Expose the agent's permission/plan mode as a control in the prompt box

[jesse-adapta]

Environment: Parallel Code 1.12.0 (deb), Linux, Claude Code agent. Code references verified against main @ d000fff65.

Problem

Claude Code has a permission-mode cycle (default → auto-accept edits → plan mode, normally cycled with Shift+Tab). Right now the only way to change it is to click into the agent's terminal pane and press Shift+Tab there.

Request

Add a mode control to PromptInput (src/components/PromptInput.tsx), next to the send button:

• Shows the agent's current mode.
• Click (or a keybinding while the prompt box is focused) cycles it — implemented by sending the agent's mode-cycle key to the PTY, so no agent-side integration is needed.
• Ideally the sent prompt uses whatever mode is selected at send time, so I can pick "plan" and send in one gesture.

[LarryHu0217]

I checked this against current main (d000fff). Sending Claude's mode-cycle key from PromptInput is straightforward (\x1b[Z), but the app currently has no authoritative permission-mode state. The only observable source is Claude's rendered PTY footer, and getAgentOutputTail() is not itself reactive. Tracking the mode only after button clicks would drift whenever the user presses Shift+Tab in the terminal.

Would you be comfortable with a Claude-only first pass that:

• parses normalized PTY frames into default, accept edits, plan, or bypass permissions when Claude exposes one of those footer labels;
• shows the control only for Claude Code, with an explicit unknown state until a recognizable footer is observed;
• sends exactly one Shift+Tab sequence per click while keeping focus in the prompt box; and
• uses recorded-frame tests for mode detection plus a focused control test for the PTY write?

I would avoid optimistically cycling local state, so the displayed value always comes back from the agent's output. If that scope matches what you want, I can take it.