oil-oil/dsh-vision
Near-native image understanding for DeepSeek Harness
dsh-vision adds image understanding to DeepSeek Harness regardless of your main model's capabilities. When your selected model supports images, it passes them through natively. When the main model is text-only, the plugin routes images to a separate vision model (ZenMux, Alibaba Cloud, TokenDance, or OpenRouter), then injects the visual description as context for the text model's final answer. Multiple images are analyzed together — comparisons and combined visual evidence work naturally.
Install
dsh plugin --profile web add github:oil-oil/dsh-visionRestart dsh web after install. Configure the Vision Recognition settings card under Settings → Plugins → Plugin configuration → Vision Recognition. Paste an API key from ZenMux, Alibaba Cloud Model Studio, TokenDance, or OpenRouter.
Compatibility
dsh 0.1.0-rc.6 and rc.7. Satisfies both legacy list Slot and current keyed Slot registration. Vision bridge for text-only main models; native passthrough for vision-capable models.
Details
- Repo: oil-oil/dsh-vision
- Category: Other
- Stars: 78
- Version: Supports dsh 0.1.0-rc.6 and rc.7
- Last push: 2026-08-18
- First seen: 2026-08-14
Recent updates
Falls back to macOS Vision or Tesseract if cloud vision is unavailable. Adds a Vision Recognition card to Settings → Plugins. API key is write-only in the browser — never exposed in chat, logs, or settings documents.
FAQ
- How do I install dsh-vision?
- Run: dsh plugin --profile web add github:oil-oil/dsh-vision — then restart dsh web and configure the Vision Recognition settings card with an API key from ZenMux, Alibaba Cloud Model Studio, TokenDance, or OpenRouter.
- Does dsh-vision work if my main model already supports images?
- Yes — when the main model is vision-capable, dsh-vision passes images through natively without any preprocessing. The bridge only activates for text-only models like deepseek-official.
- What vision providers does dsh-vision support?
- dsh-vision supports ZenMux, Alibaba Cloud Model Studio, TokenDance, and OpenRouter as cloud vision backends. Local fallbacks include macOS Vision and Tesseract. Choose Automatic to skip cloud credentials entirely.
Alternatives
Anionex/dsh-vision-toolkit · ysr666/dsh-vision-router