fix: fail fast on tools + structured output conflict for OpenRouter (Issue #7132)#7196
Open
ichbinlucaskim wants to merge 1 commit intomicrosoft:mainfrom
Open
Conversation
…Issue microsoft#7132) - Add guardrail in OpenAIChatCompletionClient to fail fast when tools are combined with Pydantic structured outputs - Add regression tests plus an OpenRouter repro sample for issue microsoft#7132.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This PR implements a short‑term, fail‑fast guardrail for issue #7132 and adds a dedicated reproduction sample plus regression tests.
When
AssistantAgentis used with an OpenRouter (OpenAI‑compatible) model andoutput_content_typeis a Pydantic model, theOpenAIChatCompletionClientcurrently routes requests throughbeta.chat.completions.parse(response_format=...). In this structured‑output mode, the OpenAI API does not support tool calling, so requests that include bothresponse_formatandtoolsresult in tool calls being silently ignored.This PR makes that incompatibility explicit and prevents the “silent tool drop” behavior.
Changes
1. Fail‑fast guardrail in
_process_create_argsFile:
python/packages/autogen-ext/src/autogen_ext/models/openai/_openai_client.pyMethod:
_process_create_argsRight after
converted_tools = convert_tools(tools), we now check for the incompatible combination of structured output and tools:Behavior:
response_format_valueis set (Pydantic‑based structured output is enabled), andconverted_toolsis non‑empty (at least one function tool is present)._process_create_argsraisesValueErrorwith a descriptive message._process_create_argsis changed:2. Regression tests for the guardrail
File:
python/packages/autogen-ext/tests/models/test_openai_model_client.pyAdded a minimal Pydantic model used only for these tests:
Added a minimal tool function:
Test 1: Pydantic
json_output+ tools →ValueErrortest_structured_output_with_tools_raises_value_errorBaseOpenAIChatCompletionClientwith aMagicMockunderlying client andmodel_info.structured_output=True.messages=[UserMessage(...)]tools=[FunctionTool.from_function(_dummy_tool_for_guardrail)]json_output=Weatherclient._process_create_args(...).ValueErroris raised."Cannot use structured output (output_content_type) together with function tools".Test 2: Pydantic
json_output+ no tools → passestest_structured_output_without_tools_passesWeathermodel.tools=[].client._process_create_args(...).create_paramshasresponse_format is Weather.len(create_params.tools) == 0.Both tests use a mocked underlying client and only exercise
_process_create_args; they do not perform any real network calls.3. Reproduction sample
File:
python/samples/agentchat_openrouter/assistant_openrouter_output_content_type.pySample that reproduces issue #7132 using
AssistantAgentwith an OpenRouter model:get_weather).output_content_typeto a Pydantic model.ValueErrordescribed above.Structural / Long‑Term Note
This PR intentionally does not attempt to make “tools + structured output in one request” work, because it is not compatible with the current OpenAI API contract.
beta.chat.completions.parse(response_format=...)is designed for returning structured JSON that matches the schema in a single step.tool_call, external code runs, and then the model is called again with tool results.Trying to combine both in a single request is structurally incompatible with the current API behavior.
A more robust long‑term solution would be to change the agent workflow to use a two‑step protocol:
response_format) to execute tools and collect their outputs.response_formatset and tools disabled, to turn tool outputs into a typed response.This PR is a short‑term, fail‑fast guardrail that makes the incompatibility explicit; it does not change the higher‑level agent workflow.
Testing
Local runs:
uv run pytest packages/autogen-ext/tests/models/test_openai_model_client.pyValueErroras expected under the guardrail.The new unit tests:
test_structured_output_with_tools_raises_value_errortest_structured_output_without_tools_passesboth pass.