The singleton event bus fans out LLMStreamChunkEvent to all registered
handlers. When multiple streaming runs execute concurrently, each run's
handler receives chunks from all other runs, causing cross-run chunk
contamination.
Fix:
- Add run_id field to LLMStreamChunkEvent
- Use contextvars.ContextVar to track the current streaming run_id
- create_streaming_state() generates a unique run_id per run and sets it
in the context var
- LLM emit paths (base_llm.py, llm.py) stamp run_id on emitted events
- Stream handler filters events by matching run_id
Tests:
- Handler rejects events from different run_id
- Two concurrent streaming states receive only their own events
- Two concurrent threads with isolated contextvars receive only their
own chunks
- run_id field defaults to None for backward compatibility
Co-Authored-By: João <joao@crewai.com>
5511f0a089