Verify UI through dev control

Choose the session

• Automated regression: use a disposable fixture profile; it makes no provider/account calls.
• Real app manual testing, including “no mocks”: use start --profile manual-review --providers live. That request authorizes the live launch; do not ask again. The window is visible and focusable. Isolated app data does not isolate provider accounts, Keychain, quotas or compute. Do not copy or inspect credentials, and do not perform paid generations merely to check startup. The launch drops your agent session's own variables (envStripped in the start JSON) and warns when the games root sits under a .claude folder (warnings).
• A new user's first launch (no Claude Code, Codex or sign-in): add --fresh-machine to that live start; shell --profile <name> is that account's terminal.
• Human review explicitly using fixture data: set STUDIO_FIXTURE_INTERACTIVE=1 on start/restart.
• Fixtures run in English. STUDIO_FIXTURE_LANGUAGE=ru (any app language) shows one; STUDIO_FIXTURE_LANGUAGE=system STUDIO_FIXTURE_SYSTEM_LANGUAGES=ja-JP,en-US lets the app detect it as on a Mac with that language list (macOS; the computer's settings are untouched).

Inspect ownership/status before reusing a profile. Preserve active user work. Leave human testing sessions running and retain their state. The stop/clean steps below apply only to disposable sessions created for your automated check. See the field guide. Use owned profiles instead of npm start or the normal profile/~/AI Games.

CLI: scripts/studio-dev.ts (parser in scripts/studio-dev/args.ts). Operations and their limits: src/main/dev/protocol.ts. Every command prints JSON (after npm's two header lines).

Automated fixture loop

npm run studio:dev -- fixtures                                   # the named fixtures
npm run studio:dev -- start --profile ui-check --fixture app-basics   # builds, waits for ready
npm run studio:dev -- snapshot --profile ui-check                # controls, state, text
npm run studio:dev -- snapshot --profile ui-check --scope '[data-stage-view]' --limit 40
npm run studio:dev -- ui --profile ui-check --json '{"method":"click","params":{"selector":"[data-stage-action=\"assets\"]"}}'
npm run studio:dev -- capture --profile ui-check --surface desktop --name after
npm run studio:dev -- logs --profile ui-check --surface desktop  # also game|core|harness|stdout|stderr
npm run studio:dev -- stop --profile ui-check
npm run studio:dev -- clean --profile ui-check                   # disposable state only

Pick the fixture that already shows the surface (app-basics, chat-history, run-controls, sidebar, build-history, game-surface, asset-library, ...). Changing source makes the session stale: restart --profile ui-check rebuilds with the same fixture. start --reuse reopens a stopped one.

Requests

Target elements by data-* hooks or aria-label, never by class names or visible copy. A selector must match exactly one visible, enabled, unobscured element (scope narrows it). Longer requests go through stdin, which avoids shell quoting:

npm run studio:dev -- ui --profile ui-check --request - <<'JSON'
{"method":"type","params":{"selector":"[data-promptbar] textarea[aria-label=\"Prompt\"]","text":"Make it rain","replace":true}}
JSON

Templates (one per request; --request FILE also works):

{"method":"click","params":{"selector":"button[aria-label=\"Hide sidebar\"]"}}
{"method":"click","params":{"selector":"[data-asset-card=\"assets/a.glb\"] button","modifiers":["Meta"]}}
{"method":"key","params":{"surface":"desktop","key":"Enter","code":"Enter"}}
{"method":"select","params":{"selector":"select[aria-label=\"Build history\"]","value":"<runId from snapshot>"}}
{"method":"scroll","params":{"surface":"desktop","deltaX":0,"deltaY":600,"selector":"[data-chat-scroll]"}}
{"method":"file.drag","params":{"selector":"[data-promptbar] textarea[aria-label=\"Prompt\"]","stage":"drop","file":"picture"}}
{"method":"game.input","params":{"actions":[{"type":"tap","keys":["ArrowRight"]},{"type":"wait","ms":300}]}}
{"method":"game.state","params":{}}

A fixture's Genex account signs in with STUDIO_FIXTURE_GENEX_ACCOUNT=1 (the full account; new, quiet, fresh are others); asset-library signs in by itself.

Hooks for the controls an unattended operator presses: [data-run-resume="<runId>"] (Resume on the result card and on the chat's paused line), [data-onboarding-action="next|skip|start"] (the welcome). key sends Arrow keys (all four), PageUp/PageDown, Home/End with real key codes, so sliders and segment groups move as they do under a keyboard. runs --profile P lists each open run and the open chat's newest (state, phase, clock, last record, stop reason, resumable); it reads the profile's records, asks the window which chat is open, and answers on a stale build or with the harness down.

Evidence

• Assert on the snapshot JSON (state, stage, controls[].label, text), not only images.
• capture returns the PNG path (under .studio-dev/evidence/); open it and look at it.
• Report source revision, build ID, profile and provider mode from start/status; include the fixture name when applicable. Say whether the session remains running for manual testing.
• Errors come back as {ok:false,error}: target-not-visible, ambiguous-selector, stale-build (restart), unsupported-in-fixture (native action; not a UI bug).
• For your disposable automated fixture only, stop then clean, including after failure; evidence is kept. Do not clean retained live profiles or stop a human testing session.