claude-code - 💡(How to fix) Fix Runtime --permission-mode switch (parity with SDK q.setPermissionMode) [3 pull requests]

Feature request: runtime --permission-mode switch

Context

The Claude Agent SDK exposes q.setPermissionMode("plan" | "acceptEdits") so a host can toggle permission mode mid-session without restarting query(). The CLI binary accepts --permission-mode only at spawn time.

Problem

Hosts that wrap the CLI under a PTY (to preserve subscription billing instead of API billing) cannot mirror SDK parity. When a user toggles plan mode in our UI mid-chat, we either:

1. Send /permissions — opens the interactive menu, doesn't actually set the mode.
2. Kill the process and respawn with new --permission-mode — loses the in-flight turn state.

Neither matches the SDK's behavior.

Request

Expose a runtime equivalent. Options:

• A slash command, e.g. /set-permission-mode plan / /set-permission-mode acceptEdits, that mutates the running session.
• A control message accepted on stdin (e.g. \x1b[control]set-permission-mode plan\n) parallel to whatever the SDK uses under the hood.
• A documented signal-based mechanism (e.g. SIGUSR1 rotates through modes).

Any of these would let CLI-wrapping hosts achieve SDK parity for plan-mode toggling.

Why this matters

Without a runtime switch, every CLI-wrapped Kanna chat session that wants to toggle plan mode must either lose state (respawn) or document a UX regression (close + reopen chat).

Repro / context

• Kanna PTY driver: https://github.com/cuongtranba/kanna/blob/main/src/server/claude-pty/driver.ts
• Parity audit: https://github.com/cuongtranba/kanna/issues/162 (item D4)