agent-resolve
Skill10 starsChanged 40 days agoMITView on GitHub
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.
- Commits and pushes
In this repository7 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- agent-loopSkill
- agent-triageSkill
- buildSkill
- deploySkill
- releaseSkill
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 1 is distinctive, and 6 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
6
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 18
Pick one: If the user named an issue/PR, use it.
- Very specific: line 21
Lock it + isolate in a worktree: Add the agent-wip label (prevents agent-loop/double-work).
- Distinctive: line 38
never: commit to / push main (feature branch only).
- Very specific: line 42
Needs human review + merge: After merge, deploy with the deploy skill." Remove agent-wip.
- Very specific: line 51
Every resolve runs in its own git worktree, never the shared main checkout.
- Very specific: line 58
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 …
- Very specific: line 62
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 …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
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-agentdraft PR that is not yetagent-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-wiplabel (preventsagent-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>" mainthencdinto<repo>/../llmux-wt/issue-<N>. (-Bresets that path's branch to currentmain.) - 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:<why>"), removeready-to-agent, addagent-blocked+ the right downgrade (not-in-repo/needs-design), dropagent-wip, leave the draft PR with the analysis. Move on. - Implement in the worktree. Respect AGENTS.md: typed errors, no
unwrap/expectin prod paths, never log raw creds (proxy::logging::mask_credentials), config writes viaconfig::update(read-merge-write). - Gate:
just checkgreen (fmt + clippy -D warnings + tests). Add/adjust tests for the change. - Kill-switches (hard — enforce, don't rationalize past them):
just checkstill red after 2 honest fix attempts → stop,agent-blocked+ comment, dropagent-wip.- diff > ~400 LOC or > ~15 files → stop, escalate
needs-design, do not commit. - touches secrets/auth,
Cargo.tomlversion,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 pushthe branch (token fallback in cd-reference if the remoteghs_token is stale). - Flip to review-ready:
gh pr ready <N>. Set the PR body: "Agent-authored.just checkgreen. Needs human review + merge. After merge, deploy with thedeployskill." Removeagent-wip. - Report: PR link, what changed,
just checkresult, 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>"thengit -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 listalready shows one for this issue, reuse it (cdin,git pull/rebase ontomain) 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 checkred, or with an emoji / AI co-author line. - Touching
main,Cargo.tomlversion, 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).
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

