Worked example: video foundation

This is an illustrative example with placeholder candidates and scores, not a captured Hub search or runtime result. Actual output depends on current metadata.

This example shows the video foundational modality with facets understanding, generation, and action. Profiles seed existing vision and general roles — no new video discovery runtime.

CLI

python3 scripts/mlx-agent research \
  --domain "sports highlight captioning" \
  --modality video --facet understanding

Or detect from domain text containing video / caption.

Resulting intent (shape)

{
  "domain": "sports highlight captioning",
  "roles": ["vision", "general"],
  "keywords": ["video", "caption", "understanding"],
  "modalities": ["video"],
  "facets": ["understanding"]
}

Pack excerpt

## Runtime preference

- Preferred: `mlx-vlm`
- Alternates: `lmstudio` when its server is up
- Rationale: Video understanding uses the selected vision runtime; review model support before loading.

## Modality foundations

- `video` — Video
  - Facets: `understanding` (Understanding / captioning)

## Candidates
...

Generation and action packs reuse the same foundation with --facet generation or --facet action.