Although today being possible to do with current demonstration I would prefer to see this implemented as supported scenario in here:
Where if you add ChatHistory as the argument type of a function it is properly injected in the function.

@rogerbarreto Thanks for feedback! Yes, I agree that such approach would work in the best way for end users, and I've investigated it in more detail. It appeared that in order to support that, we will have to add ChatHistory parameter to the abstraction in places where it doesn't look like it belongs there.

For example, we will need to update kernel.InvokeAsync(function, arguments, cancellationToken) to kernel.InvokeAsync(function, arguments, chatHistory, cancellationToken), which I'm not sure is a good approach, since a function may be invoked outside of chat completion context. Even adding an optional ChatHistory dependency to KernelFunction looks like an unnecessary high coupling.

I was also thinking about adding ChatHistory property to KernelArguments class, but again, I'm not sure if KernelArguments should be aware of any chat-related models in compile time.

Taking into account that today it's possible to pass ChatHistory through KernelArguments in runtime as one of the dictionary values, I think we can keep this as a solution for now, and we can think about improving it if we will see some problems associated with it.

Hope that makes sense!