Excuse My French - Android Compose App

CI details — unit/screenshot/R8 jobs + emulator matrix, API 34 → 37.1, plus Android CLI and Emulator Preview legs

Besides unit tests, Compose screenshot validation, and a minified R8 release build, instrumented tests run on GitHub-hosted emulators:

Legs	Image	Emulator channel	GPU	Gating
API 34, 36	default (AOSP) x86_64	stable	swiftshader	✅ blocking
API 37.0	google_apis_ps16k (16 KB page size)	stable	lavapipe	non-blocking
API 37.0	google_apis_ps16k	canary (--channel=3)	lavapipe, auto	non-blocking
API 37.1	google_apis_ps16k	canary	lavapipe, auto	non-blocking
Android CLI experiment	google_apis_ps16k 37.0	canary	emulator default	non-blocking
Emulator Preview (emulators;latest)	google_apis_ps16k 37.0	preview package	auto	non-blocking
Emulator Preview multi-run (snapshot cycles)	google_apis_ps16k 37.0	preview package	auto	non-blocking
Android CLI multi-run (snapshot cycles)	google_apis_ps16k 37.0	canary	emulator default	non-blocking

The Android CLI leg drives the whole flow with the android CLI (android sdk install --canary, android emulator create/start/stop) instead of sdkmanager/avdmanager and the emulator-runner action.

All emulator-runner legs use the pixel_6 profile, full diagnostics (-verbose -show-kernel -debug-metrics -metrics-collection), and a cmdline-tools;latest update so avdmanager writes a valid target=android-37.x; the AVD cache is keyed on the cmdline-tools version so stale AVDs with a broken target=android-0 root ini can never be restored.

This is a simple Android application built with Jetpack Compose that displays French insults and accompanying images. It fetches data from a public API.

Overview

The app consists of a single screen that displays:

• A randomly selected French insult (text).
• An associated image.
• A divider between the text and image.
• A loading indicator while data is being fetched.
• An error message if data fetching fails.
• A Text-To-Speech (TTS) feature that reads the insult aloud in French. TTS problems are shown as a small note without hiding the insult.
• A mute/unmute toggle button.
• Pause/resume and next buttons to control the automatic insult refresh.
• A level selector (1/2/3) to choose the maximum insult level: 1 = family-friendly, 2 = adds vulgar, 3 = adds offensive. Levels are cumulative, the choice is persisted across launches, and the app defaults to level 1 on first launch.
• An AppFunction (getFrenchInsult) that lets on-device AI agents and the Android system fetch an insult without opening the app UI (Android 16+).
• An adaptive launcher icon (beret-wearing robot on the French flag) with a monochrome layer for Android 13+ themed home screens.

The layout adapts to the window size. In windows narrower than 600dp (phones in portrait), the text is displayed above the image with a horizontal divider, and the text area takes up at least 15% of the window height. In wider windows (tablets, unfolded foldables, split-screen, and desktop windows), the text and image are shown side by side with a vertical divider. In both layouts the text is centered and the image keeps its aspect ratio while fitting within 90% of the available space.

Technologies Used

• Language: Kotlin
• UI Framework: Jetpack Compose
• Networking: OkHttp
• JSON Parsing: kotlinx.serialization
• Asynchronous Programming: Kotlin Coroutines (with StateFlow)
• Dependency Injection: Manual dependency injection (constructor injection for Repository and TtsService)
• ViewModel: AndroidViewModel (for access to application context)
• Text-To-Speech: Android TextToSpeech API abstracted behind a TtsService interface.
• Preferences: Jetpack DataStore (persists the mute state and insult level).
• Agent Integration: Jetpack AppFunctions (androidx.appfunctions, processed with KSP).
• Testing:
  • Unit Tests: JUnit 4, MockK, Robolectric, kotlinx-coroutines-test
  • UI Tests: androidx.compose.ui:ui-test-junit4
  • Screenshot Tests: Compose Preview Screenshot Testing (reference images for phone, foldable, tablet, and desktop; run ./gradlew validateDebugScreenshotTest to check, updateDebugScreenshotTest to re-record)
• Image Loading: painterResource (for placeholder), and manual Base64 decoding and Bitmap conversion for fetched images.

Continuous Integration

Every push and pull request to main runs GitHub Actions jobs: unit tests, screenshot validation against the committed reference images, a minified release build (guarding the R8 configuration), and instrumented tests on an emulator matrix — API 34 and 36 (blocking) plus the API 37.0 and 37.1 16 KB page-size preview images (non-blocking). Dependabot keeps GitHub Actions and Gradle dependencies up to date with weekly PRs.

Four further jobs, all non-blocking, track newer Android emulator tooling. Two of them run a snapshot multi-run experiment: the emulator is booted four times against the same AVD with quickboot snapshots enabled, the app is launched only on the first cycle, and every later cycle checks whether the snapshot brought it back by itself — still running, and still rendering. The app is deliberately never relaunched after a restore, since that is the thing being measured. One job drives the Emulator Preview package, the other the canary emulator through the android CLI, so the same experiment can be compared across both.

Because this app's UI is static, rendering is judged with android layout rather than by diffing two screenshots — a screenshot diff would report a stall on every cycle. A non-empty layout tree plus a focused window means the UI is present and enumerable. Screenshots are still captured and uploaded for every cycle.

The preview jobs share their setup through the composite action in .github/actions/preview-emulator. scripts/replay-preview-multirun.sh replays the multi-run job locally in a few minutes instead of a push cycle.

AppFunctions

On Android 16+, the app registers the getFrenchInsult AppFunction so on-device agents (assistants, voice commands, system shortcuts) can fetch an insult without opening the UI. It takes an optional maxLevel parameter (1–3, same semantics as the in-app level selector; omitted or 0 means level 1) and returns the insult text and its level. Test it from a shell:

adb shell cmd app_function execute-app-function \
  --package io.github.jpcottin.excusemyfrench \
  --function "io.github.jpcottin.excusemyfrench.appfunctions.InsultFunctions#getFrenchInsult" \
  --parameters "'{\"maxLevel\": 2}'" --brief-yaml

API

The application fetches data from the following public API endpoint:

• Insult and Image Data: https://excusemyfrench.herokuapp.com/api/v1/img

The optional level query parameter (?level=1|2|3, default 3) caps the insult level; levels are cumulative, so level=2 serves levels 1 and 2. The app always sends the user's selected level.

More information about the API, including its source code, can be found on GitHub: https://github.com/jpcottin/excusemyfrench/blob/master/README.md. The API returns a JSON response in the following format:

{
  "insult": {
    "text": "French insult text here",
    "index": 123,
    "level": 2
  },
  "image": {
    "data": "/9j/4AAQSkZJRgABAQAAAQABAAD/...",
    "mimetype": "image/jpeg",
    "indexImg": 456
  }
}