v1.0.29

v1.0.29

Vision Bridge v1.0.1

Fixed

• Images from earlier turns reached the text-only model. Describe mode only replaced the newest image-bearing message, so every older image was re-sent untouched on each request (500 image input is not supported on Ollama, or a confident description of an image the model cannot see). Anything the vision pass does not cover is now replaced by a text marker, in both modes.
• Describe mode purges the analyzed image from the chat again. Open WebUI inlines uploaded images as data URIs before filters run, so the file id was no longer in the request and purge_from_history and delete_file_record had quietly stopped firing, which is what kept the old images coming back. The image is now identified through the stored chat.
• A failed vision call no longer consumes the image: the chat message and the file record are left untouched.
• Images beyond max_images are flagged in the message instead of vanishing, and the purge is skipped while any image is left undescribed.
• strip_only markers now name a call the model can actually make, replacing the dead [Image attached, no file id available.].
• Both the filter and the tool recognise the temporary: chat id prefix, so a temporary chat is no longer looked up in the database.