Web access: web_fetch, and web_search when a search provider is configured.

Used by the Assistant harness ({Web, tools: [:fetch, :search]}) and the Librarian (bare).

Tools

• web_fetch reads one URL and returns it as markdown (default), plain text or raw text. Arguments: url, format, max_chars, offset (to page a long page) and artifact (save the full text under a name instead of returning it). There is no headers argument: an arbitrary header is a way to inject credentials or forge a request, and the tool refuses the argument if a model sends it.
• web_search runs a query on a provider (SuperPi.Web.Search). It is added to the tool list only when a provider resolves, so the model never sees a search tool that cannot work.

Both tools declare the effect net_read. SuperPi.Plugins.Permissions therefore allows them in plan mode (unless network_in_plan_mode is false), accept-edits and auto, and refuses them under network: "deny". Under network: "allowlist" the url argument decides, so web_search, which has none, is refused. SuperPi.Env.Workspace hides both from a workspace configured with "network" => false. The artifact save is internal to the tool and not a second effect.

The tools are not in the global tool catalog. The plugin adds the specs to agent_cfg["tools"] in before_agent_start and answers the calls in wrap_tool, so another harness on the same node never sees them.

Result channels

• output: title, final URL and the page text, with a note that names the offset to continue from when the text is cut.
• structured_output: final_url, status, content_type, bytes, chars, truncated, title, up to 30 absolute links (enough to crawl without a crawl tool), artifact, offset and next_offset.
• details: the redirect chain, the resolved address, the duration and cache_hit (no cache exists yet, so it is always false).

A 4xx or 5xx status, a refused content type, a policy refusal, an unreadable page or a timeout sets is_error, with the reason in output.

Size and artifacts

The page is cut at max_chars (default 20,000, at most 100,000). A text longer than auto_artifact_chars (200,000) is also saved in full as the artifact web/<host>-<hash>.<ext> and the note names it. An artifact holds the converted text in the requested format. A PDF or other binary file is refused unless artifact is given, and is then saved byte for byte. Fetches run inline in the tool task and are not durable: a half-read socket cannot resume and a GET is cheap to repeat. Crawling many pages goes through delegate_subagent.

Options

• tools: (default [:fetch, :search]). web_fetch is always offered. :search gates web_search.
• search: a provider, see SuperPi.Web.Search.
• fetch: options for SuperPi.Web.Fetch (max_bytes, timeout_ms, resolver, pin_ip, req_options). For operators and tests, never from the model.
• auto_artifact_chars: and parse_timeout_ms: (default 10 s).