deep-researchlisted

# Deep Research Verified: 2026-09-08 Primary sources over blog rewrites. Every claim traceable, every gap named. "Could not find X" is a valid result — never fabricate to close one. ## Workflow Scope (question, audience, depth, query budget) → outline (items, fields, success criteria) → pick retrieval tool → gather tiers 1→4, never start at 4 → verify per protocol → date-check staleness/version → apply stop criteria → compile output format. ## Retrieval tool | Need | Use | |---|---| | Library/API/framework docs | Docs-index tool (e.g. Context7-style MCP) — version-scoped, current, far fewer queries | | Fetch returns empty/JS shell | Headless-browser fetch — plain HTTP misses client-rendered content | | Many sources at once | Bulk-crawler — one dispatch vs. N fetches, spares budget | | 403/challenge on a known-live site | Retry browser-capable before writing it off; fingerprinting blocks curl, not the site | | Everything else | Web search + `search-techniques.md` | ## Query budget Search tools are capped per session and the cap is shared across parallel subagents, not per-agent. Divide it up front, give each agent a fixed quota in its prompt, prefer few targeted queries to many broad ones. Budget thin → finish sequentially rather than adding agents. ## Source Tiers | Tier | Sources | |---|---| | 1 Primary | Specs (RFC, W3C, WHATWG), source code, official docs, original papers, changelogs | | 2 Secondary | Peer-reviewed papers, vendor threat intel, audits, MITRE ATT&C