dispatch-subagent

dispatch-subagent — spawn an ICM-aware subagent

The subagent is a generic ICM agent, not a special tool. You tell it WHERE (which room) and WHAT (the job); it enters the workspace, reads AGENTS.md, routes to that room, reads the room CONTEXT.md, loads the Inputs it names, does the work in that room's output target, and drops markers — exactly as any ICM agent would. You do not pre-load its context; the frontdoor protocol does.

Two canonical uses

1. Worker — dispatch to a room to do real work (most common: a gated edit via the maintenance room, or running a productive room). The subagent follows that room's contract: writes to the room's output target, drops the room's markers, honors its Boundaries. It is a full ICM agent.
2. Reviewer — an auto-review pre-pass before you request human review. The subagent is transitory: it reads via absolute paths and writes only a provisional review file; it creates no markers and no state. You read its feedback, apply or discard, record the decision, and remove the provisional file.

When to invoke

• You need work completed in a room other than the one you're in (and that room is the right owner).
• expected value is high, self-assessment is low, or the output is novel → a reviewer pre-pass is worth it before the human gate.

Do NOT invoke if

• You are a subagent yourself (no recursive dispatch).
• For a reviewer: you are not at a clear post-generation, pre-human-gate point.

Args (infer from context)

• mode — worker | reviewer.
• room — the target room (worker) or the path to review (reviewer), as an absolute path or a routing target the subagent can resolve from AGENTS.md.
• model — haiku (default) | sonnet | opus, scaled to the work's value tier. For a maintenance room, the tier is canonical in maintenance/docs/model-tiers.md (one editable table) — read it and use that room's model. Otherwise scale by value: low/medium → haiku · high → sonnet · critical → opus. Model IDs: claude-haiku-4-5-20251001 · claude-sonnet-4-6 · claude-opus-4-8.

Process

1. Build the subagent prompt (fill the braces):

You are an ICM agent dispatched into this workspace, model {model}, mode {mode}.

ENTRY (do this first):
- Read {workspace_root}/AGENTS.md, then root CONTEXT.md for routing. Match your mission to a Routing row
  and navigate to {room}.
  Read that room's CONTEXT.md and load ONLY the Inputs it names. Find your own data and output target.

MISSION:
- {one-paragraph statement of WHAT to accomplish}

IF mode=worker:
- Do the job per the room's contract. Write ONLY to that room's output target. Drop the room's markers.
  Honor its Boundaries. You are a full ICM agent — follow the structure exactly.

IF mode=reviewer (transitory):
- Review the TARGET at {room}. Read everything via absolute paths.
- Write ONLY `{room}/review_provisional.md` (problems + suggested fixes, specific, citing the workspace's
  Layer-3 identity files where relevant).
- DO NOT create junctions, markers (`_handoff`/`_done`/`_rejected`/`_awaiting-human`), or touch state/memory.
- DO NOT dispatch a subagent. Terminate by returning "review complete".

2. Invoke the Agent tool — subagent_type: general-purpose, model: per the arg, prompt: the filled text, description: "{mode} in {room}".
3. Wait for it to return.
4. Post-return:
   • Worker: verify the markers/output landed where the room contract says; continue your loop.
   • Reviewer: read {room}/review_provisional.md, decide apply/discard, record the decision in the run's review sidecar (e.g. review.json auto_review[]), remove the provisional file, then proceed to the human gate.

Anti-patterns

• Dispatching when you ARE a subagent, or recursive dispatch.
• Pre-loading the subagent's room context yourself instead of letting it self-route (defeats the point).
• Reviewer leaving review_provisional.md behind, or applying feedback without recording it.
• Worker writing outside the target room's output target.

Related

• icm — the meta-skill (the frontdoor protocol the subagent follows).