Permission cards that explain themselves

A permission card shows the exact command a session wants to run, and the session's own label for it, such as "Delete the policy file". The command is often hard to read, and the label is only what the session claims. So every card also says, in plain words, what the command will really do. You read that and decide, instead of working out a line of PowerShell at midnight.

It works for Claude Code, Codex and Grok, in the studio and in the iPhone app, which is not out yet. Answering the card itself is covered in Answering permission cards and questions.

What the box says

Under the line that says why the card asks, a box What it really does holds two kinds of lines.

First, two or three plain sentences written by a model, read from the command itself. For example: "It permanently deletes one small settings file, policy.json, outside this project. Then it only checks that the file is gone. Nothing is sent anywhere." Under them: "Written by Claude from the command itself, not from the session's own label." While the model writes, the box says "Reading the command...".

Then the facts the studio knows itself, one a line. They cost nothing and are never a guess:

• what the command deletes, said as deleted for good, because a delete from the shell skips the Recycle Bin and the Trash;
• for a folder, how many files it holds; for a file that is not there, that it is not there;
• the folders outside the project it reaches;
• that it goes over the network, for commands such as curl, ssh, git push or npm publish;
• for a destructive command with none of the above, that it cannot be undone;
• which session wrote a file it deletes or changes, and when, such as "This same session wrote policy.json itself, at 00:41".

The facts show even when no model answers. Allow and Do not allow never wait for any of it, and nothing is allowed or refused because of what the box says. It is words for you only.

Who writes the sentences, and what it costs

The project's own subscription writes them: the account chosen for the project's mail and map, with its cheapest model. It is never another account and never a key of ours. The model is told that the session's label is a claim, which a confused or hostile session could make sound harmless, and that it must describe what the command does, not repeat the label. It never says whether to allow.

Each distinct command is asked about once, and the words are kept and shown again with no new question. A plain request, such as reading or writing a named file, fetching a page or a search, gets the facts only, with no model. The sentences count in your usage like any small prompt.

When the account cannot answer, because it is signed out, at its limit or the tool is missing, the card keeps the facts and says nothing false. The studio does not ask again for ten minutes, or until you choose another account for the project.

Turn the sentences off

1. Open Settings.
2. Under How sessions work, find Explain each permission request.
3. Turn it off.

With it off, no model is asked, and the card shows only the facts the studio knows itself. It is on by default. Which requests reach a card at all depends on the permission mode; see The four permission modes.

Questions