agentleFS
OvichOvich/swarmstorm/skills/execute/orchestrate-plan/SKILL.md

orchestrate-plan

Skill14 starsChanged 5 days agoMITView on GitHub

Use whenever execute-plan runs, as the plan's orchestrator: settling the mandate, running the plan phase by phase on the swarm board, dispatching the plan's workers once, posting each phase's tasks to them, watching the board, verifying merges, and handling the branch and pull request. Not for keeping the tracker or deciding when to pause (execute-plan), and not for doing a task (the swarm-worker agent).

  • Commits and pushes

2 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈0

    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. Very specific:

    Read the plan's *Mandate* section.

    line 12

≈ “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

Orchestrate a plan

Load the orchestrate skill and the swarm-board skill's orchestrator side with this one. The first holds the charter protocol, dispatch, briefs and findings; the second everything said on the board, and replaces the orchestrate skill's Watching. Here the charter is the plan's Mandate; a member is a seat's worker for the whole run, given one task per phase. The phases, seats and roles are the plan's; the cut is the write-plan skill's references/phases.md.

The mandate

Read the plan's Mandate section. Ask what it leaves open before the first phase, with the questions of references/mandate.md, and write the answers there and in the line under the tracker.

The run

  • Every phase runs on the board, from the first to the close, whatever its width. The board is the run's record, the person's view and the next phase's memory. Running a phase off it, or merging a seat's work yourself, changes the mandate, which is the person's.
  • One board run for the plan, opened before the first phase (--kind task, the plan's branch as --onto, the plan's URL as --link, --protocol build-verify when verify: is on), one row per seat for the whole run: the builders w1 … w<N>, each role naming its area, and the verifiers.
  • Leave the plan's branch checked out nowhere while the run is open (detach the main checkout): a worker fast-forwards it from its worktree. Re-run each done-when in a detached worktree of your own.
  • Every seat is dispatched once, at the run's open, all at once: builders on the plugin's swarm-worker agent in their own worktrees, verifiers on verifier, with the briefs of references/brief.md, read before the first dispatch. A worker stays for the whole run; one whose context runs long is replaced in the same seat.

A phase

  1. Write its tasks now, at most one per seat, as the orchestrate skill's references/tasks.md says, from the phase's goal, every seat's role and area, and what the run taught so far. The tracker's placeholder becomes the phase's seats, each a subgraph whose steps are its task's done-when.
  2. Open it: phase <name> "<text>", the text naming every seat's part, then post each task to its worker. A seat with no task in this phase is told done <name> now.
  3. Watch the board: it replaces the periodic check and the digest. Each seat's state (working, stopped, lost, away, ended) and the swarm stopped and lost <member> notifications are read and answered as the swarm-board skill's orchestrator side says; a drift is answered on the board, by its interrupt.
  4. A merge: re-run the task's done-when on the merged branch, and what the testing prompt asked for; tick its steps.
  5. The phase closes when every task is ticked and the full check is green on the merged branch: the board's all-done, then that check. Only then is the next phase's tasks written. A worker that stopped on a finding: fix the task (or the plan, for a decision) and post it again, or pause when the finding changes the phase.

Two seats' agreements are theirs. One that changes a decision the plan took becomes a row in the plan's decisions, as the execute-plan skill says.

Verification: the build-verify protocol

When the mandate's verify: is on, the plan runs the orchestrate skill's build-verify protocol (references/protocols/build-verify/protocol.md, read before the run opens): the plan's phases are its builds, its seats its builders, the validator its verifier v1 (one per focus when the mandate names several), dispatched at the run's open with the builders, away until a verify phase. A verify phase is yours to open after each phase or at the end, as verify: says, its message the protocol's. The verifiers' tables are judged as the orchestrate skill's references/findings.md says, and the judged table decides the next cut: the next phase as planned, amended, a fix phase of the rows that stand, or a pause when a finding breaks a decision. Write what changed into the plan before cutting. The mandate's loop: bounds the fix phases a verify may add beyond the plan's own.

A verifier that ended (its context spent) is replaced by a fresh one in the same row. Its source, a project skill or an instructed verifier defined on the fly for the job, is chosen as the protocol says.

Branch and pull request

  • One branch for the plan, one pull request, opened as a draft (gh pr create --draft) on the first push; give the person its URL and write its number into the state node. Marked ready where the pace merges or hands it to the person.
  • Never delete a branch, and never pass the option that deletes one on merge.
  • Whatever merges leaves the branch green: complete, or inert where it is not yet.

The close

Close the run on the board once the last phase closes, before the plan's close. A step the person does (a walk, a trial) stays open in the tracker and never holds the run open.

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.

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.