feat(integrations): openai-agents: add usage and response model reporting for chat and invoke_agent spans

sentry_sdk/integrations/openai_agents/patches/agent_run.py

@@ -8,6 +8,8 @@
 if TYPE_CHECKING:
     from typing import Any, Optional
 
+    from sentry_sdk.tracing import Span
+
 try:
     import agents
 except ImportError:
@@ -27,13 +29,15 @@ def _patch_agent_run():
     original_execute_final_output = agents._run_impl.RunImpl.execute_final_output
 
     def _start_invoke_agent_span(context_wrapper, agent, kwargs):
-        # type: (agents.RunContextWrapper, agents.Agent, dict[str, Any]) -> None
+        # type: (agents.RunContextWrapper, agents.Agent, dict[str, Any]) -> Span
         """Start an agent invocation span"""
         # Store the agent on the context wrapper so we can access it later
         context_wrapper._sentry_current_agent = agent
         span = invoke_agent_span(context_wrapper, agent, kwargs)
         context_wrapper._sentry_agent_span = span
 
+        return span
+
     def _end_invoke_agent_span(context_wrapper, agent, output=None):
         # type: (agents.RunContextWrapper, agents.Agent, Optional[Any]) -> None
         """End the agent invocation span"""
@@ -73,7 +77,8 @@ async def patched_run_single_turn(cls, *args, **kwargs):
                 if current_agent and current_agent != agent:
                     _end_invoke_agent_span(context_wrapper, current_agent)
 
-            _start_invoke_agent_span(context_wrapper, agent, kwargs)
+            span = _start_invoke_agent_span(context_wrapper, agent, kwargs)
+            agent._sentry_agent_span = span
 
         # Call original method with all the correct parameters
         result = await original_run_single_turn(*args, **kwargs)

sentry_sdk/integrations/openai_agents/patches/models.py

@@ -3,6 +3,7 @@
 from sentry_sdk.integrations import DidNotEnable
 
 from ..spans import ai_client_span, update_ai_client_span
+from sentry_sdk.consts import SPANDATA
 
 from typing import TYPE_CHECKING
 
@@ -33,13 +34,37 @@ def wrapped_get_model(cls, agent, run_config):
         model = original_get_model(agent, run_config)
         original_get_response = model.get_response
 
+        # Wrap _fetch_response if it exists (for OpenAI models) to capture raw response model
+        if hasattr(model, "_fetch_response"):
+            original_fetch_response = model._fetch_response
+
+            @wraps(original_fetch_response)
+            async def wrapped_fetch_response(*args, **kwargs):
+                # type: (*Any, **Any) -> Any
+                response = await original_fetch_response(*args, **kwargs)
+                if hasattr(response, "model"):
+                    agent._sentry_raw_response_model = str(response.model)
+                return response
+
+            model._fetch_response = wrapped_fetch_response
+
         @wraps(original_get_response)
         async def wrapped_get_response(*args, **kwargs):
             # type: (*Any, **Any) -> Any
             with ai_client_span(agent, kwargs) as span:
                 result = await original_get_response(*args, **kwargs)
 
-                update_ai_client_span(span, agent, kwargs, result)
+                response_model = getattr(agent, "_sentry_raw_response_model", None)
+                if response_model:
+                    agent_span = getattr(agent, "_sentry_agent_span", None)
+                    if agent_span:
+                        agent_span.set_data(
+                            SPANDATA.GEN_AI_RESPONSE_MODEL, response_model
+                        )
+
+                    delattr(agent, "_sentry_raw_response_model")
+
+                update_ai_client_span(span, agent, kwargs, result, response_model)
 
             return result
 

sentry_sdk/integrations/openai_agents/patches/runner.py

@@ -26,9 +26,11 @@ async def wrapper(*args, **kwargs):
         # Isolate each workflow so that when agents are run in asyncio tasks they
         # don't touch each other's scopes
         with sentry_sdk.isolation_scope():
-            agent = args[0]
+            # Clone agent because agent invocation spans are attached per run.
+            agent = args[0].clone()
             with agent_workflow_span(agent):
                 result = None
+                args = (agent, *args[1:])
                 try:
                     result = await original_func(*args, **kwargs)
                     return result

sentry_sdk/integrations/openai_agents/spans/ai_client.py

@@ -14,7 +14,7 @@
 
 if TYPE_CHECKING:
     from agents import Agent
-    from typing import Any
+    from typing import Any, Optional
 
 
 def ai_client_span(agent, get_response_kwargs):
@@ -35,8 +35,14 @@ def ai_client_span(agent, get_response_kwargs):
     return span
 
 
-def update_ai_client_span(span, agent, get_response_kwargs, result):
-    # type: (sentry_sdk.tracing.Span, Agent, dict[str, Any], Any) -> None
+def update_ai_client_span(
+    span, agent, get_response_kwargs, result, response_model=None
+):
+    # type: (sentry_sdk.tracing.Span, Agent, dict[str, Any], Any, Optional[str]) -> None
     _set_usage_data(span, result.usage)
     _set_output_data(span, result)
     _create_mcp_execute_tool_spans(span, result)
+
+    # Set response model if captured from raw response
+    if response_model is not None:
+        span.set_data(SPANDATA.GEN_AI_RESPONSE_MODEL, response_model)

sentry_sdk/integrations/openai_agents/spans/invoke_agent.py

@@ -10,7 +10,7 @@
 from sentry_sdk.utils import safe_serialize
 
 from ..consts import SPAN_ORIGIN
-from ..utils import _set_agent_data
+from ..utils import _set_agent_data, _set_usage_data
 
 from typing import TYPE_CHECKING
 
@@ -84,6 +84,10 @@ def update_invoke_agent_span(context, agent, output):
     span = getattr(context, "_sentry_agent_span", None)
 
     if span:
+        # Add aggregated usage data from context_wrapper
+        if hasattr(context, "usage"):
+            _set_usage_data(span, context.usage)
+
         if should_send_default_pii():
             set_data_normalized(
                 span, SPANDATA.GEN_AI_RESPONSE_TEXT, output, unpack=False