One-click handoff from agent commands to the terminal and back

[githubcustomerserviceistrash]

Agents sometimes reach a step that requires explicit human authorization, such as entering a sudo password. The current loop is slow when repeated: copy the command, open a terminal, paste and run it, then copy any failure output back into chat.

It would be useful if an agent command offered Open in terminal: open or create the thread terminal with the command prefilled, then let the user edit it or press Enter. Afterward, the user could reference that terminal in chat (for example, @terminal-1) or attach the latest command output in one action.

This would preserve the human authorization boundary while removing most of the handoff work. A useful first version could stop at prefilling rather than auto-executing.

Related: #26 added selected terminal output to chat; #1042 proposed terminal mentions/context.