The step names said "Set up Python 3.12" while the actual version is set by python-version and updated by Renovate, so the names got out of sync.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The GitHub cache is limited to 10 GB per repository. About 15 jobs were writing the Gradle
cache on develop, each adding up to several GB (dependencies, transforms and build cache),
so the cache was always at the limit and entries were evicted. A job could then restore a
Gradle User Home without its matching dependencies entry, which broke the nightly build on
all the PRs.
Only the debug build job and the unit tests job now write the cache on develop. All the other
jobs only read it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Add --continue to the unit tests, screenshot tests, detekt and ktlint
Gradle invocations, so that a failure in one module does not prevent
the other modules from being checked, and the CI reports all the
failures at once.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The enterprise repository is private, so only mention the number of new
commits in the generated PR description.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
When a PR is merged on develop with the enterprise submodule targeting a commit
which is not on the enterprise main branch, open a PR to update it to the latest
commit of main. The workflow also runs daily and manually to keep the submodule
up to date.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each Maestro run logs in to @maestroelement:matrix.org, and the sessions
pile up on the account. Run tools/mas/delete_all_sessions.py at the end
of the Maestro job, even if the tests failed, to end them all.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Recording rewrote every screenshot unconditionally, so renderer noise far below
the verification tolerance produced large "Update screenshots" commits.
Use `app.cash.paparazzi.overwriteOnMaxPercentDifference` to keep screenshots that
are within tolerance. Existing screenshots are no longer deleted before recording;
obsolete ones are pruned afterwards, only when their preview produced no
screenshot at all.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* Remove Roborazzi, use Paparazzi for Compound
The new screenshots will be recorded with the rest of the Paparazzi ones. Note the names changed, since Paparazzi enforces a `$packagename.$testname.$name.$extension` format
* Set a max width value for color previews
* Merge the `verifyPaparazziDebug` CI steps
* Update screenshots
---------
Co-authored-by: ElementBot <android@element.io>
* Set `maxParallelForks = 1` for most tests, except for screenshot ones
It turns out the Robolectric setup penalty compounds the more parallel JVMs you spawn. Using 1 JVM vs 8 was up to 2.5x faster in local tests.
* Split the roborazzi tests tasks config and CI step
Secrets are never exposed to pull_request runs from forks, so checkEnv.js
throws on MAESTRO_PASSWORD after a full APK build and emulator boot, and the
red check is indistinguishable from a real test failure. The existing guard
checked github.event.pull_request.fork, a field that does not exist on the
pull_request payload (it lives at head.repo.fork), so it was always true.
Gate both jobs at job level on head.repo.full_name == github.repository
(workflow_dispatch still runs), and drop the dead step-level condition.
Fixes#7653
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
With the default (queue: single), a run that finishes Build APK while
another run is already pending in the maestro-test group cancels that
pending run, which then never gets a Maestro verdict for its commit.
queue: max keeps the one-at-a-time constraint (no cancel-in-progress) and
lets up to 100 runs wait, at no runner cost.
Fixes#7654
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The PR metadata checks move to a new PR Checks workflow, and the lint,
ktlint and detekt reports are now surfaced as inline annotations from the
quality workflow. This removes the danger-js dependency, its ad-hoc npm
installs and the need for the ElementBot token, so the checks also run on
pull requests from forks.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
recordScreenshots.sh runs under `set -e`, and the `git commit -m "Update
screenshots"` at the end of it is unconditional. When a recording run
produces no diff — the common case once the snapshots are already up to
date — `git commit` exits 1 with "nothing to commit, working tree clean"
and takes the whole workflow down with it, reporting a red job for a
successful run.
Stage as before, then return early when the staged diff is empty. `set -e`
is left in place: it was added deliberately so that a failed Gradle run
cannot push a commit that deletes every snapshot. Using `git diff --cached
--quiet` as an `if` condition is `set -e`-safe, because its non-zero "there
are changes" status is consumed by the condition. The early return also
skips the push, which had nothing to send.
* Check if splitting the steps for testing and generating the reports improves performance and keeps the coverage stats
* Exclude Konsist tests and include `:app` ones
* Remove the extra dependency of `koverVerifyAll` on `:app:koverVerify{variant}`
New locales added via Localazy would change the locales line in
aaptDump.txt and break the manifest CI check if the reference file
is not kept in sync.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
