| name | web-research |
| description | Search and extraction strategy for the researcher specialist. Adapted to Retainer's current tools (Brave web search + news search via BRAVE_SEARCH_API_KEY; Jina Reader via JINA_API_KEY). |
| agents | researcher, cognitive |
Web research
The researcher specialist handles web work. From the cog's perspective you delegate via agent_researcher; from inside the researcher, here's the tool set.
Tools available
The researcher's registry is gated on environment keys:
| Tool | Best for | Requires |
|---|
brave_web_search | Broad discovery — titles, URLs, snippets | BRAVE_SEARCH_API_KEY |
brave_news_search | Time-sensitive queries, current events, recent developments | BRAVE_SEARCH_API_KEY |
jina_reader | Extract clean markdown from a known URL | JINA_API_KEY (or anonymous tier when RETAINER_ENABLE_JINA_ANON set) |
Decision tree
- General web search? →
brave_web_search
- Time-sensitive / news? →
brave_news_search
- Have a URL, need full content? →
jina_reader
- No keys configured? → tell the cog (the researcher won't have been started; the delegate tool wouldn't exist)
Quality signals
After extraction, note in your final summary:
- Publication date — when the source was last updated
- Primary vs secondary — did the source author the information, or aggregate it
- Contradictions with earlier results — flag these explicitly; don't paper over them
Prefer primary sources. When a snippet from search and a full extraction conflict, trust the full extraction.
Output format
When you complete your task, respond with a concise summary the orchestrator can act on:
- Sources with URLs
- Key facts (with confidence levels when relevant)
- Anything that didn't turn up (dead ends matter)
Omit raw page contents. Omit your own intermediate reasoning — the orchestrator wants the answer, not your scratchpad. The cog will pass your reply back to the user; make it readable.
What's deferred
These tools exist in Springdrift but are not yet ported to Retainer — do NOT assume you have them:
brave_answer, brave_llm_context, brave_summarizer (extra Brave endpoints; only the search clients are wired today)web_search (DuckDuckGo fallback) — no implementationfetch_url — no raw HTTP toolkagi_search, kagi_summarize — Kagi backend not configuredstore_result / retrieve_result — artifacts subsystem deferred (so big extractions go directly into your reply, capped by max_tokens)
If a task genuinely needs one of those, say so in your final reply rather than approximating.
Honesty under low signal
If a search returns no useful results, say so. Don't synthesise a plausible-sounding answer from nothing — the cog (and the user) need the truth, including "I couldn't find this." Same for partial coverage: report what you found AND what you didn't.
