Brain, ledger and undo

What the agent knows about the project, what it did, and how to get back.

The project brain

Per site, the app keeps what the work is for: the purpose, what was agreed, what must never be touched, and an open task list. Your agent reads it at the start of a conversation instead of being told again.

It cannot delete a rule you set, and it cannot promote one to apply everywhere. That stays with you, in the app. An agent that can rewrite its own constraints does not have constraints.

The ledger

Every edit is written down before it happens, with the property-level before and after. Not "the hero section was changed", but the specific property, the old value and the new one.

Thirty edits to one page inside a minute are merged into one record rather than thirty. A log nobody can read is not a log.

Checkpoints

Name a checkpoint before starting something you are unsure about, keep working, and roll back to it. The app also writes one when a conversation ends: what got done, what is in progress and what is blocked. It hands that back when the next one starts, so a context window running out is not the same as losing the thread.

Undo, in three layers

1. Drafts. Pages are built as drafts. Publishing is a separate, deliberate swap, with a backup taken immediately before it.
2. A backup before every write. Going back is restoring one file rather than reasoning about what changed. PHP that will not parse is refused before it is saved at all.
3. The watchdog. The site is checked after each change. If it stops answering and the last change was ours, it is reverted without being asked.

That last one is the difference between a bad afternoon and a phone call from a client.