Python AgentRuntimeClient can create sessions but cannot delete them

[nabrahma]

What would you like to be added:

I propose adding explicit server-side session deletion to AgentRuntimeClient
in the Python SDK to match the lifecycle behavior CodeInterpreterClient already
has. This includes:

1. Extending ControlPlaneClient (sdk-python/agentcube/clients/control_plane.py)
   to support DELETE /v1/agent-runtime/sessions/:sessionId — either by adding a
   dedicated delete_agent_runtime_session(session_id) method or by making the
   existing delete_session method accept the resource kind. The existing
   delete_session path used by CodeInterpreterClient will remain unchanged.

2. Adding optional workload_manager_url and auth_token parameters to
   AgentRuntimeClient.__init__ (sdk-python/agentcube/agent_runtime.py) with
   the same env-var fallback pattern used by CodeInterpreterClient, so existing
   callers that do not need explicit delete are not broken.

3. Adding a stop() method to AgentRuntimeClient that:

   • Closes the local data-plane connection.
   • Calls the WorkloadManager delete API for sessions the client created itself.
   • Does NOT auto-delete sessions that were attached by passing an existing
     session_id (to preserve long-running session reuse).

4. Updating AgentRuntimeClient.__exit__ to call stop() instead of close().

5. Updating sdk-python/examples/agent_runtime_usage.py to demonstrate the full
   create → invoke → stop lifecycle.

6. Adding tests covering:

   • Owned session: stop() calls the delete endpoint.
   • Attached session: stop() does not auto-delete.
   • Context manager: stop() is called on exit for owned sessions.

The backend route DELETE /v1/agent-runtime/sessions/:sessionId already exists
in pkg/workloadmanager/server.go and uses the shared handleDeleteSandbox
handler. No Go backend changes are needed.

Why is this needed:

Currently AgentRuntimeClient.close() only closes the local HTTP connection —
no WorkloadManager API call is made and the server-side sandbox is never released
through the SDK. CodeInterpreterClient already has this full lifecycle via
stop() which calls DELETE /v1/code-interpreter/sessions/:sessionId.

As agent itself may be long-running (issue #303 is removing the default 8h max
TTL for AgentRuntime), SDK users will have no path to release server-side
resources. The sandbox keeps running until idle-timeout fires or someone manually
cleans up the cluster. CodeInterpreterClient and AgentRuntimeClient should
have a consistent lifecycle contract.

This is a Python SDK-only change. No Go code, no CRD regeneration, no generated
client changes are required.

[nabrahma]

@hzxuzhonghu
Please let me know if this is the right direction. If yes, I am happy to open a PR.
I traced the relevant SDK and backend code and confirmed this gap exists in main.

/assign

[vivek41-glitch]

Hi @nabrahma, this is a great proposal. I'd like to help if needed - I have Python SDK experience. Let me know if you need assistance with the PR.

[vivek41-glitch]

/assign