Most checks run in Node: unit tests of the shared core and of the app’s screens, type checks and lint, on every pull request. What the television shows is checked on the Vega Virtual Device, by hand or from scripts, and since October 6, 2026, on a Fire TV Stick 4K Select as well (see the end of this page).
Three packages have tests, all on Node’s own test runner (node --test) with the Node version in mise.toml. Counts and approximate times are from the October 22, 2026, local check on a Mac, for the application revision released in 1.0.5, build 59 (6827541, following #490):
Core Suite test/
337 testsabout 13 s
The turn controller and its legal actions, dice, saved-game snapshots, the completed-game ledger, the rules guide against the engine, the opponents’ engine algorithms, the tutorial’s lessons, the danger search, cursor jumps and press counts, voice lines, and the English-only check.
Runner & environment:node --test --experimental-strip-types, with node:assert/strict
App Suite native/test/
711 testsabout 24 s
Board and piece rendering, remote input and focus, menus, the tutorial, saving and the ledger across relaunches, serialized opponent requests and recovery, sound, music and voices (including ducking and player lifecycle), the reduced-motion path against a stubbed platform setting, and generated assets including the launcher icon.
Runner & environment:node --test with esbuild module hooks (native/test/hooks.mjs); react-test-renderer over hand-written stubs of react-native and the @amazon-devices packages.
The tracked portrait pack is checked in CI: its version, every pinned image and the original notice. These checks are required.
Web Bench web/test/
13 testsunder 1 s
The browser test bench's configuration and board marks, plus Web Worker completion and termination contracts.
Turns follow the engine: tests check that the turn controller accepts only actions the rules engine lists. A turn must use as many dice as the position allows, promotion offers only the engine’s choices, a king capture ends the game at once, and a save or bot reply that skips or invents an action is refused.
Saved games: a snapshot is validated before it is written, and tests check that a damaged one is refused and never replaces a good save (against an in-memory stand-in for MMKV).
Completed games counted once: tests recreate what a crash between a game’s end and the ledger write would leave, then relaunch, and check that each finished game is counted exactly once.
Rules guide: 18 of the in-game rules guide’s 33 sentences, in 8 of its 9 topics, are each tied to a test that checks the engine does what the sentence says (counted on October 4, 2026). The en passant topic has none. When the guide was written (September 23, 2026), three false sentences were put in by hand, and each failed its test.
Launcher icon:native/test/splash.test.ts reads the icon’s pixels and fails if any artwork lies outside the band the launcher keeps when it crops the icon into a 3:2 tile. The band was measured on the Vega Virtual Device (FL-16).
The audio work in PR #388 passed mise run check with 325 core, 586 native and 8 browser tests: 919 in total, including 33 added native regressions. They cover readiness and delayed playback, completion and cleanup, stale events, bounded failure and retry, silence during crossfade errors, and verified MP3 tag removal. These counts describe that PR’s local checkout on October 14; the suite cards above give the separately dated October 21 snapshot.
Release packages were built for armv7, aarch64 and x86_64. Separate VVD and physical Stick runs covered ten minutes of playback, all six tutorial lessons and all 45 tutorial clips, settings and Home/relaunch, plus eight injected crossfade failure/silence cases per device. Native events and release completions support the hardware result; recorded audio is from VVD. Clean packages were installed afterward, deployed JS/Hermes hashes matched their archives, and saved data and CRC matched their pre-test backups. The validation record separates the device, automated and package evidence, including cold Home recovery and the remaining warm-focus and physical-listening limits.
Opponent and audio lifecycle checks, October 21, 2026
PR #480 added regression checks for serialized searches, foreground-only transport deadlines, late-result discard and reuse after a delayed response settles. Its final ordinary build 52 completed eight natural games on the physical Stick, two per bot and two Hot Seat games, and checked Menu/resume, cold search recovery, a 180-second same-process Home interval, settings, all tutorial lessons, Rules and About. Both Recovery choices were also exercised in a separate package with a delayed response callback. The opponent runtime record distinguishes ordinary Release, diagnostic, Virtual Device and Node evidence.
PR #479 added checks for construction after the first commit, startup speech, suspended startup, effect-player release and recreation, and music’s idle timer. Physical build 53 completed five natural games and checked first/cold launches, audio settings, same-process Home returns and a music loop seam. Build 54’s review follow-up checked one first launch, three cold launches and three Home returns; it did not repeat the full games. The audio lifecycle record identifies the tested commits and the incomplete continuous logging during the later build 53 games.
These local Release builds used a separate QA package to preserve the installed store build 30. Installed payloads were checked against archives; QA saves and CRC were restored byte-for-byte afterward. Actual device input drove the games, but both Hot Seat sides were automated. The runs do not establish human TV listening, a two-human session, measured input latency or an Amazon-signed in-place upgrade.
The sonar CI job first runs the core and app coverage tests; a failing test fails that job. Only its subsequent SonarCloud scan step has continue-on-error, so a failed scan does not fail the workflow. The workflow does not wait for the quality gate; SonarCloud reports its own status separately. The project uses SonarCloud’s default quality gate, Sonar way, and passed it on October 4, 2026. That day SonarCloud reported:
0 bugs
0 vulnerabilities
0 security hotspots
98.5% coverage. It comes from the Node tests, which run the app’s code against stubs of react-native and the Vega packages, so it shows which lines the tests reach, not that the code behaves the same on a device.
CodeRabbit reviews only when asked to, with an @coderabbitai review comment on a pull request that changes code; automatic reviews are off (.coderabbit.yaml). Its instructions for the app’s code list Vega behavior measured on the Virtual Device and ask it to flag code that contradicts it, and comments that claim behavior without evidence. Its reviews also run actionlint, zizmor, gitleaks, TruffleHog and ShellCheck.
To catch issues before they enter git history, lefthook runs these jobs on each commit (lefthook.yml; mise run setup installs them):
English only (test/english.test.ts): reads every tracked file and every new file Git does not ignore, skipping binary files, and fails on any match of the Unicode \p{Script=Cyrillic} regex. No Cyrillic character is allowed. It catches text pasted from the project’s private notes, which are not in English; it cannot detect private content written in English. CI runs the same test with the core suite.
Secret scanning: betterleaks scans the staged diff with its default rules for secrets such as API keys, tokens and private keys; nothing leaves the machine. On GitHub, secret scanning and push protection are also on for the repository.
Formatting and workflows: Prettier formats the staged files and adds the result to the commit, and actionlint checks staged workflow files. Before a push, npm run format:check checks the whole repository, as CI does.
The way the cursor moves was chosen by counting presses.
Moving a cursor across the board one square at a time takes many presses on a TV remote. A script, scripts/cursor-presses.ts (npm run presses), plays 200 seeded random games through the app’s shared core, with random legal actions for both sides, and counts the arrow and OK presses each way of moving the cursor would need. Random moves are not a person’s moves, but every way is scored on the same ones. #68 recorded the first run; the figures below are from the run of October 4, 2026, seed 68, 8,346 turns:
Navigation Model
Presses per Hot Seat Turn
Change
Square by square
23.0 presses
Baseline
Jumps between pieces and destinations
9.2 presses
60% fewer presses
Random movers take no more often than they make any other move, so this count cannot credit where the cursor lands once a piece is picked up. Landing on captures costs them 0.29 presses a turn: 9.21 against 8.92 with the cursor always landing on the central destination. The landing follows people, who take far more often than not.
Jumps must not leave a piece or destination out of reach. Over 20,000 random sets of squares, of 3 to 8 squares each, also recorded in #68, the jump rule the board uses (the cone rule) left no square out of reach. The two other rules it was compared with, axis and nearest, each left a square out of reach in some sets. That is a sample, not a proof; a unit test (test/cursor.test.ts) pins one set where the axis rule fails and the cone rule does not. The sample does not cover larger sets or every starting square: in the 200 simulated games, a queen with fifteen destinations had one that the arrows reach only from the central destination. So the cursor lands anywhere else only when every destination stays in reach from there, and test/boardInput.test.ts checks that position.
What the television shows is checked on the Vega Virtual Device, by hand or from scripts, with vega-vvd-driver (vvd), Fortemate’s open-source driver for it. CI has no Virtual Device, so these device checks do not run there.
1.0.1, build 30 (October 14, 2026): the candidate was checked separately on the Virtual Device and a physical Fire TV Stick 4K Select, including upgrade preservation, saved-turn recovery, settings, navigation, ten-minute Rampage and friend sessions on each device, and shorter Rolly/Grabby runs. Its 919 automated tests also passed. The final-main packages have the same application payload as the tested candidate; the published GitHub downloads were verified against their SHA-256 manifest. The owner additionally reported checking the Stick and finding everything working well. The release validation record preserves the exact scope and limits. The earlier checks below remain evidence for their own dates and builds.
Scripted sessions: Amazon’s vega CLI starts the Virtual Device (also without its window, with --no-gui), installs the package and launches the app. vvd enable-grpc turns on the emulator’s gRPC endpoint, and vvd press sends remote keys as Linux evdev codes, the route the Virtual Device’s own on-screen remote uses. On September 28, 2026, a script drove beta 5 this way through Settings, the rules guide, About, a resumed game and the tutorial’s first move.
Screenshots and the safe area:vvd screenshot saves the 1920x1080 screen as a PNG, and vvd wait-change exits once the screen changes, so a script can tell that a press did something. vvd safe-area counts the pixels in the outer 5% of each edge that differ from a solid background color; in the same session on September 28, 2026, it found the margins clear on the rules guide, About and the tutorial’s first lesson. A unit test, native/test/layout.test.ts, checks the 48 dp and 27 dp insets the layout keeps.
Fire TV Stick: the owner’s Fire TV Stick 4K Select has run the game since October 6, 2026. Checks there read the device log, the OS’s traces or reports from the app itself, and the owner watches the television. They include the launch and the response to a cursor key (Performance), the dice’s stagger, horizontal cycling on the board (Board selection algorithm), the return from Home (Building on Vega) and the automatic OK on the only choice (#342). Version 1.0.0, build 26 of dac3302, was released on GitHub and submitted to the Amazon Appstore on October 10, 2026. Its release notes record what was checked on the Stick, on the Appstore build of that commit: Settings with the music row and Play as, muting and unmuting the music, Main menu and Resume game, the opponent cards, and the confirmation before a game in play is replaced. Build 28 of d4ffdaa replaced it on October 11, 2026; its Appstore build was installed on the Stick over an earlier build, kept the saved game, and opened on the home menu with no errors from the app in the device log (release notes).