video-digest-slides

Input: <item_dir> with manifest.json + sessions/*.md (highlights with timestamps). Requires Pillow (uv run --with pillow if missing).

1. Deterministic (scripts/)

python3 scripts/vd_frames.py <item_dir>    # frames/<stem>/hNN_{0,1,2}.jpg + highlights.json
python3 scripts/vd_montage.py <item_dir>   # montages/<stem>_pN.jpg + montages/index.json

Frames at 960px wide, offsets −5/0/+5s per highlight. Montage grid: rows = highlights, 3 columns of candidates, labeled hNN:j HH:MM:SS, 8 rows/page.

2. Vision LLM steps (fan out ~5 talks/agent, CHEAP models)

1. Select (references/select-prompt.md, haiku/sonnet): read montages, pick per highlight the candidate j (0/1/2) that is a clean readable slide of THIS talk, or null (speaker/audience/transition/sponsor wall). Writes frames/sel_<stem>.json = {"0": j|null, ...}.
2. Verify (references/verify-prompt.md, sonnet): full-res re-check each pick against the highlight text; fix or null. Overwrites sel_<stem>.json. Don't skip — the cheap select pass overpicks.

Fireside chats / no-slides talks: all-null is correct; the digest renders text-only points.