dsh-sight: Plug-and-Play Visual Capabilities for Text-Only DSH Models
dsh-sight is an MIT-licensed plugin designed for DSH (DeepSeek Harness) text-only models, aiming to endow them with image understanding capabilities through a built-in VLM backend. By wrapping the prompt admission process, the plugin automatically saves pasted images as local-path prompts and allows the model to invoke the `vision` tool. This tool supports passing in up to 10 local paths or URLs at a time, with the backend VLM generating individually annotated text descriptions. The plugin offers free presets such as `opencode-zen` and `gemini-flash`, and also supports connecting to any OpenAI-compatible endpoint. Users can configure it via the web settings page or environment variables (which take precedence over the configuration file), and it supports
Read More@dsh-extension/dsh-vision-bridge: On-Demand Visual Capability Integration for Text-Only DSH Sessions
`@dsh-extension/dsh-vision-bridge` is a third-party plugin that provides on-demand visual capabilities for text-only DSH sessions. It addresses the issues where text models cannot process image blocks and sending long history directly to vision models is prohibitively expensive. Mechanically, the plugin preserves the original images in the session and UI, only reformatting the images into text tokens at the text model's input layer. When the model needs to understand an image, it invokes the `vision_describe` tool, sending only the specified images (1-4) and a focused question to an OpenAI-compatible vision endpoint, thereby avoiding the transmission of lengthy session histories. The plugin supports installation via npm, with configuration options including enable status, base URL, API key (with environment variable support), and vision model name. Configuration priority follows the order: settings.
Read More