From d2f249dc9ab36ab3029f27547d83d635dafc2e9d Mon Sep 17 00:00:00 2001
From: GordonWei <GordonWei@users.noreply.github.com>
Date: Sun, 1 Nov 2026 23:57:39 +0800
Subject: [PATCH] fix(a2a): bypass response cache for Agent Engine
 session-backed calls

The reply depends on server-side session history, but the response cache
key ignores user_id/session_id, so a cached reply could be served to another
key or to a later turn of the same conversation.
---
 litellm/llms/vertex_ai/agent_engine/a2a.py      |  9 ++++++++-
 .../agent_engine/test_agent_engine_a2a.py       | 17 +++++++++++++++++
 2 files changed, 25 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/vertex_ai/agent_engine/a2a.py b/litellm/llms/vertex_ai/agent_engine/a2a.py
index 4a51735648d..51e8eacc113 100644
--- a/litellm/llms/vertex_ai/agent_engine/a2a.py
+++ b/litellm/llms/vertex_ai/agent_engine/a2a.py
@@ -159,4 +159,11 @@ async def merge_a2a_session_into_litellm_params(
     existing_session: Final = _str_param(litellm_params, "session_id")
     session_id: Final = existing_session or await lookup_latest_session_id(litellm_params, user_id)
     session_update: Final[Mapping[str, object]] = {"session_id": session_id} if session_id else {}
-    return {**litellm_params, "user_id": user_id, **session_update}
+    # The reply depends on server-side session history, but the response cache key ignores
+    # user_id/session_id, so a cached reply could be served to another key or a later turn.
+    return {
+        **litellm_params,
+        "user_id": user_id,
+        **session_update,
+        "cache": {"no-cache": True, "no-store": True},
+    }
diff --git a/tests/unit/llms/vertex_ai/agent_engine/test_agent_engine_a2a.py b/tests/unit/llms/vertex_ai/agent_engine/test_agent_engine_a2a.py
index 7414c55be45..da47923e478 100644
--- a/tests/unit/llms/vertex_ai/agent_engine/test_agent_engine_a2a.py
+++ b/tests/unit/llms/vertex_ai/agent_engine/test_agent_engine_a2a.py
@@ -192,6 +192,23 @@ class TestMergeSessionIntoLitellmParams:
             await merge_a2a_session_into_litellm_params(LITELLM_PARAMS, _params(), "hash-a")
         assert LITELLM_PARAMS == original
 
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("sessions", [[], None])
+    async def test_session_backed_call_bypasses_response_cache(self, sessions):
+        user_id = derive_agent_engine_user_id("ctx-1", "hash-a")
+        existing = [{"name": f"{ENGINE_PATH}/sessions/777", "userId": user_id, "updateTime": "2026-01-01T00:00:00Z"}]
+        token_patch, client_patch, _ = _mock_sessions_api(existing if sessions is None else sessions)
+        with token_patch, client_patch:
+            merged = await merge_a2a_session_into_litellm_params(LITELLM_PARAMS, _params(), "hash-a")
+        assert merged["cache"] == {"no-cache": True, "no-store": True}
+
+    @pytest.mark.asyncio
+    async def test_noop_path_leaves_cache_settings_alone(self):
+        token_patch, client_patch, _ = _mock_sessions_api([])
+        with token_patch, client_patch:
+            merged = await merge_a2a_session_into_litellm_params(LITELLM_PARAMS, _params(None), "hash-a")
+        assert "cache" not in merged
+
 
 class TestVertexAgentEngineA2AConfig:
     def test_registered_for_agent_engine_models(self):