video-digest-summarize

Input: <item_dir> with talks.json + transcript/transcript.tsv. See video-digest for layout/config.

1. Deterministic prep (scripts/)

python3 scripts/vd_sessions.py <item_dir> [--clips]   # slices + manifest.json (+ per-talk clips)
python3 scripts/vd_clean.py <item_dir>                # cleaned transcript, glossary from digest.json

• vd_sessions.py: per talk sessions/NN_slug.txt ("[HH:MM:SS] text" lines, video offsets) + manifest.json. --clips also cuts lossless per-talk mp4s (only useful for multi-talk streams; skipped automatically when a talk spans the whole video).
• vd_clean.py: collapses Whisper loop/applause runs, applies glossary/nonspeech from the project's digest.json (walks up from item_dir) → transcript/transcript.clean.txt, sessions/*.clean.txt, transcript/corrections.md.

2. LLM steps (prompts in references/; fan out ~5 talks per agent)