From 435b103b6977c0797661202442b0e34d4d4d3c51 Mon Sep 17 00:00:00 2001
From: kime541200 <kime541200@outlook.com>
Date: Fri, 24 Apr 2026 12:41:29 +0800
Subject: [PATCH] fix(anthropic): drop output_config for Azure

Prevent the Anthropic pass-through adapter from forwarding output_config
after it has already been translated to reasoning_effort for Azure calls.

Add a regression test for azure/gpt-5.2-chat adaptive thinking so the
translated reasoning parameter is preserved without leaking the
unsupported output_config field.

Made-with: Cursor
---
 .../adapters/handler.py                       |  2 +-
 .../test_reasoning_effort_fields.py           | 24 +++++++++++++++++++
 2 files changed, 25 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/handler.py b/litellm/llms/anthropic/experimental_pass_through/adapters/handler.py
index d16f5afb45c..c34e08f92b0 100644
--- a/litellm/llms/anthropic/experimental_pass_through/adapters/handler.py
+++ b/litellm/llms/anthropic/experimental_pass_through/adapters/handler.py
@@ -225,7 +225,7 @@ class LiteLLMMessagesToCompletionTransformationHandler:
                 "include_usage": True,
             }
 
-        excluded_keys = {"anthropic_messages"}
+        excluded_keys = {"anthropic_messages", "output_config"}
         extra_kwargs = extra_kwargs or {}
         for key, value in extra_kwargs.items():
             if (
diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/test_reasoning_effort_fields.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/test_reasoning_effort_fields.py
index d42d109f21b..48a6514d363 100644
--- a/tests/test_litellm/llms/anthropic/experimental_pass_through/test_reasoning_effort_fields.py
+++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/test_reasoning_effort_fields.py
@@ -261,6 +261,30 @@ class TestAdapterAdaptiveThinking:
         else:
             assert re == "high"
 
+    def test_messages_handler_does_not_forward_output_config_to_azure(self):
+        """Adaptive thinking should translate output_config for Azure without leaking it."""
+        from litellm.llms.anthropic.experimental_pass_through.adapters.handler import (
+            LiteLLMMessagesToCompletionTransformationHandler,
+        )
+
+        completion_kwargs, _ = (
+            LiteLLMMessagesToCompletionTransformationHandler._prepare_completion_kwargs(
+                max_tokens=1024,
+                messages=[{"role": "user", "content": "hello"}],
+                model="azure/gpt-5.2-chat",
+                thinking={"type": "adaptive"},
+                extra_kwargs={
+                    "api_version": "2025-01-01-preview",
+                    "drop_params": True,
+                    "output_config": {"effort": "high"},
+                },
+            )
+        )
+
+        assert completion_kwargs["reasoning_effort"] == "high"
+        assert completion_kwargs["api_version"] == "2025-01-01-preview"
+        assert "output_config" not in completion_kwargs
+
     def test_responses_adapter_adaptive_with_output_config(self):
         """Responses adapter: adaptive thinking + output_config.effort."""
         from litellm.llms.anthropic.experimental_pass_through.responses_adapters.transformation import (