| name | deep-research |
| description | Investigate a question across many files in fleet-generator without consuming the main context. Spawn a forked subagent that reads broadly and returns a structured synthesis. |
| when_to_use | Need to answer a question that requires reading many files. The question is exploratory (no edits planned). The main thread is mid-mission and reading 5+ files would burn context. Trigger phrases: dig into, investigate, research, look across files, find examples of, understand the pattern, deep-dive, explore, where is X used. |
| status | active |
| tags | ["skill"] |
| updated | 2026-05-07T00:00:00.000Z |
Deep Research
When invoked, you are the forked general-purpose subagent. The
parent provides the research question. Your job: answer it by
reading fleet-generator, then return a structured synthesis.
You have read tools (Read, Glob, Grep) — no Edit, no Write, no
state-changing Bash.
Procedure
- Identify relevant files via Glob and Grep. Cast wide; narrow
as the picture clears.
- Read the top 5–10 most relevant files. Skim larger files
with offset/limit when only a region matters.
- Synthesize in the output shape below.
- Return one final message to the parent. Do not transcribe
your reads — the parent only needs the synthesis.
Output shape
## Summary
<one paragraph — the headline answer>
## Findings
- <finding 1> — `path/to/file.ext:line-range`
- <finding 2> — `path/to/file.ext:line-range`
- <finding 3> — `path/to/file.ext:line-range`
## Open questions
- <questions the research surfaced but couldn't resolve from
reading alone, if any>
When NOT to invoke
- Edits planned. If the user wants changes made, this is the
wrong tool — use a normal mission with the worker.
- Quick lookup. If the answer is in 1–2 known files, the parent
should just Read them. Fork overhead isn't worth it.
- Already-known patterns. If the parent has the context
already, no fork needed.
Web research
If the question has external context (RFCs, library docs, vendor
specs), prefer primary sources: official docs, RFCs, peer-reviewed
papers. Cross-reference across at least two sources. Cite with
attribution. Distinguish confirmed facts from interpretations.
