droidconKeKotlin: Instructions file for Codex

AGENTS.md

Working notes for AI agents and new contributors on the droidcon Kenya Android app.

CONTRIBUTING.md covers the human process — issues, forks, PR expectations. This file covers what you need to know to change code here without breaking it.

Commands

Run these before opening a PR. CI runs the same set.

./gradlew spotlessApply ktlintFormat          # format first — the checks below are strict
./gradlew spotlessCheck ktlintCheck detekt    # static analysis
./gradlew lint                                # Android Lint + Slack's Compose rules
./gradlew stabilityCheck                      # Compose recomposition regressions
./gradlew assembleDebug                       # build
./gradlew testDebugUnitTest                   # JVM + Robolectric tests
./gradlew verifyRoborazziDebug                # screenshot goldens
./gradlew :app:bundleRelease && config/size/check-app-size.sh   # download size, needs bundletool

Startup is measured, not asserted. See docs/performance.md — it also records why neither device on hand can produce the baseline-profile comparison.

./gradlew :app:generateBaselineProfile        # regenerate the shipped ART profile

Screenshot goldens live in src/test/screenshots/, in two families: screens/ is the light / dark / 200 %-font matrix at phone size, and form_factors/ is one image per window size (phone, foldable, tablet, desktop) from ChaiScreenshotTest.captureFormFactors, which overrides the window rather than the device so currentWindowAdaptiveInfo really sees it. Every image loads as a flat grey placeholder with no size, which ContentScale.Fit draws as nothing; a test of logos overrides placeholderImage() with wideLogoImage(). Goldens recorded on macOS must pass on Linux CI. Paths drawn on a Canvas anti-alias a few colour levels apart there, so the harness compares loosely inside any node tagged cube_hero and exactly everywhere else; it finds the cube's bounds in the semantics tree at capture, so tests pass nothing. They are outside Gradle's tracked outputs, so after changing anything visual, record with --rerun-tasks — an up-to-date test task will otherwise leave stale images on disk: