Deep Research

Overview

deep-research runs a recursive, multi-round web investigation. It fans out parallel subagents breadth-first then depth-first, runs an adversarial verification valve once per round, and synthesizes by re-deriving contested claims from retained raw sources rather than from a summary chain. The loop is a double-diamond: each round alternates diverge (explore/expand) with converge (synthesize/review), repeating until findings saturate or the user's budget runs out.

Four invariants make this safe and cheap — they are load-bearing, not optional:

1. Raw-to-disk. Every fetch is dumped verbatim to a tmpfile. Subagents return only {claim, source-url, raw-pointer, self-tag} — never raw content upward.
2. Review re-derives from raw. The verification valve checks contested/load-bearing claims against the raw tmpfiles, never against the distilled findings ledger.
3. Review fans out by claim. "Once per round" is a cadence, not a single monolithic pass: extract atomic claims, dispatch one skeptic per claim (or small batch).
4. Synthesis reads artifacts; the user reads progress. Progress prose is for the human and must never become a synthesis input — the ledger + topic-map + raw are the lossless memory.