Replaces fragile whole-screen text matching with a screenshot-diff step that
finds each dialog's actual pixel bounds, then restricts every subsequent AI
locate/click to that cropped region — eliminating false matches from text
elsewhere on screen (title bars, side panels). Adds column-boundary and
row-height detection so the patient row and Exam procedure click positions
are computed geometrically instead of relying on repeated fuzzy AI guesses
for visually similar neighbors. Also adds per-run screenshot/debug-crop
backups and a structured run.log for diagnosing failed runs, plus a
cmd:move primitive on the Windows agent for pre-click confirmation crops.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Gitead
ad2e15ec62