fix(voice): clamp the user_speaking span end to its start

[rosetta-livekit-bot]

Summary

• clamp a backdated user_speaking span end to the recorded span start
• preserve valid positive span durations
• add a patch changeset for @livekit/agents

Ports livekit/agents#7332. Fixes livekit/agents#7307.

Testing

• pnpm test agents/src/voice/audio_recognition_span.test.ts (11 passed)
• pnpm --filter @livekit/agents typecheck
• pnpm --filter @livekit/agents build
• pnpm build (40 workspace packages built)
• pnpm --filter @livekit/agents lint (passes with existing warnings)
• pnpm exec prettier --check "agents/src/**/*.{ts,tsx,md,json}"
• pnpm test agents (2,721 passed, 5 skipped; 3 unrelated agents/src/telemetry/loop_monitor.test.ts failures because this host falls back to process CPU accounting while those tests require thread CPU accounting)
• cue-cli voice validation attempted; the configured LiveKit endpoint rejected both the worker and driven turn with 401 Unauthorized - invalid API key, so framework-event validation could not complete

Source diff coverage

Classification of every file in livekit/agents#7332

Source file	Classification	Target coverage
livekit-agents/livekit/agents/voice/agent_session.py	Adapted	agents/src/voice/agent_session.ts: stores the explicit millisecond start anchor alongside the span and clamps the end against it, keeping the behavior independent of the tracer provider's span implementation.
tests/test_user_speaking_span.py	Adapted	agents/src/voice/audio_recognition_span.test.ts: ports both source cases into the existing JS span-test counterpart, covering a 550 ms backdated end and an unchanged 1,250 ms positive duration.
tests/trace_schema.py	Not applicable	This is the Python repository's offline OTLP trace-schema validator. agents-js has no equivalent trace-schema validator or span-record model, and runtime behavior does not depend on it.
tests/test_trace_schema.py	Not applicable	Its only source change tests the Python-only offline validator above; there is no target counterpart.

No source runtime behavior depends on the Python-only validator omitted above.

---

Ported from livekit/agents#7332

Original PR description

Problem: The user_speaking span starts on the STT turn anchor and ends on a VAD anchor that is backdated by silence_duration + inference_duration. The two detectors do not share a clock, so on a short burst the end lands before the start, and OpenTelemetry keeps the duration unsigned — the reporter's backend read 18446744073500970us.

Fix: The end is clamped to the span start, so the span is zero-length instead of negative. check_trace in tests/trace_schema.py now reports any span that ends before it starts.

Fixes #7307.

Context for reviewing and coding agents

How to see it

tests/test_user_speaking_span.py::test_backdated_end_is_clamped_to_the_start fails on main. It opens the span on one anchor, closes it on an anchor 0.55 s earlier, and asserts end_time == start_time. The second case in that file closes 1.25 s later and asserts the duration is kept, so the clamp cannot swallow a real one.

Why the end is clamped and not re-anchored

The backdating is deliberate and feeds the end-of-turn delay metrics, so both anchors must stay. The VAD and the STT genuinely disagree here: the VAD places the end of speech before the STT places the start, and a zero-length span is the honest record of that. _end_eou_wait_span already resolves the same conflict the same way at audio_recognition.py:2093.

Blast radius

user_speaking was the only span that ends on an anchor a different detector produced. The other explicit ends cannot go negative: audio_recognition.py:2079 and 2104 end at time.time_ns(), 2124 is the clamp that already exists, and telemetry/loop_monitor.py:462 computes its end as start_ns + duration.

Where the numbers came from

The reporter's OTLP export holds one span with start 1789557796406789376 and end 1789557796198206208, a gap of 208.6 ms. Its speaking -> listening log line is stamped 575.8 ms after that end value, which is Silero's default min_silence_duration of 0.55 s plus inference. python -m tests.trace_schema <export> over that file reports that span and none of the other 144.

[changeset-bot]

🦋 Changeset detected

Latest commit: c9835da

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 39 packages

Name	Type
@livekit/agents	Patch
@livekit/agents-plugin-anam	Patch
@livekit/agents-plugin-anthropic	Patch
@livekit/agents-plugin-assemblyai	Patch
@livekit/agents-plugin-azure	Patch
@livekit/agents-plugin-baseten	Patch
@livekit/agents-plugin-bey	Patch
@livekit/agents-plugin-cartesia	Patch
@livekit/agents-plugin-cerebras	Patch
@livekit/agents-plugin-deepgram	Patch
@livekit/agents-plugin-did	Patch
@livekit/agents-plugin-elevenlabs	Patch
@livekit/agents-plugin-fishaudio	Patch
@livekit/agents-plugin-google	Patch
@livekit/agents-plugin-hume	Patch
@livekit/agents-plugin-inworld	Patch
@livekit/agents-plugin-krisp	Patch
@livekit/agents-plugin-lemonslice	Patch
@livekit/agents-plugin-liveavatar	Patch
@livekit/agents-plugin-livekit	Patch
@livekit/agents-plugin-meta	Patch
@livekit/agents-plugin-minimax	Patch
@livekit/agents-plugin-mistral	Patch
@livekit/agents-plugin-mistralai	Patch
@livekit/agents-plugin-neuphonic	Patch
@livekit/agents-plugin-openai	Patch
@livekit/agents-plugin-perplexity	Patch
@livekit/agents-plugin-phonic	Patch
@livekit/agents-plugin-protoface	Patch
@livekit/agents-plugin-resemble	Patch
@livekit/agents-plugin-rime	Patch
@livekit/agents-plugin-runway	Patch
@livekit/agents-plugin-sarvam	Patch
@livekit/agents-plugin-silero	Patch
@livekit/agents-plugin-soniox	Patch
@livekit/agents-plugin-tavus	Patch
@livekit/agents-plugins-test	Patch
@livekit/agents-plugin-trugen	Patch
@livekit/agents-plugin-xai	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR