fix(responses): normalize non-leading system messages in chat completion transformation

Pulls system messages to the beginning of the message list and merges
multiple system messages into a single leading block to prevent HTTP 400
errors from backends requiring strict chat template ordering.

Fixes #40693

litellm/responses/litellm_completion_transformation/transformation.py

@@ -447,7 +447,86 @@ class LiteLLMCompletionResponsesConfig:
             )
         )
 
-        return messages
+        return LiteLLMCompletionResponsesConfig._normalize_system_messages(messages)
+
+    @staticmethod
+    def _normalize_system_messages(
+        messages: list[
+            AllMessageValues
+            | GenericChatCompletionMessage
+            | ChatCompletionMessageToolCall
+            | ChatCompletionResponseMessage
+            | Message
+        ],
+    ) -> list[
+        AllMessageValues
+        | GenericChatCompletionMessage
+        | ChatCompletionMessageToolCall
+        | ChatCompletionResponseMessage
+        | Message
+    ]:
+        """
+        Normalize system messages so all system content appears at the beginning.
+
+        If multiple system messages exist, merge their contents into a single leading system message
+        to comply with backend chat templates that require at most one leading system message.
+        """
+
+        def _is_system(msg: object) -> bool:
+            if isinstance(msg, dict):
+                return msg.get("role") == "system"
+            elif hasattr(msg, "role"):
+                return msg.role == "system"
+            return False
+
+        system_messages: list[
+            AllMessageValues
+            | GenericChatCompletionMessage
+            | ChatCompletionMessageToolCall
+            | ChatCompletionResponseMessage
+            | Message
+        ] = [m for m in messages if _is_system(m)]
+        if not system_messages:
+            return messages
+
+        non_system_messages: list[
+            AllMessageValues
+            | GenericChatCompletionMessage
+            | ChatCompletionMessageToolCall
+            | ChatCompletionResponseMessage
+            | Message
+        ] = [m for m in messages if not _is_system(m)]
+
+        if len(system_messages) == 1:
+            if messages and _is_system(messages[0]):
+                return messages
+            return [system_messages[0]] + non_system_messages
+
+        merged_content_parts: list[str] = []
+        for sm in system_messages:
+            raw_content: object = None
+            if isinstance(sm, dict):
+                raw_content = sm.get("content")
+            elif hasattr(sm, "content"):
+                raw_content = sm.content
+
+            if isinstance(raw_content, str):
+                if raw_content:
+                    merged_content_parts.append(raw_content)
+            elif isinstance(raw_content, list):
+                for block in raw_content:
+                    if isinstance(block, str) and block:
+                        merged_content_parts.append(block)
+                    elif isinstance(block, dict):
+                        text = block.get("text")
+                        if isinstance(text, str) and text:
+                            merged_content_parts.append(text)
+
+        merged_system_message = ChatCompletionSystemMessage(
+            role="system",
+            content="\n\n".join(merged_content_parts),
+        )
+        return [merged_system_message] + non_system_messages
 
     @staticmethod
     async def async_responses_api_session_handler(

tests/test_litellm/responses/litellm_completion_transformation/test_system_message_normalization.py

@@ -0,0 +1,170 @@
+"""
+Tests for system message normalization in Responses API -> Chat Completion transformation.
+Regression tests for issue #40693: Anthropic /v1/messages -> Responses -> Chat Completions can emit non-leading system messages.
+"""
+
+from typing import Any
+import pytest
+from litellm.responses.litellm_completion_transformation.transformation import (
+    LiteLLMCompletionResponsesConfig,
+)
+
+
+def test_reproduce_issue_40693_non_leading_system_message() -> None:
+    """
+    Reproduces issue #40693:
+    When instructions are provided and the Responses input contains a system message
+    (e.g., Claude Code harness injecting skills/agent metadata after user prompt),
+    the resulting message sequence must NOT emit non-leading system messages.
+    All system messages must be normalized into a single leading system message.
+    """
+    responses_api_request = {
+        "instructions": "You are Claude Code, an AI assistant.",
+    }
+    input_items = [
+        {"role": "user", "content": "Hello, please help with this repo."},
+        {
+            "role": "system",
+            "content": "Available skills: [git, bash, edit]\nAvailable tools: [search]",
+        },
+    ]
+
+    messages = LiteLLMCompletionResponsesConfig.transform_responses_api_input_to_messages(
+        input=input_items,
+        responses_api_request=responses_api_request,
+    )
+
+    # 1. Exactly one leading system message at index 0
+    assert len(messages) == 2
+    assert messages[0]["role"] == "system"
+    assert messages[1]["role"] == "user"
+
+    # 2. No non-leading system messages
+    assert all(
+        (m.get("role") if isinstance(m, dict) else getattr(m, "role", None)) != "system"
+        for m in messages[1:]
+    )
+
+    # 3. Content from both instructions and subsequent system message are preserved
+    system_content = messages[0]["content"]
+    assert "You are Claude Code, an AI assistant." in system_content
+    assert "Available skills: [git, bash, edit]" in system_content
+
+
+def test_single_non_leading_system_message_moved_to_start() -> None:
+    """
+    When a single system message appears after a user message without instructions,
+    it should be moved to the beginning of the message list.
+    """
+    responses_api_request: dict[str, Any] = {}
+    input_items = [
+        {"role": "user", "content": "What is the weather?"},
+        {"role": "system", "content": "Respond only in metric units."},
+    ]
+
+    messages = LiteLLMCompletionResponsesConfig.transform_responses_api_input_to_messages(
+        input=input_items,
+        responses_api_request=responses_api_request,
+    )
+
+    assert len(messages) == 2
+    assert messages[0]["role"] == "system"
+    assert messages[0]["content"] == "Respond only in metric units."
+    assert messages[1]["role"] == "user"
+    assert messages[1]["content"] == "What is the weather?"
+
+
+def test_already_leading_system_message_unchanged() -> None:
+    """
+    When a single system message is already at the beginning, it should remain untouched.
+    """
+    responses_api_request: dict[str, Any] = {}
+    input_items = [
+        {"role": "system", "content": "System prompt."},
+        {"role": "user", "content": "User prompt."},
+    ]
+
+    messages = LiteLLMCompletionResponsesConfig.transform_responses_api_input_to_messages(
+        input=input_items,
+        responses_api_request=responses_api_request,
+    )
+
+    assert len(messages) == 2
+    assert messages[0]["role"] == "system"
+    assert messages[0]["content"] == "System prompt."
+    assert messages[1]["role"] == "user"
+    assert messages[1]["content"] == "User prompt."
+
+
+def test_no_system_message() -> None:
+    """
+    When no system message is provided, messages should remain unchanged.
+    """
+    responses_api_request: dict[str, Any] = {}
+    input_items = [
+        {"role": "user", "content": "Hello!"},
+    ]
+
+    messages = LiteLLMCompletionResponsesConfig.transform_responses_api_input_to_messages(
+        input=input_items,
+        responses_api_request=responses_api_request,
+    )
+
+    assert len(messages) == 1
+    assert messages[0]["role"] == "user"
+    assert messages[0]["content"] == "Hello!"
+
+
+def test_multiple_system_messages_with_structured_blocks() -> None:
+    """
+    Handles system messages with list content blocks (e.g. text/input_text blocks).
+    """
+    responses_api_request = {
+        "instructions": "Instruction text.",
+    }
+    input_items = [
+        {
+            "role": "system",
+            "content": [
+                {"type": "text", "text": "Structured system block 1."},
+                {"type": "text", "text": "Structured system block 2."},
+            ],
+        },
+        {"role": "user", "content": "Run tests."},
+    ]
+
+    messages = LiteLLMCompletionResponsesConfig.transform_responses_api_input_to_messages(
+        input=input_items,
+        responses_api_request=responses_api_request,
+    )
+
+    assert len(messages) == 2
+    assert messages[0]["role"] == "system"
+    assert messages[1]["role"] == "user"
+
+    system_content = messages[0]["content"]
+    assert "Instruction text." in system_content
+    assert "Structured system block 1." in system_content
+    assert "Structured system block 2." in system_content
+
+
+def test_transform_responses_api_request_to_chat_completion_request_normalizes_system() -> None:
+    """
+    Verifies end-to-end transformation via transform_responses_api_request_to_chat_completion_request.
+    """
+    request = LiteLLMCompletionResponsesConfig.transform_responses_api_request_to_chat_completion_request(
+        model="openai/qwen3.8-flash-next",
+        input=[
+            {"role": "user", "content": "Hello"},
+            {"role": "system", "content": "Follow instructions"},
+        ],
+        responses_api_request={"instructions": "Be helpful"},
+    )
+
+    messages = request["messages"]
+    assert len(messages) == 2
+    assert messages[0]["role"] == "system"
+    assert "Be helpful" in messages[0]["content"]
+    assert "Follow instructions" in messages[0]["content"]
+    assert messages[1]["role"] == "user"
+    assert messages[1]["content"] == "Hello"