← ClaudeAtlas

retrolisted

Reviews a finished /orchestrate run and proposes the smallest change to the pipeline docs that would have prevented what went wrong — or says nothing needs changing. --audit reviews the pipeline docs themselves: size, retirement candidates, dead references, review-cycle trend.
Zalaras/muster · ★ 2 · Code & Development · score 71
Install: claude install-skill Zalaras/muster
> **Maintainer note:** Runs in the main session, normally at the end of the session that ran > `/orchestrate` — the orchestrator's own stumbles (a mis-stamped state, an improvised wave, a > re-sent prompt) live only in that conversation. It also works later, from the files alone; > say so in the header when it does. Authored 2026-09-03 because ad-hoc retros netted +10 to > +50 lines each into `.claude/` (`orchestrate/SKILL.md` went 393 → 545 lines in 12 days); > re-cut 2026-09-11 when the line budget turned out to be satisfied by longer lines. Argument: **$ARGUMENTS** - No argument → the plan is inferred: `plan/<name>` if HEAD is on one, else the `plans/*/orchestration-state.json` with the latest `updated_at` whose `status` is `completed`. Name it in the report header. - `<plan-name>` → that run. - `--audit` → § 5, nothing else. ## 1. Gather evidence — measured, not remembered ```bash S=.claude/skills/orchestrate/scripts/orch-state.py python3 $S <plan> timings && python3 $S <plan> show # wall-clock, retries ls plans/<plan>/ # review cycles (review.cycle<N>.md), decisions/ git log --format='%h %s' main..plan/<plan> # commit suffixes, who committed what ``` Read every `review*.md` (all cycles), each `## Fix Attempt` in the implementation logs, the `## Repairs` and `## E2E Implementation Bugs` tables in `test-specs.md`, and any `decisions/*/decision.md`. Then add what only you saw: every point in this se