Merge 3376a8464d into ef8eb98724

litellm/llms/deepseek/chat/transformation.py

@@ -72,6 +72,15 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
              (LiteLLM stores provider-specific response fields there).
           2. Otherwise inject a single space — the minimum value the API accepts.
         """
+        missing_reasoning_content: Final = any(
+            msg.get("role") == "assistant"
+            and not msg.get("reasoning_content")
+            and not (
+                isinstance(provider_fields := msg.get("provider_specific_fields"), dict)
+                and provider_fields.get("reasoning_content")
+            )
+            for msg in messages
+        )
         result: Final[list[AllMessageValues]] = []
         for msg in messages:
             if msg.get("role") == "assistant" and not msg.get("reasoning_content"):
@@ -84,20 +93,21 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
                     cleaned.pop("reasoning_content", None)
                     patched["provider_specific_fields"] = cleaned
                 else:
-                    litellm.verbose_logger.warning(
-                        "DeepSeek thinking mode: assistant message is missing "
-                        "`reasoning_content` and none was saved in "
-                        "`provider_specific_fields`. A single-space placeholder "
-                        "is being injected to satisfy API validation, but the "
-                        "model will receive a blank reasoning chain for this turn, "
-                        "which may silently degrade multi-turn response quality. "
-                        "Preserve `reasoning_content` from the original assistant "
-                        "response when building multi-turn conversation history."
-                    )
                     patched["reasoning_content"] = " "
                 result.append(cast(AllMessageValues, patched))
             else:
                 result.append(msg)
+        if missing_reasoning_content:
+            litellm.verbose_logger.warning(
+                "DeepSeek thinking mode: assistant message is missing "
+                "`reasoning_content` and none was saved in "
+                "`provider_specific_fields`. A single-space placeholder "
+                "is being injected to satisfy API validation, but the "
+                "model will receive a blank reasoning chain for this turn, "
+                "which may silently degrade multi-turn response quality. "
+                "Preserve `reasoning_content` from the original assistant "
+                "response when building multi-turn conversation history."
+            )
         return result
 
     @overload

tests/test_litellm/llms/deepseek/chat/test_deepseek_chat_transformation.py

@@ -1,3 +1,6 @@
+from copy import deepcopy
+from unittest.mock import patch
+
 from litellm.llms.deepseek.chat.transformation import DeepSeekChatConfig
 
 
@@ -103,6 +106,47 @@ async def test_async_transform_request_strips_unsupported_tools_from_body():
     assert body["tools"][0]["function"]["name"] == "shell"
 
 
+def test_transform_request_warns_once_per_replayed_history_and_preserves_history():
+    messages = [
+        {"role": "user", "content": "Use both tools."},
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [{"id": "first"}],
+            "reasoning_content": "",
+        },
+        {"role": "tool", "tool_call_id": "first", "content": "first result"},
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [{"id": "second"}],
+            "reasoning_content": "",
+        },
+        {"role": "tool", "tool_call_id": "second", "content": "second result"},
+        {"role": "user", "content": "Continue."},
+    ]
+    original_messages = deepcopy(messages)
+    config = DeepSeekChatConfig()
+
+    warning_path = "litellm.llms.deepseek.chat.transformation.litellm.verbose_logger.warning"
+    with patch(warning_path) as warning:
+        results = [
+            config.transform_request(
+                model="deepseek-reasoner",
+                messages=messages,
+                optional_params={"thinking": {"type": "enabled"}},
+                litellm_params={},
+                headers={},
+            )
+            for _ in range(2)
+        ]
+
+    assert warning.call_count == 2
+    for result in results:
+        assert [result["messages"][index]["reasoning_content"] for index in (1, 3)] == [" ", " "]
+    assert messages == original_messages
+
+
 def test_thinking_mode_active_bool_thinking_returns_false_without_crashing():
     config = DeepSeekChatConfig()
     assert config._thinking_mode_active(model="deepseek-reasoner", optional_params={"thinking": True}) is False