Safety Boundary

Default read-only behavior, approval boundaries, risky command rules, and redaction requirements for DebianClub AI Skills.

The value of AI Skills is not only knowing Debian commands. The more important goal is preventing an agent from changing a system before it has enough evidence. debian-linux-reliability defaults to read-only diagnostics.

No Automatic Modification By Default

By default, the agent must not automatically run or directly recommend these actions:

Install, remove, or upgrade packages
Edit /etc/apt/, /etc/systemd/, /etc/network/, /boot/, or /etc/fstab
Change firewall rules
Restart services or reboot the system
Modify disks, partitions, bootloaders, or permissions
Run remote scripts or pipe-to-shell commands

If the user explicitly asks for a change, the agent must first explain the impact, inspection step, verification step, and rollback path.

Read-only First

Prefer commands like:

cat /etc/os-release
dpkg --print-architecture
apt-cache policy package-name
systemctl status service-name
journalctl -u service-name --no-pager -n 100
ss -tulpen
ip route

Read-only does not mean safe to share without review. Logs and configuration can include tokens, usernames, hostnames, internal IPs, or private-key paths. Review and redact before sending output to external systems.

Risky Command Detection

These command categories need extra review:

Type	Examples
Remote scripts	`curl URL
Deletion	`rm -rf`, directory clearing, wildcard deletion
Permissions	`chmod -R 777`, recursive `chown`
Disk	`dd`, `mkfs`, `parted`, editing `/etc/fstab`
Repositories	Adding third-party sources, importing GPG keys, changing codenames
Services	Restarting ssh, display managers, networking, or firewall services

Review a single command with:

bash skills/debian-linux-reliability/scripts/risk-check.sh -- "command here"

Change Proposal Template

Any proposal that mutates the system should include:

Will change:
- ...

Why:
- ...

Inspect before:
- ...

Apply:
- ...

Verify:
- ...

Rollback:
- ...

Redaction Requirements

Before sharing logs, configuration, or agent responses, remove:

Tokens, API keys, passwords, cookies
Private keys and certificate contents
Customer names, hostnames, real usernames
Public IPs, internal topology, sensitive domains
Database connection strings and .env contents

Built-in script redaction is a helper, not a replacement for human review.

Next: Evaluation & Maintenance.