Declarative agent A2A: SessionNotFoundError due to user_id mismatch between session-create and session-get

[davidkarlsen]

Summary

Inbound A2A message/send to a Declarative ADK agent fails with
SessionNotFoundError. kagent-adk._agent_executor creates the session
through kagent-controller's API under one user_id, but
google.adk.runners._get_or_create_session then looks the same session
up under a different user_id, so the GET 404s and execution aborts.

Environment

• kagent v0.9.4 (chart + bundled kagent-adk)
• Agent CR: type: Declarative, ADK runtime, A2A enabled
• controller.auth.mode: unsecure

Repro

Any A2A request to a Declarative agent over its operator-managed A2A
Service. Inbound JSON-RPC envelope carries a non-empty context_id.

Trace (sanitized — log lines reformatted, no internal IDs)

step	call	result	user_id used
1	GET /api/sessions/?...	404	A2A_USER_
2	POST /api/tasks	201	(auth-derived)
3	POST /api/sessions	201	(auth-derived)
4	POST /api/tasks	201	(auth-derived)
5	POST /api/sessions//events?user_id=	201	← different
6	GET /api/sessions/?...	404	A2A_USER_ ← mismatch
7	SessionNotFoundError: Session not found:

<CTX> = the inbound A2A request's context_id.
<AUTH> = the user_id injected by the controller's auth layer
(in unsecure mode this is the controller's static fallback identity).

Python traceback terminates in:

google/adk/runners.py:435  _get_or_create_session
  → google.adk.errors.session_not_found_error.SessionNotFoundError

Root-cause hypothesis

• The A2A executor derives a synthetic A2A_USER_<context_id> and uses
  it for session GETs.
• Session creation (POST /api/sessions, POST /api/sessions/<CTX>/events)
  writes under the controller's auth-derived identity instead.
• The two keyspaces don't overlap, so the lookup never finds the session
  it just created.

Expected

A single user_id consistently used for both create and lookup, so the
follow-up GET resolves.

Notes

• The [EXPERIMENTAL] ResumabilityConfig UserWarning printed at startup
  is co-located but benign.
• Reproducible on every A2A invocation; not flaky.

[timfdev]

Same issue here