lanekeeper-pr-review

The reviewer-agent procedure for Lanekeeper pull requests. Re-run the gates yourself, check scope and contract boundaries, trace every S# to code and a test, check every P# against registered benchmar

SKILL.md

GitHub

Reviewing a Lanekeeper PR

You are the gate before a human sees this PR. Re-run the checks yourself; don't trust the PR text. Be specific. Each finding gives a file and line, the rule it breaks (S#, P#, D#, or an AGENTS.md rule), and the fix.

Procedure

  1. Scope

    • Every changed path is owned by the PR's prompt (see its metadata table), or is a test, benchmark or fixture under those paths.
    • Contract paths are untouched, unless the PR is labelled contract-change and contains nothing else. Use the lanekeeper-contract-change skill to judge it.
    • plans/NN-*.md exists, and the commits follow its task order (NN/Tk: …).
  2. Gates. Run these yourself:

    • just verify-NN
    • just verify
    • cargo xtask bench-check
    • just fuzz-smoke, if parsers changed

    Paste the results.

  3. Security. Use the lanekeeper-security skill.

    • For each S# in the prompt, and any the change touches, locate the implementation and the test that proves it.
    • Scan for the red flags listed in that skill.
    • Confirm that docs/threat-model.md was updated for the component.
  4. Performance. Use the lanekeeper-performance skill.

    • Every P# has a registered benchmark that measures the real path, at target scale, within budget.
    • No anti-patterns: per-request parsing, OFFSET, N+1 queries, blocking in async, unbounded collections.
  5. Domain correctness. For engine, registry or MCP changes, use the lanekeeper-config-semantics skill. Verify:

    • property sources merge setting by setting, and resource files are chosen whole;
    • rendering substitutes only keys present in the property view;
    • tenant sets are respected;
    • line endings are normalised for comparison and restored on write;
    • pickup states, not "pending restart".
  6. Tests

    • Behaviour is asserted, not implementation details.
    • Error paths are covered.
    • No sleeps; no tautological assertions; no mocking of owned code (port fakes are fine).
    • Golden and fuzz tests exist where the prompt requires them.
  7. UI changes. Use the lanekeeper-ui-from-design skill. Visual and keyboard tests are present and green. Every deviation is recorded.

  8. Open questions. Every one the prompt touches uses its stated default, and is configurable. Each assumption is listed and reasonable.

Output: one review comment, exactly this structure

## Verdict: APPROVE | CHANGES REQUESTED

### Blocking
1. <file:line> — <problem> — violates <S#/P#/D#/rule> — fix: <concrete change>

### Non-blocking
1. ...

### Evidence
- just verify-NN: <pass/fail + summary>
- just verify: <...>
- bench-check: <table rows for this PR's P#>
- Security trace: S# → impl file → test name (one line each)
- Threat model updated: yes/no (section)

Approve only when there are no blocking findings. Unverifiable evidence, such as a gate you couldn't run, is itself a blocking finding.