ValueError when finishing trace in Runner.run_streamed(): “token was created in a different Context”

[sarattha]

When using Runner.run_streamed() in async streaming mode, at the end of the stream the SDK crashes with:

    ValueError: <Token var=<ContextVar name='current_trace' default=None at 0x…> at 0x…> was created in a different Context

This comes from the tracing code’s Scope.reset_current_trace(token) call in Trace.finish(), because the token was set in the parent
task’s context but is being reset in the spawned Task’s context.

To Reproduce

    1. Call `Runner.run_streamed(...)` (e.g. from a FastAPI endpoint) on an agent that emits a streaming response.
    2. Consume it all, letting the stream reach completion.
    3. Observe the exception in the server logs.

Expected behavior
The trace should clean up without raising a ValueError, or tracing should be disabled automatically for streaming if it can’t safely reset in the child Task.

Traceback

    File ".../agents/tracing/scope.py", line 49, in reset_current_trace
        _current_trace.reset(token)
    ValueError: <Token var=<ContextVar name='current_trace' default=None at 0x…> at 0x…> was created in a different Context

Environment

    * openai‑agents‑python version: 0.0.11
    * Python: 3.13.2 on macOS (aarch64)
    * starlette / fastapi / uvicorn versions: 0.34.1

Additional context
run_streamed() does:

    1. `new_trace = trace(...); new_trace.start(mark_as_current=True)`
    2. `asyncio.create_task(_run_streamed_impl(...))`
    3. When the child task ends, `stream_events()` calls `trace.finish(reset_current=True)`, using the token created in the parent
context. Python’s ContextVar API refuses to reset with a token from another context.

Suggested fix
Either:

    * Don’t call `reset_current_trace()` in the worker Task (or catch/reset appropriately),
    * Or capture and re‐apply the correct ContextVar in the spawned Task,
    * Or disable tracing by default in `run_streamed()` if it can’t guarantee same context.