fix(router): keep the mirror flag when litellm_metadata is set and strip a client-sent is_silent_experiment

Spend rows read router metadata from litellm_metadata when a call carries it, so the router now sets the flag there too. The proxy drops a client-sent is_silent_experiment with the other router-owned keys, so real traffic cannot be recorded as mirrored.

litellm/proxy/litellm_pre_call_utils.py

@@ -469,6 +469,7 @@ _CLIENT_PRICING_METADATA_FIELDS: Final = frozenset({"model_info", "standard_logg
 _ROUTER_RESERVED_METADATA_FIELDS: Final = frozenset(
     {
         "attempted_fallbacks",
+        "is_silent_experiment",
         "original_model_group",
         "previous_models",
         "request_retry_count",

litellm/router.py

@@ -3124,6 +3124,8 @@ class Router:
         silent_kwargs["metadata"].pop("litellm_parent_otel_span", None)
 
         silent_kwargs["metadata"]["is_silent_experiment"] = True
+        if "litellm_metadata" in silent_kwargs and isinstance(silent_kwargs["litellm_metadata"], dict):
+            silent_kwargs["litellm_metadata"] = {**silent_kwargs["litellm_metadata"], "is_silent_experiment": True}
 
         # Pop logging objects and call IDs to ensure a fresh logging context
         # This prevents collisions in the Proxy's database (spend_logs)

tests/unit/proxy/test_litellm_pre_call_utils.py

@@ -8278,6 +8278,7 @@ def _reserved_stamp_key(key_metadata: dict | None = None) -> UserAPIKeyAuth:
 
 _PLANTED_STAMPS = {
     "attempted_fallbacks": 99,
+    "is_silent_experiment": True,
     "original_model_group": "spoofed-group",
     "request_retry_count": -100,
     "previous_models": [{"exception_type": "Timeout", "deployment_id": "spoofed", "attempted_retries": 101}],
@@ -8312,6 +8313,7 @@ async def test_add_litellm_data_to_request_strips_router_reserved_stamps_from_bo
 
     assert "litellm_metadata" not in updated
     assert "attempted_fallbacks" not in updated["metadata"]
+    assert "is_silent_experiment" not in updated["metadata"]
     assert "original_model_group" not in updated["metadata"]
     assert "_client_output_ceiling" not in updated["metadata"]
     assert "request_retry_count" not in updated["metadata"]

tests/unit/test_router_silent_experiment.py

@@ -1,4 +1,5 @@
 import asyncio
+import datetime
 import json
 import time
 from collections.abc import Callable, Mapping
@@ -282,6 +283,34 @@ async def test_standard_logging_payload_marks_only_the_mirrored_calls_as_silent_
     assert "is_silent_experiment" not in metadata_by_group["primary-model"]
 
 
+@pytest.mark.asyncio
+async def test_spend_rows_mark_only_the_mirrored_calls_when_the_caller_sends_litellm_metadata(recording_logger):
+    from litellm.proxy.spend_tracking.spend_tracking_utils import get_logging_payload
+
+    router: Final = Router(model_list=_streaming_model_list(["shadow-a", "shadow-b"]))
+    await router.acompletion(
+        model="primary-model",
+        messages=[{"role": "user", "content": "hi"}],
+        mock_response="pong",
+        litellm_metadata={"user_api_key": "test-key"},
+    )
+    await _wait_for_all_successes(recording_logger, expected=3)
+
+    now: Final = datetime.datetime.now(datetime.timezone.utc)
+    silent_by_group: Final = {
+        call["standard_logging_object"]["model_group"]: json.loads(
+            get_logging_payload(
+                kwargs=call,
+                response_obj=litellm.ModelResponse(id="chatcmpl-test", choices=[]),
+                start_time=now,
+                end_time=now,
+            )["metadata"]
+        )["is_silent_experiment"]
+        for call in recording_logger.success_kwargs
+    }
+    assert silent_by_group == {"primary-model": False, "shadow-a": True, "shadow-b": True}
+
+
 def test_multiple_shadow_targets_fan_out_sync(recording_logger):
     router = Router(model_list=_streaming_model_list(["shadow-a", "shadow-b"]))
     response = router.completion(