| name | perception-routing |
| description | Route behavior-analysis perception to the lightest Annolid backend that can satisfy the task, preferring existing trackers and typed artifacts before heavier open-set grounding or propagation. |
| metadata | {"annolid":{"always":false}} |
Perception Routing
Use this skill when the bot must decide how to obtain the evidence required for
behavior analysis.
Preferred order
- Reuse existing typed artifacts and replayable sidecars if available.
- Reuse Annolid tracking or segmentation outputs already attached to the video.
- Use GUI/model workflows only when the required evidence is missing.
- Escalate to heavier routing only when the task truly needs open-set grounding
or propagation across long clips.
Decision factors
- whether typed
TrackArtifact rows already exist,
- number of animals or objects,
- occlusion risk,
- need for identity continuity,
- need for open-set object discovery,
- cost versus expected gain.
Output
Return a short backend choice plus a concrete reason, not a generic model list.
