dsh-web-search-local

The built-in web_search tool in dsh is model-agnostic at the interface level, but its default provider ships every query to DeepSeek's server-side search API, which requires a DEEPSEEK_API_KEY. The moment you switch to a local model such as Ollama, search simply stops working. dsh-web-search-local fills that gap: it registers both a search provider and a page-fetch provider on the ctx.web seam, performing all HTTP with node's native modules and zero third-party runtime dependencies.

Search runs through a three-layer engine chain: a private SearXNG instance, then Google / DuckDuckGo / Mojeek, then Bing / Baidu / Sogou / 360. Engines within the same layer are queried in parallel and merged round-robin; a layer that returns no results automatically degrades to the next. The provider auto-detects local proxy ports (Clash, v2rayN, etc.) but routes only the global engines through the tunnel, while CN engines and a private SearXNG always connect directly, avoiding Baidu's verification wall, Sogou's anti-spider captcha, and 360's redirect loops that a foreign exit IP would trigger. Per-engine circuit breakers cool down walls and rate-limits, and DuckDuckGo includes a lite-endpoint fallback, so a single failing engine never takes down the whole search.

This plugin is ideal for dsh users running local models (Ollama, llama.cpp, and the like) who want keyless web search without signing up for any search API. It also serves network-restricted environments that need reliable coverage of both global and Chinese search sources. Installation auto-activates the web profile with no manual config edits required, and all settings are hot-reloadable through dsh's settings panel, taking effect on the very next query.