OptionalcontextUser: UserInfoThe user the engine config runs as (server-side).
Optionalprovider: IMetadataProviderThe request-scoped metadata provider (multi-provider safe).
OptionalresolveAPIKey: AIAPIKeyResolverKey-resolution seam deciding which vendors count as runnable. Omitted (the
voice-picker query has no run context), each vendor resolves as a session would: its platform MJ
Credential, then AI_VENDOR_API_KEY__<driver> — so the picker offers the models a session can mint.
Active realtime models, each with its driver's voices.
Enumerates the active Realtime models with each driver's supported voices — the source for the dev model/voice picker. Only models with an Active vendor + resolvable API key + ClassFactory driver are returned (a model you can't actually run isn't worth offering). Voices come from the driver (BaseRealtimeModel.SupportedVoices) — the near-term, driver-owned source of truth. A persona voice whose persona also has a preset Video binding on the same vendor names that avatar (RealtimeModelVoiceOption.AvatarID), when the model shows avatars on that vendor (RealtimeModelShowsAvatar: its Video/Output row allows video or it has none, and the driver renders avatars on its endpoint); driver voices never do. So the picker marks a voice "Comes with an avatar" only where a call can show one. A persona with a face doesn't stand in for its plain voice, so the plain voice is listed too.