Builds candidate actions from OCR and UI state for Jev to control macOS, calling a text model only when writing is needed.
Selects the next step from deterministically extracted controls and actions before desktop execution.
Separates screen reading, action selection, and text generation.
README cost and speed comparisons use the author’s limited screenshot and goal setup, with some reasoning supplied by deterministic code. They are not general savings multipliers and were not independently retested.
Explore macOS control through OCR and accessibility candidates with a separate text-writing helper.
Material analysed：
Decision.confidence takes the minimum of kind and selected target confidence for target-bearing actions. An irrelevant head does not lower confidence for another action.
Inspect the supporting source ↗The README distinguishes a one-step proposal from --act, which drives the machine. Low confidence and repeated no-ops are stopping conditions, not evidence that all harmful actions are prevented.
Inspect the supporting source ↗Prefer this when a native screen needs both OCR and accessibility data. For a DOM-heavy web app, compare a browser-specific observer before accepting screen-capture and focus overhead.
Inspect Decision.confidence and verify_typed in decide.py. For an item click, compare the two confidences; for typing, trace the additional ok Noul. Check the README’s dry-run and --act distinction before trying any live run.