From 78207064d8d426b2e408c6572f780ccda613eaf6 Mon Sep 17 00:00:00 2001
From: tin <tin@berri.ai>
Date: Thu, 30 Jul 2026 01:09:37 +0000
Subject: [PATCH] fix(complexity_router): log the classifier request on chat
 completions too

The classifier read its metadata only from litellm_metadata, which the proxy
populates just for LITELLM_METADATA_ROUTES (/v1/messages, /v1/responses, ...);
/v1/chat/completions puts it under metadata, so the classifier call arrived
unattributed and _should_track_cost_callback dropped it, leaving no spend-log
row at all for the captured request body to show up in.

Also log response_format in the wire shape litellm actually sends
(type_to_response_format_param) instead of the bare pydantic JSON schema

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../complexity_router/complexity_router.py    |  6 +++--
 .../router_strategy/test_complexity_router.py | 25 +++++++++++++++++++
 2 files changed, 29 insertions(+), 2 deletions(-)

diff --git a/litellm/router_strategy/complexity_router/complexity_router.py b/litellm/router_strategy/complexity_router/complexity_router.py
index 836c2e9d4c0..4bc847c923e 100644
--- a/litellm/router_strategy/complexity_router/complexity_router.py
+++ b/litellm/router_strategy/complexity_router/complexity_router.py
@@ -25,6 +25,7 @@ from pydantic import BaseModel
 from litellm._logging import verbose_router_logger
 from litellm.constants import RETURN_RAW_MODEL_NAME_METADATA_KEY
 from litellm.integrations.custom_logger import CustomLogger
+from litellm.llms.base_llm.base_utils import type_to_response_format_param
 from litellm.types.utils import ModelResponse
 
 from .config import (
@@ -427,13 +428,14 @@ class ComplexityRouter(CustomLogger):
         # attributed to the calling key/team instead of being dropped. Excludes the
         # parent request's budget reservation, which the routed completion (not this
         # internal classifier call) is responsible for reconciling.
-        metadata = _classifier_call_metadata((request_kwargs or {}).get("litellm_metadata"))
+        request_metadata = (request_kwargs or {}).get("litellm_metadata") or (request_kwargs or {}).get("metadata")
+        metadata = _classifier_call_metadata(request_metadata)
 
         proxy_server_request = {
             "body": {
                 "model": llm_config.model,
                 "messages": [{"role": "user", "content": classification_prompt}],
-                "response_format": TierClassification.model_json_schema(),
+                "response_format": type_to_response_format_param(TierClassification),
             }
         }
 
diff --git a/tests/test_litellm/router_strategy/test_complexity_router.py b/tests/test_litellm/router_strategy/test_complexity_router.py
index a9b86c16c33..1feeb150c87 100644
--- a/tests/test_litellm/router_strategy/test_complexity_router.py
+++ b/tests/test_litellm/router_strategy/test_complexity_router.py
@@ -1417,6 +1417,24 @@ class TestLLMClassifier:
         call_kwargs = mock_router_instance.acompletion.call_args.kwargs
         assert call_kwargs["metadata"] == request_metadata
 
+    @pytest.mark.asyncio
+    async def test_aclassify_forwards_metadata_key_used_by_chat_completions(
+        self, llm_complexity_router, mock_router_instance
+    ):
+        """/v1/chat/completions puts the request metadata under "metadata", not "litellm_metadata".
+
+        Only the routes in LITELLM_METADATA_ROUTES (/v1/messages, /v1/responses, ...) get a
+        "litellm_metadata" bucket; chat completions gets "metadata". Reading only
+        "litellm_metadata" leaves the classifier call unattributed on the most common route,
+        so _should_track_cost_callback drops it and no spend-log row is written at all,
+        which also makes the captured request body unreachable in the Logs UI.
+        """
+        mock_router_instance.acompletion = AsyncMock(return_value=_llm_response('{"tier": "SIMPLE"}'))
+        request_metadata = {"user_api_key": "sk-abc", "user_api_key_team_id": "team-1"}
+        await llm_complexity_router.aclassify("hi", request_kwargs={"metadata": request_metadata})
+        call_kwargs = mock_router_instance.acompletion.call_args.kwargs
+        assert call_kwargs["metadata"] == request_metadata
+
     @pytest.mark.asyncio
     async def test_aclassify_captures_request_body_in_proxy_server_request(
         self, llm_complexity_router, mock_router_instance
@@ -1439,6 +1457,13 @@ class TestLLMClassifier:
         assert body["model"] == "haiku-classifier"
         assert body["messages"] == call_kwargs["messages"]
         assert "explain quantum tunneling in depth" in body["messages"][0]["content"]
+        assert body["response_format"]["type"] == "json_schema"
+        assert body["response_format"]["json_schema"]["schema"]["properties"]["tier"]["enum"] == [
+            "SIMPLE",
+            "MEDIUM",
+            "COMPLEX",
+            "REASONING",
+        ]
 
     @pytest.mark.asyncio
     async def test_aclassify_strips_budget_reservation_from_classifier_metadata(