From 9c2f9e95c1cabe861d3031cb8479ad6b611f5901 Mon Sep 17 00:00:00 2001
From: Christian Sidak <csidak@gmail.com>
Date: Thu, 9 Apr 2026 19:51:25 -0700
Subject: [PATCH] fix(azure_ai): strip max_retries from Azure AI Anthropic
 /v1/messages request body (#18746)

Azure AI Foundry's Anthropic endpoint rejects max_retries, extra_body,
and stream_options with "Extra inputs are not permitted". These params
can leak into the request body from the router's default litellm params.
The chat/completions path already strips them (transformation.py); this
applies the same fix to the /v1/messages path.

Fixes https://github.com/BerriAI/litellm/issues/18746
---
 .../anthropic/messages_transformation.py      |  7 ++
 ...azure_anthropic_messages_transformation.py | 69 ++++++++++++++++++-
 2 files changed, 73 insertions(+), 3 deletions(-)

diff --git a/litellm/llms/azure_ai/anthropic/messages_transformation.py b/litellm/llms/azure_ai/anthropic/messages_transformation.py
index 59d8fb02c6d..377b5f20b02 100644
--- a/litellm/llms/azure_ai/anthropic/messages_transformation.py
+++ b/litellm/llms/azure_ai/anthropic/messages_transformation.py
@@ -1,6 +1,7 @@
 """
 Azure Anthropic messages transformation config - extends AnthropicMessagesConfig with Azure authentication
 """
+
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
 
 from litellm.llms.anthropic.experimental_pass_through.messages.transformation import (
@@ -163,4 +164,10 @@ class AzureAnthropicMessagesConfig(AnthropicMessagesConfig):
             headers=headers,
         )
         self._remove_scope_from_cache_control(anthropic_messages_request)
+
+        # Remove parameters that Azure AI Foundry's Anthropic endpoint rejects.
+        # These may leak into the request body from the router's default litellm
+        # params or from the proxy layer.
+        for unsupported_param in ("max_retries", "extra_body", "stream_options"):
+            anthropic_messages_request.pop(unsupported_param, None)
         return anthropic_messages_request
diff --git a/tests/test_litellm/llms/azure_ai/claude/test_azure_anthropic_messages_transformation.py b/tests/test_litellm/llms/azure_ai/claude/test_azure_anthropic_messages_transformation.py
index 83653bc037b..90659af9313 100644
--- a/tests/test_litellm/llms/azure_ai/claude/test_azure_anthropic_messages_transformation.py
+++ b/tests/test_litellm/llms/azure_ai/claude/test_azure_anthropic_messages_transformation.py
@@ -59,7 +59,9 @@ class TestAzureAnthropicMessagesConfig:
             assert result["x-api-key"] == "test-api-key"
             assert "api-key" not in result
 
-    def test_validate_anthropic_messages_environment_converts_api_key_to_x_api_key(self):
+    def test_validate_anthropic_messages_environment_converts_api_key_to_x_api_key(
+        self,
+    ):
         """Test that api-key header is converted to x-api-key"""
         config = AzureAnthropicMessagesConfig()
         headers = {}
@@ -231,7 +233,7 @@ class TestAzureAnthropicMessagesConfig:
         config = AzureAnthropicMessagesConfig()
         model = "claude-sonnet-4-5"
         params = config.get_supported_anthropic_messages_params(model)
-        
+
         assert "messages" in params
         assert "model" in params
         assert "max_tokens" in params
@@ -281,7 +283,9 @@ class TestAzureAnthropicMessagesConfig:
         assert "scope" not in result["system"][0]["cache_control"]
         assert result["system"][0]["cache_control"]["type"] == "ephemeral"
         assert "scope" not in result["messages"][0]["content"][0]["cache_control"]
-        assert result["messages"][0]["content"][0]["cache_control"]["type"] == "ephemeral"
+        assert (
+            result["messages"][0]["content"][0]["cache_control"]["type"] == "ephemeral"
+        )
 
 
 class TestProviderConfigManagerAzureAnthropicMessages:
@@ -327,3 +331,62 @@ class TestProviderConfigManagerAzureAnthropicMessages:
         )
 
         assert config is None
+
+
+class TestAzureAnthropicMessagesTransformRequest:
+    """Test that unsupported parameters are stripped from the request body (issue #18746)"""
+
+    def test_transform_request_removes_max_retries(self):
+        """Test that max_retries is removed from the request body.
+
+        Regression test for https://github.com/BerriAI/litellm/issues/18746
+        Azure AI Foundry rejects max_retries with 'Extra inputs are not permitted'.
+        """
+        config = AzureAnthropicMessagesConfig()
+        model = "claude-haiku-4-5"
+        messages = [{"role": "user", "content": [{"type": "text", "text": "hi"}]}]
+        optional_params = {
+            "max_tokens": 1000,
+            "max_retries": 0,
+        }
+        litellm_params = GenericLiteLLMParams()
+        headers = {}
+
+        result = config.transform_anthropic_messages_request(
+            model=model,
+            messages=messages,
+            anthropic_messages_optional_request_params=optional_params,
+            litellm_params=litellm_params,
+            headers=headers,
+        )
+
+        assert "max_retries" not in result
+        assert result["model"] == model
+        assert result["max_tokens"] == 1000
+        assert result["messages"] == messages
+
+    def test_transform_request_removes_extra_body_and_stream_options(self):
+        """Test that extra_body and stream_options are also removed."""
+        config = AzureAnthropicMessagesConfig()
+        model = "claude-haiku-4-5"
+        messages = [{"role": "user", "content": [{"type": "text", "text": "hi"}]}]
+        optional_params = {
+            "max_tokens": 1000,
+            "extra_body": {"foo": "bar"},
+            "stream_options": {"include_usage": True},
+        }
+        litellm_params = GenericLiteLLMParams()
+        headers = {}
+
+        result = config.transform_anthropic_messages_request(
+            model=model,
+            messages=messages,
+            anthropic_messages_optional_request_params=optional_params,
+            litellm_params=litellm_params,
+            headers=headers,
+        )
+
+        assert "extra_body" not in result
+        assert "stream_options" not in result
+        assert result["model"] == model
+        assert result["max_tokens"] == 1000