name: 🔎 REVIEW.md Drift Audit

on:
  pull_request:
    types: [opened, ready_for_review, synchronize]
    paths-ignore:
      - "docs/**"
      - ".changeset/**"
      - ".server-changes/**"

concurrency:
  group: review-md-drift-${{ github.event.pull_request.number }}
  cancel-in-progress: true

jobs:
  audit:
    # Set the ENABLE_CLAUDE_CODE repository variable to 'false' to turn off Claude
    # jobs; leave it unset (the default) to keep them enabled.
    if: >-
      vars.ENABLE_CLAUDE_CODE != 'false' &&
      github.event.pull_request.draft == false &&
      github.event.pull_request.head.repo.full_name == github.repository
    runs-on: warp-ubuntu-latest-x64-2x
    permissions:
      contents: read
      pull-requests: write
      id-token: write
    steps:
      - name: Checkout repository
        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
        with:
          fetch-depth: 0
          persist-credentials: false

      - name: Run Claude Code
        id: claude
        uses: anthropics/claude-code-action@428971d2ecd6e3a7cb0ee0da2a3a8b33fdb3678d # v1.0.157
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
          use_sticky_comment: true
          allowed_bots: "devin-ai-integration[bot]"

          claude_args: |
            --max-turns 30
            --allowedTools "Read,Glob,Grep,Bash(git diff:*)"

          prompt: |
            You are auditing this PR for drift against `.claude/REVIEW.md`.

            ## Context

            `.claude/REVIEW.md` is the repo's source of truth for what AI / agent code reviewers should treat as critical findings (rolling-deploy safety, hot-table indexes, recovery-path queries, testcontainers usage, Lua versioning, etc.). It is consumed by review agents to calibrate severity. If REVIEW.md goes stale, every future agent review degrades.

            ## Strategy — read this first

            You have a hard turn budget. Spend it on signal, not coverage. The audit is allowed to miss things; it is NOT allowed to time out.

            1. Read `.claude/REVIEW.md` once, in full.
            2. Run `git diff origin/main...HEAD --name-only` to get the list of changed files. Do NOT read the diff content yet.
            3. Scan the file-list for relevance to REVIEW.md scope. Relevance signals: changes to Prisma schema, Redis / queue / Lua code, hot tables, recovery / restart loops, new packages, deletions of paths REVIEW.md cites. Skim everything else.
            4. Open at most **5 files** total — only the ones most likely to surface a real signal. If nothing in the file-list looks relevant to any REVIEW.md rule, do NOT read any files; go straight to the verdict.
            5. Form a verdict and stop. Do not exhaust the turn budget exploring.

            Large PRs (>50 files changed) are a strong signal to be MORE selective, not more thorough. Pick 3-5 files at most.

            ## What to look for

            - **Stale references** — does any REVIEW.md rule cite a file, directory, function, table, Prisma model, or package name that has been removed or renamed in this PR (or is already gone from `main`)?
            - **Contradictions** — does code in this PR clearly violate a current REVIEW.md rule? (Don't re-review the PR. Only flag if REVIEW.md and the PR plainly disagree.)
            - **Missing rules** — does this PR introduce a new pattern future reviewers should know about? Examples: a new hot table, a new Lua-script versioning convention, a new safety wrapper, a new "must always check" invariant.
            - **Obsolete rules** — has the repo moved past a constraint REVIEW.md still asserts? (e.g. a deprecated path is gone, a pattern is now linted, V1 code is deleted.)

            ## Response format

            If nothing needs changing:

            ✅ REVIEW.md looks current for this PR.

            Otherwise:

            📝 **REVIEW.md updates suggested:**

            - **[stale]** `<rule excerpt>` — <what's stale and why>
            - **[contradiction]** `<rule excerpt>` — <what in this PR disagrees>
            - **[missing]** under `## <section>` — <one-sentence draft rule>
            - **[obsolete]** `<rule excerpt>` — <why this rule no longer applies>

            ## Rules

            - Maximum 3 suggestions per audit. Pick the highest-signal ones.
            - Only flag things that would actually mislead a future reviewer. Style and wording do not count.
            - Do NOT review the PR itself. Do NOT propose rules outside REVIEW.md's existing sections.
            - Do NOT propose rules for one-off PR specifics that don't generalize to future PRs.
            - If REVIEW.md does not exist in the repo, respond with `(skip)` and stop.
            - When in doubt between "one more file read" and "finish now" — finish now.