Research

Runtime skill — a playbook the deployed @context agent runs for its owner, invoked in natural language. Not a coding-agent workflow; those live in .agents/skills/.

Turn a topic into a durable, sourced brief filed in the knowledge base. Like process-today, this one writes — it ends in an update_knowledge call, not just a summary. It still never acts on the outside world: it reads and files, and nothing leaves.

Procedure

1. Pin the topic and the question. Settle what the owner actually wants to know ("research how teams handle multi-tenant RBAC" → the question is how do teams do this, and where does it leave us). If the topic is genuinely ambiguous and a wrong guess wastes the work, ask one clarifying question.
2. Sweep what we already know first — internal before web. Always do this, even for a topic that looks purely external: it's how you avoid duplicating a page that already exists and how you learn where the brief should land. A quick sweep that comes back empty is fine; skipping it is not.
   • query_knowledge — route through the index to any spec or page already on the topic. If one exists, the brief extends it; it doesn't shadow it.
   • query_crm — contacts, notes, projects, meetings tagged to the topic (our relationship and history with it).
   • query_slack (when connected) — recent threads where the team discussed it; often the freshest internal context.
3. Go to the web for current and external background. query_web for what we can't know from our own notes: public facts, recent developments, how others approach it. Keep queries to the public subject, not the owner's private notes. Skip the web only when the topic is purely internal and the internal sweep already answers it.
4. Read the codebase when the topic is framework-related. If the question touches how @context is built or how Agno does something (a toolkit, a provider, the scheduler, AgentOS), read the repo's files (list_files, read_file) to see how it uses it, and query_web for anything the code doesn't answer.
5. Synthesize a clear, sourced brief. Lead with the answer. Attribute every claim to where it came from, and keep web background visibly separate from our own record: tag lines (from web), (on file), or (auto-summarized, unverified) for anything you distilled but haven't confirmed. Never blend an unverified web claim into our known facts.
6. File it with update_knowledge. Filing is frictionless (no approval gate), so land it directly:
   • Topic already covered → file into that spec's right sub-file (a decision → decisions.md, design detail → design.md, status → the README.md) and keep the status table current.
   • New topic that's a real spec → a new folder-per-spec following _template/, plus a row in the root README.md index.
   • A standalone write-up that isn't a spec → a loose prose page ("what we know about X"). The write sub-agent carries the provenance labels through, so keep them in the prose you hand it.
7. Make the brief findable — don't orphan your own research. Filing isn't done until the new page or spec has a row in the root README.md index. A new spec earns its row through the write tool's own rules; for a new loose page, make a second update_knowledge call that adds a one-line index row pointing at the path you just wrote. Skipping this lands the brief as exactly the kind of orphan the knowledge-review sweep exists to flag.
8. Report what you filed and where. One line per file touched (what + path), then one line on what you leaned on — "web: 4 sources; on file: the pgvector-standard spec" — so the owner can see the brief's footing at a glance. Mirror process-today's digest.

Format

The filed brief, only the sections with content:

• Answer — the tight take on the question, up top.
• What we already knew — from knowledge / crm / slack, tagged (on file).
• What the web adds — public background, tagged (from web) and dated where it matters; marked external / unverified.
• Where this leaves us — the so-what for the owner: implications, open questions, a next step if there's an obvious one.
• Sources — the web results cited and the internal pages drawn on.

Edge cases

• Read and file only. No mail, no calendar, no Slack send. This skill never reaches outward, so a scheduled run completes with no approval gate.
• Nothing on the web (or nothing useful): say so plainly and file what the internal sweep gave you, rather than padding the brief with generic results.
• Thin topic, nothing anywhere: don't fabricate a brief. File a short stub page that frames the question and says what's missing, and tell the owner.
• Already well covered: extend the existing spec or page in place and say what you added; don't create a parallel duplicate.
• Label every fact. A reader six months out has to know whether a line is our own record, an unverified summary, or a web claim. Keep the labels light (tag the claim or the section, not every sentence), but never drop them.