The HuggingFace provider profile — the per-provider knobs the shared OpenAIRealtime
protocol implementation runs with when driving a self-hosted speech-to-speech endpoint:
No default input-transcription model: the cascade's STT stage transcribes the user
natively; a transcription block is only sent when the Config bag supplies
inputTranscriptionModel (and only names a model if the compat layer supports selecting one).
Config deferred to session.created — the endpoint validates session.update against the
GA shape once the session exists; the readiness gate below rides the same frame.
InitialContext folds into the system prompt under a "Prior context" heading — the compat
protocol has no guaranteed history-seeding channel.
GA feature gates OFF: self-hosted stacks lag the GA reasoning/MCP surface; feature keys in
a shared co-agent config are scrubbed, never sent raw. Flipping a gate here is all it takes
once the compat layer supports one.
The HuggingFace provider profile — the per-provider knobs the shared OpenAIRealtime protocol implementation runs with when driving a self-hosted speech-to-speech endpoint:
inputTranscriptionModel(and only names a model if the compat layer supports selecting one).session.created— the endpoint validatessession.updateagainst the GA shape once the session exists; the readiness gate below rides the same frame.InitialContextfolds into the system prompt under a "Prior context" heading — the compat protocol has no guaranteed history-seeding channel.