Add execution-time user input checkpoints outside Plan mode

Feature request

Please consider adding an execution-time user input checkpoint tool that agents can use outside Plan mode.

Today, Plan mode can ask the user for structured input without fully ending the planning flow. However, many important uncertainties only appear during actual execution, for example:

• choosing between two similar config files discovered mid-task
• confirming before deletion, overwrite, bulk edits, encoding changes, or global configuration changes
• resolving ambiguous UI or image judgment
• asking for account/environment-specific choices discovered after tools have already run

In Default/execution mode, the agent can still ask a normal chat question, but that ends the current assistant turn. For longer work, this creates extra stop-start friction and makes execution feel less continuous.

Proposed behavior

Add an inline confirmation/checkpoint capability available during execution mode, similar in spirit to Plan mode's structured user input:

• allow 1-3 short questions during an active run
• support both required answers and optional auto-resolve behavior after a timeout
• let the agent attach concise context such as file paths, diff summaries, screenshots, or risk level
• after the user responds, continue the same execution flow with the existing tool/context state
• make it suitable for high-risk or user-preference-dependent operations

Why this helps

This would let agents remain proactive while still being conservative around risky or subjective decisions. It would also improve workflows where the right behavior cannot be known until after code inspection, logs, screenshots, or filesystem state have been examined.

In short: Plan mode is useful for deciding the initial approach, but execution mode also needs lightweight user checkpoints for uncertainties discovered mid-task.