菜单
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).
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
How the observer specialist works in Retainer — role, expected inputs, output shape, quality bar, common pitfalls. The observer reads this at the start of every dispatch.
How the Retainer scheduler specialist works — role, expected inputs, output shape, common pitfalls. The scheduler reads this at the start of every dispatch.
When the Retainer cog should delegate to the observer, how to brief it, what to expect back, and how to verify the result. Observer is the knowledge gateway — recent (hot index) and deep (full archive). Read before dispatching to agent_observer.
When the Retainer cog should delegate to the scheduler, how to brief it, what to expect back, and how to verify the result. Read before dispatching to agent_scheduler.
Decision framework for when and how to delegate work to Retainer's specialist agents. Covers the two specialists in this reference cut (observer, scheduler) plus the cog's direct tools.
How dispatching to specialist agents works in Retainer's current shape — the agent_<name> tool shape, what the reply envelope looks like, how errors propagate. Read with `delegation-strategy` (which covers when to delegate).
| 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 |
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.
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) |
brave_web_searchbrave_news_searchjina_readerAfter extraction, note in your final summary:
Prefer primary sources. When a snippet from search and a full extraction conflict, trust the full extraction.
When you complete your task, respond with a concise summary the orchestrator can act on:
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.
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.
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.