Dispatch — parallel task execution + re-integration

Use the harness's session tools to fan a pool of interdependent tasks out across parallel sessions (one session per task) in dependency-ordered waves, collect each session's branch + commit + evidence, merge the work back into one integration branch, and archive the finished sessions. You are the orchestrator: you plan waves, spawn workers, coordinate, integrate, and verify — you do not implement the tasks yourself.

Dispatch is about the orchestration mechanics (create_session, send_session_message, delete_item, SQL tracking, re-integration), not a fixed recipe. What each worker does is whatever the user's prompt specifies — dispatch just carries those instructions into each session, scoped to one task. Do not inject a quality chain (tests, reviews, smoke, etc.) the user did not ask for.

When to use