Auto-checkpoint raises PydanticSerializationError when a Task has a Python-callable guardrail

Summary

With auto-checkpointing enabled, CheckpointListener._do_checkpoint calls state.model_dump(mode="json"), which fails on any Task whose guardrail is a Python callable (a function/lambda, as opposed to a string guardrail). Pydantic can't JSON-serialize the callable, so:

1. A UserWarning: Callable 'guardrail' cannot be JSON-serialized and will be dropped during checkpointing; restored checkpoints will not run this guardrail. fires, then
2. model_dump(mode="json") still raises PydanticSerializationError: Unable to serialize unknown type: <class 'function'> (also seen as <class 'pydantic._internal._model_construction.ModelMetaclass'>), which is re-raised out of _do_checkpoint and logged by the event handler as a full WARNING:crewai.state.checkpoint_listener:Auto-checkpoint failed for event task_completed traceback.

Net effect: every checkpointed event on a guardrail-bearing task emits a traceback, and no usable checkpoint is written for that task.

Version

crewai 1.15.10 (also reproduces on 1.15.21). Python 3.12.

Actual behavior (real traceback)

UserWarning: Callable 'guardrail' cannot be JSON-serialized and will be dropped
during checkpointing; restored checkpoints will not run this guardrail.
WARNING:crewai.state.checkpoint_listener:Auto-checkpoint failed for event task_completed
Traceback (most recent call last):
  File ".../crewai/state/checkpoint_listener.py", line 242, in _on_any_event
    _do_checkpoint(state, cfg, event)
  File ".../crewai/state/checkpoint_listener.py", line 144, in _do_checkpoint
    payload = state.model_dump(mode="json")
pydantic_core._pydantic_core.PydanticSerializationError: Unable to serialize unknown type: <class 'function'>

Steps to reproduce

Configure a Crew with auto-checkpointing where on_events includes task_completed, and give one Task a callable guardrail (e.g. guardrail=my_fn where my_fn(output) -> (bool, Any)). Run kickoff(). Each task_completed triggers the failure above. String guardrails are unaffected.

Expected behavior

The checkpoint serializer should degrade gracefully on non-serializable fields (callable guardrails, and other non-JSON state) rather than raising: drop them from the snapshot with a single, quiet notice (not a per-event traceback), and still write the rest of the checkpoint. Callable guardrails are a first-class, documented feature, so checkpointing a task that uses one shouldn't fail or spam tracebacks.

Root cause

checkpoint_listener.py::_do_checkpoint -> state.model_dump(mode="json") has no default=/exclusion handling for callable/non-serializable fields; the except Exception at ~line 156 emits CheckpointFailedEvent and then raises, so the caller logs the whole traceback.

Workaround (downstream)

Restricting on_events to failure-only events (task_failed, crew_kickoff_failed) avoids the per-successful-task noise, but the underlying serialization gap remains for any checkpointed event on a guardrail-bearing task.

Start in checkpoint_listener.py, especially CheckpointListener._do_checkpoint and the _on_any_event path described in the traceback. Run the reported auto-checkpoint reproduction with a callable Task guardrail and verify that the remaining state is written without a per-event traceback, while the callable field is omitted with only one quiet notice.