[BUG] [MCPClient] AIFunction arguments of complex types are passed as string representation of

[SGStino]

Describe the bug
For an MCP function with complex arguments, the json serialization doesn't match the required jsonschema of the server when called through a KernelFunction of SemanticKernel.

To Reproduce
Steps to reproduce the behavior:

1. create an MCP function:

(TodoListItems values, string scope) => "success"

public record TodoListItems(string[] Items);

2. Use that function in semantic kernel like in the sample:
   https://github.com/microsoft/semantic-kernel/blob/fac316e7dc2cd7dabbaab0ee899c04187c301093/dotnet/samples/Demos/ModelContextProtocolClientServer/MCPClient/Samples/MCPToolsSample.cs#L42

3. Let the model call the KernelFunction, which calls the AIFunction, which forwards it to the MCP client and passes the following along to the MCP server

"{\"items\":[\"todo1\", \"todo2\"]}"

And that obviously doesn't end well.

Expected behavior
It should pass:

{"items":["todo1", "todo2"]}

Logs
N/A

Additional context
AIFunction itself also presents the same issue: dotnet/extensions#6544

I'm honestly not sure if it goes wrong in SemanticKernel that should already deserialize it (since the Ollama client sends even a JsonElement along to the KernelFunction argument instead of a string like in the OpenAI client)

Or if it's something wrong in AIFunction, or in the MCP Client.

What's the design requirement around the AIFunction/KernelFunction? Since both seem to be implemented differently and packed around themselves.

[SGStino]

Side note: when using Ollama with a mistral-small instead of AzureOpenAI, I got passed a JsonElement instead of a string representation. But then I wasn't using MCP yet, and now I don't have access to the GPU to test that.

But i suspect the JsonElement as argument gets passed along nicely to the MCPClient's serializer?

[eiriktsarpalis]

Appears to be a duplicate of #274?