Permissions
Claude asks before it acts. In Vitrino the asking is a card pinned to the bottom of the transcript, and it is the one part of the interface designed to be impossible to skim past.
What a card shows
A permission card cannot be collapsed and does not scroll away. It stays until you answer, and it shows what will actually happen — not a summary of it:
- For a command, the whole command. Never a shortened form of it.
- For an edit, the diff it would make.
- For a new file, the contents it would write.
- For a fetch, the URL, and what Claude means to ask of the page.
- For anything else, every argument, in full.
Whatever the tool, All arguments opens the complete raw input. A card never truncates the thing you are being asked to approve.
Commands worth a second look
Some shell commands are marked: deleting recursively, escalating privilege, piping a downloaded script into a shell, force‑pushing. The risky part of the command is highlighted within it and named in plain words above it.
This is presentation only. The choices are the same, in the same order — a pattern match is nowhere near good enough to justify making a button harder to reach. It is there so a dangerous command cannot look like every other one.
An edit is shown as the change it would make, not the arguments that describe it:
Your choices
| Choice | What it does |
|---|---|
| Allow | Runs this, once. |
| Always allow | Runs this and saves a rule so you are not asked again. |
| Deny | Refuses, and Claude carries on and tries something else. |
| Deny and stop | Refuses and ends the turn — the equivalent of Esc in the terminal. |
| Add a reason | Tells Claude why, or what to do instead. Works with either Deny. |
Always allow lists exactly what it will save and where before you press it — which rules, in which settings file. Nothing is written that you have not been shown.
Important
Permission decisions deliberately have no keyboard shortcut. Approving tool use takes a deliberate click, so it cannot be fired by muscle memory.
A tool Vitrino has no card for still shows everything it was asked to do:
Permission modes
The mode chip beside the message box changes how much Claude asks. These are Claude Code's own modes, described in its words:
| Mode | Behaviour |
|---|---|
| Ask before acting (default) | Prompts before operations that need permission |
| Accept edits | Accepts file edits without asking |
| Plan mode | Plans only; no actual tool execution |
| Auto | A model classifier approves or denies permission prompts |
| Don't ask | Never prompts; denies anything not already allowed |
Claude Code's bypassPermissions — approve everything, unconditionally — is not offered.
Mode is per session. Changing it in one tab does not change another.
Rules you already have
Vitrino reads the permission rules in your project and user settings exactly as Claude Code does, so anything you have already allowed stays allowed. See Project configuration.
When you are not looking
If Claude needs you and its window is not focused — or not open — Vitrino posts a system notification naming the project and what is wanted. Clicking it opens that window. Only waiting states notify; finished turns and background work do not. You can turn this off in Settings → Notifications.
Next
Writing messages covers the other half of the conversation.