review-conformance

Conformance Review (first pass, before quality)

Wrong-but-clean code is worse than ugly-but-right code. Check conformance FIRST.

Three-level check

1. Requirement conformance — For each acceptance criterion of the task: point at the exact code/test that satisfies it. No pointer = finding. Check scope both ways: nothing promised is missing, and nothing out-of-scope was touched (the diff.status list shows every file — files outside the task's declared Files: list need justification).
2. Plan/architecture conformance — Does the implementation follow the approved design: same components, same data flow, same chosen option? Check the code against PLAN.md's Contracts — signatures, data shapes, and API/event contracts must match what was agreed, because other tasks were sliced and reviewed assuming them; a silently changed contract breaks callers this task can't see. Codex substituting its own design is a finding even when it works — the plan was reviewed, the substitute wasn't. If the deviation is genuinely better, surface it to the user as a plan change, don't wave it through.
3. Structure conformance — Files in the right places, naming matching conventions, integration through existing seams (not parallel copies of existing helpers), tests where the project keeps tests.

Evidence rules

• Check the tool result's status field first: only review a success run normally. A partial run (no completion marker / parse errors) means the result may be incomplete — verify everything from the diff and your own check runs, or re-run the task (see review-feedback).
• When the run carried a verifyCommand, read the verification field before anything else: passed: true proves the acceptance command exited 0 in the workspace; passed: false or skipped is an automatic unmet criterion — quote its outputTail in the finding.
• Work from the tool result's diff field (status + patch); read full files when the patch lacks context. Never review from Codex's agentMessage alone — it describes intent, not reality.
• Do not re-run the suite per task: read accepted and verification (the server-run verifyCommand) as the acceptance evidence, and run a manual probe only when the acceptance is not a command. commands in the result shows what Codex ran and exit codes — confirm the targeted tests plus build/typecheck ran green; Codex is not expected to run the full suite (it runs once per merged parallel wave and once in the whole-feature review).