self-review

Self-review loop

You cannot review your own work with your own eyes: the mistakes you just made are the ones you were blind to a minute ago, and re-reading with the same context re-derives the same conclusions. What works is what engineering teams do: hand the change to people who were not there, let them attack it from different angles, check their claims, fix what is real, and look again — because a fix is a new change that can itself be wrong.

This skill is that process, with subagents as the people. It runs at the end of every turn that changed code, and on demand for anything else. It is budgeted, and the section below is where that budget is set and why.

What a review costs, and the two rules that keep it cheap

1. Wait inside one call, never across turns. After launching reviewers, the next call is "${CLAUDE_PLUGIN_ROOT}/scripts/wait.mjs" --work <work> --round <n> [names…], with the Bash tool's timeout set to 600000. It blocks until every named reviewer's transcript shows a finished report or has been silent past the stale limit, then prints one line per reviewer and exits: 0 — all settled, collect now (§2c); 1 — some still active and the round's 30-minute wait budget is not spent, call it again as the very next tool call; 3 — budget spent, treat the active ones as dead (§2f). A reviewer it lists as stalled is neither: see §2f. Do not end the turn to wait, and do not check on a reviewer any other way: ListAgents, TaskOutput(block=false), Monitor, sleep — every one is a full-context turn that tells you nothing the wait did not, and poll-guard denies the first two from the third call on. The harness's wake-ups still arrive — an idle notification for a named agent, a <task-notification> for an unnamed one — but they are a courtesy, not the signal: they carry no report, and on 2026-09-03 three reviewers reported into the lead's inbox with success:true and nothing was delivered for 2h49m while the transcripts on disk said finished within ten minutes. If one wakes you between calls, do not act on it; go back to the same call. The Stop gate still releases a turn while subagents run; that is the fallback for a wait.mjs that cannot run — exit 2 with "no subagent transcripts", not an exit 2 naming a bad flag, which you fix and re-run — and never the way to wait.
2. Act once per round, in one go. When wait.mjs exits 0, act once: verify, write the directives, dispatch the applier, wait again on its name, then pre-flight, ledger, record, next round or marker — each stretch in as few tool calls as it takes, except the marker, which gets a message of its own (§4). Never start fixing from partial results while reviewers are still out: fixes move the tree under the reviewers still reading it. Read cited line ranges (sed -n 'a,bp'), not whole files.