By name: 1k-ui-verify description: AI-agent-driven UI verification for OneKey. Use to actually drive the running app and confirm a visual/interactive change works — Electron desktop via Chrome DevTools Protocol (CDP) on port 9222 with playwright-core, and React Native (iOS/Android) via callstack agent-device. Triggers on "verify the UI", "drive the app", "screenshot the change", "check it on desktop/simulator", "CDP 9222", "agent-device", "UI 验证", "跑一下看看", "截图确认".
UI Verification (Desktop CDP + RN agent-device)
Drive the running OneKey app to confirm a UI change instead of guessing from code. Pick the backend by the platform under test — never assume one tool covers all.
Decision: which backend?
| Platform under test | Backend | Why |
|---|---|---|
| Desktop (Electron) | CDP 9222 + playwright-core | Renderer is Chromium; already enabled in dev. See electron-cdp.md |
| iOS / Android (native RN) | agent-device | Real simulator/emulator/device, RN component tree. See agent-device-rn.md |
| Web / Extension | Out of scope here | Use plain Playwright/Puppeteer against the dev URL |
playwright-core (^1.50.0) is already a root dependency — no install needed for desktop.
agent-device is an external CLI/MCP — see its rule file for setup.
Workflow
- [ ] 1. Confirm platform + expected vs actual (ask the user if a visual bug — per CLAUDE.md)
- [ ] 2. Launch the app for that platform (desktop: yarn app:desktop; RN: build/run the dev app)
- [ ] 3. Connect the backend (CDP connectOverCDP / agent-device open)
- [ ] 4. Navigate to the target screen, capture a BEFORE screenshot if comparing
- [ ] 5. Reproduce the interaction, capture AFTER screenshot + console/logs
- [ ] 6. Report with the screenshot as evidence — do not claim "fixed" without it
Always anchor to testID props, never brittle text/CSS selectors. Add a testID to the
component first if one is missing (search grep -rn 'testID' packages/kit/src/views/<area>).
Desktop quick recipe (CDP 9222)
Prereq: yarn app:desktop running — its dev:main launches Electron with
--remote-debugging-port=9222 --remote-debugging-address=127.0.0.1.
node .claude/skills/1k-ui-verify/scripts/cdp-shot.mjs --out .tmp/ui/after.png
node .claude/skills/1k-ui-verify/scripts/cdp-shot.mjs --testid header-gift-action --click --out .tmp/ui/clicked.png
The script connects to the main window (the page with [data-testid^="tab-modal"]), captures a
screenshot, optionally clicks a testID, and prints any console errors. Full pattern, console
capture, and pitfalls (DevTools closed, finding the right window): electron-cdp.md.
RN quick recipe (agent-device)
agent-device open --platform ios # boot/attach simulator + app
agent-device metro reload # after editing src/node_modules; wait ~10-13s
agent-device click 'id="home-more-button"' # click by testID — works on iOS AND Android
agent-device fill 'id="amount-input"' "0.5" # type by testID
agent-device screenshot --out .tmp/ui/rn.png # read the PNG — it's the source of truth
testID works on iOS (click/is visible/get attrs with id="...") even though snapshot looks
empty. Gotcha: the dev CPU/RAM/fps perf-overlay is a full-screen window that swallows taps over
its HUD box — if a click prints "Tapped" but nothing changes, disable the overlay or get attrs
click <x y>on a clear spot. Command isclick, nottap. Full details, the proven edit→reload→verify workflow, and pitfalls: agent-device-rn.md.
Prefer the MCP integration so the agent calls these directly.
Regression scenario series
Reproducible freeze/regression scenarios live in one root runner: scenarios/regression.mjs.
Each scenario declares a backend (cdp for desktop/web, agent-device for RN) and the cdp
backend is parameterized by port, so web reuses it (no slow per-run Playwright launch — connect
over CDP to an already-running Chrome).
node scenarios/regression.mjs list
node scenarios/regression.mjs gift-storm-desktop # CDP 9222 (yarn app:desktop)
node scenarios/regression.mjs gift-storm-web # CDP 9223 (Chrome --remote-debugging-port=9223 on app:web)
node scenarios/regression.mjs gift-storm-rn --platform ios # agent-device
REGRESSION=1 node scenarios/regression.mjs gift-storm-desktop # exit 1 if reproduced (CI gate)
The gift-storm-rn scenario's selectors (tab labels, testIDs) are mirrored from desktop and
marked VERIFY — confirm against agent-device snapshot --json on a live device before relying on it.
Related Skills
/verify— generic "run the app and confirm a change" harness (this skill is the OneKey-specific how)/run— launch/screenshot the app/1k-dev-commands—yarn app:desktop/yarn app:ios/yarn app:android/1k-cross-platform— platform-specific behavior to verify separately