Struct HardwareSession 

pub struct HardwareSession {
    pub device_id: String,
    pub device_type: DeviceType,
    pub caps: Caps,
    pub mode: Mode,
    pub ambient_session_id: Option<String>,
    pub conversation_id: String,
    pub agent_id: Option<String>,
    /* private fields */
}

Live state for one connected device.

Fields

device_id: String

device_type: DeviceType

caps: Caps

mode: Mode

ambient_session_id: Option<String>

The ambient long-running meeting id, if this device is ambient-capable.

conversation_id: String

Stable per-device conversation id used for each chat turn’s trace + (future) history binding. NOTE: v1 hardware chat is per-turn STATELESS — the turn runs run_text_turn(persist=false) with only the current user message, so prior turns are not replayed into the model. The stable id is here so a later persist=true / history-prefill upgrade has a durable key to hang on. Public so the Core WS pump (which owns the kernel-welded chat turn) can build the turn’s key without reaching into the session.

agent_id: Option<String>

The agent that handles device chat turns (None = the default LLM path). Resolved once at connect by the Core WS handler and carried here so the pump can build the kernel-welded ChatTurn from a pure TurnInput.

Implementations

impl HardwareSession

pub fn new( device_id: String, device_type: DeviceType, caps: Caps, ambient_session_id: Option<String>, meetings: Arc<dyn MeetingIngest>, agent_id: Option<String>, ) -> Result<Self>

Create a session from the device’s hello. ambient_session_id is the resumed/opened long-running meeting (set by the WS handler when the device is ambient-capable); None for interactive-only devices.

pub fn tts_format() -> AudioFormat

The TTS downlink format advertised back to the device in hello_ack.

pub fn set_mode(&mut self, mode: Mode)

Switch operating mode. On entering chat we clear any stale turn buffer.

pub fn on_listen_start(&mut self)

Begin a chat turn: drop any buffered audio so the turn starts clean.

pub async fn on_audio( &mut self, opus_packet: &[u8], ) -> Result<Vec<SessionOutput>>

Handle a decoded uplink Opus packet for the current mode.

In chat it accumulates the turn (the model runs on listen:stop). In ambient it buffers ~1 s then feeds a WAV chunk to the meetings pipeline, returning ambient_ack/ambient_skip. Idle mode ignores audio.

pub fn take_voice_turn(&mut self) -> Option<TurnInput>

Take the buffered chat turn on listen:stop. Returns the captured 16 kHz mono PCM as a pure TurnInput, or None when nothing was captured. The Core WS pump wraps this in its kernel-welded ChatTurn (with the session deps + conversation_id) and spawns the ASR → model → TTS turn off the recv loop so the loop stays live for barge-in.

pub fn take_text_turn(&mut self, content: &str) -> Option<TurnInput>

Build a text-fallback turn input (skips ASR). None for empty input.

pub fn emotion_for_phase(&self) -> Emotion

Map a chat/processing phase to the face emotion to push. Used by the WS handler when it wants to nudge the face outside a full turn.