agents

Upgrade to a LiveKit Agents release containing PR #7580, then start the worker with LIVEKIT_AGENTS_PRELOAD_EOT=0. Accepted disabling values are 0, false, no, and off.

Temporary shell start:

export LIVEKIT_AGENTS_PRELOAD_EOT=0
python your_worker.py

For systemd workers, add:

[Service]
Environment=LIVEKIT_AGENTS_PRELOAD_EOT=0

This skips the ~240 MB end-of-turn model preload while keeping the lightweight VAD preload. If a local TurnDetector is actually used later, its weights are loaded lazily on first use. Under Linux forkserver, EOT preloading remains enabled by default. If multiple workers run on one host, using one AgentServer with a larger num_idle_processes can also share a single copy instead of loading it once per worker.

Upgrade livekit-agents to **1.8.3 or newer**. This shutdown race was fixed by PR #7225; versions including 1.8.1 can hang when idle prewarmed THREAD runners exit before their shutdown acknowledgement is read.

pip install --upgrade 'livekit-agents>=1.8.3'

If using uv:

uv add 'livekit-agents>=1.8.3'

Then restart the worker and send SIGTERM/Ctrl+C after the idle runners are initialized. The worker should shut down promptly without repeated job did not ack shutdown in time errors.

If you cannot upgrade immediately, use JobExecutorType.PROCESS, which does not exhibit this hang. Sending a second signal only force-terminates the process and is not a graceful workaround.

This occurs in affected versions such as livekit-agents==1.8.4: _ProcClient.run() creates an asyncio event loop but does not cancel pending tasks, shut down async generators, reset the thread loop, or close the loop.

**Preferred fix:** upgrade livekit-agents to a release containing the PR fix(ipc): close the job loop when the runner exits (fixes issue #7629).

If you cannot upgrade immediately, patch the finally block in _ProcClient.run() to perform the same shutdown sequence as asyncio.run():

finally:
    loop_monitor.stop_monitoring(loop)
    try:
        pending = [t for t in asyncio.all_tasks(loop) if not t.done()]
        for task in pending:
            task.cancel()
        if pending:
            loop.run_until_complete(
                asyncio.gather(*pending, return_exceptions=True)
            )
        loop.run_until_complete(loop.shutdown_asyncgens())
        loop.run_until_complete(loop.shutdown_default_executor())
    finally:
        asyncio.set_event_loop(None)
        loop.close()

As a temporary workaround, use process isolation instead of threads:

server = AgentServer(job_executor_type=JobExecutorType.PROCESS)

After the fix, completed job loops should report is_closed() == True, and Python should no longer emit ResourceWarning: unclosed event loop.

Root cause: In SpeechStream._send_audio (livekit-plugins-meta), the pacing clock (pacing_origin) starts when the first packet is *sent*, not when the stream starts. Since _run awaits _connect_ws() first, any audio frames queued while the Muse websocket is connecting create a permanent offset — every packet for the rest of the stream is sent late by the connect duration (e.g., 300 ms connect time ≈ 221 ms constant delay), because pacing never sends faster than real time and the backlog is never caught up. This delays transcripts and end-of-turn events by the same amount.

Fix (PR #7890): Start the pacing clock when the stream starts running instead of at the first send. The buffered backlog is flushed immediately once the socket is ready, and subsequent packets are paced against real time as before.

Workaround until the patched release: monkey-patch _send_audio to skip the pacing sleep while _input_ch still has queued frames (only sleep once the buffer is drained). Confirmed affected: livekit-agents==1.8.6, livekit-plugins-meta==1.8.6 with muse-voice-transcribe-1.0.

Affected versions: livekit-agents 1.8.3 and main @ 57b3227. _transcripts_equivalent() calls split_words(..., ignore_punctuation=True), but the punctuation removal list in livekit/agents/tokenize/_basic_word.py is ASCII-only. Therefore Arabic ، ؛ ؟, Urdu ۔, Devanagari ।, and fullwidth CJK ，。 remain attached to words, causing the final transcript to appear different and triggering preemptive generation invalidated ... because the transcript ... changed.

Fix the tokenizer to remove all Unicode punctuation categories (P*) in addition to the existing ASCII punctuation:

import unicodedata

def _is_punctuation(char: str) -> bool:
    return (
        char in tokenize.tokenizer.PUNCTUATIONS
        or unicodedata.category(char).startswith("P")
    )

Use this check in _basic_word.py wherever ignore_punctuation=True filters characters. This preserves symbols such as $ and +, which have category S, and Arabic diacritics, which have category Mn.

Verify with:

from livekit.agents.voice.agent_activity import _transcripts_equivalent

assert _transcripts_equivalent("مرحبا كيف حالك", "مرحبا، كيف حالك؟")
assert _transcripts_equivalent("你好 我很好", "你好，我很好。")
assert _transcripts_equivalent("आप कैसे हैं", "आप कैसे हैं।")

Rebuild or reinstall the patched package. Once the upstream fix is released, upgrade with pip install -U livekit-agents.