fix(responses-bridge): map system-only chat request to system input item

litellm/completion_extras/litellm_responses_transformation/transformation.py

@@ -402,6 +402,20 @@ def transform_request(
             instructions,
         ) = self.convert_chat_completion_messages_to_responses_api(messages)
 
+        # OpenAI's Responses API rejects an empty input. For a system-only
+        # request, carry the system message as a system-role input item instead
+        # of instructions, mirroring how non-string system content is already
+        # handled in convert_chat_completion_messages_to_responses_api.
+        if not input_items and instructions is not None:
+            input_items = [
+                {
+                    "type": "message",
+                    "role": "system",
+                    "content": [{"type": "input_text", "text": instructions}],
+                }
+            ]
+            instructions = None
+
         optional_params = self._extract_extra_body_params(optional_params)
 
         # Build responses API request using the reverse transformation logic

tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py

@@ -13,6 +13,9 @@
     0, os.path.abspath("../../..")
 )  # Adds the parent directory to the system-path
 import litellm
+from litellm.completion_extras.litellm_responses_transformation.transformation import (
+    LiteLLMResponsesTransformationHandler,
+)
 
 
 def test_convert_chat_completion_messages_to_responses_api_image_input():
@@ -860,6 +863,39 @@ def test_extract_extra_body_params_reasoning_effort_override():
     assert "extra_body" not in result
 
 
+def test_transform_request_system_only_message_maps_to_system_input_item():
+    """System-only requests must not send input=[] to the Responses API.
+
+    OpenAI rejects both input=[] and input="". When the only message is a
+    system message, carry it as a system-role input item (single copy, correct
+    role) rather than leaving input empty or duplicating it into instructions.
+    """
+    handler = LiteLLMResponsesTransformationHandler()
+    logging_obj = Mock()
+    messages = [{"role": "system", "content": "You are a helpful assistant."}]
+
+    result = handler.transform_request(
+        model="gpt-5.3-codex",
+        messages=messages,
+        optional_params={},
+        litellm_params={},
+        headers={},
+        litellm_logging_obj=logging_obj,
+    )
+
+    assert result["input"] == [
+        {
+            "type": "message",
+            "role": "system",
+            "content": [
+                {"type": "input_text", "text": "You are a helpful assistant."}
+            ],
+        }
+    ]
+    # System content lives in input only; not duplicated into instructions.
+    assert not result.get("instructions")
+
+
 def test_transform_request_single_char_keys_not_matched():
     """Test that single-character keys are not incorrectly matched to 'metadata' or 'previous_response_id'