fix: using local variable to track start time

Author: constantinius

sentry_sdk/integrations/openai_agents/patches/models.py

@@ -150,18 +150,16 @@ async def wrapped_stream_response(*args: "Any", **kwargs: "Any") -> "Any":
 
                     streaming_response = None
                     ttft_recorded = False
+                    # Capture start time locally to avoid race conditions with concurrent requests
+                    start_time = time.perf_counter()
 
                     async for event in original_stream_response(*args, **kwargs):
                         # Detect first content token (text delta event)
                         if not ttft_recorded and hasattr(event, "delta"):
-                            start_time = getattr(
-                                agent, "_sentry_chat_ttft_start_time", None
+                            ttft = time.perf_counter() - start_time
+                            span.set_data(
+                                SPANDATA.GEN_AI_RESPONSE_TIME_TO_FIRST_TOKEN, ttft
                             )
-                            if start_time is not None:
-                                ttft = time.perf_counter() - start_time
-                                span.set_data(
-                                    SPANDATA.GEN_AI_RESPONSE_TIME_TO_FIRST_TOKEN, ttft
-                                )
                             ttft_recorded = True
 
                         # Capture the full response from ResponseCompletedEvent

sentry_sdk/integrations/openai_agents/spans/ai_client.py

@@ -1,5 +1,3 @@
-import time
-
 import sentry_sdk
 from sentry_sdk.consts import OP, SPANDATA
 
@@ -38,9 +36,6 @@ def ai_client_span(
     # TODO-anton: remove hardcoded stuff and replace something that also works for embedding and so on
     span.set_data(SPANDATA.GEN_AI_OPERATION_NAME, "chat")
 
-    # Store start time for TTFT calculation on the agent object
-    agent._sentry_chat_ttft_start_time = time.perf_counter()
-
     _set_agent_data(span, agent)
     _set_input_data(span, get_response_kwargs)
 

tests/integrations/openai_agents/test_openai_agents.py

@@ -2276,15 +2276,12 @@ async def stream_response(self, *args, **kwargs):
             span.set_data(SPANDATA.GEN_AI_RESPONSE_STREAMING, True)
 
             ttft_recorded = False
-            start_time = getattr(test_agent, "_sentry_chat_ttft_start_time", None)
+            # Capture start time locally (same as production code after race condition fix)
+            start_time = time.perf_counter()
 
             async for event in mock_model.stream_response():
                 # This is the same logic used in the actual integration
-                if (
-                    not ttft_recorded
-                    and hasattr(event, "delta")
-                    and start_time is not None
-                ):
+                if not ttft_recorded and hasattr(event, "delta"):
                     ttft = time.perf_counter() - start_time
                     span.set_data(SPANDATA.GEN_AI_RESPONSE_TIME_TO_FIRST_TOKEN, ttft)
                     ttft_recorded = True