Propagate cache_control through Responses API → Chat Completion transform

litellm/responses/litellm_completion_transformation/transformation.py

@@ -988,14 +988,16 @@ class LiteLLMCompletionResponsesConfig:
             # Since guardrails skip None content anyway, we return empty list to exclude it from structured messages
             if content is None:
                 return []
-            return [
-                GenericChatCompletionMessage(
-                    role=input_item.get("role") or "user",
-                    content=LiteLLMCompletionResponsesConfig._transform_responses_api_content_to_chat_completion_content(
-                        content
-                    ),
-                )
-            ]
+            message = GenericChatCompletionMessage(
+                role=input_item.get("role") or "user",
+                content=LiteLLMCompletionResponsesConfig._transform_responses_api_content_to_chat_completion_content(
+                    content
+                ),
+            )
+            cache_control = input_item.get("cache_control")
+            if cache_control is not None:
+                message["cache_control"] = cache_control  # type: ignore[typeddict-unknown-key]
+            return [message]
 
     @staticmethod
     def _is_input_item_tool_call_output(input_item: Any) -> bool:
@@ -1294,14 +1296,16 @@ class LiteLLMCompletionResponsesConfig:
                         text_value = item.get("text")
                         if text_value is None:
                             continue
-                        content_list.append(
-                            {
-                                "type": LiteLLMCompletionResponsesConfig._get_chat_completion_request_content_type(
-                                    item.get("type") or "text"
-                                ),
-                                "text": text_value,
-                            }
-                        )
+                        block = {
+                            "type": LiteLLMCompletionResponsesConfig._get_chat_completion_request_content_type(
+                                item.get("type") or "text"
+                            ),
+                            "text": text_value,
+                        }
+                        cache_control = item.get("cache_control")
+                        if cache_control is not None:
+                            block["cache_control"] = cache_control
+                        content_list.append(block)
             return content_list
         else:
             raise ValueError(f"Invalid content type: {type(content)}")

tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py

@@ -992,6 +992,58 @@ class TestContentTypeTransformation:
         assert result[0]["text"] == "valid text"
         assert result[1]["text"] == "another valid"
 
+    def test_cache_control_propagated_to_content_block(self):
+        """
+        Test that a cache_control field on a Responses API content item is
+        propagated onto the Chat Completion content block.
+        """
+        cache_control = {"type": "ephemeral"}
+        content = [
+            {"type": "text", "text": "cached text", "cache_control": cache_control},
+            {"type": "text", "text": "uncached text"},
+        ]
+        result = LiteLLMCompletionResponsesConfig._transform_responses_api_content_to_chat_completion_content(
+            content
+        )
+        assert len(result) == 2
+        assert result[0]["text"] == "cached text"
+        assert result[0]["cache_control"] == cache_control
+        assert result[1]["text"] == "uncached text"
+        assert "cache_control" not in result[1]
+
+    def test_cache_control_propagated_to_message(self):
+        """
+        Test that a cache_control field on a Responses API input item is
+        propagated onto the resulting GenericChatCompletionMessage.
+        """
+        cache_control = {"type": "ephemeral"}
+        input_item = {
+            "role": "user",
+            "content": [{"type": "input_text", "text": "hello"}],
+            "cache_control": cache_control,
+        }
+        result = LiteLLMCompletionResponsesConfig._transform_responses_api_input_item_to_chat_completion_message(
+            input_item
+        )
+        assert len(result) == 1
+        assert result[0]["role"] == "user"
+        assert result[0]["cache_control"] == cache_control
+
+    def test_cache_control_absent_when_not_provided(self):
+        """
+        Test that no cache_control key is added when the input item does not
+        carry one.
+        """
+        input_item = {
+            "role": "user",
+            "content": [{"type": "input_text", "text": "hello"}],
+        }
+        result = LiteLLMCompletionResponsesConfig._transform_responses_api_input_item_to_chat_completion_message(
+            input_item
+        )
+        assert len(result) == 1
+        assert "cache_control" not in result[0]
+
 
 class TestToolTransformation:
     """Test cases for tool transformation from Responses API to Chat Completion format"""