shainedemo/dsh-vision-pro-bridge
Vision bridge for DeepSeek-V4-Pro: images are described by Flash-Vision then forwarded to Pro — no model switch, no extra API key, zero new dependencies.
Gives the text-only deepseek-v4-pro model image understanding without switching models or adding a second API key. It registers a twin provider route that intercepts image inputs, describes them using deepseek-v4-flash-vision-exp, and forwards the description to deepseek-v4-pro for the actual response — so pasting a screenshot into a Pro session works instead of failing with the does-not-accept-image-input error.
Install
dsh plugin --profile web add dsh-vision-pro-bridgeInstalls as dsh-vision-pro-bridge 1.0.0 into the web profile; restart dsh web to activate. No new API key is required — the plugin reuses the existing DEEPSEEK_API_KEY. It registers a twin provider route deepseek-vision-pro (model deepseek-v4-pro-vision) alongside the standard deepseek-v4-pro route. Zero third-party dependencies beyond the harness's own @deepseek-ai/dsh-llm and @deepseek-ai/dsh-llm-deepseek.
Compatibility
Images go only to DeepSeek (deepseek-v4-flash-vision-exp) through the same DEEPSEEK_API_KEY — no second VLM vendor, no local model, no Ollama. The plugin subclasses the official DeepSeekAdapter so image normalization, Files-API upload, and inline-base64 fallback are identical to official behavior. After the vision step the image block in the conversation is replaced with a text description; the now text-only message is then streamed to deepseek-v4-pro.
Details
- Repo: shainedemo/dsh-vision-pro-bridge
- Category: Vision & Multimodal
- Stars: 0
- Version: 1.0.0 (npm)
Alternatives
shinegeer/dsh-vision-window · Einskyle/dsh-llm-vision-bridge · ethanweave/dsh-vision-better