Frequently asked questions

Yes. Evaluation requires one runnable, representative package; interface names alone do not establish runtime behavior. We can sign an NDA first, or start with a stripped package that retains logic while removing production art.

Three games (PixiJS, Phaser, and native Canvas) each run 3 times on a Mate 30 Pro; see the performance page. All 80 shared conformance assertions pass across the Linux reference implementation, Android emulator, and physical device. Raw data is in migo-bench.

• Entry-level device comparisonIn progressCurrent measurements use a Mate 30 Pro. The same harness will run on an entry-level Redmi A-series device.
• Synchronized device videoIn progressOne uninterrupted recording of the same game on Migo and WebView, overlaid with PSS and CPU.
• Cross-OEM conformanceIn progressAdd one physical-device result for Xiaomi, OPPO, and vivo.
• Interface referenceIn progressPublish a searchable reference entry for each of the 369 interfaces.
• Additional benchmark gamesPlannedExtend the same-device harness beyond the current PixiJS, Phaser, and native Canvas workloads.

Yes. This is the primary path. With migo-wx-adapter, existing wx.* calls remain unchanged. Cocos, Egret, and Laya mini-game exports load directly; Unity uses its WebGL export converted to a mini-game target. Browser targets use migo-web-adapter. Calls outside the published surface return an error instead of failing silently.

With migo-web-adapter, browser names map to runtime functions as follows:

Interfaces are exposed under migo.*. With wx-adapter, wx.* references the same functions. The following twelve groups do not overlap:

Another 197 cover canvas and images, audio/video capture and streaming, device and system information, files, Bluetooth and Beacon, location, logging and performance reporting, Worker, mini-program navigation, and lifecycle.

11 are stubs (dashed outline): the names are published and calls do not throw, but they perform no operation. The list comes from data/surface.txt and is verified at build time.

Yes. Since v0.9.4, nojni.aar provides a 0.40 MB bootstrap package without the engine. The host downloads the engine on first launch through Play Feature Delivery or its own channel, verifies SHA-256, and caches it locally. The engine and native binaries are identical to the embedded AAR build. Failed downloads can be retried without restarting.

The Bunnymark sweep scales from 40,000 to 220,000 sprites and averages two runs at each level. Both runtimes reach the knee at 40,000; the difference appears after that point:

Both runtimes fall below 55 fps at the same level. Migo is 1 fps faster at 70k, 100k, and 180k. At 220k—well beyond the measured production workloads—both reach 13 fps.

The same conformance assertions run against the Linux reference implementation, Android emulator, and Mate 30 Pro. On 2026-08-30, all 80 shared assertions passed against v0.9.6; platform-only assertions are always reported when skipped. The first run found three real defects: an empty getImageData buffer, two sizes accepted in one ctx.font assignment, and an empty readback after the first clear of a new WebGL context. The suite is in migo-conformance.

• AndroidReleasedarm64-v8a / x86_64. AAR with Java/Kotlin SDK; minSdk 26.
• LinuxReleasedx86_64 / aarch64. Static and shared libraries, pkg-config, CMake, and Qt 6 / X11 host references.
• WindowsReleasedx86_64 / aarch64. migo.dll, import library, and CMake.
• HarmonyOS NEXTReleasedaarch64 / x86_64. Third-party VMs cannot use JIT; see the next answer.
• iOSReleasedaarch64 device and simulator framework with C ABI.
• macOSReleasedaarch64 / x86_64.
• C ABI clientCandidateConnects a custom render loop directly through C without the Java layer. The interface is not frozen.

Artifacts are published in Releases, each with an .attestation.json.

From 5.0.0(12), third-party VMs cannot use JIT, so V8 runs in interpreter mode. Android simulation with --jitless shows that content containing .wasm does not run. CPU-bound pure JavaScript reaches 1/2.4 to 1/16 of JIT frame rate, and the knee moves from 40,000 sprites to 2,000. Lightweight pure JavaScript on a high-end device is unaffected and uses 5–9% less memory. Data is in migo-bench JITLESS.md.

The runtime version is fixed in each app build. Rolling back to the preceding build also restores the preceding runtime. A system WebView version is outside host control and cannot be rolled back this way.

The host controls all three. The runtime bundles no ad network, holds no identity or merchant credentials, and only issues requests through explicit callbacks. Missing capabilities fail as documented, and the runtime never declares a payment successful. The host callback reference gives the fallback behavior for all 9 groups.

The source is available under BSL 1.1 and may be modified within its terms. Before production use, read the license text. Commercial licensing is handled through licensing@minigame-labs.com; security reports follow SECURITY.md.