fix(memory): preserve round directives and component route discovery

litellm/proxy/_lazy_features.py

@@ -56,11 +56,6 @@ class LazyFeature:
 
 
 LAZY_FEATURES: Final[tuple[LazyFeature, ...]] = (
-    LazyFeature(
-        name="memory_v2",
-        module_path="litellm.proxy.memory.management",
-        path_prefixes=("/v2/memory",),
-    ),
     LazyFeature(
         name="guardrails",
         module_path="litellm.proxy.guardrails.guardrail_endpoints",

litellm/proxy/memory/gateway.py

@@ -137,8 +137,21 @@ class GatewayMemoryLoop:
 
     async def _call(self) -> AsyncGenerator[bytes, None]:
         self.stream.begin_round()
+        # Claude output directives control the next generated turn. Repeat them
+        # on outgoing rounds without adding pending directives to saved history.
+        directives: Final = (
+            transcript_items(self.original, self.route)[-self.replaced_input :] if self.replaced_input else ()
+        )
+        messages: Final = transcript_items(self.data, self.route)
         body: Final = {  # mutable-ok: Native provider JSON containers.
             **self.data,
+            **(
+                {  # mutable-ok: Native provider JSON containers.
+                    "messages": [*messages, *directives],  # mutable-ok: Provider request JSON.
+                }
+                if directives and messages[-len(directives) :] != directives
+                else {}  # mutable-ok: Native provider JSON containers.
+            ),
             "cache": {  # mutable-ok: Native provider JSON containers.
                 **object_value(self.data.get("cache")),
                 "no-cache": True,

litellm/proxy/proxy_server.py

@@ -18802,6 +18802,9 @@ app.include_router(auto_router_management_router)
 app.include_router(tag_management_router)
 app.include_router(workflow_management_router)
 app.include_router(memory_router)
+from litellm.proxy.memory.management import router as memory_v2_router
+
+app.include_router(memory_v2_router)
 app.include_router(plugin_router)
 app.include_router(cost_tracking_settings_router)
 app.include_router(router_settings_router)

tests/test_litellm/proxy/memory/test_memory_v2_boundaries.py

@@ -579,6 +579,44 @@ async def test_trailing_system_messages_survive_client_tool_continuation(prisma_
     assert original["messages"] == [prefix, directive]
 
 
+@pytest.mark.asyncio
+async def test_claude_output_directives_reach_search_answer_and_reflection_rounds(prisma_edge: MagicMock) -> None:
+    provider = FastAPI()
+    observed = []
+    directive = {"role": "system", "content": [], "output_config": {"effort": "low"}}
+
+    @provider.post("/v1/messages")
+    async def model(incoming: Request):
+        body = await incoming.json()
+        observed.append(body)
+        assert body["messages"][-1] == directive
+        if len(observed) == 1:
+            content = [
+                {"type": "tool_use", "id": "search", "name": "litellm_memory_search", "input": {"query": "demo"}}
+            ]
+        elif len(observed) == 2:
+            content = [{"type": "text", "text": "The port is 8347"}]
+        else:
+            content = [
+                {"type": "tool_use", "id": "reflect", "name": "litellm_memory_capture", "input": {"observations": []}}
+            ]
+        return {
+            "id": "msg_" + str(len(observed)),
+            "stop_reason": "end_turn" if len(observed) == 2 else "tool_use",
+            "content": content,
+        }
+
+    original = {"messages": [{"role": "user", "content": "My demo port?"}, directive]}
+    loop = GatewayMemoryLoop(provider, request(), original, "anthropic_messages", store(prisma_edge))
+    async for _ in loop.run():
+        pass
+    assert len(observed) == 3
+    assert observed[1]["messages"][-2]["content"][0]["tool_use_id"] == "search"
+    assert "reflect once" in observed[2]["messages"][-2]["content"]
+    assert observed[0]["messages"][0] == original["messages"][0]
+    assert sum(message["role"] == "system" for message in object_items(loop.data.get("messages"))) == 1
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("bad_id,count", [(True, 1), (False, 17)])
 async def test_invalid_model_calls_are_rejected_before_storage(