agent-reviewlisted

# Agent Review Use this skill when Codex wants Claude Code, Reasonix, or a manual external agent to review work without taking over the main workspace. ## Rules - Prefer read-only review. - Findings come before suggestions. - Ask for file paths, line references, severity, and concrete risk. - Optional fixes must be written as small suggestions or `patch.diff`. - Codex decides whether findings are valid. - For mandatory dual code or strategy review, use `New-AgentDiscussion.ps1 -AuditProfile scientific -Protocol adversarial-discussion -Agents claude-code,reasonix`. While Reasonix is non-blocking, use `-Agents claude-code -OptionalAgents reasonix` and report that the audit is not dual when Reasonix is absent. - Round 1 is blind; Round 2 addresses only disputed, blocking, or weak-evidence findings. - Do not treat Reasonix Desktop launch, instruction files, or single-reviewer completion as review evidence. - Do not decide by model vote. Use `findings.json`, `disagreement-matrix.md`, and verified calibration outcomes. ## Review Packet Contents Include: - user request - files or diff under review - expected behavior - tests already run - tests still needed - severity scale - forbidden actions ## Required Review Output The external agent should write `review.md` with: - Findings - Evidence - Suggested fixes - Residual risks - Tests recommended For scientific review, use the finding block required by `brief.md`, including evidence, confidence, counter-evidence, blocking sta