What problem does it solve? Answering questions that require evidence from outside the codebase usually means ad-hoc web searches, uncited summaries, and no record of what was found. This Skill runs a structured deep research pipeline that decomposes a query, fetches and stores sources, drafts a cited report, adversarially critiques it, patches findings, and ships only after a deterministic gate passes. ## Core Features & Use Cases - Six-step gated pipeline: Decompose the query into atomic items, sweep the web with parallel fetcher subagents, draft one report, critique it, patch findings, and ship only when the ship gate passes. - Citation contract and report format: Enforces [n] citations, a Sources table with search lenses (breadth, canonical, adversarial), typed Key findings, an answer-first Summary, and a paragraph cap, all checked by ship_gate.py. - Knowledge graph ingestion: Converts typed Key findings into graph memories with URL evidence tagged by fetch date and content hash, so conclusions persist across sessions. - Use Case: Ask "deep research on the current state of free-threaded CPython" and receive a cited report.md with an answer-first summary, per-question sections, open questions, and conclusions stored in the project knowledge graph. ## Quick Start Enable deep research in the Navigator config, then ask the assistant to do deep research on your question and wait for the cited report to be written under .agent/research.
