claude-parallel-writers

how a claude code orchestrator runs parallel writers, such as PR owners or dependency workers, each in its own worktree. codex does not use this skill: it runs parallel writers as new threads

• the app refuses start_session in any turn the user's own message didn't start, including turns a writer's message or note wakes. so start every writer your calling skill plans in the turn that invoked it
• start_session with initiation: "user_asked", use_worktree: true and a short name
• a session's worktree can start from a stale local branch, so each prompt names the writer's base and tells it to reset its new branch to the freshly fetched base before it starts
• add your session id from get_session with self to each prompt, and tell the writer to send its result there with SendMessage before it ends its turn. the app's notes on writers' turns arrive only once every writer is idle and quote 400 characters, so read them as session status, never as results
• follow up with SendMessage to the writer's local_… session id
• for a writer that comes up later, or when start_session fails, use the first that works:
  1. start_session as above, in a turn the user's message started
  2. a writer session you started whose work has merged: send it the new task with its prompt values, and tell it to switch its worktree to a new branch from the freshly fetched base before it starts
  3. a sub-agent: read and follow sub-agents.md

complete when every planned writer has a session or a sub-agent, apart from writers your calling skill defers