fix(langfuse): trace successful decisions calls with their token counts

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

litellm/integrations/langfuse/langfuse.py

@@ -15,6 +15,7 @@ from packaging.version import Version
 import litellm
 from litellm._logging import verbose_logger
 from litellm.constants import MAX_LANGFUSE_INITIALIZED_CLIENTS
+from litellm.cost_calculator import get_usage_object
 from litellm.integrations.langfuse.langfuse_mock_client import (
     create_mock_langfuse_client,
     should_use_langfuse_mock,
@@ -32,6 +33,7 @@ from litellm.litellm_core_utils.redact_messages import redact_user_api_key_info
 from litellm.llms.custom_httpx.http_handler import get_httpx_client
 from litellm.responses.utils import ResponseAPILoggingUtils
 from litellm.secret_managers.main import str_to_bool
+from litellm.types.decisions import DecisionsResponse, OpenAIDecisionResponse
 from litellm.types.integrations.langfuse import *
 from litellm.types.llms.openai import HttpxBinaryResponseContent, ResponseAPIUsage, ResponsesAPIResponse
 from litellm.types.utils import (
@@ -584,7 +586,9 @@ class LangFuseLogger:
         | TranscriptionResponse
         | RerankResponse
         | HttpxBinaryResponseContent
-        | ResponsesAPIResponse,
+        | ResponsesAPIResponse
+        | DecisionsResponse
+        | OpenAIDecisionResponse,
         prompt: dict,
         level: str,
         status_message: str | None,
@@ -634,6 +638,9 @@ class LangFuseLogger:
         elif response_obj is not None and isinstance(response_obj, litellm.ResponsesAPIResponse):
             input = prompt
             output = self._get_responses_api_content_for_langfuse(response_obj)
+        elif response_obj is not None and isinstance(response_obj, (DecisionsResponse, OpenAIDecisionResponse)):
+            input = prompt
+            output = response_obj.model_dump(mode="json", exclude={"usage"})
         elif (
             kwargs.get("call_type") is not None
             and kwargs.get("call_type") == "_arealtime"
@@ -859,10 +866,14 @@ class LangFuseLogger:
             usage = None
             usage_details = None
             if response_obj is not None:
-                if hasattr(response_obj, "id") and response_obj.get("id", None) is not None:
+                if isinstance(response_obj, _ResponseWithId) and getattr(response_obj, "id", None) is not None:
                     generation_id = _logging_id(start_time, response_obj)
                 _raw_usage_obj: Final[_UsageObject | None] = getattr(response_obj, "usage", None)
-                _usage_obj: Final = _chat_usage(_raw_usage_obj)
+                _usage_obj: Final = (
+                    get_usage_object(response_obj)
+                    if isinstance(response_obj, (DecisionsResponse, OpenAIDecisionResponse))
+                    else _chat_usage(_raw_usage_obj)
+                )
 
                 if _usage_obj:
                     # Safely get usage values, defaulting None to 0 for Langfuse compatibility.

tests/unit/integrations/langfuse/test_langfuse.py

@@ -433,6 +433,49 @@ async def test_langfuse_logging_vllm_embedding(
     assert route.call_count == 1
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("response_id", [None, "decision-1"])
+async def test_langfuse_logging_systemone(
+    langfuse_export: tuple[respx.Route, str], respx_mock: respx.MockRouter, response_id: str | None
+) -> None:
+    route, trace_id = langfuse_export
+    answers: Final = {"urgent": {"type": "noul", "noul": 0.87}}
+    upstream: Final = respx_mock.post("https://openrouter.ai/api/alpha/decisions").respond(
+        json={
+            **({"id": response_id} if response_id is not None else {}),
+            "model": "typesafe/jev-1.13",
+            "answers": answers,
+            "usage": {"input_tokens": 283, "output_tokens": 20},
+        }
+    )
+    await litellm.asystemone(
+        model="openrouter/typesafe/jev-1.13",
+        state="Customer writes: my payout has failed three days in a row.",
+        questions={"urgent": {"type": "noul", "instructions": "Is this urgent?"}},
+        api_key="caller-key",
+        metadata={"trace_id": trace_id},
+    )
+    await _flush_langfuse()
+    assert upstream.call_count == 1
+    generations: Final = tuple(
+        TypeAdapter(dict[str, object]).validate_python(span["attributes"])
+        for span in _exported_spans(route)
+        if span["trace_id"] == resolve_trace_id(trace_id)
+        and TypeAdapter(dict[str, object]).validate_python(span["attributes"]).get("langfuse.observation.type")
+        == "generation"
+    )
+    assert len(generations) == 1
+    assert generations[0]["langfuse.observation.usage_details"] == {
+        "input": 283,
+        "output": 20,
+        "total": 303,
+        "cache_creation_input_tokens": 0,
+        "cache_read_input_tokens": 0,
+    }
+    output: Final = TypeAdapter(dict[str, object]).validate_python(generations[0]["langfuse.observation.output"])
+    assert output["answers"] == answers
+
+
 @pytest.mark.asyncio
 async def test_langfuse_logging_with_router(langfuse_export: tuple[respx.Route, str]) -> None:
     route, trace_id = langfuse_export