菜单
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).
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
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 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).
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-delivery |
| description | 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). |
Run ONE bounded task as a controlled loop. You are the worker and the only writer.
Subagents (planner, explorer, verifier) run in isolated context and return text — you
persist their output. All artifacts go under docs/loen/<run-id>/ (the loop-guard hook
enforces the layout).
<today> via date +%F; run-id = <today>-<topic>
(<topic> must be lowercase kebab-case, e.g. fix-auth-bug — the loop-guard hook
only accepts run-ids matching ^\d{4}-\d{2}-\d{2}-[a-z0-9-]+$).
With Bash (not the Write tool) create the run dir and the pointer BEFORE any
Write-tool artifact — the loop-guard hook blocks docs/loen writes when no run is active:
mkdir -p docs/loen/<run-id>/iterations && ln -sfn <run-id> docs/loen/current.
Copy the state skeleton from this skill's assets/state.template.md (the skill base
directory is printed when this skill is invoked; assets/ is a sibling of this SKILL.md)
into docs/loen/<run-id>/state.md.planner subagent (isolated), passing the task and
the ABSOLUTE path to this skill's assets/loop.template.yaml (resolved from the skill
base directory — not ${CLAUDE_PLUGIN_ROOT}, which is unset outside hooks). Instruct the
planner to emit mutable_scope/protected_scope as block-style YAML lists. It returns a
filled loop.yaml + a plan. Validate the YAML parses, then write
docs/loen/<run-id>/loop.yaml and docs/loen/<run-id>/plan.md.loen:audit plan — must return OK before Act.mutable_scope (the hook
blocks otherwise). Save the iteration diff to
docs/loen/<run-id>/iterations/iter-NN/diff.patch (git diff > …), where iter-NN
is zero-padded to two digits (e.g. iter-01) — the hook and check_layout.sh reject
anything else. Use explorer when you need code evidence without loading files into
this context.quality_gates from loop.yaml; capture output to
iterations/iter-NN/gates.log. Then run loen:audit check — it dispatches the
verifier and writes iterations/iter-NN/verifier.md.budget.max_iterations.loen:audit result
to finalize report.html, write pr-summary.md, and mark the docs/TODO.md row.budget exceeded → stop, report the best result and the blocker.handoff_conditions trigger (schema / PII / license / architecture / prod creds) →
hard stop, ask the human. Never auto-merge.protected_scope path → stop; the scope IS the contract.handoff_conditions trigger (schema / PII / license / architecture / prod-creds) → hard stop, ask the human.budget → stop; report the best result and the blocker.