dsh-web-search-tavily

The built-in web search in dsh fires a full auxiliary model call for every query, burning model tokens and adding noticeable latency. dsh-web-search-tavily swaps that backend for a single Tavily REST call, returning results in roughly one second at basic depth with 1 credit and zero model-token cost. The plugin registers a WebSearchProvider through the official ctx.web capability seam, so the model-side web_search tool contract, citation formatting, result caps, and timeout budget all stay exactly as shipped. Only the transport underneath changes.

Key capabilities: four search depths (basic, advanced, fast, ultra-fast); topic filtering (general, news, finance); Tavily generated answers mapped directly into result content for downstream citation; and error handling that follows the ctx.web contract, distinguishing aborts, missing credentials, and provider failures for easy diagnosis. All configuration is managed through the patch layer, keeping API keys out of plugin code.

This plugin suits developers running dsh RC/stable who want to cut search latency and token overhead in agent workflows where web_search is called frequently. It requires only a Tavily API key (free tier available) and falls back cleanly to the shipped DeepSeek provider on uninstall, leaving no residual state behind.