dots3-note-prev (Image-Text-to-Text) incorrectly detected as non-vision, images rejected by unsupportedParts() #43392

Description

When uploading an image while using dots3-note-prev, the model receives this error instead of the image:

ERROR: Cannot read clipboard (this model does not support image input). Inform the user.

This is wrong — dots3-note-prev is a multimodal omni model (280B-A16B omni MoE) that natively supports image, audio, and video input. It should accept images.

Root Cause

unsupportedParts() in packages/opencode/src/provider/transform.ts:438-440 strips image parts when model.capabilities.input.image is false. Capability detection in packages/opencode/src/provider/provider.ts:1267 derives this from the external model config:

image: model.modalities?.input?.includes("image") ?? false,

If dots3-note-prev is not listed in the model config with modalities.input including "image", it defaults to false and images get rejected — even though the model actually supports them.

Steps to Reproduce

1. Configure dots3-note-prev as the model
2. Upload/paste an image
3. Observe: error text instead of image processing

Expected Behavior

Images should pass through to the model since dots3-note-prev natively supports image input (as well as audio and video).

Evidence

• Official model page: https://huggingface.co/dots-studio/dots3-note-prev
• Requantized variant: https://huggingface.co/JANGQ-AI/dots3-note-prev-NVFP4 (280B / 16B total/active params, NVFP4 for Blackwell/GB10)
• Model card lists: Multimodal (image · video · audio), pipeline task: image-text-to-text

Environment

• Model: dots3-note-prev (280B-A16B omni MoE, Image-Text-to-Text + audio + video)

Related

• unsupportedParts() "Inform the user" instruction overrides system prompt, preventing fallback to vision MCP for non-vision models #40495 — same error message for non-vision models
• Non-vision models blocked from passing images to vision-capable MCP tools #29216 — non-vision models blocked from passing images to vision MCP
• fix(provider): emit file metadata instead of error when model lacks image/file support #29279 — previous PR to fix this (closed, not merged)