Bug [langgraph_checkpoint_aws] - agentcore serializer miss class serialization in writes event #712
Description
The helper for serialize_event(), in the case the event is WritesEvent does not mantain the original class, serializing it as dictionary.
ex. event:
WritesEvent(
event_type='writes',
checkpoint_id='1f0ac518-084c-6d84-8001-d72be2d5493b',
writes=[
WriteItem(
task_id='d3c2cbaa-5b30-0fae-94c3-f8a41da849d4',
channel='__interrupt__',
value=(Interrupt(
value="Trying to call tool xxxx. Please approve or suggest edits.",
id='0d1508ae263740a8ffc819cbe41e36ff'
),),
task_path='~__pregel_push, 0000000000, 0000000000'
)
]
)when serialized and deserialized, the value gets converted to dictionary:
WritesEvent(
event_type="writes",
checkpoint_id="1f0ac518-084c-6d84-8001-d72be2d5493b",
writes=[
WriteItem(
task_id="d3c2cbaa-5b30-0fae-94c3-f8a41da849d4",
channel="__interrupt__",
value=[
{
"value": "Trying to call tool xxxx. Please approve or suggest edits.",
"id": "0d1508ae263740a8ffc819cbe41e36ff",
}
],
task_path="~__pregel_push, 0000000000, 0000000000",
)
],
)This creates an issue when ad example an Interrupt gets deserialized as dictionary and langgraph try to access a class property like ".id"
proposed solution in serialize_event() from
elif isinstance(event, WritesEvent):
# The writes field is already properly serialized by model_dump()
# We just need to serialize the value field in each write
for write in event_dict["writes"]:
val = write.get("value", EMPTY_CHANNEL_VALUE)
write["value"] = self.serialize_value(val)to
elif isinstance(event, WritesEvent):
event_dict["writes"] = [
{
**write.model_dump(exclude_none=True),
"value": self.serialize_value(write.value)
}
for write in event.writes
]