[yuting0624]

Summary

Fixes #4930

VertexAiSessionService.append_event() was not including rewind_before_invocation_id in the actions dict sent to the Vertex AI Agent Engine API. This caused rewind events to lose their target invocation ID when persisted, making session rewind non-functional for Agent Engine deployments.

Changes

• src/google/adk/sessions/vertex_ai_session_service.py: Added rewind_before_invocation_id to the actions dict in append_event() (3 lines)
• tests/unittests/sessions/test_vertex_ai_session_service.py: Added test_append_event_with_rewind to verify the field round-trips correctly through append and get

Root Cause

The actions dict construction in append_event() explicitly maps each field from EventActions to the API payload, but rewind_before_invocation_id was omitted. Other fields like transfer_to_agent, state_delta, skip_summarization, etc. were all present.

[adk-bot]

Response from ADK Triaging Agent

Hello @yuting0624, thank you for your contribution!

It looks like the pyink-check is failing. You can fix this by running the auto-formatting script from the root of the repository:

./autoformat.sh

This will ensure your changes adhere to the project's coding style. Thanks!

[lucasbarzotto-axonify]

Hi @yuting0624, does this fix works?

I tried it and got the following error:

Error rewinding session 137352779248173056: 1 validation error for _AppendAgentEngineSessionEventRequestParameters

config.actions.rewind_before_invocation_id

Extra inputs are not permitted [type=extra_forbidden, input_value='p-146156f2-8f8e-42ae-af2a-75d00a48130c', input_type=str]

For further information visit https://errors.pydantic.dev/2.12/v/extra_forbidden

Looking at this, it looks like the rewind_before_invocation_id field is not allowed?

Should we use custom_metadata instead? The same way it's done for event compaction?

    # Store compaction data in custom_metadata since the Vertex AI service
    # does not yet support the compaction field.
    # TODO: Stop writing to custom_metadata once the Vertex AI service
    # supports the compaction field natively in EventActions.
    if event.actions and event.actions.compaction:
      compaction_dict = event.actions.compaction.model_dump(
          exclude_none=True, mode='json'
      )
      _set_internal_custom_metadata(
          metadata_dict,
          key=_COMPACTION_CUSTOM_METADATA_KEY,
          value=compaction_dict,
      )

[yuting0624]

Thanks for catching this @lucasbarzotto-axonify! You're absolutely right — the Vertex AI SDK doesn't support rewind_before_invocation_id in EventActions yet.

I've pushed a fix that stores it in custom_metadata following the same pattern as compaction:

• Write path: rewind_before_invocation_id is now stored via _set_internal_custom_metadata() instead of directly in the actions dict
• Read path: Extracted from custom_metadata and restored into EventActions on retrieval

All 25 session service tests pass. The Pydantic validation error should no longer occur.

[DeanChensj]

Thanks for the contribution!

Instead of introducing a new custom piggyback field (or extending the schema with a new custom parameter for this), we should leverage the underlying SDK's existing raw_event (or raw/extra data) field. This keeps the ADK schema cleaner and natively aligns with how the SDK already passes through arbitrary underlying event properties.

Could you update the implementation to map this data using the SDK's raw event field instead? Let us know if you need any pointers on where it's exposed!