[dotnet-port-api] agent-as-tool: session propagation and argument customization #1139

Parity gap surfaced comparing tool/agenttool against .NET/Python.

Gap

agenttool hardcodes a single query string argument and runs the sub-agent with an independent context, returning only resp.String(). It cannot:

1. Propagate the parent session to the sub-agent (so a sub-agent-as-tool starts stateless every call).
2. Customize the argument name/description.

Note: the tool name/description defaults and the query arg name already match .NET (AgentExtensions.cs), so those are intentionally aligned — only session propagation and arg customization are missing.

Reference

• Python Agent.as_tool(arg_name="task", arg_description=..., propagate_session=..., ...) (_agents.py).
• .NET AsAIFunction(this AIAgent, options, session) accepts a session (AgentExtensions.cs).

Proposed shape (needs maintainer input)

Add optional config to the agenttool constructor, e.g.:

agenttool.New(a, agenttool.Config{
    ArgName:        "task",         // default "query" (matches .NET)
    ArgDescription: "...",
    PropagateSession: true,          // thread the caller's *agent.Session into the sub-run
})

Open questions:

1. How to obtain the caller session at Call time — thread it via context, or capture from the invoking run? (Go tool Call(ctx, argsJSON) has no session param today.)
2. Should session propagation be opt-in (default off, matching a stateless tool) or default on (matching Python's common usage)?
3. Whether to also surface sub-agent approval/user-input requests (Python raises UserInputRequiredException; .NET returns .Text — Go currently matches .NET).

Filing issue-first per CONTRIBUTING (new public API). Happy to implement once the direction is settled.