loop-repair

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).

Loop Repair (mode: repair)

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).

Steps

1. Bootstrap — identical to loop-delivery steps 1–3, with these deltas:
   • mode: repair in loop.yaml.
   • Record the failing command (from the user's invocation — the source of truth) in the Baseline section of docs/loen/<run-id>/state.md BEFORE loen:audit plan; the plan stage deterministically checks that this recorded command appears among quality_gates.
   • The 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.
2. Human approval gate, then loen:audit plan — must return OK before any edit.
3. Reproduce first (iter-01). BEFORE any edit, run the failing command; capture output
   • exit code into 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.
4. Isolate + minimal fix. The smallest diff that makes the failing command pass. Save 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.
5. Regression coverage. Either (a) a new/extended test in the diff, or (b) the originally-failing test IS the regression test. For case (a) YOU (the worker — the verifier stays read-only) produce inversion evidence into gates.log: git stash push -- <fix files> → run the regression test (must FAIL) → git stash pop → run it again (must PASS).
6. Check. Run the quality_gates into iterations/iter-NN/gates.log, then run loen:audit check — it dispatches the verifier and writes iterations/iter-NN/verifier.md.
7. Fix. Address only verifier-confirmed issues. Repeat Act→Check within budget.max_iterations (default 3 — the methodology default for repair).
8. Report. On green gates + verifier APPROVE run loen:audit result. pr-summary.md MUST state the root cause and a rollback note.

Done condition (gated by loen:audit result)

1. The originally-failing command exits 0 (evidence in the final gates.log).
2. Regression coverage evidenced — case (a) new/extended test with logged inversion evidence, or case (b) the originally-failing test itself; the verifier states which case applies.
3. The diff is minimal — no test changes except the added regression test; every non-test hunk required for the originally-failing command to pass.
4. Verifier APPROVE.

Stop conditions

• The failure does not reproduce → stop BEFORE any edit, report.
• budget.max_iterations exhausted → stop; report root-cause analysis, best attempt, and the blocker.
• Any 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.

Red Flags — STOP

• "Fixing" a failure you have not reproduced → stop; no reproduction, no fix.
• A non-test hunk not required for the failing command to pass → out of scope, drop it.
• Changing tests beyond ADDING the regression test → not allowed.
• Claiming regression coverage without the logged inversion evidence (stash, regression test FAILS, pop, it PASSES) → not proven.
• Reporting the failure fixed without the originally-failing command exiting 0 in the final gates.log → not fixed.
• A 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.