agentleFS
prestariusprestarius/jet-skills-stack/skills/coordinator-workers/SKILL.md

coordinator-workers

Skill1 starsChanged 2 months agoMITView on GitHub

Run a plan's issues with one subagent worker per issue, each in its own git worktree and branch, while this session stays the coordinator that briefs, reviews every diff, merges, closes issues and waits on outside events. Use when a repo has issues split from a plan (plan-to-issues) and more than one can run at once, or when the user says "let your agents build it", "fan out the issues", "one worker per issue".

  • Reads credentials
  • Commits and pushes

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 5 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read the repo's AGENTS.md/CLAUDE.md, style guide and ADRs.

    line 16
  2. Distinctive:

    Every brief holds, in this order:

    line 25
  3. Distinctive:

    Write the numbers in the brief from the data, not from memory: a worker once caught the coordinator's "ok 11" when the file held 10.

    line 40
  4. Distinctive:

    Check the issue's "Done when" yourself, the way the repo says (route, render, byte compare), not through the worker.

    line 51
  5. Very specific:

    Remove the worktree: git worktree remove ../<repo>-wt-<n> --force.

    line 55
  6. Distinctive:

    Never a tight loop; take the delay the other side gives.

    line 64

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

coordinator-workers

A Coordinator and Workers that never talk to each other. Workers get a brief, a worktree and a branch; they push and report; the Coordinator reads the diff, merges and closes. The Agent tool, git worktrees and gh are enough.

1. Before any worker starts

  1. Read the repo's AGENTS.md/CLAUDE.md, style guide and ADRs; they are the rules every brief repeats.
  2. Land the shared contract first, on main, in one commit: the types file, the signature, the schema. Workers start against that commit, never against a promise.
  3. From the issue list, mark which issues block which. Only unblocked issues go out together.
  4. Pick the model per issue the way the repo says: the strongest for logic, a cheaper one for views and tests.

2. The brief (one per worker, Agent tool, model chosen above)

Every brief holds, in this order:

  • the repo path, the issue number, and "do only that issue";
  • the worktree command, verbatim: git -C <repo> fetch -q && git -C <repo> worktree add ../<repo>-wt-<n> -b issue-<n> origin/main, then "work only inside that folder", then the install command;
  • the files to read (rules, the contract, the files it will touch, gh issue view <n>);
  • what to build, with the edge cases spelled out (missing data, empty input, bad input) and where in a shared file to put its lines, so two workers' edits to one file merge clean;
  • what not to touch (fixtures, .env, other workers' files, the contract);
  • the checks to run, with the exact expected outputs;
  • the commit message shape (..., closes #<n>), the attribution lines, git push -u origin issue-<n>;
  • "do NOT merge, do NOT close the issue, do NOT comment on it";
  • the report shape: commit sha, files changed, the checks' output verbatim, what it was unsure about, at most 12 lines.

Write the numbers in the brief from the data, not from memory: a worker once caught the coordinator's "ok 11" when the file held 10. When a worker reports a mismatch, count again before deciding.

3. When a worker reports

  1. git fetch, then read the whole diff: git diff main...origin/issue-<n>. Accept nothing on the report alone.
  2. Merge (--ff-only when it is first, --no-edit after). Run typecheck and tests on main. Two workers who import the same thing leave two import lines after a merge: fix that in the merge commit.
  3. Check the issue's "Done when" yourself, the way the repo says (route, render, byte compare), not through the worker.
  4. Push, then close with the repo's closing comment, crediting the worker's harness, model and effort, e.g. worker: claude-code · sonnet-5.5 · low, commit: <sha>.
  5. Remove the worktree: git worktree remove ../<repo>-wt-<n> --force. Branches may stay on origin.

4. Issues that wait on the outside

An issue whose "Done when" is someone else's act (a review, a reply, a scheduled run) stays open until that act lands, with a comment saying what it waits on. Closing it at send time has been refused by a reviewer ("you closed it before I looked"). The clean-up issue is blocked by it.

Waiting: a background sleep <retryAfter> in Bash, then the same call again with the same answer. Never a tight loop; take the delay the other side gives.

5. Before a public link goes out

Fetch it the way the other side will: a plain curl with no sign-in, byte-compared with the file you uploaded. Where a preview lives is a rule of the working repo (CLAUDE.md or an ADR); a preview is never a branch of the repo unless the repo says so.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

Reviews can't be read right now.

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.