fix: merge system messages and place them first to comply with Qwen/v… (#39136)

dify-agent/src/dify_agent/adapters/llm/model.py

@@ -304,16 +304,40 @@ def _map_messages_to_prompt_messages(
         for part in (Model._get_instruction_parts(messages, model_request_parameters) or [])
         if part.content.strip()
     ]
-    if instruction_messages:
-        insert_at = next(
-            (index for index, message in enumerate(prompt_messages) if not isinstance(message, SystemPromptMessage)),
-            len(prompt_messages),
-        )
-        prompt_messages[insert_at:insert_at] = instruction_messages
+    prompt_messages = _order_system_messages_first(prompt_messages, instruction_messages)
 
     return prompt_messages
 
 
+def _order_system_messages_first(
+    prompt_messages: Sequence[PromptMessage],
+    instruction_messages: Sequence[SystemPromptMessage],
+) -> list[PromptMessage]:
+    """Merge all system content into a single leading system message.
+
+    Some providers (e.g. vLLM serving Qwen3.5/3.6 chat templates) reject any
+    system message that is not exactly the first message, so sorting alone is
+    not enough: multiple system messages must be merged into one.
+    """
+    system_contents: list[str] = []
+    non_system_messages: list[PromptMessage] = []
+    for message in prompt_messages:
+        if isinstance(message, SystemPromptMessage):
+            text = message.get_text_content()
+            if text.strip():
+                system_contents.append(text)
+        else:
+            non_system_messages.append(message)
+    for instruction in instruction_messages:
+        text = instruction.get_text_content()
+        if text.strip():
+            system_contents.append(text)
+
+    if not system_contents:
+        return non_system_messages
+    return [SystemPromptMessage(content="\n\n".join(system_contents)), *non_system_messages]
+
+
 def _map_model_request_to_prompt_messages(message: ModelRequest) -> list[PromptMessage]:
     prompt_messages: list[PromptMessage] = []
 

dify-agent/tests/local/dify_agent/adapters/llm/test_model.py

@@ -171,12 +171,11 @@ class DifyLLMAdapterModelTests(unittest.IsolatedAsyncioTestCase):
             self.assertEqual(set(tools_by_name), {"weather", "incident_summary"})
             self.assertEqual(tools_by_name["incident_summary"]["parameters"]["required"], ["title"])
             self.assertEqual(data["prompt_messages"][0]["role"], "system")
-            self.assertEqual(data["prompt_messages"][0]["content"], "request system")
-            self.assertEqual(data["prompt_messages"][1]["content"], "be concise")
-            self.assertEqual(data["prompt_messages"][2]["content"], "hello")
+            self.assertEqual(data["prompt_messages"][0]["content"], "request system\n\nbe concise")
+            self.assertEqual(data["prompt_messages"][1]["content"], "hello")
+            self.assertEqual(data["prompt_messages"][2]["role"], "tool")
             self.assertEqual(data["prompt_messages"][3]["role"], "tool")
-            self.assertEqual(data["prompt_messages"][4]["role"], "tool")
-            self.assertEqual(data["prompt_messages"][5]["role"], "assistant")
+            self.assertEqual(data["prompt_messages"][4]["role"], "assistant")
             return build_stream_response(
                 LLMResultChunk(
                     model="demo-model",
@@ -291,6 +290,78 @@ class DifyLLMAdapterModelTests(unittest.IsolatedAsyncioTestCase):
         self.assertEqual(usage.time_to_first_token, 0.2)
         self.assertEqual(usage.time_to_generate, 0.6)
 
+    async def test_request_merges_system_messages_before_history(self) -> None:
+        messages = [
+            ModelRequest(parts=[UserPromptPart("previous user")]),
+            ModelResponse(parts=[TextPart(content="previous answer")]),
+            ModelRequest(parts=[SystemPromptPart("current system"), UserPromptPart("current user")]),
+        ]
+        request_parameters = ModelRequestParameters(instruction_parts=[InstructionPart(content="runtime instruction")])
+
+        def handler(request: httpx.Request) -> httpx.Response:
+            payload = json.loads(request.content.decode("utf-8"))
+            prompt_messages = payload["target"]["prompt_messages"]
+
+            self.assertEqual(
+                [message["role"] for message in prompt_messages],
+                ["system", "user", "assistant", "user"],
+            )
+            self.assertEqual(prompt_messages[0]["content"], "current system\n\nruntime instruction")
+            self.assertEqual(prompt_messages[1]["content"], "previous user")
+            self.assertEqual(prompt_messages[2]["content"], "previous answer")
+            self.assertEqual(prompt_messages[3]["content"], "current user")
+            return build_stream_response(*single_text_chunk("adapter response", prompt_tokens=11, completion_tokens=7))
+
+        async with self.mock_gateway_stream(httpx.MockTransport(handler)):
+            adapter = DifyLLMAdapterModel(
+                "demo-model",
+                self.make_provider(),
+                model_provider="openai",
+            )
+
+            response = await adapter.request(
+                messages,
+                model_settings=None,
+                model_request_parameters=request_parameters,
+            )
+
+        self.assertEqual(response.model_name, "demo-model")
+        self.assertEqual(cast(TextPart, response.parts[0]).content, "adapter response")
+
+    async def test_request_merges_scattered_system_messages_without_instructions(self) -> None:
+        messages = [
+            ModelRequest(parts=[SystemPromptPart("first system"), UserPromptPart("hello")]),
+            ModelResponse(parts=[TextPart(content="answer")]),
+            ModelRequest(parts=[SystemPromptPart("second system"), UserPromptPart("follow up")]),
+        ]
+
+        def handler(request: httpx.Request) -> httpx.Response:
+            payload = json.loads(request.content.decode("utf-8"))
+            prompt_messages = payload["target"]["prompt_messages"]
+
+            self.assertEqual(
+                [message["role"] for message in prompt_messages],
+                ["system", "user", "assistant", "user"],
+            )
+            self.assertEqual(prompt_messages[0]["content"], "first system\n\nsecond system")
+            return build_stream_response(*single_text_chunk("adapter response", prompt_tokens=11, completion_tokens=7))
+
+        async with self.mock_gateway_stream(httpx.MockTransport(handler)):
+            adapter = DifyLLMAdapterModel(
+                "demo-model",
+                self.make_provider(),
+                model_provider="openai",
+            )
+
+            response = await adapter.request(
+                messages,
+                model_settings=None,
+                model_request_parameters=ModelRequestParameters(),
+            )
+
+        self.assertEqual(response.model_name, "demo-model")
+        self.assertEqual(cast(TextPart, response.parts[0]).content, "adapter response")
+
     async def test_request_maps_tool_call_only_assistant_history_to_empty_string_content(self) -> None:
         messages = [
             ModelRequest(parts=[SystemPromptPart("request system"), UserPromptPart("hello")]),