deep-research

Investigate an open question to a well-supported, cited answer. Modeled on the orchestrator-worker research system (a lead plans → parallel subagents gather in their own contexts → lead synthesizes → a citation pass), which beats single-agent research because breadth is parallelized and each angle is blind to the others.

Procedure (AI-first, token-light)

1. Decompose. Break the question into independent sub-questions and list the source modalities to sweep — web, docs, codebase, data, primary sources. One angle alone always misses things.
2. Sweep (fan out). Spawn ONE subagent per sub-question/modality, in parallel. Each runs in its OWN context (the search burns the subagent's tokens, not yours) and returns only a small structured findings list: {claim, source, confidence}. They're blind to each other — that's the point.
3. Triangulate. A claim is "supported" only if ≥2 independent sources agree. Single-source = unverified. Prefer primary sources; down-rank low-credibility ones. Note disagreements explicitly.
4. Refute. Run /grill-ai + /doubt-driven-development on the surviving claims — try to break each. Confidence-filter: verify a weak claim up, or drop it. Better 3 solid findings than 30 maybes.
5. Synthesize. Integrate into a coherent answer with a citation per claim and a confidence level; honest UNKNOWN — needs X where evidence is thin. A report, not a link dump.
6. Completeness critic. Ask: what modality didn't I run? what claim is still single-source? what key source is unread? what contradiction is unresolved? Each gap becomes a targeted next sweep.
7. Loop until saturation. Repeat sweeps until K rounds surface nothing new (loop-until-dry). Record every settled finding in the decided ledger (key=finding) so later rounds never re-search the same ground.

Rules

• Triangulate (≥2 independent) before asserting; cite every claim; confidence-scored; honest UNKNOWN.
• Parallel + subagent-delegated (token-light) — synthesis reads only the findings, never the raw search context.
• Refute BEFORE synthesis (attack findings, don't just confirm). Loop-until-dry, not one-shot.
• Record findings in the ledger so rounds don't repeat. AI-first; only HUMAN-GATE actions crossing a real gate.