uss-parks/lamprey-harness/fan-out

Use this skill when the task decomposes into independent reasoning passes that can be checked separately.

Prefer multi_agent_run when the branches do not need tools: for example planner plus reader, reviewer plus verifier, or multiple architecture comparisons over supplied context. Keep each sub-agent prompt narrow. Provide bounded context, explicit output format, and only the facts needed for that role.

Use the active chat model for every sub-agent. Do not use fan-out to create a visible multi-model conversation or to bypass tool approval. Sub-agents must not call tools; tool use stays in the main run after their outputs return.

Good role sets: planner for decomposition, reader for fact extraction, verifier for checking claims, reviewer for code risk, and coworker for a second implementation angle.

After fan-out, synthesize the results yourself. Name disagreements and choose one path.

Stop when the independent angles converge or when the disagreement identifies the next concrete read or verification step.