Code Review

Two modes, chosen by request and risk:

Normal review (default) — you read the changes and the touched files yourself and report evidence-backed findings in chat. No subagents, no report files.
Deep review — parallel reviewer agents, adversarial verification, and a persistent review bundle (review.json, events.jsonl, report.html, summary.md). Run it only when the user explicitly asks for a thorough or deep review, or the diff carries substantive cross-module, security, or concurrency risk. Full process: references/deep-review.md.

Use normal review unless the request or a concrete risk warrants deeper review. The evidence protocol applies in both modes.

code-review