Run the same suites locally and in GitHub Actions:
| Suite | Command | Coverage |
|---|---|---|
| Android | bundle exec fastlane android checks |
JVM logic, MockWebServer HTTP integration, Robolectric database/preferences/provider integration, Compose interactions, localization, manifest security, light/dark contrast, debug/release lint and APK builds, unsigned release verification |
| Emulator | ANDROID_HOME=/path/to/sdk ANDROID_SERIAL=emulator-5554 bundle exec fastlane android device_tests |
Real Android Keystore, encrypted preferences/SQLite queue and retry identity, Android HTTP/ACK/redirect behavior, FileProvider access boundaries on API 26/35 |
| Python | PYTHON=.venv/bin/python bundle exec fastlane android python_checks |
DevServer HTTP test-client behavior against a local HTTP server, release automation with mocked APIs/local Git, pinned Black/isort style checks |
| n8n | bundle exec fastlane android server_tests |
Live receive/validation, notification/SMS/test payloads, HTTP failure, invalid ACK and timeout |
For Python, create .venv with python3 -m venv .venv and install requirements-dev.txt.
For n8n, first run docker compose -f DevServer/compose.yaml up -d --wait --wait-timeout 300.
The server suite uses synthetic data and retains it in the development execution history.
.github/workflows/ci.yml runs Android, Python, n8n and separate API 26/35 emulator jobs on pull requests, main pushes and manual dispatch.
Manual dispatch with release_dry_run_version runs only the release dry run and skips these suites.
Android XML/HTML test and lint reports are uploaded even if a check fails. Successful Android jobs
also publish separate debug APK and unsigned release APK artifacts, expiring after 14 days.
The trusted pr-artifacts-description.yml workflow keeps their links in a marked PR-description
block, preserving author text and rejecting old heads/runs. It never executes PR code or posts
comments. It starts working after merge into the default branch. These checks never sign release artifacts.
Local success does not establish a GitHub run result for uncommitted/unpushed changes.
The applicable categories match MegaProxy: JVM logic and Android integration, Compose interactions under Robolectric, resource/security/UI contracts, Python tests and formatting, lint and builds. Message487 also has a live Docker/n8n integration suite. MegaProxy's Go race tests and optional native-parser fuzz lane apply to its Go/JNI networking core; Message487 has no native core. Message487 has its own release-automation Python tests, but no MegaProxy change classifier or ancestor-check history lookup.
Compose tests use the real navigation, screens, ViewModel and preferences with a test Application that suppresses startup workers and process-wide crash-handler installation. An explicit ViewModel factory binds each test to its own Application; background completions are drained before assertions. Tests cover navigation/back, connection validation and persistence, bulk application selection and diagnostic deletion. They do not prove notification/SMS permission delivery, WorkManager/OS scheduling, real Android Keystore, process death handling or email-client behavior. Those remain device/emulator checks; see diagnostics for the crash/report scenario. The separate emulator checks exercise real Keystore, storage, HTTP and provider APIs; Robolectric still covers most screen interactions. Emulator CI requires KVM; startup or test failure keeps the check red, with no automatic assertion retries.
Signed release verification runs the Android test/lint suite separately with signing inputs. PR checks reject those inputs and remain unsigned. See release automation.
Use a disposable Google APIs AVD named message487-tests or message487-tests-26 /
message487-tests-35. Tests clear Message487 settings and the queue on that emulator.
Connect only that emulator and set ANDROID_SERIAL and ANDROID_HOME; the lane rejects
physical devices, other AVD names and additional connected devices. CI creates a fresh AVD
for each API and builds only debug/test APKs, without release signing secrets.
To create a local AVD, choose arm64-v8a on Apple Silicon or x86_64 on Intel/Linux:
sdkmanager "system-images;android-35;google_apis;arm64-v8a"
avdmanager create avd --name message487-tests-35 --package "system-images;android-35;google_apis;arm64-v8a" --device pixel_5
emulator -avd message487-tests-35 -no-snapshot
# In another terminal, using the actual serial from adb devices:
ANDROID_SERIAL=emulator-5554 bundle exec fastlane android device_testsRepeat with API 26. Reopening stores is tested; process death, real SMS/notification capture, WorkManager scheduling, OEM restrictions and upgrade of a signed APK remain device acceptance scenarios. These debug-emulator results do not certify a release candidate.
Android uses Gradle JUnit XML. Python and the live n8n smoke test use unittest-xml-reporting;
n8n Node tests use Node's native JUnit reporter. Install requirements-dev.txt for both Python
and server lanes. Python XML is in test-results/python/, n8n XML in test-results/server/,
and device XML in app/build/outputs/androidTest-results/connected/. CI uploads reports even
after test failures; emulator evidence includes HTML and Logcat (14-day retention).
test-reports.yml publishes each suite as GitHub Checks and a collapsible Actions summary
as soon as its job finishes. Internal CI/release jobs call it directly; fork PR reports use
a trusted workflow_run after merge into the default branch. The reporter parses artifacts
without checking out or executing PR code. A test/build failure without XML remains a failed
CI job, never a passing test report. Signed-release JVM results are published separately.
For signed candidates, F-Droid and device checks, follow Release testing.
Update regression tests cover the entry dialog, weekly snooze/skip, F-Droid routing, GitHub navigation without automatic download, persisted background results, transient-error retry policy, cancellation and stale-source handling. HTTP 429/5xx are retryable; 404 and invalid responses are not. These JVM/Compose checks do not establish real JobScheduler timing.