From d441be63f60d9b8013e13f035f11c7afa8e38afd Mon Sep 17 00:00:00 2001
From: "devin-ai-integration[bot]"
 <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Sun, 8 Nov 2026 01:34:51 +0000
Subject: [PATCH] fix(router): let the sync responses() honor the router's
 num_retries and fallbacks like aresponses()

---
 litellm/router_rust/invoke/endpoints.py       |  4 +-
 .../test_router_sync_responses_retries_sdk.py | 74 +++++++++++++++++++
 2 files changed, 77 insertions(+), 1 deletion(-)
 create mode 100644 tests/integration/sdk/test_router_sync_responses_retries_sdk.py

diff --git a/litellm/router_rust/invoke/endpoints.py b/litellm/router_rust/invoke/endpoints.py
index 7a5b52761e2..3041ba608a6 100644
--- a/litellm/router_rust/invoke/endpoints.py
+++ b/litellm/router_rust/invoke/endpoints.py
@@ -1652,7 +1652,7 @@ class RustEndpoints:
                     raise TypeError(
                         "_generic_api_call_with_fallbacks() missing 1 required positional argument: 'model'"
                     )
-                return self._core.call_sync(plan, kwargs, _SINGLE_ATTEMPT)
+                return self._core.call_sync(plan, kwargs, None if name in _SYNC_RETRIED else _SINGLE_ATTEMPT)
             provider_kwargs: dict[  # mutable-ok: the request kwargs the Python router mutates in place  # rebind-ok: rebound as the Python router rebinds it
                 str, object
             ] = (
@@ -2465,6 +2465,8 @@ class RustEndpoints:
 
 
 _SINGLE_ATTEMPT: Final[Mapping[str, object]] = MappingProxyType({"num_retries": 0, "disable_fallbacks": True})
+# The sync endpoints whose walk honors the router's `num_retries` and `fallbacks` like their async twin (#49938).
+_SYNC_RETRIED: Final = frozenset({"responses"})
 
 
 async def _answers_none(*_args: object, **_kwargs: object) -> None:  # kwargs-ok: any request
diff --git a/tests/integration/sdk/test_router_sync_responses_retries_sdk.py b/tests/integration/sdk/test_router_sync_responses_retries_sdk.py
new file mode 100644
index 00000000000..d179bc306be
--- /dev/null
+++ b/tests/integration/sdk/test_router_sync_responses_retries_sdk.py
@@ -0,0 +1,74 @@
+from __future__ import annotations
+
+import uuid
+from collections.abc import Callable
+from typing import Final
+
+from integration._support.anthropic_sse import Attempts
+from integration._support.openai_wire import openai_error, posted_targets, responses_reply
+from integration._support.wire import Reply, Request, wire_server
+from litellm import Router
+from litellm.types.llms.openai import ResponsesAPIResponse
+from litellm.types.router import DeploymentTypedDict
+
+_MODEL: Final = "gpt-5.4"
+_API_KEY: Final = "sync-responses-retry-test-key"
+_TEXT: Final = "served after the blip"
+
+
+def _deployment(model_name: str, api_base: str) -> DeploymentTypedDict:
+    return DeploymentTypedDict(
+        model_name=model_name,
+        litellm_params={"model": f"openai/{_MODEL}", "api_base": api_base, "api_key": _API_KEY},
+    )
+
+
+def _upstream(fails: Callable[[str, int], bool]) -> Callable[[Request], Reply]:
+    attempts: Final = Attempts()
+
+    def respond(request: Request) -> Reply:
+        group, _, route = request.target.lstrip("/").partition("/")
+        assert (request.method, route) == ("POST", "responses"), request
+        if fails(group, attempts.record(group)):
+            return openai_error(503)
+        return responses_reply(f"resp_{group}_{uuid.uuid4().hex}", _MODEL, _TEXT, stream=False)
+
+    return respond
+
+
+def _output_text(response: object) -> str:
+    assert isinstance(response, ResponsesAPIResponse), response
+    return str(response.output[0].content[0].text)  # pyright: ignore[reportUnknownMemberType, reportAttributeAccessIssue]  # the output item shape is the wire's
+
+
+def test_sync_responses_retries_the_group_after_a_503() -> None:
+    with wire_server(_upstream(lambda _group, attempt: attempt == 1)) as wire:
+        router: Final = Router(
+            model_list=[_deployment("my-model", f"{wire.url}/my-model")],
+            num_retries=1,
+            retry_after=0,
+            disable_cooldowns=True,
+        )
+        response: Final = router.responses(model="my-model", input=f"retry {uuid.uuid4().hex}")
+        targets: Final = posted_targets(wire)
+
+    assert _output_text(response) == _TEXT
+    assert targets == ("/my-model/responses", "/my-model/responses")
+
+
+def test_sync_responses_falls_back_when_the_primary_group_keeps_failing() -> None:
+    with wire_server(_upstream(lambda group, _attempt: group == "primary")) as wire:
+        router: Final = Router(
+            model_list=[
+                _deployment("primary", f"{wire.url}/primary"),
+                _deployment("backup", f"{wire.url}/backup"),
+            ],
+            fallbacks=[{"primary": ["backup"]}],
+            num_retries=0,
+            disable_cooldowns=True,
+        )
+        response: Final = router.responses(model="primary", input=f"fallback {uuid.uuid4().hex}")
+        targets: Final = posted_targets(wire)
+
+    assert _output_text(response) == _TEXT
+    assert targets == ("/primary/responses", "/backup/responses")