| name | subagents |
| description | Use when a plan has parallel tasks and the harness supports dispatching subagents. |
Subagents
Fresh context per task. No session history.
Dispatch: reference, don't copy
Subagents have file-read tools. Don't inline what they can read.
TASK: <one sentence>
FILES: <paths>
SPEC: <docs/staging/specs/...#anchor>
ACCEPTANCE: <test or cmd>
EXTRA: <only what's NOT in referenced files>
Pasting >10 lines from the spec? Stop - let the subagent read it.
Expand milestone -> atomic steps at dispatch time, not in the plan.
Loop per task
implementer -> on DONE: spec-reviewer (matches spec?) -> quality-reviewer (review) -> mark complete, continue.
The coordinator marks - [x], never the subagent.
Status
DONE_WITH_CONCERNS - address if correctness; note if observation.
NEEDS_CONTEXT - supply the fact, re-dispatch.
BLOCKED - diagnose (missing context / stronger model / too big / plan wrong). Never silently retry.
Model
1-2 files mechanical -> cheap. Multi-file integration -> standard. Design/review -> strongest.
All done -> review whole diff -> ship.