Fix JSON logging for asyncio exceptions

[neubig]

Problem

When JSON_LOGS=true is set, asyncio exceptions (like "Task exception was never retrieved") were being logged as plain text with each line of the traceback as a separate log entry. This made it difficult to parse and analyze errors in Datadog.

Example of the problem in Datadog:

{"message": "Task exception was never retrieved", "attributes": {"environment": "production"}}
{"message": "Traceback (most recent call last):", "attributes": {"environment": "production"}}
{"message": "  File \"/usr/lib/python3.13/site-packages/litellm/proxy/hooks/key_management_event_hooks.py\", line 48,", "attributes": {"environment": "production"}}
...

Notice that:

1. Each line of the traceback is a separate log entry
2. There's no level or timestamp attributes (indicating JSON formatter is not being used)
3. There's no stacktrace field

Root Cause

The asyncio exception handler was being set on the wrong event loop at module import time. When litellm/_logging.py is imported, it calls asyncio.get_event_loop().set_exception_handler(). However, at import time, the event loop may not be the one that uvicorn will use. When uvicorn creates a new event loop, the exception handler is not set on it.

Additionally, the original async_json_exception_handler was setting exc_info=None, which meant the traceback was not being included in the JSON output.

Solution

1. Add _setup_asyncio_json_exception_handler() function - This function sets the asyncio exception handler on the running event loop. It should be called AFTER the event loop is created by uvicorn.

2. Call the function in proxy_startup_event - This ensures the handler is set on the correct event loop that uvicorn uses.

3. Include full traceback in exc_info - Extract the traceback from the exception object using exception.__traceback__ and pass it to the LogRecord.

4. Handle non-exception errors - Log them as JSON too instead of falling back to the default handler.

Expected Result

After this fix, asyncio exceptions will be logged as single JSON objects with the stacktrace field:

{
  "message": "Task exception was never retrieved",
  "level": "ERROR",
  "timestamp": "2026-01-02T17:57:31.144636",
  "stacktrace": "Traceback (most recent call last):\n  File \"/tmp/test.py\", line 68, in failing_task\n    raise RuntimeError(\"Async task failed!\")\nRuntimeError: Async task failed!"
}

Testing

Verified with a standalone test that:

1. The JSON formatter correctly includes stacktrace in a single JSON object
2. The asyncio exception handler is set on the running event loop
3. Exceptions from fire-and-forget tasks are logged with full traceback in JSON format

@neubig can click here to continue refining the PR

[openhands-ai]

Looks like there are a few issues preventing this PR from being merged!

• GitHub Actions are failing:
  • LiteLLM Mock Tests (folder - tests/litellm)
  • LiteLLM Linting

If you'd like me to help, just leave a comment, like

@OpenHands please fix the failing actions on PR #3 at branch `fix-json-logging-asyncio`

Feel free to include any additional details that might help me get this PR into a better state.

_{^{You can manage your notification settings}}