turn-timing.md

Reading reply timing

The agent emits codetrial turn_timing when the first reply audio chunk is accepted into its output queue. Later chunks do not emit another summary. A pending measurement that ends without that chunk emits outcome=no_audio, interrupted, replaced, paused, held, or ended instead. Replacements finish the old measurement before the socket number changes; the new socket starts with no inherited timestamps or tool counters.

For example, this synthetic measurement has one tool call before audio:

codetrial turn_timing room=interview-example session=42 socket=1 turn=7 at=0:03.000 start_source=pcm_level wait_ms=3000 elapsed_ms=3000 transcript_wait_ms=2900 turn_close_ms=1500 turn_close_source=voice_activity_received turn_close_reason=observed vad_audio_offset_ms=99000 tool_calls=1 pre_tool_ms=2000 tool_phase_ms=1000 tool_exec_ms=400 tool_response_ms=100 tool_response_errors=0 tool_wait_ms=500 tool_after_close_ms=500 other_ms=1000 outcome=audio

room, session, socket and turn identify a measurement. turn is a local sequence, not a provider turn ID or the usage_event counter. at is elapsed interview time at emission. The line contains no transcript, code, tool arguments, credentials or tokens.

Field	Boundary or meaning
start_source=pcm_level	Last candidate PCM frame above the local voice-level threshold before the measurement freezes
start_source=last_transcript	Last input transcription receipt that armed a candidate reply wait, when no PCM proxy is available
start_source=unknown	No usable start proxy was observed, as in a greeting, platform prompt reply, or activity-end signal without a recent PCM frame or transcript
wait_ms	From that start proxy to the first accepted reply audio chunk; unknown when no audio arrived or no start was observed
elapsed_ms	From the same proxy to emission, including outcomes with no audio
transcript_wait_ms	From the last input transcript that armed a candidate reply wait to accepted reply audio; a separate receipt-clock cross-check
turn_close_ms	From the start proxy to receipt of Gemini voiceActivity.ACTIVITY_END; unknown if absent or ordered before the start
turn_close_source	voice_activity_received when that interval is available, otherwise unknown
turn_close_reason	observed, no_start, not_observed, before_start, or after_audio; missing and unusably ordered observations remain unknown
vad_audio_offset_ms	Provider audio-stream offset accompanying that signal, if valid; never mixed with the local monotonic clock
tool_calls	Calls locally executed before the first accepted audio chunk
pre_tool_ms	From the start proxy to entry into the first tool handler, if both are observed in that order
tool_phase_ms	From that first tool handler to accepted reply audio, including sends, board preparation, subsequent calls and model continuation; unknown without tools or reply audio
tool_exec_ms	Sum of local tool execution durations, excluding network sends
tool_response_ms	Sum of tool-response WebSocket send durations, including failed attempts
tool_response_errors	Failed tool-response sends in that window
tool_wait_ms	Union of local execution and response-send intervals clipped to the speech-proxy-to-audio window; overlapping work is counted once
tool_after_close_ms	The same union clipped to the valid turn-close-to-audio window; unknown without a usable close or reply audio
other_ms	Wait remaining after turn-close and tool work after close; unknown when either component is unavailable or inconsistent

A transcription or provider activity observation establishes candidate speech; microphone levels alone cannot assign a candidate wait to a greeting. A PCM frame is usable when it precedes the speech observation by less than two seconds; an older frame cannot stand in for a quiet answer's speech ending. Before endpointing or model output, PCM may advance the speech-end proxy as the candidate continues. The first activity-end signal or model output (including a tool request) freezes the primary start. Later PCM and late transcripts cannot shorten that wait or erase its endpoint. Accepted transcripts still update the separate transcription cross-check.

A provider ACTIVITY_START can reopen speech before generation starts. During generation, an activity-start observation is retained for the next measurement until the old model turn ends. The original voice-activity event order and interruption policy are preserved. A raw level crossing is never enough to reopen a frozen measurement. Output transcripts received after turn completion cannot create another measurement for the completed reply.

Durations are rounded down to whole milliseconds. A fast tool may report zero milliseconds while still incrementing tool_calls. Tools handled during a pause, thinking hold, or discarded generation are excluded, as are calls after the first reply audio. Board-image preparation, provider inference and event queue delays are not local tool execution.

tool_phase_ms includes tool_exec_ms and tool_response_ms; these fields are not additive. A long tool phase with fast local execution can indicate waiting for the model's continuation. It does not prove which provider or network operation caused the wait. Without a PCM observation, a late transcript fallback can put the start proxy after a tool request. Then pre_tool_ms is unknown, and only the portion of tool work inside the wait is attributed to it.

When all observations are usable, the decomposition is:

wait = turn_close + tool_after_close + other

The example above reads as 3000 = 1500 + 500 + 1000 milliseconds. Work before close is already covered by turn_close_ms and must not be added again. The unclipped execution/send sums and tool_phase_ms describe operations, while the clipped unions describe their contribution to waiting. other_ms includes model continuation, transport, board preparation and event-processing delays; it is not a measurement of provider inference alone. Missing close observations do not become zero or a residual labeled as VAD. Independently rounded fields can differ from the total by up to two milliseconds.

The PCM threshold is a speech proxy: noise or echo can cross it, and quiet speech can fall below it. Transcription receipt can lag the final spoken word. The end signal's receipt includes transport and room-event processing delay; turn_close_ms is an observed endpointing interval, not pure provider VAD processing time. Missing signals remain unknown. Subtracting tool durations from wait_ms does not measure VAD latency. The audio queue's acceptance is also earlier than browser playout, so this is not an end-to-end audible-delay measurement.

Gemini turnComplete closes the model's turn, not the candidate's speech. The automatic detection setting and its silence window are unchanged by this instrumentation. Timing observations never feed model prompts or recovery deadlines. See the Live API reference for the optional activity signal and its separate audio clock.

For a manual check, use a rebuilt binary, speak a complete answer, and compare the next summary with one where the interviewer reads the editor. Check that the latter has tool activity, that later audio chunks add no summaries, and that a cut, pause or replacement leaves the next measurement with fresh timestamps. A naturally slow reply is useful evidence, not a prerequisite for testing the arithmetic with controlled timestamps in tests/unit/livekit/.