| name | dispatch |
| description | Lightweight sub-agent dispatcher — fast fan-out that skips formal restatement and review-by-default |
Phase 0 — EVOLVE
Read evolution.md in this skill's folder. Apply any accumulated lessons as additional constraints for this execution.
Phase 1 — ASSESS
- Parse the request.
- If genuinely ambiguous (unclear target files, conflicting instructions), use
AskUserQuestion to clarify before proceeding.
- Identify the units of work to dispatch, noting any dependencies between them.
Phase 2 — DISPATCH
Every sub-agent is a senior peer, not a subordinate — it reads files, explores, and decides the HOW within its unit. Your job is to transfer the mental model (WHY and WHAT), then step back on the HOW. Peer-at-the-unit-level does not mean hands-off at the pipeline level: the dispatcher still coordinates across units — selecting subagent_type, relaying hand-offs between serialized agents, and handling failures.
Self-check before sending any dispatch prompt: "Am I telling the agent what to think, or giving it what it needs to think for itself?" If the former, cut.
For each unit from Phase 1, spawn a sub-agent via Agent:
- Point, don't summarize: give file paths, the user's motivation, and hard constraints the agent couldn't infer — omit everything else.
- Parallelize where possible: launch independent units in a single message; serialize dependent ones. Units that edit overlapping files count as dependent — dispatch them sequentially. For serialized dependencies, relay a concise hand-off pointer (paths + key findings) from agent N into agent N+1's prompt — never the full output. Keeps the dispatcher's context lean.
- Prefer dispatching over reading: keep the dispatcher's context lean. A single small read is fine; dispatch a sub-agent for anything larger.
- Pick
subagent_type: default to general-purpose; use Explore/Plan only when the task genuinely needs their restricted toolset (user's standing preference — see CLAUDE.md). The type is a dispatcher-level routing choice — never prescribe the HOW inside the prompt itself.
- Handle failures actively: retry a transient failure once; surface persistent failures in REPORT without auto-fixing.
Permitted tools: Agent, AskUserQuestion, Glob, Grep, ToolSearch, Skill, Read (prefer dispatch; single small reads OK).
Phase 3 — REPORT
After all sub-agents complete, deliver a consolidated summary:
- What was done, per agent (concise — enough for the reader to pick up cold)
- Any issues or items that need attention
Optional: REVIEW
Triggered when the user signals they want review — "with review" is the canonical phrase, close variants qualify. After REPORT, dispatch reviewer agent(s) to audit all changes. Report the reviewer's findings. No auto-fix loop — surface issues for the user to decide. If the user then requests fixes, continue the original author agent via SendMessage (its context is intact) rather than briefing a cold agent.
