| name | research-partner |
| description | Use when the user opts into a local research-interest model, wants personalized paper/preprint/trial updates, saves or rejects research suggestions, asks for a proactive briefing, or wants workflow prediction across sessions. |
Privacy-first research partner
This skill turns repeated research interactions into a local, inspectable profile. It is not permission to retain conversation text.
Consent and data boundary
- Before the first
research_interest_observe call, obtain explicit opt-in for local learning. Pass consent=true only within that consent scope. A user asking a biomedical question is not consent by itself.
- Send only structured public concepts such as HGNC symbols, disease terms, pathway names, PMID/DOI/NCT identifiers, and short workflow slugs. Never send raw prompts, abstracts, notes, filenames, PHI, or arbitrary metadata.
- Default storage is HMAC-pseudonymized aggregates. No event history is retained; time is reduced to coarse weekday/weekend buckets.
- Honor inspection and deletion immediately with
research_interest_inspect and research_interest_delete.
Learning events
- Saving a paper:
paper_saved with its public ID and canonical topics.
- Repeated gene/target lookup:
entity_queried.
- Accepted or rejected suggestion:
suggestion_accepted / suggestion_rejected; rejection is weak negative feedback, not proof that a topic is irrelevant.
- Completed workflow:
workflow_observed with a short task slug.
- Shown recommendation:
recommendation_shown so cooldown deduplication can work.
Proactive loop
- At the beginning of a research session, call
research_session_brief only if a local profile exists. Build topic_catalog from concepts already present in the local project, saved-paper index, or local KG.
- The returned refresh actions are a plan, not executed requests. Keep
allow_remote_queries=false until the user or an existing policy explicitly authorizes outbound public-database queries.
- Execute authorized actions through
bio-lit / bio-trials, normalize results into the candidate schema, and call research_updates_rank locally.
- Explain why each suggestion matched, distinguish relevance from evidence quality, and include source IDs. Do not present a high relevance score as scientific validity.
- Record
recommendation_shown, then record later acceptance or rejection if the user provides that feedback.
Active behavior boundary
This version supports session-start proactive briefings and scheduler-ready refresh actions. It does not silently run a background daemon, contact external services, or issue OS notifications. Any host scheduler must preserve the same consent, sensitive-mode, endpoint, cooldown, and deletion guarantees.
