Skip to content

Web research

Milimarty edited this page Oct 19, 2026 · 1 revision

Web research

web-researcher (0.18.20) is a read-only agent for outside docs: APIs, libraries, webhooks, standards. It searches first, then reads the best sources (developer portal, API reference, webhook pages, changelog, OpenAPI, SDKs on GitHub, Postman, Stack Overflow). It tries at least five sources before it calls a fact unconfirmed. Its answer has the confirmed facts (URL + quote), every URL it tried, what a person must supply, and CONFIDENCE.

 spike:    frame ─ web_docs ─ researchers per angle (they cite its URLs) ─ findings (confirmed / not confirmed, sources)
 feature:  web_check ─ WEB: yes ─ web_docs ─ spec      (WEB: no ─ spec)
  • spike: the web step always runs. The findings list their sources and say what is confirmed and what is not.
  • feature (0.18.21): web_check decides first. WEB is yes when the request or the code it touches depends on a third-party API, webhook, library or standard whose details the repository does not hold. It is no for work inside the code (a refactor, a page) and for a request that already carries web research (a spike's findings). On yes, web_docs runs, and the spec cites what is confirmed. An unconfirmed detail becomes an Open question with what a person must supply (a sample payload, a sandbox account, a docs page), never an invented field name.
  • change: a change whose criteria rest on outside docs is triaged as a feature, so the feature flow does the research.
  • Web tools: an agent whose tools: line names WebSearch or WebFetch gets them (Claude Code, opencode). On Codex, Copilot and API keys it reads the web with curl -sL. Other agents get no web tools.
  • A read-only run can read the web (curl -sL, wget -qO-) but not download files: curl -o, curl -O and wget <url> are refused.

From docs/GUIDE.md · All pages

Clone this wiki locally