Skip to content

Testing and releases

Local verification:

./gradlew checkAll
./gradlew checkContractTargets
./gradlew serverSmoke
./gradlew clientSmoke
./gradlew sameLoaderSmoke
./gradlew releaseChecksums cyclonedxBom
./gradlew checkApiCompatibility -Penderfall.previousApi=/path/to/previous-api.jar

test-mod is both a normal API compile fixture and a standalone nine-target consumer build. Run ./gradlew checkContractTargets to publish the current SDK into an isolated workspace repository and build that full-feature mod unchanged for every target. Unit suites cover identifiers/builders, registration freezing, event priority and fault isolation, strict codecs and manifests, TOML repair/backups, and deterministic data paths. Gradle TestKit creates a fresh consumer, builds two targets, and inspects both JAR metadata formats.

serverSmoke launches every generated dedicated server sequentially, waits for the portable SERVER_STARTED lifecycle marker, sends stop, and requires a clean zero-code exit. It creates EULA and offline/ephemeral-port settings only inside ignored test run directories. Per-target logs and a JSON summary are written under build/reports/runtime-smoke. Use -Penderfall.smokeTarget=1.21.4-fabric to exercise one catalog target.

clientSmoke launches every generated client sequentially, waits for portable lifecycle and client-tick assertions after the initial resource reload, requests a native graceful shutdown, and requires a clean zero-code exit. It uses --refresh-dependencies for the nested consumer build because workspace publications deliberately reuse the coordinated development version. Per-target logs and client.json are written under the same report directory. Use -Penderfall.smokeTarget=<target> to exercise one catalog target.

generatedBridgeServerSmoke and generatedBridgeClientSmoke exercise only targets implemented by the central bridge compiler. They resolve SDK artifacts from build/generated-bridge-repository, where handwritten runtime publications are absent, and use build/generated-bridge-gradle-user-home to isolate Loom caches from IDE imports. Preparation and game-launch timeouts are separate. Their reports are written under build/reports/generated-bridge-runtime-smoke. When requested together, generated server/client smoke lanes run in that order to avoid concurrent native preparation and launches for the same consumer target. Modern NeoForge and legacy Forge-family targets use the root prepareClient or prepareServer proxy, which materializes only the selected consumer target and finishes native assets and run artifacts before the timed game phase. All nine catalog targets, including 1.20.1 Forge and NeoForge, have passed both generated-only process-smoke lanes. These are lifecycle/resource checks, not interactive feature or connection tests.

uiAcceptanceMatrix is the named automated UI lane. It uses the generated-runtime client smoke on all selected targets and only emits its completion marker after the portable screen has successfully invoked fills, aligned text, clipping, sprites, tiling, nine-slicing, opacity/depth stacks, pose transforms, item/tooltips and a living-entity preview. It also requires initial and requested resource reloads. Use -Penderfall.smokeTarget=1.20.1-forge,26.2-neoforge for a focused legacy/extraction pair. The complete nine-target matrix passed locally on Windows on 2026-09-28 using generated runtimes only. The machine-readable evidence is written to build/reports/generated-bridge-runtime-smoke/client.json on each run; the focused pair is still useful for a faster oldest/newest development check.

Automated completion proves that the native callbacks executed without an exception; it does not prove pixel-perfect composition or human input feel. Release-candidate UI acceptance must still manually open a standard screen and authored inventory screen on the oldest and newest rendering ABIs and verify pointer input, keyboard focus/typing, scrolling, clipping, slot interaction, entity rendering, resource reload and GUI scales 2/3/4.

For connected generated-runtime menu actions and workbench slot crafting, run generatedBridgeGameplaySmoke; see the scenario and evidence gates. verifyGameplayMatrix -Penderfall.gameplayReports=<oldest.json>,<newer.json>,... audits explicitly ordered reports and their actual logs for every catalog target. It audits historical focused-scenario evidence, not release readiness or the latest edited binaries; newer failures supersede older passes.

The release workflow uses a stricter, fresh-evidence path. Each target job runs the generated-only dedicated-server, client/UI, and connected gameplay lanes, stamps its artifact with the workflow Git SHA, and uploads its reports plus logs. The downstream verifyRuntimeMatrix task accepts exactly nine artifacts, rereads all checkpoints, requires the same portable-source hash, and rejects stale revisions, missing lanes, unknown assertion keys, false checks, failure text, dirty exit codes, and failure markers. This task intentionally has no permissive local default.

sameLoaderSmoke launches a dedicated server and a separate client for the same exact Minecraft/loader target, auto-connects over loopback, requires a three-leg SDK packet round trip, opens and renders the synchronized contract screen, and then shuts both processes down cleanly. It never pairs different loaders. Per-target logs and connection.json are written under the runtime-smoke report directory. The full nine-target matrix passed locally on Windows on 2026-09-04; CI runs the same task as one isolated Linux/Xvfb job per target.

The release gate additionally requires all nine target contract suites, Linux/Xvfb client and dedicated-server smoke tests, every same-loader external connection test, Windows/Linux fresh-template builds, reproducibility comparison, signature/checksum/SBOM inspection, and a manual launcher/dedicated-server release-candidate pass. The exact manual steps are recorded in the release-candidate checklist.

A green Java/Gradle build is not evidence that Minecraft started or that a real client exchanged packets with a dedicated server. The process-level server report, client report, and same-loader connection report are recorded separately. Coordinated publication is blocked if any supported target fails.