dsh-web-fetch

DeepSeek Harness raises WEB_PROVIDER_AMBIGUOUS when multiple providers are registered for the same capability via ctx.web.registerSearchProvider, leaving the model unable to pick between them. dsh-web-fetch takes a different approach: instead of cramming every source into a single provider interface, each data source is registered as a standalone DSH tool. web_fetch_cdp connects to a remote Chrome over CDP (via cloakbrowser) to render JS-heavy pages and SPAs, while web_fetch_tavily calls the Tavily Extract API for fast, browser-free extraction. The LLM decides which tool to invoke based on context. Each tool can be independently enabled or disabled; a disabled tool is entirely invisible to the model. The architecture centers on a FetchStrategy contract with two built-in strategies shipping out of the box. Adding a new data source requires writing one strategy file and one line of registration, with no changes to the router or core logic. The plugin attaches as a Cordis bundle with zero DSH core patches, depends only on the official @deepseek-ai settings, tools, and schemastery packages, and supports live configuration through a settings UI card or hot-reloaded ~/.dsh/settings.yaml without a restart. It is well suited for developers working in DSH workflows who need to fetch dynamically rendered pages, extract structured web content, or plug in a custom data source as a first-class LLM tool.