Deep research

Accept /deep-research me or /deep-research <name>. If the target is missing, ask who to research. This invocation requests discovery and saving cited findings; use normal tool permissions, without an extra blanket saving question.

Resolve the target

• For me, read my_profile and the connected-account context. Use research_person with person_id: "me". Never substitute another person. Self research works before GitHub or LinkedIn is connected; ask only for a missing identity anchor that prevents reliable research.
• For a name, use search_people in the selected Team and read get_person. Only the explicit argument me selects self research. A named argument must resolve through the Team even when it matches the current user’s display name. Never combine person_id: "me" with a Team ID. Resolve ambiguous matches with one short question before dispatch. Other-person research needs a Team: use the selected Team or ask which Team when unclear. Never silently switch Teams or start a self run for a named person. If no person record exists, resolve their identity and ask to add them first.

Run once and follow through

Call research_person once for the resolved target and Team when applicable. Let the native engine use its production model, default budgets and saved account anchors. Ask it to discover deep, non-obvious public professional information, including earlier history, overlooked creations and original sources in the person's own language. Do not seed expected personal facts or fixed results. Do not replace the engine with a couple of web searches or a meeting-prep brief.

Retain the returned run ID. Check research_status for that same run; never start another run to check progress. Explain that discovery can continue in the background. If still running when this turn ends, give its status and run ID so the next turn can resume it. On completion, read committed findings through my_profile or the selected Team's get_person, following section cursors. Use memory_get for individual claim details. Report partial coverage honestly, including publication versus indexing status. Never say research is complete just because dispatch succeeded or a source request failed.

Findings and review

Show the most useful or surprising findings first, with the actual source links and dates. Separate new discoveries from known information and repeated facts. Keep historical roles historical. Account links already rejected by the user remain rejected; shared names or institutions alone do not establish identity.

For self research, ask_person_fact presents a saved public-research claim in a yes/no/text card. Use it only when user review would resolve real uncertainty; do not re-confirm facts already supported by their connected profile or LinkedIn. Ask one at a time. The card persists the answer, so do not duplicate its write. For a plain-text answer, reread the current claim and use correct_person_claim only for an explicit correction/rejection/confirmation with its current revision. Never infer consent or truth from silence. Preserve useful partial results and all prior corrections. No outreach, identity merging or recursive network crawl.