dsh-watch-video

dsh-watch-video

中文文档 | Full Chinese documentation

Subtitle-first video transcription for DeepSeek Harness. It uses subtitles when available, falls back to faster-whisper large-v3, and writes transcript, SRT, summary, and metadata artifacts without retaining the original video.

Features

• HTTP(S) video input through yt-dlp.
• Subtitle-first processing with Whisper fallback.
• SRT, text, JSON, Markdown summary, and metadata artifacts.
• Cancellable jobs with status and list tools.
• Windows, Linux, and macOS support with process-tree cleanup.
• No external LLM calls and no original-video retention.
• Stale running jobs become interrupted after restart; the default staleJobAfterMs is 6 hours, and ambiguous heartbeats are skipped conservatively.

Child-process output is bounded independently per stream: stdout keeps a prefix and stderr keeps a tail, with limits measured in UTF-8 bytes while both streams continue draining. Timeout, cancellation, and output-limit termination clean the entire process tree and return bounded diagnostics. maxOutputBytes takes precedence over the legacy outputLimitBytes name.

Requirements

Node.js 20+, Python 3.10+, yt-dlp, and ffmpeg. Install Node and Python dependencies, run scripts/bootstrap.ps1, then npm run doctor. The first real Whisper fallback downloads the large-v3 model.

Tools

• dsh_watch_video
• dsh_watch_video_status
• dsh_watch_video_list
• dsh_watch_video_cancel

Verification

npm run verify:pack

This clean-package check runs entirely offline: it does not download a Whisper model or access real video sites.

Job status and cancellation are process-local. See README.zh.md for full configuration and platform details.