fix: make FunctionParameters and LogitBias OpenAI-compatible

[markstur]

Misc PR

Type of PR

• Bug Fix
• New Feature
• Documentation
• Other

Description

• Link to Issue: Fixes This "RootModel" key is non-standard and will cause any real OpenAI-compatible client to fail. #971

fix: make FunctionParameters and LogitBias OpenAI-compatible

Use Pydantic's RootModel to accept bare JSON Schema dicts instead of
requiring a non-standard RootModel wrapper field. This makes m serve
compatible with standard OpenAI clients.

BREAKING CHANGE: FunctionParameters(RootModel={...}) is now FunctionParameters({...})

Testing

• Tests added to the respective file if code was changed
• New code has 100% coverage if code as added
• Ensure existing tests and github automation passes (a maintainer will kick off the github automation when the rest of the PR is populated)

Attribution

• AI coding assistants used

[github-actions]

The PR description has been updated. Please fill out the template for your PR to be reviewed.

[markstur]

@planetf1 don't want to assign you work but fyi this is an issue raised from your feedback

[planetf1]

The core fix looks solid — switching FunctionParameters and LogitBias to RootModel[dict] is exactly right, and the test updates follow through consistently. Left a few notes below: one thing worth sorting before this lands, and a couple of minor observations that are fine to pick up in a follow-up if you'd rather keep this focused.

[planetf1]

Minor / non-blocking: small gotcha with the alias — Field(alias="schema") means .model_dump() emits schema_ (the Python attribute name), not schema:

j = JsonSchemaFormat(name="test", schema={"type": "object"})
j.model_dump()
# → {"name": "test", "schema_": {"type": "object"}}

For the current code path — where json_schema.schema_ is only read internally — this is harmless. But if it ever gets serialised and sent anywhere (a client, a log), the wrong key comes out. Worth either using .model_dump(by_alias=True) at the call site, or adding "serialize_by_alias": True to model_config to make it the default. Fine to leave for a follow-up.

[planetf1]

Minor / non-blocking: one gap worth noting in the test coverage — there's no HTTP-layer test that sends the old {"RootModel": {...}} payload to /v1/chat/completions and confirms it gets a sensible error back. Given this is a breaking API change, a short TestClient test for the legacy envelope format would make the regression boundary explicit. Totally fine as a follow-up if you'd rather keep this PR focused.

[planetf1]

One issue worth sorting before this lands — the silent {"RootModel": {...}} pass-through on FunctionParameters (see inline comment). For a breaking API change, clients that haven't updated yet deserving a clear error rather than a confusing inference failure downstream.

The other comments are minor and non-blocking.

[planetf1]

Both issues addressed — happy with the validator approach and the LogitBias cleanup. LGTM.