InvalidEventBytes exception should be caught and handled as recoverable

[S-W-Leong]

Bug Description

When using aws.realtime.RealtimeModel (Nova Sonic), the InvalidEventBytes exception from smithy_aws_event_stream is treated as an unrecoverable error, causing the entire AgentSession to terminate. This exception should be handled as recoverable and trigger a session restart, similar to how ModelStreamErrorException and ModelTimeoutException are handled.

Expected Behavior

The InvalidEventBytes exception should be caught and handled as a recoverable error, triggering _restart_session() to attempt recovery, similar to how other stream errors are handled.

Reproduction Steps

Create an agent using aws.realtime.RealtimeModel with Nova Sonic
Start a conversation session
Wait for the InvalidEventBytes error to occur (typically happens during barge-in events or network interruptions)

Operating System

macOS

Models Used

Nova 2 Sonic (AWS)

Package Versions

livekit-agents: 1.3.11
livekit-plugins-aws: 1.3.11
smithy-aws-event-stream: 0.2.1
aws_sdk_bedrock_runtime: 0.3.0
Python: 3.13

Session/Room/Call IDs

No response

Proposed Solution

from smithy_aws_event_stream.exceptions import InvalidEventBytes

# In _process_responses():
except (
    ThrottlingException,
    ModelNotReadyException,
    ModelErrorException,
    ModelStreamErrorException,
    InvalidEventBytes,  # Add this
) as re:
    logger.warning(f"Retryable error: {re}\nAttempting to recover...")
    await self._restart_session(re)
    break

Additional Context

The InvalidEventBytes error typically occurs when:

The user interrupts the model (barge-in)
Network connection is interrupted mid-stream
The bidirectional stream is closed unexpectedly
The log message "Nova Sonic handles interruption automatically via barge-in detection" appears right after the error, suggesting this may be related to barge-in handling where the stream is intentionally interrupted but the exception isn't being handled gracefully.

Workaround (temp)

As a temporary workaround, I've created a monkey-patch that intercepts InvalidEventBytes at the EventReceiver.receive() level and converts it to ModelStreamErrorException:

from smithy_aws_event_stream.exceptions import InvalidEventBytes
from aws_sdk_bedrock_runtime.models import ModelStreamErrorException
from smithy_aws_event_stream.aio import EventReceiver

original_receive = EventReceiver.receive

async def patched_receive(self):
    try:
        return await original_receive(self)
    except InvalidEventBytes as e:
        raise ModelStreamErrorException(
            message=f"Stream error (InvalidEventBytes): {e}",
            fault="server",
            is_retry_safe=True,
        ) from e

EventReceiver.receive = patched_receive

Screenshots and Recordings

No response

[chenghao-mou]

I could not reproduce the issue with interruptions, even after 10+ consecutive barge-ins. Could you share your model setup?

[S-W-Leong]

thanks for trying to reproduce. heres my setup

session = AgentSession(
    llm=aws.realtime.RealtimeModel.with_nova_sonic_2(
        voice="tiffany",
        region="us-east-1",  # Cross-region call from ap-southeast-1
        tool_choice="auto",  # Enable automatic tool calling
    ),
    allow_interruptions=True,  # Required for server-side turn detection
)

[chenghao-mou]

I saw this issue in only one out of the many tests I did. So most likely a random server-side issue. Will create a PR to catch it in agents.