AnyIO worker exception due to sync history processor function in TemporalAgent

[M0dEx]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

When a sync history processor function is used in conjunction with a TemporalAgent instance, the function is executed within an AnyIO loop used by PydanticAI, which causes issues due to the event loop being overridden by Temporal (the history processor is executed within the workflow).

This results in an exception being thrown from the worker thread, due to the Temporal SDK not implementing is_closed:

<redacted>  | Exception in thread AnyIO worker thread:
<redacted>  | Traceback (most recent call last):
<redacted>  |   File "/<redacted>/cpython-3.13.9-linux-aarch64-gnu/lib/python3.13/threading.py", line 1043, in _bootstrap_inner
<redacted>  |     self.run()
<redacted>  |     ~~~~~~~~^^
<redacted>  |   File "/<redacted>/site-packages/anyio/_backends/_asyncio.py", line 973, in run
<redacted>  |     if not self.loop.is_closed():
<redacted>  |            ~~~~~~~~~~~~~~~~~~~^^
<redacted>  |   File "/<redacted>/cpython-3.13.9-linux-aarch64-gnu/lib/python3.13/asyncio/events.py", line 251, in is_closed
<redacted>  |     raise NotImplementedError
<redacted>  | NotImplementedError

The workaround is to make the history processor async, which makes it be executed within the workflow, but that can potentially cause issue if the history processor did any non-deterministic actions, as Temporal might raise an error then.

Example Code

def dummy_history_processor(messages: list[ModelMessage]) -> list[ModelMessage]:
    return messages[1:] # some dummy processing


MY_AGENT = TemporalAgent(
    Agent(
        ...
        history_processors=[dummy_history_processor],
    )
)

Python, Pydantic AI & LLM client version

Python: CPython 3.13.9
Pydantic: 1.23.0
LLM client: `anthropic 0.72.0`

[M0dEx]

I now found a warning in the documentation, but would it be possible to detect this and throw an exception/error that actually describes what the issue is? This took us a while to debug, and I do not think we are the only ones.

[DouweM]

@M0dEx Thanks for the report, sorry to hear you lost time debugging this.

would it be possible to detect this and throw an exception/error that actually describes what the issue is?

Yep, TemporalAgent.__init__ could look at agent.history_processors (and agent._instructions: #2805), detect sync callables, and complain.

Alternatively, the run_in_executor util that runs a sync function in a thread could be overridden to check whether we're inside Temporal, and if so run the function in the regular blocking fashion. Or instead of having that function directly know about Temporal, we could have it check a ContextVar that can be set from inside the TemporalAgent._temporal_overrides context manager.

Would you be up for submitting a PR?