, status_width),
+                _fit(findings, _FINDINGS_WIDTH),
+                _fit(run, run_width),
+            ]
+        )
+
+    def _row(self, run: RunSummary, selected: bool, widths: tuple[int, int, int]) -> Text:
+        target_width, status_width, run_width = widths
+        primary = "bold" if selected else ""
+        muted = "" if selected else "dim"
+        status_style = _STATUS_STYLES.get(run.status, "")
+        if not run.resumable:
+            primary = muted = status_style = "dim"
+        line = Text()
+        line.append(_CURSOR if selected else " " * len(_CURSOR), style=_GREEN)
+        line.append(_fit(relative_time(run.started_at, self.now), _STARTED_WIDTH), style=muted)
+        line.append("  ")
+        line.append(_fit(run.target, target_width), style=primary)
+        line.append("  ")
+        line.append(_fit(_status_text(run), status_width), style=status_style)
+        line.append("  ")
+        line.append(_fit(str(run.findings), _FINDINGS_WIDTH), style=muted)
+        line.append("  ")
+        line.append(_fit(run.run_name, run_width), style=muted)
+        return line
+
+    def draw(self) -> None:
+        self.clear()
+        lines = self.render()
+        for line in lines:
+            self.console.print(line, soft_wrap=True, overflow="crop", end="\n")
+        self.console.file.flush()
+        self._drawn = len(lines)
+
+    def clear(self) -> None:
+        if self._drawn:
+            self.console.file.write(f"\x1b[{self._drawn}A\r{_CLEAR_BELOW}")
+            self.console.file.flush()
+            self._drawn = 0
+
+    def handle(self, key: str) -> tuple[bool, RunSummary | None]:
+        """Apply one key: (done, run), where done with no run means cancelled."""
+        self.notice = ""
+        if key == KEY_INTERRUPT:
+            return True, None
+        if key == KEY_ESCAPE:
+            if not self.filter:
+                return True, None
+            self.filter = ""
+            self.cursor = 0
+            return False, None
+        if key == KEY_ENTER:
+            return self._choose()
+        if key == KEY_BACKSPACE:
+            self.filter = self.filter[:-1]
+            self.cursor = 0
+        elif len(key) == 1 and key.isprintable():
+            self.filter += key
+            self.cursor = 0
+        else:
+            self._move(key)
+        self.cursor = max(0, min(self.cursor, max(0, len(self.rows) - 1)))
+        return False, None
+
+    def _choose(self) -> tuple[bool, RunSummary | None]:
+        rows = self.rows
+        if not rows:
+            return False, None
+        run = rows[self.cursor]
+        if not run.resumable:
+            self.notice = f"{run.run_name} has no saved agent state to resume from"
+            return False, None
+        return True, run
+
+    def _move(self, key: str) -> None:
+        steps = {
+            KEY_UP: -1,
+            KEY_DOWN: 1,
+            KEY_PAGE_UP: -self._visible(),
+            KEY_PAGE_DOWN: self._visible(),
+        }
+        if key in steps:
+            self.cursor += steps[key]
+        elif key == KEY_HOME:
+            self.cursor = 0
+        elif key == KEY_END:
+            self.cursor = len(self.rows) - 1
+
+    def run(self, next_key: Callable[[], str]) -> RunSummary | None:
+        self.console.file.write(_CURSOR_HIDE)
+        try:
+            self.draw()
+            while True:
+                done, chosen = self.handle(next_key())
+                if done:
+                    return chosen
+                self.draw()
+        except KeyboardInterrupt:
+            return None
+        finally:
+            self.clear()
+            self.console.file.write(_CURSOR_SHOW)
+            self.console.file.flush()
+
+
+def pick_run(
+    runs: list[RunSummary],
+    *,
+    runs_dir: str,
+    console: Console | None = None,
+    stdin: TextIO | None = None,
+) -> RunSummary | None:
+    stdin = stdin or sys.stdin
+    console = console or Console()
+    if not (hasattr(stdin, "isatty") and stdin.isatty() and console.is_terminal):
+        raise PickerUnavailableError(
+            "--resume needs a run name when there is no interactive terminal"
+        )
+    picker = ResumePicker(runs, console=console, runs_dir=runs_dir)
+    return picker.run(lambda: read_key(stdin))

strix/interface/scan_setup.py

@@ -11,6 +11,7 @@ from __future__ import annotations
 
 import asyncio
 import logging
+import unicodedata
 from datetime import UTC, datetime
 from typing import TYPE_CHECKING, Any
 
@@ -45,6 +46,8 @@ from strix.utils.api_spec import (
 if TYPE_CHECKING:
     import argparse
 
+    from agents.models.interface import Model
+
 logger = logging.getLogger(__name__)
 
 HOST_GATEWAY_HOSTNAME = "host.docker.internal"
@@ -58,43 +61,112 @@ class ModelConnectionError(RuntimeError):
         self.model_name = model_name
 
 
+def _first_non_ascii(value: str) -> tuple[int, str] | None:
+    for position, char in enumerate(value):
+        if ord(char) > 0x7F:
+            return position, char
+    return None
+
+
+def _header_candidates(
+    prefix: str, model: str | None, api_key: str | None, extra_headers: dict[str, str] | None
+) -> list[tuple[str, str]]:
+    candidates: list[tuple[str, str]] = []
+    if api_key and not codex.subscription_model(model):
+        candidates.append((f"{prefix}LLM_API_KEY", api_key))
+    for header, value in (extra_headers or {}).items():
+        candidates.append((f"{prefix}LLM_EXTRA_HEADERS header name {header!r}", header))
+        candidates.append((f"{prefix}LLM_EXTRA_HEADERS value for {header!r}", value))
+    return candidates
+
+
+def check_header_safe_credentials(settings: Settings) -> None:
+    llm = settings.llm
+    dedupe = settings.dedupe
+    candidates = _header_candidates("", llm.model, llm.api_key, llm.extra_headers)
+    if dedupe.model:
+        candidates += _header_candidates(
+            "DEDUPE_", dedupe.model, (dedupe.api_key or "").strip(), dedupe.extra_headers
+        )
+    for setting, value in candidates:
+        found = _first_non_ascii(value)
+        if found is None:
+            continue
+        position, char = found
+        raise ValueError(
+            f"{setting} contains a character that cannot be sent in an HTTP header: "
+            f"U+{ord(char):04X} ({unicodedata.name(char, 'unnamed character')}) "
+            f"at position {position + 1} of {len(value)}. Re-enter the value without it."
+        )
+
+
 async def preflight_model_connection(
     model_name: str,
     *,
     settings: Settings | None = None,
 ) -> None:
     """Verify the configured model route before starting a scan."""
-    from agents.models.interface import ModelTracing
-
     from strix.config.models import StrixProvider, configure_sdk_model_defaults
-    from strix.core.inputs import make_model_settings
 
     resolved_settings = load_settings() if settings is None else settings
+    check_header_safe_credentials(resolved_settings)
     configure_sdk_model_defaults(resolved_settings)
     model = StrixProvider().get_model(model_name)
+    await preflight_request(
+        model,
+        model_name=model_name,
+        extra_headers=resolved_settings.llm.extra_headers,
+        timeout=resolved_settings.llm.preflight_timeout,
+        api_base_setting="LLM_API_BASE",
+    )
+
+
+async def preflight_request(
+    model: Model,
+    *,
+    model_name: str,
+    extra_headers: dict[str, str] | None,
+    timeout: int,
+    api_base_setting: str,
+) -> None:
+    """Send one tiny request to ``model`` and fail if it does not answer in ``timeout`` seconds.
+
+    ``api_base_setting`` names the environment variable that points at this
+    model's endpoint, so the timeout message sends the user to the right one.
+    """
+    from agents.models.interface import ModelTracing
+
+    from strix.core.inputs import make_model_settings
+
     request_settings = make_model_settings(
         None,
         model_name=model_name,
-        request_timeout=resolved_settings.llm.timeout,
+        request_timeout=timeout,
         prompt_cache=False,
-        extra_headers=resolved_settings.llm.extra_headers,
+        extra_headers=extra_headers,
         has_tools=False,
     )
-    await asyncio.wait_for(
-        model.get_response(
-            system_instructions="You are a helpful assistant.",
-            input="Reply with just 'OK'.",
-            model_settings=request_settings,
-            tools=[],
-            output_schema=None,
-            handoffs=[],
-            tracing=ModelTracing.DISABLED,
-            previous_response_id=None,
-            conversation_id=None,
-            prompt=None,
-        ),
-        timeout=resolved_settings.llm.timeout,
-    )
+    try:
+        await asyncio.wait_for(
+            model.get_response(
+                system_instructions="You are a helpful assistant.",
+                input="Reply with just 'OK'.",
+                model_settings=request_settings,
+                tools=[],
+                output_schema=None,
+                handoffs=[],
+                tracing=ModelTracing.DISABLED,
+                previous_response_id=None,
+                conversation_id=None,
+                prompt=None,
+            ),
+            timeout=timeout,
+        )
+    except TimeoutError:
+        raise TimeoutError(
+            f"{model_name} did not answer within {timeout}s (LLM_PREFLIGHT_TIMEOUT). "
+            f"Check {api_base_setting} and that the endpoint is reachable."
+        ) from None
 
 
 def build_targets_info(args: argparse.Namespace) -> None:

strix/interface/tui/backend/controller.py

@@ -11,7 +11,6 @@ from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
 from strix.config import load_settings
-from strix.config.models import is_recommended_or_frontier_model
 from strix.config.settings import DEFAULT_MAX_TURNS
 from strix.interface.tui.backend.live_view import TuiLiveView
 from strix.interface.tui.backend.projection import (
@@ -153,11 +152,6 @@ class TuiController:
         ]
         self.notify_changed()
 
-    def begin_preparation(self) -> None:
-        """Mark a directly-launched run as preparing behind the live TUI."""
-        self.scan_state = "preparing"
-        self.notify_changed()
-
     def fail_preparation(self, detail: str) -> None:
         self.scan_state = "failed"
         self.error = detail
@@ -189,11 +183,6 @@ class TuiController:
         subscription = False
         with contextlib.suppress(Exception):
             subscription = is_subscription_run(self.report_state)
-        model_warning = ""
-        if model and not is_recommended_or_frontier_model(model):
-            model_warning = (
-                f"{model} is not a recommended frontier model. Pentest quality could be degraded."
-            )
         state = {
             "setup_mode": self.setup_mode,
             "scan_started": self.scan_started,
@@ -211,7 +200,6 @@ class TuiController:
             "scope_mode": self.scope_mode,
             "diff_base": terminal_projection(self.diff_base, max_string=256),
             "model": terminal_projection(model, max_string=256),
-            "model_warning": terminal_projection(model_warning, max_string=512),
             "caido_url": terminal_projection(
                 getattr(self.report_state, "caido_url", None), max_string=1024
             ),

strix/interface/tui/backend/projection.py

@@ -150,7 +150,6 @@ def bounded_state_projection(state: dict[str, Any]) -> dict[str, Any]:
         key: state["usage"][key] for key in ("total_tokens", "cost") if key in state["usage"]
     }
     state["error"] = terminal_projection(state["error"], max_string=512)
-    state["model_warning"] = terminal_projection(state["model_warning"], max_string=256)
     state["caido_url"] = terminal_projection(state["caido_url"], max_string=256)
     state["viewer_url"] = terminal_projection(state["viewer_url"], max_string=256)
     if encoded_size(state) <= STATE_TARGET_BYTES:
@@ -173,7 +172,6 @@ def bounded_state_projection(state: dict[str, Any]) -> dict[str, Any]:
         "scope_mode": state["scope_mode"],
         "diff_base": state["diff_base"],
         "model": state["model"],
-        "model_warning": "",
         "caido_url": None,
         "messages": [],
         "usage": state["usage"],

strix/interface/tui/internal/app/agents.go

@@ -209,8 +209,7 @@ func (m *Model) ensureAgentVisible() {
 		m.agentOffset = 0
 		return
 	}
-	_, _, _, agentHeight := m.sidebarHeights()
-	rows := max(1, agentHeight-4)
+	rows := m.agentPageSize()
 	row := selectedAgentRow(entries, m.selectedAgent)
 	if row < m.agentOffset {
 		m.agentOffset = row

strix/interface/tui/internal/app/mcp_test.go

@@ -27,11 +27,14 @@ func mcpModel(t *testing.T) Model {
 func TestMcpPanelShowsHealthyAndOffline(t *testing.T) {
 	m := mcpModel(t)
 	out := ansi.Strip(m.mcpConnectionsView(40, 6))
-	for _, want := range []string{"MCP Connections (2)", "supabase", "3 tools", "vercel", "offline"} {
+	for _, want := range []string{"supabase", "3 tools", "vercel", "offline"} {
 		if !strings.Contains(out, want) {
 			t.Fatalf("panel missing %q:\n%s", want, out)
 		}
 	}
+	if sidebar := ansi.Strip(m.sidebarView(26, 40)); !strings.Contains(sidebar, "MCP (2)") {
+		t.Fatalf("sidebar header did not carry the connection count:\n%s", sidebar)
+	}
 }
 
 // A roster longer than the panel height shows a window of rows rather than every
@@ -46,9 +49,8 @@ func TestMcpPanelWindowsLargeRosterAndCountsAll(t *testing.T) {
 	}
 	m.snapshot.Connections = conns
 
-	// rows = 6 → one header line + five roster rows.
 	out := ansi.Strip(m.mcpConnectionsView(40, 6))
-	if !strings.Contains(out, "MCP Connections (12)") {
+	if !strings.Contains(ansi.Strip(m.sidebarView(26, 40)), "MCP (12)") {
 		t.Fatalf("header did not carry the full connection count:\n%s", out)
 	}
 	if !strings.Contains(out, "conn-00") {
@@ -58,14 +60,13 @@ func TestMcpPanelWindowsLargeRosterAndCountsAll(t *testing.T) {
 		t.Fatalf("a roster past the panel height should be windowed, not fully drawn:\n%s", out)
 	}
 	if got := strings.Count(out, "\n") + 1; got != 6 {
-		t.Fatalf("panel rendered %d lines, want 6 (header + five rows)", got)
+		t.Fatalf("panel rendered %d lines, want 6", got)
 	}
 
-	// Scrolling the roster brings the tail into view while the header count holds.
-	m.mcpOffset = 7
+	m.mcpOffset = 6
 	scrolled := ansi.Strip(m.mcpConnectionsView(40, 6))
-	if !strings.Contains(scrolled, "conn-11") || !strings.Contains(scrolled, "MCP Connections (12)") {
-		t.Fatalf("scrolled window did not reveal the tail with the count intact:\n%s", scrolled)
+	if !strings.Contains(scrolled, "conn-11") {
+		t.Fatalf("scrolled window did not reveal the tail:\n%s", scrolled)
 	}
 }
 

strix/interface/tui/internal/app/model.go

@@ -100,6 +100,9 @@ type Model struct {
 	filtered               []string
 	cursor                 int
 	collapsedAgents        map[string]bool
+	collapsedPanels        map[sidebarPanel]bool
+	zoomedPanel            sidebarPanel
+	sidebarHidden          bool
 	expandedEvents         map[string]bool
 	blockCache             map[string]renderedBlock
 	eventSpans             []eventSpan
@@ -278,7 +281,7 @@ func New(client *Client) Model {
 	input.Focus()
 	return Model{
 		client: client, input: input, viewport: viewport.New(80, 20), vulnViewport: viewport.New(80, 20),
-		collapsedAgents: map[string]bool{}, expandedEvents: map[string]bool{}, blockCache: map[string]renderedBlock{}, showSplash: true, splashStarted: time.Now(), followOutput: true,
+		collapsedAgents: map[string]bool{}, collapsedPanels: map[sidebarPanel]bool{}, zoomedPanel: panelNone, expandedEvents: map[string]bool{}, blockCache: map[string]renderedBlock{}, showSplash: true, splashStarted: time.Now(), followOutput: true,
 		collectionRevisions: map[string]int{}, collectionAssemblies: map[string]*collectionAssembly{}, resyncRequested: map[string]bool{}, resyncRequests: map[string]string{},
 		seenMessages: map[string]bool{},
 	}
@@ -384,7 +387,16 @@ func (m Model) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 			m.vulnerabilityCopyError = msg.err.Error()
 		}
 		return m, nil
+	case tea.ResumeMsg:
+		// Suspend turns mouse tracking off with the rest of the terminal state,
+		// but the restore brings back only the alt screen, so turn it back on.
+		return m, tea.EnableMouseCellMotion
 	case tea.KeyMsg:
+		// Raw mode clears ISIG, so ctrl+z arrives as a key instead of SIGTSTP.
+		// Suspend on every screen, the way a shell job would.
+		if msg.Type == tea.KeyCtrlZ {
+			return m, tea.Suspend
+		}
 		if m.showSplash {
 			switch msg.String() {
 			case "ctrl+c", "ctrl+q", "q", "esc":

strix/interface/tui/internal/app/model_test.go

@@ -370,17 +370,6 @@ func TestStartedSnapshotTransitionsToLiveView(t *testing.T) {
 	}
 }
 
-func TestSplashModelWarningRendersTheBackendSentenceOnce(t *testing.T) {
-	warning := "openai/glm-5.3 is not a recommended frontier model. Pentest quality could be degraded."
-	got := ansi.Strip(splashModelWarning("openai/glm-5.3", warning))
-	if got != "⚠ "+warning {
-		t.Fatalf("splash warning = %q, want %q", got, "⚠ "+warning)
-	}
-	if got := ansi.Strip(splashModelWarning("other/model", warning)); got != "⚠ "+warning {
-		t.Fatalf("splash warning with unrelated model = %q", got)
-	}
-}
-
 func TestSetupStartScreenFitsNarrowTerminal(t *testing.T) {
 	model := New(nil)
 	model.width, model.height = 40, 18
@@ -901,6 +890,14 @@ func TestRunningViewerShowsCompleteWrappedURL(t *testing.T) {
 	if want := strings.Count(model.viewerView(model.viewerContentWidth()), "\n") + 3; model.viewerHeight() != want {
 		t.Fatalf("viewer height = %d, want %d", model.viewerHeight(), want)
 	}
+
+	raw := model.viewerView(18)
+	if got, want := strings.Count(raw, ansi.SetHyperlink(url)), len(urlLines); got != want {
+		t.Fatalf("every wrapped line should link to the full URL: got %d links for %d lines in %q", got, want, raw)
+	}
+	if got := strings.Count(raw, ansi.ResetHyperlink()); got != len(urlLines) {
+		t.Fatalf("hyperlinks not closed: %d resets for %d lines", got, len(urlLines))
+	}
 }
 
 func TestVerticalScrollbarThumbTracksScrollOffset(t *testing.T) {
@@ -962,6 +959,17 @@ func TestFillBackgroundRestoresBaseForegroundAfterReset(t *testing.T) {
 	}
 }
 
+func TestFillBackgroundRestoresBaseColorsAfterBareReset(t *testing.T) {
+	view := "\x1b[38;2;115;115;115mModel\x1b[m   padding\x1b[0m"
+	filled := fillBackground(view)
+	if !strings.Contains(filled, "\x1b[m"+baseFrameColors+"   padding") {
+		t.Fatalf("base colors not restored after bare reset: %q", filled)
+	}
+	if strings.Contains(filled, "\x1b[m   ") {
+		t.Fatalf("cells after a bare reset show the terminal background: %q", filled)
+	}
+}
+
 func TestMainTraceTreeAndFindingsRenderScrollbars(t *testing.T) {
 	model := New(nil)
 	model.width, model.height = 150, 35
@@ -1439,3 +1447,31 @@ func TestNarrowTerminalKeepsTheFrameIntact(t *testing.T) {
 		}
 	}
 }
+
+func TestCtrlZSuspendsFromEveryScreen(t *testing.T) {
+	for name, prepare := range map[string]func(*Model){
+		"splash": func(m *Model) { m.showSplash = true },
+		"modal":  func(m *Model) { m.showSplash = false; m.openModal(modalHelp) },
+		"main":   func(m *Model) { m.showSplash = false },
+	} {
+		model := New(nil)
+		prepare(&model)
+		_, cmd := model.Update(tea.KeyMsg{Type: tea.KeyCtrlZ})
+		if cmd == nil {
+			t.Fatalf("%s: ctrl+z returned no command", name)
+		}
+		if _, ok := cmd().(tea.SuspendMsg); !ok {
+			t.Fatalf("%s: ctrl+z did not suspend", name)
+		}
+	}
+}
+
+func TestResumeReenablesMouse(t *testing.T) {
+	_, cmd := New(nil).Update(tea.ResumeMsg{})
+	if cmd == nil {
+		t.Fatal("resume returned no command")
+	}
+	if msg := cmd(); msg != tea.EnableMouseCellMotion() {
+		t.Fatalf("resume did not re-enable mouse tracking: %#v", msg)
+	}
+}

strix/interface/tui/internal/app/panels.go

@@ -0,0 +1,311 @@
+package app
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/charmbracelet/lipgloss"
+)
+
+type sidebarPanel int
+
+const (
+	panelNone sidebarPanel = iota - 1
+	panelAgents
+	panelFindings
+	panelMcp
+	panelStats
+)
+
+const (
+	panelOpenGlyph      = "▾"
+	panelCollapsedGlyph = "▸"
+	panelZoomGlyph      = "⤢"
+	panelUnzoomGlyph    = "⤡"
+	sidebarHideGlyph    = "»"
+	sidebarShowGlyph    = "«"
+	panelGlyphZone      = 3
+)
+
+type panelRect struct {
+	panel  sidebarPanel
+	top    int
+	height int
+}
+
+func (m Model) panelControls() bool {
+	return len(m.snapshot.Vulnerabilities) > 0 || len(m.snapshot.Connections) > 0
+}
+
+func (m Model) panelShrunk(panel sidebarPanel) bool {
+	if panel == panelStats || !m.panelControls() {
+		return false
+	}
+	return m.collapsedPanels[panel] || (m.zoomedPanel != panelNone && m.zoomedPanel != panel)
+}
+
+func (m Model) panelHeight(panel sidebarPanel, open int) int {
+	if m.panelShrunk(panel) {
+		return 1
+	}
+	return open
+}
+
+func panelFocus(panel sidebarPanel) (focusMode, bool) {
+	switch panel {
+	case panelAgents:
+		return focusAgents, true
+	case panelFindings:
+		return focusVulnerabilities, true
+	case panelMcp:
+		return focusMcp, true
+	default:
+		return focusInput, false
+	}
+}
+
+func (m Model) sidebarPanels() []panelRect {
+	statsHeight, vulnHeight, mcpHeight, agentHeight := m.sidebarHeights()
+	top := m.viewerHeight()
+	rects := []panelRect{{panelAgents, top, agentHeight}}
+	top += agentHeight
+	if vulnHeight > 0 {
+		rects = append(rects, panelRect{panelFindings, top, vulnHeight})
+		top += vulnHeight
+	}
+	if mcpHeight > 0 {
+		rects = append(rects, panelRect{panelMcp, top, mcpHeight})
+		top += mcpHeight
+	}
+	top += m.sidebarGap()
+	return append(rects, panelRect{panelStats, top, statsHeight})
+}
+
+func (m Model) sidebarGap() int {
+	statsHeight, vulnHeight, mcpHeight, agentHeight := m.sidebarHeights()
+	return max(0, m.height-m.viewerHeight()-agentHeight-vulnHeight-mcpHeight-statsHeight)
+}
+
+func (m Model) panelAt(y int) (panelRect, bool) {
+	for _, rect := range m.sidebarPanels() {
+		if y >= rect.top && y < rect.top+rect.height {
+			return rect, true
+		}
+	}
+	return panelRect{}, false
+}
+
+func (m Model) panelTop(panel sidebarPanel) int {
+	for _, rect := range m.sidebarPanels() {
+		if rect.panel == panel {
+			return rect.top
+		}
+	}
+	return 0
+}
+
+func (m Model) panelTitle(panel sidebarPanel) string {
+	switch panel {
+	case panelAgents:
+		return fmt.Sprintf("Agents (%d)", len(m.snapshot.Agents))
+	case panelFindings:
+		return fmt.Sprintf("Findings (%d)", len(m.snapshot.Vulnerabilities))
+	case panelMcp:
+		return fmt.Sprintf("MCP (%d)", len(m.snapshot.Connections))
+	default:
+		return m.snapshot.Model
+	}
+}
+
+func (m Model) panelHeader(panel sidebarPanel, width int) string {
+	style := lipgloss.NewStyle().Foreground(dim)
+	if !m.panelControls() {
+		return truncate(style.Render(m.panelTitle(panel)), max(1, width))
+	}
+	glyph := panelZoomGlyph
+	if m.zoomedPanel == panel {
+		glyph = panelUnzoomGlyph
+	}
+	label := truncate(style.Render(panelOpenGlyph+" "+m.panelTitle(panel)), max(1, width-2))
+	gap := max(1, width-lipgloss.Width(label)-1)
+	return label + strings.Repeat(" ", gap) + style.Render(glyph)
+}
+
+func (m Model) collapsedPanelRow(panel sidebarPanel, width int) string {
+	title := m.panelTitle(panel)
+	if panel != panelStats {
+		title = panelCollapsedGlyph + " " + title
+	}
+	label := lipgloss.NewStyle().Foreground(dim).Render(title)
+	return " " + truncate(label, max(1, width-1))
+}
+
+func (m Model) panelBox(panel sidebarPanel, body string, width, height int, focused bool) string {
+	if height <= 1 {
+		return m.collapsedPanelRow(panel, width)
+	}
+	border := dark
+	if focused {
+		border = green
+	}
+	content := body
+	if panel != panelStats {
+		content = m.panelHeader(panel, width-4)
+		if body != "" {
+			content += "\n\n" + body
+		}
+	}
+	return lipgloss.NewStyle().Width(width-2).Height(height-2).Border(lipgloss.RoundedBorder()).
+		BorderForeground(border).Padding(0, 1).Render(content)
+}
+
+const (
+	toggleButtonWidth = 3
+	railButtonWidth   = 5
+	railButtonHeight  = 3
+	sidebarRailWidth  = railButtonWidth + 1
+)
+
+var toggleButtonFill = lipgloss.Color("#262626")
+
+func sidebarToggleButton(glyph string) string {
+	return lipgloss.NewStyle().
+		Background(toggleButtonFill).
+		Foreground(brightWhite).
+		Bold(true).
+		Width(toggleButtonWidth).
+		Align(lipgloss.Center).
+		Render(glyph)
+}
+
+func (m Model) railVisible() bool {
+	return m.sidebarHidden && m.width >= 120
+}
+
+func (m Model) sidebarRail(height int) string {
+	button := lipgloss.NewStyle().
+		Background(toggleButtonFill).
+		Foreground(brightWhite).
+		Bold(true).
+		Width(railButtonWidth).
+		Height(railButtonHeight).
+		Align(lipgloss.Center, lipgloss.Center).
+		Render(sidebarShowGlyph)
+	return lipgloss.NewStyle().
+		Width(sidebarRailWidth).
+		Height(height).
+		Align(lipgloss.Right).
+		Render(button)
+}
+
+func (m Model) toggleButtonHit(x, y int) bool {
+	if m.railVisible() {
+		return y < railButtonHeight && x >= m.width-railButtonWidth
+	}
+	return y == 1 && x >= m.width-2-toggleButtonWidth && x < m.width-2
+}
+
+func (m Model) viewerBox(width int) string {
+	textWidth := max(1, width-toggleButtonWidth-1)
+	text := m.viewerView(textWidth)
+	rows := strings.Count(text, "\n") + 1
+	return lipgloss.JoinHorizontal(
+		lipgloss.Top,
+		fixedPanelBody(text, textWidth, rows),
+		" ",
+		sidebarToggleButton(sidebarHideGlyph),
+	)
+}
+
+func (m *Model) toggleSidebar() {
+	m.sidebarHidden = !m.sidebarHidden
+	m.resizeViewport()
+	m.panelsChanged()
+}
+
+func (m *Model) togglePanelCollapsed(panel sidebarPanel) {
+	if m.collapsedPanels[panel] {
+		delete(m.collapsedPanels, panel)
+	} else {
+		m.collapsedPanels[panel] = true
+		if m.zoomedPanel == panel {
+			m.zoomedPanel = panelNone
+		}
+	}
+	m.panelsChanged()
+}
+
+func (m *Model) togglePanelZoom(panel sidebarPanel) {
+	if m.zoomedPanel == panel {
+		m.zoomedPanel = panelNone
+	} else {
+		m.zoomedPanel = panel
+		delete(m.collapsedPanels, panel)
+	}
+	m.panelsChanged()
+}
+
+func (m *Model) revealPanel(panel sidebarPanel) {
+	delete(m.collapsedPanels, panel)
+	m.zoomedPanel = panelNone
+	m.panelsChanged()
+}
+
+func (m *Model) clickPanel(rect panelRect, x, y int) bool {
+	switch {
+	case rect.panel == panelStats || !m.panelControls():
+		return false
+	case rect.height <= 1 && !m.panelShrunk(rect.panel):
+		m.togglePanelZoom(rect.panel)
+	case rect.height <= 1:
+		m.revealPanel(rect.panel)
+	case y-rect.top != 1:
+		return false
+	case x >= m.width-2-panelGlyphZone:
+		m.togglePanelZoom(rect.panel)
+	default:
+		m.togglePanelCollapsed(rect.panel)
+	}
+	if focus, ok := panelFocus(rect.panel); ok && !m.panelShrunk(rect.panel) {
+		m.focus = focus
+		m.input.Blur()
+	}
+	return true
+}
+
+func (m *Model) panelsChanged() {
+	showSidebar, _, _, _ := m.layout()
+	for _, rect := range m.sidebarPanels() {
+		if focus, ok := panelFocus(rect.panel); ok && m.focus == focus && (!showSidebar || rect.height <= 1) {
+			m.focus = focusInput
+			m.input.Focus()
+		}
+	}
+	m.ensureAgentVisible()
+	totalRows, _ := m.vulnerabilityScrollRows()
+	m.vulnOffset = clampVulnerabilityOffset(m.vulnOffset, totalRows, m.vulnerabilityPageSize())
+	m.mcpOffset = m.clampMcpOffset(m.mcpOffset)
+}
+
+func (m *Model) scrollPanel(panel sidebarPanel, delta int) {
+	switch panel {
+	case panelAgents:
+		m.focus = focusAgents
+		m.input.Blur()
+		rows := m.agentPageSize()
+		total := len(agentTreeEntries(m.snapshot.Agents, m.collapsedAgents))
+		m.agentOffset = min(max(0, total-rows), max(0, m.agentOffset+delta))
+		m.keepAgentSelectionInWindow()
+		m.refreshViewport()
+	case panelFindings:
+		m.focus = focusVulnerabilities
+		m.input.Blur()
+		totalRows, _ := m.vulnerabilityScrollRows()
+		m.vulnOffset = min(max(0, totalRows-m.vulnerabilityPageSize()), max(0, m.vulnOffset+delta))
+		m.keepVulnerabilitySelectionInWindow()
+	case panelMcp:
+		m.focus = focusMcp
+		m.input.Blur()
+		m.mcpOffset = m.clampMcpOffset(m.mcpOffset + delta)
+	}
+}

strix/interface/tui/internal/app/panels_test.go

@@ -0,0 +1,357 @@
+package app
+
+import (
+	"fmt"
+	"strings"
+	"testing"
+
+	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
+	"github.com/charmbracelet/x/ansi"
+	"github.com/usestrix/strix/tui/internal/protocol"
+)
+
+func panelsModel(t *testing.T) Model {
+	t.Helper()
+	m := New(nil)
+	m.width, m.height = 130, 40
+	m.showSplash = false
+	m.ready = true
+	root := "root"
+	agents := []protocol.Agent{{ID: root, Name: "Root Agent", Status: "running"}}
+	for i := 0; i < 30; i++ {
+		name := fmt.Sprintf("Worker %02d", i)
+		agents = append(agents, protocol.Agent{ID: name, Name: name, ParentID: &root, Status: "running"})
+	}
+	conns := make([]protocol.Connection, 0, 12)
+	for i := 0; i < 12; i++ {
+		conns = append(conns, protocol.Connection{Name: fmt.Sprintf("conn-%02d", i), ToolCount: 2})
+	}
+	url := "http://127.0.0.1:57388/?token=abc"
+	m.snapshot = protocol.Snapshot{
+		ScanStarted: true, ScanState: "running", Agents: agents, Connections: conns,
+		ViewerStatus: "running", ViewerURL: &url, Model: "openrouter/z-ai/glm-5.3",
+		Vulnerabilities: []map[string]any{{"id": "v1", "title": "Finding one", "severity": "high"}},
+	}
+	m.resizeViewport()
+	return m
+}
+
+func click(t *testing.T, m Model, x, y int) Model {
+	t.Helper()
+	updated, _ := m.updateMouse(tea.MouseMsg{X: x, Y: y, Button: tea.MouseButtonLeft, Action: tea.MouseActionPress})
+	return updated.(Model)
+}
+
+func panelRectOf(t *testing.T, m Model, panel sidebarPanel) panelRect {
+	t.Helper()
+	for _, rect := range m.sidebarPanels() {
+		if rect.panel == panel {
+			return rect
+		}
+	}
+	t.Fatalf("panel %d not in sidebar", panel)
+	return panelRect{}
+}
+
+func TestSidebarPanelsRenderHeadersWithControls(t *testing.T) {
+	m := panelsModel(t)
+	_, sidebarWidth, _, _ := m.layout()
+	view := ansi.Strip(m.sidebarView(sidebarWidth, m.height))
+	for _, want := range []string{"▾ Agents (31)", "▾ Findings (1)", "▾ MCP (12)", "⤢"} {
+		if !strings.Contains(view, want) {
+			t.Fatalf("sidebar missing %q:\n%s", want, view)
+		}
+	}
+	if strings.Contains(view, "⤡") {
+		t.Fatalf("nothing is zoomed, yet a restore glyph is drawn:\n%s", view)
+	}
+}
+
+func TestClickingPanelHeaderCollapsesAndExpands(t *testing.T) {
+	m := panelsModel(t)
+	_, _, chatWidth, _ := m.layout()
+	rect := panelRectOf(t, m, panelFindings)
+	before := panelRectOf(t, m, panelAgents).height
+
+	m = click(t, m, chatWidth+4, rect.top+1)
+	collapsed := panelRectOf(t, m, panelFindings)
+	if collapsed.height != 1 || !m.collapsedPanels[panelFindings] {
+		t.Fatalf("header click did not collapse the panel: %+v", collapsed)
+	}
+	if got := panelRectOf(t, m, panelAgents).height; got <= before {
+		t.Fatalf("agent panel did not grow into the freed rows: %d -> %d", before, got)
+	}
+	_, sidebarWidth, _, _ := m.layout()
+	view := ansi.Strip(m.sidebarView(sidebarWidth, m.height))
+	if !strings.Contains(view, "▸ Findings (1)") || strings.Contains(view, "Finding one") {
+		t.Fatalf("collapsed panel should be a one-line header without its rows:\n%s", view)
+	}
+
+	m = click(t, m, chatWidth+4, collapsed.top)
+	if panelRectOf(t, m, panelFindings).height == 1 || m.collapsedPanels[panelFindings] {
+		t.Fatalf("clicking the collapsed row did not expand the panel")
+	}
+	if m.focus != focusVulnerabilities {
+		t.Fatalf("expanded panel did not take focus: %v", m.focus)
+	}
+}
+
+func TestClickingZoomGlyphGivesPanelTheSidebar(t *testing.T) {
+	m := panelsModel(t)
+	_, sidebarWidth, chatWidth, _ := m.layout()
+	rect := panelRectOf(t, m, panelMcp)
+
+	m = click(t, m, m.width-3, rect.top+1)
+	if m.zoomedPanel != panelMcp {
+		t.Fatalf("zoom glyph click did not zoom the panel: %d", m.zoomedPanel)
+	}
+	for _, other := range []sidebarPanel{panelAgents, panelFindings} {
+		if got := panelRectOf(t, m, other).height; got != 1 {
+			t.Fatalf("panel %d should shrink to its header while another is zoomed, got %d", other, got)
+		}
+	}
+	zoomed := panelRectOf(t, m, panelMcp)
+	rects := m.sidebarPanels()
+	if last := rects[len(rects)-1]; last.top+last.height != m.height || zoomed.height < m.height-12 {
+		t.Fatalf("zoomed panel does not fill the sidebar: %+v screen=%d", rects, m.height)
+	}
+	view := ansi.Strip(m.sidebarView(sidebarWidth, m.height))
+	if !strings.Contains(view, "⤡") || !strings.Contains(view, "conn-11") {
+		t.Fatalf("zoomed roster should show every connection and the restore glyph:\n%s", view)
+	}
+	if m.focus != focusMcp {
+		t.Fatalf("zoomed panel did not take focus: %v", m.focus)
+	}
+
+	m = click(t, m, m.width-3, zoomed.top+1)
+	if m.zoomedPanel != panelNone || panelRectOf(t, m, panelAgents).height == 1 {
+		t.Fatalf("restore glyph did not unzoom")
+	}
+
+	m = click(t, m, m.width-3, panelRectOf(t, m, panelAgents).top+1)
+	m = click(t, m, chatWidth+4, panelRectOf(t, m, panelFindings).top)
+	if m.zoomedPanel != panelNone || m.collapsedPanels[panelFindings] {
+		t.Fatalf("clicking a shrunk row should restore the whole sidebar")
+	}
+}
+
+func TestToggleButtonsHideAndShowSidebar(t *testing.T) {
+	m := panelsModel(t)
+	m.focus = focusAgents
+	showSidebar, sidebarWidth, chatWidth, _ := m.layout()
+	if !showSidebar {
+		t.Fatalf("precondition: sidebar visible")
+	}
+	rows := strings.Split(ansi.Strip(m.sidebarView(sidebarWidth, m.height)), "\n")
+	if !strings.HasSuffix(rows[1], " »  │") || strings.Contains(rows[2], "»") {
+		t.Fatalf("viewer panel lacks the one-row hide button:\n%s", strings.Join(rows[:3], "\n"))
+	}
+	if strings.Contains(ansi.Strip(m.View()), "«") || strings.Contains(ansi.Strip(m.statusView(chatWidth)), "sidebar") {
+		t.Fatalf("show button or text hint drawn while the sidebar is visible")
+	}
+
+	m = click(t, m, m.width-4, 1)
+	if showSidebar, _, width, _ := m.layout(); showSidebar || width != m.width-sidebarRailWidth-1 {
+		t.Fatalf("hide button click did not hide the sidebar: show=%v chatWidth=%d", showSidebar, width)
+	}
+	if m.focus != focusInput {
+		t.Fatalf("focus stayed on a hidden panel: %v", m.focus)
+	}
+	frame := ansi.Strip(m.View())
+	rows = strings.Split(frame, "\n")
+	if !strings.HasSuffix(rows[1], "  «  ") || strings.Contains(rows[0], "«") || strings.Contains(rows[3], "«") || lipgloss.Width(rows[0]) != m.width {
+		t.Fatalf("hidden sidebar should leave a %d-row show button in the rail:\n%s", railButtonHeight, strings.Join(rows[:4], "\n"))
+	}
+	if !m.toggleButtonHit(m.width-railButtonWidth, railButtonHeight-1) || m.toggleButtonHit(m.width-railButtonWidth-1, 0) || m.toggleButtonHit(m.width-1, railButtonHeight) {
+		t.Fatalf("show button hit zone does not match the drawn button")
+	}
+	if strings.Contains(frame, "»") {
+		t.Fatalf("hide button drawn while the sidebar is hidden")
+	}
+
+	m = click(t, m, m.width-3, 2)
+	if showSidebar, _, _, _ := m.layout(); !showSidebar {
+		t.Fatalf("show button click did not bring the sidebar back")
+	}
+}
+
+func TestFocusCyclingSkipsShrunkPanels(t *testing.T) {
+	m := panelsModel(t)
+	m.collapsedPanels[panelFindings] = true
+	seen := map[focusMode]bool{}
+	for range 6 {
+		m.cycleFocus(1)
+		seen[m.focus] = true
+	}
+	if seen[focusVulnerabilities] || !seen[focusAgents] || !seen[focusMcp] {
+		t.Fatalf("tab order wrong with findings collapsed: %v", seen)
+	}
+
+	m.zoomedPanel = panelAgents
+	seen = map[focusMode]bool{}
+	for range 6 {
+		m.cycleFocus(1)
+		seen[m.focus] = true
+	}
+	if seen[focusMcp] || !seen[focusAgents] {
+		t.Fatalf("tab order wrong with agents zoomed: %v", seen)
+	}
+
+	m.sidebarHidden = true
+	seen = map[focusMode]bool{}
+	for range 4 {
+		m.cycleFocus(1)
+		seen[m.focus] = true
+	}
+	if seen[focusAgents] || len(seen) != 2 {
+		t.Fatalf("tab order should stay in the chat column with the sidebar hidden: %v", seen)
+	}
+}
+
+func TestScrollbarHitTestFollowsPanelState(t *testing.T) {
+	m := panelsModel(t)
+	showSidebar, _, chatWidth, chatHeight := m.layout()
+	agents := panelRectOf(t, m, panelAgents)
+	at := func(y int) scrollbarTarget {
+		return m.scrollbarAt(tea.MouseMsg{X: m.width - 3, Y: y}, showSidebar, chatWidth, chatHeight)
+	}
+	if at(agents.top+2) != scrollbarAgents {
+		t.Fatalf("overflowing agent tree should expose its scrollbar")
+	}
+	if at(agents.top+1) != scrollbarNone {
+		t.Fatalf("the header row is a control, not a scrollbar")
+	}
+	findings := panelRectOf(t, m, panelFindings)
+	if at(findings.top+2) != scrollbarNone {
+		t.Fatalf("a findings list that fits has no scrollbar to grab")
+	}
+
+	m.collapsedPanels[panelAgents] = true
+	agents = panelRectOf(t, m, panelAgents)
+	if agents.height != 1 || at(agents.top) != scrollbarNone {
+		t.Fatalf("collapsed panel still reports a scrollbar")
+	}
+}
+
+func TestSidebarFitsShortTerminal(t *testing.T) {
+	m := panelsModel(t)
+	m.height = 19
+	m.snapshot.Connections = m.snapshot.Connections[:1]
+	m.resizeViewport()
+	_, sidebarWidth, _, _ := m.layout()
+	rects := m.sidebarPanels()
+	last := rects[len(rects)-1]
+	if last.top+last.height > m.height {
+		t.Fatalf("panels run past the screen: %+v (height %d)", rects, m.height)
+	}
+	if got := lipgloss.Height(m.sidebarView(sidebarWidth, m.height)); got > m.height {
+		t.Fatalf("sidebar renders %d rows on a %d-row terminal", got, m.height)
+	}
+	squeezed := panelRectOf(t, m, panelMcp)
+	if squeezed.height != 1 {
+		t.Fatalf("expected the MCP panel squeezed to its header, got %+v", squeezed)
+	}
+	if last.panel != panelStats || last.height < 3 {
+		t.Fatalf("Model panel must keep its box on short terminals: %+v", last)
+	}
+	m = click(t, m, m.width-6, squeezed.top)
+	if m.zoomedPanel != panelMcp {
+		t.Fatalf("clicking a squeezed header should zoom it, zoomed=%v", m.zoomedPanel)
+	}
+}
+
+func TestModelPanelHasNoControlsAndNeverShrinks(t *testing.T) {
+	m := panelsModel(t)
+	_, sidebarWidth, _, _ := m.layout()
+	stats := panelRectOf(t, m, panelStats)
+	rows := strings.Split(ansi.Strip(m.sidebarView(sidebarWidth, m.height)), "\n")
+	if first := rows[stats.top+1]; strings.ContainsAny(first, "▾▸⤢⤡") || strings.Contains(first, "Model") || !strings.Contains(first, m.snapshot.Model[:8]) {
+		t.Fatalf("Model panel should start with the model name and carry no controls: %q", first)
+	}
+	before := m.sidebarPanels()
+	m = click(t, m, m.width-3, stats.top+1)
+	m = click(t, m, m.width-10, stats.top+1)
+	if m.zoomedPanel != panelNone || len(m.collapsedPanels) != 0 {
+		t.Fatalf("clicks on the Model header changed panel state: zoom=%v collapsed=%v", m.zoomedPanel, m.collapsedPanels)
+	}
+	m = click(t, m, m.width-3, panelRectOf(t, m, panelAgents).top+1)
+	if m.zoomedPanel != panelAgents {
+		t.Fatalf("precondition: agents zoomed")
+	}
+	after := panelRectOf(t, m, panelStats)
+	if after.height != before[len(before)-1].height {
+		t.Fatalf("Model panel shrank under zoom: %+v -> %+v", before[len(before)-1], after)
+	}
+}
+
+func TestModelPanelSticksToBottom(t *testing.T) {
+	m := panelsModel(t)
+	m = click(t, m, m.width-10, panelRectOf(t, m, panelAgents).top+1)
+	if !m.collapsedPanels[panelAgents] {
+		t.Fatalf("precondition: agents collapsed")
+	}
+	stats := panelRectOf(t, m, panelStats)
+	if stats.top+stats.height != m.height {
+		t.Fatalf("Model panel not at the bottom: %+v (height %d)", stats, m.height)
+	}
+	_, sidebarWidth, _, _ := m.layout()
+	rows := strings.Split(ansi.Strip(m.sidebarView(sidebarWidth, m.height)), "\n")
+	if len(rows) != m.height {
+		t.Fatalf("sidebar renders %d rows, want %d", len(rows), m.height)
+	}
+	if !strings.Contains(rows[stats.top+1], m.snapshot.Model[:8]) {
+		t.Fatalf("Model panel not where its rect says:\n%s", strings.Join(rows[stats.top-2:], "\n"))
+	}
+}
+
+func TestStatsPanelKeepsAssignedHeight(t *testing.T) {
+	m := panelsModel(t)
+	m.snapshot.Model = strings.Repeat("openrouter/some-vendor/a-very-long-model-name ", 12)
+	m.resizeViewport()
+	_, sidebarWidth, _, _ := m.layout()
+	statsHeight, _, _, _ := m.sidebarHeights()
+	if statsHeight != 15 {
+		t.Fatalf("stats panel should hit its cap, got %d", statsHeight)
+	}
+	if got := lipgloss.Height(m.sidebarView(sidebarWidth, m.height)); got != m.height {
+		t.Fatalf("sidebar renders %d rows, want %d", got, m.height)
+	}
+}
+
+func TestHeadersAreFollowedByABlankRow(t *testing.T) {
+	m := panelsModel(t)
+	_, sidebarWidth, _, _ := m.layout()
+	rows := strings.Split(ansi.Strip(m.sidebarView(sidebarWidth, m.height)), "\n")
+	for _, panel := range []sidebarPanel{panelAgents, panelFindings, panelMcp} {
+		rect := panelRectOf(t, m, panel)
+		header, blank, first := rows[rect.top+1], rows[rect.top+2], rows[rect.top+3]
+		if !strings.Contains(header, m.panelTitle(panel)) || strings.TrimSpace(strings.Trim(blank, "│")) != "" || strings.TrimSpace(strings.Trim(first, "│")) == "" {
+			t.Fatalf("panel %d should render header, blank row, content:\n%s", panel, strings.Join(rows[rect.top:rect.top+4], "\n"))
+		}
+	}
+}
+
+func TestLonePanelHasNoControls(t *testing.T) {
+	m := panelsModel(t)
+	m.snapshot.Vulnerabilities = nil
+	m.snapshot.Connections = nil
+	_, sidebarWidth, _, _ := m.layout()
+	view := ansi.Strip(m.sidebarView(sidebarWidth, m.height))
+	if strings.ContainsAny(view, "▾▸⤢⤡") || !strings.Contains(view, "Agents (31)") {
+		t.Fatalf("a lone Agents panel should carry no collapse or zoom controls:\n%s", view)
+	}
+	rect := panelRectOf(t, m, panelAgents)
+	m = click(t, m, m.width-3, rect.top+1)
+	m = click(t, m, m.width-10, rect.top+1)
+	if m.zoomedPanel != panelNone || len(m.collapsedPanels) != 0 || panelRectOf(t, m, panelAgents) != rect {
+		t.Fatalf("header clicks changed a lone panel: zoom=%v collapsed=%v", m.zoomedPanel, m.collapsedPanels)
+	}
+	if m.sidebarGap() != 0 {
+		t.Fatalf("lone Agents panel should take the spare rows, gap=%d", m.sidebarGap())
+	}
+	m.snapshot.Connections = panelsModel(t).snapshot.Connections
+	if !strings.Contains(ansi.Strip(m.sidebarView(sidebarWidth, m.height)), "▾ Agents (31)") {
+		t.Fatalf("controls should return once a second panel exists")
+	}
+}

strix/interface/tui/internal/app/selection.go

@@ -11,6 +11,7 @@ import (
 
 	tea "github.com/charmbracelet/bubbletea"
 	"github.com/charmbracelet/x/ansi"
+	"github.com/usestrix/strix/tui/internal/render"
 )
 
 type selectionCopiedMsg struct{ err error }
@@ -206,7 +207,7 @@ func (m *Model) toggleEventAtLine(line int) {
 
 func (m Model) selectedText() string {
 	fromLine, fromCol, toLine, toCol := m.selection.bounds()
-	source := m.viewportContent
+	source := render.StopSpinners(m.viewportContent)
 	if m.selection.region == regionInput {
 		source = m.inputText()
 	}

strix/interface/tui/internal/app/setup.go

@@ -126,11 +126,14 @@ func (m Model) statusVisible() bool {
 }
 
 func (m Model) layout() (showSidebar bool, sidebarWidth, chatWidth, chatHeight int) {
-	showSidebar = m.width >= 120
-	if showSidebar {
+	showSidebar = m.width >= 120 && !m.sidebarHidden
+	switch {
+	case showSidebar:
 		sidebarWidth = max(24, m.width/5)
 		chatWidth = m.width - sidebarWidth - 1
-	} else {
+	case m.railVisible():
+		chatWidth = m.width - sidebarRailWidth - 1
+	default:
 		chatWidth = m.width
 	}
 	statusH := 0

strix/interface/tui/internal/app/spinner_test.go

@@ -0,0 +1,24 @@
+package app
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/charmbracelet/x/ansi"
+	"github.com/usestrix/strix/tui/internal/protocol"
+)
+
+func TestParkedWaitSpins(t *testing.T) {
+	model := New(nil)
+	model.width, model.height, model.showSplash, model.ready = 130, 30, false, true
+	model.snapshot = protocol.Snapshot{
+		Agents: []protocol.Agent{{ID: "one", Name: "Agent", Status: "waiting"}},
+		Events: []protocol.Event{{ID: "1", AgentID: "one", Type: "tool", Data: map[string]any{"tool_name": "wait_for_agents"}}},
+	}
+	model.resizeViewport()
+	before := ansi.Strip(model.View())
+	model.sweepFrame += 2
+	if after := ansi.Strip(model.View()); strings.Contains(before, "○ waiting") || before == after {
+		t.Fatalf("wait line did not spin:\n%s", before)
+	}
+}

strix/interface/tui/internal/app/update.go

@@ -162,12 +162,8 @@ func (m Model) updateMouse(msg tea.MouseMsg) (tea.Model, tea.Cmd) {
 		return m.updateSetupMouse(msg)
 	}
 	showSidebar, _, chatWidth, chatHeight := m.layout()
-	viewerHeight := m.viewerHeight()
-	_, vulnHeight, mcpHeight, agentHeight := m.sidebarHeights()
 	x, y := msg.X, msg.Y
-	if m.updateMainScrollbarMouse(
-		msg, showSidebar, chatWidth, chatHeight, viewerHeight, agentHeight, vulnHeight, mcpHeight,
-	) {
+	if m.updateMainScrollbarMouse(msg, showSidebar, chatWidth, chatHeight) {
 		return m, nil
 	}
 	if m.selection.dragging {
@@ -198,24 +194,8 @@ func (m Model) updateMouse(msg tea.MouseMsg) (tea.Model, tea.Cmd) {
 	switch msg.Button {
 	case tea.MouseButtonWheelUp:
 		if showSidebar && x >= chatWidth+1 {
-			switch {
-			case y < viewerHeight:
-				return m, nil
-			case y < viewerHeight+agentHeight:
-				m.focus = focusAgents
-				m.input.Blur()
-				m.agentOffset = max(0, m.agentOffset-3)
-				m.keepAgentSelectionInWindow()
-				m.refreshViewport()
-			case vulnHeight > 0 && y < viewerHeight+agentHeight+vulnHeight:
-				m.focus = focusVulnerabilities
-				m.input.Blur()
-				m.vulnOffset = max(0, m.vulnOffset-3)
-				m.keepVulnerabilitySelectionInWindow()
-			case mcpHeight > 0 && y < viewerHeight+agentHeight+vulnHeight+mcpHeight:
-				m.focus = focusMcp
-				m.input.Blur()
-				m.mcpOffset = m.clampMcpOffset(m.mcpOffset - 3)
+			if rect, ok := m.panelAt(y); ok && rect.height > 1 {
+				m.scrollPanel(rect.panel, -3)
 			}
 			return m, nil
 		}
@@ -226,26 +206,8 @@ func (m Model) updateMouse(msg tea.MouseMsg) (tea.Model, tea.Cmd) {
 		return m, nil
 	case tea.MouseButtonWheelDown:
 		if showSidebar && x >= chatWidth+1 {
-			switch {
-			case y < viewerHeight:
-				return m, nil
-			case y < viewerHeight+agentHeight:
-				m.focus = focusAgents
-				m.input.Blur()
-				rows := m.agentPageSize()
-				m.agentOffset = min(max(0, len(agentTreeEntries(m.snapshot.Agents, m.collapsedAgents))-rows), m.agentOffset+3)
-				m.keepAgentSelectionInWindow()
-				m.refreshViewport()
-			case vulnHeight > 0 && y < viewerHeight+agentHeight+vulnHeight:
-				m.focus = focusVulnerabilities
-				m.input.Blur()
-				totalRows, _ := m.vulnerabilityScrollRows()
-				m.vulnOffset = min(max(0, totalRows-m.vulnerabilityPageSize()), m.vulnOffset+3)
-				m.keepVulnerabilitySelectionInWindow()
-			case mcpHeight > 0 && y < viewerHeight+agentHeight+vulnHeight+mcpHeight:
-				m.focus = focusMcp
-				m.input.Blur()
-				m.mcpOffset = m.clampMcpOffset(m.mcpOffset + 3)
+			if rect, ok := m.panelAt(y); ok && rect.height > 1 {
+				m.scrollPanel(rect.panel, 3)
 			}
 			return m, nil
 		}
@@ -288,21 +250,32 @@ func (m Model) updateMouse(msg tea.MouseMsg) (tea.Model, tea.Cmd) {
 		return m, nil
 	}
 
-	if !showSidebar || x < chatWidth+1 {
+	if x < chatWidth+1 {
 		return m, nil
 	}
-	// Sidebar: viewer, agents, vulnerabilities, then stats.
-	switch {
-	case y < viewerHeight:
+	if m.toggleButtonHit(x, y) {
+		m.selection.active = false
+		m.toggleSidebar()
+		return m, nil
+	}
+	if !showSidebar {
+		return m, nil
+	}
+	if y < m.viewerHeight() {
 		return m, send(m.client, "viewer.open", map[string]any{})
-	case y < viewerHeight+agentHeight:
+	}
+	rect, ok := m.panelAt(y)
+	if !ok || m.clickPanel(rect, x, y) {
+		return m, nil
+	}
+	localY := y - rect.top
+	switch rect.panel {
+	case panelAgents:
 		m.focus = focusAgents
 		m.input.Blur()
-		// Content starts after the top border (1) and vertical padding (1).
 		entries := agentTreeEntries(m.snapshot.Agents, m.collapsedAgents)
-		start := windowStart(m.agentOffset, len(entries), max(1, agentHeight-4))
-		localY := y - viewerHeight
-		if row := start + localY - 2; localY >= 2 && localY < agentHeight-2 && row < len(entries) {
+		start := windowStart(m.agentOffset, len(entries), m.agentPageSize())
+		if row := start + localY - 2; localY >= 2 && localY < rect.height-1 && row < len(entries) {
 			m.selectedAgent = entries[row].index
 			agentID := m.snapshot.Agents[m.selectedAgent].ID
 			if hasAgentChildren(agentID, m.snapshot.Agents) {
@@ -311,24 +284,22 @@ func (m Model) updateMouse(msg tea.MouseMsg) (tea.Model, tea.Cmd) {
 			}
 			m.refreshViewport()
 		}
-	case vulnHeight > 0 && y < viewerHeight+agentHeight+vulnHeight:
+	case panelFindings:
 		m.focus = focusVulnerabilities
 		m.input.Blur()
-		// Content starts after the top border (1); clicking a row opens its detail.
-		row := y - viewerHeight - agentHeight - 1
-		if idx := m.vulnerabilityIndexAtRow(row); row >= 0 && row < vulnHeight-2 && idx >= 0 {
+		row := localY - 2
+		if idx := m.vulnerabilityIndexAtRow(row); row >= 0 && row < rect.height-3 && idx >= 0 {
 			m.selectedVuln = idx
 			m.openModal(modalVulnerability)
 		}
+	case panelMcp:
+		m.focus = focusMcp
+		m.input.Blur()
 	}
 	return m, nil
 }
 
-func (m *Model) updateMainScrollbarMouse(
-	msg tea.MouseMsg,
-	showSidebar bool,
-	chatWidth, chatHeight, viewerHeight, agentHeight, vulnHeight, mcpHeight int,
-) bool {
+func (m *Model) updateMainScrollbarMouse(msg tea.MouseMsg, showSidebar bool, chatWidth, chatHeight int) bool {
 	if msg.Action == tea.MouseActionRelease {
 		if m.draggingScrollbar == scrollbarNone {
 			return false
@@ -337,18 +308,18 @@ func (m *Model) updateMainScrollbarMouse(
 		return true
 	}
 	if msg.Action == tea.MouseActionMotion && m.draggingScrollbar != scrollbarNone {
-		m.scrollFromMouse(m.draggingScrollbar, msg.Y, chatHeight, viewerHeight, agentHeight, vulnHeight)
+		m.scrollFromMouse(m.draggingScrollbar, msg.Y, chatHeight)
 		return true
 	}
 	if msg.Action != tea.MouseActionPress || msg.Button != tea.MouseButtonLeft {
 		return false
 	}
-	target := m.scrollbarAt(msg, showSidebar, chatWidth, chatHeight, viewerHeight, agentHeight, vulnHeight, mcpHeight)
+	target := m.scrollbarAt(msg, showSidebar, chatWidth, chatHeight)
 	if target == scrollbarNone {
 		return false
 	}
 	m.draggingScrollbar = target
-	m.scrollFromMouse(target, msg.Y, chatHeight, viewerHeight, agentHeight, vulnHeight)
+	m.scrollFromMouse(target, msg.Y, chatHeight)
 	return true
 }
 
@@ -361,32 +332,28 @@ func nearColumn(x, column int) bool {
 	return x >= column-scrollbarGrab && x <= column+scrollbarGrab
 }
 
-// scrollbarAt reports which scrollbar, if any, the pointer is over.
-func (m Model) scrollbarAt(
-	msg tea.MouseMsg,
-	showSidebar bool,
-	chatWidth, chatHeight, viewerHeight, agentHeight, vulnHeight, mcpHeight int,
-) scrollbarTarget {
-	mcpTop := viewerHeight + agentHeight + vulnHeight
-	switch {
-	case nearColumn(msg.X, chatWidth-2) && msg.Y >= 1 && msg.Y < chatHeight-1 &&
-		m.viewport.TotalLineCount() > m.viewport.VisibleLineCount():
+func (m Model) scrollbarAt(msg tea.MouseMsg, showSidebar bool, chatWidth, chatHeight int) scrollbarTarget {
+	if nearColumn(msg.X, chatWidth-2) && msg.Y >= 1 && msg.Y < chatHeight-1 &&
+		m.viewport.TotalLineCount() > m.viewport.VisibleLineCount() {
 		return scrollbarTrace
-	case showSidebar && nearColumn(msg.X, m.width-3) && msg.Y >= viewerHeight+2 &&
-		msg.Y < viewerHeight+agentHeight-2 &&
-		len(agentTreeEntries(m.snapshot.Agents, m.collapsedAgents)) > m.agentPageSize():
-		return scrollbarAgents
-	case showSidebar && vulnHeight > 0 && nearColumn(msg.X, m.width-3) &&
-		msg.Y >= viewerHeight+agentHeight+1 &&
-		msg.Y < viewerHeight+agentHeight+vulnHeight-1:
-		totalRows, _ := m.vulnerabilityScrollRows()
-		if totalRows > m.vulnerabilityPageSize() {
+	}
+	if !showSidebar || !nearColumn(msg.X, m.width-3) {
+		return scrollbarNone
+	}
+	rect, ok := m.panelAt(msg.Y)
+	if !ok || rect.height <= 1 || msg.Y < rect.top+2 || msg.Y >= rect.top+rect.height-1 {
+		return scrollbarNone
+	}
+	switch rect.panel {
+	case panelAgents:
+		if len(agentTreeEntries(m.snapshot.Agents, m.collapsedAgents)) > m.agentPageSize() {
+			return scrollbarAgents
+		}
+	case panelFindings:
+		if totalRows, _ := m.vulnerabilityScrollRows(); totalRows > m.vulnerabilityPageSize() {
 			return scrollbarFindings
 		}
-	// The roster scrolls below a fixed header, so its bar starts two rows into
-	// the panel (border then header) rather than one.
-	case showSidebar && mcpHeight > 0 && nearColumn(msg.X, m.width-3) &&
-		msg.Y >= mcpTop+2 && msg.Y < mcpTop+mcpHeight-1:
+	case panelMcp:
 		if len(m.snapshot.Connections) > m.mcpPageSize() {
 			return scrollbarMcp
 		}
@@ -394,10 +361,7 @@ func (m Model) scrollbarAt(
 	return scrollbarNone
 }
 
-func (m *Model) scrollFromMouse(
-	target scrollbarTarget,
-	y, chatHeight, viewerHeight, agentHeight, vulnHeight int,
-) {
+func (m *Model) scrollFromMouse(target scrollbarTarget, y, chatHeight int) {
 	switch target {
 	case scrollbarTrace:
 		height := max(1, chatHeight-2)
@@ -411,7 +375,7 @@ func (m *Model) scrollFromMouse(
 		total := len(agentTreeEntries(m.snapshot.Agents, m.collapsedAgents))
 		m.focus = focusAgents
 		m.input.Blur()
-		m.agentOffset = scrollbarOffset(y-viewerHeight-2, height, total, height)
+		m.agentOffset = scrollbarOffset(y-m.panelTop(panelAgents)-2, height, total, height)
 		m.keepAgentSelectionInWindow()
 		m.refreshViewport()
 	case scrollbarFindings:
@@ -420,15 +384,14 @@ func (m *Model) scrollFromMouse(
 		m.focus = focusVulnerabilities
 		m.input.Blur()
 		// The offset is a row, so dragging moves the list continuously.
-		m.vulnOffset = scrollbarOffset(y-viewerHeight-agentHeight-1, height, totalRows, height)
+		m.vulnOffset = scrollbarOffset(y-m.panelTop(panelFindings)-2, height, totalRows, height)
 		m.keepVulnerabilitySelectionInWindow()
 	case scrollbarMcp:
 		height := m.mcpPageSize()
 		total := len(m.snapshot.Connections)
 		m.focus = focusMcp
 		m.input.Blur()
-		// The bar starts two rows into the panel (border then the fixed header).
-		m.mcpOffset = scrollbarOffset(y-viewerHeight-agentHeight-vulnHeight-2, height, total, height)
+		m.mcpOffset = scrollbarOffset(y-m.panelTop(panelMcp)-2, height, total, height)
 	}
 }
 
@@ -579,13 +542,11 @@ func labelHitAt(panel, label string, left, top, x, y int) bool {
 
 func (m *Model) cycleFocus(delta int) {
 	available := []focusMode{focusInput, focusChat}
-	if m.width >= 120 {
-		available = append(available, focusAgents)
-		if len(m.snapshot.Vulnerabilities) > 0 {
-			available = append(available, focusVulnerabilities)
-		}
-		if len(m.snapshot.Connections) > 0 {
-			available = append(available, focusMcp)
+	if showSidebar, _, _, _ := m.layout(); showSidebar {
+		for _, rect := range m.sidebarPanels() {
+			if focus, ok := panelFocus(rect.panel); ok && rect.height > 1 {
+				available = append(available, focus)
+			}
 		}
 	}
 	idx := 0

strix/interface/tui/internal/app/view.go

@@ -28,15 +28,16 @@ type renderedBlock struct {
 	version    int
 	width      int
 	expanded   bool
+	live       bool
 	wrapped    string
 	expandable bool
 	height     int
 }
 
-func (m *Model) renderEvent(event protocol.Event, width int) renderedBlock {
+func (m *Model) renderEvent(event protocol.Event, width int, live bool) renderedBlock {
 	expanded := m.expandedEvents[event.ID]
 	if cached, ok := m.blockCache[event.ID]; ok &&
-		cached.version == event.Version && cached.width == width && cached.expanded == expanded {
+		cached.version == event.Version && cached.width == width && cached.expanded == expanded && cached.live == live {
 		return cached
 	}
 	var block string
@@ -48,7 +49,10 @@ func (m *Model) renderEvent(event protocol.Event, width int) renderedBlock {
 		name := render.StringValue(event.Data["tool_name"])
 		block, expandable = render.CollapseTool(render.Tool(event.Data), name, expanded)
 	}
-	entry := renderedBlock{version: event.Version, width: width, expanded: expanded, expandable: expandable}
+	if !live {
+		block = render.StopSpinners(block)
+	}
+	entry := renderedBlock{version: event.Version, width: width, expanded: expanded, live: live, expandable: expandable}
 	if block != "" {
 		entry.wrapped = wrapBlock(block, width)
 		entry.height = strings.Count(entry.wrapped, "\n") + 1
@@ -96,6 +100,15 @@ func (m *Model) chatContent() string {
 	// to width-2 and indent every line by one cell.
 	contentWidth := max(1, m.viewport.Width-2)
 	render.SetImageWidth(contentWidth - 2)
+	// A parked agent is waiting on its latest tool call.
+	parkedOn := ""
+	if m.snapshot.Agents[m.selectedAgent].Status == "waiting" {
+		for _, event := range events {
+			if event.AgentID == agentID && event.Type == "tool" {
+				parkedOn = event.ID
+			}
+		}
+	}
 	var blocks []string
 	var spans []eventSpan
 	line := 0
@@ -103,7 +116,7 @@ func (m *Model) chatContent() string {
 		if event.AgentID != agentID {
 			continue
 		}
-		entry := m.renderEvent(event, contentWidth)
+		entry := m.renderEvent(event, contentWidth, event.ID == parkedOn)
 		if entry.wrapped == "" {
 			continue
 		}
@@ -164,6 +177,18 @@ func wrapBlock(value string, width int) string {
 	return strings.Join(out, "\n")
 }
 
+// hyperlinkBlock wraps a URL to the column width and marks every wrapped line
+// as an OSC 8 hyperlink to the whole URL. Terminals that linkify by text only
+// see the first line of a wrapped URL, which for the viewer meant opening it
+// with a truncated token.
+func hyperlinkBlock(url string, width int, style lipgloss.Style) string {
+	lines := strings.Split(wrapBlock(url, width), "\n")
+	for i, line := range lines {
+		lines[i] = ansi.SetHyperlink(url) + style.Render(line) + ansi.ResetHyperlink()
+	}
+	return strings.Join(lines, "\n")
+}
+
 // scrollbarThumb brightens the bar being dragged so the grab reads as taking
 // hold of it.
 func (m Model) scrollbarThumb(target scrollbarTarget) lipgloss.Color {
@@ -362,10 +387,11 @@ const (
 
 // fillBackground paints the whole frame black like Textual's Screen background.
 // Bubble Tea has no screen compositor, so any cell the view does not explicitly
-// color shows the terminal's default background. lipgloss emits a full reset
-// (\x1b[0m) at the end of every styled span, which clears both foreground and
-// background. Reasserting only black made uncolored and faint text inherit the
-// terminal profile's foreground; light profiles therefore rendered that text
+// color shows the terminal's default background. lipgloss emits a reset
+// (\x1b[0m, or the bare \x1b[m that x/ansi uses) at the end of every styled
+// span, which clears both foreground and background. Reasserting only black
+// made uncolored and faint text inherit the terminal profile's foreground;
+// light profiles therefore rendered that text
 // black-on-black. Reapply both base colors after each reset (and at the start).
 // Spans that set their own colors — inline code, selected rows, buttons — keep
 // them, because their color is emitted after the base style.
@@ -373,9 +399,14 @@ func fillBackground(view string) string {
 	if view == "" {
 		return view
 	}
-	return baseFrameColors + strings.ReplaceAll(view, "\x1b[0m", "\x1b[0m"+baseFrameColors)
+	return baseFrameColors + baseColorRestorer.Replace(view)
 }
 
+var baseColorRestorer = strings.NewReplacer(
+	"\x1b[0m", "\x1b[0m"+baseFrameColors,
+	"\x1b[m", "\x1b[m"+baseFrameColors,
+)
+
 func (m Model) splashView() string {
 	shine := "Starting Strix Agent"
 	chars := []rune(shine)
@@ -410,33 +441,19 @@ func (m Model) splashView() string {
 	content := wordmark() + "\n\n" +
 		welcome + "\n" + version + "\n" + tagline + "\n\n" +
 		start.String() + "\n\n" + url
-	if warn := m.snapshot.ModelWarning; warn != "" {
-		content += "\n\n" + splashModelWarning(m.snapshot.Model, warn)
-	}
 	panel := lipgloss.NewStyle().Border(lipgloss.RoundedBorder()).BorderForeground(green).Padding(1, 6).Align(lipgloss.Center).Render(content)
 	// #splash_screen background is solid black.
 	return lipgloss.Place(m.width, m.height, lipgloss.Center, lipgloss.Center, panel,
 		lipgloss.WithWhitespaceBackground(black))
 }
 
-// splashModelWarning renders the backend's full warning sentence, with the
-// model name highlighted when the sentence leads with it.
-func splashModelWarning(model, warning string) string {
-	yellow := lipgloss.Color("#eab308")
-	out := lipgloss.NewStyle().Bold(true).Foreground(yellow).Render("⚠ ")
-	if model != "" && strings.HasPrefix(warning, model) {
-		out += lipgloss.NewStyle().Bold(true).Foreground(render.Cyan).Render(model)
-		warning = strings.TrimPrefix(warning, model)
-	}
-	return out + lipgloss.NewStyle().Foreground(yellow).Render(warning)
-}
-
 // chatPaneKey identifies everything the bordered trace depends on.
 type chatPaneKey struct {
 	offset        int
 	width, height int
 	border        lipgloss.Color
 	selection     selectionState
+	spinnerFrame  int
 }
 
 // chatPane memoizes the bordered trace: slicing, scrollbar padding and border
@@ -450,12 +467,18 @@ var chatPane struct {
 }
 
 func (m Model) renderChatPane(width, height int, border lipgloss.Color) string {
-	key := chatPaneKey{offset: m.viewport.YOffset, width: width, height: height, border: border, selection: m.selection}
+	visible := visibleContent(m.viewportContent, m.viewport.YOffset, height)
+	// Only a trace with a spinner on screen changes with the tick.
+	spinnerFrame := 0
+	if strings.Contains(visible, render.SpinnerMarker) {
+		spinnerFrame = m.sweepFrame / 2
+	}
+	key := chatPaneKey{offset: m.viewport.YOffset, width: width, height: height, border: border, selection: m.selection, spinnerFrame: spinnerFrame}
 	if chatPane.out != "" && chatPane.key == key && chatPane.content == m.viewportContent {
 		return chatPane.out
 	}
 	trace := withVerticalScrollbar(
-		m.highlightSelection(visibleContent(m.viewportContent, m.viewport.YOffset, height), m.viewport.YOffset),
+		render.AnimateSpinners(m.highlightSelection(visible, m.viewport.YOffset), spinnerFrame),
 		width,
 		height,
 		m.viewport.TotalLineCount(),
@@ -499,6 +522,8 @@ func (m Model) mainView() string {
 	body := leftColumn
 	if showSidebar {
 		body = lipgloss.JoinHorizontal(lipgloss.Top, leftColumn, " ", m.sidebarView(sidebarWidth, m.height))
+	} else if m.railVisible() {
+		body = lipgloss.JoinHorizontal(lipgloss.Top, leftColumn, " ", m.sidebarRail(m.height))
 	}
 	return lipgloss.NewStyle().Background(black).Foreground(textColor).Render(body)
 }
@@ -509,75 +534,108 @@ func (m Model) mainView() string {
 // and so never applied - honoring it made the outline vanish on the one panel
 // that had just become active.
 func (m Model) sidebarView(width, height int) string {
-	// Stats box height fits its content (auto, max 15); vulns panel max-height 12.
-	statsBody := m.statsView()
 	statsHeight, vulnHeight, mcpHeight, agentHeight := m.sidebarHeights()
-	agentBorder := dark
-	if m.focus == focusAgents {
-		agentBorder = green
-	}
-	// #agents_tree padding: 1 (all sides); interior lines = box - border - v.padding.
-	agentRows := max(1, agentHeight-4)
-	agentEntries := agentTreeEntries(m.snapshot.Agents, m.collapsedAgents)
-	agents := withVerticalScrollbar(
-		m.agentsView(max(1, width-5), agentRows),
-		width-4,
-		agentRows,
-		len(agentEntries),
-		agentRows,
-		m.agentOffset,
-		m.scrollbarThumb(scrollbarAgents),
-	)
-	parts := []string{
-		lipgloss.NewStyle().Width(width-2).Height(m.viewerHeight()-2).Border(lipgloss.RoundedBorder()).BorderForeground(dark).Padding(0, 1).Render(m.viewerView(width - 4)),
-		lipgloss.NewStyle().Width(width-2).Height(agentHeight-2).Border(lipgloss.RoundedBorder()).BorderForeground(agentBorder).Padding(1, 1).Render(agents),
-	}
-	if vulnHeight > 0 {
-		vulnBorder := dark
-		if m.focus == focusVulnerabilities {
-			vulnBorder = green
-		}
-		vulnRows := max(1, vulnHeight-2)
-		totalRows, offsetRows := m.vulnerabilityScrollRows()
-		findings := withVerticalScrollbar(
-			m.vulnerabilitiesView(m.vulnerabilityListWidth(), vulnRows),
+	parts := []string{lipgloss.NewStyle().
+		Width(width-2).
+		Height(m.viewerHeight()-2).
+		Border(lipgloss.RoundedBorder()).
+		BorderForeground(dark).
+		Padding(0, 1).
+		Render(m.viewerBox(width - 4))}
+	agents := ""
+	if agentHeight > 1 {
+		agentRows := max(1, agentHeight-4)
+		agents = withVerticalScrollbar(
+			m.agentsView(max(1, width-5), agentRows),
 			width-4,
-			vulnRows,
-			totalRows,
-			vulnRows,
-			offsetRows,
-			m.scrollbarThumb(scrollbarFindings),
+			agentRows,
+			len(agentTreeEntries(m.snapshot.Agents, m.collapsedAgents)),
+			agentRows,
+			m.agentOffset,
+			m.scrollbarThumb(scrollbarAgents),
 		)
-		parts = append(parts, lipgloss.NewStyle().Width(width-2).Height(vulnRows).Border(lipgloss.RoundedBorder()).BorderForeground(vulnBorder).Padding(0, 1).Render(findings))
+	}
+	parts = append(parts, m.panelBox(panelAgents, agents, width, agentHeight, m.focus == focusAgents))
+	if vulnHeight > 0 {
+		findings := ""
+		if vulnHeight > 1 {
+			vulnRows := max(1, vulnHeight-4)
+			totalRows, offsetRows := m.vulnerabilityScrollRows()
+			findings = withVerticalScrollbar(
+				m.vulnerabilitiesView(m.vulnerabilityListWidth(), vulnRows),
+				width-4,
+				vulnRows,
+				totalRows,
+				vulnRows,
+				offsetRows,
+				m.scrollbarThumb(scrollbarFindings),
+			)
+		}
+		parts = append(parts, m.panelBox(panelFindings, findings, width, vulnHeight, m.focus == focusVulnerabilities))
 	}
 	if mcpHeight > 0 {
-		mcpBorder := dark
-		if m.focus == focusMcp {
-			mcpBorder = green
+		roster := ""
+		if mcpHeight > 1 {
+			roster = m.mcpConnectionsView(width-4, max(1, mcpHeight-4))
 		}
-		mcpRows := max(1, mcpHeight-2)
-		parts = append(parts, lipgloss.NewStyle().Width(width-2).Height(mcpRows).Border(lipgloss.RoundedBorder()).BorderForeground(mcpBorder).Padding(0, 1).Render(m.mcpConnectionsView(width-4, mcpRows)))
+		parts = append(parts, m.panelBox(panelMcp, roster, width, mcpHeight, m.focus == focusMcp))
 	}
-	parts = append(parts, lipgloss.NewStyle().Width(width-2).Height(statsHeight-2).Border(lipgloss.RoundedBorder()).BorderForeground(dark).Padding(0, 1).Render(statsBody))
-	return strings.Join(parts, "\n")
+	if gap := m.sidebarGap(); gap > 0 {
+		parts = append(parts, lipgloss.NewStyle().Width(width).Height(gap).Render(""))
+	}
+	stats := ""
+	if statsHeight > 1 {
+		stats = fixedPanelBody(lipgloss.NewStyle().Width(width-4).Render(m.statsView()), width-4, statsHeight-2)
+	}
+	parts = append(parts, m.panelBox(panelStats, stats, width, statsHeight, false))
+	return lipgloss.JoinVertical(lipgloss.Left, parts...)
 }
 
 func (m Model) sidebarHeights() (statsHeight, vulnHeight, mcpHeight, agentHeight int) {
 	// Measure the stats panel the way its box will render it: a long model name
 	// wraps inside the sidebar, and counting only its newlines would size the
 	// box short and push the whole frame past the bottom of the terminal.
-	statsRows := lipgloss.Height(lipgloss.NewStyle().Width(m.viewerContentWidth()).Render(m.statsView()))
-	statsHeight = min(15, statsRows+2)
+	statsRows := lipgloss.Height(lipgloss.NewStyle().Width(m.sidebarInnerWidth()).Render(m.statsView()))
+	statsHeight = m.panelHeight(panelStats, min(15, statsRows+2))
 	if len(m.snapshot.Vulnerabilities) > 0 {
-		vulnHeight = min(12, len(m.vulnerabilityRows(m.vulnerabilityListWidth()))+2)
+		vulnHeight = m.panelHeight(panelFindings, min(13, len(m.vulnerabilityRows(m.vulnerabilityListWidth()))+4))
 	}
-	// One header line + one line per connection + the box border (2). Capped so a
+	// Header line + one line per connection + the box border (2). Capped so a
 	// long roster cannot crowd out the agent tree; a roster past the cap scrolls
 	// inside the panel. Absent entirely when the run has no MCP connections.
 	if len(m.snapshot.Connections) > 0 {
-		mcpHeight = min(9, len(m.snapshot.Connections)+3)
+		mcpHeight = m.panelHeight(panelMcp, min(10, len(m.snapshot.Connections)+4))
+	}
+	agentHeight = m.panelHeight(panelAgents, 5)
+	for _, p := range []struct {
+		h     *int
+		floor int
+	}{{&statsHeight, 4}, {&mcpHeight, 5}, {&vulnHeight, 5}} {
+		over := m.viewerHeight() + statsHeight + vulnHeight + mcpHeight + agentHeight - m.height
+		if over <= 0 {
+			break
+		}
+		if *p.h > p.floor {
+			*p.h -= min(over, *p.h-p.floor)
+		}
+	}
+	for _, h := range []*int{&mcpHeight, &vulnHeight, &statsHeight} {
+		if m.viewerHeight()+statsHeight+vulnHeight+mcpHeight+agentHeight <= m.height {
+			break
+		}
+		if *h > 1 {
+			*h = 1
+		}
+	}
+	spare := max(0, m.height-m.viewerHeight()-statsHeight-vulnHeight-mcpHeight-agentHeight)
+	switch {
+	case m.zoomedPanel == panelFindings && vulnHeight > 1:
+		vulnHeight += spare
+	case m.zoomedPanel == panelMcp && mcpHeight > 1:
+		mcpHeight += spare
+	case agentHeight > 1:
+		agentHeight += spare
 	}
-	agentHeight = max(3, m.height-m.viewerHeight()-statsHeight-vulnHeight-mcpHeight)
 	return
 }
 
@@ -585,7 +643,7 @@ func (m Model) viewerHeight() int {
 	return strings.Count(m.viewerView(m.viewerContentWidth()), "\n") + 3
 }
 
-func (m Model) viewerContentWidth() int {
+func (m Model) sidebarInnerWidth() int {
 	_, sidebarWidth, _, _ := m.layout()
 	if sidebarWidth == 0 {
 		sidebarWidth = 24
@@ -593,13 +651,16 @@ func (m Model) viewerContentWidth() int {
 	return max(1, sidebarWidth-4)
 }
 
+func (m Model) viewerContentWidth() int {
+	return max(1, m.sidebarInnerWidth()-toggleButtonWidth-1)
+}
+
 func (m Model) viewerView(width int) string {
 	switch m.snapshot.ViewerStatus {
 	case "running":
 		status := lipgloss.NewStyle().Foreground(green).Render("● Viewer running")
 		if m.snapshot.ViewerURL != nil && strings.TrimSpace(*m.snapshot.ViewerURL) != "" {
-			url := wrapBlock(strings.TrimSpace(*m.snapshot.ViewerURL), width)
-			return status + "\n" + lipgloss.NewStyle().Foreground(dim).Render(url)
+			return status + "\n" + hyperlinkBlock(strings.TrimSpace(*m.snapshot.ViewerURL), width, lipgloss.NewStyle().Foreground(dim))
 		}
 		return status
 	case "unavailable":
@@ -646,29 +707,22 @@ func (m Model) statsView() string {
 	return b.String()
 }
 
-// mcpConnectionsView renders the sidebar MCP panel: a header carrying the total
-// connection count, then one row per connection with a status glyph and its tool
-// count (or "offline").
+// mcpConnectionsView renders the sidebar MCP roster: one row per connection with
+// a status glyph and its tool count (or "offline").
 //   - a solid green dot marks an attached, idle connection;
 //   - a green cycling quarter-circle (◐ ◓ ◑ ◒) marks a call running against it;
 //   - a red dot plus "offline" marks a connection whose live session has died.
 //
-// The header stays fixed while the roster below it scrolls: when there are more
-// connections than the panel can show, the visible window is chosen by
-// m.mcpOffset and withVerticalScrollbar draws a thumb in the reserved last
-// column, exactly as the agent tree and findings list scroll.
+// When there are more connections than the panel can show, the visible window
+// is chosen by m.mcpOffset and withVerticalScrollbar draws a thumb in the
+// reserved last column, exactly as the agent tree and findings list scroll.
 //
 // "In use" is derived from the connection-tagged tool-call events in the stream,
 // not carried on the connection roster, so a call in flight shows motion without
 // any extra backend signal. The quarter-circle rides the shared sweepFrame tick.
 func (m Model) mcpConnectionsView(width, rows int) string {
 	conns := m.snapshot.Connections
-	header := truncate(lipgloss.NewStyle().Foreground(dim).Render(
-		fmt.Sprintf("MCP Connections (%d)", len(conns))), width)
-	bodyRows := max(0, rows-1)
-	if bodyRows == 0 {
-		return header
-	}
+	bodyRows := max(1, rows)
 	inUse := m.mcpInUse()
 	frames := []rune{'◐', '◓', '◑', '◒'}
 	// Reserve the scrollbar column whether or not the bar is showing, so the
@@ -705,7 +759,7 @@ func (m Model) mcpConnectionsView(width, rows int) string {
 		m.mcpOffset,
 		m.scrollbarThumb(scrollbarMcp),
 	)
-	return header + "\n" + roster
+	return roster
 }
 
 // mcpPageSize is how many connection rows the roster shows at once, below its

strix/interface/tui/internal/app/vulnerabilities.go

@@ -135,7 +135,7 @@ func clampVulnerabilityOffset(offset, total, height int) int {
 
 func (m Model) vulnerabilityPageSize() int {
 	_, vulnHeight, _, _ := m.sidebarHeights()
-	return max(1, vulnHeight-2)
+	return max(1, vulnHeight-4)
 }
 
 // vulnerabilityPageItems is how many findings a page step should move by: the
@@ -192,7 +192,7 @@ func (m Model) modalView() string {
 	switch m.modal {
 	case modalHelp:
 		title := lipgloss.NewStyle().Bold(true).Foreground(green).Width(34).Align(lipgloss.Center).Render("Strix Help")
-		body := lipgloss.NewStyle().Foreground(textColor).Render("F1        Help\nCtrl+O    Open viewer\nCtrl+Q/C  Quit\nESC       Stop Agent\nEnter     Send / expand node\nCtrl+J    Newline in message\nTab       Switch panels\n↑/↓       Navigate tree\nDrag      Select & copy text\nClick     Expand/collapse tool")
+		body := lipgloss.NewStyle().Foreground(textColor).Render("F1        Help\nCtrl+O    Open viewer\nCtrl+Q/C  Quit\nESC       Stop Agent\nEnter     Send / expand node\nCtrl+J    Newline in message\nTab       Switch panels\n↑/↓       Navigate tree\nDrag      Select & copy text\nClick     Expand/collapse tool\nClick ▾/⤢ Collapse / zoom panel\nClick »/« Hide / show sidebar")
 		content := title + "\n\n" + body
 		return lipgloss.NewStyle().Width(38).Border(lipgloss.RoundedBorder()).BorderForeground(green).Background(black).Padding(1, 2).Render(content)
 	case modalQuit:

strix/interface/tui/internal/protocol/protocol.go

@@ -71,7 +71,6 @@ type Snapshot struct {
 	ScopeMode           string           `json:"scope_mode"`
 	DiffBase            string           `json:"diff_base"`
 	Model               string           `json:"model"`
-	ModelWarning        string           `json:"model_warning"`
 	CaidoURL            string           `json:"caido_url"`
 	Messages            []Message        `json:"messages"`
 	Agents              []Agent          `json:"-"`

strix/interface/tui/internal/render/agent_message.go

@@ -11,8 +11,6 @@ import (
 // Markdown (agent_message_renderer.py)
 // ---------------------------------------------------------------------------
 
-var blankLineRuns = regexp.MustCompile(`\n\s*\n`)
-
 type mdHeader struct {
 	prefix string
 	strip  int
@@ -33,13 +31,36 @@ func renderAssistantMarkdown(content string) string {
 	if content == "" {
 		return ""
 	}
-	cleaned := strings.TrimSpace(blankLineRuns.ReplaceAllString(content, "\n\n"))
+	cleaned := strings.TrimSpace(collapseBlankLines(content))
 	if cleaned == "" {
 		return ""
 	}
 	return applyMarkdownStyles(cleaned)
 }
 
+func collapseBlankLines(content string) string {
+	lines := strings.Split(content, "\n")
+	out := make([]string, 0, len(lines))
+	inCode := false
+	blank := false
+	for _, line := range lines {
+		if strings.HasPrefix(line, "```") {
+			inCode = !inCode
+		}
+		if !inCode && strings.TrimSpace(line) == "" {
+			if blank {
+				continue
+			}
+			blank = true
+			out = append(out, "")
+			continue
+		}
+		blank = false
+		out = append(out, line)
+	}
+	return strings.Join(out, "\n")
+}
+
 func applyMarkdownStyles(text string) string {
 	var out strings.Builder
 	lines := strings.Split(text, "\n")

strix/interface/tui/internal/render/agents_graph.go

@@ -55,7 +55,8 @@ func renderAgentGraphTool(name string, args map[string]any, result any) string {
 			b.WriteString("\n  " + Dim().Render("Completing task..."))
 		}
 	case "wait_for_agents":
-		b.WriteString(Col(Gray).Render("○ ") + Dim().Render("waiting"))
+		// The chat pane animates the marker while the agent is parked here.
+		b.WriteString(Col(Gray).Render(SpinnerMarker+" ") + Dim().Render("waiting"))
 		if reason := StringValue(args["reason"]); reason != "" {
 			b.WriteString("\n  " + Dim().Render(reason))
 		}

strix/interface/tui/internal/render/dependency.go

@@ -51,7 +51,7 @@ func renderDependencyReport(args map[string]any, result any) string {
 	field("Target", StringValue(args["target"]))
 	section := func(label, value string) {
 		if value != "" {
-			b.WriteString("\n\n" + Bold(Field).Render(label) + "\n" + value)
+			b.WriteString("\n\n" + Bold(Field).Render(label) + "\n" + renderAssistantMarkdown(value))
 		}
 	}
 	section("Description", StringValue(args["description"]))

strix/interface/tui/internal/render/render_test.go

@@ -321,3 +321,36 @@ func TestCollapseToolOnlyOutputHeavyTools(t *testing.T) {
 		t.Fatal("respond_to_user must never collapse")
 	}
 }
+
+func TestReportSectionsRenderMarkdown(t *testing.T) {
+	body := "**Systemic themes:** zero auth on `/share`\n\n# Recommendations\n1. Merge the **ejs 3.1.10** fix"
+	cases := map[string]string{
+		"finish_scan":                 renderFinishScan(map[string]any{"executive_summary": body}),
+		"create_vulnerability_report": renderVulnerabilityReport(map[string]any{"title": "x", "description": body}, nil),
+		"create_dependency_report":    renderDependencyReport(map[string]any{"title": "x", "description": body}, nil),
+	}
+	for name, out := range cases {
+		plain := ansi.Strip(out)
+		for _, raw := range []string{"**", "`", "# Recommendations"} {
+			if strings.Contains(plain, raw) {
+				t.Fatalf("%s left raw markdown %q in %q", name, raw, plain)
+			}
+		}
+		for _, want := range []string{"Systemic themes:", "/share", "Recommendations", "1. Merge the ejs 3.1.10 fix"} {
+			if !strings.Contains(plain, want) {
+				t.Fatalf("%s lost %q in %q", name, want, plain)
+			}
+		}
+	}
+}
+
+func TestReportSectionsKeepBlankLinesInFencedCode(t *testing.T) {
+	body := "Intro.\n\n\n\nRequest:\n```http\nGET /share HTTP/1.1\nHost: target\n\n\nbody\n```\n\n\n\nAfter."
+	plain := ansi.Strip(renderVulnerabilityReport(map[string]any{"title": "x", "technical_analysis": body}, nil))
+	if !strings.Contains(plain, "Host: target\n\n\nbody") {
+		t.Fatalf("fenced code lost its blank lines in %q", plain)
+	}
+	if strings.Contains(plain, "Intro.\n\n\n") || strings.Contains(plain, "\n\n\nAfter.") {
+		t.Fatalf("blank-line runs outside fenced code were not collapsed in %q", plain)
+	}
+}

strix/interface/tui/internal/render/report.go

@@ -89,7 +89,7 @@ func renderReport(args map[string]any, result any, heading, pending string) stri
 
 	section := func(label, value string) {
 		if value != "" {
-			b.WriteString("\n\n" + Bold(Field).Render(label) + "\n" + value)
+			b.WriteString("\n\n" + Bold(Field).Render(label) + "\n" + renderAssistantMarkdown(value))
 		}
 	}
 	if confidence := StringValue(args["confidence"]); confidence != "" {

strix/interface/tui/internal/render/scan.go

@@ -13,7 +13,7 @@ func renderFinishScan(args map[string]any) string {
 	b.WriteString(Col(Green).Render("◆ ") + Bold(Green).Render("Penetration test completed"))
 	section := func(label, value string) {
 		if value != "" {
-			b.WriteString("\n\n" + Bold(Field).Render(label) + "\n" + value)
+			b.WriteString("\n\n" + Bold(Field).Render(label) + "\n" + renderAssistantMarkdown(value))
 		}
 	}
 	es := StringValue(args["executive_summary"])

strix/interface/tui/internal/render/spinner.go

@@ -0,0 +1,18 @@
+package render
+
+import "strings"
+
+// SpinnerMarker is a one-cell placeholder for a spinner. Rendered blocks are
+// cached, so the chat pane swaps in the current frame (AnimateSpinners) or a
+// still circle once the wait is over (StopSpinners).
+const SpinnerMarker = "\uE000"
+
+var spinnerFrames = []string{"⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏"}
+
+func StopSpinners(s string) string {
+	return strings.ReplaceAll(s, SpinnerMarker, "○")
+}
+
+func AnimateSpinners(s string, tick int) string {
+	return strings.ReplaceAll(s, SpinnerMarker, spinnerFrames[tick%len(spinnerFrames)])
+}

strix/interface/tui/runtime.py

@@ -11,7 +11,7 @@ import shutil
 import sys
 from copy import deepcopy
 from pathlib import Path
-from typing import TYPE_CHECKING, Any
+from typing import TYPE_CHECKING, Any, TextIO
 
 from strix.config import load_settings, persist_current
 from strix.core.agents import AgentCoordinator
@@ -58,6 +58,10 @@ class GoTuiPreActivationError(RuntimeError):
     """A sidecar failure raised before the Go TUI activates."""
 
 
+def _open_output_sink() -> TextIO:
+    return Path(os.devnull).open("a", buffering=1, encoding="utf-8")
+
+
 class GoTuiRuntime:
     def __init__(self, args: argparse.Namespace) -> None:
         self.args = args
@@ -69,6 +73,7 @@ class GoTuiRuntime:
         self.scan_error: BaseException | None = None
         self._last_sync_fingerprint = ""
         self._error_noted_agents: set[str] = set()
+        self._output_syncs: set[asyncio.Task[None]] = set()
         self.model_verified = False
         self._setup_preflight: asyncio.Task[None] | None = None
         self.controller = TuiController(
@@ -122,8 +127,8 @@ class GoTuiRuntime:
     async def check_setup_model(self) -> None:
         """Verify the model route as soon as the start screen is up.
 
-        The same round trip a direct launch makes in prepare_and_start, run in
-        the background so the screen paints first and the outcome lands in the
+        The same round trip main() makes before a direct launch, run in the
+        background so the screen paints first and the outcome lands in the
         setup log before the user has finished typing.
         """
         if not (load_settings().llm.model or "").strip():
@@ -157,13 +162,13 @@ class GoTuiRuntime:
         await preflight_model_connection(model)
         self.model_verified = True
 
-    def _start_preparation(self) -> asyncio.Task[None]:
+    def _start_preparation(self) -> asyncio.Task[None] | None:
         """Kick off the work that runs behind the freshly painted TUI."""
         if self.controller.setup_mode:
             self._setup_preflight = asyncio.create_task(self.check_setup_model())
             return self._setup_preflight
-        self.controller.begin_preparation()
-        return asyncio.create_task(self.prepare_and_start())
+        self.start_scan()
+        return None
 
     async def start_from_setup(self) -> None:
         candidate = deepcopy(self.args)
@@ -203,34 +208,6 @@ class GoTuiRuntime:
         self.init_run_state()
         self.start_scan()
 
-    async def prepare_and_start(self) -> None:
-        """Prepare a directly-launched scan once the TUI is on screen.
-
-        The model round trip and run preparation run here rather than before
-        launch so the interface appears immediately.
-        """
-        model = (load_settings().llm.model or "").strip()
-        set_scan_phase("preflight")
-        try:
-            await preflight_model_connection(model)
-        except Exception as exc:
-            logger.exception("Go TUI scan preparation failed")
-            report_error("model_connection_failed", exc)
-            self.controller.fail_preparation(str(exc))
-            return
-        try:
-            persist_current()
-            prepare_run(self.args)
-            telemetry_start(self.args)
-        except Exception as exc:
-            logger.exception("Go TUI scan preparation failed")
-            report_error("scan_preparation_failed", exc)
-            self.controller.fail_preparation(str(exc))
-            return
-        self.controller.scan_state = "running"
-        self.init_run_state()
-        self.start_scan()
-
     def start_scan(self) -> None:
         if self.scan_task is None:
             self.scan_task = asyncio.create_task(self._run_scan())
@@ -276,6 +253,18 @@ class GoTuiRuntime:
 
     def capture_event(self, agent_id: str, event: Any) -> None:
         self.live_view.ingest_sdk_event(agent_id, event)
+        if getattr(getattr(event, "item", None), "type", "") == "tool_call_output_item":
+            # A tool that parks its agent has already set the agent's status by
+            # the time it returns; sync it now so both reach the TUI together.
+            task = asyncio.get_running_loop().create_task(self._sync_and_notify())
+            self._output_syncs.add(task)
+            task.add_done_callback(self._output_syncs.discard)
+            return
+        self.controller.notify_changed()
+
+    async def _sync_and_notify(self) -> None:
+        with contextlib.suppress(Exception):
+            await self._sync_agent_state()
         self.controller.notify_changed()
 
     def capture_mcp_status(self, roster: list[dict[str, Any]]) -> None:
@@ -420,7 +409,7 @@ class GoTuiRuntime:
         # only the Python-level bindings change.
         original_stdout = sys.stdout
         original_stderr = sys.stderr
-        output_sink = Path(os.devnull).open("a", buffering=1)  # noqa: SIM115
+        output_sink = _open_output_sink()
         sys.stdout = output_sink
         sys.stderr = output_sink
         backend_socket: socket.socket | None = None
@@ -428,6 +417,8 @@ class GoTuiRuntime:
         prepare_task: asyncio.Task[None] | None = None
         process: asyncio.subprocess.Process | subprocess.Popen[bytes] | None = None
         try:
+            if not self.controller.setup_mode:
+                self.init_run_state()
             env = child_environment()
             env["STRIX_VERSION"] = package_version()
             command = self.binary_command()

strix/interface/tui/sidecar.py

@@ -177,9 +177,17 @@ async def _launch_windows_tui_process(
     return windows_process, connection
 
 
+class TuiProcessExitedError(RuntimeError):
+    """The Go TUI process ended with a non-zero status."""
+
+    def __init__(self, return_code: int) -> None:
+        super().__init__(f"Bubble Tea TUI exited with status {return_code}")
+        self.return_code = return_code
+
+
 def check_return_code(return_code: int) -> None:
     if return_code != 0:
-        raise RuntimeError(f"Bubble Tea TUI exited with status {return_code}")
+        raise TuiProcessExitedError(return_code)
 
 
 def package_version() -> str:

strix/interface/utils.py

@@ -1599,20 +1599,21 @@ def clone_repository(repo_url: str, run_name: str, dest_name: str | None = None)
 
 
 def check_docker_connection() -> Any:
-    import docker
-    from docker.errors import DockerException
+    from strix.runtime.docker_connection import DockerConnectionError, connect_docker
 
     try:
-        return docker.from_env()
-    except DockerException as exc:
-        report_error("docker_unavailable", exc)
+        return connect_docker()
+    except DockerConnectionError as exc:
+        report_error("docker_unavailable", exc.cause)
         console = Console()
         error_text = Text()
         error_text.append("DOCKER NOT AVAILABLE", style="bold red")
         error_text.append("\n\n", style="white")
-        error_text.append("Cannot connect to Docker daemon.\n", style="white")
+        error_text.append(f"Cannot connect to Docker at {exc.endpoint.label}.\n", style="white")
+        error_text.append(f"{exc.detail}\n\n", style="dim red")
         error_text.append(
-            "Please ensure Docker Desktop is installed and running, and try running strix again.\n",
+            "Make sure Docker is running. If `docker info` works in this shell, strix uses "
+            "the same daemon. Otherwise set DOCKER_HOST.\n",
             style="white",
         )
 
@@ -1624,7 +1625,7 @@ def check_docker_connection() -> Any:
             padding=(1, 2),
         )
         console.print("\n", panel, "\n")
-        raise RuntimeError("Docker not available") from None
+        sys.exit(1)
 
 
 def image_exists(client: Any, image_name: str) -> bool:

strix/report/runs.py

@@ -0,0 +1,86 @@
+"""Summaries of the runs under ./strix_runs, for pickers and listings."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any
+
+from strix.core.paths import run_record_path, runs_base_dir, runtime_state_dir
+
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+
+@dataclass(frozen=True)
+class RunSummary:
+    run_name: str
+    target: str
+    started_at: str
+    ended_at: str
+    status: str
+    findings: int
+    resumable: bool
+
+
+def list_run_summaries(*, cwd: Path | None = None) -> list[RunSummary]:
+    """Every run with a run.json, newest activity first (run.json mtime)."""
+    base = runs_base_dir(cwd=cwd)
+    if not base.is_dir():
+        return []
+    rows: list[tuple[float, RunSummary]] = []
+    for child in base.iterdir():
+        record_path = run_record_path(child)
+        try:
+            if not record_path.is_file():
+                continue
+            modified = record_path.stat().st_mtime
+        except OSError:
+            continue
+        rows.append((modified, _summarize(child, _load_json(record_path, default={}))))
+    rows.sort(key=lambda row: (row[0], row[1].run_name), reverse=True)
+    return [summary for _, summary in rows]
+
+
+def _summarize(run_dir: Path, record: Any) -> RunSummary:
+    if not isinstance(record, dict):
+        record = {}
+    findings = _load_json(run_dir / "vulnerabilities.json", default=[])
+    return RunSummary(
+        run_name=run_dir.name,
+        target=_describe_target(record),
+        started_at=str(record.get("start_time") or ""),
+        ended_at=str(record.get("end_time") or ""),
+        status=str(record.get("status") or "unknown"),
+        findings=len(findings) if isinstance(findings, list) else 0,
+        resumable=(runtime_state_dir(run_dir) / "agents.json").is_file(),
+    )
+
+
+def _describe_target(record: dict[str, Any]) -> str:
+    targets = record.get("targets_info")
+    originals = [
+        str(entry["original"])
+        for entry in targets or []
+        if isinstance(entry, dict) and entry.get("original")
+    ]
+    if originals:
+        if len(originals) == 1:
+            return originals[0]
+        return f"{originals[0]} +{len(originals) - 1} more"
+    mount = record.get("workspace_mount")
+    if isinstance(mount, str) and mount:
+        return f"{mount} (workspace)"
+    instruction = str(record.get("user_instruction") or record.get("instruction") or "").strip()
+    return instruction.splitlines()[0] if instruction else ""
+
+
+def _load_json(path: Path, *, default: Any) -> Any:
+    try:
+        return json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError):
+        return default
+
+
+__all__ = ["RunSummary", "list_run_summaries"]

strix/report/state.py

@@ -657,6 +657,35 @@ class ReportState:
     def record_observed_llm_cost(self, cost: float) -> None:
         self._llm_usage.record_observed_cost(cost)
 
+    def record_llm_provider(
+        self,
+        provider: str,
+        *,
+        agent_id: str | None,
+        input_tokens: int,
+        cached_tokens: int,
+        cost: float,
+    ) -> None:
+        self._llm_usage.record_provider(
+            provider,
+            agent_id=agent_id,
+            input_tokens=input_tokens,
+            cached_tokens=cached_tokens,
+            cost=cost,
+            cache_block_tokens=load_settings().llm.cache_block_tokens,
+        )
+
+    def get_process_llm_providers(self) -> dict[str, dict[str, float]]:
+        """Per-provider usage since this process started, like get_process_llm_usage."""
+        baseline = self._telemetry_llm_usage_baseline.get("providers") or {}
+        providers: dict[str, dict[str, float]] = {}
+        for name, tally in (self._llm_usage.to_record().get("providers") or {}).items():
+            before = baseline.get(name) or {}
+            delta = {key: max(0, value - _number(before.get(key))) for key, value in tally.items()}
+            if delta["requests"]:
+                providers[name] = delta
+        return providers
+
     def get_total_llm_usage(self) -> dict[str, Any]:
         return dict(self.run_record.get("llm_usage") or self._build_llm_usage_record())
 
@@ -990,6 +1019,30 @@ class StreamedOpenRouterCosts:
 streamed_openrouter_costs = StreamedOpenRouterCosts()
 
 
+def record_openrouter_provider(provider: Any, usage: Any) -> None:
+    """Tally which upstream provider served a stream, from its final usage chunk.
+
+    OpenRouter spreads one model across many providers whose prices, quantization
+    and prompt caching differ, so this is what shows where a scan's tokens went.
+    """
+    # Deferred: request_log pulls in the agents SDK, which strix.report must not import.
+    from strix.llm.request_log import current_call_context
+
+    report_state = get_global_report_state()
+    if report_state is None or not isinstance(usage, dict):
+        return
+    details = usage.get("prompt_tokens_details")
+    report_state.record_llm_provider(
+        provider if isinstance(provider, str) and provider else "unknown",
+        agent_id=current_call_context().agent_id,
+        input_tokens=int(_number(usage.get("prompt_tokens"))),
+        cached_tokens=int(_number(details.get("cached_tokens")))
+        if isinstance(details, dict)
+        else 0,
+        cost=openrouter_stream_cost(usage) or 0.0,
+    )
+
+
 def litellm_cost_callback(
     kwargs: Any,
     completion_response: Any,

strix/report/usage.py

@@ -6,6 +6,7 @@ import logging
 from typing import Any
 
 from agents.usage import Usage, deserialize_usage, serialize_usage
+from pydantic import BaseModel, TypeAdapter, ValidationError
 
 from strix.report.pricing import resolve_litellm_model
 
@@ -13,6 +14,22 @@ from strix.report.pricing import resolve_litellm_model
 logger = logging.getLogger(__name__)
 
 
+class ProviderUsage(BaseModel):
+    """Running totals for one upstream provider OpenRouter routed calls to."""
+
+    requests: int = 0
+    input_tokens: int = 0
+    cached_tokens: int = 0
+    cost: float = 0.0
+    # Calls that didn't find the agent's whole previous prompt cached, and the
+    # previous-prompt tokens they had to pay for again.
+    cache_misses: int = 0
+    missed_tokens: int = 0
+
+
+_PROVIDER_USAGE = TypeAdapter(dict[str, ProviderUsage])
+
+
 class LLMUsageLedger:
     """Aggregate SDK ``Usage`` objects and attach best-effort cost estimates."""
 
@@ -23,6 +40,10 @@ class LLMUsageLedger:
         self._observed_cost = 0.0
         self._estimated_cost = 0.0
         self._has_observed_cost = False
+        # Keyed by upstream provider name, e.g. "Z.AI" or "DeepInfra".
+        self._providers: dict[str, ProviderUsage] = {}
+        # Each agent's last prompt size, which its next call should find cached.
+        self._last_input_tokens: dict[str, int] = {}
         # When True, tokens are still tracked but cost stays $0 — the run is on a
         # model subscription, so there is no metered per-token charge to report.
         self.zero_cost = False
@@ -62,6 +83,33 @@ class LLMUsageLedger:
             self._observed_cost += float(cost)
             self._has_observed_cost = True
 
+    def record_provider(
+        self,
+        provider: str,
+        *,
+        agent_id: str | None,
+        input_tokens: int,
+        cached_tokens: int,
+        cost: float,
+        cache_block_tokens: int,
+    ) -> None:
+        tally = self._providers.setdefault(provider, ProviderUsage())
+        tally.requests += 1
+        tally.input_tokens += input_tokens
+        tally.cached_tokens += cached_tokens
+        if agent_id:
+            previous = self._last_input_tokens.get(agent_id, 0)
+            # The previous prompt is a prefix of this one, so all of it but a
+            # partial last block should read back cached. A shrinking prompt means
+            # compaction rewrote it, so a miss is expected.
+            missed = previous - cached_tokens
+            if input_tokens >= previous and missed >= cache_block_tokens:
+                tally.cache_misses += 1
+                tally.missed_tokens += missed
+            self._last_input_tokens[agent_id] = input_tokens
+        if not self.zero_cost:
+            tally.cost = _round_cost(tally.cost + cost)
+
     @property
     def total_cost(self) -> float:
         if self.zero_cost:
@@ -71,6 +119,7 @@ class LLMUsageLedger:
     def to_record(self) -> dict[str, Any]:
         record = serialize_usage(self._total_usage)
         record["cost"] = self.total_cost
+        record["providers"] = {name: tally.model_dump() for name, tally in self._providers.items()}
         record["agents"] = []
 
         agent_tokens = {aid: _resolve_total_tokens(u) for aid, u in self._agent_usage.items()}
@@ -102,10 +151,16 @@ class LLMUsageLedger:
         self._observed_cost = 0.0
         self._estimated_cost = 0.0
         self._has_observed_cost = False
+        self._providers = {}
 
         if not isinstance(raw_usage, dict):
             return
 
+        try:
+            self._providers = _PROVIDER_USAGE.validate_python(raw_usage.get("providers") or {})
+        except ValidationError:
+            logger.exception("Failed to hydrate llm_usage providers from run.json")
+
         try:
             self._total_usage = deserialize_usage(raw_usage)
         except Exception:

strix/runtime/backends.py

@@ -27,9 +27,10 @@ async def _docker_backend(
     """Bring up a session backed by the local Docker daemon.
 
     Uses :class:`StrixDockerSandboxClient` to inject NET_ADMIN /
-    NET_RAW caps + ``host.docker.internal`` host-gateway. Imports
-    ``docker`` lazily so deployments that target a non-Docker
-    backend don't need the docker-py library installed.
+    NET_RAW caps + ``host.docker.internal`` host-gateway, on the same
+    endpoint the startup check resolved (``DOCKER_HOST``, docker context,
+    default socket). Imports ``docker`` lazily so deployments that target
+    a non-Docker backend don't need the docker-py library installed.
 
     ``session.start()`` is what materializes the manifest into the running
     container — the SDK's ``client.create()`` only builds the inner session
@@ -37,12 +38,12 @@ async def _docker_backend(
     Strix manages session lifetime explicitly via ``client.delete()`` so we
     trigger ``start()`` ourselves.
     """
-    import docker
     from agents.sandbox.sandboxes.docker import DockerSandboxClientOptions
 
     from strix.runtime.docker_client import StrixDockerSandboxClient
+    from strix.runtime.docker_connection import connect_docker
 
-    client = StrixDockerSandboxClient(docker.from_env())
+    client = StrixDockerSandboxClient(connect_docker())
     client.strix_bind_mounts = bind_mounts or []
     options = DockerSandboxClientOptions(image=image, exposed_ports=exposed_ports)
     session = await client.create(options=options, manifest=manifest)

strix/runtime/docker_connection.py

@@ -0,0 +1,90 @@
+from __future__ import annotations
+
+import os
+from dataclasses import dataclass
+from typing import Any
+
+import docker  # type: ignore[import-untyped, unused-ignore]
+from docker.context import ContextAPI  # type: ignore[import-untyped, unused-ignore]
+from docker.context.config import (  # type: ignore[import-untyped, unused-ignore]
+    get_current_context_name,
+)
+from docker.errors import DockerException  # type: ignore[import-untyped, unused-ignore]
+
+
+DEFAULT_CONTEXT = "default"
+
+
+@dataclass(frozen=True)
+class DockerEndpoint:
+    """Where strix talks to Docker and how that address was chosen."""
+
+    host: str | None
+    source: str
+    tls: Any = None
+
+    @property
+    def label(self) -> str:
+        return f"{self.host} ({self.source})" if self.host else self.source
+
+
+class DockerConnectionError(RuntimeError):
+    """The daemon at ``endpoint`` could not be reached."""
+
+    def __init__(self, endpoint: DockerEndpoint, cause: BaseException) -> None:
+        self.endpoint = endpoint
+        self.cause = root_cause(cause)
+        super().__init__(f"Cannot connect to Docker at {endpoint.label}: {self.detail}")
+
+    @property
+    def detail(self) -> str:
+        text = str(self.cause).strip()
+        name = type(self.cause).__name__
+        return text if name in text else f"{name}: {text}" if text else name
+
+
+def resolve_docker_endpoint(environ: dict[str, str] | None = None) -> DockerEndpoint:
+    """``DOCKER_HOST``, else the current docker context, else the SDK default."""
+    env = os.environ if environ is None else environ
+    host = env.get("DOCKER_HOST", "").strip()
+    if host:
+        return DockerEndpoint(host, "DOCKER_HOST")
+
+    name = get_current_context_name()
+    if name == DEFAULT_CONTEXT:
+        return DockerEndpoint(None, "default socket")
+
+    source = f"docker context '{name}'"
+    try:
+        context = ContextAPI.get_context(name)
+    except DockerException as exc:
+        raise DockerConnectionError(DockerEndpoint(None, source), exc) from exc
+    if context is None:
+        missing = DockerException(f"{source} is selected but does not exist")
+        raise DockerConnectionError(DockerEndpoint(None, source), missing)
+    return DockerEndpoint(context.Host, source, context.TLSConfig)
+
+
+def connect_docker() -> Any:
+    """Return a ``docker.DockerClient`` for the resolved endpoint or raise DockerConnectionError."""
+    endpoint = resolve_docker_endpoint()
+    try:
+        if endpoint.source == "DOCKER_HOST" or endpoint.host is None:
+            return docker.from_env()  # also reads DOCKER_TLS_VERIFY and DOCKER_CERT_PATH
+        return docker.DockerClient(base_url=endpoint.host, tls=endpoint.tls or False)
+    except DockerException as exc:
+        raise DockerConnectionError(endpoint, exc) from exc
+
+
+def root_cause(exc: BaseException) -> BaseException:
+    """The innermost exception: the SDK wraps the OSError in requests/urllib3 errors."""
+    current = exc
+    seen: list[BaseException] = []
+    while not any(current is s for s in seen):
+        seen.append(current)
+        nested = [arg for arg in current.args if isinstance(arg, BaseException)]
+        following = current.__cause__ or current.__context__ or (nested[-1] if nested else None)
+        if following is None:
+            break
+        current = following
+    return current

strix/skills/cloud/aws.md

@@ -94,6 +94,8 @@ curl https://BUCKET.s3.amazonaws.com/
 
 ### IAM Privilege Escalation
 
+**Resource control policies:** AWS Organizations RCPs impose an organization-level limit on supported resources, including access by external principals. Map resource-account RCPs as well as principal-account SCPs, identity/resource policies, session policies, and permission boundaries when assessing a cross-account or leaked-key path. RCPs do not grant access and do not apply uniformly to every service; an identity-policy allow alone does not describe the effective permission ([AWS RCP documentation](https://docs.aws.amazon.com/organizations/latest/userguide/orgs_manage_policies_rcps.html)).
+
 Common escalation paths (verify with `aws iam simulate-principal-policy` when possible):
 
 | Permission | Escalation |

strix/skills/cloud/kubernetes.md

@@ -32,6 +32,10 @@ Kubernetes clusters expose a large attack surface through their API server, kube
 
 ## Key Vulnerabilities
 
+### Ingress-NGINX Admission
+
+Inventory validating-webhook reachability from the pod network and the controller's service-account permissions. IngressNightmare illustrates controller execution without a Kubernetes account ([advisory](https://kubernetes.io/blog/2025/03/24/ingress-nginx-cve-2025-1974/)). Identify the controller implementation/image, then verify affected builds, backports, and maintenance status against current project or vendor notices; distinguish community ingress-nginx from other NGINX controllers and the Ingress API ([project notice](https://kubernetes.io/blog/2026/01/29/ingress-nginx-statement/)).
+
 ### RBAC Misconfigurations
 
 - Wildcard verbs or resources in ClusterRole/Role bindings: `verbs: ["*"]`, `resources: ["*"]`

strix/skills/frameworks/django.md

@@ -57,6 +57,11 @@ Map endpoints, authentication classes, and permission classes per route.
 
 ## Key Vulnerabilities
 
+### Request and Spatial Input Handling
+
+- Under ASGI, compare underscore/hyphen forms of proxy-injected identity headers through `ASGIRequest` normalization into `request.META`. Resolve the installed branch, vendor backports, and support status from package metadata and current Django release/advisory information ([header-collision advisory](https://www.djangoproject.com/weblog/2026/apr/07/security-releases/)).
+- For GeoDjango, trace attacker-controlled spatial lookup strings/dictionaries into `GDALRaster`: driver behavior can turn them into server-side requests or file writes. Include custom filter APIs and admin changelist filtering; direct model-field assignment is a separate input path ([spatial lookup advisory](https://www.djangoproject.com/weblog/2026/aug/04/security-releases/)).
+
 ### Authentication & Authorization
 
 **Permission Class Gaps**
@@ -207,7 +212,7 @@ Static analysis is the fastest way to reach the sinks above in white-box scope.
 - **pip-audit** (PyPA) — dependency CVE scanner for known-vuln Django/DRF/simplejwt versions: `pipx install pip-audit && pip-audit -r requirements.txt`
 - **ast-grep** (preinstalled) — quick structural grep for risky calls without a full SAST run: `ast-grep run -p 'mark_safe($X)' -l python`
 
-For the `SECRET_KEY` → signed-cookie/reset-token forgery path noted under Session Issues, Django's own `django.core.signing` is the "tool": with a leaked key you can mint valid `signing.dumps()` values (session cookies, password-reset tokens, and `PickleSerializer`-backed session RCE).
+For the `SECRET_KEY` → signed-cookie forgery path noted under Session Issues, Django's own `django.core.signing` is the "tool": with a leaked key you can mint valid `signing.dumps()` values using the consumer's serializer and signing salt. Inspect `SESSION_SERIALIZER` and the installed implementation: the session-RCE path requires a reachable pickle-backed serializer; JSON sessions do not provide it. Password-reset tokens instead use `PasswordResetTokenGenerator`, with user state and timestamp in the digest, rather than the generic `signing.dumps()` format ([Django 5.0 removals](https://docs.djangoproject.com/en/5.0/releases/5.0/#features-removed-in-5-0), [token implementation](https://github.com/django/django/blob/stable/5.2.x/django/contrib/auth/tokens.py)).
 
 ## Summary
 

strix/skills/frameworks/fastapi.md

@@ -58,6 +58,12 @@ For each route, identify:
 
 ## Key Vulnerabilities
 
+### Starlette Request Handling
+
+Resolve Starlette's version independently of FastAPI and identify the ASGI server/front proxy. Check current upstream advisories and any vendor backports for the installed build before treating a parser issue as applicable. Compare middleware authorization against the actual routed path: malformed Host values can alter reconstructed `request.url` without changing routing ([URL parsing advisory](https://github.com/Kludex/starlette/security/advisories/GHSA-86qp-5c8j-p5mr)).
+
+For `request.form()`, test URL-encoded and multipart limits separately; a limit enforced on one parser may not constrain the other. Check whether crossing an upload's memory-to-disk spool threshold blocks the event loop ([form limits](https://github.com/Kludex/starlette/security/advisories/GHSA-82w8-qh3p-5jfq), [file spooling](https://github.com/Kludex/starlette/security/advisories/GHSA-2c2j-9gv5-cj73)).
+
 ### Authentication & Authorization
 
 **Dependency Injection Gaps**

strix/skills/frameworks/nestjs.md

@@ -78,6 +78,13 @@ For each controller and method, identify:
 
 ## Key Vulnerabilities
 
+### Adapter and Schema Version Boundaries
+
+- **Express adapter:** resolve the installed Nest/Express versions and effective query-parser setting from configuration or a harmless nested-query probe. Apply `qs`/nested-operator attacks only when the parser constructs nested objects. Check the matching router documentation and test both root and child paths against authentication middleware ([migration reference](https://nestjs.io/tutorials/what-s-new-in-express-5-eeb51579)).
+- **Fastify adapter:** inspect the installed version's schema requirements and any custom validator configuration. Verify whether the route schema is actually enforced before relying on a DTO or shorthand-schema declaration ([migration reference](https://fastify.dev/docs/v5.0.x/Guides/Migration-Guide-V5/)).
+- **Standard Schema, where supported:** `@Body({ schema })`, `@Query({ schema })`, `@Param(..., { schema })`, and `@RawBody({ schema })` attach metadata; enforcement requires `StandardSchemaValidationPipe`. Probe schema-decorated routes for accepted invalid values when the pipe is absent or transport-local. Outgoing schema enforcement uses `StandardSchemaSerializerInterceptor`; check actual output rather than relying only on class-transformer decorators.
+- **GraphQL transport:** inspect installed packages, enabled IDE, and negotiated WebSocket subprotocol. Verify supported/deprecated transports against the matching Nest documentation, then recheck connection and per-operation authentication on the deployed transport ([migration guide](https://docs.nestjs.com/migration-guide)).
+
 ### Guard Bypass
 
 **Decorator Stack Gaps**

strix/skills/frameworks/nextjs.md

@@ -12,7 +12,7 @@ Security testing for Next.js applications. Focus on authorization drift across r
 **Routers**
 - App Router (`app/`) and Pages Router (`pages/`) often coexist
 - Route Handlers (`app/api/**`) and API routes (`pages/api/**`)
-- Middleware: `middleware.ts` at project root
+- Middleware/proxy: inspect `middleware.ts`, `proxy.ts`, installed Next.js metadata, and runtime configuration. Check the matching framework docs for supported filenames, Node/Edge behavior, and feature status before selecting runtime-specific probes ([migration guide](https://nextjs.org/docs/app/guides/upgrading/version-16)).
 
 **Runtimes**
 - Node.js (full API access)
@@ -43,6 +43,8 @@ Security testing for Next.js applications. Focus on authorization drift across r
 
 ## Reconnaissance
 
+At each assessment, establish the deployed build from package metadata, lockfiles, or runtime evidence. Verify release/support status and applicable fixes through current official docs, advisories, or upstream source; linked advisories are starting points, not a complete or permanently current list. Recheck when the target build or proposed remediation changes, and mark status unverified if evidence is unavailable.
+
 **Route Discovery**
 
 ```javascript
@@ -83,10 +85,15 @@ Inspect Network tab for POST requests with `Next-Action` header. Extract action
 
 ## Key Vulnerabilities
 
+### RSC and Image Processing
+
+- **React2Shell (CVE-2025-55182):** App Router applications can expose the vulnerable RSC decoder without explicit Server Actions. Identify the bundled `react-server-dom-*` implementation and trace requests into decoding; client-only React and Pages-only apps have different exposure. Check the framework's patched branch, including source-disclosure and DoS fixes ([RCE advisory](https://react.dev/blog/2025/12/03/critical-security-vulnerability-in-react-server-components), [RSC advisories](https://react.dev/blog/2025/12/11/denial-of-service-and-source-code-exposure-in-react-server-components)).
+- Treat `/_next/image` optimization and Node `next/og` `ImageResponse` as separate input paths. Trace attacker-controlled image bytes into native decoders, and SVG content/attributes/styles into image construction. URL allowlisting does not make image contents trusted; match the decoder, runtime, and Next.js version to the relevant advisory ([optimizer](https://nextjs.org/blog/august-2026-security-release), [ImageResponse](https://github.com/vercel/next.js/security/advisories/GHSA-vcvr-r3jv-pc5j)).
+
 ### Middleware Bypass
 
 **Known Techniques**
-- `x-middleware-subrequest` header crafting (CVE-class bypass)
+- `x-middleware-subrequest` header crafting (CVE-2025-29927): middleware-only authorization can be skipped on affected deployments. Resolve the installed branch and hosting protections from the advisory and provider configuration, then check whether external headers reach the origin and the destination enforces authorization independently ([advisory](https://github.com/vercel/next.js/security/advisories/GHSA-f82v-jwr5-mffw)).
 - `x-nextjs-data` probing
 - Look for 307 + `x-middleware-rewrite`/`x-nextjs-redirect` headers
 
@@ -118,7 +125,7 @@ Middleware checks first value, handler uses last or array.
 **Cache Boundary Failures**
 - User-bound data cached without identity keys (ETag/Set-Cookie unaware)
 - Personalized content served from shared cache/CDN
-- Missing `no-store` on sensitive fetches
+- Missing `no-store` on sensitive fetches matters only when they enter a shared cache. Determine `fetch` and GET Route Handler defaults for the deployed version from docs/configuration and observed responses. Where Cache Components are enabled, inspect `use cache` arguments/closed-over values, `cacheTag`, `cacheLife`, and invalidation for user/tenant separation ([caching documentation](https://nextjs.org/docs/app/getting-started/cache-components)).
 
 **Flight Data Leakage**
 

strix/skills/protocols/graphql.md

@@ -134,6 +134,10 @@ Parser precedence varies; may bypass validation. Also test default argument valu
 
 Send unexpected keys in input objects; backends may pass them to resolvers or downstream logic.
 
+### OneOf Inputs
+
+Discover `@oneOf` input types through the schema or `__Type.isOneOf`. They require exactly one non-null field, with every other field omitted. Test both inline literals and variables with zero fields, two selectors, and an extra null-valued selector; all must fail coercion. Then test authorization independently for each valid selector (ID, username, organization/email), since schema exclusivity does not ensure identical tenant checks in each resolver branch. Confirm the deployed implementation supports OneOf before treating acceptance as a specification violation ([GraphQL September 2025](https://spec.graphql.org/September2025/#sec-OneOf-Input-Objects)).
+
 ### Cursor Manipulation
 
 Decode cursors (usually base64) to:

strix/skills/protocols/oauth.md

@@ -77,7 +77,7 @@ com.app://callback  (mobile custom scheme)
 
 ### State and Nonce
 
-- Missing, predictable, or reusable `state` → CSRF on OAuth login (session fixation, account linking)
+- Missing, predictable, or reusable `state` → test CSRF on OAuth login (session fixation, account linking); RFC 9700 also permits correctly bound PKCE, or OIDC `nonce`, to provide CSRF protection, so establish whether that protection survives a cross-session callback
 - Missing `nonce` in OIDC → ID token injection/replay
 - `state` not bound to client session or PKCE verifier
 
@@ -102,7 +102,7 @@ com.app://callback  (mobile custom scheme)
 ### Scope and Token Issues
 
 - Scope escalation: request `admin`/`offline_access`/`openid profile email` beyond app need; server grants all requested scopes
-- Refresh token not rotated or reuse not detected → persistent access
+- Public-client refresh tokens neither sender-constrained nor rotated with reuse detection → persistent access; a non-rotating token bound to the client's key is permitted by RFC 9700, so test replay without that key
 - Access token accepted across services (missing audience/resource binding)
 - Token introspection returns `active:true` without proper auth on introspection endpoint
 
@@ -113,6 +113,10 @@ com.app://callback  (mobile custom scheme)
 - Userinfo endpoint returns PII without matching access token scope
 - `sub` collision across issuers if `iss` not validated
 
+### OAuth Security BCP (RFC 9700)
+
+Authorization servers must support PKCE, public clients must use it, and confidential clients are also recommended to use it. Test challenge stripping and verifier injection across both client types. Resource-owner-password grants must not be used; implicit access-token responses are discouraged because tokens can leak or be replayed. For mix-up defenses, bind the selected issuer and endpoints to the authorization transaction; validate an authorization-response `iss` when that defense is used, rather than merely checking the issuer of a later ID token ([RFC 9700](https://www.rfc-editor.org/rfc/rfc9700.html)).
+
 ## Advanced Techniques
 
 **Referer Leakage**

strix/skills/technologies/active_directory.md

@@ -100,6 +100,13 @@ certipy find -u <USER>@<DOMAIN> -p <PASS> -dc-ip <DC_IP> -vulnerable -stdout
 - **ESC8** — NTLM relay to the CA web-enrollment endpoint (coerce a DC, relay to `/certsrv`) → DC certificate → DCSync.
 - **ESC others** — ESC2/3 (any-purpose/enrollment-agent), ESC4 (writable template DACL → make it ESC1), ESC6 (`EDITF_ATTRIBUTESUBJECTALTNAME2` on the CA), ESC7 (CA officer rights), ESC9/10 (weak cert mapping), ESC11 (RPC relay), ESC13 (issuance-policy→group), ESC15 (app-policy on v1 templates). `certipy find -vulnerable` flags each.
 
+**Strong certificate mapping:** establish the DC patch level, mapping policy, and vendor backports; consult Microsoft's current enforcement guidance before assuming Compatibility mode is available. Inspect template SID extensions, explicit mappings, and the principal selected at authentication. A requested privileged UPN alone does not establish a working ESC chain ([KB5014754](https://support.microsoft.com/en-us/servicing/os/windows-server/2022/05/kb5014754-certificate-based-authentication-changes-on-windows-domain-controllers)).
+
+### Windows Server 2025 dMSA / BadSuccessor
+
+- **CVE-2025-53779:** on unpatched Server 2025 DCs, control sufficient to create/modify a delegated Managed Service Account can establish a one-way migration link to another principal and obtain its authority/keys through the KDC. Enumerate dMSA creation rights, object ACLs, and migration attributes rather than limiting discovery to conventional service accounts and delegation flags.
+- **Patched behavior:** the KDC requires mutual dMSA↔target linkage. Writing the dMSA-side attribute still succeeds, so LDAP write success is not proof of exploitation. Determine whether the tester also controls the target object's reciprocal link and whether the KDC actually issues the relevant ticket. Post-patch abuse is a different prerequisite chain from the original low-privilege OU-control escalation ([researchers' patch analysis](https://www.akamai.com/blog/security-research/badsuccessor-is-dead-analyzing-badsuccessor-patch)).
+
 ### NTLM Coercion & Relay
 
 Force a privileged machine to authenticate to you, then relay that NTLM auth to a service that doesn't enforce signing/EPA (LDAP, AD CS, SMB).

strix/skills/technologies/auth0.md

@@ -86,12 +86,14 @@ Authorization: Bearer <token_with_audience_A>
 
 ### Rules and Actions Abuse
 
+Inventory which Rules, Hooks, and Actions actually execute in the tenant. Check the current Auth0 lifecycle notice and tenant capabilities for retirement or read-only restrictions. During migration, compare claim assignment, MFA, denial decisions, and account-linking checks across every connection; distinguish secret/configuration access from source-code modification ([lifecycle notice](https://auth0.com/docs/troubleshoot/product-lifecycle/deprecations-and-migrations#rules-and-hooks-deprecations)).
+
 **Post-Login Rule/Action Injection**
 - Rules that add claims based on unvalidated user metadata:
   ```javascript
   user.app_metadata.role = 'admin'  // if user can set app_metadata via signup/API
   ```
-- `context.authorization` manipulation in Actions
+- Rules use `context`; Actions receive `event` and enforce token/MFA/denial changes through `api` methods. Test migrated checks that only mutate event data: those mutations do not propagate to other Actions or substitute for the corresponding enforcement API ([migration behavior](https://auth0.com/docs/customize/actions/migrate/migrate-from-rules-to-actions)).
 - Secrets in Rule code exposed to tenant admins or via Management API leak
 
 **Signup / Registration Actions**

strix/skills/technologies/grafana_prometheus.md

@@ -79,6 +79,13 @@ Unauthenticated view (and, with `public_mode`, delete) of the lowest-key snapsho
 ### Prometheus / Alertmanager — exposure is the vuln (no auth by default)
 Prometheus and Alertmanager ship with **no authentication**; the docs explicitly say do not expose them. There is rarely a CVE — reachability itself is the finding, and the payoff is recon + credential leakage + pivoting (below).
 
+## Plugin and MCP Boundaries
+
+Resolve installed Grafana, plugin, and MCP server builds separately; check current upstream advisories and vendor backports before applying a listed attack path or recommending a release.
+
+- Plugin archives are extracted before signature verification. Test chained symlinks and containment before trusting a signature failure to prevent filesystem writes; archive installation is the required trigger ([extraction advisory](https://grafana.com/security/security-advisories/cve-2026-15815/)).
+- Inventory mcp-grafana separately from Grafana. Trace `grafana_api_request` and `X-Grafana-URL` into destination selection; preventing token forwarding alone does not prevent SSRF ([MCP advisory](https://grafana.com/security/security-advisories/cve-2026-19516/)).
+
 ## Pivoting: Observability → Deeper Compromise
 
 This is the core value. Chain each exposure into something that matters. Always articulate the pivot in the finding, not just the exposed endpoint.

strix/skills/technologies/llm_applications.md

@@ -78,11 +78,11 @@ Record both forward and reverse reachability: attacker-controlled input to privi
 
 ## Optional Tool Routing
 
-Use tools only when they match the deployed surface. Treat generated cases and scanner labels as leads until the application-side boundary is validated.
+Use tools only when they match the deployed surface. Resolve compatible versions, runtime requirements, maintenance status, and advisories from registry metadata and current upstream documentation, then pin the exact reviewed versions for the run. Treat generated cases and scanner labels as leads until the application-side boundary is validated.
 
-- **[Promptfoo](https://github.com/promptfoo/promptfoo)** — use for repeatable model/application trials, custom adversarial cases, graders, provider comparisons, and success-rate regression. Install the reviewed version locally with `npm install --save-dev --save-exact promptfoo@0.122.0`, then invoke `./node_modules/.bin/promptfoo redteam run`. Define explicit plugins, assertions, `numTests`, `maxConcurrency`, and `delay`; provider calls may transmit test data and incur cost. Its `owasp:llm` preset still uses the 2025 category mapping in version 0.122.0, so build or select tests from the 2026 matrix above and do not present the preset report as complete 2026 coverage.
-- **[MCP Inspector](https://github.com/modelcontextprotocol/inspector)** — use for LLM01/LLM03 surface mapping when MCP servers are present. Install the reviewed version with `npm install --save-dev --save-exact @modelcontextprotocol/inspector@2.2.0`, then use `./node_modules/.bin/mcp-inspector --cli --config <reviewed-config> --server <name> --method tools/list` and the equivalent `resources/list` / `prompts/list` operations. Starting a stdio server executes that configured process, initialization/list handlers may have side effects, and `tools/call` can perform the real action; inspect the target and credentials before invoking it.
-- **[ModelScan](https://github.com/protectai/modelscan)** — use for LLM04 static triage of supported H5, Pickle, and SavedModel artifacts before loading them, for example `uvx modelscan==0.8.8 -p <artifact>`. Run it as an untrusted-file parser in an isolated analysis environment. A clean result covers only the scanner's supported formats and signatures; it does not establish artifact provenance, integrity, or absence of behavioral backdoors.
+- **[Promptfoo](https://github.com/promptfoo/promptfoo)** — use for repeatable model/application trials, custom adversarial cases, graders, provider comparisons, and success-rate regression. Install the reviewed version locally with `npm install --save-dev --save-exact promptfoo@<reviewed-version>`, then invoke `./node_modules/.bin/promptfoo redteam run`. Define explicit plugins, assertions, `numTests`, `maxConcurrency`, and `delay`; provider calls may transmit test data and incur cost. Inspect the installed `owasp:llm` preset's category mapping and actual tests before claiming coverage; add cases for gaps against the assessment's chosen taxonomy.
+- **[MCP Inspector](https://github.com/modelcontextprotocol/inspector)** — use for LLM01/LLM03 surface mapping when MCP servers are present. Install the reviewed version with `npm install --save-dev --save-exact @modelcontextprotocol/inspector@<reviewed-version>`, then use `./node_modules/.bin/mcp-inspector --cli --config <reviewed-config> --server <name> --method tools/list` and the equivalent `resources/list` / `prompts/list` operations. Starting a stdio server executes that configured process, initialization/list handlers may have side effects, and `tools/call` can perform the real action; inspect the target and credentials before invoking it.
+- **[ModelScan](https://github.com/protectai/modelscan)** — use for LLM04 static triage of supported H5, Pickle, and SavedModel artifacts before loading them, for example `uvx modelscan==<reviewed-version> -p <artifact>`. Run it as an untrusted-file parser in an isolated analysis environment. A clean result covers only the scanner's supported formats and signatures; it does not establish artifact provenance, integrity, or absence of behavioral backdoors.
 
 ## LLM01:2026 Prompt Injection
 

strix/skills/technologies/supabase.md

@@ -35,7 +35,7 @@ Security testing for Supabase applications. Focus on mis-scoped Row Level Securi
 - Functions: `https://<ref>.functions.supabase.co/`
 
 **Headers**
-- `apikey: <anon-or-service>` — identifies project
+- `apikey: <publishable-or-secret-key>` — identifies the application component; legacy `anon` / `service_role` JWT keys can still coexist
 - `Authorization: Bearer <JWT>` — binds user context
 
 **Roles**
@@ -45,6 +45,13 @@ Security testing for Supabase applications. Focus on mis-scoped Row Level Securi
 **Key Principle**
 `auth.uid()` returns current user UUID from JWT. Policies must never trust client-supplied IDs over server context.
 
+### API Keys and Signing Keys
+
+- Opaque keys start with `sb_publishable_` or `sb_secret_`; JWT decoding will not classify them. Search bundles, server configuration, and responses for both formats. Publishable keys are intended for public clients; secret keys use `service_role` and bypass RLS. A browser 401 for a leaked secret key is not revocation: the browser restriction uses `User-Agent` and does not stop server-side use.
+- Creating new keys does not disable legacy keys. After migration, test whether the old `anon` / `service_role` credential remains accepted; dashboard revocation is a separate action. Verify supported key formats and migration/deprecation status from project settings and current Supabase documentation ([API keys](https://supabase.com/docs/guides/getting-started/api-keys)).
+- Edge Functions use user JWTs in `Authorization` and API keys in `apikey`. Verify the deployed platform's `verify_jwt` behavior with current docs and controlled requests, including API keys on either header; passing a platform check with a publishable key does not establish a signed-in user. Test whether the handler verifies user identity before using a privileged client. Conversely, `verify_jwt=false` with explicit secret-key or webhook verification is not necessarily public. Trace the configured `@supabase/server` auth mode or custom verification ([function authorization](https://supabase.com/docs/guides/functions/auth-headers)).
+- Asymmetric signing keys are discoverable at `/auth/v1/.well-known/jwks.json`. Test custom API/Edge Function validators for issuer/algorithm binding and stale cached keys after rotation/revocation. Measure cache lifetimes from response headers, SDK configuration, and controlled rotation tests; check service-specific revocation behavior in current documentation rather than assuming platform and custom validators share a cache ([signing keys](https://supabase.com/docs/guides/auth/signing-keys)).
+
 ## High-Value Targets
 
 - Tables with sensitive data (users, orders, payments, PII)

strix/skills/vulnerabilities/agentic_system_security.md

@@ -106,10 +106,16 @@ Classify each discovered integration by data read, data write, external communic
 - For each tool, validate the same authorization and argument checks through every supported transport.
 - Treat server-launched subprocess configuration, environment variables, and working directories as sensitive executable configuration.
 - For HTTP/SSE transports, validate OAuth issuer, signature, expiry, audience/resource, tenant, and scope claims at the server boundary. Reject tokens minted for the wrong audience, and do not treat a session ID as identity.
-- For downstream APIs, do not pass through the same bearer token unless the target explicitly authorizes that audience and principal. Separate upstream MCP authentication from downstream target authorization.
+- MCP forbids token passthrough: reject access tokens not issued for the MCP server, and use a separate downstream authorization flow instead of forwarding a client's token to another API ([MCP authorization](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization)).
 - For browser or loopback OAuth, review redirect URI, state/PKCE handling, localhost binding, and consent proxying. Treat metadata fetches and tool discovery on remote servers as SSRF-relevant surfaces.
 - For stdio servers, the launch command and environment are already code execution. Discovery must not execute an unreviewed server binary or mutable package tag.
 
+#### MCP Protocol Boundaries
+
+- Identify the protocol revision and client/server SDK builds from configuration and wire traffic, then consult the matching specification and current SDK advisories. Determine whether the transport uses initialization/session IDs or independent requests with client metadata and discovery. Test authorization on every request; client-supplied identity/capabilities in `_meta` are not authenticated identity.
+- Where multi-round tool requests use `input_required` and `inputResponses`, trace the retry/resume binding. Test whether a response from another user, tool invocation, or approval round can be substituted, and whether a retry repeats a consequential action ([MCP protocol specification](https://blog.modelcontextprotocol.io/posts/2026-07-28/)).
+- For OAuth callbacks, bind the expected issuer from validated discovery to the PKCE transaction. A present `iss` must match exactly, even if metadata did not advertise support; an absent `iss` must be rejected when `authorization_response_iss_parameter_supported` is true. Test issuer changes between discovery, registration, and callback, including error callbacks. Bind client credentials to their authorization-server issuer. Require the MCP server's `resource` in authorization and token requests and validate the token's audience at the server ([authorization requirements](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization)).
+
 ### Executable Component Supply Chain
 
 Every skill, plugin, MCP server, model adapter, package, and update channel is an executable or behavior-shaping dependency. Record:
@@ -158,7 +164,7 @@ npx @modelcontextprotocol/inspector@<reviewed-version> --cli \
   --method tools/list --format json
 ```
 
-- Current upstream requirements should be checked before pinning; as of August 12, 2026, MCP Inspector 2.1.0 requires Node.js `>=22.19.0`.
+- Before choosing MCP Inspector, inspect package-registry metadata, engine requirements, release notes, and advisories. Select a compatible reviewed version and pin that exact version for the run; do not infer safety from a release tag.
 - Prefer CLI/TUI and loopback binding over exposing the web UI.
 - Preserve the generated API token; never disable authentication or bind the process-spawning backend to an external interface.
 - Do not publish ports 6274/6277 or pass through the Docker socket/host devices.
@@ -173,7 +179,7 @@ npx @modelcontextprotocol/inspector@<reviewed-version> --cli \
 npx promptfoo@<reviewed-version> eval
 ```
 
-- Current upstream engine constraints should be checked before pinning; as of August 12, 2026, Promptfoo documents Node.js `^20.20.0` or `>=22.22.0`.
+- Resolve Promptfoo's runtime requirements, supported features, and security status from registry metadata and current upstream docs before choosing and pinning a version.
 - Use synthetic prompts/data and a dedicated test provider/project.
 - Provider calls transmit data externally and can incur cost even when evaluation orchestration is local. Set request/concurrency and spending ceilings.
 - Pin model, provider, prompt, tool schema, retrieval corpus revision, and evaluator versions.

strix/skills/vulnerabilities/browser_security.md

@@ -104,6 +104,12 @@ Prove the strongest reliable capability first. If escalation requires a user ges
 - Treat scriptless disclosure of a nonce or trusted URL as a primitive; prove a second controllable sink before claiming bypass.
 - For response splitting, consider whether a same-origin endpoint can be turned into a script resource with a controlled body length or framing.
 
+### Local Network and Loopback Access
+
+Identify the browser build, enabled flags, permission state, and enterprise policy; consult its current Local Network Access documentation and verify actual behavior with controlled requests. Test DNS rebinding and browser-to-local-service chains in both permission-denied and permission-granted states. A successful server-side request does not establish browser reachability ([Chrome LNA](https://developer.chrome.com/blog/local-network-access)).
+
+Discover which permission names and aliases the target supports, including `local-network`, `loopback-network`, and `local-network-access`. Record which destination class was permitted before claiming access to both LAN and localhost services ([permission reference](https://developer.chrome.com/release-notes/145#local_network_access_split_permissions)).
+
 ### JavaScript Gadget Discovery
 
 - When direct calls are blocked, inspect implicit coercions (`toString`, `valueOf`, iterators, getters, proxies) and callbacks invoked by accessible library functions.

strix/skills/vulnerabilities/http_request_smuggling.md

@@ -97,6 +97,12 @@ transfer-encoding: chunked
 SMUGGLED
 ```
 
+### CL.0, 0.CL, and Double Desync
+
+In CL.0, the front end honors `Content-Length` while the back end ignores the body; in 0.CL the front end ignores it while the back end expects it. The latter commonly deadlocks until an early-response gadget responds without consuming the body and keeps the connection open. Test redirects and early errors for that specific connection behavior. A subsequent controlled request can expose the boundary shift; two successive desyncs can convert 0.CL into CL.0. A lone timeout or 400 does not establish that chain.
+
+Include `Expect: 100-continue` handling and interim/final response sequencing in differential tests. Calculate offsets from what the back end receives, including proxy-added headers. Use controlled follow-up requests to establish which bytes are consumed and which response belongs to each request ([HTTP/1.1 must die](https://portswigger.net/research/http1-must-die), [0.CL walkthrough](https://portswigger.net/blog/http-1-1-must-die-conquering-the-0-cl-challenge)).
+
 ## Key Vulnerabilities
 
 ### Front-End Security Control Bypass

strix/skills/vulnerabilities/insecure_deserialization.md

@@ -79,6 +79,8 @@ JNDI injection is not itself a serialization format. It becomes part of this wor
 
 ### Python Pickle
 
+**Model checkpoints:** `torch.load(..., weights_only=True)` does not rule out parser/storage memory corruption. Resolve the installed loader build and check current upstream advisories/backports before trusting that flag. Trace untrusted checkpoints through the exact loader and inspect safe-global allowlists ([example advisory](https://github.com/pytorch/pytorch/security/advisories/GHSA-63cw-57p8-fm3p)).
+
 Pickle executes arbitrary code during unpickling by design:
 ```python
 import pickle, os, base64
@@ -101,13 +103,15 @@ When `yaml.load` used instead of `yaml.safe_load`.
 - POP chains through framework classes (Laravel, Symfony, WordPress plugins)
 
 **Phar Deserialization**
-- Upload or reference `phar://` wrapper triggering metadata deserialization on file operations
+- Trace `phar://` file operations and explicit `Phar::getMetadata()` / `PharFileInfo::getMetadata()` calls. Verify the installed PHP version's metadata-deserialization behavior and `allowed_classes` handling from source/docs; archive opening alone does not establish a deserialization sink ([migration reference](https://www.php.net/manual/en/migration80.incompatible.php#migration80.incompatible.phar)).
 
 ### .NET Deserialization
 
 **BinaryFormatter / LosFormatter**
 - Never safe on untrusted input; full RCE with known gadget chains (ysoserial.net)
 
+Inspect the target framework and resolved serialization packages to establish whether `BinaryFormatter` executes, throws, or is restored through a compatibility package. Check the matching runtime documentation before choosing gadget chains; `LosFormatter` and other serializers are separate surfaces ([Microsoft guide](https://learn.microsoft.com/en-us/dotnet/standard/serialization/binaryformatter-migration-guide/)).
+
 **Json.NET**
 ```json
 {"$type":"System.Windows.Data.ObjectDataProvider, PresentationFramework", ...}
@@ -193,7 +197,7 @@ Payload generation is the practitioner's core tool here. The sandbox has `git`/`
 | **ysoserial** (frohoff) | Java native | Gadget-chain payloads: `CommonsCollections1-7`, `Groovy1`, `Spring1/2`, and `URLDNS` for a safe no-exec DNS oracle. Needs a JRE. |
 | **phpggc** (ambionics) | PHP `unserialize` / Phar | Framework POP chains (Laravel, Symfony, WordPress, Drupal, Monolog). Needs `php-cli`. |
 | **ysoserial.net** | .NET `BinaryFormatter` / Json.NET | Windows/.NET gadget payloads. Needs .NET/mono — usually out of scope in a Linux sandbox. |
-| **marshalsec** | Java Hessian/Burlap, Kryo, JSON, and JNDI reference tooling | Use only from a reviewed, pinned upstream commit when a non-native Java marshaller requires it. It has no stable release and intentionally bundles historical gadget dependencies; do not treat it as a globally installed default tool. |
+| **marshalsec** | Java Hessian/Burlap, Kryo, JSON, and JNDI reference tooling | Use only from a reviewed, pinned upstream commit when a non-native Java marshaller requires it. Check upstream release status and bundled gadget dependencies before selecting a commit; do not treat it as a globally installed default tool. |
 
 ```
 # Java: prove the sink with a no-exec DNS oracle BEFORE any RCE chain

strix/skills/vulnerabilities/nosql_injection.md

@@ -91,7 +91,7 @@ Binary search the character space to minimize requests. Works on any string fiel
 
 ### `$where` JavaScript Injection
 
-If `$where` operator is enabled (disabled by default in MongoDB 7.0+; MongoDB 4.4–6.x deprecated it but left `javascriptEnabled` defaulting to `true`), inject arbitrary server-side JavaScript:
+If `$where` is enabled, inject server-side JavaScript. Inspect `security.javascriptEnabled` / `--noscripting`, managed-service restrictions, and the installed engine version. Verify operator availability, defaults, and supported functions against that build's documentation or controlled probes; deprecation alone does not mean execution is disabled ([MongoDB documentation](https://www.mongodb.com/docs/manual/reference/operator/query/where/)):
 ```json
 {"$where": "function(){return this.role == 'admin'}"}                          // direct filter — returns matching documents
 {"$where": "function(){return this.username == 'admin' && sleep(2000)}"}       // timing oracle only — sleep() returns undefined (falsy), so no documents are returned; observe latency

strix/skills/vulnerabilities/path_traversal_lfi_rfi.md

@@ -148,6 +148,8 @@ Improper file path handling and dynamic inclusion enable sensitive file disclosu
 - Verify symlink handling and path canonicalization prior to write
 - Impact: overwrite config/templates or drop webshells into served directories
 
+**Python extraction filters:** Determine the effective `filter` and `TarFile.extraction_filter` from the installed Python runtime, caller configuration, and matching documentation; verify patch/backport status against current advisories ([documentation](https://docs.python.org/3/library/tarfile.html#extraction-filters)). Filters do not eliminate link-handling bugs: test archive link ordering and post-extraction reads for outside-file disclosure or permission/timestamp changes, separately from content overwrite ([security advisory](https://mail.python.org/archives/list/[email protected]/thread/EFJWGAZJA56AKSBR2WHMHQZO7RRLZPRH/)).
+
 ### File Write to Execution
 
 Characterize the write primitive before choosing a payload:

strix/skills/vulnerabilities/prototype_pollution.md

@@ -51,7 +51,7 @@ Prototype pollution corrupts shared object prototypes (`Object.prototype`, `Arra
 **Common Sinks**
 - `lodash.merge`, `lodash.defaultsDeep`, `deep-extend`, `merge-options`
 - Express/query parsers accepting nested objects
-- YAML `load()` (not `safeLoad`) with prototype keys
+- YAML merge-key handling with prototype keys: identify the js-yaml version/schema and check applicable upstream advisories. Test whether `<<` merges alter the parsed result's prototype, then trace inherited values into a sensitive consumer; global `Object.prototype` modification is not required. Verify the installed API rather than assuming `load` versus `safeLoad` determines safety ([merge advisory](https://github.com/nodeca/js-yaml/security/advisories/GHSA-mh29-5h37-fv8m)).
 - JSON.parse → merge into existing object without null prototype
 
 **RCE Gadget Chains (Node.js)**

strix/skills/vulnerabilities/ssti.md

@@ -76,6 +76,8 @@ When output isn't reflected:
 
 ### Jinja2 / Mako (Python)
 
+**Jinja sandbox:** resolve the installed Jinja build and check current sandbox advisories before selecting indirect `str.format` or `|attr` gadgets. Establish template-source control and inspect custom filters; user data passed only as a variable is a different surface ([release notes](https://jinja.palletsprojects.com/en/stable/changes/)).
+
 The classic Python class walk — every object exposes its method-resolution-order, which leads to `object`, which exposes every subclass loaded in the interpreter, which includes things like `subprocess.Popen`:
 
 ```jinja
@@ -137,7 +139,7 @@ Twig sandbox bypasses are version-specific. The canonical historical gadget (Twi
 {{_self.env.registerUndefinedFilterCallback("system")}}{{_self.env.getFilter("id")}}
 ```
 
-This was patched — in Twig 2.x / 3.x `_self` returns the template name as a string and no longer exposes `.env`. Modern bypasses depend on which extensions are loaded and the active sandbox policy; consult Twig's published security advisories for the current state and probe with the version-specific gadgets (filter/function abuse, reflection on `_context` in some configs).
+In Twig 2.x / 3.x, `_self` returns the template name as a string and does not expose `.env`. Bypasses depend on which extensions are loaded and the active sandbox policy; consult Twig's published security advisories and probe with the version-specific gadgets (filter/function abuse, reflection on `_context` in some configs).
 
 Smarty `{php}...{/php}` was the historical RCE primitive; deprecated in Smarty 3 and removed in 4. On modern Smarty, the surface is static-method invocation and template-object reflection — `{$smarty.template_object->smarty->...}` walks back to the Smarty engine, and direct static calls on whitelisted classes (e.g. `{Smarty_Internal_Write_File::writeFile(...)}` on misconfigured installs) reach the filesystem. Probe both before assuming Smarty is hardened.
 

strix/skills/vulnerabilities/weak_password_detection.md

@@ -49,13 +49,15 @@ Weak or default credentials remain one of the most prevalent and high-impact vul
 
 ### Weak Password Policies
 
-- No minimum length or complexity requirements
+- Insufficient minimum length for the authentication mode; NIST SP 800-63B-4 requires 15 characters for single-factor passwords and permits a minimum of eight when the password is only used as part of MFA
 - Allowing common passwords: `password`, `123456`, `qwerty`, `admin`, `letmein`
 - Not checking against breached password databases (Have I Been Pwned)
 - Case-insensitive password storage
-- No password history enforcement
+- Forced periodic changes without compromise evidence, which encourage predictable password changes
 - Excessively short maximum length (indicates plaintext or weak hashing)
 
+NIST SP 800-63B-4 recommends allowing a maximum of at least 64 characters and requires screening against common/compromised passwords and checking the entire password without truncation. Missing character-class rules is not a weakness: the standard prohibits mandatory composition rules and periodic resets without evidence of compromise. Apply these requirements when that standard is the target's policy baseline ([NIST SP 800-63B-4](https://pages.nist.gov/800-63-4/sp800-63b.html#passwordver)).
+
 ### Default and Hardcoded Credentials
 
 - Vendor defaults: `admin/admin`, `admin/password`, `root/root`, `guest/guest`
@@ -191,7 +193,7 @@ No password wordlists ship in the sandbox by default — download what you need
 5. Test for password spraying (one password, many users) before targeted brute-force
 6. Check for concurrent session limits; successful logins may kick out legitimate users
 7. GraphQL batching can test multiple credentials in a single request, bypassing per-request limits
-8. Document the password policy and recommend minimum standards (length, complexity, breach checking)
+8. Document the password policy and recommend minimum standards (length appropriate to single-factor/MFA use, breach checking, and resistance to online guessing)
 9. For web logins prefer `ffuf`; for other services use `nmap` NSE `*-brute` scripts or custom scripts with equivalent logic
 10. Combine with MFA testing: weak passwords plus missing MFA is a critical finding
 

strix/skills/vulnerabilities/xss.md

@@ -53,6 +53,10 @@ Cross-site scripting persists because context, parser, and framework edges are c
 
 ## Key Vulnerabilities
 
+### DOMPurify Live-Node Sanitization
+
+Distinguish HTML-string input from live DOM objects sanitized with `IN_PLACE`. Resolve the deployed DOMPurify build and check relevant upstream advisories; a controlled observable `nodeName` is one live-node bypass mechanism. Trace same-origin foreign nodes or `adoptNode()` into this mode before selecting the bypass ([example advisory](https://github.com/cure53/DOMPurify/security/advisories/GHSA-x4vx-rjvf-j5p4)).
+
 ### DOM XSS
 
 **Sources**

strix/telemetry/logging.py

@@ -59,6 +59,7 @@ _NOISY_LIBS: tuple[str, ...] = (
 
 
 _HANDLER_TAG = "_strix_scan_handler"
+_STREAM_TAG = "_strix_stream_handler"
 
 
 # ``openai.agents`` is the openai-agents SDK's canonical logger root.
@@ -87,35 +88,64 @@ def configure_dependency_logging() -> None:
     logging.getLogger("asyncio").setLevel(logging.CRITICAL)
     logging.getLogger("asyncio").propagate = False
     warnings.filterwarnings("ignore", category=RuntimeWarning, module="asyncio")
-    _silence_urllib3_finalizer_noise()
+    _route_unraisable_to_log()
 
 
-_unraisable_hook_installed = False
-
-
-def _is_urllib3_closed_file_noise(unraisable: sys.UnraisableHookArgs) -> bool:
-    return (
-        isinstance(unraisable.exc_value, ValueError)
-        and "I/O operation on closed file" in str(unraisable.exc_value)
-        and type(unraisable.object).__module__.split(".")[0] == "urllib3"
-    )
-
-
-def _silence_urllib3_finalizer_noise() -> None:
-    global _unraisable_hook_installed  # noqa: PLW0603
-    if _unraisable_hook_installed:
-        return
-    _unraisable_hook_installed = True
-    previous = sys.unraisablehook
-
-    def hook(unraisable: sys.UnraisableHookArgs) -> None:
-        if _is_urllib3_closed_file_noise(unraisable):
-            return
-        previous(unraisable)
+def _route_unraisable_to_log() -> None:
+    def hook(u: sys.UnraisableHookArgs) -> None:
+        with contextlib.suppress(BaseException):
+            logger = logging.getLogger("strix.telemetry")
+            if logger.hasHandlers():
+                logger.warning(
+                    u.err_msg or f"Exception ignored in {u.object!r}",
+                    exc_info=(u.exc_type, u.exc_value, u.exc_traceback),  # type: ignore[arg-type]
+                )
 
     sys.unraisablehook = hook
 
 
+class _CurrentStderrHandler(logging.StreamHandler):  # type: ignore[type-arg]
+    def __init__(self, level: int) -> None:
+        logging.Handler.__init__(self, level)
+
+    @property
+    def stream(self) -> object:
+        return sys.stderr
+
+    @stream.setter
+    def stream(self, _value: object) -> None:
+        pass
+
+
+def _debug_enabled(debug: bool | None) -> bool:
+    if debug is not None:
+        return debug
+    return (os.environ.get("STRIX_DEBUG") or "").strip().lower() in {"1", "true", "yes", "on"}
+
+
+def _stream_handler(debug: bool) -> logging.Handler:
+    handler = _CurrentStderrHandler(logging.DEBUG if debug else logging.ERROR)
+    handler.setFormatter(logging.Formatter(_FORMAT, datefmt=_DATEFMT))
+    handler.addFilter(_StrixContextFilter())
+    handler.addFilter(_StdoutQuietFilter())
+    setattr(handler, _STREAM_TAG, True)
+    return handler
+
+
+def _has_stream_handler(tracked: logging.Logger) -> bool:
+    return any(getattr(handler, _STREAM_TAG, False) for handler in tracked.handlers)
+
+
+def setup_console_logging(*, debug: bool | None = None) -> None:
+    configure_dependency_logging()
+    for name in _TRACKED_ROOTS:
+        tracked = logging.getLogger(name)
+        tracked.setLevel(logging.DEBUG)
+        tracked.propagate = False
+        if not _has_stream_handler(tracked):
+            tracked.addHandler(_stream_handler(_debug_enabled(debug)))
+
+
 def setup_scan_logging(run_dir: Path, *, debug: bool | None = None) -> Callable[[], None]:
     """Attach scan-scoped handlers; return a teardown callable.
 
@@ -134,38 +164,25 @@ def setup_scan_logging(run_dir: Path, *, debug: bool | None = None) -> Callable[
     """
     configure_dependency_logging()
 
-    if debug is None:
-        debug = (os.environ.get("STRIX_DEBUG") or "").strip().lower() in {
-            "1",
-            "true",
-            "yes",
-            "on",
-        }
+    debug = _debug_enabled(debug)
 
     run_dir.mkdir(parents=True, exist_ok=True)
     log_path = run_dir / "strix.log"
 
-    formatter = logging.Formatter(_FORMAT, datefmt=_DATEFMT)
-    context_filter = _StrixContextFilter()
-
     file_handler = logging.FileHandler(log_path, mode="a", encoding="utf-8")
     file_handler.setLevel(logging.DEBUG)
-    file_handler.setFormatter(formatter)
-    file_handler.addFilter(context_filter)
+    file_handler.setFormatter(logging.Formatter(_FORMAT, datefmt=_DATEFMT))
+    file_handler.addFilter(_StrixContextFilter())
     setattr(file_handler, _HANDLER_TAG, True)
 
-    stream_handler = logging.StreamHandler()
-    stream_handler.setLevel(logging.DEBUG if debug else logging.ERROR)
-    stream_handler.setFormatter(formatter)
-    stream_handler.addFilter(context_filter)
-    stream_handler.addFilter(_StdoutQuietFilter())
-    setattr(stream_handler, _HANDLER_TAG, True)
-
     tracked_loggers = [logging.getLogger(name) for name in _TRACKED_ROOTS]
     for tracked in tracked_loggers:
         tracked.setLevel(logging.DEBUG)
         tracked.addHandler(file_handler)
-        tracked.addHandler(stream_handler)
+        if not _has_stream_handler(tracked):
+            stream_handler = _stream_handler(debug)
+            setattr(stream_handler, _HANDLER_TAG, True)
+            tracked.addHandler(stream_handler)
         tracked.propagate = False
 
     for name in _NOISY_LIBS:

strix/telemetry/posthog.py

@@ -118,6 +118,7 @@ def end(report_state: "ReportState", exit_reason: str = "completed") -> None:
             }
     except (TypeError, ValueError, AttributeError):
         pass
+    providers = report_state.get_process_llm_providers()
 
     report_state.posthog_scan_ended_sent = _send(
         "scan_ended",
@@ -129,6 +130,7 @@ def end(report_state: "ReportState", exit_reason: str = "completed") -> None:
             "vulnerabilities_total": len(report_state.vulnerability_reports),
             **{f"vulnerabilities_{k}": v for k, v in vulnerabilities_counts.items()},
             **llm_props,
+            **({"llm_providers": providers} if providers else {}),
             "skills": get_loaded_skill_names(),
         },
     )

strix/telemetry/scarf.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import json
 import logging
 import urllib.parse
 from typing import TYPE_CHECKING, Any
@@ -120,6 +121,7 @@ def end(report_state: ReportState, exit_reason: str = "completed") -> None:
             }
     except (TypeError, ValueError, AttributeError):
         pass
+    providers = report_state.get_process_llm_providers()
 
     report_state.scarf_scan_ended_sent = _send(
         "scan_ended",
@@ -132,6 +134,8 @@ def end(report_state: ReportState, exit_reason: str = "completed") -> None:
             "vulnerabilities_total": len(report_state.vulnerability_reports),
             **{f"vulnerabilities_{k}": v for k, v in vulnerabilities_counts.items()},
             **llm_props,
+            # Query params are flat, so the per-provider tally travels as JSON.
+            **({"llm_providers": json.dumps(providers)} if providers else {}),
             "skills": ",".join(get_loaded_skill_names()),
         },
     )

strix/tools/agents_graph/tools.py

@@ -12,16 +12,13 @@ from typing import Any, Literal, get_args
 
 from agents import RunContextWrapper, function_tool
 
-from strix.core.agents import Status, coordinator_from_context
+from strix.core.agents import ACTIVE_STATUSES, Status, coordinator_from_context
 from strix.core.execution import notify_parent_on_terminal
 from strix.core.hooks import LLM_TURN_KEY
 from strix.report.state import get_global_report_state
 from strix.skills import validate_requested_skills
 
 
-_ACTIVE_STATUSES: frozenset[str] = frozenset({"running", "waiting"})
-
-
 logger = logging.getLogger(__name__)
 
 
@@ -810,13 +807,13 @@ async def stop_agent(
         )
 
     current_status = statuses[target_agent_id]
-    if current_status not in _ACTIVE_STATUSES:
+    if current_status not in ACTIVE_STATUSES:
         return json.dumps(
             {
                 "success": False,
                 "error": (
                     f"Agent {target_agent_id} is already '{current_status}'; "
-                    "stop_agent only acts on running/waiting agents — use "
+                    "stop_agent only acts on running/waiting/paused agents — use "
                     "view_agent_graph to find still-active descendants and "
                     "stop them individually, or send_message_to_agent if you "
                     "want to wake this one with new instructions"

strix/tools/finish/tool.py

@@ -238,12 +238,17 @@ async def finish_scan(
           remediation steps. End with retest/validation guidance.
 
     - **Formatting — use markdown in every field.** These fields may be
-      rendered into generated reports, so structure them clearly: lead
-      each section with a short ``# Heading``, use ``**bold**`` for labels/emphasis,
-      ``inline code`` for identifiers/paths/parameters, bullet or
-      numbered lists for enumerations, and fenced code blocks
-      (```` ```language ````) for any code/payload excerpts. Never emit
-      one flat wall of prose or leave code unformatted.
+      rendered into generated reports, so structure them clearly: use
+      ``**bold**`` for labels/emphasis, ``inline code`` for
+      identifiers/paths/parameters, bullet or numbered lists for
+      enumerations, and fenced code blocks (```` ```language ````) for
+      any code/payload excerpts. Never emit one flat wall of prose or
+      leave code unformatted.
+    - **Each field is a section body, not a document.** The report adds
+      the section title ("Executive Summary", "Methodology", ...) itself,
+      so do NOT start a field with a heading such as ``# Executive
+      Summary`` — it would print twice. Sub-headings (``##``) inside a
+      field are fine.
     - If **zero** vulnerabilities were found, say so plainly and
       characterize the posture positively; ``technical_analysis`` should
       summarize the areas tested and confirm no issues, and
@@ -252,8 +257,6 @@ async def finish_scan(
     Example (abbreviated — mirror this structure, not the wording)::
 
         executive_summary:
-            # Executive Summary
-
             An external assessment of the **Acme Customer Portal**
             identified multiple weaknesses that could lead to
             unauthorized access to customer data.
@@ -269,8 +272,6 @@ async def finish_scan(
             - Potential exposure of customer records across tenants.
 
         methodology:
-            # Methodology
-
             Conducted per the **OWASP WSTG**.
 
             **Engagement type:** Gray-box external test.
@@ -280,8 +281,6 @@ async def finish_scan(
             and tenant-isolation testing, input/SSRF testing.
 
         technical_analysis:
-            # Technical Analysis
-
             **Severity model** reflects exploitability x impact.
 
             1. **SSRF in URL preview** (Critical) — insufficient
@@ -293,8 +292,6 @@ async def finish_scan(
             no deny-by-default egress policy.
 
         recommendations:
-            # Recommendations
-
             **Immediate**
             1. Remediate SSRF: enforce a destination allowlist,
                deny-by-default, re-validate on every redirect hop.

strix/tools/reporting/tool.py

@@ -1075,11 +1075,11 @@ async def create_vulnerability_report(
       "Techniques" that read like an engineering runbook rather than a
       client deliverable.
     - **Use markdown in every text field**: ``**bold**`` for emphasis,
-      ``inline code`` for identifiers/values/parameters, and fenced
-      code blocks (```` ```language ````) for any code/payload/HTTP
+      ``inline code`` for identifiers/values/parameters, and
+      language-tagged fenced code blocks for any code/payload/HTTP
       excerpt. Never leave code bare/unformatted. When referencing a
-      file, annotate the fence, e.g.
-      ```` ```python title=app.py startLineNumber=42 endLineNumber=50 ````.
+      file, annotate the opening fence with
+      ``title=app.py startLineNumber=42 endLineNumber=50`` after the language.
     - Field discipline: ``poc_description`` is steps only — NO code (all
       code goes in ``poc_script_code``); ``remediation_steps`` is prose
       only — NO code/diffs (code fixes go in ``code_locations``).
@@ -1191,6 +1191,57 @@ async def create_vulnerability_report(
       Broken / Risky Crypto, CWE-311 Missing Encryption, CWE-916 Weak
       Password Hashing.
 
+    Example (abbreviated — mirror this structure)::
+
+        title: "Reflected XSS in /search q parameter"
+        description:
+            The **`q`** parameter of `/search` reflects user input into
+            the HTML response without encoding, allowing script
+            injection.
+        technical_analysis:
+            The handler interpolates `q` directly into the page body:
+
+            ```python title=views.py startLineNumber=42 endLineNumber=44
+            html = f"<h2>Results for {q}</h2>"
+            return HttpResponse(html)
+            ```
+
+            No output encoding is applied, so `<script>` executes.
+        poc_description:
+            1. Navigate to `/search?q=<payload>`.
+            2. Observe the payload executes in the victim's browser.
+        poc_script_code:
+            ```
+            GET /search?q=<script>alert(document.domain)</script>
+            ```
+        evidence:
+            Response echoes the payload verbatim:
+
+            ```html
+            <h2>Results for <script>alert(document.domain)</script></h2>
+            ```
+        assumptions:
+            Assumes a victim can be induced to open a crafted link.
+        remediation_steps:
+            Context-encode all user input rendered into HTML; prefer the
+            template engine's auto-escaping over string interpolation.
+        counterevidence:
+            No output encoding, CSP, or WAF observed on this response;
+            payload executed in a current browser. The parameter is
+            reflected on an unauthenticated route, so no privileged
+            position is required.
+        confidence: "high"
+        severity_change_conditions:
+            A restrictive CSP that blocks inline script execution would
+            reduce impact and lower the severity.
+        fix_effort: "low"
+
+    Nice to have: for code findings, if the checkout has git history, a quick
+    ``git blame`` (quote the paths) on the vulnerable line is worth weaving into
+    ``technical_analysis`` — who last touched it, when, and in which commit, as
+    part of the prose, not a separate section. Skip it if the line is
+    uncommitted or the command fails.
+
     Args:
         title: Specific finding title (e.g.
             ``"SQL Injection in /api/users login parameter"``). Don't
@@ -1359,57 +1410,6 @@ async def create_vulnerability_report(
             fix (summary + rationale). Prose/markdown only — the code
             change itself belongs in ``code_locations``. Omit for
             black-box findings.
-
-    Example (abbreviated — mirror this structure)::
-
-        title: "Reflected XSS in /search q parameter"
-        description:
-            The **`q`** parameter of `/search` reflects user input into
-            the HTML response without encoding, allowing script
-            injection.
-        technical_analysis:
-            The handler interpolates `q` directly into the page body:
-
-            ```python title=views.py startLineNumber=42 endLineNumber=44
-            html = f"<h2>Results for {q}</h2>"
-            return HttpResponse(html)
-            ```
-
-            No output encoding is applied, so `<script>` executes.
-        poc_description:
-            1. Navigate to `/search?q=<payload>`.
-            2. Observe the payload executes in the victim's browser.
-        poc_script_code:
-            ```
-            GET /search?q=<script>alert(document.domain)</script>
-            ```
-        evidence:
-            Response echoes the payload verbatim:
-
-            ```html
-            <h2>Results for <script>alert(document.domain)</script></h2>
-            ```
-        assumptions:
-            Assumes a victim can be induced to open a crafted link.
-        remediation_steps:
-            Context-encode all user input rendered into HTML; prefer the
-            template engine's auto-escaping over string interpolation.
-        counterevidence:
-            No output encoding, CSP, or WAF observed on this response;
-            payload executed in a current browser. The parameter is
-            reflected on an unauthenticated route, so no privileged
-            position is required.
-        confidence: "high"
-        severity_change_conditions:
-            A restrictive CSP that blocks inline script execution would
-            reduce impact and lower the severity.
-        fix_effort: "low"
-
-    Nice to have: for code findings, if the checkout has git history, a quick
-    ``git blame`` (quote the paths) on the vulnerable line is worth weaving into
-    ``technical_analysis`` — who last touched it, when, and in which commit, as
-    part of the prose, not a separate section. Skip it if the line is
-    uncommitted or the command fails.
     """
     (
         http_exchange_ids,

tests/test_budget_pause_policy.py

@@ -0,0 +1,540 @@
+"""``budget_policy="pause"``: agents park before a paid call and never hear about budgets."""
+
+from __future__ import annotations
+
+import asyncio
+from typing import TYPE_CHECKING, Any
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from strix.core import execution
+from strix.core.agents import AgentCoordinator
+from strix.core.execution import _start_child_runner, run_agent_loop
+from strix.core.hooks import (
+    BudgetExceededError,
+    BudgetPausedError,
+    ReportUsageHooks,
+    recomputed_budget_flags,
+)
+from strix.core.sessions import open_agent_session
+
+
+if TYPE_CHECKING:
+    from collections.abc import AsyncIterator, Callable
+    from pathlib import Path
+
+
+COST_PER_CALL = 1.0
+_CALL_LATENCY_S = 0.005
+
+
+class _FakeLedger:
+    def __init__(self) -> None:
+        self.cost = 0.0
+        self.calls: list[str] = []
+        self.remaining: dict[str, int] = {}
+        self.warned_inputs: list[list[Any]] = []
+        self.gate: asyncio.Event | None = None
+        self.in_flight = 0
+
+    def record_sdk_usage(self, **_kwargs: Any) -> None:
+        return
+
+    def get_total_llm_cost(self) -> float:
+        return self.cost
+
+
+class _FakeStream:
+    """One ``Runner.run_streamed`` call: several LLM turns, each guarded by the hooks.
+
+    Mirrors the SDK's ordering: ``on_llm_start`` runs before the paid request,
+    ``on_llm_end`` after it. A ``BudgetPausedError`` from ``on_llm_start`` ends
+    the stream without spending, exactly like the SDK surfacing a hook error.
+    """
+
+    def __init__(
+        self,
+        *,
+        ledger: _FakeLedger,
+        hooks: ReportUsageHooks,
+        context: dict[str, Any],
+        agent: Any,
+        coordinator: AgentCoordinator,
+    ) -> None:
+        self._ledger = ledger
+        self._hooks = hooks
+        self._context = context
+        self._agent = agent
+        self._coordinator = coordinator
+        self.run_loop_exception: BaseException | None = None
+        self.final_output = None
+
+    async def stream_events(self) -> AsyncIterator[Any]:
+        agent_id = str(self._context.get("agent_id"))
+        ctx_wrapper = MagicMock()
+        ctx_wrapper.context = self._context
+        while self._ledger.remaining.get(agent_id, 0) > 0:
+            input_items: list[Any] = []
+            try:
+                await self._hooks.on_llm_start(ctx_wrapper, self._agent, None, input_items)
+            except BudgetPausedError as exc:
+                self.run_loop_exception = exc
+                return
+            self._ledger.warned_inputs.append(input_items)
+            if self._ledger.gate is not None:
+                self._ledger.in_flight += 1
+                await self._ledger.gate.wait()
+                self._ledger.in_flight -= 1
+            self._ledger.cost += COST_PER_CALL
+            self._ledger.calls.append(agent_id)
+            self._ledger.remaining[agent_id] -= 1
+            await self._hooks.on_llm_end(ctx_wrapper, self._agent, MagicMock())
+            await asyncio.sleep(_CALL_LATENCY_S)
+        if self._coordinator.statuses.get(agent_id) == "running":
+            await self._coordinator.set_status(agent_id, "completed")
+        items: tuple[Any, ...] = ()
+        for item in items:
+            yield item
+
+    def cancel(self, mode: str = "immediate") -> None:  # noqa: ARG002
+        return
+
+
+def _fake_runner(ledger: _FakeLedger, coordinator: AgentCoordinator) -> Any:
+    class _FakeRunner:
+        @staticmethod
+        def run_streamed(
+            agent: Any,
+            input: Any,  # noqa: A002, ARG004
+            *,
+            run_config: Any,  # noqa: ARG004
+            context: dict[str, Any],
+            max_turns: int,  # noqa: ARG004
+            session: Any,  # noqa: ARG004
+            hooks: ReportUsageHooks,
+        ) -> _FakeStream:
+            return _FakeStream(
+                ledger=ledger,
+                hooks=hooks,
+                context=context,
+                agent=agent,
+                coordinator=coordinator,
+            )
+
+    return _FakeRunner
+
+
+async def _noop_compact(*_args: Any, **_kwargs: Any) -> bool:
+    return False
+
+
+async def _wait_until(predicate: Callable[[], bool], *, timeout: float = 5.0) -> None:
+    async def _poll() -> None:
+        while not predicate():
+            await asyncio.sleep(0.001)
+
+    await asyncio.wait_for(_poll(), timeout=timeout)
+
+
+def _all_parked(coordinator: AgentCoordinator, *agent_ids: str) -> bool:
+    return all(coordinator.statuses.get(aid) == "budget_paused" for aid in agent_ids)
+
+
+class _Scan:
+    """Root + children driven through the real non-interactive loops."""
+
+    def __init__(
+        self,
+        tmp_path: Path,
+        monkeypatch: pytest.MonkeyPatch,
+        *,
+        max_budget_usd: float | None,
+    ) -> None:
+        self.ledger = _FakeLedger()
+        self.hooks = ReportUsageHooks(
+            model="test-model", max_budget_usd=max_budget_usd, budget_policy="pause"
+        )
+        self.coordinator = AgentCoordinator()
+        self.coordinator.set_budget_policy("pause")
+        self.coordinator.set_budget_limit_setter(self.hooks.set_max_budget_usd)
+        monkeypatch.setattr(execution, "Runner", _fake_runner(self.ledger, self.coordinator))
+        monkeypatch.setattr(execution, "_compact_session", _noop_compact)
+        self.db_path = tmp_path / "agents.sqlite"
+        self.sessions: list[Any] = []
+        self.run_config = MagicMock()
+        self.root_ctx: dict[str, Any] = {
+            "agent_id": "root",
+            "parent_id": None,
+            "coordinator": self.coordinator,
+        }
+        self.root_task: asyncio.Task[Any] | None = None
+
+    async def start_root(self, *, calls: int) -> None:
+        self.ledger.remaining["root"] = calls
+        await self.coordinator.register("root", "strix", parent_id=None)
+        session = open_agent_session("root", self.db_path)
+        self.sessions.append(session)
+        self.root_task = asyncio.create_task(
+            run_agent_loop(
+                agent=MagicMock(),
+                initial_input=[],
+                run_config=self.run_config,
+                context=self.root_ctx,
+                max_turns=500,
+                coordinator=self.coordinator,
+                agent_id="root",
+                interactive=False,
+                session=session,
+                hooks=self.hooks,
+            )
+        )
+
+    async def start_child(self, child_id: str, *, calls: int) -> None:
+        self.ledger.remaining[child_id] = calls
+        await self.coordinator.register(child_id, "recon", parent_id="root")
+        await _start_child_runner(
+            parent_ctx=self.root_ctx,
+            coordinator=self.coordinator,
+            agents_db_path=self.db_path,
+            sessions_to_close=self.sessions,
+            run_config=self.run_config,
+            max_turns=500,
+            interactive=False,
+            child_agent=MagicMock(),
+            child_id=child_id,
+            name=f"recon-{child_id}",
+            parent_id="root",
+            task="probe things",
+            initial_input=[],
+            hooks=self.hooks,
+        )
+
+    def tasks(self) -> list[asyncio.Task[Any]]:
+        tasks = [self.root_task] if self.root_task is not None else []
+        tasks.extend(rt.task for rt in self.coordinator.runtimes.values() if rt.task is not None)
+        return tasks
+
+    async def teardown(self) -> None:
+        for task in self.tasks():
+            task.cancel()
+        await asyncio.gather(*self.tasks(), return_exceptions=True)
+        for session in self.sessions:
+            session.close()
+
+
+@pytest.mark.asyncio
+async def test_pause_policy_parks_every_agent_at_the_limit_and_resumes_in_place(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    scan = _Scan(tmp_path, monkeypatch, max_budget_usd=5.0)
+    agents = ("root", "child-a", "child-b")
+
+    with patch("strix.core.hooks.get_global_report_state", return_value=scan.ledger):
+        await scan.start_root(calls=100)
+        await scan.start_child("child-a", calls=100)
+        await scan.start_child("child-b", calls=100)
+
+        await _wait_until(lambda: _all_parked(scan.coordinator, *agents))
+        assert scan.ledger.cost == pytest.approx(5.0)
+        assert len(scan.ledger.calls) == 5
+        assert scan.coordinator.budget_paused is False
+        assert scan.coordinator.budget_stopped is False
+        assert scan.coordinator.reserve_stopped is False
+        assert all(not task.done() for task in scan.tasks())
+
+        woken = await scan.coordinator.resume_budget(max_budget_usd=8.0)
+        assert sorted(woken) == sorted(agents)
+        assert scan.hooks.max_budget_usd == 8.0
+        await _wait_until(lambda: scan.ledger.cost >= 8.0)
+        await _wait_until(lambda: _all_parked(scan.coordinator, *agents))
+        assert scan.ledger.cost == pytest.approx(8.0)
+
+        await scan.coordinator.resume_budget(max_budget_usd=9.0)
+        await _wait_until(lambda: scan.ledger.cost >= 9.0)
+        await _wait_until(lambda: _all_parked(scan.coordinator, *agents))
+        assert scan.ledger.cost == pytest.approx(9.0)
+        assert len(scan.ledger.calls) == 9
+        assert all(not task.done() for task in scan.tasks())
+
+        # The model never saw a budget message: no warning band, no resume note.
+        assert scan.ledger.warned_inputs
+        assert all(items == [] for items in scan.ledger.warned_inputs)
+        for session in scan.sessions:
+            assert await session.get_items() == []
+
+        # Stop while parked: an individual stop wakes that loop and it exits.
+        child_a_task = scan.coordinator.runtimes["child-a"].task
+        assert child_a_task is not None
+        await scan.coordinator.request_stop("child-a")
+        await asyncio.wait_for(child_a_task, timeout=5.0)
+        assert scan.coordinator.statuses["child-a"] == "stopped"
+        assert scan.coordinator.statuses["root"] == "budget_paused"
+        assert scan.coordinator.statuses["child-b"] == "budget_paused"
+
+        # A scan-wide cancel while parked tears the rest down cleanly.
+        await scan.teardown()
+        assert all(task.done() for task in scan.tasks())
+        assert scan.ledger.cost == pytest.approx(9.0)
+
+
+@pytest.mark.asyncio
+async def test_pause_policy_operator_pause_and_resume_without_a_limit(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    scan = _Scan(tmp_path, monkeypatch, max_budget_usd=None)
+    agents = ("root", "child-a", "child-b")
+
+    with patch("strix.core.hooks.get_global_report_state", return_value=scan.ledger):
+        await scan.start_root(calls=6)
+        await scan.start_child("child-a", calls=6)
+        await scan.start_child("child-b", calls=6)
+
+        await _wait_until(lambda: scan.ledger.cost >= 3.0)
+        await scan.coordinator.pause_budget()
+        spent_at_pause = scan.ledger.cost
+        await _wait_until(lambda: _all_parked(scan.coordinator, *agents))
+        await _wait_until(lambda: scan.coordinator.budget_paused, timeout=0.1)
+        assert scan.ledger.cost == pytest.approx(spent_at_pause)
+        assert all(not task.done() for task in scan.tasks())
+
+        await asyncio.sleep(0.05)
+        assert scan.ledger.cost == pytest.approx(spent_at_pause)
+
+        woken = await scan.coordinator.resume_budget()
+        assert sorted(woken) == sorted(agents)
+        await _wait_until(lambda: not scan.coordinator.budget_paused, timeout=0.1)
+        await asyncio.wait_for(asyncio.gather(*scan.tasks(), return_exceptions=True), timeout=5.0)
+        assert scan.ledger.cost == pytest.approx(18.0)
+        assert {aid: str(s) for aid, s in scan.coordinator.statuses.items()} == {
+            "root": "completed",
+            "child-a": "completed",
+            "child-b": "completed",
+        }
+        assert all(items == [] for items in scan.ledger.warned_inputs)
+
+    for session in scan.sessions:
+        session.close()
+
+
+@pytest.mark.asyncio
+async def test_pause_policy_keeps_in_flight_overshoot(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    scan = _Scan(tmp_path, monkeypatch, max_budget_usd=1.0)
+    scan.ledger.gate = asyncio.Event()
+    await scan.coordinator.register("root", "strix", parent_id=None)
+
+    with patch("strix.core.hooks.get_global_report_state", return_value=scan.ledger):
+        await scan.start_child("child-a", calls=100)
+        await scan.start_child("child-b", calls=100)
+
+        # Both calls were dispatched under the limit; neither is cancelled.
+        await _wait_until(lambda: scan.ledger.in_flight == 2)
+        assert scan.ledger.cost == 0.0
+        scan.ledger.gate.set()
+
+        await _wait_until(lambda: _all_parked(scan.coordinator, "child-a", "child-b"))
+        assert scan.ledger.cost == pytest.approx(2.0)
+        assert scan.hooks.max_budget_usd is not None
+        assert scan.ledger.cost > scan.hooks.max_budget_usd
+        assert scan.coordinator.budget_stopped is False
+        assert all(not task.done() for task in scan.tasks())
+
+        # Resuming at a limit that is already spent parks again without a call.
+        await scan.coordinator.resume_budget(max_budget_usd=1.5)
+        await asyncio.sleep(0.05)
+        await _wait_until(lambda: _all_parked(scan.coordinator, "child-a", "child-b"))
+        assert scan.ledger.cost == pytest.approx(2.0)
+
+        await scan.teardown()
+
+
+@pytest.mark.asyncio
+async def test_resume_between_park_decision_and_wait_is_not_missed() -> None:
+    coordinator = AgentCoordinator()
+    coordinator.set_budget_policy("pause")
+    await coordinator.register("a", "strix", parent_id=None)
+
+    parked_epoch = coordinator.resume_epoch
+    await coordinator.park_for_budget("a")
+    assert _all_parked(coordinator, "a")
+    await coordinator.resume_budget()
+
+    await asyncio.wait_for(
+        coordinator.wait_for_budget_resume("a", parked_epoch=parked_epoch), timeout=1.0
+    )
+    assert coordinator.statuses["a"] == "running"
+
+
+@pytest.mark.asyncio
+async def test_parked_wait_returns_on_stop_signals() -> None:
+    coordinator = AgentCoordinator()
+    coordinator.set_budget_policy("pause")
+    await coordinator.register("a", "strix", parent_id=None)
+    await coordinator.register("b", "recon", parent_id="a")
+    await coordinator.park_for_budget("a")
+    await coordinator.park_for_budget("b")
+
+    wait_a = asyncio.create_task(
+        coordinator.wait_for_budget_resume("a", parked_epoch=coordinator.resume_epoch)
+    )
+    wait_b = asyncio.create_task(
+        coordinator.wait_for_budget_resume("b", parked_epoch=coordinator.resume_epoch)
+    )
+    await asyncio.sleep(0.02)
+    assert not wait_a.done()
+    assert not wait_b.done()
+
+    await coordinator.request_stop("b")
+    await asyncio.wait_for(wait_b, timeout=1.0)
+    assert coordinator.statuses["b"] == "stopped"
+    assert not wait_a.done()
+
+    await coordinator.trigger_budget_stop()
+    await asyncio.wait_for(wait_a, timeout=1.0)
+    assert coordinator.statuses["a"] == "budget_paused"
+
+
+@pytest.mark.asyncio
+async def test_park_never_overwrites_a_stop_that_landed_first() -> None:
+    coordinator = AgentCoordinator()
+    coordinator.set_budget_policy("pause")
+    await coordinator.register("a", "strix", parent_id=None)
+
+    parked_epoch = coordinator.resume_epoch
+    await coordinator.request_stop("a")
+    assert await coordinator.park_for_budget("a") is False
+    assert coordinator.statuses["a"] == "stopped"
+
+    await asyncio.wait_for(
+        coordinator.wait_for_budget_resume("a", parked_epoch=parked_epoch), timeout=1.0
+    )
+    assert coordinator.statuses["a"] == "stopped"
+
+
+@pytest.mark.asyncio
+async def test_parked_children_keep_the_scan_open() -> None:
+    coordinator = AgentCoordinator()
+    coordinator.set_budget_policy("pause")
+    await coordinator.register("root", "strix", parent_id=None)
+    await coordinator.register("child", "recon", parent_id="root")
+    assert await coordinator.park_for_budget("child") is True
+
+    active = await coordinator.active_agents_except("root")
+    assert [a["agent_id"] for a in active] == ["child"]
+    assert active[0]["status"] == "budget_paused"
+
+
+@pytest.mark.asyncio
+async def test_resume_budget_replaces_the_limit_and_validates_it() -> None:
+    hooks = ReportUsageHooks(model="m", max_budget_usd=10.0, budget_policy="pause")
+    coordinator = AgentCoordinator()
+    coordinator.set_budget_policy("pause")
+    coordinator.set_budget_limit_setter(hooks.set_max_budget_usd)
+
+    await coordinator.resume_budget(max_budget_usd=25.0)
+    assert hooks.max_budget_usd == 25.0
+    await coordinator.resume_budget()
+    assert hooks.max_budget_usd == 25.0
+    with pytest.raises(ValueError, match="greater than 0"):
+        await coordinator.resume_budget(max_budget_usd=0.0)
+    with pytest.raises(ValueError, match="finite"):
+        await coordinator.resume_budget(max_budget_usd=float("inf"))
+    assert hooks.max_budget_usd == 25.0
+
+
+def _ctx(coordinator: AgentCoordinator | None, *, parent_id: str | None = None) -> MagicMock:
+    wrapper = MagicMock()
+    wrapper.context = {"agent_id": "x", "parent_id": parent_id}
+    if coordinator is not None:
+        wrapper.context["coordinator"] = coordinator
+    return wrapper
+
+
+@pytest.mark.asyncio
+async def test_pause_hooks_never_warn_and_park_only_at_the_limit() -> None:
+    ledger = _FakeLedger()
+    hooks = ReportUsageHooks(model="m", max_budget_usd=10.0, budget_policy="pause")
+    coordinator = AgentCoordinator()
+    coordinator.set_budget_policy("pause")
+
+    with patch("strix.core.hooks.get_global_report_state", return_value=ledger):
+        for cost in (7.0, 8.5, 9.5, 9.99):
+            ledger.cost = cost
+            for parent_id in (None, "root"):
+                items: list[Any] = []
+                await hooks.on_llm_start(
+                    _ctx(coordinator, parent_id=parent_id), MagicMock(), None, items
+                )
+                assert items == []
+                await hooks.on_llm_end(
+                    _ctx(coordinator, parent_id=parent_id), MagicMock(), MagicMock()
+                )
+
+        ledger.cost = 10.0
+        await hooks.on_llm_end(_ctx(coordinator, parent_id="root"), MagicMock(), MagicMock())
+        with pytest.raises(BudgetPausedError) as at_limit:
+            await hooks.on_llm_start(_ctx(coordinator), MagicMock(), None, [])
+        assert at_limit.value.resume_epoch == coordinator.resume_epoch
+
+        ledger.cost = 13.7
+        with pytest.raises(BudgetPausedError):
+            await hooks.on_llm_start(_ctx(coordinator, parent_id="root"), MagicMock(), None, [])
+
+        hooks.set_max_budget_usd(20.0)
+        items = []
+        await hooks.on_llm_start(_ctx(coordinator), MagicMock(), None, items)
+        assert items == []
+
+        await coordinator.pause_budget()
+        with pytest.raises(BudgetPausedError, match="paused"):
+            await hooks.on_llm_start(_ctx(coordinator), MagicMock(), None, [])
+
+
+@pytest.mark.asyncio
+async def test_pause_hooks_do_not_count_a_parked_turn() -> None:
+    ledger = _FakeLedger()
+    ledger.cost = 10.0
+    hooks = ReportUsageHooks(model="m", max_budget_usd=10.0, budget_policy="pause")
+    coordinator = AgentCoordinator()
+    ctx = _ctx(coordinator)
+    with patch("strix.core.hooks.get_global_report_state", return_value=ledger):
+        with pytest.raises(BudgetPausedError):
+            await hooks.on_llm_start(ctx, MagicMock(), None, [])
+        assert "llm_turn" not in ctx.context
+        hooks.set_max_budget_usd(11.0)
+        await hooks.on_llm_start(ctx, MagicMock(), None, [])
+        assert ctx.context["llm_turn"] == 1
+
+
+@pytest.mark.asyncio
+async def test_stop_policy_is_unchanged() -> None:
+    ledger = _FakeLedger()
+    hooks = ReportUsageHooks(model="m", max_budget_usd=10.0)
+    assert hooks.budget_policy == "stop"
+
+    with patch("strix.core.hooks.get_global_report_state", return_value=ledger):
+        ledger.cost = 7.0
+        items: list[Any] = []
+        await hooks.on_llm_start(_ctx(None), MagicMock(), None, items)
+        assert len(items) == 1
+        assert "Scan cost budget" in str(items[0])
+
+        ledger.cost = 10.0
+        with pytest.raises(BudgetExceededError):
+            await hooks.on_llm_end(_ctx(None), MagicMock(), MagicMock())
+
+    assert recomputed_budget_flags(10.0, 10.0, interactive=False, budget_policy="stop") == (
+        True,
+        True,
+    )
+    assert recomputed_budget_flags(10.0, 10.0, interactive=False, budget_policy="pause") == (
+        False,
+        False,
+    )
+
+
+def test_budget_policy_is_validated() -> None:
+    with pytest.raises(ValueError, match="budget_policy"):
+        ReportUsageHooks(model="m", budget_policy="later")  # type: ignore[arg-type]

tests/test_cli_fail_on.py

@@ -0,0 +1,127 @@
+"""Tests for the --fail-on headless severity gate."""
+
+from __future__ import annotations
+
+import importlib
+import sys
+from types import SimpleNamespace
+from typing import Any
+
+import pytest
+
+
+cli_main: Any = importlib.import_module("strix.interface.main")
+cli_args: Any = importlib.import_module("strix.interface.cli_args")
+
+
+def _stub_settings(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setattr(
+        cli_main,
+        "load_settings",
+        lambda: SimpleNamespace(runtime=SimpleNamespace(max_local_copy_mb=1024)),
+    )
+
+
+def _reports(*severities: str | None) -> list[dict[str, Any]]:
+    return [{"id": f"vuln-{i:04d}", "severity": sev} for i, sev in enumerate(severities, 1)]
+
+
+def test_no_findings_never_fail() -> None:
+    assert not cli_main.findings_fail_build([], None)
+    assert not cli_main.findings_fail_build([], "info")
+
+
+@pytest.mark.parametrize("severity", ["critical", "high", "medium", "low", "info", "none"])
+def test_without_threshold_any_finding_fails(severity: str) -> None:
+    assert cli_main.findings_fail_build(_reports(severity), None)
+
+
+@pytest.mark.parametrize(
+    ("fail_on", "severity", "expected"),
+    [
+        ("high", "critical", True),
+        ("high", "high", True),
+        ("high", "medium", False),
+        ("high", "low", False),
+        ("high", "info", False),
+        ("critical", "high", False),
+        ("critical", "critical", True),
+        ("info", "info", True),
+        ("info", "low", True),
+        ("medium", "HIGH", True),
+        ("medium", " Low ", False),
+    ],
+)
+def test_threshold_compares_severity(fail_on: str, severity: str, expected: bool) -> None:
+    assert cli_main.findings_fail_build(_reports(severity), fail_on) is expected
+
+
+def test_one_finding_at_threshold_fails_a_mixed_run() -> None:
+    assert cli_main.findings_fail_build(_reports("info", "low", "high"), "high")
+
+
+@pytest.mark.parametrize("severity", ["severe", "", None])
+def test_unrecognized_severity_fails_closed(severity: str | None) -> None:
+    assert cli_main.findings_fail_build(_reports(severity), "critical")
+
+
+def test_none_severity_passes_any_threshold() -> None:
+    assert not cli_main.findings_fail_build(_reports("none"), "info")
+
+
+def test_parse_fail_on_is_case_insensitive(monkeypatch: pytest.MonkeyPatch) -> None:
+    _stub_settings(monkeypatch)
+    monkeypatch.setattr(
+        sys, "argv", ["strix", "-t", "https://test.com/", "-n", "--fail-on", "HIGH"]
+    )
+
+    args = cli_main.parse_arguments()
+
+    assert args.fail_on == "high"
+
+
+def test_parse_fail_on_defaults_to_none(monkeypatch: pytest.MonkeyPatch) -> None:
+    _stub_settings(monkeypatch)
+    monkeypatch.setattr(sys, "argv", ["strix", "-t", "https://test.com/", "-n"])
+
+    assert cli_main.parse_arguments().fail_on is None
+
+
+def test_parse_fail_on_rejects_unknown_severity(
+    monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
+) -> None:
+    _stub_settings(monkeypatch)
+    monkeypatch.setattr(
+        sys, "argv", ["strix", "-t", "https://test.com/", "-n", "--fail-on", "severe"]
+    )
+
+    with pytest.raises(SystemExit):
+        cli_main.parse_arguments()
+
+    assert "--fail-on" in capsys.readouterr().err
+
+
+def test_parse_fail_on_requires_non_interactive(
+    monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
+) -> None:
+    _stub_settings(monkeypatch)
+    monkeypatch.setattr(cli_args, "terminal_attached", lambda: True)
+    monkeypatch.setattr(sys, "argv", ["strix", "-t", "https://test.com/", "--fail-on", "high"])
+
+    with pytest.raises(SystemExit):
+        cli_main.parse_arguments()
+
+    assert "--fail-on only applies to headless runs" in capsys.readouterr().err
+
+
+def test_parse_fail_on_without_a_terminal_waits_for_the_headless_fallback(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    _stub_settings(monkeypatch)
+    monkeypatch.setattr(cli_args, "terminal_attached", lambda: False)
+    monkeypatch.setattr(sys, "argv", ["strix", "-t", "https://test.com/", "--fail-on", "high"])
+
+    args = cli_main.parse_arguments()
+
+    assert args.fail_on == "high"
+    assert args.non_interactive is False

tests/test_cli_resume_picker.py

@@ -0,0 +1,123 @@
+"""Tests for `strix --resume` without a run name."""
+
+from __future__ import annotations
+
+import importlib
+import json
+import sys
+from typing import TYPE_CHECKING, Any
+
+import pytest
+
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+
+cli_main: Any = importlib.import_module("strix.interface.main")
+
+
+def _write_run(base: Path, name: str, *, state: bool = True) -> None:
+    run_dir = base / name
+    (run_dir / ".state").mkdir(parents=True)
+    (run_dir / "run.json").write_text(
+        json.dumps(
+            {
+                "run_name": name,
+                "status": "completed",
+                "start_time": "2026-10-04T10:00:00+00:00",
+                "targets_info": [
+                    {
+                        "type": "web_application",
+                        "details": {"target_url": "https://example.com"},
+                        "original": "https://example.com",
+                    }
+                ],
+            }
+        ),
+        encoding="utf-8",
+    )
+    if state:
+        (run_dir / ".state" / "agents.json").write_text("{}", encoding="utf-8")
+
+
+def test_bare_resume_defers_to_the_picker(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.chdir(tmp_path)
+    _write_run(tmp_path / "strix_runs", "example-com_1111")
+    monkeypatch.setattr(sys, "argv", ["strix", "--resume"])
+
+    args = cli_main.parse_arguments()
+
+    assert args.resume is None
+    assert args.resume_picker is True
+    assert args.needs_setup is False
+    assert args.targets_info == []
+
+
+def test_bare_resume_keeps_the_new_instruction_as_resume_guidance(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    monkeypatch.chdir(tmp_path)
+    _write_run(tmp_path / "strix_runs", "example-com_1111")
+    monkeypatch.setattr(sys, "argv", ["strix", "--resume", "--instruction", "focus on auth"])
+
+    args = cli_main.parse_arguments()
+
+    assert args.resume_picker is True
+    assert args.user_explicit_instruction == "focus on auth"
+
+
+def test_bare_resume_with_no_runs_is_an_error(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
+) -> None:
+    monkeypatch.chdir(tmp_path)
+    monkeypatch.setattr(sys, "argv", ["strix", "--resume"])
+
+    with pytest.raises(SystemExit):
+        cli_main.parse_arguments()
+
+    assert "no runs in ./strix_runs" in capsys.readouterr().err
+
+
+def test_bare_resume_headless_lists_the_runs_instead_of_prompting(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
+) -> None:
+    monkeypatch.chdir(tmp_path)
+    _write_run(tmp_path / "strix_runs", "example-com_1111")
+    _write_run(tmp_path / "strix_runs", "example-com_2222", state=False)
+    monkeypatch.setattr(sys, "argv", ["strix", "-n", "--resume"])
+
+    with pytest.raises(SystemExit):
+        cli_main.parse_arguments()
+
+    err = capsys.readouterr().err
+    assert "--resume needs a run name in headless mode" in err
+    assert "example-com_1111  completed  2026-10-04T10:00:00  https://example.com" in err
+    assert "example-com_2222  completed  2026-10-04T10:00:00  https://example.com" in err
+    assert "example-com_2222" in err
+
+
+def test_bare_resume_still_rejects_targets(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
+) -> None:
+    monkeypatch.chdir(tmp_path)
+    _write_run(tmp_path / "strix_runs", "example-com_1111")
+    monkeypatch.setattr(sys, "argv", ["strix", "--resume", "-t", "https://example.com"])
+
+    with pytest.raises(SystemExit):
+        cli_main.parse_arguments()
+
+    assert "Cannot combine --resume with --target" in capsys.readouterr().err
+
+
+def test_named_resume_without_agent_state_is_an_error(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
+) -> None:
+    monkeypatch.chdir(tmp_path)
+    _write_run(tmp_path / "strix_runs", "example-com_1111", state=False)
+    monkeypatch.setattr(sys, "argv", ["strix", "--resume", "example-com_1111"])
+
+    with pytest.raises(SystemExit):
+        cli_main.parse_arguments()
+
+    assert "never reached its first agent snapshot" in capsys.readouterr().err

tests/test_console_logging.py

@@ -0,0 +1,105 @@
+from __future__ import annotations
+
+import io
+import logging
+import sys
+from typing import TYPE_CHECKING
+
+import pytest
+
+from strix.telemetry import logging as strix_logging
+
+
+if TYPE_CHECKING:
+    from collections.abc import Iterator
+    from pathlib import Path
+
+
+@pytest.fixture
+def clean_loggers() -> Iterator[None]:
+    tracked = [logging.getLogger(name) for name in strix_logging._TRACKED_ROOTS]
+    saved = [(logger.handlers[:], logger.level, logger.propagate) for logger in tracked]
+    for logger in tracked:
+        logger.handlers = []
+    try:
+        yield
+    finally:
+        for logger, (handlers, level, propagate) in zip(tracked, saved, strict=True):
+            logger.handlers = handlers
+            logger.setLevel(level)
+            logger.propagate = propagate
+
+
+def _stderr_handlers(logger: logging.Logger) -> list[logging.Handler]:
+    return [h for h in logger.handlers if getattr(h, strix_logging._STREAM_TAG, False)]
+
+
+@pytest.mark.usefixtures("clean_loggers")
+def test_console_logging_hides_warnings_and_shows_errors(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.delenv("STRIX_DEBUG", raising=False)
+    err = io.StringIO()
+    monkeypatch.setattr(sys, "stderr", err)
+    strix_logging.setup_console_logging()
+
+    logger = logging.getLogger("strix.llm.request_log")
+    logger.warning("llm_request outcome=error")
+    logger.error("boom")
+
+    out = err.getvalue()
+    assert "llm_request" not in out
+    assert "ERROR" in out
+    assert "strix.llm.request_log: boom" in out
+
+
+@pytest.mark.usefixtures("clean_loggers")
+def test_console_logging_follows_the_current_stderr(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.delenv("STRIX_DEBUG", raising=False)
+    first = io.StringIO()
+    monkeypatch.setattr(sys, "stderr", first)
+    strix_logging.setup_console_logging()
+    second = io.StringIO()
+    monkeypatch.setattr(sys, "stderr", second)
+
+    logging.getLogger("strix.test").error("after swap")
+
+    assert first.getvalue() == ""
+    assert "after swap" in second.getvalue()
+
+
+@pytest.mark.usefixtures("clean_loggers")
+def test_console_logging_is_idempotent() -> None:
+    strix_logging.setup_console_logging()
+    strix_logging.setup_console_logging()
+
+    for name in strix_logging._TRACKED_ROOTS:
+        assert len(_stderr_handlers(logging.getLogger(name))) == 1
+
+
+@pytest.mark.usefixtures("clean_loggers")
+def test_scan_logging_reuses_the_console_handler(
+    monkeypatch: pytest.MonkeyPatch, tmp_path: Path
+) -> None:
+    monkeypatch.delenv("STRIX_DEBUG", raising=False)
+    err = io.StringIO()
+    monkeypatch.setattr(sys, "stderr", err)
+    strix_logging.setup_console_logging()
+    teardown = strix_logging.setup_scan_logging(tmp_path)
+
+    strix_logger = logging.getLogger("strix")
+    assert len(_stderr_handlers(strix_logger)) == 1
+    logging.getLogger("strix.test").error("once")
+    assert err.getvalue().count("once") == 1
+    assert "once" in (tmp_path / "strix.log").read_text(encoding="utf-8")
+
+    teardown()
+    assert len(_stderr_handlers(strix_logger)) == 1
+    assert not any(isinstance(h, logging.FileHandler) for h in strix_logger.handlers)
+
+
+@pytest.mark.usefixtures("clean_loggers")
+def test_scan_logging_alone_still_attaches_a_stderr_handler(tmp_path: Path) -> None:
+    teardown = strix_logging.setup_scan_logging(tmp_path)
+    strix_logger = logging.getLogger("strix")
+    assert len(_stderr_handlers(strix_logger)) == 1
+    teardown()
+    assert _stderr_handlers(strix_logger) == []

tests/test_cost_tracking.py

@@ -2,9 +2,12 @@
 
 from __future__ import annotations
 
+import uuid
 from types import SimpleNamespace
-from unittest.mock import MagicMock, patch
+from typing import TYPE_CHECKING, Any
+from unittest.mock import MagicMock, call, patch
 
+import httpx
 import litellm
 import pytest
 from litellm.types.utils import LlmProviders
@@ -14,6 +17,7 @@ from strix.config.models import (
     _configure_litellm_compatibility,
     _install_openrouter_stream_cost_capture,
 )
+from strix.llm import request_log
 from strix.report.state import (
     ReportState,
     litellm_cost_callback,
@@ -21,6 +25,11 @@ from strix.report.state import (
     set_global_report_state,
     streamed_openrouter_costs,
 )
+from strix.report.usage import LLMUsageLedger
+
+
+if TYPE_CHECKING:
+    from litellm.types.llms.openai import AllMessageValues
 
 
 @pytest.fixture(autouse=True)
@@ -257,3 +266,126 @@ def test_openrouter_stream_handler_records_cost() -> None:
     assert streamed_openrouter_costs.take(SimpleNamespace(id="gen-stream")) == pytest.approx(
         0.0035055
     )
+
+
+def test_openrouter_tallies_provider() -> None:
+    _install_openrouter_stream_cost_capture()
+    config = ProviderConfigManager.get_provider_chat_config(
+        model="z-ai/glm-5.3", provider=LlmProviders.OPENROUTER
+    )
+    assert config is not None
+    handler = config.get_model_response_iterator(streaming_response=iter([]), sync_stream=True)
+    report_state = MagicMock()
+    usage = {
+        "prompt_tokens": 1000,
+        "completion_tokens": 10,
+        "cost": 0.002,
+        "prompt_tokens_details": {"cached_tokens": 900},
+    }
+    with patch("strix.report.state.get_global_report_state", return_value=report_state):
+        handler.chunk_parser(
+            {
+                "id": "gen-a",
+                "created": 1,
+                "model": "z-ai/glm-5.3",
+                "provider": "Together",
+                "choices": [{"index": 0, "delta": {"content": None}}],
+                "usage": usage,
+            }
+        )
+        # Non-streamed replies (LLM_DISABLE_STREAMING) carry the same fields.
+        reply = {
+            "choices": [{"message": {"role": "assistant"}}],
+            "provider": "Together",
+            "usage": usage,
+        }
+        config.transform_response(
+            "z-ai/glm-5.3",
+            httpx.Response(200, json=reply),
+            litellm.ModelResponse(),
+            MagicMock(),
+            {},
+            [],
+            {},
+            {},
+            None,
+        )
+
+    tally = call("Together", agent_id=None, input_tokens=1000, cached_tokens=900, cost=0.002)
+    assert report_state.record_llm_provider.call_args_list == [tally, tally]
+
+
+def test_provider_tally_survives_run_record_round_trip() -> None:
+    ledger = LLMUsageLedger()
+    for input_tokens, cached_tokens, cost in [(1000, 900, 0.002), (500, 0, 0.001)]:
+        ledger.record_provider(
+            "Together",
+            agent_id=None,
+            input_tokens=input_tokens,
+            cached_tokens=cached_tokens,
+            cost=cost,
+            cache_block_tokens=128,
+        )
+
+    restored = LLMUsageLedger()
+    restored.hydrate(ledger.to_record())
+
+    assert restored.to_record()["providers"] == {
+        "Together": {
+            "requests": 2,
+            "input_tokens": 1500,
+            "cached_tokens": 900,
+            "cost": 0.003,
+            "cache_misses": 0,
+            "missed_tokens": 0,
+        }
+    }
+
+
+def test_provider_tally_counts_cache_misses_per_agent() -> None:
+    ledger = LLMUsageLedger()
+    calls = [
+        ("Z.AI", "a1", 1000, 0),  # first call: nothing to miss
+        ("Z.AI", "a1", 1200, 960),  # 40 short of the previous 1000: within a block
+        ("DeepInfra", "a1", 1500, 200),  # 1000 of the previous 1200 lost
+        ("Z.AI", "a2", 800, 0),  # another agent's first call
+        ("Z.AI", "a1", 600, 0),  # prompt shrank: compaction, not a miss
+    ]
+    for provider, agent_id, input_tokens, cached_tokens in calls:
+        ledger.record_provider(
+            provider,
+            agent_id=agent_id,
+            input_tokens=input_tokens,
+            cached_tokens=cached_tokens,
+            cost=0.0,
+            cache_block_tokens=128,
+        )
+
+    providers = ledger.to_record()["providers"]
+    assert providers["DeepInfra"]["cache_misses"] == 1
+    assert providers["DeepInfra"]["missed_tokens"] == 1000
+    assert providers["Z.AI"]["cache_misses"] == 0
+
+
+def test_openrouter_request_carries_agent_session_id() -> None:
+    _install_openrouter_stream_cost_capture()
+    config = ProviderConfigManager.get_provider_chat_config(
+        model="moonshotai/kimi-k3", provider=LlmProviders.OPENROUTER
+    )
+    assert config is not None
+    messages: list[AllMessageValues] = [{"role": "user", "content": "hi"}]
+
+    def body() -> dict[str, Any]:
+        return config.transform_request("moonshotai/kimi-k3", messages, {}, {}, {})
+
+    assert "session_id" not in body()
+    token = request_log.bind_call_context("a1b2c3d4", "root")
+    try:
+        assert "session_id" not in body()
+        with patch("strix.config.models.load_settings") as settings:
+            settings.return_value.llm.openrouter_sticky_sessions = True
+            session_id = body()["session_id"]
+            assert str(uuid.UUID(session_id)) == session_id
+            assert body()["session_id"] == session_id
+    finally:
+        request_log.reset_call_context(token)