Autopilot

import { Aside } from ‘@astrojs/starlight/components’;

Autopilot controls how much an agent can do on your behalf without pausing for approval. All settings are off by default — Ptah asks before touching anything that writes to disk or runs a shell command.

Where it lives

Settings → Autopilot, or autopilot block in ~/.ptah/settings.json:

{
  "autopilot": {
    "enabled": false,
    "autoApproveReads": true,
    "autoApproveWrites": false,
    "autoApproveShell": false
  }
}

Toggles

Toggle	Effect when enabled
enabled	Master switch. When false, every other toggle is ignored and all actions prompt.
autoApproveReads	Allow Read, Glob, and Grep without prompting. Safe for almost all workflows.
autoApproveWrites	Allow Write and Edit to workspace files without prompting.
autoApproveShell	Allow Bash / PowerShell commands without prompting. High risk.

`autoApproveShell` lets agents run arbitrary commands on your machine. Only enable it inside throwaway environments (dev containers, VMs) or when you fully trust the agent configuration.

Per-session override

Each chat has an “Autopilot” toggle in its header. It inherits the global setting but can be disabled for a single session — useful when you want tighter oversight on a production workspace.

Hooks

Approval prompts are themselves hooks. You can override the behavior per-workspace by adding a permissions block to .claude/settings.json:

{
  "permissions": {
    "allow": ["Read", "Grep", "Glob"],
    "deny": ["Bash(rm*:*)"]
  }
}

Workspace permissions are merged with global autopilot settings. A deny rule always wins over an autopilot allow.