Self-healing builds
AI-written code is checked, repaired and verified before anything reaches your users, and a broken change never replaces a working app. This page explains what runs behind the Building… screen and what to do on the rare occasion a build still fails.
What is checked before deployment#
Every build and every chat edit passes through the same gates, in order:
| Gate | What it catches | What happens |
|---|---|---|
| Prompt safety | Instructions hidden in a description that try to make the AI reach for secrets or other customers' data | The request is refused with an explanation |
| Syntax | Python files that do not compile, JSX with unclosed tags, malformed code | Repaired mechanically, then re-checked |
| Imports and dependencies | Components that are imported but were never written; packages used but not declared | Missing pieces are generated or declared; undeclared packages are added |
| Routing contract | Pages written where the framework cannot find them, dead links in navigation | Normalised, with the navigation fixed to match |
| Build | Anything that only shows up when the app is actually compiled | A real build runs; failures go to an AI repair loop with a strict budget |
| Render | Pages that compile but crash when rendered | Each page is rendered once; a failing page is repaired on its own |
| Pre-publish security check | Sign-up open by accident, missing legal pages, a draft that answers publicly | Reported on the Security tab and in the launch check |
Only when every gate passes does the new version go live.
What happens when a repair is not possible#
- On a first build, if the custom work cannot be made to pass, the builder deploys the instant version of the same starter: a working, seeded app of the right kind, and tells you what it could not do.
- On a change to a working app, the change is not deployed. The previous version keeps serving, the files are reverted to the last snapshot, and the chat names the file that failed so you can rephrase. See Branches, versions and locked pages.
- If the app stops working later (for example after an integration change), the dashboard shows an incident with the reason and a Fix on web link. The same incident appears in the phone's Inbox.
Diagnose and fix on demand#
The dashboard has Diagnose and Fix all controls. Diagnose runs the same gates against the current files and returns a report: a health score, each finding with severity, the files involved and whether it can be fixed automatically. Fix all applies every automatic repair, snapshots first, and redeploys.
| Severity | Meaning |
|---|---|
| Critical | The app will not build or start |
| Error | A feature does not work |
| Warning | Works, but something is off (an unused dependency, a missing hint) |
| Info | A suggestion |
Keeping running apps healthy#
Deployed apps are watched. An app whose server keeps crashing is parked rather than left burning resources, and you are told why. Apps that go unused are put to sleep and woken on the next visit. Nightly, starters are rebuilt from scratch to catch drift between what a template says and what it ships, so a new app never starts from a broken base.
When to write to support#
Write to support@autoflowly.com with the app's address if:
- Diagnose reports a critical finding that Fix all does not clear.
- An incident has been open for more than an hour with no change.
- The app is published but the address shows the loading page for longer than a minute.
Include what you asked for last; it is usually the fastest way to find the cause.