pipeline-retrolisted

# Pipeline Retro Independent retrospective for a completed (or aborted) dev-pipeline run. The dev-pipeline scores its own eval at Stage 9+ — this skill exists because **the executor grading its own homework is structurally generous**. Everything here is scored from on-disk and on-GitHub artifacts by fresh context, never from the executing session's memory of itself. **Usage:** `/pipeline-retro <issue-number>` — or no argument to use the most recently updated file in `.claude/pipeline-state/*.json`. **Hard rules:** - `eval-criteria.md` is LOCKED — this skill never edits it. Criteria problems become a _proposal_ in the report, acted on by the human between optimization loops. - The original `{issue}-eval.json` is never mutated. The retro writes its own artifact. - The independent scorer is a **fresh subagent** with no access to this conversation — it sees only the artifacts listed below. ## Step 1: Gather run artifacts ```bash ISSUE=<n> S=../dev-pipeline/statectl.sh cat .claude/pipeline-state/${ISSUE}.json # state: stages, checkpoints, deviations, failureContext cat .claude/pipeline-state/${ISSUE}-eval.json # the run's SELF-score # The run report — Stage 9's durable narrative, written before the terminal # narration so an API disconnect cannot destroy it. Absent = either a pre-schema # run or a run that never reached Stage 9's pr-add. [ -f ".claude/pipeline-state/${ISSUE}-report.md" ] && cat ".claude/pipeline-state/${ISSUE}-report.md" bash ../dev-pipeline/too