MJ-normalized turn-detection mode vocabulary — provider-neutral by design so a shared model
catalog is safe on every provider:
'default' — let the provider profile decide (byte-for-byte today's behavior).
'serverVad' — classic silence-based server VAD.
'semanticVad' — the provider's semantic end-of-utterance detection (OpenAI semantic_vad).
'native' — deliberately open-ended: "this model's smartest documented turn/duplex mode,
whatever the profile maps it to." The forward slot for full-duplex reasoning voice models
(e.g. Grok Voice Think Fast) — a profile that maps 'native' ships the mapping once and the
catalog opts models in via metadata, no driver release per model.
A profile that does NOT support a requested mode logs and falls back to its default — a wrong
inherited value degrades safely, it never rejects a session.
MJ-normalized turn-detection mode vocabulary — provider-neutral by design so a shared model catalog is safe on every provider:
'default'— let the provider profile decide (byte-for-byte today's behavior).'serverVad'— classic silence-based server VAD.'semanticVad'— the provider's semantic end-of-utterance detection (OpenAIsemantic_vad).'native'— deliberately open-ended: "this model's smartest documented turn/duplex mode, whatever the profile maps it to." The forward slot for full-duplex reasoning voice models (e.g. Grok Voice Think Fast) — a profile that maps'native'ships the mapping once and the catalog opts models in via metadata, no driver release per model.A profile that does NOT support a requested mode logs and falls back to its default — a wrong inherited value degrades safely, it never rejects a session.