From 33f60b640079b2384864a7608325478ed591e641 Mon Sep 17 00:00:00 2001
From: jesco-absolut <team@srswti.com>
Date: Fri, 3 Jul 2026 21:19:21 -0400
Subject: [PATCH] fix(proxy): preserve reasoning fields on streaming cache hits

---
 .../convert_dict_to_response.py               |  7 +--
 .../test_convert_dict_to_chat_completion.py   | 54 +++++++++++++++++++
 2 files changed, 55 insertions(+), 6 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 a6f10e1ede3..87e50a3046c 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
@@ -204,12 +204,7 @@ async def convert_to_streaming_response_async(
                     t["index"] = index
                 pydantic_tool_calls.append(ChatCompletionDeltaToolCall(**t))
             choice["message"]["tool_calls"] = pydantic_tool_calls
-        delta = Delta(
-            content=choice["message"].get("content", None),
-            role=choice["message"]["role"],
-            function_call=choice["message"].get("function_call", None),
-            tool_calls=choice["message"].get("tool_calls", None),
-        )
+        delta = Delta(**choice["message"])
         finish_reason = choice.get("finish_reason", None)
 
         if finish_reason is None:
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 5683d973ac9..f091b4453e9 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
@@ -2005,6 +2005,60 @@ class TestConvertToStreamingResponseAsync:
         assert chunks[-1].choices[0].finish_reason == "stop"
         assert chunks[-1].usage.prompt_tokens == 3
 
+    def test_preserves_reasoning_fields(self):
+        import asyncio
+
+        from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response import (
+            convert_to_streaming_response_async,
+        )
+
+        thinking_blocks = [
+            {
+                "type": "thinking",
+                "thinking": "cached reasoning",
+                "signature": "sig-cache",
+            }
+        ]
+        response_object = {
+            "id": "msg_async_reasoning_cache",
+            "model": "claude-3",
+            "created": 1700000000,
+            "choices": [
+                {
+                    "finish_reason": "stop",
+                    "index": 0,
+                    "message": {
+                        "content": "Final answer",
+                        "role": "assistant",
+                        "reasoning_content": "cached reasoning",
+                        "thinking_blocks": thinking_blocks,
+                    },
+                }
+            ],
+        }
+
+        async def run():
+            return [
+                chunk
+                async for chunk in convert_to_streaming_response_async(
+                    response_object=response_object
+                )
+            ]
+
+        chunks = asyncio.run(run())
+
+        assert (
+            "".join(c.choices[0].delta.content or "" for c in chunks) == "Final answer"
+        )
+        assert chunks[0].choices[0].delta.reasoning_content == "cached reasoning"
+        assert chunks[0].choices[0].delta.thinking_blocks == thinking_blocks
+        assert not any(
+            hasattr(c.choices[0].delta, "reasoning_content") for c in chunks[1:]
+        )
+        assert not any(
+            hasattr(c.choices[0].delta, "thinking_blocks") for c in chunks[1:]
+        )
+
 
 class TestHandleInvalidParallelToolCalls:
     def test_none_input(self):