How it works

You keep a task board in Linear. The dispatcher is a Claude Code session you start with /dispatcher:start <milestone>. It moves each card from Ready to Human Review with agents doing the work in between. You decide what is ready, and you merge.

Click any stage for more info

The board's columns in its own colours. Solid arrows are the main flow; dashed arrows are the two detours.

A card moves left to right. You put it in Ready; the agents take it through In Progress; when the review passes it lands in Human Review; your merge moves it to Done. Problems found in review, or changes you request, send it back to Changes Requested. A question for you parks it in Question.

Inside In Progress

Click any stage for more info

Each step names who does it. Solid arrows are the round every task takes; dashed arrows are the returns.

What stays yours

• Deciding what is ready: moving a task from Backlog to Ready.
• Priority: the board's manual order is the only ranking.
• Answering the questions an agent could not decide alone.
• Reviewing and merging. Your merge is what marks a task Done.
• Anything you want to build yourself: assign yourself with no agent, and the dispatcher leaves that task alone (how).

Stop the loop at any time with /dispatcher:stop. Everything that matters is on the board and on GitHub, so the next start resumes where it left off.

Why it is safe to leave running

• Agent commits and PRs are made by a developer GitHub App and reviews by a separate reviewer app, so GitHub lets you approve every PR and no agent can approve anything.
• A worker's report is checked against GitHub before the board moves.
• Nothing hidden lives in the session: claims, states and PRs are all visible on the board and in the repository.

Next: The board. Everything technical is in the System breakdown.