Bug: Unable to create session for Agent Engine deployment when deploying with event compaction/context caching

[juszzz]

Describe the bug
When trying to create a session after deploying an agent with event compaction and/or context caching (which requires creating an App that gets passed to AdkApp), session creation fails due to an invalid app_name. This seems to occur because Pydantic expects you to supply a name when creating an App (note that the problem will occur even if you set name=None), however this name is not a valid reasoning engine resource name (e.g., projects/1234567/locations/us-central1/reasoningEngines/456789) and so attempts to create a session fail.

To Reproduce

1. Deploy an agent to Agent Engine using an App to configure event compaction and/or context caching:

   app = google.adk.apps.App(
       root_agent=my_agent,
       name="my_agent_name",
       events_compaction_config=EventsCompactionConfig(),
       context_cache_config=ContextCacheConfig(),
   )
   
   adk_app = vertexai.agent_engines.AdkApp(
       app=app,
       enable_tracing=False,
   )
   
   vertexai.agent_engines.create(
       agent_engine=adk_app,
       ...
   )
   

2. Attempt to create a session:

   remote_agent: vertexai.agent_engines.AgentEngine = agent_engines.get(resource_name="projects/12345/...")
   remote_session = await remote_agent.async_create_session(user_id="my_user_id")
   

Actual behavior
The async_create_session() call fails with the error message:

Error Details: {"detail":"Agent Engine Error: An error occurred during invocation. Exception: App name my_agent_name is not valid. It should either be the full ReasoningEngine resource name, or the reasoning engine id.\nRequest Data: {'user_id': 'my_user_id'}"}"

Expected behavior
A session is successfully created, with Agent Engine supplying the necessary resource name as-needed in order for session creation to occur.

Desktop (please complete the following information):

• OS: macOS
• Python version(python -V): 3.11
• ADK version(pip show google-adk): 1.18.0

Model Information:

• Are you using LiteLLM: No
• Which model is being used: N/A

Additional context
Add any other context about the problem here.

[juszzz]

Just to be clear: I don't think the issue has anything to do with event compaction and/or context caching, it's just that creating a google.adk.apps.App is needed in order to configure those things, and passing that App (instead of just the root_agent) to vertexai.agent_engines.AdkApp causes the problem to appear. As far as I'm aware, there is no way to configure event compaction or context caching directly via instantiation of an AdkApp; if I'm wrong, I'd love to know how to do so.

[sarojrout]

I think, it’s caused by how App.name is interpreted once the object is passed through vertexai.agent_engines.AdkApp. ADK validates App.name as a simple identifier (e.g. my_agent_name), but Agent Engine assumes the same field already contains the reasoning-engine resource name (projects/.../reasoningEngines/...) or the engine ID. So when an App is supplied to AdkApp, the deployment stores that friendly name and session creation later rejects it as invalid. At the moment Agent Engine doesn’t consume the extra App-level config, so the recommended workaround is to deploy with AdkApp(root_agent=my_agent, …) and omit the App wrapper

[juszzz]

Thanks for the clarification, @sarojrout. Regarding the workaround: does this mean it's not currently possible to deploy to Agent Engine with context caching or event compaction enabled? I only see a way to configure those options via an App wrapper, i.e., not directly via AdkApp.

[sarojrout]

Unfortunately, event compaction and context caching are currently not supported when deploying to Agent Engine.
These features are handled by ADK's Runner using App-level configs (events_compaction_config and context_cache_config). They require the Runner to process the App and apply these configs during execution.

[sarojrout]

@juszzz , i have raised the pr for the fix. once its merged, you should be able to validate

[juszzz]

@sarojrout thanks for the super quick turnaround on this; happy to test once it's merged!