dsh-video-lens

Many DeepSeek Harness agents are text-only and cannot directly interpret local video files, making tasks like video summarization, retrieval, or clip-based question answering difficult to automate. dsh-video-lens fills this gap by giving text-only models eyes and ears on video.

The plugin is built around three tools: video_probe quickly returns metadata such as container, duration, resolution, frame rate, and codecs via ffprobe; video_analyze uses ffmpeg scene detection for content-aware frame sampling, optionally adds speech transcription, and fuses the visual information with any OpenAI-compatible vision model into structured JSON evidence with timestamps; video_ask supports time-anchored Q&A by parsing expressions like "at 3:20" or locating relevant speech through transcript keyword matching, then re-sampling frames from the matched windows to produce grounded answers.

It avoids provider lock-in: both vision and ASR endpoints are configurable through base URL, model name, and environment variables. All media processing is delegated to the local ffmpeg/ffprobe binaries, so the agent itself does not need native decoding support.

This plugin is ideal for developers who want to add video understanding to their DeepSeek Harness workflows—whether building video summarization and indexing pipelines, or developing time-based retrieval and Q&A assistants for local video content.