| name | agent-resolve |
| description | Use when the user says "resolve", "해결해줘", "픽스해줘", or wants to take one ready-to-agent llmux issue/PR, implement the fix on its branch, get `just check` green, and hand back a review-ready PR. Does NOT merge or deploy by default. |
agent-resolve — one issue → green, review-ready PR
Core safety boundary: agent-resolve produces a green, review-ready PR and STOPS. It does not
merge to main and does not 배포. In this repo "배포" pushes to a public preview channel
(main → CI preview → brew); that recovery cost is asymmetric, so a human gates merge + deploy.
The original request was "resolve and 배포". That is intentionally downgraded to a human gate
(validated by strategist consult). Auto-ship exists only behind an explicit opt-in (bottom) and is
OFF by default. Build/deploy mechanics live in .claude/skills/_shared/cd-reference.md.
Steps
- Pick one. If the user named an issue/PR, use it. Else the oldest
ready-to-agent draft PR
that is not yet agent-wip:
gh pr list --repo 2lab-ai/llmux --draft --label ready-to-agent --json number,headRefName,createdAt.
- Lock it + isolate in a worktree. Add the
agent-wip label (prevents agent-loop/double-work).
Then work in a dedicated git worktree — never the main checkout, so parallel resolves and
the user's live tree never collide (see Worktree isolation below):
git -C <repo> worktree add -B agent/issue-<N>-<slug> "<repo>/../llmux-wt/issue-<N>" main
then cd into <repo>/../llmux-wt/issue-<N>. (-B resets that path's branch to current main.)
- Re-verify it is actually fixable in-repo (gate item #1). If mid-work you discover it is not
(the gpt-5.5 / out-of-repo pattern): STOP — do not ship a non-fix. Post a root-cause comment
(
file:line, or "out-of-repo: "), remove ready-to-agent, add agent-blocked + the right
downgrade (not-in-repo / needs-design), drop agent-wip, leave the draft PR with the analysis. Move on.
- Implement in the worktree. Respect AGENTS.md: typed errors, no
unwrap/expect in prod
paths, never log raw creds (proxy::logging::mask_credentials), config writes via config::update
(read-merge-write).
- Gate:
just check green (fmt + clippy -D warnings + tests). Add/adjust tests for the change.
- Kill-switches (hard — enforce, don't rationalize past them):
just check still red after 2 honest fix attempts → stop, agent-blocked + comment, drop agent-wip.- diff > ~400 LOC or > ~15 files → stop, escalate
needs-design, do not commit.
- touches secrets/auth,
Cargo.toml version, v* tags, brew, or .github/workflows/ → stop, needs-human.
- never commit to / push
main (feature branch only).
- Commit (conventional, lowercase, no emoji, no AI co-author line) and
git push the branch
(token fallback in cd-reference if the remote ghs_ token is stale).
- Flip to review-ready:
gh pr ready <N>. Set the PR body: "Agent-authored. just check green.
Needs human review + merge. After merge, deploy with the deploy skill." Remove agent-wip.
- Report: PR link, what changed,
just check result, and the explicit next human step
(review → merge → deploy).
- Tear down the worktree once the PR is open — the branch lives on the remote, the worktree is
disposable:
git -C <repo> worktree remove "<repo>/../llmux-wt/issue-<N>" then git -C <repo> worktree prune. On a kill-switch stop, leave the worktree in place for human inspection.
Worktree isolation
Every resolve runs in its own git worktree, never the shared main checkout. This is what makes
parallel agent-loop resolves and the user's live editing safe simultaneously — no branch-switching
or index races in the primary tree.
- Path convention (standard):
<repo>/../llmux-wt/issue-<N> — a sibling of the main checkout,
never nested inside it. (<repo> = git -C <repo> rev-parse --show-toplevel.)
- Create/reset:
git -C <repo> worktree add -B agent/issue-<N>-<slug> "<repo>/../llmux-wt/issue-<N>" main.
- A branch can be checked out in only one worktree. If
git worktree list already shows one for
this issue, reuse it (cd in, git pull/rebase onto main) instead of creating a second.
- Remove when done (
git worktree remove + prune); the pushed branch + PR are the durable artifact.
Only if the user explicitly says "resolve and 배포" / "ship it" / passes --ship: after the PR is
green and a human has merged it, run the deploy skill. agent-resolve never merges its own PR
and never deploys unmerged/unreviewed code, even with --ship.
Red flags — STOP
- Auto-merging or 배포-ing a self-authored PR by default. The public channel is human-gated.
- Shipping a "fix" for an out-of-repo issue (the gpt-5.5 trap). If you can't point at the llmux line
you changed and show it moves the behavior, it is not fixed.
- Committing with
just check red, or with an emoji / AI co-author line.
- Touching
main, Cargo.toml version, tags, workflows, or secrets.
- Pushing past a kill-switch threshold "because it's almost done."
- Working in the main checkout instead of a dedicated worktree (races other agents / the user's live tree).
