From 5313a9c67f59a16809462b77d8379e0d4075416c Mon Sep 17 00:00:00 2001
From: nehaaprasaad <neh960112rxs@gmail.com>
Date: Thu, 7 May 2026 07:56:42 +0530
Subject: [PATCH] fix(cache): align Anthropic reasoning_content with live
 response on replay

---
 .../convert_dict_to_response.py               |  5 +-
 .../test_convert_dict_to_chat_completion.py   | 46 ++++++++++++++++++-
 2 files changed, 47 insertions(+), 4 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 5fd42fe0d36..b0ba237c627 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
@@ -595,10 +595,9 @@ def convert_to_model_response_object(  # noqa: PLR0915
                         thinking_blocks = choice["message"]["thinking_blocks"]
                         provider_specific_fields["thinking_blocks"] = thinking_blocks
 
+                    # Align with live provider responses: reasoning_content is message-level only.
                     if reasoning_content:
-                        provider_specific_fields["reasoning_content"] = (
-                            reasoning_content
-                        )
+                        provider_specific_fields.pop("reasoning_content", None)
 
                     message = Message(
                         content=content,
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 66a1a4d74af..944ee275369 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
@@ -902,7 +902,51 @@ def test_convert_to_model_response_object_with_thinking_content():
 
     resp: ModelResponse = convert_to_model_response_object(**args)
     assert resp is not None
-    assert resp.choices[0].message.reasoning_content is not None
+    msg = resp.choices[0].message
+    assert msg.reasoning_content is not None
+    assert "reasoning_content" not in (msg.provider_specific_fields or {})
+
+
+def test_convert_to_model_response_object_strips_reasoning_content_from_provider_specific_fields():
+    """
+    Stale or merged provider_specific_fields must not duplicate top-level reasoning_content.
+    Duplication breaks disk cache keys on later turns (see Anthropic extended thinking).
+    """
+    duplicate = "same reasoning"
+    args = {
+        "response_object": {
+            "id": "chatcmpl-dup-reasoning",
+            "created": 1741057687,
+            "model": "claude-4-sonnet-20250514",
+            "object": "chat.completion",
+            "choices": [
+                {
+                    "finish_reason": "stop",
+                    "index": 0,
+                    "message": {
+                        "content": "Answer.",
+                        "role": "assistant",
+                        "reasoning_content": duplicate,
+                        "provider_specific_fields": {
+                            "citations": None,
+                            "reasoning_content": duplicate,
+                            "thinking_blocks": [],
+                        },
+                    },
+                }
+            ],
+            "usage": {
+                "completion_tokens": 1,
+                "prompt_tokens": 1,
+                "total_tokens": 2,
+            },
+        },
+        "model_response_object": ModelResponse(),
+    }
+    resp = convert_to_model_response_object(**args)
+    msg = resp.choices[0].message
+    assert msg.reasoning_content == duplicate
+    assert "reasoning_content" not in (msg.provider_specific_fields or {})
 
 
 def test_convert_to_model_response_object_with_empty_error_object():