fix(llmobs): skip downstream OpenAI span check for litellm_proxy/ models

ddtrace/llmobs/_integrations/litellm.py

@@ -158,6 +158,9 @@ def _has_downstream_openai_span(self, kwargs: dict[str, Any], model: Optional[st
         """
         stream = kwargs.get("stream", False)
         model_lower = model.lower() if model else ""
+        # litellm_proxy/ requests route through a proxy — the OpenAI integration never fires for these
+        if model_lower.startswith("litellm_proxy/"):
+            return False
         # best effort attempt to check if Open AI or Azure since model_provider is unknown until request completes
         is_openai_model = any(prefix in model_lower for prefix in ("gpt", "openai", "azure"))
         return is_openai_model and not stream and LLMObs._integration_is_enabled("openai")

releasenotes/notes/litellm-proxy-model-llmobs-fix-a1b2c3d4e5f6a7b8.yaml

@@ -0,0 +1,7 @@
+---
+fixes:
+  - |
+    litellm: Fix missing LLMObs spans when routing through a litellm proxy.
+    Models with the ``litellm_proxy/`` prefix were incorrectly suppressed by
+    the OpenAI downstream span check, resulting in empty or missing LLMObs
+    spans. Proxy requests are now always handled by the litellm integration.

tests/contrib/litellm/test_litellm_llmobs.py

@@ -611,6 +611,35 @@ def test_completion_with_reasoning(self, litellm, request_vcr, llmobs_events, te
         assert event_metrics["reasoning_output_tokens"] == 15
 
 
+@pytest.mark.parametrize(
+    "model,stream,openai_enabled,expected",
+    [
+        # litellm_proxy/ prefix always suppresses downstream check regardless of model name or OpenAI enabled
+        ("litellm_proxy/azure-gpt-5-nano", False, True, False),
+        ("litellm_proxy/gpt-4o", False, True, False),
+        ("litellm_proxy/openai/gpt-4", False, True, False),
+        # normal OpenAI/Azure models with OpenAI integration enabled (non-streamed) should return True
+        ("gpt-4o", False, True, True),
+        ("azure/gpt-4", False, True, True),
+        ("openai/gpt-4", False, True, True),
+        # streaming disables downstream check
+        ("gpt-4o", True, True, False),
+        # OpenAI integration disabled disables downstream check
+        ("gpt-4o", False, False, False),
+        # non-OpenAI models are unaffected
+        ("anthropic/claude-3", False, True, False),
+    ],
+)
+def test_has_downstream_openai_span(model, stream, openai_enabled, expected):
+    from ddtrace import config
+    from ddtrace.llmobs._integrations import LiteLLMIntegration
+
+    integration = LiteLLMIntegration(integration_config=config.litellm)
+    kwargs = {"stream": stream}
+    with mock.patch("ddtrace.llmobs._integrations.litellm.LLMObs._integration_is_enabled", return_value=openai_enabled):
+        assert integration._has_downstream_openai_span(kwargs, model) is expected
+
+
 def test_enable_llmobs_after_litellm_was_imported(run_python_code_in_subprocess):
     """
     Test that LLMObs.enable() logs a warning if litellm is imported before LLMObs.enable() is called.