DeepSeek brain + automatic image recognition
Keep DeepSeek as the brain — paste images anyway.
Project overview
dsh-vision-proxy Keep DeepSeek as the brain — paste images anyway. GUI image attachments auto-transcribed for text-only DeepSeek on DeepSeek Harness. Why this exists
DeepSeek Harness natively gates image attachments on the selected model's declared inputModalities. DeepSeek's chat-completions line is text-only, so attaching an image with DeepSeek selected is rejected by design. Tool-based vision plugins exist, but GUI image attachments still fail with a text-only model.
A real conversation on the deepseek-vision route (DeepSeek-V4-Flash as the brain): the user pasted a meme and asked "你看到了什么" (what do you see?); the image was auto-transcribed by the VLM and DeepSeek answered from the text — one step, 7.6 s. Left: the model picker showing the deepseek-vision route (DeepSeek + 自动识图) selected — that is what admits image attachments. Right: DeepSeek's full answer derived from the transcribed image text.