fix(closes OPEN-8001): handle ValueError in _current_step.reset to manage context variable mismatches

[viniciusdsmello]

Pull Request

Summary

Fix context variable ValueError that occurs when async generators cross context boundaries in multi-agent systems. This issue manifests when using isolated execution contexts (e.g., Google ADK's AgentTool pattern, LangGraph sub-graphs, CrewAI delegation) where cleanup attempts to reset a context variable token in a different context than where it was created.

Changes

• Added try-catch block in tracer.py:405-420 to gracefully handle context variable mismatch during async generator cleanup

Context

Problem: When agent frameworks create isolated execution contexts (AgentTool pattern), async generator cleanup fails with:
ValueError: <Token var= was created in a different ContextRoot Cause:

1. Parent agent creates trace step with context variable token in Context A
2. AgentTool/sub-agent runs in isolated Context B
3. Async generator yields events across contexts
4. Cleanup attempts _current_step.reset(token) in Context B with token from Context A
5. Python's contextvars enforces context boundaries → ValueError

Why This Fix Works:

• Traces are complete and sent before cleanup error occurs
• Child context is destroyed immediately after, making stale references harmless

Testing

• Unit tests (existing tests pass, error only occurs in cleanup)
• Manual testing
  • Created AG-UI FastAPI server with Google ADK + AgentTool pattern
  • Test client sends requests triggering multiple AgentTool invocations
  • Verified no ValueError in logs after fix
  • Confirmed traces sent successfully to Openlayer
• Postman CI/CD (not applicable - internal tracing fix)