video-digest — orchestrator

Turn videos into a digest site: overview card grid → per-talk detail (TL;DR, "take this to work", slides with captions), optional bilingual toggle + synthesis essay. Generalized from the AIE WF2026 pipeline.

Stage skills (mix & match)

Stage	Skill	In → Out
1. Ingest	`video-digest-ingest`	source (URL/playlist/folder) → `video.mp4`, transcript, 30s timeline per item
2. Segment	`video-digest-segment`	ONLY multi-talk streams: timeline → `talks.json`. Single-video items: auto (1-talk talks.json, done by ingest)
3. Summarize	`video-digest-summarize`	transcript slices → `sessions/.md`, `.take.json`, optional `zh.json`
4. Slides	`video-digest-slides`	highlights → candidate frames → montages → vision-picked `frames/sel_*.json`
5. Site	`video-digest-site`	everything → `index.html`/`dist/` (+ optional essay), publish via here-now or Artifact

Routing by source shape:

YouTube playlist / folder of videos → ingest each video as its own item (skip segment). One group per playlist/folder.
Long multi-talk stream → one item, then segment.
Single talk video → one item, skip segment.