dsh-screen-reader

A large model, no matter how capable, has no eyes that can land on your screen. You ask it to verify a setting change, and it cannot see. You tell it to check a script's output, and it cannot read the window. Worse, when you hand two screenshots to a vision model and ask what changed, it will fabricate a plausible-sounding diff, because probabilistic inference is precisely the worst tool for pixel-level comparison. dsh-screen-reader takes both jobs out of the model's hands and hands them to deterministic code.

The plugin ships two tools. see_screen captures the full screen, a specific window, or a region you define, then injects the image directly into the current conversation as a content block. No second model transcribes it, no extra API call, no layer of interpretive loss. see_diff runs a local pixel-level comparison: the change regions are detected and bounded by code, and the model is confined to explaining only the areas it has already proven changed. "Did it change?" stops being a probability question and becomes a fact.

Who is it for? You are running agent automation on Windows and need the agent to look at the screen mid-task, verify a render, or confirm whether a button was actually pressed. You are tired of manually capturing screenshots and pasting them into the chat, or of hard-coding coordinates into prompts and hoping the layout holds. Who is it not for? Cross-platform use, pixel-precise measurement, or low-contrast visual QA. The README states its limits with unusual honesty: it is a descriptor, not a caliper; it sees only the present moment; it runs on Windows only.