review-coding-agent-design

Provide the design document and the code or diff, then the review builds a traceability matrix mapping each design claim to implementation and verification evidence. Findings are reported by severity with a final verdict such as Approve, Request redesign, or Insufficient evidence.

It checks both directions: promises in the design document that the code does not fulfill, and behavior, state, or failure modes introduced by code but absent from the document. Each material claim gets a status of implemented, partially implemented, contradicted, or unverified.

Yes, but the review is limited to the design itself and will not claim implementation alignment. With only code, the intent is reconstructed from authoritative project sources and explicitly labeled as inferred.

No. The review is read-only by default and does not edit code, commit, push, approve changes, or write to external systems unless you separately request that action.

Typical issues include invented fields or APIs, scope drift across layers, hard-coded heuristics replacing semantic design, swallowed errors, duplicate mechanisms bypassing authoritative paths, and verification claims without reproducible evidence. Each finding must be tied to repository evidence.

Use design review when judging structural soundness and design-to-code consistency of agent-produced work. Use change-risk-review when the main goal is classifying the delivery risk of current Git changes before commit.