feat: Structured output for Responses API (#2324)

# What does this PR do?

This adds the missing `text` parameter to the Responses API that is how
users control structured outputs. All we do with that parameter is map
it to the corresponding chat completion response_format.

## Test Plan

The new unit tests exercise the various permutations allowed for this
property, while a couple of new verification tests actually use it for
real to verify the model outputs are following the format as expected.

Unit tests:

`python -m pytest -s -v
tests/unit/providers/agents/meta_reference/test_openai_responses.py`

Verification tests:

```
llama stack run llama_stack/templates/together/run.yaml
pytest -s -vv 'tests/verifications/openai_api/test_responses.py' \
  --base-url=http://localhost:8321/v1/openai/v1 \
  --model meta-llama/Llama-4-Scout-17B-16E-Instruct
```

Note that the verification tests can only be run with a real Llama Stack
server (as opposed to using the library client via
`--provider=stack:together`) because the Llama Stack python client is
not yet updated to accept this text field.

Signed-off-by: Ben Browning <bbrownin@redhat.com>

llama_stack/apis/agents/agents.py

@@ -37,6 +37,7 @@ from .openai_responses import (
     OpenAIResponseInputTool,
     OpenAIResponseObject,
     OpenAIResponseObjectStream,
+    OpenAIResponseText,
 )
 
 # TODO: use enum.StrEnum when we drop support for python 3.10
@@ -603,6 +604,7 @@ class Agents(Protocol):
         store: bool | None = True,
         stream: bool | None = False,
         temperature: float | None = None,
+        text: OpenAIResponseText | None = None,
         tools: list[OpenAIResponseInputTool] | None = None,
         max_infer_iters: int | None = 10,  # this is an extension to the OpenAI API
     ) -> OpenAIResponseObject | AsyncIterator[OpenAIResponseObjectStream]:

llama_stack/apis/agents/openai_responses.py

@@ -7,6 +7,7 @@
 from typing import Annotated, Any, Literal
 
 from pydantic import BaseModel, Field
+from typing_extensions import TypedDict
 
 from llama_stack.schema_utils import json_schema_type, register_schema
 
@@ -126,6 +127,32 @@ OpenAIResponseOutput = Annotated[
 register_schema(OpenAIResponseOutput, name="OpenAIResponseOutput")
 
 
+# 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
+# default values which the strong typing code requires for discriminators.
+class OpenAIResponseTextFormat(TypedDict, total=False):
+    """Configuration for Responses API text format.
+
+    :param type: Must be "text", "json_schema", or "json_object" to identify the format type
+    :param name: The name of the response format. Only used for json_schema.
+    :param schema: The JSON schema the response should conform to. In a Python SDK, this is often a `pydantic` model. Only used for json_schema.
+    :param description: (Optional) A description of the response format. Only used for json_schema.
+    :param strict: (Optional) Whether to strictly enforce the JSON schema. If true, the response must match the schema exactly. Only used for json_schema.
+    """
+
+    type: Literal["text"] | Literal["json_schema"] | Literal["json_object"]
+    name: str | None
+    schema: dict[str, Any] | None
+    description: str | None
+    strict: bool | None
+
+
+@json_schema_type
+class OpenAIResponseText(BaseModel):
+    format: OpenAIResponseTextFormat | None = None
+
+
 @json_schema_type
 class OpenAIResponseObject(BaseModel):
     created_at: int
@@ -138,6 +165,9 @@ class OpenAIResponseObject(BaseModel):
     previous_response_id: str | None = None
     status: str
     temperature: float | None = None
+    # Default to text format to avoid breaking the loading of old responses
+    # before the field was added. New responses will have this set always.
+    text: OpenAIResponseText = OpenAIResponseText(format=OpenAIResponseTextFormat(type="text"))
     top_p: float | None = None
     truncation: str | None = None
     user: str | None = None

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

@@ -29,6 +29,7 @@ from llama_stack.apis.agents import (
     Session,
     Turn,
 )
+from llama_stack.apis.agents.openai_responses import OpenAIResponseText
 from llama_stack.apis.common.responses import PaginatedResponse
 from llama_stack.apis.inference import (
     Inference,
@@ -324,11 +325,12 @@ class MetaReferenceAgentsImpl(Agents):
         store: bool | None = True,
         stream: bool | None = False,
         temperature: float | None = None,
+        text: OpenAIResponseText | None = None,
         tools: list[OpenAIResponseInputTool] | None = None,
         max_infer_iters: int | None = 10,
     ) -> OpenAIResponseObject:
         return await self.openai_responses_impl.create_openai_response(
-            input, model, instructions, previous_response_id, store, stream, temperature, tools, max_infer_iters
+            input, model, instructions, previous_response_id, store, stream, temperature, text, tools, max_infer_iters
         )
 
     async def list_openai_responses(

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

@@ -37,6 +37,8 @@ from llama_stack.apis.agents.openai_responses import (
     OpenAIResponseOutputMessageFunctionToolCall,
     OpenAIResponseOutputMessageMCPListTools,
     OpenAIResponseOutputMessageWebSearchToolCall,
+    OpenAIResponseText,
+    OpenAIResponseTextFormat,
 )
 from llama_stack.apis.inference.inference import (
     Inference,
@@ -50,7 +52,12 @@ from llama_stack.apis.inference.inference import (
     OpenAIChoice,
     OpenAIDeveloperMessageParam,
     OpenAIImageURL,
+    OpenAIJSONSchema,
     OpenAIMessageParam,
+    OpenAIResponseFormatJSONObject,
+    OpenAIResponseFormatJSONSchema,
+    OpenAIResponseFormatParam,
+    OpenAIResponseFormatText,
     OpenAISystemMessageParam,
     OpenAIToolMessageParam,
     OpenAIUserMessageParam,
@@ -158,6 +165,21 @@ async def _convert_chat_choice_to_response_message(choice: OpenAIChoice) -> Open
     )
 
 
+async def _convert_response_text_to_chat_response_format(text: OpenAIResponseText) -> OpenAIResponseFormatParam:
+    """
+    Convert an OpenAI Response text parameter into an OpenAI Chat Completion response format.
+    """
+    if not text.format or text.format["type"] == "text":
+        return OpenAIResponseFormatText(type="text")
+    if text.format["type"] == "json_object":
+        return OpenAIResponseFormatJSONObject()
+    if text.format["type"] == "json_schema":
+        return OpenAIResponseFormatJSONSchema(
+            json_schema=OpenAIJSONSchema(name=text.format["name"], schema=text.format["schema"])
+        )
+    raise ValueError(f"Unsupported text format: {text.format}")
+
+
 async def _get_message_type_by_role(role: str):
     role_to_type = {
         "user": OpenAIUserMessageParam,
@@ -180,6 +202,7 @@ class ChatCompletionContext(BaseModel):
     mcp_tool_to_server: dict[str, OpenAIResponseInputToolMCP]
     stream: bool
     temperature: float | None
+    response_format: OpenAIResponseFormatParam
 
 
 class OpenAIResponsesImpl:
@@ -343,10 +366,12 @@ class OpenAIResponsesImpl:
         store: bool | None = True,
         stream: bool | None = False,
         temperature: float | None = None,
+        text: OpenAIResponseText | None = None,
         tools: list[OpenAIResponseInputTool] | None = None,
         max_infer_iters: int | None = 10,
     ):
         stream = False if stream is None else stream
+        text = OpenAIResponseText(format=OpenAIResponseTextFormat(type="text")) if text is None else text
 
         output_messages: list[OpenAIResponseOutput] = []
 
@@ -355,6 +380,9 @@ class OpenAIResponsesImpl:
         messages = await _convert_response_input_to_chat_messages(input)
         await self._prepend_instructions(messages, instructions)
 
+        # Structured outputs
+        response_format = await _convert_response_text_to_chat_response_format(text)
+
         # Tool setup
         chat_tools, mcp_tool_to_server, mcp_list_message = (
             await self._convert_response_tools_to_chat_tools(tools) if tools else (None, {}, None)
@@ -369,6 +397,7 @@ class OpenAIResponsesImpl:
             mcp_tool_to_server=mcp_tool_to_server,
             stream=stream,
             temperature=temperature,
+            response_format=response_format,
         )
 
         # Fork to streaming vs non-streaming - let each handle ALL inference rounds
@@ -379,6 +408,7 @@ class OpenAIResponsesImpl:
                 input=input,
                 model=model,
                 store=store,
+                text=text,
                 tools=tools,
                 max_infer_iters=max_infer_iters,
             )
@@ -389,6 +419,7 @@ class OpenAIResponsesImpl:
                 input=input,
                 model=model,
                 store=store,
+                text=text,
                 tools=tools,
                 max_infer_iters=max_infer_iters,
             )
@@ -400,6 +431,7 @@ class OpenAIResponsesImpl:
         input: str | list[OpenAIResponseInput],
         model: str,
         store: bool | None,
+        text: OpenAIResponseText,
         tools: list[OpenAIResponseInputTool] | None,
         max_infer_iters: int | None,
     ) -> OpenAIResponseObject:
@@ -416,6 +448,7 @@ class OpenAIResponsesImpl:
                 tools=ctx.tools,
                 stream=False,
                 temperature=ctx.temperature,
+                response_format=ctx.response_format,
             )
             current_response = OpenAIChatCompletion(**inference_result.model_dump())
 
@@ -470,6 +503,7 @@ class OpenAIResponsesImpl:
             object="response",
             status="completed",
             output=output_messages,
+            text=text,
         )
         logger.debug(f"OpenAI Responses response: {response}")
 
@@ -489,6 +523,7 @@ class OpenAIResponsesImpl:
         input: str | list[OpenAIResponseInput],
         model: str,
         store: bool | None,
+        text: OpenAIResponseText,
         tools: list[OpenAIResponseInputTool] | None,
         max_infer_iters: int | None,
     ) -> AsyncIterator[OpenAIResponseObjectStream]:
@@ -503,6 +538,7 @@ class OpenAIResponsesImpl:
             object="response",
             status="in_progress",
             output=output_messages.copy(),
+            text=text,
         )
 
         # Emit response.created immediately
@@ -520,6 +556,7 @@ class OpenAIResponsesImpl:
                 tools=ctx.tools,
                 stream=True,
                 temperature=ctx.temperature,
+                response_format=ctx.response_format,
             )
 
             # Process streaming chunks and build complete response
@@ -645,6 +682,7 @@ class OpenAIResponsesImpl:
             model=model,
             object="response",
             status="completed",
+            text=text,
             output=output_messages,
         )