code-review
Code Review - Multi-Agent Local Review
Overview
Runs five specialized review agents in parallel, each focused on exactly one concern, then runs a verifier pass that keeps only the evidence-backed findings, then merges them into a single report sorted by severity. Keeps each reviewer in its lane so nothing drifts into generic "looks fine" commentary.
Core principle: Narrow-scope reviewers find more real issues than one broad reviewer trying to cover everything.
Model tier: Designed for Opus or Sonnet-tier models. Haiku may need more guidance — prefer Sonnet+ for the reviewer agents.
The Five Reviewers
- basics — surface-level hygiene (unused imports, dead code, stale comments, debug statements, orphaned symbols)
- architecture — pattern fit with siblings (must read 3-5 sibling files first; flags structural holes peers would have filled)
- clarity — readability, naming, function length, nesting (focus on what changed; same-scope pre-existing issues surface as
[Pre-existing]) - testing — coverage of the change and assertion strength (membership checks, weak truthiness, missing error-path tests)
- repo-hygiene — secrets, env var documentation drift, manifest/lockfile alignment, README/CLAUDE.md/AGENTS.md drift
All five run concurrently as a single Task dispatch. A sixth verifier pass then filters the merged candidates down to the evidence-backed findings before they reach REVIEW.md at the repo root. Full per-lane prompts in Phase 2; verifier prompt in Phase 2.5.