메뉴
Use when a specific test, CI job, or regression is failing and must be fixed under a reproduce-first controlled loop with proven regression coverage. Not for delivering a change (use loop-delivery) or metrics (use loop-autoresearch).
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when improving ONE numeric metric under a controlled research loop with a fixed eval and kept/reverted experiments. Not for a feature (use loop-delivery) or a failing test (use loop-repair).
Use when delivering ONE bounded change — a feature, refactor, or chore — as a controlled, audited loop in any repo. Not for a failing test (use loop-repair) or a numeric metric (use loop-autoresearch).
Use when a loen loop stage — plan, act, check, or result — must be validated and gated before the next one. Mode-aware for delivery/repair/research; the execution-loop analog of check-chain.
Use when you need a cross-run dashboard over all docs/loen/ runs, or --triage to turn failing runs into proposed next actions (proposals only; never launches loops or edits runs).
Use when an active, human-approved loen run should keep going multi-turn on its own — wraps it in Claude's native /goal from loop.yaml. Optional; never bootstraps a run or submits /goal itself.
Use BEFORE superpowers:brainstorming for any non-trivial work (new module, CLI flag, API change, arch decision). Triggers on "/fix-intent", "/idd", "capture intent", "intent doc". If brainstorming would start on non-trivial work without an approved intent doc in docs/superpowers/intents/, run IDD first. Brainstorm proceeds only when intent Status: approved.
| name | loop-repair |
| description | Use when a specific test, CI job, or regression is failing and must be fixed under a reproduce-first controlled loop with proven regression coverage. Not for delivering a change (use loop-delivery) or metrics (use loop-autoresearch). |
Fix ONE failing test / CI / regression through the cycle
failure → reproduce → isolate → minimal fix → regression test. You are the worker
and the only writer. This skill is a specialization of loop-delivery: same run
layout, contract, audit gates, subagents, and hook. Shared templates live at
<skill-base>/../loop-delivery/assets/ (single source — never copy them).
loop-delivery steps 1–3, with these deltas:
mode: repair in loop.yaml.docs/loen/<run-id>/state.md BEFORE loen:audit plan; the
plan stage deterministically checks that this recorded command appears among
quality_gates.planner dispatch prompt MUST instruct: mode: repair; a NARROW mutable_scope
(the failing area + its tests); the recorded failing command among quality_gates;
block-style YAML scope lists (guard_protected.sh parses block-style only); leave
eval_command empty.loen:audit plan — must return OK before any edit.docs/loen/<run-id>/iterations/iter-01/gates.log. For a suspected
flaky failure run the command up to 3 times (attempts recorded in state.md); any
failing run counts as reproduced. No reproduction → stop and report — never "fix"
what cannot be reproduced. Python targets: purge __pycache__ (or point
PYTHONPYCACHEPREFIX at a temp dir) before repro, inversion, and gate runs — a stale
.pyc of identical size and mtime second can mask the fix or the failure.iterations/iter-NN/diff.patch (git diff > …, iter-NN zero-padded). Every non-test
hunk must be required for the failing command to pass; tests change only by ADDING the
regression test. Use explorer for code evidence without loading files here.gates.log:
git stash push -- <fix files> → run the regression test (must FAIL) →
git stash pop → run it again (must PASS).quality_gates into iterations/iter-NN/gates.log, then run
loen:audit check — it dispatches the verifier and writes
iterations/iter-NN/verifier.md.budget.max_iterations (default 3 — the methodology default for repair).loen:audit result.
pr-summary.md MUST state the root cause and a rollback note.loen:audit result)gates.log).APPROVE.budget.max_iterations exhausted → stop; report root-cause analysis, best attempt, and
the blocker.handoff_conditions trigger → hard stop, ask the human. Never auto-merge.No new artifacts: iterations/iter-NN/{diff.patch,gates.log,verifier.md} suffice. The
research streams (metrics.jsonl, experiments.jsonl) are simply absent in repair — the
hook allows them, no audit stage requires or reads them.
gates.log → not fixed.handoff_conditions trigger → hard stop, ask the human; never auto-merge.budget.max_iterations exhausted → stop; report the root cause, best attempt, and the blocker.