dsh-preview

The most common failure mode after an agent ships frontend code is not that it cannot write the code, but that it never looks at the result. It re-reads its own source, declares the work done, and leaves you to open a browser, take a screenshot, and describe what broke to it. dsh-preview targets exactly that blind spot: it installs a deterministic verification loop on the agent that opens the page, inspects the console, asserts layout facts from computed styles, exercises the UI, saves screenshots as evidence, fixes what it finds, and re-verifies, all before it is allowed to claim completion.

The heart of the design is browser_read, which returns rendered text, outer HTML, and the bounding box plus key computed styles for a given selector in plain text. This lets a text-only model verify box sizes, color values, display properties, and rendered copy deterministically instead of hallucinating over pixels. browser_interact handles clicks, typing, and key presses and reports any console errors the interaction caused. browser_screenshot saves viewport, full-page, or single-element PNGs into the workspace for the human in the loop. All six tools serve one discipline: open, console, read, interact, screenshot, fix, re-verify. The bundled frontend-verify skill enforces that sequence and requires the agent to name what it could not verify rather than implying full coverage.

On the security side, localhost and 127.0.0.1 are always allowed because that is what frontend verification needs, while every other host is refused by default unless explicitly granted via allowedHosts. Local file paths are served read-only on an ephemeral loopback port with path containment. The skill forbids screenshotting pages that contain secrets and never types credentials. Known limitations include headless rendering differences from a desktop browser for pointer lock and certain GPU code paths, which the skill instructs the agent to flag explicitly. This plugin fits developers using dsh on frontend work who want their agent to actually open, check, and evidence its own output before reporting that the work is finished.