From ea29b28f404417f0ea36be3db4af3dd314f74c12 Mon Sep 17 00:00:00 2001
From: Filippo Menghi <113345637+Cyberfilo@users.noreply.github.com>
Date: Wed, 10 Jun 2026 12:42:44 +0200
Subject: [PATCH] fix(sdk): stop mirroring reasoning_content in
 provider_specific_fields (#30110)

The dict-to-response conversion path mirrored reasoning_content into
provider_specific_fields, while live provider transforms (Anthropic's
_build_provider_specific_fields) only set it top-level on the Message.
Cache-replayed messages therefore serialized differently from live
ones, breaking disk cache key stability for multi-turn conversations
with extended thinking.

The mirror was added for DeepSeek before Message.reasoning_content
existed as a top-level attribute. The top-level field is still set by
the converter, so DeepSeek's request-side promotion is unaffected.

Fixes #27337.
---
 .../convert_dict_to_response.py               |  5 ---
 .../test_convert_dict_to_chat_completion.py   | 36 +++++++++++++++++++
 2 files changed, 36 insertions(+), 5 deletions(-)

diff --git a/litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py b/litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py
index 2547fd4d8c6..4e5b53a13d7 100644
--- a/litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py
+++ b/litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py
@@ -633,11 +633,6 @@ def convert_to_model_response_object(  # noqa: PLR0915
                         thinking_blocks = choice["message"]["thinking_blocks"]
                         provider_specific_fields["thinking_blocks"] = thinking_blocks
 
-                    if reasoning_content:
-                        provider_specific_fields["reasoning_content"] = (
-                            reasoning_content
-                        )
-
                     message = Message(
                         content=content,
                         role=choice["message"]["role"] or "assistant",
diff --git a/tests/llm_translation/test_llm_response_utils/test_convert_dict_to_chat_completion.py b/tests/llm_translation/test_llm_response_utils/test_convert_dict_to_chat_completion.py
index 01bcb1a247a..9a69f513069 100644
--- a/tests/llm_translation/test_llm_response_utils/test_convert_dict_to_chat_completion.py
+++ b/tests/llm_translation/test_llm_response_utils/test_convert_dict_to_chat_completion.py
@@ -2425,6 +2425,42 @@ class TestConvertToModelResponseObjectCompletion:
         assert result.choices[0].message.content == "The answer is 4."
         assert result.choices[0].message.reasoning_content == "2+2=4"
 
+    def test_reasoning_content_not_mirrored_into_provider_specific_fields(self):
+        """Mirroring reasoning_content into provider_specific_fields made
+        cache-replayed messages diverge from live Anthropic messages, which
+        only set it top-level, breaking cache key stability (issue #27337)."""
+        response_object = {
+            "id": "chatcmpl-5",
+            "model": "claude-sonnet-4-5",
+            "choices": [
+                {
+                    "finish_reason": "stop",
+                    "index": 0,
+                    "message": {
+                        "content": "The answer is 4.",
+                        "role": "assistant",
+                        "reasoning_content": "2+2=4",
+                        "thinking_blocks": [
+                            {
+                                "type": "thinking",
+                                "thinking": "2+2=4",
+                                "signature": "sig",
+                            }
+                        ],
+                    },
+                }
+            ],
+            "usage": {"prompt_tokens": 5, "completion_tokens": 10, "total_tokens": 15},
+        }
+
+        result = convert_to_model_response_object(
+            response_object=response_object,
+            model_response_object=ModelResponse(),
+        )
+        message = result.choices[0].message
+        assert message.reasoning_content == "2+2=4"
+        assert "reasoning_content" not in (message.provider_specific_fields or {})
+
     def test_response_none_raises(self):
         with pytest.raises(Exception):
             convert_to_model_response_object(