programmiere

The maintainer loop

An AI maintains my plugin. It has to ask me first.

An AI reads the support forum, writes fixes, drafts replies and prepares releases for a WordPress plugin with about 4,000 installs. Then it stops and sends me a message. I answer with one word, "ja", and exactly that one thing happens. Nothing else.

How the maintainer loop worksA scout reads forums and reviews, a worker writes a fix on its own branch, deterministic checks verify it, and nothing is released, merged or posted until the owner approves that single action from his phone.Scoutreads the forumsWorkerwrites the fixChecksno model involvedGate„ja“one word per action,from my phonethe model only ever proposes —it never ships anything itself

The main question

How much should an AI agent be allowed to do?

Enough to do real work, but not enough to do damage on its own.

My answer is a strict split. The model may propose anything. Deciding and executing happen in plain code, and every step that can't be undone needs a human, every time.

Four roles

How the work is split.

The parts only talk to each other through files. There is no shared memory and nothing hidden.

  1. Scout rules

    Checks the plugin's forum and reviews, competitors' changelogs, form-builder forums and its own live spam data. Anything new goes into an inbox. No AI involved.

  2. Worker AI

    An AI coding session that runs unattended and implements one item on its own branch. Then a verifier in plain code, not the AI, decides whether the result may be proposed at all.

  3. Concierge AI + rules

    Talks to me over Signal. The model may word the messages and suggest "go" or "veto" for one existing item, but it can never execute anything.

  4. Gate rules

    The only part that can act on the outside world: releases, merges, forum posts. One approval per action, tied to my phone number. An approval never carries over to the next action.

Untrusted input

I assume someone will try to give the AI instructions.

Forum posts are written by strangers, and some of them may try to tell an AI what to do.

Outside text is fenced off

Anything from outside reaches the model clearly marked as untrusted data, never as instructions.

The approval question isn't AI-written

The text of what I approve comes from fixed templates. No model and no forum post can change the wording of the question I answer with "ja".

One question at a time

Only one confirmation can be open, and only for a short time. If a new notification arrives while a question is open, the old question is cancelled, so a "ja" can't land on the wrong thing.

Releases need a different word

A release to WordPress.org can't be confirmed with the word that approves a merge. The risk is different, so the word is too.

No reminders

A pending release shows up in the daily summary, but there is no reminder that gets more urgent, because the pace of changes is partly driven by text from strangers.

Fails closed

If the stored state can't be read, the model is broken or a marker is missing, the loop always takes the safe option.

From the phone

Releasing by text message.

I type "release 5.2.0". The loop prepares everything in an isolated copy: version bump, changelog from the commit history, validation. Then it shows me what will go out and waits.

An AI may improve the English of the changelog once, if I want. The raw text is shown right next to it, so I compare the two instead of trusting a rewrite. If the model doesn't answer, the release still goes out with the raw text. A model never blocks a release.

Monitoring

"Not measured" is not the same as "fine".

Small test runs, so-called canaries, check every day that the AI's sandbox still holds and that the model still behaves as expected. When a check can't run, it says so. It never reports "healthy" for something it didn't test.

Something I learned in practice: a loop that reports its own failures correctly still needs a second channel to reach me. In agent systems, a lot of the real engineering is not in the agent but in the monitoring around it.

By the numbers

How big it is.

~30,000lines of loop code across almost 100 modules
1,889automated tests; the test runner refuses any real model call
4roles: scout, worker, concierge, gate
1 wordfrom me before a prepared change goes out

The general idea

The same split works elsewhere.

The model proposes, plain code decides what is possible at all, and a human approves every step that can't be undone. I use it for a WordPress plugin, but it fits anything where an AI should act, just not alone.

Email me about it