Colab Video Pipeline

Use this skill to process Jiang Lens video sources through the Colab notebooks in ops/notebooks/colab/.

Safety Boundary

• Do not commit Google cookies, YouTube cookies, HuggingFace tokens, Colab browser profiles, audio, or video.
• Local browser auth belongs under ops/secrets/browser-profiles/colab/.
• YouTube yt-dlp cookies belong in Google Drive at /content/drive/MyDrive/jianglens/cookies.txt.
• HuggingFace auth should use Colab userdata key HF_TOKEN when possible.
• Stop and ask the maintainer on Google login, 2FA, CAPTCHA, account chooser ambiguity, quota exhaustion, or unexpected paid-credit prompts.

Drive Layout

The canonical Drive root is:

/content/drive/MyDrive/jianglens/
  _colab_envs/
  _hf_home/
  cookies.txt
  youtube/
    _config.json
    <channel-or-handle>/
      _channel.json
      <video-id>/
        audio.wav
        metadata.youtube.json   # optional; local import can create this by video id
        dump.json
        grouped.json
        transcription.json

Local text artifact sync uses:

./ops/notebooks/colab/sync-drive.sh --dry-run
./ops/notebooks/colab/sync-drive.sh

Notebook Order

1. YouTube_Manager.ipynb: register channels, filter, download audio.wav into Drive.
2. Pyannote_4_Pipeline-GPT-5.3.ipynb: produce dump.json and grouped.json.
3. Whisper_Transcription.ipynb: produce transcription.json.
4. sync-drive.sh: copy text artifacts from Drive to content/sources/raw/youtube.

Automation Contract

Before running browser automation:

1. Confirm the maintainer has created MyDrive/jianglens.
2. Confirm the target notebook is already saved in the maintainer's Colab account.
3. Use a persistent Playwright profile under ops/secrets/browser-profiles/colab/.
4. Run one notebook at a time. Do not start parallel Colab runtimes unless explicitly asked.
5. Capture screenshots and console/log evidence under ops/staging/.

Success means expected Drive artifacts exist:

• YouTube stage: audio.wav
• Metadata stage: metadata.youtube.json in synced raw source artifacts when YouTube metadata was resolvable. If the notebook did not write it, ops/scripts/import-colab-video.mjs can fetch and cache it by video id.
• Diarization stage: dump.json and grouped.json
• Transcription stage: transcription.json with turns[].words[] word timestamps when Whisper produced them.

The website transcript reader uses those word timestamps for phrase-level highlighting. If transcription.json has no turns[].words[], do not silently proceed as if phrase tracking exists; rerun Whisper_Transcription.ipynb with word_timestamps: True or mark the import as block-timed only.

If a notebook has RUN_BATCH = False, do not assume it processed videos. Either the maintainer must set it in Colab, or the automation must intentionally set it before running the batch cell.

Import Handoff

After syncing Drive artifacts into the repo, import a processed video with:

node ops/scripts/import-colab-video.mjs --video-id VIDEO_ID --channel @PredictiveHistory

The importer uses raw metadata.youtube.json when present. If it is missing, it tries yt-dlp by video id and caches compact metadata back into the raw source artifact folder. Publication dates are required for canon promotion; undated imports are review/search material only.

If YouTube blocks metadata-only access, provide a runtime-only cookie file. The importer accepts --youtube-cookies PATH, checks JIANGLENS_YOUTUBE_COOKIES_FILE, YOUTUBE_COOKIES_FILE, and YTDLP_COOKIES_FILE, and also looks for ignored local files at ops/secrets/youtube.com_cookies.txt or /var/lib/spawnfile/secrets/youtube.com_cookies.txt. Never commit the cookie file or print its contents.