fix(anthropic): skip the SDK num_retries retry for an attempt that is sent again

litellm/_internal_context.py

@@ -118,7 +118,8 @@ def current_service_caller() -> str | None:
 @contextmanager
 def resent_on(rejection: Callable[[Exception], bool]) -> Generator[None]:
     """An attempt its caller sends again when it fails with ``rejection``. That failure is not the request's
-    outcome, so nothing logs it and its budget reservation and rate-limit slot stay with the attempt sent again."""
+    outcome, so it is neither logged as the request's failure nor retried, and its budget reservation and rate-limit
+    slot stay with the attempt sent again."""
     token: Final = _resent_on.set(rejection)
     try:
         yield

litellm/utils.py

@@ -1985,7 +1985,7 @@ def client(original_function):
 
                 is_completion_litellm_router_call: Final = _is_litellm_router_call(kwargs, is_async=False)
                 if (
-                    num_retries and not is_completion_litellm_router_call
+                    num_retries and not is_completion_litellm_router_call and not is_resent_failure(e)
                 ):  # only enter this if call is not from litellm router/proxy. router has it's own logic for retrying
                     if (
                         isinstance(e, openai.APIError)
@@ -2019,7 +2019,7 @@ def client(original_function):
 
                 is_responses_litellm_router_call: Final = _is_litellm_router_call(kwargs, is_async=False)
                 if (
-                    num_retries and not is_responses_litellm_router_call
+                    num_retries and not is_responses_litellm_router_call and not is_resent_failure(e)
                 ):  # only enter this if call is not from litellm router/proxy. router has it's own logic for retrying
                     if (
                         isinstance(e, openai.APIError)
@@ -2300,7 +2300,10 @@ def client(original_function):
                 is_acompletion_litellm_router_call: Final = _is_litellm_router_call(kwargs, is_async=True)
 
                 if (
-                    num_retries and not is_acompletion_litellm_router_call and not isinstance(e, ImportError)
+                    num_retries
+                    and not is_acompletion_litellm_router_call
+                    and not isinstance(e, ImportError)
+                    and not is_resent_failure(e)
                 ):  # only enter this if call is not from litellm router/proxy. router has it's own logic for retrying
                     try:
                         litellm.num_retries = None  # set retries to None to prevent infinite loops
@@ -2331,7 +2334,10 @@ def client(original_function):
                 is_aresponses_litellm_router_call: Final = _is_litellm_router_call(kwargs, is_async=True)
 
                 if (
-                    num_retries and not is_aresponses_litellm_router_call and not isinstance(e, ImportError)
+                    num_retries
+                    and not is_aresponses_litellm_router_call
+                    and not isinstance(e, ImportError)
+                    and not is_resent_failure(e)
                 ):  # only enter this if call is not from litellm router/proxy. router has it's own logic for retrying
                     try:
                         litellm.num_retries = None  # set retries to None to prevent infinite loops

tests/unit/llms/anthropic/pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py

@@ -5584,6 +5584,51 @@ def test_sync_messages_retry_once_without_reasoning_the_backend_cannot_decrypt(
     assert [block["text"] for block in response["content"] if block["type"] == "text"] == ["Oslo"]
 
 
+@pytest.mark.asyncio
+@pytest.mark.usefixtures("_chat_bridge_httpx")
+async def test_async_messages_with_num_retries_send_the_rejected_reasoning_only_once(
+    respx_mock: respx.MockRouter,
+) -> None:
+    route: Final = respx_mock.post(url__regex=_RETRY_URL).mock(
+        side_effect=[_RETRY_REJECTION, httpx.Response(200, json=_RETRY_REPLY)]
+    )
+
+    await litellm.anthropic.messages.acreate(
+        model="azure/gpt-5.4-nano",
+        api_base=_RETRY_API_BASE,
+        api_key="fake-key",
+        api_version="2025-04-01-preview",
+        max_tokens=64,
+        tools=_RETRY_TOOLS,
+        messages=_history_with_foreign_reasoning(),
+        num_retries=2,
+    )
+
+    assert [len(_sent_reasoning_items(call.request)) for call in route.calls] == [1, 0]
+
+
+@pytest.mark.usefixtures("_chat_bridge_httpx")
+def test_sync_messages_with_num_retries_send_the_rejected_reasoning_only_once(
+    respx_mock: respx.MockRouter,
+) -> None:
+    route: Final = respx_mock.post(url__regex=_RETRY_URL).mock(
+        side_effect=[_RETRY_REJECTION, httpx.Response(200, json=_RETRY_REPLY)]
+    )
+
+    litellm.anthropic.messages.create(
+        model="azure/gpt-5.4-nano",
+        api_base=_RETRY_API_BASE,
+        api_key="fake-key",
+        api_version="2025-04-01-preview",
+        max_tokens=64,
+        tools=_RETRY_TOOLS,
+        messages=_history_with_foreign_reasoning(),
+        num_retries=2,
+    )
+
+    assert [len(_sent_reasoning_items(call.request)) for call in route.calls] == [1, 0]
+
+
 class _FailureLog(CustomLogger):
     def __init__(self) -> None:
         super().__init__()

tests/unit/llms/anthropic/pass_through/responses_adapters/test_responses_adapters_handler.py

@@ -800,6 +800,39 @@ def test_sync_openai_turn_is_sent_again_once_without_replayed_reasoning_the_back
     assert cast(Mapping[str, object], response)["content"] == [{"type": "text", "text": "done"}]
 
 
+@pytest.mark.asyncio
+async def test_openai_turn_with_num_retries_sends_the_rejected_reasoning_only_once(
+    respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    upstream: Final = _upstream(respx_mock, monkeypatch, httpx.Response(400, json=REJECTED_REPLAYED_REASONING), False)
+
+    await litellm.anthropic.messages.acreate(
+        max_tokens=64,
+        messages=_history(replayed_reasoning=True),
+        model="openai/gpt-5.6-luna",
+        tools=[dict(BASH_TOOL)],
+        num_retries=2,
+    )
+
+    assert _replayed_reasoning_counts(upstream) == [1, 0]
+
+
+def test_sync_openai_turn_with_num_retries_sends_the_rejected_reasoning_only_once(
+    respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    upstream: Final = _upstream(respx_mock, monkeypatch, httpx.Response(400, json=REJECTED_REPLAYED_REASONING), False)
+
+    litellm.anthropic.messages.create(
+        max_tokens=64,
+        messages=_history(replayed_reasoning=True),
+        model="openai/gpt-5.6-luna",
+        tools=[dict(BASH_TOOL)],
+        num_retries=2,
+    )
+
+    assert _replayed_reasoning_counts(upstream) == [1, 0]
+
+
 class _FailureLog(CustomLogger):
     def __init__(self) -> None:
         super().__init__()