[API Proposal]: Dynamic Approval Required Capabilities

[gjz22]

Background and motivation

ApprovalRequiredAIFunctions mean that a tool is either always requires approval or never does. However, most coding agents have a more fine grained permission than that. They allow the functions to be approved based on the arguments passed. For example, a "Read" tool might need permission for /etc/passwd but not for /tmp/x.

API Proposal

The goal is to use middleware for this purpose so that a permissioning system could be injected into a middleware to ensure backwards compatibility and allow the most flexibility.

In FunctionInvocationContext simply add two fields:

public bool ApprovalRequired { get; set; }

The ApprovalRequired would be set at the beginning of the function invocation to whether the target function is an ApprovalRequiredAIFunction.

public string ApprovalRequiredMessage { get; set; }

This allows a dynamic message to be sent to the client.

API Usage

An example of how the a policy could be injected to a middleware to affect the approval required dynamically with the new field.

public sealed class PolicyPermissionMiddleware
{
    private readonly IToolPermissionPolicy _policy;

    public PolicyPermissionMiddleware(IToolPermissionPolicy policy)
    {
        _policy = policy;
    }

    public async Task InvokeAsync(
        FunctionInvocationContext context,
        Func<FunctionInvocationContext, Task> next)
    {
        PermissionDecision decision = _policy.Evaluate(context);

        if (decision.ApprovalRequired)
        {
            context.ApprovalRequired = true;
            context.ApprovalRequiredMessage = decision.Message;

            // Do not invoke the tool. The framework/client should surface
            // the approval request and only continue if approval is granted.
            return;
        }

        await next(context);
    }
}

public sealed record PermissionDecision(
    bool ApprovalRequired,
    string? Message = null);

public interface IToolPermissionPolicy
{
    PermissionDecision Evaluate(FunctionInvocationContext context);
}

Alternative Designs

No response

Risks

No response