// Use when Claude CLI is the chosen local review worker for a bounded diff review or extraction pass and a saved /tmp artifact is required.
Use when a local Codex CLI review, pull-request opening or update compliance check, pull-request comment review loop, or saved Codex review artifact is required for this repo.
Use when editing AGENTS.md, CLAUDE.md, or .codex/skills/** policy, workflow, and manager-skill files. Keeps durable operating rules concise, separates general policy from manager-only policy, and requires review for distorted meaning or regressions.
Use when a task needs a durable active requirement queue and archive workflow. Captures only real user requests or requirements with proofread, sanitized requirement text, source pointers, category, intent, status, next action, and linked artifacts; avoids noisy transcript dumps without erasing user intent.
Use when a manager delegates to subagents, scopes staged worker tasks, decides whether to reuse or rotate workers, or reviews delegated output before relying on it.
| name | claude-cli-review |
| description | Use when Claude CLI is the chosen local review worker for a bounded diff review or extraction pass and a saved /tmp artifact is required. |
Use Claude CLI only when AGENTS.md and Tool And Model Policy allow it. Treat it as weaker evidence than GPT-5.5; managers must verify its output before final decisions.
The current repo-approved Claude review model is Claude Opus 4.7. Keep the concrete CLI model ID in the health-check command variable below so future model updates have one obvious edit point.
CLAUDE_REVIEW_MODEL=claude-opus-4-7
claude -p --model "$CLAUDE_REVIEW_MODEL" --effort high "Reply exactly with: ok"
Expected output: ok.
/tmp/claude_review_<short_sha>.txt.subprocess.run(...) wrapper with a hard timeout.Prompt shape:
Review this git diff for real correctness, regression, security, data-boundary, policy, repo-rule, PR compliance, review-gate, test/QA evidence, excessive-scope, and unintentional-drift issues. Treat real issues as P0/P1/P2 findings by risk under AGENTS.md. Ignore style nits. Return exactly "No findings." if clean.
Report whether Claude returned No findings., whether it timed out, and any concrete findings with file paths.