feat(otel): stamp litellm.request.route on the LLM call span

litellm/integrations/otel/mappers/genai.py

@@ -89,6 +89,7 @@ class GenAIMapper:
         f"{LiteLLM.COST_PREFIX}margin_percent": lambda d: d.cost.margin_percent,
         f"{LiteLLM.COST_PREFIX}margin_total_amount": lambda d: d.cost.margin_total_amount,
         LiteLLM.REQUEST_STREAMING: lambda d: d.is_streaming,
+        LiteLLM.REQUEST_ROUTE: lambda d: d.identity.request_route or None,
     }
 
     _TOOL_ATTRS: dict[str, Callable[[ToolDefinition], AttrValue | None]] = {

litellm/integrations/otel/model/metadata.py

@@ -64,6 +64,9 @@ class RequestIdentity:
     # completes (routing has picked a deployment), so it's absent from the
     # auth-time seed and filled only from the payload.
     provider_model: str | None = None
+    # The normalized proxy route (``/v1/chat/completions``), set by auth after the
+    # Baggage seed, so like ``provider_model`` it is only filled from the payload.
+    request_route: str | None = None
     metadata: Mapping[str, str] = field(default_factory=dict)
 
     @classmethod
@@ -87,6 +90,7 @@ def from_payload(cls, payload: StandardLoggingPayload) -> RequestIdentity:
             key_hash=as_str(raw_meta.get("user_api_key_hash")),
             end_user=as_str(payload.get("end_user")) or as_str(raw_meta.get("user_api_key_end_user_id")),
             provider_model=resolve_provider_model(payload),
+            request_route=as_str(raw_meta.get("user_api_key_request_route")),
             metadata=metadata,
         )
 

litellm/integrations/otel/model/semconv.py

@@ -295,6 +295,7 @@ class LiteLLM:
     # ``litellm_params.model``), distinct from the user-facing ``gen_ai.request.model``.
     PROVIDER_MODEL: Final = "litellm.provider.model"
     REQUEST_STREAMING: Final = "litellm.request.streaming"
+    REQUEST_ROUTE: Final = "litellm.request.route"
     TOOLS_DECLARED: Final = "litellm.request.tools.declared"
     GUARDRAIL_NAME: Final = "litellm.guardrail.name"
     GUARDRAIL_MODE: Final = "litellm.guardrail.mode"

tests/test_litellm/integrations/otel/test_otel_v2_sources_of_truth.py

@@ -722,6 +722,26 @@ def test_request_identity_falls_back_to_legacy_team_keys():
     assert ident.team_alias == "legacy"
 
 
+def test_llm_span_carries_proxy_request_route():
+    """The LLM span records the proxy route the request arrived on, so it can be
+    filtered by endpoint (``/v1/responses`` vs ``/v1/chat/completions``) without
+    joining back to the root SERVER span's ``http.route``."""
+    data = LLMCallSpanData.from_standard_logging_payload(
+        _sample_payload(metadata={"user_api_key_request_route": "/v1/responses"})
+    )
+    attrs = GenAIMapper().map(data)
+
+    assert data.identity.request_route == "/v1/responses"
+    assert attrs[LiteLLM.REQUEST_ROUTE] == "/v1/responses"
+
+
+def test_llm_span_omits_request_route_off_the_proxy():
+    """An SDK call has no inbound route, so the key is absent rather than empty."""
+    attrs = GenAIMapper().map(LLMCallSpanData.from_standard_logging_payload(_sample_payload(metadata={})))
+
+    assert LiteLLM.REQUEST_ROUTE not in attrs
+
+
 def test_guardrail_span_data_block_carries_verdict_and_error():
     from litellm.integrations.otel.model.payloads import GuardrailSpanData