Python: Exclude ClassVar annotations from generated JSON schemas

[Pushpak Siva Sai (Pushpak731)]

Fixes #14527

Description

KernelJsonSchemaBuilder.build_model_schema builds the schema from get_type_hints, which includes ClassVar annotations, so a class constant became a required object-valued property in the generated schema. Pydantic excludes ClassVars from model_fields and from serialized instances, so the resulting schema rejected valid serialized payloads ({'value': 1} failed because marker was required). The same wrong schema reached KernelParameterMetadata, generated function-call tool definitions and structured_output=True schemas.

The fix skips ClassVar-annotated hints before the required/property loop — covering the parameterized form (ClassVar[str]), the bare form (ClassVar, whose get_origin is None), and inherited ClassVars, for KernelBaseModel subclasses, plain annotated classes and model instances alike.

class Payload(KernelBaseModel):
    marker: ClassVar[str] = "constant"
    value: int

KernelParameterMetadata(name="payload", type_object=Payload).schema_data
# before: properties include marker (type object) and required: [marker, value]
# after:  {'properties': {'value': ...}, 'required': ['value']}

Contribution Checklist

• I have read the contribution guidelines.
• The code builds clean without any errors or warnings
• The PR follows the SK Contribution Guidelines and the pre-submit framework
• Tests are included and pass locally: tests/unit/schema/ (45 passed) and tests/unit/functions/test_kernel_parameter_metadata.py (7 passed) — 10 new regression tests cover pydantic models, bare/inherited ClassVar, plain classes, instances, structured_output, descriptions, optionals and the serialized-payload contract
• ruff check and ruff format --check are clean on both changed files
• Documentation updates are not needed (no public API change; behavior now matches model_dump)

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Pushpak Siva Sai (Pushpak731)]

@microsoft-github-policy-service agree

[PRABHU KIRAN VANDRANKI (VANDRANKI)]

Community review. This does not clear the merge gate, it is one reader's check.

I reproduced the bug and checked the fix by running code. I loaded the installed semantic_kernel/schema/kernel_json_schema_builder.py (no ClassVar handling in it), then built a copy with your two added lines applied by string patch, and compared KernelJsonSchemaBuilder.build output for three classes.

• class M(KernelBaseModel): marker: ClassVar[str] = "c"; value: int
  • current code: properties marker: {"type": "object"} and value, required ["marker", "value"]
  • with the patch: only value, required ["value"]
  • M(value=1).model_dump() is {"value": 1} and M.model_fields is ["value"], so the current schema demands a field the model never serializes, and it also types it wrongly as object.
• Annotated[ClassVar[int], "x"] on a KernelBaseModel: same result, the patch drops it.
• A plain (non-pydantic) class with a string annotation "ClassVar[dict]": also dropped, because get_type_hints resolves the string before the check.

The check field_type is ClassVar or get_origin(field_type) is ClassVar covers both the bare ClassVar and the subscripted form, which are the two shapes get_type_hints returns. It also fits with the loop: nothing after the continue depends on the skipped field, so required and properties stay consistent.

Notes:

• I did not run the repo's test file or the full suite, only the builder on the three classes above. The new tests (inherited ClassVar, structured output with additionalProperties, via-instance) read correctly to me, but I did not execute them.
• I did not check typing_extensions.ClassVar on older Pythons. It is the same object as typing.ClassVar on current versions, so I expect it to work, but that is unverified.
• Nit: the ten new tests overlap a lot. Two or three would cover the same branches (bare, subscripted, inherited).

Approving from my side.