add support for instructions parameter in response object

llama_stack/apis/agents/openai_responses.py

@@ -327,6 +327,35 @@ OpenAIResponseOutput = Annotated[
 register_schema(OpenAIResponseOutput, name="OpenAIResponseOutput")
 
 
+@json_schema_type
+class OpenAIResponseInputFunctionToolCallOutput(BaseModel):
+    """
+    This represents the output of a function call that gets passed back to the model.
+    """
+
+    call_id: str
+    output: str
+    type: Literal["function_call_output"] = "function_call_output"
+    id: str | None = None
+    status: str | None = None
+
+
+OpenAIResponseInput = Annotated[
+    # Responses API allows output messages to be passed in as input
+    OpenAIResponseOutputMessageWebSearchToolCall
+    | OpenAIResponseOutputMessageFileSearchToolCall
+    | OpenAIResponseOutputMessageFunctionToolCall
+    | OpenAIResponseInputFunctionToolCallOutput
+    | OpenAIResponseMCPApprovalRequest
+    | OpenAIResponseMCPApprovalResponse
+    | OpenAIResponseOutputMessageMCPCall
+    | OpenAIResponseOutputMessageMCPListTools
+    | OpenAIResponseMessage,
+    Field(union_mode="left_to_right"),
+]
+register_schema(OpenAIResponseInput, name="OpenAIResponseInput")
+
+
 # This has to be a TypedDict because we need a "schema" field and our strong
 # typing code in the schema generator doesn't support Pydantic aliases. That also
 # means we can't use a discriminator field here, because TypedDicts don't support
@@ -545,6 +574,7 @@ class OpenAIResponseObject(BaseModel):
     :param tools: (Optional) An array of tools the model may call while generating a response.
     :param truncation: (Optional) Truncation strategy applied to the response
     :param usage: (Optional) Token usage information for the response
+    :param instructions: (Optional) System message inserted into the model's context
     """
 
     created_at: int
@@ -564,6 +594,7 @@ class OpenAIResponseObject(BaseModel):
     tools: list[OpenAIResponseTool] | None = None
     truncation: str | None = None
     usage: OpenAIResponseUsage | None = None
+    instructions: str | list[OpenAIResponseInput] | None = None
 
 
 @json_schema_type
@@ -1237,35 +1268,6 @@ OpenAIResponseObjectStream = Annotated[
 register_schema(OpenAIResponseObjectStream, name="OpenAIResponseObjectStream")
 
 
-@json_schema_type
-class OpenAIResponseInputFunctionToolCallOutput(BaseModel):
-    """
-    This represents the output of a function call that gets passed back to the model.
-    """
-
-    call_id: str
-    output: str
-    type: Literal["function_call_output"] = "function_call_output"
-    id: str | None = None
-    status: str | None = None
-
-
-OpenAIResponseInput = Annotated[
-    # Responses API allows output messages to be passed in as input
-    OpenAIResponseOutputMessageWebSearchToolCall
-    | OpenAIResponseOutputMessageFileSearchToolCall
-    | OpenAIResponseOutputMessageFunctionToolCall
-    | OpenAIResponseInputFunctionToolCallOutput
-    | OpenAIResponseMCPApprovalRequest
-    | OpenAIResponseMCPApprovalResponse
-    | OpenAIResponseOutputMessageMCPCall
-    | OpenAIResponseOutputMessageMCPListTools
-    | OpenAIResponseMessage,
-    Field(union_mode="left_to_right"),
-]
-register_schema(OpenAIResponseInput, name="OpenAIResponseInput")
-
-
 class ListOpenAIResponseInputItem(BaseModel):
     """List container for OpenAI response input items.
 

llama_stack/providers/inline/agents/meta_reference/responses/openai_responses.py

@@ -123,6 +123,17 @@ class OpenAIResponsesImpl:
                 # Use stored messages directly and convert only new input
                 message_adapter = TypeAdapter(list[OpenAIMessageParam])
                 messages = message_adapter.validate_python(previous_response.messages)
+                # When managing conversation state with the previous_response_id parameter,
+                # the instructions used on previous turns will not be carried over in the context
+                previous_instructions = previous_response.instructions
+                if previous_instructions:
+                    if (isinstance(previous_instructions, str) and
+                        previous_instructions == messages[0].content and
+                        messages[0].role == "system"):
+                        # Omit instructions from previous response
+                        del messages[0]
+                    else:
+                        raise ValueError("Instructions from the previous response could not be validated")
                 new_messages = await convert_response_input_to_chat_messages(input, previous_messages=messages)
                 messages.extend(new_messages)
             else:
@@ -359,6 +370,7 @@ class OpenAIResponsesImpl:
             tool_executor=self.tool_executor,
             safety_api=self.safety_api,
             guardrail_ids=guardrail_ids,
+            instructions=instructions,
         )
 
         # Stream the response

llama_stack/providers/inline/agents/meta_reference/responses/streaming.py

@@ -112,6 +112,7 @@ class StreamingResponseOrchestrator:
         tool_executor,  # Will be the tool execution logic from the main class
         safety_api,
         guardrail_ids: list[str] | None = None,
+        instructions: str,
     ):
         self.inference_api = inference_api
         self.ctx = ctx
@@ -133,6 +134,8 @@ class StreamingResponseOrchestrator:
         self.accumulated_usage: OpenAIResponseUsage | None = None
         # Track if we've sent a refusal response
         self.violation_detected = False
+        # system message that is inserted into the model's context
+        self.instructions = instructions
 
     async def _create_refusal_response(self, violation_message: str) -> OpenAIResponseObjectStream:
         """Create a refusal response to replace streaming content."""
@@ -176,6 +179,7 @@ class StreamingResponseOrchestrator:
             tools=self.ctx.available_tools(),
             error=error,
             usage=self.accumulated_usage,
+            instructions=self.instructions,
         )
 
     async def create_response(self) -> AsyncIterator[OpenAIResponseObjectStream]: