@cbg33695

Let a text-only model see the screen: capture desktop/window/region, vision transcription, a few minutes of rolling screen memory, exact local pixel diff, and vision self-calibration. Windows only. Experimental - the README states what is measured and what is not.