What Protuno Can And Cannot Change
This is the part worth reading before anything else. Protuno does the work on staging. Your live site is only read. On top of that habit, the safety model rests on one setting you control per site, plus a clear-eyed view of what can and cannot be undone afterwards.
Read only and Execute: the two site modes
Every connected site is in one of two modes, changed from the site’s own screen:
| Mode |
What it means |
| Read only |
Analyses and reports. Changes nothing. |
| Execute |
May change files, the database and settings. |
A site you connect as a live site starts in Read only. A site you connect as a staging site starts in Execute, because staging is where Protuno does the work. Leave your live site in Read only. Fix on staging, check the result there, then move the fixes to live in your own way, with your host’s push-to-live or the tools you already use. That last step is yours, not Protuno’s. If you are connecting a staging install that matters, check the mode after connecting rather than assuming it. Only you can switch a site to Execute, and only in Protuno. Your AI app can set a site to Read only, never to Execute.
The mode is the control
The site’s mode is what decides whether anything can be written. A site set to Execute grants write access to any playbook that runs on it. A site set to Read only does not get written to: every tool that could change the site is removed from the run’s inventory before the model sees it, so it cannot make the change rather than being asked not to.
The Read Only / Can Write label on a playbook card is real and is enforced, but only for scheduled runs: a routine for a playbook that can write will not fire on a read-only site, and the routine row says so. 17 of the 27 only read. The other 10 can change something.
Two limits worth knowing, because the guardrail is good rather than absolute. Tools are removed by matching their names against a list of changing verbs, so an ability the plugin adds under an unfamiliar name would survive. And the REST tool is deliberately kept, because it is the main way a site is read at all; on a read-only site it is held to reads by instruction rather than by removal. Neither is a reason to distrust the mode. Both are reasons to do the work on staging and keep your live site in Read only, rather than treating the mode as the only thing between an agent and the site you care about.
You approve before anything changes
When you ask a Super Agent for a job, your AI app shows you the plan, and the plan appears as a card in Protuno. Only you can approve it, by pressing Approve there. Your AI app cannot approve for you. Each agent also has a trust level on each site, set on the agent’s page under On this site:
- Read. Looks, never touches. Runs only playbooks that read the site.
- Ask. Plans, then waits for your yes. Every step that changes the site asks again first, with Approve this change on the job card.
- Auto. After you approve the plan, it runs without stopping to ask. Playbooks that need a backup make one as their first step.
New agents start at Ask. On top of that, the work runs on your staging site, that your live site stays in Read only unless you change it, and that the sequence about to run was written down in full before it ran, so you can read it first.
A playbook you build yourself can carry an approval checkpoint, which holds the run at that step until somebody approves it. The run then shows as Waiting on you in Activity. None of the 27 ready-made playbooks uses one.
If you want a conversation before a change, have it in your AI app. It proposes, you approve in Protuno, and nothing is applied until you do.
Rolling a run back
Every run records what it changed, as it changes it, by re-reading the target and confirming it is different. The run’s report carries a Rollback this run button, and opening it shows exactly what would be put back and what would not, before you commit to anything.
The record is the honest part. A change made through a named capability is recorded with enough to reverse it. A change made by running arbitrary PHP is recorded as “cannot be undone”, because nothing observed what the snippet did. So a report will tell you “every change this run made can be undone”, or “4 changes can be undone; 2 cannot”, rather than offering a button that might mean either.
What rollback fully covers
Three playbooks are known to make only reversible changes, and they are exempt from the pre-run backup gate for that reason:
| Playbook |
How it goes back |
| Plugin Safe Update |
Reinstalls the previous version from wordpress.org. It also rolls itself back during the run: each plugin folder is stashed before its update and restored if the site stops working. |
| Spam Comment Deleter |
It never deletes. It moves comments between moderation states, and rollback moves them back. |
| Broken Link Checker |
Each edit lands as a post update that records the post’s previous body, so rollback restores the text. The edit is also a WordPress revision, so you can restore it by hand from the post history. |
What rollback does not cover
Everything else. WP Doctor, WP Hardening, Database Optimisation and AI SEO Optimiser all change things by running PHP on the site, which records “cannot be undone”. Those runs get the backup gate instead: the run refuses to start until a restorable backup exists.
Database Optimisation is the clearest case, and the one worth reading twice. Its deletes are permanent. It does not run table optimisation and it does not use raw SQL; it removes revisions, spam and other bloat through WordPress’s own delete functions, and it skips any category that looks anomalous. Its second and third steps take a database backup and prove that backup is sound, before anything is deleted. That backup is the way back. Take your own as well if the site matters.
Be precise about what is checked, because it is more than it sounds and less than a guarantee.
A backup step does not just report that it ran. Protuno re-opens what it wrote: the archive has to open and contain files, the database dump has to be large enough to be real, decompress, carry the statements a dump carries, and be free of the specific corruption a past run produced. For a database-only backup, every table that exists on the site has to appear in the dump, and any that are missing are named. Only then does the verdict read Verified, and a run that needs a restore point will not proceed past a Failed one.
What it does not do is restore the backup to see what comes back. Nothing is restored, extracted or written during the check.
So it is a verified file rather than a rehearsed recovery. Run anything destructive on staging first, and before you move the result to a site that matters, take your own backup of that site as well and know it works.
The credential Protuno holds is an application password issued by your own WordPress user. You can revoke it from Users, Profile at any moment, without contacting us, and the connection stops at once.
Leave a Reply