Audit another agent’s work, given a Codex session ID, a Claude Code transcript, a PR, a branch, a CI run, or a pasted summary. The skill describes itself as “a reviewer with a pager who is also a second investigator.” It has four modes, inferred from the request: watch only, audit, audit and fix, and compare. “If authority is unclear, default to audit-only.”
The idea worth taking: investigate independently, then diff. “Anchoring is the failure mode: reading their work first and nodding along.” The watchdog forms its own hypotheses, ideally before reading the other agent’s conclusions, and uses the wait while that agent is still running to pre-map the problem. It prioritizes evidence the watched agent probably did not look at: production ledgers, session replays, error trackers, deploy state, and other worktrees. It then diffs the two investigations (what each found that the other missed, where the approaches diverge, and “any product or design fork they took silently”) and turns the difference into concrete add-on directions to pass along.
Relay sparingly. A mid-run agent “re-plans around every note it receives”, so the watchdog interrupts only for a defect in code being touched now, a correction, or an answer the agent is blocked on. Everything else waits for a checkpoint and goes as one batch. “Never hand a mid-run agent an unranked list… an unranked backlog reliably produces several half-finished surfaces instead of a few complete ones.” It also reports what it verified as correct, “so they stop re-litigating settled ground.”
It classifies each issue as a gap, a bug, a verification miss, scope drift, or no issue. stage-review, with the reviewer set up as a separate agent (often a different vendor) instead of a subagent of the author: pattern-cross-model-review at the level of whole sessions.
See Also
- builder-plan-arbiter — the same cross-agent stance, applied to plans instead of finished work.
- gstack-codex — a second opinion from another model; this goes further and reconstructs the whole run.
- stage-review — the canonical stage this implements.