Overnight Self-Healing

Requires brAIn 1.45

This page documents a release that is on its way. Everything below is built and tested; it is not in the current add-on yet.

Every other thing brAIn does to your house is a press you made on a specific card. This one is not: it runs while the house is asleep and nobody is looking, which is the whole point and also the reason it is the narrowest thing in the add-on.

It is off by default. self_healing ships false, and a house does not start healing itself because it was updated.

The three, and nothing else

Each one is keyed to a finding the house checks already file, and each is a call you would have made yourself:

The finding	What it does	Why it is safe to do unattended
An add-on set to start at boot is not running	Starts it, through the Supervisor	It was meant to be running; this is what boot would have done
A Z-Wave node the controller has marked dead	Pings it	A read with a side effect — the controller re-marks the node if it answers. Nothing in the house moves
An integration that failed to set up	Reloads that config entry	A reload of something already broken cannot make it more broken

That is the complete list. A finding from any other producer is not something self-healing touches, and there is no way to add one.

What it will not do

No power-cycling of anything. A plug switched off and on again has a freezer behind it.
No restart of Home Assistant, and no restart of brAIn.
No Claude run anywhere on this path. A model choosing what to restart in your house at 3am is a guess wearing a repair.
An add-on in an error state is never started. The Supervisor could not start it, or it started and exited; asking again gets the same answer. That is a case for you and a log, not for a loop at 3am.
A boot: manual add-on is never touched. You stopped it on purpose.

When it runs

Once a night, and only inside a window you have already defined:

Your quiet hours, if you have set them. They are the house saying when it is asleep, and a second answer derived from anything else would disagree with them on the nights the two drift apart.
An hour after this house settles, if you have not. That is the settle time brAIn has measured — see rhythm — plus an hour, because bedtime is when people are still moving about. The window stays open 45 minutes, so an add-on restarted at 04:00 does not heal at lunchtime.
With neither, it does not run at all — and ⚙ Diagnostics says so, in as many words. Running at an hour nobody set would be acting on a guess about when nobody is looking, and a self-healer that has never run looks exactly like one that had nothing to do.

A night is keyed around local midnight rather than at it, so 23:40 and 03:10 are the same night. A key that changed at midnight would let a pass at 23:50 and a pass at 00:10 both run, twenty minutes apart.

The floors

Every one of these is a refusal, and each is tested against the mutation that removes it.

At most three repairs a night. A house with nine broken things at once is not a house to fix unattended; it is a house to look at. The plan is ordered oldest-first, so a night that hits the cap takes the same three every time rather than whatever order the store happened to be in.
One attempt per problem per night, written to disk after every attempt rather than at the end — so the add-on restarting at three in the morning does not start the same add-on twice. The unit is the finding rather than the target, because the finding is what clears and the clearing is the verification: a house with three add-ons down gets one started and keeps its row.
Never something you have already answered. Only open findings. If you pressed Fix it, Wrong or Remind me later on that row, it is a conversation you are already having with brAIn.
Never a protected entity, area or device. protected_entities is enforced at the MCP server, the chokepoint every Claude path to your house crosses — and neither a Supervisor request nor an unattended service call is one of its callers, so this path asks the same question with the same patterns. A ping reaches the box, so every entity on that node’s device is checked, not only the sensor named on the finding.
An unresolvable target is skipped, not guessed. If brAIn cannot work out which entities a repair would touch — an older Core carries no config-entry link on its registry rows — and your protected list is not empty, it skips. “I could not tell” and “nothing is protected” are different answers, and only one of them is safe to act on.
Never off something it could not read. If the Supervisor did not answer this pass, that is not every add-on having come back. A snapshot brAIn could not fetch means I could not look, which is a skip.

A failed call is recorded and not retried the same night. That needs no machinery: the attempt is written down whatever happened to it, and the store is what the next pass reads.

It does not check its own work

Nothing here reads back whether the call worked, deliberately. A 200 from the Supervisor is a request being accepted, which is a different claim from an add-on that is running.

What proves a repair is the next house-checks pass: the finding clears, or it does not. And the morning brief is where you find out which:

brAIn started the Mosquitto broker add-on at 03:10; it is working now.

brAIn pinged the Z-Wave node Back door sensor at 03:11; it has not cleared yet.

brAIn tried to fix something overnight and could not: it reloaded the Hue integration at 03:10 — the call failed.

Something mended while the house was asleep is itself a reason to send that morning’s brief, which is otherwise silent on the mornings there is nothing worth saying.

Where to see it

⚙ Diagnostics in the panel: whether it is on, when it would next run, what last night attempted and skipped — and, when it did nothing, which of the three silences that was.
The run journal, which records every attempt as healed, heal_failed or heal_skipped, with the skips and their reasons.
brain report, which bundles both for a bug report.

There is no button that runs a healing pass on demand. The window is the point: the only moment this is allowed to act is the one you said nobody would be looking.

Turning it on

self_healing in the add-on’s configuration, described there as Fix three things overnight. Turning it on is a decision about how much you trust everything above it, and that is not a decision an add-on should make on your behalf.

Set your quiet hours at the same time, unless you are happy for brAIn to wait until it has measured your house going quiet.