Migrate openai framework to use any_llm

[HareeshBahuleyan]

• Introduce AnyllmModel based on LitellmModel from openai agents SDK: https://github.com/openai/openai-agents-python/blob/main/src/agents/extensions/models/litellm_model.py#L64
• Updated tests to use appropriate mock depending on the llm library (litellm / any_llm) used by the respective frameworks

[codecov]

Codecov Report

❌ Patch coverage is 74.65753% with 37 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/any_agent/frameworks/openai.py	64.42%	26 Missing and 11 partials ⚠️

Files with missing lines	Coverage Δ
src/any_agent/testing/helpers.py	40.98% <ø> (-29.51%)	⬇️
src/any_agent/tools/wrappers.py	94.56% <100.00%> (ø)
tests/conftest.py	79.74% <100.00%> (-11.40%)	⬇️
tests/unit/conftest.py	100.00% <100.00%> (ø)
tests/unit/frameworks/test_any_agent.py	100.00% <100.00%> (ø)
tests/unit/frameworks/test_openai.py	100.00% <100.00%> (ø)
tests/unit/tools/test_exception_run.py	100.00% <100.00%> (ø)
tests/unit/tools/test_unit_wrappers.py	96.36% <100.00%> (ø)
src/any_agent/frameworks/openai.py	63.22% <64.42%> (-15.73%)	⬇️

... and 47 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[HareeshBahuleyan]

Integration tests passing on CI: https://github.com/mozilla-ai/any-agent/actions/runs/18492592870

Status of running [OPENAI] framework integration tests locally:

• ✅ Mistral
• ✅ OpenAI
• ❌ Gemini (also fails when tested on main)
  • Reason: Gemini has no support for OpenAI json_schema for structured output; they use a different format: https://googleapis.github.io/python-genai/index.html#json-response-schema
• ❌ XAI (also fails when tested on main)
  • Reason: XAI SDK does not support OpenAI structured output format: https://docs.x.ai/docs/guides/structured-outputs
• ❌ HF Inference Endpoint
  • Reason: The completion fails with ClientResponseError / Unprocessable Entity when the request payload contains both tools and response_format!
    • However, if either one of them are popped out from the dict, it works. Could it be an [model-specific issue](https://www.reddit.com/r/LocalLLaMA/comments/1h2y7ys/why_can_you_not_use_structured_output_and_tool/) with the hosted Qwen-3 model?
  • When tested with tinyagent (instead of OpenAI), the payload prepared by tinyagent contains either only tool or only response_format but not both.

@daavoo I guess Gemini and XAI not supporting OpenAI structured output format is something that we cannot fix.
Hence I think we based on the code review, we could proceed with the PR.

[daavoo]

@daavoo I guess Gemini and XAI not supporting OpenAI structured output format is something that we cannot fix.

@HareeshBahuleyan I don't think we can fix it in any-agent but I don't see why we couldn't make this work with changes in any-llm

[HareeshBahuleyan]

but I don't see why we couldn't make this work with changes in any-llm

@daavoo Do you mean implementing a helper function in any-llm to convert openAI JSON schema format to Pydantic?

[daavoo]

but I don't see why we couldn't make this work with changes in any-llm

@daavoo Do you mean implementing a helper function in any-llm to convert openAI JSON schema format to Pydantic?

no, I mean that any-llm should support receiving as response_format any of the argument types that openai supports and take care of converting them to the appropriate format for each provider under the hood (in this case there should be some conversion for gemini and xai)

[HareeshBahuleyan]

Sounds reasonable, created issues in any-llm to track this: mozilla-ai/any-llm#541 and mozilla-ai/any-llm#542

[daavoo]

I run the tests with the debugger and implementation looks good.

We weed to update pyproject.toml and would be great to try the class-based AnyLLM API.

[HareeshBahuleyan]

Rerun of integration tests after class-based AnyLLM definition:
https://github.com/mozilla-ai/any-agent/actions/runs/18588782499

[daavoo]

Thanks for addressing all the comments @HareeshBahuleyan 🙏