shuxu-zhao/dsh-image-dscriber

README 摘要

📖 English Docs  •  📖 中文文档 English Giving text models eyes to see in DeepSeek Harness. dsh-image-describer is a native plugin for DeepSeek Harness that provides dual-mode visual understanding capabilities for text-only LLMs (such as deepseek-v4-flash , deepseek-v4-pro ). 🌟 Dual-Mode Architecture The plugin seamlessly supports two complementary visual interaction scenarios: Direct UI Chat Image Pasting (Attachment Mode) and Active File Inspection (Tool Mode) . Scenario 1: Direct UI Chat Image Pasting (Attachment Mode) User Action : Direct Ctrl + V screenshot paste or drag-and-drop into the chat box. Workflow : 1. The image enters the session as an ImageBlock attachment in the user's message. 2. The plugin intercepts the request at the DSH llm/stream pipeline. 3. The plugin calls the configured multimodal vision model (default MiniMax-M3 ) for OCR and comprehensive visual recognition. 4. The plugin replaces the ImageBlock with an explicit, structured system prompt: 5. The text-only model receives clean context and responds immediately without crashing or unnecessary workspace searches. Scenario 2: Active Tool Calling on Local Files (Tool Mode) User Action : Mentioning loc…

分类