fix(router): consume _target_order at deployment selection so it never reaches a provider

Reading _target_order with .get left it in the request kwargs after selection, and only
nine provider boundaries stripped it. _atext_completion and _aadapter_completion spread
the raw kwargs, so an order-2 hop on /completions sent _target_order upstream, which
real providers reject as an unknown argument. Popping at selection strips it for every
path in one place; the PR's retry-keeping test already passed with pop because each retry
hands the callee its own kwargs copy.

Claude-Session: https://claude.ai/code/session_01XKkTFa6g7Rmd6vtHL91GMn

litellm/router.py

@@ -558,11 +558,6 @@ RETRY_BREADCRUMB_EXCLUDED_KWARGS: Final = frozenset(
 )
 
 
-def _without_target_order(kwargs: Mapping[str, object]) -> Mapping[str, object]:
-    """Drop the router-internal order-fallback target so it never reaches a provider call."""
-    return MappingProxyType({k: v for k, v in kwargs.items() if k != "_target_order"})
-
-
 class Router:
     model_names: set = set()
     cache_responses: bool | None = False
@@ -2177,7 +2172,7 @@ class Router:
                 "messages": messages,
                 "caching": self.cache_responses,
                 "client": model_client,
-                **_without_target_order(kwargs),
+                **kwargs,
             }
             response: Final = litellm.completion(**input_kwargs)
             verbose_router_logger.info("litellm.completion(model=%s)\x1b[32m 200 OK\x1b[0m", model_name)
@@ -3198,7 +3193,7 @@ class Router:
                 "messages": messages,
                 "caching": self.cache_responses,
                 "client": model_client,
-                **_without_target_order(kwargs),
+                **kwargs,
             }
             input_kwargs.pop("silent_model", None)
             input_kwargs.pop("include_fallback_errors", None)
@@ -4075,7 +4070,7 @@ class Router:
                     "prompt": prompt,
                     "caching": self.cache_responses,
                     "client": model_client,
-                    **_without_target_order(kwargs),
+                    **kwargs,
                 }
             )
             self.success_calls[model_name] += 1
@@ -4135,7 +4130,7 @@ class Router:
                     "prompt": prompt,
                     "caching": self.cache_responses,
                     "client": model_client,
-                    **_without_target_order(kwargs),
+                    **kwargs,
                 }
             )
 
@@ -4239,7 +4234,7 @@ class Router:
                     "file": file,
                     "caching": self.cache_responses,
                     "client": model_client,
-                    **_without_target_order(kwargs),
+                    **kwargs,
                 }
             )
 
@@ -4892,7 +4887,7 @@ class Router:
             response_kwargs: Final = {
                 **data,
                 "caching": self.cache_responses,
-                **_without_target_order(kwargs),
+                **kwargs,
                 "model": model_name,
             }
             # Only set custom_llm_provider if it's not None
@@ -5342,7 +5337,7 @@ class Router:
                     **data,
                     "custom_llm_provider": custom_llm_provider,
                     "caching": self.cache_responses,
-                    **_without_target_order(kwargs),
+                    **kwargs,
                 }
             )
 
@@ -5408,7 +5403,7 @@ class Router:
                     "input": input,
                     "caching": self.cache_responses,
                     "client": model_client,
-                    **_without_target_order(kwargs),
+                    **kwargs,
                 }
             )
             self.success_calls[model_name] += 1
@@ -5471,7 +5466,7 @@ class Router:
                     "input": input,
                     "caching": self.cache_responses,
                     "client": model_client,
-                    **_without_target_order(kwargs),
+                    **kwargs,
                 }
             )
 
@@ -11933,7 +11928,7 @@ class Router:
         )
 
         ## ORDER FILTERING ## -> if user set 'order' in deployments, return deployments with lowest order (e.g. order=1 > order=2)
-        _target_order: Final = (request_kwargs or {}).get("_target_order")
+        _target_order: Final = (request_kwargs or {}).pop("_target_order", None)
         healthy_deployments = litellm.utils._get_order_filtered_deployments(
             cast(list[dict], healthy_deployments), target_order=_target_order
         )
@@ -12698,7 +12693,7 @@ class Router:
             )
 
         ## ORDER FILTERING ## -> if user set 'order' in deployments, return deployments with lowest order (e.g. order=1 > order=2)
-        _target_order: Final = (request_kwargs or {}).get("_target_order")
+        _target_order: Final = (request_kwargs or {}).pop("_target_order", None)
         healthy_deployments = litellm.utils._get_order_filtered_deployments(
             healthy_deployments, target_order=_target_order
         )

tests/test_litellm/test_router_order_fallback.py

@@ -6,8 +6,10 @@ should be tried first, and higher order deployments should be used as fallbacks
 when lower order deployments fail.
 """
 
+import json
 from typing import Final, Optional
 
+import httpx
 import pytest
 
 import litellm
@@ -580,6 +582,64 @@ async def test_generic_api_call_strips_target_order_from_provider_kwargs():
     assert "_target_order" not in captured
 
 
+@pytest.mark.asyncio
+async def test_text_completion_order_fallback_hop_does_not_send_target_order_upstream():
+    upstream_bodies: Final[list[dict]] = []
+
+    def _upstream(request: httpx.Request) -> httpx.Response:
+        upstream_bodies.append(json.loads(request.content))
+        return httpx.Response(
+            200,
+            json={
+                "id": "cmpl-1",
+                "object": "text_completion",
+                "created": 0,
+                "model": "gpt-3.5-turbo-instruct",
+                "choices": [{"text": "ok from order 2", "index": 0, "logprobs": None, "finish_reason": "stop"}],
+                "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+            },
+        )
+
+    session: Final = httpx.AsyncClient(transport=httpx.MockTransport(_upstream))
+    litellm.in_memory_llm_clients_cache.flush_cache()
+    litellm.aclient_session = session
+    router = Router(
+        model_list=[
+            {
+                "model_name": "test-model",
+                "litellm_params": {
+                    "model": "text-completion-openai/gpt-3.5-turbo-instruct",
+                    "api_key": "key",
+                    "mock_response": Exception("fail order 1"),
+                    "order": 1,
+                },
+                "model_info": {"id": "1"},
+            },
+            {
+                "model_name": "test-model",
+                "litellm_params": {
+                    "model": "text-completion-openai/gpt-3.5-turbo-instruct",
+                    "api_key": "key",
+                    "api_base": "http://upstream.test",
+                    "order": 2,
+                },
+                "model_info": {"id": "2"},
+            },
+        ],
+        num_retries=0,
+    )
+    try:
+        response = await router.atext_completion(model="test-model", prompt="hi")
+    finally:
+        litellm.aclient_session = None
+        litellm.in_memory_llm_clients_cache.flush_cache()
+        await session.aclose()
+
+    assert response._hidden_params["model_id"] == "2"
+    assert upstream_bodies
+    assert all("_target_order" not in body for body in upstream_bodies)
+
+
 def test_check_non_standard_fallback_format():
     from litellm.router_utils.fallback_event_handlers import (
         _check_non_standard_fallback_format,