Safety: sandbox & approvals

Command sandbox

By default, commands run in a sandbox — a working copy of the project in a temp folder: the agent experiments freely, and changes are transferred to the main folder after successful completion. This protects your files from unexpected side effects.

Heavy directories (Unreal Engine build folders: Content, Binaries, Intermediate, Saved; IDE caches like .vs; node_modules, etc.) are not copied — they are linked via junctions, so commands, including builds, work with them directly and no extra disk space is used. Folders larger than 512 MB are linked automatically. Orphaned sandboxes (if the app was closed mid-command) are cleaned up automatically.

The copy budget is configured in Settings → Agent (2 GB by default): folders above the budget are linked instead of copied, so even huge projects take no extra disk space.

File operations (read/write) are always confined to the workspace folder — the agent cannot escape it.

Approvals

• Normal mode — every potentially risky action (command, write, delete) shows an approval card: the tool name, a summary, a before/after diff for file edits, and «Allow» / «Reject» buttons.
• Full access — a toggle in the chat header that lets the agent act without per-step approvals. Enable it consciously, when you trust the task.
• Decisions are protected by one-time tokens (they cannot be replayed), and an approval card stays valid for 2 minutes — after that the buttons expire and the action must be requested again.

Plan approval

In Plan mode, risky steps start only after the plan is approved: the plan card appears right in the chat, and the decision takes one click.

SSH host keys

On first connection to a server, its host key fingerprint is stored. Later connections verify the fingerprint — protection against server impersonation (MITM).