rustwright 0.1.0a2__tar.gz

rustwright-0.1.0a2/.dockerignore

@@ -0,0 +1,35 @@
+.git
+.audit-playwright
+.venv
+.venv_browser_harness_compare
+.pytest_cache
+.mypy_cache
+.ruff_cache
+target
+dist
+build
+*.egg-info
+__pycache__
+*.py[cod]
+*.so
+*.dylib
+*.dll
+*.log
+.DS_Store
+downloads
+test-results
+playwright-report
+videos
+traces
+screenshots
+.benchmark-data
+benchmark-data
+datasets
+data/Mind2Web
+data/mind2web
+data/WebVoyager
+data/webvoyager
+data/WebVoyager_data.jsonl
+data/GAIA_web.jsonl
+data/reference_answer.json
+benchmarks/external

rustwright-0.1.0a2/.github/ISSUE_TEMPLATE/bug_report.md

@@ -0,0 +1,34 @@
+---
+name: Bug report
+about: Report a reproducible Rustwright bug
+title: "[Bug]: "
+labels: bug
+assignees: ""
+---
+
+## What happened?
+
+
+## Reproduction
+
+```python
+# Minimal code or commands that reproduce the issue
+```
+
+## Expected behavior
+
+
+## Environment
+
+- OS:
+- Python version:
+- Rust version:
+- Node version, if relevant:
+- Browser:
+- Rustwright commit:
+
+## Logs
+
+```text
+
+```

rustwright-0.1.0a2/.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,5 @@
+blank_issues_enabled: false
+contact_links:
+  - name: Questions and setup help
+    url: https://discord.gg/fG2XXEuQX3
+    about: Use the Skyvern Discord for questions, troubleshooting, and discussion.

rustwright-0.1.0a2/.github/ISSUE_TEMPLATE/feature_request.md

@@ -0,0 +1,18 @@
+---
+name: Feature request
+about: Suggest a Rustwright capability or compatibility improvement
+title: "[Feature]: "
+labels: enhancement
+assignees: ""
+---
+
+## What should Rustwright support?
+
+
+## Why is this useful?
+
+
+## Playwright behavior or reference
+
+
+## Additional context

rustwright-0.1.0a2/.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,16 @@
+## What
+
+
+## Why
+
+
+## Testing
+
+- [ ] `cargo check --locked`
+- [ ] Targeted pytest, using the focused `-k` subset from `CONTRIBUTING.md`
+
+## Checklist
+
+- [ ] I kept this change focused.
+- [ ] I added or updated tests for behavior changes.
+- [ ] I did not include private credentials, tokens, or internal-only artifacts.

rustwright-0.1.0a2/.github/dependabot.yml

@@ -0,0 +1,26 @@
+version: 2
+updates:
+  - package-ecosystem: cargo
+    directory: "/"
+    schedule:
+      interval: weekly
+
+  - package-ecosystem: cargo
+    directory: "/node"
+    schedule:
+      interval: weekly
+
+  - package-ecosystem: pip
+    directory: "/"
+    schedule:
+      interval: weekly
+
+  - package-ecosystem: npm
+    directory: "/node"
+    schedule:
+      interval: weekly
+
+  - package-ecosystem: github-actions
+    directory: "/"
+    schedule:
+      interval: weekly

rustwright-0.1.0a2/.github/prompts/code-review.md

@@ -0,0 +1,56 @@
+Please review this pull request for the **Rustwright** repository and give concise, actionable feedback.
+
+Rustwright is a Rust engine (PyO3 for Python, napi-rs for Node) that speaks raw Chrome DevTools Protocol and exposes a **Playwright-compatible** API. It is alpha, Chromium-only, MIT-licensed. The Rust core lives in `src/lib.rs`; the Python API in `python/rustwright/`; the Node bindings in `node/`.
+
+Review for:
+
+- **Correctness & bugs** — logic errors, unhandled edge cases, error handling. Pay special attention to:
+  - the FFI boundaries — PyO3 `Bound<'py, T>` usage and GIL handling (`py.detach` around blocking work), and napi async/threadsafe-function usage;
+  - the CDP layer — session/frame routing (OOPIF), the tokio-tungstenite WebSocket transport, and message dispatch/correlation.
+- **Memory & concurrency safety** — any new `unsafe`; `unwrap()`/`expect()`/`panic!` on fallible paths that could crash a user's process across the FFI boundary; races or deadlocks in the async CDP client; holding the GIL across I/O.
+- **Security** (this is browser-automation software) — untrusted-input handling, JS/CDP injection, and anything that silently changes the **trusted-input-by-default** behavior. Rustwright must **not overclaim on anti-detection**: flag any change that adds "undetectable" / "bypass Cloudflare" / "stealth" style claims, or that weakens the honest signal-hygiene framing (it currently passes 3/4 public fingerprint targets and says so plainly).
+- **API compatibility** — the Python/Node surface mirrors Playwright. Flag divergence from Playwright's shape/behavior that isn't justified, and missing parity coverage.
+- **Performance** — avoidable allocations/copies on the per-CDP-message hot path; blocking calls inside async paths.
+- **Tests** — is the change covered by the pytest suite (`tests/`)? Does it warrant a parity test against real Playwright?
+- **Honesty of claims** — if the PR touches README/docs/benchmarks, verify the claims match what the code and tests actually support (no inflated numbers, no unverified platform/behavior claims, MSRV kept accurate).
+
+Use `CONTRIBUTING.md` (and `CLAUDE.md` if present) for conventions. Be constructive.
+
+Keep the review **concise and scannable**:
+- Start with a 1–2 sentence overall assessment.
+- Group feedback by severity using GitHub `<details>`/`<summary>` collapsible sections.
+- Expand 🔴 Critical by default; keep 🟡 Suggestions and 📝 Minor collapsed.
+- If there are no issues, post a brief LGTM (with the marker).
+
+Format:
+
+```
+<!-- claude-code-review -->
+## Summary
+Brief overall assessment.
+
+<details open>
+<summary>🔴 Critical Issues (X)</summary>
+
+- ...
+
+</details>
+
+<details>
+<summary>🟡 Suggestions (X)</summary>
+
+- ...
+
+</details>
+
+<details>
+<summary>📝 Minor / Style (X)</summary>
+
+- ...
+
+</details>
+```
+
+IMPORTANT: Start your comment with the exact marker `<!-- claude-code-review -->` on the first line — this lets a later commit replace the previous review instead of stacking comments.
+
+Use `gh pr comment` (via your Bash tool) to post the review as a single comment on the PR.

rustwright-0.1.0a2/.github/workflows/benchmark-testbox.yml

@@ -0,0 +1,51 @@
+name: Rustwright Benchmark Testbox
+
+on:
+  workflow_dispatch:
+    inputs:
+      testbox_id:
+        type: string
+        description: "Testbox session ID"
+        required: true
+
+jobs:
+  benchmark:
+    if: github.repository == 'Skyvern-AI/rustwright'
+    runs-on: blacksmith-4vcpu-ubuntu-2404
+    timeout-minutes: 120
+    env:
+      RUSTWRIGHT_DOCKER_IMAGE: rustwright-verify-testbox
+      TEST_DOCKER_MEMORY_LIMIT: 8g
+      INSTALL_BROWSER_HARNESS: "1"
+      INSTALL_PUPPETEER: "1"
+      DOCKER_BUILDKIT: "1"
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+
+      - name: Begin Testbox
+        uses: useblacksmith/begin-testbox@233448af4bfdc6fca509a7f0974411ac6d8a8043 # v2
+        with:
+          testbox_id: ${{ inputs.testbox_id }}
+
+      - uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6.3.0
+        with:
+          python-version: "3.13"
+
+      - name: Verify Docker
+        run: |
+          docker version
+          docker info
+
+      - name: Build benchmark Docker image
+        run: |
+          tools/docker_test.sh build .
+
+      - name: Run Testbox
+        uses: useblacksmith/run-testbox@5ca05834db1d3813554d1dd109e5f2087a8d7cbc # v2
+        if: always()
+        env:
+          RUSTWRIGHT_DOCKER_IMAGE: rustwright-verify-testbox
+          TEST_DOCKER_MEMORY_LIMIT: 8g
+          INSTALL_BROWSER_HARNESS: "1"
+          INSTALL_PUPPETEER: "1"
+          DOCKER_BUILDKIT: "1"

rustwright-0.1.0a2/.github/workflows/benchmark.yml

@@ -0,0 +1,297 @@
+name: Rustwright Benchmarks
+
+on:
+  pull_request:
+    types:
+      - opened
+      - synchronize
+      - reopened
+      - ready_for_review
+      - labeled
+  workflow_dispatch:
+    inputs:
+      benchmark:
+        type: choice
+        description: "Benchmark workload to run"
+        required: true
+        default: strict
+        options:
+          - strict
+          - equivalent
+          - mind2web-sharded
+          - defensible-speed
+      implementations:
+        type: string
+        description: "Comma-separated implementations, or all"
+        required: true
+        default: "rustwright,playwright"
+      repetitions:
+        type: string
+        description: "Benchmark repetitions"
+        required: true
+        default: "1"
+      iterations:
+        type: string
+        description: "Iterations per benchmark run"
+        required: true
+        default: "1"
+      lifecycle:
+        type: choice
+        description: "bench-full lifecycle"
+        required: true
+        default: warm-browser
+        options:
+          - warm-browser
+          - warm-page
+          - cold-browser
+          - cold-container
+      case:
+        type: string
+        description: "Optional strict/equivalent case name"
+        required: false
+        default: ""
+      max_tasks:
+        type: string
+        description: "Optional Mind2Web task cap"
+        required: false
+        default: ""
+      shard_size:
+        type: string
+        description: "Mind2Web shard size"
+        required: false
+        default: "25"
+      defensible_lifecycles:
+        type: string
+        description: "Comma-separated lifecycles for defensible-speed"
+        required: false
+        default: "warm-browser,warm-page,cold-browser"
+      strict_repetitions:
+        type: string
+        description: "Strict-suite repetitions for defensible-speed; defaults to repetitions"
+        required: false
+        default: ""
+      strict_iterations:
+        type: string
+        description: "Strict-suite iterations for defensible-speed"
+        required: false
+        default: "1"
+      enforce_phase2:
+        type: boolean
+        description: "Fail the workflow if strict Phase 2 acceptance checks fail"
+        required: false
+        default: false
+      runner_label:
+        type: choice
+        description: "Runner label for provenance/debug runs; Blacksmith remains the default"
+        required: false
+        default: blacksmith-4vcpu-ubuntu-2404
+        options:
+          - blacksmith-4vcpu-ubuntu-2404
+          - ubuntu-latest
+
+jobs:
+  benchmark:
+    if: >-
+      ${{
+        github.repository == 'Skyvern-AI/rustwright' &&
+        (
+          github.event_name != 'pull_request' ||
+          github.event.pull_request.head.repo.full_name == github.repository
+        ) &&
+        (
+          github.event_name == 'workflow_dispatch' ||
+          vars.RUSTWRIGHT_RUN_PR_BENCHMARKS == '1' ||
+          (github.event_name == 'pull_request' && contains(github.event.pull_request.labels.*.name, 'run-benchmarks'))
+        )
+      }}
+    runs-on: ${{ github.event_name == 'workflow_dispatch' && inputs.runner_label || vars.RUSTWRIGHT_BENCHMARK_RUNNER || 'blacksmith-4vcpu-ubuntu-2404' }}
+    timeout-minutes: 480
+    env:
+      RUSTWRIGHT_DOCKER_IMAGE: rustwright-verify-github-benchmark
+      TEST_DOCKER_MEMORY_LIMIT: 8g
+      INSTALL_BROWSER_HARNESS: "1"
+      INSTALL_PUPPETEER: "1"
+      DOCKER_BUILDKIT: "1"
+      BENCHMARK_KIND: ${{ github.event_name == 'workflow_dispatch' && inputs.benchmark || vars.RUSTWRIGHT_PR_BENCHMARK || 'equivalent' }}
+      IMPLEMENTATIONS: ${{ github.event_name == 'workflow_dispatch' && inputs.implementations || vars.RUSTWRIGHT_PR_BENCHMARK_IMPLEMENTATIONS || 'all' }}
+      REPETITIONS: ${{ github.event_name == 'workflow_dispatch' && inputs.repetitions || vars.RUSTWRIGHT_PR_BENCHMARK_REPETITIONS || '1' }}
+      ITERATIONS: ${{ github.event_name == 'workflow_dispatch' && inputs.iterations || vars.RUSTWRIGHT_PR_BENCHMARK_ITERATIONS || '1' }}
+      BENCHMARK_FULL_ITERATIONS: ${{ github.event_name == 'workflow_dispatch' && inputs.iterations || vars.RUSTWRIGHT_PR_BENCHMARK_ITERATIONS || '1' }}
+      LIFECYCLE: ${{ github.event_name == 'workflow_dispatch' && inputs.lifecycle || vars.RUSTWRIGHT_PR_BENCHMARK_LIFECYCLE || 'warm-browser' }}
+      CASE_NAME: ${{ github.event_name == 'workflow_dispatch' && inputs.case || vars.RUSTWRIGHT_PR_BENCHMARK_CASE || '' }}
+      MAX_TASKS: ${{ github.event_name == 'workflow_dispatch' && inputs.max_tasks || vars.RUSTWRIGHT_PR_MIND2WEB_MAX_TASKS || '' }}
+      SHARD_SIZE: ${{ github.event_name == 'workflow_dispatch' && inputs.shard_size || vars.RUSTWRIGHT_PR_MIND2WEB_SHARD_SIZE || '25' }}
+      DEFENSIBLE_LIFECYCLES: ${{ github.event_name == 'workflow_dispatch' && inputs.defensible_lifecycles || vars.RUSTWRIGHT_BENCHMARK_DEFENSIBLE_LIFECYCLES || 'warm-browser,warm-page,cold-browser' }}
+      DEFENSIBLE_STRICT_REPETITIONS: ${{ github.event_name == 'workflow_dispatch' && inputs.strict_repetitions || vars.RUSTWRIGHT_BENCHMARK_DEFENSIBLE_STRICT_REPETITIONS || '' }}
+      DEFENSIBLE_STRICT_ITERATIONS: ${{ github.event_name == 'workflow_dispatch' && inputs.strict_iterations || vars.RUSTWRIGHT_BENCHMARK_DEFENSIBLE_STRICT_ITERATIONS || '1' }}
+      ENFORCE_PHASE2: ${{ github.event_name == 'workflow_dispatch' && inputs.enforce_phase2 || vars.RUSTWRIGHT_BENCHMARK_ENFORCE_PHASE2 || '0' }}
+      BENCHMARK_RUNNER_LABEL: ${{ github.event_name == 'workflow_dispatch' && inputs.runner_label || vars.RUSTWRIGHT_BENCHMARK_RUNNER || 'blacksmith-4vcpu-ubuntu-2404' }}
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+
+      - uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6.3.0
+        with:
+          python-version: "3.13"
+
+      - name: Verify Docker
+        run: |
+          docker version
+          docker info
+
+      - name: Print benchmark configuration
+        run: |
+          {
+            echo "event=${GITHUB_EVENT_NAME}"
+            echo "benchmark=${BENCHMARK_KIND}"
+            echo "implementations=${IMPLEMENTATIONS}"
+            echo "repetitions=${REPETITIONS}"
+            echo "iterations=${ITERATIONS}"
+            echo "lifecycle=${LIFECYCLE}"
+            echo "case=${CASE_NAME}"
+            echo "max_tasks=${MAX_TASKS}"
+            echo "shard_size=${SHARD_SIZE}"
+            echo "defensible_lifecycles=${DEFENSIBLE_LIFECYCLES}"
+            echo "defensible_strict_repetitions=${DEFENSIBLE_STRICT_REPETITIONS}"
+            echo "defensible_strict_iterations=${DEFENSIBLE_STRICT_ITERATIONS}"
+            echo "enforce_phase2=${ENFORCE_PHASE2}"
+            echo "runner_label=${BENCHMARK_RUNNER_LABEL}"
+            echo "docker_memory=${TEST_DOCKER_MEMORY_LIMIT}"
+          }
+
+      - name: Build benchmark Docker image
+        run: |
+          tools/docker_test.sh build .
+
+      - name: Prepare Mind2Web manifest
+        if: ${{ env.BENCHMARK_KIND == 'mind2web-sharded' }}
+        run: |
+          set -euo pipefail
+          python tools/download_mind2web.py --all-train --retries 5 --timeout 300 --json
+          python tools/import_mind2web.py \
+            --source .benchmark-data/raw/mind2web/data/train \
+            --output .benchmark-data/manifests/mind2web-train-100pct-action-fixtures.json \
+            --include-action-fixtures \
+            --json
+
+      - name: Run benchmark
+        run: |
+          set -euo pipefail
+          mkdir -p .benchmark-data/results
+          timestamp="$(date -u +%Y%m%dT%H%M%SZ)"
+
+          impl_args=()
+          if [ "$IMPLEMENTATIONS" != "all" ]; then
+            IFS=',' read -ra impls <<< "$IMPLEMENTATIONS"
+            for impl in "${impls[@]}"; do
+              impl="$(echo "$impl" | xargs)"
+              if [ -n "$impl" ]; then
+                impl_args+=(--impl "$impl")
+              fi
+            done
+          fi
+
+          case_args=()
+          if [ -n "$CASE_NAME" ]; then
+            case_args+=(--case "$CASE_NAME")
+          fi
+
+          if [ "$BENCHMARK_KIND" = "defensible-speed" ]; then
+            if [ -n "$CASE_NAME" ]; then
+              echo "defensible-speed uses the full predeclared case set; do not pass case filters" >&2
+              exit 2
+            fi
+
+            IFS=',' read -ra lifecycles <<< "$DEFENSIBLE_LIFECYCLES"
+            for benchmark_lifecycle in "${lifecycles[@]}"; do
+              benchmark_lifecycle="$(echo "$benchmark_lifecycle" | xargs)"
+              if [ -z "$benchmark_lifecycle" ]; then
+                continue
+              fi
+              case "$benchmark_lifecycle" in
+                warm-browser|warm-page|cold-browser|cold-container)
+                  ;;
+                *)
+                  echo "Unsupported defensible lifecycle: $benchmark_lifecycle" >&2
+                  exit 2
+                  ;;
+              esac
+
+              BENCHMARK_FULL_ITERATIONS="$ITERATIONS" tools/docker_test.sh bench-full \
+                "${impl_args[@]}" \
+                --suite equivalent \
+                --lifecycle "$benchmark_lifecycle" \
+                --repetitions "$REPETITIONS" \
+                --output ".benchmark-data/results/defensible-speed-equivalent-${benchmark_lifecycle}-${REPETITIONS}x${ITERATIONS}-${timestamp}.json" \
+                --json
+            done
+
+            strict_repetitions="$DEFENSIBLE_STRICT_REPETITIONS"
+            if [ -z "$strict_repetitions" ]; then
+              strict_repetitions="$REPETITIONS"
+            fi
+            BENCHMARK_FULL_ITERATIONS="$DEFENSIBLE_STRICT_ITERATIONS" tools/docker_test.sh bench-full \
+              --impl rustwright \
+              --impl playwright \
+              --suite strict \
+              --lifecycle warm-browser \
+              --repetitions "$strict_repetitions" \
+              --output ".benchmark-data/results/defensible-speed-strict-warm-browser-${strict_repetitions}x${DEFENSIBLE_STRICT_ITERATIONS}-${timestamp}.json" \
+              --json
+          elif [ "$BENCHMARK_KIND" = "mind2web-sharded" ]; then
+            max_task_args=()
+            if [ -n "$MAX_TASKS" ]; then
+              max_task_args+=(--max-tasks "$MAX_TASKS")
+            fi
+            python tools/run_mind2web_sharded.py \
+              "${impl_args[@]}" \
+              --shard-size "$SHARD_SIZE" \
+              --repetitions "$REPETITIONS" \
+              --iterations "$ITERATIONS" \
+              "${max_task_args[@]}" \
+              --output ".benchmark-data/results/mind2web-sharded-${timestamp}.json" \
+              --json
+          else
+            tools/docker_test.sh bench-full \
+              "${impl_args[@]}" \
+              --suite "$BENCHMARK_KIND" \
+              --lifecycle "$LIFECYCLE" \
+              --repetitions "$REPETITIONS" \
+              "${case_args[@]}" \
+              --output ".benchmark-data/results/bench-full-${BENCHMARK_KIND}-${LIFECYCLE}-${timestamp}.json" \
+              --json
+          fi
+
+      - name: Check strict Phase 2 benchmark artifacts
+        if: always()
+        run: |
+          set -euo pipefail
+          mkdir -p .benchmark-data/reports
+
+          enforce=0
+          case "$ENFORCE_PHASE2" in
+            true|True|TRUE|1|yes|YES)
+              enforce=1
+              ;;
+          esac
+
+          args=(
+            --source github-actions
+            --runner "$BENCHMARK_RUNNER_LABEL"
+            --artifact rustwright-benchmark-results
+            --run-url "${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/actions/runs/${GITHUB_RUN_ID}"
+            --json
+          )
+          if [ "$enforce" = "1" ]; then
+            args+=(--enforce-phase2)
+          fi
+          python tools/check_benchmark_artifacts.py "${args[@]}"
+
+      - name: Upload benchmark results
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        if: always()
+        with:
+          name: rustwright-benchmark-results
+          path: |
+            .benchmark-data/results/
+            .benchmark-data/reports/
+          if-no-files-found: warn

rustwright-0.1.0a2/.github/workflows/claude-code-review.yml

@@ -0,0 +1,87 @@
+name: Claude Code Review
+
+# Automated PR review by Claude, on every opened/updated pull request.
+# Requires a repository secret CLAUDE_CODE_OAUTH_TOKEN (create with `claude setup-token`).
+# If the secret is not set, the review is skipped (the job still succeeds) instead of failing.
+# The review prompt is loaded from the BASE branch so untrusted PRs cannot alter
+# review behavior by editing the prompt file in the same PR.
+
+on:
+  pull_request:
+    types: [opened, synchronize]
+
+jobs:
+  claude-review:
+    # Skip PRs opened by Claude itself or by GitHub Actions to avoid review loops.
+    if: |
+      !contains(fromJSON('["claude","claude[bot]","github-actions[bot]"]'), github.event.pull_request.user.login)
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      pull-requests: write
+      issues: read
+      id-token: write
+    steps:
+      - name: Check for Claude token
+        id: gate
+        env:
+          CLAUDE_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
+        run: |
+          if [ -n "$CLAUDE_TOKEN" ]; then
+            echo "enabled=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "enabled=false" >> "$GITHUB_OUTPUT"
+            echo "::notice::CLAUDE_CODE_OAUTH_TOKEN is not set — skipping Claude review. Add it with 'gh secret set CLAUDE_CODE_OAUTH_TOKEN'."
+          fi
+
+      - name: Checkout repository
+        if: steps.gate.outputs.enabled == 'true'
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+        with:
+          fetch-depth: 1
+
+      - name: Delete previous Claude review comment
+        if: steps.gate.outputs.enabled == 'true'
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          # Remove any prior review comment (identified by the marker) so each
+          # push replaces the review instead of stacking comments.
+          gh api "repos/${{ github.repository }}/issues/${{ github.event.pull_request.number }}/comments" \
+            --jq '.[] | select(.body | contains("<!-- claude-code-review -->")) | .id' \
+          | while read -r id; do
+              if [ -n "$id" ]; then
+                echo "Deleting previous Claude review comment: $id"
+                gh api "repos/${{ github.repository }}/issues/comments/$id" -X DELETE
+              fi
+            done
+
+      - name: Load review prompt from base branch
+        if: steps.gate.outputs.enabled == 'true'
+        id: prompt
+        run: |
+          # Load the prompt from the PR's BASE commit (trusted), not the PR head.
+          git fetch --no-tags --depth=1 origin "${{ github.event.pull_request.base.sha }}"
+          BASE="${{ github.event.pull_request.base.sha }}"
+          DELIM="PROMPT_EOF_$(openssl rand -hex 16)"
+          {
+            printf 'content<<%s\n' "$DELIM"
+            git show "${BASE}:.github/prompts/code-review.md"
+            printf '%s\n' "$DELIM"
+          } >> "$GITHUB_OUTPUT"
+
+      - name: Run Claude Code Review
+        if: steps.gate.outputs.enabled == 'true'
+        uses: anthropics/claude-code-action@ff9acae5886d41a99ed4ec14b7dc147d55834722 # v1.0.77
+        with:
+          claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
+          allowed_bots: "dependabot[bot],github-actions[bot],renovate[bot]"
+          prompt: |
+            REPO: ${{ github.repository }}
+            PR NUMBER: ${{ github.event.pull_request.number }}
+
+            ${{ steps.prompt.outputs.content }}
+          # Read-only review: Claude may inspect the repo and the PR, and post one comment.
+          claude_args: >-
+            --max-turns 30
+            --allowed-tools "Read,Glob,Grep,Bash(gh pr comment:*),Bash(gh pr diff:*),Bash(gh pr view:*),Bash(gh pr list:*),Bash(gh issue view:*)"