Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 15 additions & 6 deletions src/strands/models/openai_responses.py
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,7 @@
"""

import base64
import copy
import json
import logging
import mimetypes
Expand Down Expand Up @@ -356,13 +357,21 @@ async def structured_output(
ContextWindowOverflowException: If the input exceeds the model's context window.
ModelThrottledException: If the request is throttled by OpenAI (rate limits).
"""
request = self._format_request(prompt, system_prompt=system_prompt)

excluded_keys = {"model", "input", "stream"}
parse_kwargs: dict[str, Any] = {
"model": request["model"],
"input": request["input"],
"text_format": output_model,
}
for key, value in request.items():
if key not in excluded_keys:
parse_kwargs.setdefault(key, value)
Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Issue: The setdefault loop pattern works correctly but is a bit indirect. Since excluded_keys already filters out model, input, and stream, and the only additional key in parse_kwargs is text_format (which won't be in request), setdefault behaves identically to a direct assignment here.

Suggestion: Consider simplifying to a dict comprehension for clarity:

parse_kwargs: dict[str, Any] = {
    key: value for key, value in request.items() if key not in excluded_keys
}
parse_kwargs["text_format"] = output_model

This makes the logic more explicit: take everything from request except the excluded keys, then add text_format.


async with openai.AsyncOpenAI(**self.client_args) as client:
try:
response = await client.responses.parse(
model=self.get_config()["model_id"],
input=self._format_request(prompt, system_prompt=system_prompt)["input"],
text_format=output_model,
)
response = await client.responses.parse(**parse_kwargs)
except openai.BadRequestError as e:
if hasattr(e, "code") and e.code == "context_length_exceeded":
logger.warning(_CONTEXT_WINDOW_OVERFLOW_MSG)
Expand Down Expand Up @@ -404,7 +413,7 @@ def _format_request(
"model": self.config["model_id"],
"input": input_items,
"stream": True,
**cast(dict[str, Any], self.config.get("params", {})),
**copy.deepcopy(cast(dict[str, Any], self.config.get("params", {}))),
Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Issue: The copy.deepcopy added to _format_request is a broader change that affects all callers (including stream()), but neither stream() nor the new structured_output() code mutates the returned request dict's nested values. The deepcopy adds runtime cost on every request for a mutation scenario that doesn't currently exist.

Suggestion: If the intent is to protect self.config["params"] from potential mutation by the OpenAI SDK client, consider scoping the defensive copy to structured_output() only (e.g., deepcopy the request dict there), rather than changing _format_request which is a shared method. Alternatively, add a brief comment explaining why deepcopy is needed here so future contributors understand the rationale.

}

if system_prompt:
Expand Down
48 changes: 48 additions & 0 deletions tests/strands/models/test_openai_responses.py
Original file line number Diff line number Diff line change
Expand Up @@ -633,6 +633,54 @@ async def test_structured_output(openai_client, model, test_output_model_cls, al
assert tru_result == exp_result


@pytest.mark.asyncio
async def test_structured_output_passes_config_params(openai_client, test_output_model_cls, alist):
"""Test that structured_output passes config params (max_output_tokens, reasoning, etc.) to responses.parse."""
model = OpenAIResponsesModel(
model_id="gpt-5.4",
params={
"max_output_tokens": 500,
"reasoning": {"effort": "high"},
},
)

messages = [{"role": "user", "content": [{"text": "Generate a person"}]}]

mock_parsed_instance = test_output_model_cls(name="Alice", age=25)
mock_response = unittest.mock.Mock(output_parsed=mock_parsed_instance)
openai_client.responses.parse = unittest.mock.AsyncMock(return_value=mock_response)

events = await alist(model.structured_output(test_output_model_cls, messages))

assert events[-1] == {"output": test_output_model_cls(name="Alice", age=25)}

call_kwargs = openai_client.responses.parse.call_args.kwargs
assert call_kwargs["model"] == "gpt-5.4"
assert call_kwargs["text_format"] == test_output_model_cls
assert call_kwargs["max_output_tokens"] == 500
assert call_kwargs["reasoning"] == {"effort": "high"}
assert "stream" not in call_kwargs


@pytest.mark.asyncio
async def test_structured_output_passes_instructions(openai_client, test_output_model_cls, alist):
"""Test that structured_output passes system_prompt as instructions to responses.parse."""
model = OpenAIResponsesModel(model_id="gpt-4o")

messages = [{"role": "user", "content": [{"text": "Generate a person"}]}]

mock_parsed_instance = test_output_model_cls(name="Bob", age=40)
mock_response = unittest.mock.Mock(output_parsed=mock_parsed_instance)
openai_client.responses.parse = unittest.mock.AsyncMock(return_value=mock_response)

events = await alist(model.structured_output(test_output_model_cls, messages, system_prompt="Be helpful"))

assert events[-1] == {"output": test_output_model_cls(name="Bob", age=40)}

call_kwargs = openai_client.responses.parse.call_args.kwargs
assert call_kwargs["instructions"] == "Be helpful"


@pytest.mark.asyncio
async def test_stream_context_overflow_exception(openai_client, model, messages):
"""Test that OpenAI context overflow errors are properly converted to ContextWindowOverflowException."""
Expand Down
Loading