CompletionUsage validation crashes session when provider returns usage.completion_tokens=None mid-stream (OpenAI-compatible path)

[onur-yildirim-infinitusai]

Bug Description

When using the OpenAI-compatible LLM path (livekit.plugins.openai.llm.LLM, which inherits its streaming _run loop from livekit.agents.inference.llm.LLMStream), if the upstream provider sends a final streaming chunk where chunk.usage is a non-null object but one of its fields (e.g. completion_tokens) is null, the SDK crashes.

The code at livekit-agents/livekit/agents/inference/llm.py only guards on if chunk.usage is not None: before constructing:

usage=llm.CompletionUsage(
    completion_tokens=chunk.usage.completion_tokens,
    prompt_tokens=chunk.usage.prompt_tokens,
    total_tokens=chunk.usage.total_tokens,
    ...
)

CompletionUsage.completion_tokens is a required int (no Optional), so a null value raises a pydantic ValidationError, which gets wrapped and re-raised as APIConnectionError(retryable=False). Because it's marked non-retryable, this eventually surfaces as an unrecoverable llm_error that terminates the whole AgentSession — killing an in-progress voice call — even though the actual issue is a malformed/partial usage payload, not a real connection failure.

We're hitting this against an OpenAI-compatible proxy in front of a custom model; the proxy occasionally omits completion_tokens in the usage object on stream completion. This is analogous to #5404, which fixed a similar "usage present but incomplete on final chunk" issue for the Google/Gemini plugin — but the equivalent guard was never added to the shared OpenAI-compatible streaming path.

Expected Behavior

A null/missing field inside chunk.usage should not raise an unhandled, non-retryable exception that ends the session. At minimum, missing numeric usage fields should default to 0 (or the chunk should be treated as retryable) rather than crashing.

Reproduction Steps

1. Configure livekit.plugins.openai.LLM against an OpenAI-compatible endpoint (e.g. a proxy in front of a non-OpenAI model).
2. Have that endpoint return a streamed response whose final SSE chunk includes a usage object with completion_tokens: null (rather than omitting usage entirely).
3. Observe the ValidationError for CompletionUsage, wrapped into APIConnectionError(retryable=False), terminating the session.

Operating System

Linux (GKE)

Models Used

OpenAI-compatible proxy in front of a custom-hosted model

Package Versions

livekit==1.1.13
livekit-agents==1.6.5
livekit-plugins-openai==1.6.5

Proposed Solution

if chunk.usage is not None:
    usage = chunk.usage
    usage_chunk = llm.ChatChunk(
        id=chunk.id,
        usage=llm.CompletionUsage(
            completion_tokens=usage.completion_tokens or 0,
            prompt_tokens=usage.prompt_tokens or 0,
            prompt_cached_tokens=cached_tokens or 0,
            total_tokens=usage.total_tokens or 0,
            service_tier=getattr(chunk, "service_tier", None),
        ),
    )
    self._event_ch.send_nowait(usage_chunk)

Additional Context

Traceback:

Traceback (most recent call last):
  File ".../livekit/agents/inference/llm.py", line 445, in _run
    usage=llm.CompletionUsage(
  File ".../pydantic/main.py", line 250, in __init__
    validated_self = self.__pydantic_validator__.validate_python(data, self_instance=self)
pydantic_core._pydantic_core.ValidationError: 1 validation error for CompletionUsage
completion_tokens
  Input should be a valid integer [type=int_type, input_value=None, input_type=NoneType]

livekit.agents._exceptions.APIConnectionError: Connection error.