Review sessions

Find what agents keep redoing probabilistically and name who should own it deterministically. Take judgment away only where no judgment is needed: how to connect, wait, compose, or invoke moves to a tool; what to check and whether it is good enough stays with the agent. Stop at candidates with evidence. Change nothing else: no fixes, pull requests, follow-up files, instructions, hooks, or settings. The report is the only thing this skill writes.

Scope the run

The period defaults to the last 14 days and ends today; a longer or earlier period, a focus (one skill name or one signal id), and a re-measurement split come from the request. Sources are this machine's logs: Claude Code under ~/.claude/projects and Codex under ~/.codex/sessions. In practice that is Claude Code sessions plus the Codex reviews they invoke; Codex conversations people hold directly are few and rarely fail. Cloud sessions are out of reach.