fix: remove race condition in LC integration tests

.github/workflows/ci_python.yml

@@ -137,7 +137,7 @@ jobs:
         run: just --set ci true --set output_dir "${{ github.workspace }}" test-python
 
       - name: Run Python LangChain integration tests
-        if: ${{ inputs.run_integration_langchain == 'true' }}
+        if: ${{ inputs.run_integration_langchain }}
         working-directory: ${{ env.NEMO_RELAY_CI_WORKSPACE }}
         run: just --set ci true --set output_dir "${{ github.workspace }}" test-python-langchain
 

python/tests/integrations/deepagents_tests/test_deepagents_integration.py

@@ -75,6 +75,7 @@ def test_before_agent_emits_configuration_mark(
     with nemo_relay.scope.scope("request", nemo_relay.ScopeType.Agent):
         middleware.before_agent(MagicMock(name="mock_state"), MagicMock(name="mock_runtime"))
 
+    nemo_relay.subscribers.flush()
     marks = _filter_mark_events(subscribed_events)
     assert [mark.name for mark in marks] == ["DeepAgents Skills Configured"]
     assert _mark_metadata(marks[0])["deepagents_kind"] == "skill"
@@ -121,6 +122,7 @@ def test_callback_handler_emits_human_in_the_loop_marks(
             )
         )
 
+    nemo_relay.subscribers.flush()
     marks = _filter_mark_events(subscribed_events)
     assert [mark.name for mark in marks] == [
         "DeepAgents Human In The Loop Interrupt",
@@ -159,6 +161,7 @@ def test_callback_handler_falls_back_for_non_hitl_interrupt(
             )
         )
 
+    nemo_relay.subscribers.flush()
     marks = _filter_mark_events(subscribed_events)
     assert [mark.name for mark in marks] == ["Graph Interrupt", "Graph Resume"]
     assert _mark_metadata(marks[0])["integration"] == "langgraph"
@@ -255,6 +258,7 @@ def test_e2e_agent(
     with nemo_relay.scope.scope("deepagents-request", nemo_relay.ScopeType.Agent):
         result = agent.invoke({"messages": [{"role": "user", "content": "Create a file named turtle."}]})
 
+    nemo_relay.subscribers.flush()
     assert (tmp_path / "turtle").read_text() == "shell"
     assert result["messages"][-1].content == "created turtle after reviewer verified turtle"
     found_write_file_message = False

python/tests/integrations/langchain_tests/test_middleware.py

@@ -404,10 +404,12 @@ def event_recorder(event):
             else:
                 result = agent.invoke(input_payload)
     finally:
+        nemo_relay.subscribers.flush()
         nemo_relay.subscribers.deregister("event_recorder")
 
     assert any(
         message.content == "The weather in San Francisco is sunny and 72 degrees." for message in result["messages"]
     )
     assert result["messages"][-1].content == _DEFAULT_MOCK_RESPONSE_MSG
+
     assert events == expected_events

python/tests/integrations/langgraph_tests/test_langgraph_integration.py

@@ -104,6 +104,8 @@ def test_sync(
         with nemo_relay.scope.scope("request", nemo_relay.ScopeType.Agent):
             result = sync_graph.invoke({"value": 1}, config={"callbacks": [callback_handler]})
 
+        nemo_relay.subscribers.flush()
+
         assert result == {"value": 2}
         assert _events_to_strings(subscribed_events) == self._expected_events
 
@@ -116,6 +118,8 @@ async def test_async(
         with nemo_relay.scope.scope("request", nemo_relay.ScopeType.Agent):
             result = await async_graph.ainvoke({"value": 1}, config={"callbacks": [callback_handler]})
 
+        nemo_relay.subscribers.flush()
+
         assert result == {"value": 2}
         assert _events_to_strings(subscribed_events) == self._expected_events
 
@@ -156,6 +160,7 @@ def test_graph_lifecycle_callbacks_emit_marks(
             )
         )
 
+    nemo_relay.subscribers.flush()
     assert _events_to_strings(subscribed_events) == expected_event_strings
 
     interrupt_event = subscribed_events[1]