Support AskHuman-like behavior

[xxchan]

The agent can implement some tools that ask the user for opinions. So the tool call will block until the user replies in the UI (ACP client here). This is something similar to Permission Request, but with a text/options reply (like a more advanced approval request).

One real example of this feature is Claude Code's interactive question tool.

Note that this is not equivalent to:

• agent asks the user as a text reply
• plan mode (some kind of workflow)

The main problem is that: last turn's LLM's thinking might be dropped. So the human preference must be passed as a tool call result (so we are still in the same agent turn).

[xxchan]

link zed-industries/claude-code-acp#136