DSH-Vision: A Plugin Adding Image Recognition Capabilities to DeepSeek Harness

dsh-vision is a DeepSeek Harness (DSH) bundle plugin maintained by Terry12138qy, licensed under the MIT License. This plugin is designed to address the issue that the DSH main model lacks native image recognition capabilities. By encapsulating the image recognition workflow, it sends local image paths or network URLs to Alibaba Cloud Bailian or Zhipu multimodal models, ultimately returning Chinese text descriptions for use by agents or Code Mode. Core features include automatically reading and encoding images, prioritizing the main model `qwen3.5-omni-plus`, and automatically switching to the fallback model `glm-4.6v-flash` for retries upon failures (e.g., quota exhaustion, rate limiting). It supports invocation via `await tools.vision(...)`.

Read More
dsh-video-lens: Equipping Pure Text DSH Agents with Video Understanding Capabilities

dsh-video-lens is a DeepSeek Harness plugin maintained by dundunhan (v0.3.1, MIT). It utilizes three tools—video_probe, video_analyze, and video_ask—to read local video metadata via ffprobe/ffmpeg, perform scene-aware frame extraction with optional ASR, and interface with OpenAI-compatible vision models, outputting structured evidence JSON with timelines. Installation requires running `pnpm add dsh-video-lens` in the DSH Profile directory and registering the bundle, along with configuring environment variables such as `VIDEO_LENS_API_KEY`. It is suitable for video review, meeting recording summaries, and time-anchored Q&A; users are advised to review the source code and permission notes before use.

Read More