Assistance Plugin Development

PluginSDK defines separate protocols for box, label, segmentation, track correction, annotation quality, active-learning priority, and dataset balance output. Each async request is bounded by a deadline and explicit frame/context values. Implementations should check task cancellation and return promptly.

All outputs must be Codable, version 1, provenance-aware, within normalized bounds, confidence and uncertainty in zero through one, under 10,000 records, and unexpired. Segmentation suggestions require a mask. Invalid output throws pluginOutputInvalid. No protocol exposes direct access to verified annotation storage.

The built-in fixture is deterministic and needs no third-party license. External models require the registry fields described in Adding Assistance Models.