Add internal interactive questioning tool 'usr_ask'

Describe the feature or problem you'd like to solve

In older versions like v1.0.6, ask_user worked nicely to display interactive multiple-choice menus in the terminal.

However, in current releases (tested on v1.0.80), the agent no longer triggers interactive question prompts, even when explicitly asked to offer choices or clarify next steps. Instead, it just outputs plain text lists and ends the turn.

Proposed solution

Restore and polish the built-in ask_user (or usr_ask) tool so the CLI can reliably render interactive selection prompts:

• Simple arrow-key navigation (↑/↓ + Enter) for choices.
• Single and multi-select support.
• Built-in "Other / custom text" option so users can write in answers.
• Ensure the model reliably triggers it when ambiguity arises instead of falling back to raw text.
• Proper handling in autonomous/YOLO/allow-all modes so interactive prompts don't unexpectedly block unattended or automated workflows (as addressed in #2109).

Example prompts or workflows

• Choosing frameworks or tools (e.g. asking which test runner to install with selectable items).
• Confirming dangerous operations before execution.
• Clarifying requirements during complex tasks without ending the session early.

Additional context

Looking at the v1.0.6 release (app.js), ask_user was defined with { question, choices, allow_freeform } and rendered via registerPrompt. On v1.0.80 this seems broken or dormant.

#2109 is especially relevant here, as it lays out solid proposals to handle ask_user gracefully across ACP and automated modes without breaking YOLO / unattended execution.

Related issues: #2109, #3323, #4050, #4260, #4450, #3722, #2067, #2929, #1898.

Inizia confrontando l'implementazione di ask_user in app.js di v1.0.6 e il suo utilizzo di registerPrompt con la CLI attuale, quindi leggi le proposte in #2109. Traccia dove verrebbero registrati e attivati usr_ask o ask_user. Il lavoro è completato quando sono supportati prompt interattivi a selezione singola e multipla, risposte in formato libero e un comportamento sicuro nelle modalità autonomous, YOLO, ACP e unattended.