Upkeep

Most of brAIn is about your house. House → Upkeep is about Home Assistant itself: the names nobody tidied, the devices in no room, the updates you are not sure about, the error log nobody reads, and the knowledge of how the house works that lives in one person’s head.

It has five sections, and every one of them is a press. Each section says what it last did, when, and — for the parts that also run on a schedule — why a scheduled pass was held back (automatic insights are paused, nothing it reads has changed). A section whose store could not be read says so, rather than showing up empty.

Nothing in Home Assistant changes on its own

Each button spends one Claude run. Suggestions are a table you tick; a verdict on an update is advice; the overnight check files findings; the house book is a page. The only things on this tab that write to Home Assistant are Apply and Undo in the names-and-rooms table.

Names, rooms and aliases

Press Suggest names and rooms and brAIn reads the house’s own naming style first — a sample of the names you chose — and then proposes one table of changes:

Kind	What it fixes
Rename	Names that are still the hardware’s — 0x00158d0001a2b3c4, Shelly Plus 1PM 8A4C — which are unfindable in a picker and unsayable to Assist
Room	Entities and devices in no area, which “turn off the kitchen” and every area card cannot see
Alias	Spoken aliases for the things people actually say

Every row has a tick box. Nothing changes until you press Apply, and Apply goes through Home Assistant’s own registry.

A room move says what it changes

An automation that targets an area follows the room. Moving a lamp out of the lounge stops the lounge’s “evening” automation reaching it — so a room move lists the automations that will start or stop reaching that entity, before you tick it. It matches on what each automation actually does, so a light service never claims a sensor in the same room.

What it refuses

A suggestion is dropped, never offered, when it:

• names an entity it was not shown;
• moves something into a room that does not exist;
• touches anything on your protected_entities list, or a device that carries one;
• is still shaped like a hardware name;
• gives a name another entity already answers to — in the house, or earlier in the same table;
• changes nothing.

Undo, for 30 days

An applied batch can be undone for 30 days, because a registry change is the kind of thing somebody notices a fortnight later. Undo is field by field, and only puts back a field that still holds what brAIn wrote: a rename you made since is yours, and an undo that overwrote it would be a second change nobody reviewed. Only the aliases this batch added are removed.

Apply re-reads every field from Home Assistant before writing and records only what Home Assistant accepted, so a batch that half-applied undoes exactly the half that happened.

From a finding

The named after its hardware (reg.hardware_name) and no room (reg.no_area) findings offer Fix it, which opens this table rather than planning one rename at a time.

Updates waiting

For each pending update, press Is it safe tonight?. brAIn reads the release notes, searches your configuration for the integrations and options they mention, and answers one of:

• Safe tonight;
• Wait — with the reason, and the edit you would need to make first;
• Can’t tell.

A safe or wait has to quote its evidence: one line from the release notes and one line from your configuration, and both are checked word for word against what was actually read. An answer that cannot quote both is shown as can’t tell, and so are notes that could not be fetched — a verdict on notes nobody read is a guess, so no run is spent on one.

The release notes are treated as data, never as instructions, and the run that reads them has no tools at all.

What it never does

• It never installs anything. There is no button here that runs an update.
• It never reads secrets.yaml. Any line in your configuration whose key looks like a credential is blanked before it is sent, and Home Assistant’s own storage and the recorder’s database are skipped.

Overnight health check

Once a night, between 03:00 and 06:00, and whenever you press Run the check now, brAIn reads:

• Home Assistant’s error log;
• the Zigbee mesh — routers, weak links, devices not seen;
• Z-Wave node statistics and node status;
• automations that switch a smart plug feeding a Zigbee router.

It asks for at most five root causes, named as causes rather than symptoms:

The hallway plug that the upstairs Zigbee sensors route through keeps being switched off by the bedtime automation.

Every cause must cite the records it rests on, by an id brAIn handed out — a log line, a Zigbee device, a Z-Wave node, an automation. A cause that cites nothing, or cites a record that was not there, is dropped and counted. The causes reach the Findings tab through the same first look every finding takes.

A cause re-worded tomorrow is the same finding, not a new card every night: it is keyed by the records it cites. And on a night a source could not be read, the findings that rest on it stay put — I could not look is not it went away.

The scheduled run answers to the same switches every scheduled run does: a Claude sign-in, Automatic insights, and the usage budget. A night with nothing in the log or the mesh worth a look costs nothing.

House book

Press Write the house book and brAIn writes a short manual of how this house works: what the automations do, how the heating runs, what the alarms do, where the shutoffs are, and anything else worth knowing — from your own automations, scripts, scenes and what brAIn has learned. It is meant for a partner, a sitter, or the person buying the house.

• Every sentence is cited. Each one carries the automations, scripts or facts it came from, and a sentence with no source is dropped.
• Secrets are left out, always. Codes, PINs, passwords and anything shaped like a token are removed from every sentence, whether or not the book is ever shared.
• Gaps become questions. Where it cannot answer from the house — where is the water shutoff? — it asks on the Findings tab, at most three per run and five open at once. Your typed answer goes into memory as your own words, so the next book knows. A subject asked once is never asked again.
• It keeps itself current. Once you have asked for one, it is rewritten weekly — but only when something it reads has changed. A held week says why.

Sharing it, and taking it back

Publish a link puts a read-only, redacted copy with no script at an unguessable address under your Home Assistant’s /local folder, and asks search engines to keep away. While it is published, every rewrite updates the shared copy.

Take the link down deletes the page and changes the address, so a link somebody kept is dead. The book’s link has its own secret, separate from the one your shared insight cards use, so taking it down never breaks a card on a dashboard.

Who can reach the house

Four house checks look at access, and file ordinary findings:

Check	What it reports
sec.lock_cloud_voice	A lock or alarm panel exposed to Alexa or Google through Home Assistant Cloud
sec.addon_unprotected	An add-on whose protection mode has been switched off
sec.login_bans	Recent failed logins that got an address banned
sec.brain_posture	A terminal session still acting without asking after Let brAIn act without asking was turned off — turning it on is your choice and files nothing

Each says nothing when it could not look, and Wrong, because… stops it for that lock or add-on.

Once a week brAIn also writes one sentence about who and what can reach the house, from what the last checks pass read — no new look at the house of its own. It shows here, with when it was written; Review now writes it on demand. It answers to the same switches as every scheduled run, and a review that failed waits six hours before trying again rather than spending a run every few minutes.