How fixes stay safe
What the system may do, what it never does, and how you stay in control.
What the AI may do
- Read your servers and websites with checks that change nothing.
- Propose a fix, with each step and its risk, for an owner to approve or reject.
What it can never do
- Run anything without an approved plan.
- Run a destructive step. Those are shown as advice only.
- Edit SSH configuration. That is the management path, so it is never edited.
- Run config lines that load code, such as an nginx module load or a php extension.
Risk tiers
Every action has one tier. The tier decides whether it can run, and how.
T0 Read-onlyRead-only
- read disk usageRuns on its own
- read memoryRuns on its own
- read loadRuns on its own
- read uptimeRuns on its own
- read cpu countRuns on its own
- read systemd failedRuns on its own
- read service statusRuns on its own
- read log tailRuns on its own
- validate nginx configRuns on its own
T1 Low impactLow impact
- restart serviceRuns only with your approval
T2 Config changeConfig change
- install packageNot enabled yet
- upgrade packagesNot enabled yet
- edit configNot enabled yet
- restore configNot enabled yet
T3 DestructiveDestructive
- restart management unitAdvice only, never runs
- rebootAdvice only, never runs
- delete pathAdvice only, never runs
- format diskAdvice only, never runs
- create userAdvice only, never runs
- firewall allowAdvice only, never runs
How approval works
- Only an owner can approve a plan.
- The approval link works once and is valid for 30 minutes. Or re-enter your password within 5 minutes.
- An approval holds for 30 minutes. Steps run only while the server still matches the plan.
What happens after you approve
- Each step is checked again right before it runs.
- The fix is verified 3 times after it runs.
- A config edit that fails its check is undone automatically, and the undo needs its own approval.
The emergency stop
An owner can stop every fix that has not finished. No further step starts, and no approval link works. A running command finishes its current step. Nothing is undone automatically by a stop.
What is logged, and what is never stored
- Every check, AI step, command, approval and notification goes to the audit log. Entries cannot be edited or deleted.
- Passwords and keys are never stored in plain text.
- Config file contents are stored sealed, and members never see them.