Cross-agent review¶
Have a second agent review the first agent's work before you do. Especially useful across runtimes — a Claude Code session writing a feature, a Codex session reviewing it for security or simplification.
The flow¶
- Builder peer finishes a chunk of work and commits or stages it.
- Builder calls
ask("reviewer", "review the diff on feat/foo for the auth path"). - Reviewer peer reads the diff (
git diff,gh pr diff), runs whatever checks it cares about, and acks back with findings. - Builder addresses the findings in a follow-up turn.
Why two agents¶
The second agent has no commitment to the implementation. It hasn't spent five turns convincing itself the current approach is fine. That neutrality is the value — not raw capability difference.
Pair runtimes for an additional check:
- Claude Code + Codex — different model lineages catch different things.
- Claude Code + Gemini — Gemini tends to flag different patterns.
- Same runtime, fresh session — works too; the new session loads the diff cold.
Concrete prompts¶
For the builder:
ask("reviewer", "I just pushed feat/auth-rework. Review the diff at
https://github.com/me/repo/pull/42 with a security pass:
authentication, authorization, session handling. Skim everything else.")
For the reviewer (no special prompt needed — repowire-aware agents will run the review on receipt). Tell the reviewer what to focus on, not just "review this."
Mark it reviewed¶
The reviewer can call mark_reviewed(pr_url) after the pass. The PR drops off the reviewer's review_queue at the recorded SHA. If the builder pushes new commits later, the PR resurfaces as re-review-suggested.
See also¶
- Orchestrator coordination for review at fleet scale.
mark_reviewedandreview_queuereference.