Review

Invoke the review skill against the current epic; it reads the epic description, runs baseline quality gates, then spawns specialized reviewers in parallel based on diff size. Findings are consolidated, deduplicated, and classified by severity before a final approval gate.

Small diffs under 100 lines use 4 core reviewers covering security, test coverage, simplicity, and Codex patterns. Medium diffs of 100-500 lines add architecture, performance, and scenario reviewers for 9 total, while large diffs over 500 lines run all reviewers.

Each reviewer is pre-loaded with 3-5 relevant past lessons retrieved via memory search, filtered by domain match and recency. If a finding contradicts a high-severity lesson, the contradiction is escalated to a human rather than auto-resolved.

Runtime verification and Playwright screenshots run only when the epic's Verification Contract requires evidence such as runtime_startup, browser_evidence, or responsive_check. Non-UI projects without these contract entries skip the step and record an informational note.

A missing Acceptance Criteria section is flagged as a P1 process finding, since the plan phase should have generated the table. The review still proceeds against baseline quality gates and risk-based evidence requirements.

No. All P1 findings must be fixed, all acceptance criteria must pass, and the implementation-reviewer must return APPROVED before the phase gate allows proceeding. P1 findings also create tracked issues for follow-up.