Consider an AIContent type for content that must be roundtripped

[stephentoub]

In cases where middleware needs to create content that will be sent back to the consumer which must then include that content in a subsequent request, we should consider introducing some IncludeContent type like:

class IncludeContent : UserInputRequestContent
{
    public IList<AIContent>? Content { get; set; } 
}

This could be used, for example, by FunctionInvokingChatClient to work around:

extensions/src/Libraries/Microsoft.Extensions.AI/ChatCompletion/FunctionInvokingChatClient.cs

Lines 48 to 53 in a4a3d80

	/// <para>
	/// Due to the nature of interactions with an underlying <see cref="IChatClient"/>, if any <see cref="FunctionCallContent"/> is received
	/// for a function that requires approval, all received <see cref="FunctionCallContent"/> in that same response will also require approval,
	/// even if they were not <see cref="ApprovalRequiredAIFunction"/> instances. If this is a concern, consider requesting that multiple tool call
	/// requests not be made in a single response, by setting <see cref="ChatOptions.AllowMultipleToolCalls"/> to <see langword="false"/>.
	/// </para>

cc: @westey-m, @jozkee

[westey-m]

In the case of ChatClientAgent, would it make sense to temporarily store such content in the AgentThread and automatically include it in the next request, so that the caller doesn't have to deal with it? This could be based on a configurable option that is opt-out.

[stephentoub]

In the case of ChatClientAgent, would it make sense to temporarily store such content in the AgentThread and automatically include it in the next request, so that the caller doesn't have to deal with it? This could be based on a configurable option that is opt-out.

How would that work if ordering was important? And how would someone know whether they needed to respond to given input or not?