cross-review

name	cross-review
description	Get a second model's independent review of work this session produced (or review another agent's work here) — blind, structured, and adjudicated rather than averaged. Use when the user wants "a second opinion", "have codex review this", cross-model review before merge, or when stakes justify two independent reviewers.

Cross-review — a second brain, not an echo

Same-model review inherits the same blind spots. A different model family reviewing the same diff catches a different error distribution — but only if the review is blind and structured.

Protocol

1. Blind the reviewer. Delegate via the bridge in a read mode; give the diff/paths and the requirements, never your reasoning or self-assessment: maaaw bridge run --agent codex --mode review-only --task "Review <scope> against <requirements>. Report findings with file:line evidence, severity, and a fix sketch." --run (cross-model-prompting skill applies to the task text.)
2. Structured findings only. The worker-result contract already forces Status/Findings sections; findings should mirror schemas/finding.schema.json (severity, title, file, line, evidence, recommendation, confidence).
3. Adjudicate, don't average. For each finding: verify against the actual code, then rule confirmed / refuted-with-evidence / needs-user. Two models disagreeing is signal — investigate, never split the difference.
4. Close the loop. Confirmed findings become fixes (smallest real fix, /loop if there's an oracle); durable ones become memory (/learn, e.g. failure-pattern records). Refuted findings with a lesson about WHY also go to memory — they save the next review.

When to use which reviewer

• This kit's code-reviewer agent: cheap, fast, same-model — default for routine diffs.
• Cross-model via bridge: security-sensitive changes, gnarly concurrency, architecture calls, anything about to ship broadly, or when the user asks.
• Both in parallel (orchestration skill): highest stakes; adjudicate the union of findings.

Rules

• The reviewer reviews the code, not you. Never send your justifications.
• A review without file:line evidence is an opinion; ask once for evidence, then discard unevidenced findings.
• You own the final verdict — cross-review informs, the orchestrator decides.