dsh-screen-eye

The screen is the last piece of context an agent needs to reason about a running environment, yet on macOS screen-capture permission fails in ways that look like bugs: the system dialog only adds the app to a list with the switch still off, and when the host process lacks an app identity macOS refuses silently without ever showing a dialog. Most screenshot tooling assumes the hard part is grabbing pixels; in practice the hard part is diagnosing the denial, naming the correct entry, and guiding the user to flip one switch. dsh-screen-eye is built for exactly this gap: it lets the agent capture the screen and see the resulting image within the same tool call, while treating permission diagnosis as a first-class concern.

The plugin exposes two model-callable tools. screenshot supports full-screen, display-by-index, rectangular region, multi-display inventory, and interactive window or rectangle selection, and can fire multi-frame bursts to observe animation or use wait_for_change to let the picture itself decide when to stop. screen_permission checks the current grant state on macOS and, when access is missing, opens the exact System Settings pane and names the entry to toggle, so the agent returns a fix rather than a stack trace. The plugin has zero dependencies and no build step; a tiny helper is compiled from bundled source on the first capture.

This plugin is aimed at developers running automation tasks inside DeepSeek Harness who need the agent to look at a running application, a modal dialog, or its own UI output without being asked for a screenshot. Whether the friction is a stubborn macOS permission or a DPI-unaware Windows process producing a downscaled frame, dsh-screen-eye turns seeing the screen into a routine tool call.