Agent & Workflowdsh-plugin
dsh-video-lens
dsh-video-lens is a DeepSeek Harness plugin that provides video understanding tools for text-only agents working with local video files. It includes video_probe for ffprobe metadata, video_analyze for scene-change-aware frame sampling with optional ASR and OpenAI-compatible vision analysis, and video_ask for time-anchored Q&A with grounded evidence.
Stars
112
Forks
2
Open issues
0
Last push
Sep 18, 2026
Latest release
v0.3.1
24h Growth
+0
7d Growth
+0
Installation
dsh plugin --profile <name> remove dsh-video-lensOverview
dsh-video-lens is a DeepSeek Harness plugin that provides video understanding tools for text-only agents working with local video files. It includes video_probe for ffprobe metadata, video_analyze for scene-change-aware frame sampling with optional ASR and OpenAI-compatible vision analysis, and video_ask for time-anchored Q&A with grounded evidence.
- Provides video_probe for fast ffprobe metadata: container, duration, resolution, fps, codecs, audio tracks, and subtitles.
- video_analyze uses ffmpeg scdet scene-change-aware frame sampling and optional timestamped ASR, fused with an OpenAI-compatible vision model into evidence JSON.
- video_ask answers time-anchored questions by parsing explicit time references or matching transcript keywords, re-sampling frames, and returning grounded eviden
- Vision and ASR are configured through OpenAI-compatible baseUrl, model, and key environment variables; the plugin does not lock users into a provider.
- Delegates media work to ffmpeg/ffprobe on PATH; the agent does not perform native decoding.
- ASR is additive: if unset or failing, visual analysis still completes and transcript is null.
- Payload size is bounded by options such as maxFrames, frameMaxWidth, and frameQuality.
- Requires Node.js >= 20; scdet scene detection needs ffmpeg >= 6.0, with older versions falling back to uniform sampling.