fix: Create graph tracker once per run, not twice

run() and _build_agents() each called create_tracker() on the graph,
producing two tracker instances. Now run() creates the tracker once
and passes it to _build_agents() so handoff callbacks and run-level
tracking share the same instance.

Tests now assert graph.create_tracker is called exactly once per run
and node create_tracker is called exactly once per node.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: jsonbailey

packages/ai-providers/server-ai-openai/src/ldai_openai/openai_agent_graph_runner.py

@@ -81,7 +81,7 @@ async def run(self, input: Any) -> AgentGraphResult:
         state = _RunState(last_handoff_ns=start_ns, last_node_key=root_key)
         try:
             from agents import Runner
-            root_agent = self._build_agents(path, state)
+            root_agent = self._build_agents(path, state, tracker)
             result = await Runner.run(root_agent, str(input))
             self._flush_final_segment(state, result)
             self._track_tool_calls(result)
@@ -119,7 +119,9 @@ async def run(self, input: Any) -> AgentGraphResult:
                 metrics=LDAIMetrics(success=False),
             )
 
-    def _build_agents(self, path: List[str], state: _RunState) -> Any:
+    def _build_agents(
+        self, path: List[str], state: _RunState, tracker: Any
+    ) -> Any:
         """
         Build the agent tree from the graph definition via reverse_traverse.
 
@@ -128,6 +130,7 @@ def _build_agents(self, path: List[str], state: _RunState) -> Any:
 
         :param path: Mutable list to accumulate the execution path
         :param state: Shared run state for tracking handoff timing and last node
+        :param tracker: Graph-level tracker shared across the entire run
         :return: The root Agent instance
         """
         try:
@@ -143,7 +146,6 @@ def _build_agents(self, path: List[str], state: _RunState) -> Any:
                 "Install it with: pip install openai-agents"
             ) from exc
 
-        tracker = self._graph.create_tracker()
         name_map: Dict[str, str] = {}
         tool_name_map: Dict[str, str] = {}
         node_trackers: Dict[str, Any] = {}

packages/ai-providers/server-ai-openai/tests/test_openai_agent_graph_runner.py

@@ -14,13 +14,15 @@ def _make_graph(enabled: bool = True) -> AgentGraphDefinition:
     """Build a minimal single-node AgentGraphDefinition for testing."""
     node_tracker = MagicMock()
     graph_tracker = MagicMock()
+    node_factory = MagicMock(return_value=node_tracker)
+    graph_factory = MagicMock(return_value=graph_tracker)
     root_config = AIAgentConfig(
         key='root-agent',
         enabled=enabled,
         model=ModelConfig(name='gpt-4'),
         provider=ProviderConfig(name='openai'),
         instructions='You are a helpful assistant.',
-        create_tracker=lambda: node_tracker,
+        create_tracker=node_factory,
     )
     graph_config = AIAgentGraphConfig(
         key='test-graph',
@@ -34,7 +36,7 @@ def _make_graph(enabled: bool = True) -> AgentGraphDefinition:
         nodes=nodes,
         context=MagicMock(),
         enabled=enabled,
-        create_tracker=lambda: graph_tracker,
+        create_tracker=graph_factory,
     )
 
 
@@ -83,7 +85,7 @@ async def test_openai_agent_graph_runner_run_raises_when_agents_not_installed():
 @pytest.mark.asyncio
 async def test_openai_agent_graph_runner_run_tracks_invocation_failure_on_exception():
     graph = _make_graph()
-    tracker = graph.create_tracker()
+    tracker = graph.create_tracker.return_value
     runner = OpenAIAgentGraphRunner(graph, {})
 
     with patch.dict('sys.modules', {'agents': None}):
@@ -97,7 +99,7 @@ async def test_openai_agent_graph_runner_run_tracks_invocation_failure_on_except
 @pytest.mark.asyncio
 async def test_openai_agent_graph_runner_run_success():
     graph = _make_graph()
-    tracker = graph.create_tracker()
+    tracker = graph.create_tracker.return_value
 
     mock_result = MagicMock()
     mock_result.final_output = "agent answer"
@@ -140,8 +142,19 @@ async def test_openai_agent_graph_runner_run_success():
 
     # The runner caches one tracker per node — verify it is the same instance
     # returned by create_tracker() and that all tracking calls hit it.
+    node_factory = graph.get_node('root-agent').get_config().create_tracker
+
+    # The runner caches one tracker per node — verify it is the same instance
+    # returned by create_tracker and that all tracking calls hit it.
     cached = runner._node_trackers['root-agent']
-    assert cached is graph.get_node('root-agent').get_config().create_tracker()
+    assert cached is node_factory.return_value
     cached.track_duration.assert_called_once()
     cached.track_tokens.assert_called_once()
     cached.track_success.assert_called_once()
+
+    # Graph-level create_tracker is called exactly once per run (not twice)
+    # so that handoff callbacks and run() share the same tracker instance.
+    graph.create_tracker.assert_called_once()
+
+    # Node-level create_tracker is called exactly once per node.
+    node_factory.assert_called_once()