review-plan

Review Plan

Responsibility

Judge one plan. Return findings or a clean verdict; do not revise the plan, implement code, or mutate GitHub.

Isolation

Run in a fresh context. Receive only:

• original task;
• trusted root/scoped instructions and architecture docs;
• policy classification;
• the plan artifact and current source needed to verify it.

Do not receive the planner's conversation. Bind the result to the plan digest and exact base/head SHA.

Review Lenses

• Correct module owner and production seam.
• Complete affected paths, consumers, serialized boundaries and lifecycle.
• Policy classification, risk, reviewers, checks and human-review decision.
• Compatibility assumptions and explicit user decisions.
• Two real analogous anchors.
• One claim-to-test row per behavior with a production entry point, revert-failing assertion and meaningful negative cases.
• Failure, cancellation, cleanup, concurrency, security, UX and recovery cases appropriate to the scope.
• Verification commands that cover every classified area.

Output

Produce JSON matching .agents/schemas/review.schema.json with subject: "plan" and reviewer_role: "review-plan". clean means zero findings. Every blocked finding needs a stable ID, severity, category, path/line evidence and required fix.

An iteration ceiling remains blocked. Never lower a finding to let work start.