subagent view: permission prompts should appear inline instead of forcing exit to main screen

Summary

When viewing a specific subagent's output in the subagent view, if that subagent requests permission for a tool call, the user must exit the subagent view and return to the main screen to approve or deny it. The approval prompt should be accessible directly within the subagent view so the user doesn't lose their context.

Environment

• opencode version: v0.0.0-beta-17639
• OS: Linux 7.0.11-76070011-generic (linux x64)
• Terminal: TERM=xterm-256color, COLORTERM=truecolor
• Shell: /bin/bash
• Install/channel: beta
• Active plugins: context-cleaner, context-forge, @tarquinen/opencode-dcp@latest

Reproduction

1. Launch a subagent (e.g., edit or general agent)
2. Enter the subagent view to watch its progress
3. The subagent triggers a tool call that requires permission (e.g., file write, shell command)
4. The permission prompt appears on the main screen, not in the subagent view
5. User must exit the subagent view to go back to the main screen to approve/deny
6. After approving, user has to navigate back to the subagent view to continue watching

Expected Behavior

• When a subagent requests permission for a tool call, the approval prompt should appear directly within the subagent view
• The user should be able to approve, deny, or always-allow the tool call without leaving the subagent view
• This should work for both foreground and background subagents
• After responding to the prompt, the user should remain in the subagent view and continue watching progress

Actual Behavior

• Permission prompts only appear on the main screen
• User must exit the subagent view to interact with the permission prompt
• This breaks the flow of monitoring a subagent's work, especially during long agentic runs where multiple permission prompts may occur

Additional Context

This is particularly disruptive when:

• Running multiple subagents in parallel and watching one closely
• Long agentic runs where the subagent makes several tool calls that need approval in sequence
• The user has to repeatedly switch between subagent view and main screen for each permission prompt

A solution could be:

• Render permission prompts inline within the subagent view (as an overlay or inline message)
• Allow keyboard shortcuts to approve/deny without leaving the current view
• Optionally show a notification/badge in the subagent view when a permission prompt is pending