web.research

web.research is web.search + web.fetch + a synthesis pass, composed into one skill so an agent gets a real, cited answer instead of needing to orchestrate three tool calls itself. The Studio planner’s mandatory research task (see Briefs) is built on this skill.

Schema

Input = {
  question: string,
  depth?: 1 | 2 | 3,            // default 1 — "1 = quick, 3 = thorough"
  recency?: 'day' | 'week' | 'month' | 'year',
}

Output = {
  answer: string,
  sources: Array<{ title: string, url: string }>,
  notes: string,
  queries: string[],             // every search actually run, including follow-ups
  costUsd: number,
}

Depth controls everything

Depth	Pages read	Follow-up queries	Chars/page	Search results fetched	Synthesis tier
1 (quick)	3	0	5,000	6	cheap
2	6	2	6,000	8	cheap
3 (thorough)	10	3	7,000	10	standard

At depth 2 and 3, before reading anything, one cheap-tier call proposes up to followUps additional search queries based on the first page of results — filling gaps like specifics, opposing views, or primary sources — and those results are merged in (deduplicated by URL) before the top N pages are chosen and fetched in parallel.

Synthesis, and what happens when a page fails

Every chosen page is read via web.fetch; a page that fails to fetch still contributes its search snippet rather than being dropped silently, so it can still be cited. The corpus (title + URL + text per page) goes to one structured-free completion instructed to cite claims inline as [2]-style bracketed numbers, say plainly when sources disagree, and say what’s missing if the sources don’t answer the question — capped at roughly 350 words for depth 1, 700 for deeper research.

Without an OpenRouter key, or if synthesis fails, extractiveAnswer() returns a plain listing of the top 6 sources with their snippets instead of prose — the skill always resolves with something useful, never an error, as long as at least one page was found.

Example

curl -s -X POST http://localhost:4000/api/studio/skills/invoke \
  -H 'content-type: application/json' -b cookies.txt \
  -d '{"name":"web.research","input":{"question":"What changed in Node.js 22 for TypeScript projects?","depth":1}}'

{
  "ok": true,
  "output": {
    "answer": "Node.js 22 stabilises the built-in TypeScript stripping (`--experimental-strip-types`)... [1][2]",
    "sources": [
      { "title": "Node.js 22 release notes", "url": "https://nodejs.org/en/blog/release/v22.0.0" },
      { "title": "TypeScript support in Node.js", "url": "https://nodejs.org/en/learn/typescript/run-natively" }
    ],
    "notes": "search provider: duckduckgo",
    "queries": ["What changed in Node.js 22 for TypeScript projects?"],
    "costUsd": 0.0009
  }
}