Adversarial Review

Automated critique-fix loop using fresh-eyes review. Each round, the code is reviewed cold — no knowledge of why decisions were made, no attachment to the implementation. Critical and Suggestion findings get fixed, then a fresh review checks again. Loop stops when only Nitpicks remain or after 3 rounds.

Execution — Harness-Neutral

The contract is the fresh-eyes critique-fix loop, not any one mechanism:

• Claude Code — spawn a new reviewer subagent each round (Agent tool, subagent_type: general-purpose) so the reviewer literally has no memory of the code or prior rounds. This is the strongest form of "fresh eyes."
• Codex / other single-agent harnesses — run the loop yourself: each round, re-read the diff cold and review it as if you'd never seen it, deliberately setting aside what you know about why it was written. No subagent required; the fresh-eyes discipline is what matters, not the process boundary.