Web research

web_search uses Serper when SERPER_API_KEY is configured, then Brave when only BRAVE_SEARCH_API_KEY is configured, otherwise keyless DuckDuckGo HTML search. Serper returns structured Google search results through https://google.serper.dev/search; the key is sent in the X-API-KEY header. Results retain their titles, URLs and associated snippets, with duplicate and unsupported destinations removed. Configured provider failures are reported directly without silently switching providers. An unavailable or invalid response is an error, not a claim that no sources exist.

fetch_url opens a public page and preserves its extracted text in the owning task. Results include the requested URL, final URL, title, capture timestamp, source ID, total length and exact next offset. source_read reads more of that captured source without another network request. This is task evidence; other actors, tasks and locations cannot reuse the handle.

SafeWebClient accepts public HTTP(S) destinations on standard ports, rejects URL credentials, validates all DNS answers and pins each connection to a validated address. Redirects receive the same checks. Provider credentials are not forwarded across redirects. Requests have a 20-second deadline and a 2 MiB transport limit. Private, loopback, link-local, reserved and mapped addresses are excluded. The client does not send cookies, execute scripts or load page subresources.

ReadablePage uses LinkeDOM to parse HTML and preserve readable structure and links. HTTP connections use the Node request API. Unsupported binary content belongs in the isolated file workflow. Pages that require JavaScript or return unsupported compression report that limitation.

The deterministic suite tests destination checks, redirects, credential stripping, snippet attribution and owned pagination. A separate opt-in test verifies a real HTTPS page through the pinned transport. Search-provider availability can still vary and must not be confused with a completed investigation.