post-implementation-review

Trigger the review once the implementation agent records turnStatus: done. The supervisor loads review-checklist-manifest.json, spawns the parallel swarm sized to the change, and converges findings into a PASS, PARTIAL, or FAIL verdict.

It is a fixed manifest of checklist items with preset severities that every review agent evaluates identically each time. This replaces free-form prose critique, which causes LLM reviewers to invent new criteria and produce inconsistent results.

Skip it when the turn modified zero files, when skipReview: true is set in the turn summary, or when the agent is not an implementation agent. Running it in those cases wastes review capacity on nothing to inspect.

The verdict depends only on blocking items: zero blocking and zero concerns yields PASS, zero blocking with one to three concerns yields PARTIAL, and one or more blocking findings yields FAIL. Concerns and nits are logged but never trigger fix loops.

Each agent is capped at three free-form findings and every finding must cite a checklist item ID. This forces checklist discipline, prevents nit-chasing and over-polishing, and keeps re-reviews consistent via the review fingerprint.

Delta review is designed but not implemented; no persistence layer exists, so every review runs full and stateless. The delta_review_rules.enabled flag in the manifest is false to match that reality.