DeepSeek Harness Plugin: Compresses a video into a single grid-style storyboard image with numbered cells, accompanied by a text list containing timecodes, dialogue, and cut counts, enabling the model to answer "What happened at minute 3?"; the 0.4.5 released by 10-30 primarily focuses on bug fixes.
Star count 195
Reason
- 10-28 has been received; it is listed again here because 0.4.5 fixed the issue where video_zoom could not find cells when zooming by cell number (a problem persisting since 0.4.4 due to storyboard cell mappings not being written to disk). Both remote links and local paths after download can now be queried successfully.
- The same version also fixed issues such as Bilibili DASH audio/video not merging, the same video opening twice during cold startup, and incorrect index retrieval for the local median of cut points; it added a local chat box chat.js (allowing interruptions even when tools are blocking), attachOnly (preventing plugins from launching a browser), prioritization of edge-tts neural voices with fallback to system SAPI on failure.
- Four tools: video_probe reads only file headers without consuming image tokens, video_storyboard generates images and lists, video_zoom zooms into a specific segment, and video_locate reverse-searches timecodes using a line of dialogue.
Usage
- winget install Gyan.FFmpeg (required); to paste Bilibili / YouTube links, run pip install --user -U yt-dlp; install the plugin via node scripts/install-dev.mjs and then restart DSH.
- First run video_probe, then video_storyboard({ input: "local path or link" }); if a segment is deemed important, use video_zoom({ from, to, cells }) to zoom in; if subtitles are already available, use transcriptFile to skip ASR.
Risks
- The author states that a single image uses 384 tokens, one call uses approximately 1,900 tokens, and a 90-second video takes about 10 seconds to produce results; these are all self-tested by the author. Obtaining dialogue requires configuring a separate ASR service or using existing subtitles.
- High-definition / member-only videos on Bilibili require exporting cookies, where SESSDATA is equivalent to login credentials; reading playback progress for "Watch Together" relies on Windows system media controls.