mirror of
https://github.com/hanzoai/ai.git
synced 2026-11-20 14:13:23 +00:00
The builtin registry held five TIME tools. An agent on /v1/responses asking to search the web had nothing to call, while /v1/websearch/search, /v1/crawl and /v1/ask had been live the whole time. The capability existed and no agent could reach it. Three tools, registered unconditionally so an agent's tool list does not change shape with deployment: web_search ranked hits with titles, URLs, snippets fetch_url pages as clean Markdown (this module's own crawl) deep_research a cited report synthesised across sources EVERY BACKEND IS A SEAM, and that is forced rather than chosen. agent/builtin_tool/web must stay a LEAF: `object` imports `agent`, `agent` imports the registry, so a tool reaching into object for Crawl closes the loop — agent → builtin_tool → web → object → agent — and the module stops compiling. Search and research are the HOST's anyway (hanzoai/cloud serves them and imports this module, so it cannot be imported back). Crawl is installed by this module's bootstrap; the host installs the other two at mount time, beside the balance and tier readers it already installs. That shape is also the lesson this repo already paid for: the ai module once reached its own platform over api.hanzo.ai, which carries a CUSTOMER credential and answers 401 to a service, and every completion fail-closed at 503 on a healthy pod. The edge is for customers; in-process is for us. AN UNAVAILABLE CAPABILITY IS AN ERROR, NEVER AN EMPTY RESULT — the property worth the most here. A model reads ordinary tool text as a FINDING: hand it "no results" when the backend was never installed and it concludes the web holds nothing on the subject, then answers from memory in a confident voice. That is a fabricated answer caused by a deployment gap, and nothing downstream can tell it from a real one. A genuinely empty search still returns a normal result, because that IS a finding. Bounded on purpose: limit clamped to 10 (a confused call must not exhaust the agent's context window), 5 URLs per fetch, 20k markdown per page, and a failed page is NAMED rather than omitted — an agent cannot tell "unreachable" from "said nothing" if it silently disappears. fetch_url accepts only absolute http(s). The URLs it is handed come from PAGES the model read, so following file:// or a custom scheme would read the host's own disk on behalf of whoever wrote that page. It also accepts a bare string, which is the shape a model sends most often and is unambiguous. 13 tests, including the refuse-vs-empty property for all three tools, the scheme filter, the clamps, and that the tools are both REGISTERED and exposed in GetToolsAsProtocolTools — a tool absent from that list is invisible to agents however well it is registered. Co-authored-by: Hanzo Dev <[email protected]>