fix(openai-agents): preserve response metadata during tracing

py/src/braintrust/integrations/openai_agents/test_openai_agents.py

@@ -79,16 +79,29 @@ def export(self):
 
 @pytest.mark.asyncio
 @pytest.mark.vcr
-async def test_openai_agents_integration_setup_creates_spans(memory_logger):
+async def test_openai_agents_integration_setup_creates_spans(memory_logger, monkeypatch):
     import agents
     from agents import Agent
     from agents.run import AgentRunner
+    from braintrust.integrations.openai_agents import tracing
 
     assert not memory_logger.pop()
 
     assert OpenAIAgentsIntegration.setup() is True
     assert agents.tracing.get_trace_provider()._disabled is False
 
+    response_metadata = []
+    response_log_data = tracing.BraintrustTracingProcessor._response_log_data
+
+    def capture_response_metadata(self, span):
+        response = span.span_data.response
+        if response is not None:
+            response.metadata["user"] = "preserve-me"
+            response_metadata.append((response.metadata, dict(response.metadata or {})))
+        return response_log_data(self, span)
+
+    monkeypatch.setattr(tracing.BraintrustTracingProcessor, "_response_log_data", capture_response_metadata)
+
     agent = Agent(name="test-agent", model=TEST_MODEL, instructions=TEST_AGENT_INSTRUCTIONS)
     result = await AgentRunner().run(agent, TEST_PROMPT)
 
@@ -116,6 +129,15 @@ async def test_openai_agents_integration_setup_creates_spans(memory_logger):
     assert any(metrics.get("prompt_cached_tokens") == 0 for metrics in llm_metrics)
     assert any(metrics.get("completion_reasoning_tokens") == 0 for metrics in llm_metrics)
 
+    assert response_metadata
+    for metadata, original_metadata in response_metadata:
+        assert metadata == original_metadata
+
+    response_spans = [span for span in llm_spans if span.get("span_attributes", {}).get("name") == "Response"]
+    assert response_spans
+    assert any(span.get("metadata", {}).get("user") == "preserve-me" for span in response_spans)
+    assert all(value is not None for span in response_spans for value in span.get("metadata", {}).values())
+
     # v0.14.0+ produces TaskSpanData and TurnSpanData wrapping agent runs. The
     # tracing processor must give them proper names, metadata, and usage
     # metrics rather than returning "Unknown" / empty dicts.

py/src/braintrust/integrations/openai_agents/tracing.py

@@ -192,9 +192,9 @@ def _response_log_data(self, span: tracing.Span[tracing.ResponseSpanData]) -> di
         if span.span_data.response is not None:
             data["output"] = span.span_data.response.output
         if span.span_data.response is not None:
-            data["metadata"] = span.span_data.response.metadata or {}
+            data["metadata"] = dict(span.span_data.response.metadata or {})
             data["metadata"].update(
-                span.span_data.response.model_dump(exclude={"input", "output", "metadata", "usage"})
+                span.span_data.response.model_dump(exclude={"input", "output", "metadata", "usage"}, exclude_none=True)
             )
 
         data["metrics"] = {}