Add support for Action Classification from body movement

[JamesDale]

Summary

Adds Action Classification to PrototypeKit — classifying a person's action from their body movement using a Create ML / Core ML Action Classifier model.

The new ActionClassifierView is a camera-based SwiftUI view that:

• detects human body-pose keypoints per frame with Vision's VNDetectHumanBodyPoseRequest,
• accumulates a sliding window of frames (an action unfolds over time), and
• feeds a [window, 3, 18] multi-array into the model, publishing the predicted label through a Binding<String>.

It mirrors the API and behaviour of the existing model-backed views (ImageClassifierView, ObjectDetectorView, HandPoseClassifierView) and works on both iOS and macOS.

@State var latestPrediction = ""

ActionClassifierView(modelURL: ActionClassifier.urlOfModelInThisBundle,
                     latestPrediction: $latestPrediction)

Details

• ActionClassifierConfiguration exposes posesFeatureName ("poses"), labelFeatureName ("label"), predictionWindowSize (60), and predictionInterval (15) for models that differ from the Create ML template. The window size is taken from the model's own input constraint when it advertises one.
• Graceful degradation — a bad/missing model shows the camera feed with no predictions, logs the failure, and reports it via the optional onError closure. Accepts a camera: CameraOptions?.
• Robust windowing — frames without a detected body are zero-padded (matching how Create ML pads missing frames), so a person briefly leaving the frame doesn't corrupt the temporal window. Predictions are throttled to every predictionInterval frames once the window is full.

Docs & tests

• Graceful-failure tests for the new view and receiver (bad URL, onError, nil-model ignores frames).
• Documented across the README (Live Action Classification section), DocC catalog, CHANGELOG.md, the Examples/ gallery, and the plugin (skill guidance, README command table, new /prototypekit:action-classifier command).

Notes

No Swift toolchain is available in this environment, so the package was not compiled here; the code follows the established patterns of the sibling views.

🤖 Generated with Claude Code

---

Generated by Claude Code