The E2E workflow runs Tests/UI/OnboardingE2ETests.swift against a real
Home Assistant every night, mirroring home-assistant/android's own e2e
workflow. The flow onboards the app from the welcome screen, then drives the frontend into opening
the app's native settings screen, which is what proves the external message bus works end to
end.
homeassistant/ is a Home Assistant configuration directory with its .storage pre-seeded, so an
instance started against it comes up already onboarded, with a fixed administrator account:
| Username | Password |
|---|---|
citest |
h7jk99&U |
It is a copy of the fixture in home-assistant/android, minus the go2rtc config
entry, which needs a binary the container ships and a pip-installed core does not. Keeping the two
in step means both apps are tested against the same instance, with the same entities and credentials.
Start an instance. Use a copy: Home Assistant rewrites .storage and writes its database and logs
into whichever config dir it is given. The interpreter has to satisfy core's requires-python,
which is 3.14.2 or newer; pip refuses the install and names the version it wanted if it is older,
so there is no need to check the patch level up front.
cp -R .github/e2e/homeassistant /tmp/homeassistant && python3.14 -m venv /tmp/ha && /tmp/ha/bin/pip install homeassistant && /tmp/ha/bin/hass --config /tmp/homeassistantThen run the flow. The lane checks the instance first, walking the same login exchange the app performs during onboarding, and erases the simulator so the app starts at the welcome screen:
bundle exec fastlane e2eIt takes url:, username:, password: and device: options, all defaulting to the fixture above
and an iPhone 17. To check the instance without running the flow:
python3 Tools/home_assistant_e2e_auth.py --url http://localhost:8123 --username citest --password 'h7jk99&U' --timeout 300 --require-component mobile_appThe iOS Simulator shares the host's network stack, so http://localhost:8123 is also the address to
type into onboarding by hand on a simulator running on the same machine. This is the one place iOS
has it easier than Android, whose emulators need an egress tunnel and a DNS override to reach the
same instance.
Nothing enforces that this stays level with Android's copy, so when they regenerate theirs it drifts silently. To compare:
for f in $(cd .github/e2e/homeassistant/.storage && ls); do curl -fsS "https://raw.githubusercontent.com/home-assistant/android/main/.github/e2e/homeassistant/.storage/$f" | diff -q - ".github/e2e/homeassistant/.storage/$f" || echo "differs: $f"; donecore.config_entries is expected to differ: the go2rtc entry is deliberately removed here.
A failing nightly opens an issue titled "E2E: the nightly onboarding test is failing", labelled
automated, naming the step that failed and linking the run. Later failures are added to that same
issue as comments, and a closed one is reopened, so a bad week leaves one thread to read rather than
seven issues. Only the scheduled run reports: a manual dispatch is usually someone working on the
workflow itself. GitHub's own notification for a failing schedule reaches only whoever last edited
the cron line, which is why this exists.
The e2e-artifacts artifact carries a screen recording of the run under e2e-recording/, which is
usually the quickest way to see where the flow went wrong. Alongside it are Home Assistant's own log,
the /api/config it reported, and the full .xcresult.
Recordings are kept for passing runs as well, through uiTestingScreenshotsLifetime in
Tests-UI.xctestplan; Xcode would otherwise discard them on
success. That is what the test plan exists for. It costs roughly 160 MB per run, so the artifact is
kept for 14 days rather than the 90-day default.
- The test is not repeatable on a simulator that has already onboarded: it starts from the welcome screen, and the location and notification prompts it answers only appear once. Erase the device between runs, which is why the lane never retries a failure.
- The elements it drives are identified by
AccessibilityIdentifier, which is compiled into both the app and the UI test bundle, so a rename breaks the build rather than the nightly run. Only the frontend's own elements are matched on their copy, since the app does not own those.