v2-refactor-temp/docs/breaking-changes/2026-07-30-chat-image-ocr-native-fallback.md
When an image is attached in Home Chat and addressed to a model without declared image-recognition support, Cherry Studio still runs OCR first — but if OCR finds no text, is not configured, or fails, the image is now forwarded to the provider as a base64-backed native image instead of being replaced with a [could not read this file] note.
Photos, charts, and other images without recognizable text are no longer silently reduced to an unreadable-file note on "non-vision" models. Models whose vision capability is under-declared in their metadata now receive the actual image; providers that truly cannot process images may ignore or reject it according to their own behavior.
Nothing. Images with recognizable text keep the existing OCR-to-text behavior; configuring an image_to_text processor still improves results for text-heavy images on non-vision models.
Only the empty-OCR / OCR-unavailable path changed; OCR-with-text, explicit OCR features (translation workflow), and the read_file tool are unchanged.
#18297 replaced this fallback with a localized error that failed the turn before the request (shipped in v2.0.5 with an "action required" release note); the behavior described above is what v2 ships. Users on a gateway or proxy that accepts images could not send any image while the block was in place, and because attachment routing replays the whole conversation, one such image failed every later turn of it too. Describe the net behavior once — do not carry the v2.0.5 "action required" wording into the release note.