Problem

After upgrading from openai-agents==0.9.3 to 0.12.5, streamed runs with parallel tool calls fail with:

KeyError: <Task finished name='None' exception=KeyError(...)>

This only reproduces under asyncio.eager_task_factory (used by Textual TUI framework on Python 3.12+).

Root Cause

In _FunctionToolBatchExecutor._create_tool_task():

task = asyncio.create_task(self._run_single_tool(...))  # ← task starts running here
self.task_states[task] = task_state                      # ← registered AFTER

Under eager_task_factory, create_task() starts executing the coroutine immediately before returning. The task can complete and be processed by _partition_pending_tasks → self.task_states[task] before self.task_states[task] = task_state executes, causing KeyError.

The issue reporter confirmed this with a minimal reproduction showing the scheduling problem on Python 3.13.

Fix

Wrap _run_single_tool in an inner async def so that eager_task_factory runs the wrapper's synchronous preamble (which is empty) and yields at the first await, returning control to _create_tool_task to finish task_states registration.

This is preferred over "register before create_task" because the task object (needed as the dict key) is only available from create_task()'s return value.

Testing

The fix ensures that when eager_task_factory eagerly executes the wrapper:

1. The wrapper's synchronous body runs (nothing happens — it's empty)
2. The wrapper hits await self._run_single_tool(...) and suspends
3. Control returns to _create_tool_task
4. self.task_states[task] = task_state executes — state is now registered
5. When the event loop resumes the wrapper, _run_single_tool runs safely