.Net: Python: KernelJsonSchemaBuilder emits {"type": "object"} with no enum for typing.Literal parameters and fields #14511

What happens

KernelJsonSchemaBuilder has no handling for typing.Literal. A parameter typed Literal["fast", "slow"] is advertised to the model as a bare {"type": "object"} with no enum. The model is told the argument is an object and is never told the allowed values.

This affects @kernel_function parameters, Pydantic model fields, and Literal nested inside list[...] or Optional[...].

Where

python/semantic_kernel/schema/kernel_json_schema_builder.py, KernelJsonSchemaBuilder.handle_complex_type() and get_json_schema().

Literal[...] has __args__, so build() sends it to handle_complex_type(). That method handles list, set, dict, tuple and Union, but not Literal, so it reaches the final fallback:

schema = cls.get_json_schema(parameter_type)   # TYPE_MAPPING.get(Literal[...], "object")

TYPE_MAPPING has no entry for a Literal alias, so the default "object" is returned. build_enum_schema() already produces the right shape for Enum classes ({"type": ..., "enum": [...]}), but nothing does that for Literal.

Repro

Tested with semantic-kernel==1.44.1. I diffed kernel_json_schema_builder.py against current main and the file is identical.

import json
from typing import Annotated, Literal
from pydantic import BaseModel
from semantic_kernel import Kernel
from semantic_kernel.functions import kernel_function
from semantic_kernel.connectors.ai.function_calling_utils import kernel_function_metadata_to_function_call_format

class Opts(BaseModel):
    level: Literal["low", "high"]

class Plugin:
    @kernel_function(name="run")
    def run(self, mode: Annotated[Literal["fast", "slow"], "speed mode"], opts: Annotated[Opts, "options"]) -> str:
        return mode

k = Kernel(); k.add_plugin(Plugin(), "p")
fn = k.get_function("p", "run")
print(json.dumps(kernel_function_metadata_to_function_call_format(fn.metadata)["function"]["parameters"], indent=1))

Output:

{
 "type": "object",
 "properties": {
  "mode": {"type": "object", "description": "speed mode"},
  "opts": {
   "type": "object",
   "properties": {"level": {"type": "object"}},
   "required": ["level"],
   "description": "options"
  }
 },
 "required": ["mode", "opts"]
}

Direct builder calls:

KernelJsonSchemaBuilder.build(Literal["fast", "slow"])            # {'type': 'object'}
KernelJsonSchemaBuilder.build(Optional[Literal["fast", "slow"]])  # {'type': ['object', 'null']}
KernelJsonSchemaBuilder.build(list[Literal["a", "b"]])            # {'type': 'array', 'items': {'type': 'object'}}

Expected

{"type": "string", "enum": ["fast", "slow"], "description": "speed mode"}

This is the same shape build_enum_schema() already returns for an Enum subclass. Optional[Literal[...]] should be nullable and list[Literal[...]] should have an enum on items.

Actual

{"type": "object"} with no allowed values. The model can emit any value, including a JSON object, and the kernel then tries to bind it to a str parameter. Tools that use Literal to constrain choices (a very common pattern) lose that constraint entirely.

Suggested fix

Add a Literal branch in handle_complex_type():

if origin is Literal:
    values = list(args)
    types = {TYPE_MAPPING.get(type(v), "string") for v in values}
    schema = {"type": types.pop() if len(types) == 1 else sorted(types), "enum": values}
    if description:
        schema["description"] = description
    return schema

The Optional[...] branch also needs to keep working with an enum schema (add None to the enum when making it nullable).

I searched open and closed issues and PRs for Literal, KernelJsonSchemaBuilder, and schema builder problems. #14239, #14241 and #14310 are about string forward references and only make sure Literal values are left alone while resolving them. #14482 and #14155 are about NoneType in unions. None of them adds Literal support.