web — fetch a web page as a Katari tool

An HTTP GET exposed as a tool the model can call. Pure Katari over the prelude's http.fetch: no FFI sidecar, and no capability to provide, since it reads public pages.

• web.fetch_page(url, max_characters ?= 16000) — fetches the page and returns its body text, cut to max_characters code points. The budget covers the whole answer, truncation marker included, and a cut page says how long the real page was.

That is the entire surface.

Failures

A non-2xx status throws http.api_failure — http.auth_error for a 401/403, http.api_error for everything else — carrying status, context ("web.fetch_page <url>") and the server's own body as message. A request that never completes throws http.fetch_error. A failure is a throw rather than a returned string, so an AI loop hands the model the error and records tool_failed.

Secrets / env

None.

Usage

import web

agent read(url: string) -> string with io | prelude.throw[http.api_failure | http.fetch_error] {
  web.fetch_page(url = url)
}