dsh-web-fetch-playwright

The built-in web_fetch tool returns raw HTML. When the target is a single-page application, most of the rendered content is simply missing, and the returned markup is still cluttered with navigation bars, ads, and forms that only burn tokens and confuse the model. dsh-web-fetch-playwright replaces plain HTTP fetching with a real browser backend: pages are rendered in Playwright or a remote CDP browser, then passed through Readability for article extraction, DOMPurify to strip layout noise, and Turndown with the GFM plugin to produce clean, structured Markdown. For Cloudflare challenge interstitials, the plugin waits in the same tab for the browser to clear the check on its own—no clicking, no captcha injection, no state spoofing—and returns a distinct error code when the budget expires rather than handing back the interstitial as content.

Two backends are available: a local Playwright browser or a remote browser already running with a DevTools Protocol endpoint. In CDP profile mode each fetch is a tab in the browser's real session, inheriting cookies and localStorage just like a tab you open by hand; switched to isolated mode, every fetch gets a fresh anonymous context. Concurrency limits, the denoise toggle, and challenge-wait settings are editable live in a settings card and apply to the next fetch without a restart.

This plugin is for teams that need reliable, browser-level web scraping in DeepSeek Harness: the target sites are React or Vue SPAs, you need logged-in access to protected content, or you want clean Markdown output ready for an LLM. If you already run DSH web, installing this package upgrades web_fetch from a half-baked HTML dumper to a full browser-grade fetcher.