| name | relay-fanout |
| description | Run a fan-out relay pattern for independent subtasks. Use when the same kind of work can be split across files, components, services, or targets with minimal coordination. |
| argument-hint | [task] |
| disable-model-invocation | true |
Run a relay fan-out for this task:
$ARGUMENTS
How spawning works
Workers are spawned with Claude Code's built-in Agent tool. The relay is only used for communication between agents.
- Use
subagent_type: "relay-worker". Only relay-worker subagents get the Agent Relay MCP server, the inbox-polling hooks, and the worker protocol. Other subagent types (researcher, general-purpose, …) cannot talk over the relay.
- Run all workers in background mode (
run_in_background: true) so they work concurrently.
- Workers inherit the workspace automatically — the relay MCP server resolves the workspace pinned to this project. Do not put the workspace key in a worker prompt. It is an administrative credential, and copying it into N prompts puts it in N transcripts. If a worker reports no workspace, fix the pin (step 2) rather than pasting the key.
- One relay team per checkout. The pin is per-project and last-writer-wins, so two leads running teams from the same directory will fight over it and a worker can register into the other lead's workspace. Run concurrent teams from separate checkouts or git worktrees. The ACK gate below is what catches this: a worker that landed in the wrong workspace finds no assignment and cannot ACK.
- The
SubagentStart hook injects the relay bootstrap (register, check inbox, ACK, DONE) into every worker.
- Use the relay MCP tools (
send_dm, check_inbox) to monitor progress.
- Do not add setup scripts or dependencies. Use the plugin's existing hooks, MCP tools, and
relay-worker agent definition.
Protocol
- Pick a stable coordinator name —
relay-lead. Pass as: "relay-lead" on every relay tool call you make, so your messages, inbox reads, and reactions stay attributed to the lead.
- Set up the workspace. Call
register_agent with relay-lead. If it fails with "Workspace key not configured", call create_workspace, then register_agent again. Both create_workspace and set_workspace_key pin the workspace to this project, which is how workers pick it up.
- Give the user a link to follow along. Call
get_observer_url and print the URL it returns — or run agent-relay observer if your session does not expose that tool. Either way the link is backed by a read-only token that expires, so it is safe to share. Never build an observer URL from the workspace key, and never print the key.
- Confirm the work is genuinely parallelizable. Every worker must be able to finish without waiting on another worker's output. If that is not true, use the pipeline pattern instead.
- Pick the worker count from the task shape. Prefer 2–8, and stay low enough that you can still track every ACK and DONE.
- Partition the work into independent units — each with its own files, target, or scope boundary, and no shared intermediate state.
- Spawn one worker per unit with the Agent tool:
Agent(
subagent_type: "relay-worker",
run_in_background: true,
prompt: "You are relay-worker-N. Your lead is relay-lead.
CRITICAL: pass as: \"relay-worker-N\" on every relay tool call, or your messages
can be attributed to another agent.
Your unit: [specific target/scope].
Files: [list of files/directories].
Deliver: [concrete output].
When done, DM your lead a DONE message with the evidence for your scope.
Do not call remove_agent on yourself — the lead releases you once the work
is accepted, so it can send you review findings to fix."
)
- Wait for an ACK from every worker with
check_inbox(as: "relay-lead"). A missing ACK means that worker is not working — re-DM it.
- Keep a live worker table in your notes: name, unit, ACK, blocked, DONE.
- Let workers run independently. Only DM them for blockers, missing ACKs, or a global decision that changes every unit.
- Collect every DONE, verify the outputs yourself, and merge the summary. Call out units that finished partially or hit blockers.
Rules
- Do not use this pattern when stage N depends on stage N-1. That is a pipeline.
- Do not give two workers the same files unless duplicate review is the point.
- Keep the task wording uniform across units so the outputs are easy to compare and merge.
- Workers cannot spawn their own subagents — only the lead spawns.