dsh-fetch-data

read_url hands you flattened prose. On a Bilibili ranking page the string "60125" might mean 60 likes plus 125 favorites, or something else entirely. Once HTML is squashed into text, the link between a number and its row is irrecoverably lost. dsh-fetch-data closes that gap by intercepting the page's real XHR and fetch JSON calls through Playwright and returning structured fields—titles, view counts, prices, comment counts—with exact row-level attribution.

The tool works in two phases. Structure mode returns a compact endpoint menu (path, payload size, array presence) so the model can identify the right call; field mode then extracts values along explicit paths like data.list[].title, truncating arrays at maxItems and capping each value at 200 characters. Auto-scroll triggers lazy-loaded feeds, JSONP is unwrapped automatically, and oversized config endpoints are filtered out, so no manual endpoint guessing is required.

Best suited for workflows where exact numbers matter more than approximate content: ranking-board parsing, comment-count harvesting, price-table extraction, article metadata collection, and any task where read_url's text output leaves number-to-row association ambiguous. Playwright is the required browser engine; login-walled APIs and pure server-side-rendered pages return a clear, actionable error message rather than a silent guess.