PermissionRequest hook auto-approves ExitPlanMode, causing plans to execute without user confirmation

[aikenahac]

SuperIsland version 1.0.8 build 9

What happened

After installing SuperIsland, I noticed that Claude Code stopped asking me to confirm plans before executing them. Even when I explicitly told Claude not to auto-accept, it would finish writing a plan and immediately start executing it - no prompt, no pause.

After some digging, I traced it back to the PermissionRequest hook that the agents-status extension installs.

Root cause

The extension registers a hook that intercepts all of Claude Code's permission requests and routes them through a local bridge server at http://127.0.0.1:7823/hooks/permission.

/Applications/SuperIsland.app/Contents/Resources/BundledExtensions/agents-status/server/server.py
The handler in server.py only blocks on AskUserQuestion - everything else gets silently approved:

def _handle_permission_hook(body):
    tool_name = (body.get("tool_name") or "").strip()
    ...

    if tool_name != "AskUserQuestion":
        return {
            "hookSpecificOutput": {
                "hookEventName": "PermissionRequest",
                "decision": {"behavior": "allow"},
            }
        }

This includes ExitPlanMode - the tool Claude Code uses to present a finished plan and wait for the user to approve or reject it before execution begins. Because the hook auto-approves it, the confirmation step never surfaces.

Steps to reproduce

1. Install SuperIsland with the agents-status extension and hooks enabled
2. Start a Claude Code session in plan mode
3. Let Claude finish writing a plan
4. Observe: execution begins immediately, no confirmation prompt appears

Expected behavior

The plan review prompt should appear as normal, letting the user approve or reject before anything runs.

Actual behavior

ExitPlanMode is approved automatically by the hook bridge. The user never sees the prompt.

Why this matters

Plan mode exists specifically so users can review and control what Claude is about to do. Silently bypassing that - with no indication it's happening - undermines the whole point. I only found out because I kept noticing Claude executing things I hadn't agreed to and went digging through the config.

Suggested fix

Treat ExitPlanMode the same way AskUserQuestion is handled: block the hook thread and surface the decision in the SuperIsland UI so the user can actually respond to it.