Retrospective

Run across several sessions, not one. The job is to reduce mistakes, wasted tokens, and restarts in future sessions by improving the skills and config they run under. The headline metric is token-weighted wasted effort — calls and tokens spent on wrong paths, misdirections, revisions, and incorrect tool use that a correctly-applied actuator would have avoided — with two findings that keep standing regardless of token cost: wrong-outcome mistakes (a wrong answer accepted, a premature "done") and abandoned-and-restarted sessions.

This is a closed control loop. Sessions are the only sensor. The actuators (the things you can change) are, ordered by standing cost — how much they cost to keep in place over many sessions, whether or not they ever fire: