Insert Instructions at the end of System Prompt

pydantic_ai_slim/pydantic_ai/models/anthropic.py

@@ -853,7 +853,7 @@ async def _map_message(  # noqa: C901
             else:
                 assert_never(m)
         if instructions := self._get_instructions(messages, model_request_parameters):
-            system_prompt_parts.insert(0, instructions)
+            system_prompt_parts.append(instructions)
         system_prompt = '\n\n'.join(system_prompt_parts)
 
         # Add cache_control to the last message content if anthropic_cache_messages is enabled

pydantic_ai_slim/pydantic_ai/models/bedrock.py

@@ -615,7 +615,7 @@ async def _map_messages(  # noqa: C901
             last_message = cast(dict[str, Any], current_message)
 
         if instructions := self._get_instructions(messages, model_request_parameters):
-            system_prompt.insert(0, {'text': instructions})
+            system_prompt.append({'text': instructions})
 
         return system_prompt, processed_messages
 

pydantic_ai_slim/pydantic_ai/models/cohere.py

@@ -271,7 +271,8 @@ def _map_messages(
             else:
                 assert_never(message)
         if instructions := self._get_instructions(messages, model_request_parameters):
-            cohere_messages.insert(0, SystemChatMessageV2(role='system', content=instructions))
+            system_prompt_count = sum(1 for m in cohere_messages if isinstance(m, SystemChatMessageV2))
+            cohere_messages.insert(system_prompt_count, SystemChatMessageV2(role='system', content=instructions))
         return cohere_messages
 
     def _get_tools(self, model_request_parameters: ModelRequestParameters) -> list[ToolV2]:

pydantic_ai_slim/pydantic_ai/models/gemini.py

@@ -363,7 +363,7 @@ async def _message_to_gemini_content(
             else:
                 assert_never(m)
         if instructions := self._get_instructions(messages, model_request_parameters):
-            sys_prompt_parts.insert(0, _GeminiTextPart(text=instructions))
+            sys_prompt_parts.append(_GeminiTextPart(text=instructions))
         return sys_prompt_parts, contents
 
     async def _map_user_prompt(self, part: UserPromptPart) -> list[_GeminiPartUnion]:

pydantic_ai_slim/pydantic_ai/models/google.py

@@ -588,7 +588,7 @@ async def _map_messages(
             contents = [{'role': 'user', 'parts': [{'text': ''}]}]
 
         if instructions := self._get_instructions(messages, model_request_parameters):
-            system_parts.insert(0, {'text': instructions})
+            system_parts.append({'text': instructions})
         system_instruction = ContentDict(role='user', parts=system_parts) if system_parts else None
 
         return system_instruction, contents

pydantic_ai_slim/pydantic_ai/models/groq.py

@@ -428,7 +428,10 @@ def _map_messages(
             else:
                 assert_never(message)
         if instructions := self._get_instructions(messages, model_request_parameters):
-            groq_messages.insert(0, chat.ChatCompletionSystemMessageParam(role='system', content=instructions))
+            system_prompt_count = sum(1 for m in groq_messages if m.get('role') == 'system')
+            groq_messages.insert(
+                system_prompt_count, chat.ChatCompletionSystemMessageParam(role='system', content=instructions)
+            )
         return groq_messages
 
     @staticmethod

pydantic_ai_slim/pydantic_ai/models/huggingface.py

@@ -361,7 +361,8 @@ async def _map_messages(
             else:
                 assert_never(message)
         if instructions := self._get_instructions(messages, model_request_parameters):
-            hf_messages.insert(0, ChatCompletionInputMessage(content=instructions, role='system'))  # type: ignore
+            system_prompt_count = sum(1 for m in hf_messages if getattr(m, 'role', None) == 'system')
+            hf_messages.insert(system_prompt_count, ChatCompletionInputMessage(content=instructions, role='system'))  # type: ignore
         return hf_messages
 
     @staticmethod

pydantic_ai_slim/pydantic_ai/models/mistral.py

@@ -557,7 +557,8 @@ def _map_messages(
             else:
                 assert_never(message)
         if instructions := self._get_instructions(messages, model_request_parameters):
-            mistral_messages.insert(0, MistralSystemMessage(content=instructions))
+            system_prompt_count = sum(1 for m in mistral_messages if isinstance(m, MistralSystemMessage))
+            mistral_messages.insert(system_prompt_count, MistralSystemMessage(content=instructions))
 
         # Post-process messages to insert fake assistant message after tool message if followed by user message
         # to work around `Unexpected role 'user' after role 'tool'` error.

pydantic_ai_slim/pydantic_ai/models/openai.py

@@ -840,7 +840,10 @@ async def _map_messages(
             else:
                 assert_never(message)
         if instructions := self._get_instructions(messages, model_request_parameters):
-            openai_messages.insert(0, chat.ChatCompletionSystemMessageParam(content=instructions, role='system'))
+            system_prompt_count = sum(1 for m in openai_messages if m.get('role') == 'system')
+            openai_messages.insert(
+                system_prompt_count, chat.ChatCompletionSystemMessageParam(content=instructions, role='system')
+            )
         return openai_messages
 
     @staticmethod
@@ -1313,7 +1316,10 @@ async def _responses_create(  # noqa: C901
             # > Response input messages must contain the word 'json' in some form to use 'text.format' of type 'json_object'.
             # Apparently they're only checking input messages for "JSON", not instructions.
             assert isinstance(instructions, str)
-            openai_messages.insert(0, responses.EasyInputMessageParam(role='system', content=instructions))
+            system_prompt_count = sum(1 for m in openai_messages if m.get('role') == 'system')
+            openai_messages.insert(
+                system_prompt_count, responses.EasyInputMessageParam(role='system', content=instructions)
+            )
             instructions = OMIT
 
         if verbosity := model_settings.get('openai_text_verbosity'):

tests/test_agent.py

@@ -6313,3 +6313,43 @@ def llm(messages: list[ModelMessage], _info: AgentInfo) -> ModelResponse:
             ]
         )
         assert run.all_messages_json().startswith(b'[{"parts":[{"content":"Hello",')
+
+
+def test_instructions_inserted_after_system_prompt():
+    """Tests that instructions are inserted after system prompts."""
+
+    agent = Agent('test')
+
+    @agent.system_prompt
+    def system_prompt_1() -> str:
+        return 'System prompt 1'
+
+    @agent.system_prompt
+    def system_prompt_2() -> str:
+        return 'System prompt 2'
+
+    @agent.instructions
+    def instructions() -> str:
+        return 'Instructions'
+
+    result = agent.run_sync('Hello')
+    assert result.all_messages()[0] == snapshot(
+        ModelRequest(
+            parts=[
+                SystemPromptPart(
+                    content='System prompt 1',
+                    timestamp=IsNow(tz=timezone.utc),
+                ),
+                SystemPromptPart(
+                    content='System prompt 2',
+                    timestamp=IsNow(tz=timezone.utc),
+                ),
+                UserPromptPart(
+                    content='Hello',
+                    timestamp=IsNow(tz=timezone.utc),
+                ),
+            ],
+            instructions='Instructions',
+            run_id=IsStr(),
+        )
+    )