| name | agent-loop |
| description | Use when the user says "loop", "반복", "쭉 돌려줘", or wants to run triage then resolve repeatedly until there are no agent-workable llmux issues/PRs left. Thin orchestrator over the agent-triage and agent-resolve skills; dry-run by default. |
agent-loop — orchestrate triage → resolve until drained
agent-loop has no judgment of its own. It runs agent-triage once, then agent-resolve on the
resulting ready PRs one at a time, and stops on clear conditions. Labels are its memory — it
never re-evaluates issues already marked needs-* / not-in-repo / agent-blocked. Inherits
agent-resolve's boundary: it does not merge or 배포.
Steps
- Mode. Default is DRY-RUN: run agent-triage's evaluation and print the resolve plan, change
nothing. Act only if the user said "apply" / "실제로 해" / passed
--apply.
- Run
agent-triage once. (Labels issues, opens draft PRs for ready ones, skips already-labeled.)
- Resolve. Over
ready-to-agent draft PRs, oldest first, run agent-resolve — each in its own
git worktree (per that skill), so resolves are isolated and may run in parallel. Verify
each PR genuinely green (just check + CI) independently before counting it done. Parallelize
only across issues that touch disjoint modules; if their diffs would overlap, serialize them to
avoid merge conflicts.
- Stop when ANY:
- no
ready-to-agent draft PRs remain (the user's "no issues/PRs left" condition),
--max-iterations reached (default 3),- 2 consecutive
agent-resolve runs end in agent-blocked,
- an
agent-resolve run hit a kill-switch needing a human (needs-design / needs-human / security / release).
- Report a run summary: issues triaged (+ labels), PRs made review-ready, PRs blocked (+ why),
and exactly what is left for a human (review → merge →
deploy).
Anti-thrashing
Never re-triage an issue that already carries a needs-* / not-in-repo / agent-blocked label —
a human must clear the label for it to re-enter the pipeline. Without this, the same un-fixable issue
(e.g. the gpt-5.5 client-limitation one) gets re-evaluated to the same dead end every iteration.
Red flags — STOP
- Running in apply mode without explicit user opt-in.
- Re-triaging labeled issues (thrashing) — labels are memory.
- Merging or deploying —
agent-loop inherits agent-resolve's human-gated boundary.
- Running unbounded — always honor
--max-iterations and the stop conditions.
