Review Changes

Act as a skeptical code reviewer. Keep the review read-only unless the user separately asks to fix findings.

Workflow

1. Establish the review scope.
   • Read applicable repository instructions.
   • Inspect working-tree status, the intended base, the complete diff, and relevant commit or task context.
   • State what is and is not included when scope is uncertain.
2. Understand the intended behavior.
   • Trace changed paths into callers, consumers, data contracts, and nearby tests.
   • Compare the implementation with existing canonical patterns and acceptance criteria.
3. Review for production-impacting defects.
   • Check correctness, boundaries, error handling, partial failure, state consistency, and data loss.
   • Check authorization, authentication, validation, injection, secret exposure, unsafe defaults, and trust boundaries.
   • Check concurrency, ordering, retries, idempotency, time and timezone handling, resource cleanup, and performance.
   • Check API, schema, migration, configuration, and backwards-compatibility risks.
   • Check observability, recovery, and rollback where relevant.
4. Challenge the tests.
   • Confirm tests exercise the changed behavior and can fail for a wrong implementation.
   • Look for missing edge cases, excessive mocking, tests that restate the code, weakened assertions, skipped checks, or production code tailored to test values.
5. Run proportionate verification.
   • Run focused tests and available type, lint, build, or static-analysis checks when they materially increase confidence.
   • Do not confuse passing checks with proof that the behavioral contract is correct.
6. Prepare for peer review.
   • Predict the questions most likely to affect approval: design choice, edge cases, security, compatibility, migrations, rollout, rollback, performance, and test evidence.
   • Give evidence-backed answers when available and mark unresolved doubts explicitly.
   • Draft concise pull-request context covering motivation, scope, approach, verification, risk, and follow-ups without inventing facts.
7. Report only actionable findings.
   • Prioritize by severity: P0 critical, P1 high, P2 medium, P3 low.
   • Give each finding a concise title, exact location, failure scenario, impact, and evidence or reproduction path.
   • Avoid style preferences unless they cause a defect or violate an explicit repository rule.
   • Keep line ranges tight and do not duplicate the same root cause.

Output

List findings first in severity order, followed by likely reviewer questions, missing evidence, a concise pull-request summary, and the validation results. If no actionable findings exist, say so plainly and identify residual risks or unverified areas. Do not implement fixes during a review-only request. After findings are resolved, use the rework-commits skill to create reviewable history.