web-search

Expose a web_search tool that accepts a query and max_results, then return titles, snippets, and URLs from search providers. The model calls it whenever users ask about real-time information, and optionally calls web_fetch to read full page text.

It queries DuckDuckGo HTML and Lite endpoints, Bing, 360 Search for Chinese queries, Exa when an API key is configured, and the GitHub repository search API. Backends are ranked by health and latency with automatic fallback.

web_fetch only allows http and https schemes, resolves every hostname and blocks private, loopback, link-local, and cloud-metadata IP ranges. Redirects are followed one hop at a time with re-validation, and response bodies are capped before parsing.

Yes. URLs are classified into channels: GitHub content uses the GitHub REST API, RSS and Atom feeds are parsed with the standard XML library, and plain-text or Markdown files are read directly. Disabled channels like YouTube or Twitter return an explicit error.

Providers may rate-limit or return anti-bot pages, in which case the circuit breaker opens that backend after repeated failures and routes to alternatives. When all backends fail, the tool reports the failure honestly instead of fabricating results.

Skip it for casual conversation and for queries handled by dedicated skills, such as dictionary lookups or internal documentation. Reserve it for explicit search requests, real-time information, external URLs, and questions the assistant cannot answer confidently.