SKILL: smoke-test — "done" includes that it STARTS (automated smoke)

Principle (P-10 · any stack/domain). "Done" is not just that it compiles and unit tests pass: it includes an artifact that STARTS, proven by an automated smoke with a clear pass/fail. The studio AUTHORS the smoke; the PO/CI runs it with ONE command. Applies to any domain and any stack (desktop, web, CLI, service): the smoke confirms the binary/process boots and shows/exposes the essentials.

Why it exists: headless QA and unit tests do NOT catch startup/integration failures of the packaged artifact (init order, production paths, bundle load, handler registration). An automated smoke does — and it runs with no human watching.

What it tests (and what it does NOT)

• YES: the artifact STARTS + renders/exposes the ESSENTIALS (window/root + 2-3 key points).
• NO: business logic — that's unit tests. Nor exhaustive coverage. The smoke is about STARTUP.
• P-11: whatever can't be validated without a real environment (hardware, live data, network) is DECLARED deferred, not faked.

Contract (invariant, any stack)

• ONE command (smoke script) → exit 0 = pass / exit 1 = fail. CI-ready.
• Machine-readable evidence: on finish it writes tests/smoke-result.json = { "pass": true|false, "ts": <epoch>, "screenshot": "tests/smoke.png" }. The RELEASE gate verifies it mechanically (DoD): without a recent smoke-result.json with pass:true, release does NOT close.
• Also leave the startup screenshot (PNG).
• Headless / no display (no human intervention required).

Current recipe — Electron (the current one, NOT the only one)

@playwright/test with _electron.launch:

1. const app = await electron.launch({ args: ['.'] }) → starts the Electron project.
2. const win = await app.firstWindow() → the main window opens (if not, FAIL).
3. Assert 2-3 key data-testid elements visible (root + main views; placed by 06 FRONTEND).
4. await win.screenshot({ path: 'tests/smoke.png' }) → visual evidence.
5. Write tests/smoke-result.json { pass, ts: Date.now(), screenshot: 'tests/smoke.png' } (in finally, so a failure also leaves pass:false). await app.close(). Pass/fail via the runner's exit code.

• Single script: "smoke": "playwright test tests/smoke.spec.ts". DevDep: @playwright/test.
• CI without display: on Linux wrap with xvfb-run; on Windows it starts directly.

Other stacks (same principle, different recipe)

• Web: Playwright/Chromium headless → the page loads + key selectors + screenshot + smoke-result.json.
• CLI: run the binary (--version/--help or a no-op) → exit 0 + expected output + smoke-result.json.
• Service/API: start + health-check (e.g. /health 200) → exit 0 + smoke-result.json. The studio picks the recipe by stack; the contract (one command · pass/fail · smoke-result.json · evidence · startup-not-logic) is identical.

Who

• 08 QA AUTHORS the smoke and leaves it running pass/fail; attaches the screenshot to QA_REPORT.md and leaves tests/smoke-result.json for the release-gate DoD.
• 06 FRONTEND places the data-testid on the key elements (root + main views) the smoke needs.
• 04/07 wire the smoke script (ONE command) into the project's scripts/build.