[deepanwadhwa]

An ordinary image upload could enter the app's bounded analysis_depth=fast path, which disabled visual analysis and left the final text model with only an OCR status message. The image OCR bridge also expected a top-level text field even though native Tesseract returns a lines array, so successful OCR was reported as unavailable.

Every image plan now gathers both literal OCR and visual evidence. The gateway joins Tesseract line objects into bounded text, invokes the installed visual specialist (preferring Molmo2), and sends separate OCR: and MOLMO IMAGE DESCRIPTION: sections to the selected final chat model. Empty OCR and failed OCR remain distinct. Fast image handling stays bounded by using an overview Molmo pass; documents retain their existing fast reader behavior.

Validation:

• make test
• make test-molmo2-gateway
• make ocr-test test-pdf-ocr-runtime
• node tests/test_composer_ui.mjs
• node tests/test_composer_perf.mjs
• Manual native Tesseract read of the supplied workout screenshot recovered its workout title, duration, distance, elevation, heart-rate, and pace text.