Proposal: allow image uploads in GUI when the main model is text-only but vision tools are installed

[fenghui12]

Summary

When using DeepSeek Harness with a text-only main model (e.g. DeepSeek V4 Flash via the deepseek-official provider) together with the vision plugin (dsh-vision-toolkit: vision_glance / vision_detect / vision_pixel_diff etc.), the GUI rejects image uploads with:

当前模型不支持图片，请切换支持图片的模型 (reason: MODEL_DOES_NOT_SUPPORT_IMAGES)

Why this is a gap

The vision tools read image paths from the session workspace and work perfectly with a text-only model — we use them daily. But the GUI upload path in dsh-host-apiproxy assumes images can only be consumed as model input: when a prompt contains image parts it checks the current model's inputModalities and refuses when the model does not declare "image".

So a text-only main model can never receive user images, even though the harness has vision tooling that could read them from disk.

Suggested behavior

When the selected model does not support image input, do not reject the upload. Instead:

1. Save the uploaded image into the session workspace (e.g. .dsh-uploads/)
2. Hand the agent a text reference (file path) in the user message, e.g. [用户上传了一张图片，已保存到 <path>。请使用 vision_glance 等视觉工具查看。]

This matches the existing "images are for tools, not the model" architecture and makes the text-only + vision-tools combination fully usable from the GUI.

Workaround

We currently apply a local patch to dsh-host-apiproxy that converts image parts into saved files + path references. It works well; it would be great to have this behavior built in (ideally configurable, e.g. only when the model lacks image input, or always when vision tools are present).

Environment

• DeepSeek Harness latest rc
• DeepSeek V4 Flash (text-only) + dsh-vision-toolkit
• Web profile (browser GUI)

[yuqingsh]

Feel free to try our zero-core-patch plugin for this, verified on 0.1.0-rc.6. repo: https://github.com/yuqingsh/dsh-image-subagent
We do three things:

1. Admission: wraps the llm service via cordis's internal/get waterfall so resolveModelInfo reports image input for text-only routes — pasting images no longer fails with MODEL_DOES_NOT_SUPPORT_IMAGES (listModels stays truthful, the model picker is not misled).
2. LLM boundary: wraps prepareCall/stream as lazy generators that clone the (deep-frozen) request and project image blocks into explicit text placeholders carrying the attachmentId — no more UNSUPPORTED_CONTENT, the turn completes.
3. Delegation: the text-only main model hands the image to an in-session multimodal subagent, which reads the durable attachment via read_attachment — native multimodal comprehension (multi-turn follow-ups), no vision-API credentials, and the image stays inside the harness's own attachment store with session-scoped authorization.