Vicky Research

Vicky Research

Fetch new external data, hand it to learn. The user gets one verb, research, and the queue drains as a side effect.

Invoke: /vicky:research "<topic>"

Example: /vicky:research "GPU-driven indirect draw stream compaction"

Contract

1. Gap check — call research-gap "<topic>". If the KB already answers it, surface that and stop (unless the user asked for a refresh).
2. Enqueue pending — call enqueue question="<topic>" priority=med sources=[<any KB sources that mention it>]. Writes vicky/pending/<slug>.md with tags: [research, pending] and a sources: frontmatter block linking to upstream notes.
3. Web search — call web-search "<topic>". For each high-signal result, append a ## Sources entry to the pending note: title, URL, author, date, key passages quoted verbatim. Reject paywalls, off-topic, link-farm pages. Dedupe by canonical URL.
4. Extract follow-ups — for each unresolved sub-question, call enqueue question="<follow-up>" requested_by=research sources=[<topic-slug>]. Tag jargon → "what is X?" follow-ups, conflicting claims → "reconcile X vs Y".
5. Absorb — call the learn MCP tool (tier-progressive, sets up monitors). It drains the pending stub into vicky/sources/<slug>.md, relinks, and creates Claude Code monitors on dependent files (graph.json, pending/, sources/) for auto-reactions. No conclusion is created here — that step is for the caller, once a real synthesis exists. Mandatory — research without learn leaves orphan pending stubs.
6. Synthesise (optional, recommended) — read the new source, then call conclude title="<slug>" sources=[<slug>, ...] with the takeaway. Skip only if the source data is too thin to draw a conclusion yet; the dashboard's "Sources awaiting synthesis" section will surface it later.

When to call

• User: research <topic>
• A research-gap came back as a gap and the user wants fresh sources
• Topic in WORKFLOW.md → priority_tags is under-sourced

Do not call for questions the KB already answers (use query or research-gap). Do not call for bulk drain with no new fetch (use /vicky:learn directly).

Quality gates

• min_sources_per_conclusion (WORKFLOW.md, default 2) — if fewer high-signal sources survive scoring, mark the pending stub confidence: low in frontmatter and ask the user to retry / abort.
• Reject sources older than 3 years for fast-moving fields unless the topic is foundational.

State touched

• vicky/pending/*.md topic stub + follow-ups (created, drained by step 5)
• vicky/sources/*.md added by step 5
• vicky/conclusions/*.md added by step 6 if you call conclude
• vicky/graphs/*.json rebuilt by step 5

Failure modes

• No usable results — report the failed search, suggest broader / narrower terms, do not enqueue a stub, skip step 5.
• All sources behind paywalls — save citations + abstracts only, mark confidence: low, still run step 5.
• Topic too broad — split into sub-topics, enqueue each as a separate pending question, run step 5 once at the end.