vllm_omni.engine.duplex.commands ¶

The command vocabulary one duplex session accepts, plus its mailbox rendering.

DuplexSessionHandle.submit() takes one of these; the websocket handler and InlineDuplexClient build them with :func:command_from_realtime.

The wire half of each command --- which client event it decodes from and what fields survive decoding --- now lives in vllm_omni.protocol.duplex.commands, which carries the whole vocabulary: the eight Tier 1 classes re-exported from vllm_omni.protocol.realtime.commands, the two Tier 2 ones and the seven Tier 3 ones. The engine imports that module and never the Tier 1 one directly, so a command that later grows a duplex extension changes one file (RFC #6592 P0a).

The engine half stays here, because it is not wire contract at all: payload() renders the session-internal mailbox dictionary the runner bodies were written against, and its type is the mailbox channel rather than the client event. Those two genuinely differ --- session.update and conversation.item.create / .delete / .truncate all travel on the turn.signal channel --- which is why the halves are separated instead of the classes being relocated wholesale.

REALTIME_COMMAND_TYPES: frozenset[str] = frozenset(
    {
        "input_audio_buffer.append",
        "input_audio_buffer.commit",
        "input_audio_buffer.clear",
        "output_audio_buffer.clear",
        "response.create",
        "response.cancel",
        "conversation.item.create",
        "conversation.item.delete",
        "conversation.item.truncate",
        "session.update",
        "playback.ack",
        "session.heartbeat",
        "session.close",
        "turn.signal",
        "input.text.append",
        "input.cancel",
        "barge_in",
    }
)

type: str = 'playback.ack'

audio: bytes = b''

type: str = 'input_audio_buffer.append'

payload() -> dict[str, object]

type: str = 'input.text.append'

type: str = 'barge_in'

type: str = 'input.cancel'

type: str = 'response.cancel'

type: str = 'input_audio_buffer.clear'

type: str = 'output_audio_buffer.clear'

type: str = 'session.close'

type: str = 'input_audio_buffer.commit'

payload() -> dict[str, object]

type: str = 'turn.signal'

payload() -> dict[str, object]

type: str = 'response.create'

payload() -> dict[str, object]

type: str = 'turn.signal'

payload() -> dict[str, object]

type: str = ''

payload() -> dict[str, object]

type: str = 'session.heartbeat'

event_id: str | None = None

wire_type: str = ''

type: str = 'turn.signal'

payload() -> dict[str, object]

type: str = 'turn.signal'

payload() -> dict[str, object]

type: str = 'turn.signal'

payload() -> dict[str, object]

command_from_realtime(
    payload: Mapping[str, object],
    *,
    defaults: RealtimeInputDefaults | None = None,
) -> DuplexCommand