dispatch

Dispatch — Orchestrator as Manager

The orchestrator's job is experience sharing and worker health, not prompt engineering. Workers are specialists. Trust them. Relay what they learn. Unblock them when stuck. Synthesize what they produce.

For the delegation prompt template and pre-send verification, activate the /agent-orchestration:agent-orchestration skill.

Two Dispatch Modes

flowchart TD
    Start(["Work to dispatch"]) --> Q{"Is there a SAM task<br>for this work?"}
    Q -->|"Yes — SAM plan exists"| SAM["SAM Dispatch<br>Minimal prompt — task has everything"]
    Q -->|"No — ad-hoc work"| AdHoc["Ad-Hoc Dispatch<br>Use delegation template from<br>/agent-orchestration:agent-orchestration"]
    SAM --> SAMPrompt["Agent prompt:<br>'You are working on P{N}/T{M}'<br>Agent loads task via sam_task —<br>acceptance criteria, context, verification steps all there"]
    AdHoc --> AdHocPrompt["Write OBSERVATIONS + DEFINITION OF SUCCESS +<br>CONTEXT per agent-orchestration template"]
    SAMPrompt --> Spawn["Spawn workers directly"]
    AdHocPrompt --> Spawn

Manager Responsibilities

1 — Spawn Workers

Fetch-once rule: Before spawning any workers, call backlog_view once per issue that will be worked in this session. Store each result in context keyed by issue number. Do NOT call backlog_view again for any issue already fetched — use the stored data for all wave iterations, prompt construction, and relay building. If a backlog_update changes an item's state mid-session, replace the cached value with a single new backlog_view call for that issue only.

Each worker gets exactly the context needed — no more.

SAM dispatch (the task is the delegation):

Agent(
  name="T42-worker",
  prompt="Your ROLE_TYPE is sub-agent. You are working on Pf1a2b3c4/T42."
)

The agent calls sam_task to load the task. All acceptance criteria, verification steps, and context live in the task.

Ad-hoc dispatch: follow the delegation template from /agent-orchestration:agent-orchestration — OBSERVATIONS, DEFINITION OF SUCCESS, CONTEXT.