---
name: deep-research
description: Use when a question needs external evidence rather than local code - "research this", "find everything about", "deep dive", "compare options/approaches", "state of the art", "is X better than Y", competitive analysis, due diligence, literature review, or any decision that depends on facts you must verify online. Not for quick factual lookups or questions answerable from the repo (use exploring-codebase)
---

# Deep Research

Evidence over vibes. Plan → parallel search → assess gaps → cite → synthesize.
Adapted from dzhng/deep-research, open_deep_research, and gpt-researcher.

```mermaid
flowchart LR
    S[Scope: 1 question + 3-5 sub-questions] --> P[Plan depth mode]
    P --> B[Breadth-first search, parallel subagents]
    B --> A{think: what is missing or contested?}
    A -- gaps and budget left --> B
    A -- done --> C[Compress: dedupe, keep cited claims]
    C --> R[Synthesize report]
    R --> V[Entailment: each claim checked on its fetched page]
```

## Depth modes (pick one, announce it, never ask the user to choose)

| Mode | Rounds | Sub-questions | Sources target | Use when |
|---|---|---|---|---|
| quick | 1 pass | 3–5 | ~5–10 | one focused factual question |
| standard (default) | 1 wide + 1 gap round | 3–5 | ~10–20 | comparisons, "what is the state of X" |
| deep | 2–3 rounds | 4–6 | 20–40+ | high-stakes decisions, due diligence |

## 1. Scope

Turn the request into **one explicit research question**; list **3–5
non-overlapping sub-questions**; state in-scope and out-of-scope in one line.
Ask at most **one** clarifying question, and only if the goal is genuinely
ambiguous — otherwise state your interpretation and proceed.

## 2. Search (start wide, then narrow)

- 2–3 keyword variants per sub-question; broaden first, then drill in.
- **Prefer primary/authoritative sources** over aggregators and SEO content:
  official docs, papers, filings, first-party sites, maintainers' own pages.
  See `references/source-quality.md` (evidence hierarchy + SIFT + red flags).
- For independent sub-questions, dispatch the `researcher` subagent — **all in
  one message** so they run in parallel (fan-out; contract in
  `dispatching-parallel-agents`). Give each: objective, output schema, source
  boundary, and search budget. Subagents return distilled findings, **never
  raw pages** (keeps context small). Extra personas are optional — never
  required; do not multiply perspectives.
- Reuse `acquiring-capabilities` trust rules: no random blogs, no unsourced
  forums, no AI-generated listicles.

## 3. Assess (`think_tool`)

After each round, write one short gap analysis: what is now supported, what is
missing, what conflicts. **Close the loop**: each round's queries derive from
what was just read; a round that does not change the next query ends the loop.
Use **angles, not synonyms**; shrink breadth as depth increases.

## 4. Compress (fan-in)

Dedupe by URL; merge duplicate facts; keep only claims that carry a citation;
preserve citation numbers; never drop a source that supports a kept claim. At
fan-in **one writer** merges: parallel section writers produce disjoint
reports, so style and citations stay consistent. Resolve conflicts per
`dispatching-parallel-agents`; list what remains unanswered.

## 5. Synthesize + verify

- Output per `references/report-format.md` and `communicating-concisely`:
  executive summary first, then themed sections (mermaid/table where it
  clarifies). Long reports go to a file; return summary + path +
  `sources_reviewed: N`.
- **Per-claim entailment**: every citation `[n]` points to a page you actually
  fetched and read — never the search engine, never a snippet-only hit, never
  an invented URL. Compare each claim against that page
  (`references/citation-check.md`): supported → keep, partial → qualify,
  unsupported → cut or mark "unverified". Count citation accuracy
  (supported ÷ checked) in the report.
- End with confidence labels, unresolved conflicts, and open questions.
- When the report compares options, end with a decision: hand off to
  `option-selection` (matrix → pick → record). A comparison with no choice is
  an unfinished report. BrowseComp is a single-short-answer benchmark; do not
  optimize for it.

## Citation discipline

- Every substantive claim carries an inline `[n]`; `## Sources` lists them
  sequentially with no gaps.
- Label confidence: **high** (3+ independent sources agree), **medium** (2),
  **low** (1 source, or sources conflict).
- Single-source claims are marked "unverified". Separate fact / estimate /
  opinion. If evidence is missing, write "insufficient data found" — never pad.

## Hard budgets (stop even if not done)

Per subagent: 3–5 searches. Global: the mode's source target. Stop when 3+
independent sources agree, a round adds no new material, or the last two
searches return the same information. Report what remains unanswered.

## Anti-patterns

- No subagents for a quick lookup; no synonym queries as separate agents.
- Never treat a search hit as validated; never cite the engine instead of the page.
- No blog-only sourcing for factual claims; no repeated sections; budget out, then report gaps honestly.