dsh-deepseek-vision: Adding image recognition to DeepSeek text models
dsh-deepseek-vision is an MIT-licensed DSH plugin designed to provide image recognition capabilities for text-only DeepSeek models. By registering the `deepseek_vision` tool, the plugin allows the model to submit local image paths (supporting PNG, JPG, WebP, and GIF, but not HEIC) and optional prompts. When invoked, the plugin utilizes the DSH subprocess service to safely call the local `deepseek-vision-cli` via array parameters, sends the image to the DeepSeek Web endpoint for recognition, and returns the descriptive text to the model. Users need to establish a Web session via manual login to the assistant or Token login; the Token is stored only in local cache. Installation supports local directories or Releases.
Read Moredsh-deepseek-vision: Adding Image Input Routing for DeepSeek Harness's Pure Text Model
dsh-deepseek-vision is a DSH plugin designed to enhance pure-text DeepSeek models with image input capabilities, addressing the gap where existing models cannot directly process visual content. The plugin registers an independent `DeepSeek + Vision` route, supporting multiple input sources such as chat window image pasting, screenshots, and tool-returned images. Its core mechanism employs a分流 processing approach: if the model directory declares visual support (e.g., deepseek-v4-flash-vision-exp), images are directed straight to the native visual endpoint; otherwise, a configurable VL model is called to convert the image description into text before passing it to DeepSeek for processing. The plugin offers an LRU cache-based description reuse mechanism to avoid duplicate billing, retains original images in session logs, and supports OpenAI-style gateways and
Read More