Skip to main content

Run the loop

Start it​

Open Claude Code at the repository root, ideally on a strong model at low effort (the agents pin their own models):

claude --model opus --effort low
/dispatcher:start v1.1.0

Name one or more milestones; that is the only thing the loop will touch. Without one it lists the board's milestones and asks. Optional, in another terminal: dispatcher listen lets board and PR changes wake the loop within seconds instead of at its timer (setup).

What you will see​

  • On the board: a dispatched task moves to In Progress with the developer agent shown as its agent and one claim comment on the issue. When the developer finishes, the agent flips to the reviewer; when the review passes, the task lands in Human Review with no agent.
  • In the session: one line of status per round saying whether it is busy, out of work, or waiting on you, plus anything that needs you: a question, a task it skipped, follow-up work an agent noticed.
  • On GitHub: branches named task/<id>-<slug>, PRs opened by the developer app, a comment-only review from the reviewer app.

What you do​

  • Move tasks from Backlog to Ready, and reorder the board; the top is worked first.
  • Review the PRs at Human Review. Merge, or request changes on the PR; the task goes back to the developer with your review word for word.
  • Answer questions: reply on the issue, then move the task to Ready or Changes Requested.
  • Put Confirm with user on anything you want a word about before it starts.
  • Take a task yourself by assigning yourself with no agent (details).

Stop it​

/dispatcher:stop        # finish what is running, take no new work
/dispatcher:stop now    # stop the running agents and put their tasks back

Run it in the same session. Nothing is lost either way: the next /dispatcher:start resumes from the board. A session that dies without stopping leaves claims behind; they expire after 90 minutes and the next dispatcher takes them over.