Home / Blog

The claude-garage plugin confirm rule

By Saiful Islam (Adar), creator of claude-garage ·

A claude-garage plugin can ask to close a session. It can never confirm. The wall only arms the action and shows a prompt, and you press a key or click a second time. This post explains the rule, the code-level reason it holds, and what it does not protect.

What is the confirm rule?

The plugin docs state it in one line: a plugin can ask, only the human can confirm. The list of dangerous actions lives in core_actions.rs.

Some core actions are dangerous. The list is session.close, session.delete, session.restart, session.restore_all, workspace.restart, workspace.remove, workspace.remove.kill, daemon.restart and hooks.install. When a plugin triggers one of them, through a button, a widget click or the ui/action API, the wall does not run it. It arms it and shows a prompt. You must press a key or click again.

Why can a plugin not just send “confirmed”?

Because that field does not exist. The request a plugin sends has no confirmed property, and the daemon never reads one, even if a plugin adds it. Confirmation lives only in the wall, where it comes from your keyboard or mouse. The decision is not something a plugin can pass in.

This is the same pattern the wall uses for its own dangerous keys. For example, x x closes a session and d d discards a worktree: the second press confirms.

What other limits protect where I type?

Two more rules close the obvious gaps:

Plugin actions also pass through limits: a plugin gets 5 ui/action requests a second, and a request must name a session or workspace that exists, or it fails with 422 before reaching the wall.

How do permissions fit in?

A permission is a word in the manifest’s permissions list. garage shows it before install, and the daemon checks it on each API call. A call without the right permission gets 403 with a message that names the missing one. Two permissions are dangerous: sessions.input and core.actions. Installing a plugin that has one needs --allow-dangerous in scripts, and an update that adds one asks again.

Use the least you need. A widget that only jumps to a blocked session needs ui.navigate, not core.actions.

What does the rule not protect?

The security model is direct about this, and so should this post be. Plugins are not sandboxed. A plugin is a program that runs as you. It can read your files, use the network, and call core routes outside the plugin API. The confirm rule controls what a plugin can make the wall do, not what the program can do on its own. permissions is a declaration that garage shows and checks on its own API. It is not a boundary.

garage also does not sign or review plugins, and finding one on GitHub by topic is not a recommendation. So:

If you find a way for a plugin to confirm a dangerous action itself, the security page asks you to report it to the maintainers as a security issue.

How do I write a plugin that respects the rule?

Do not fight it. If your action must be confirmed, mark it confirm = true in the manifest. That only adds a prompt for plugin-origin triggers, and a key press still runs it at once because a key press is you. If an action needs your sign-off beyond that, ask in your own pane.

The Build your first plugin walkthrough shows a complete example, and the security model and actions pages cover the rule in full.

Why did I build plugins this way?

I wanted claude-garage to stay very simple, with a flexible plugin architecture so anyone can add what they need. The confirm rule is what lets that stay safe to use: a plugin can add a button or a key, but the final yes on a dangerous action is always yours.

Sources

Keep reading

Install claude-garageMore posts