Struct SessionTelemetry 

pub struct SessionTelemetry { /* private fields */ }

Zero-overhead telemetry collector for speech-to-speech sessions.

Designed for the three-lane processor model:

• Fast lane (sync, <1ms): No telemetry calls — pure audio/text forwarding.
• Router: Calls record_* methods on every event, in event order. These use only atomic operations — no allocations, no locks, no syscalls.
• Control lane (async): Calls snapshot() at turn boundaries to get aggregated stats as a JSON value ready to send to the browser.

Implementations

impl SessionTelemetry

pub fn new() -> Self

Create a new telemetry tracker, starting the session clock.

pub fn record_audio_out(&self, byte_len: usize) -> Option<Duration>

Record an outgoing audio chunk. Called from the router.

Returns the response latency when this chunk is the model’s first output after the user’s end of speech (or text send) — once per turn, via a CAS so only the first chunk wins.

pub fn record_vad_end(&self)

Record VAD end (user stopped speaking).

pub fn record_text_send(&self)

Record that user sent a text message (for text-input latency tracking).

pub fn record_text_out(&self) -> Option<Duration>

Record first model text output (TextDelta). Tracks text-input latency.

Returns the response latency when this delta is the model’s first output after a text send.

pub fn record_interruption(&self)

Record an interruption (barge-in).

pub fn record_turn_complete(&self)

Record turn completion for duration tracking.

pub fn record_usage( &self, total: Option<u32>, prompt: Option<u32>, response: Option<u32>, cached: Option<u32>, thoughts: Option<u32>, )

Record token usage from a UsageMetadata event.

pub fn record_turn_usage(&self, usage: &UsageMetadata)

Add one turn’s usage, by modality, to the session totals, and to the gemini_genai_rs_tokens_total metric. Pass the last usage report of the turn: each report covers the turn so far.

pub fn mark_turn_start(&self)

Mark the beginning of a new turn (e.g., when model starts responding).

pub fn latency(&self) -> LatencyStats

Per-turn response latency: end of user speech (or text send) to the model’s first output, as a distribution over the session so far.

Cheap enough to call once per turn (it sorts at most LATENCY_RECENT_WINDOW samples); not for the audio hot path.

pub fn snapshot(&self) -> Value

Snapshot all metrics as a JSON value.

The flat *_response_latency_ms keys are kept for existing dashboards; response_latency carries the full LatencyStats (percentiles and histogram included).

Trait Implementations

impl Default for SessionTelemetry

fn default() -> Self

Returns the “default value” for a type.