# Verify Atlas > Drive the Atlas workspace UI in a running Harbor Labs desktop build via CDP. Use for screenshots, clicks, sending messages, accessibility-tree inspection, JS evaluation, or visual-state checks. - Skill: `poteto/verify-atlas` (Agent Skill, multi-file: 36 files) - Install (CLI): `npx skillmds@latest add poteto/verify-atlas` - Raw SKILL.md: https://api.skillmd.com/api/skills/poteto/verify-atlas/raw - Safety review: pending (external: skill-scanner PASS, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Web & Frontend - Author: poteto (https://skillmd.com/u/poteto) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/poteto/verify-atlas --- # Verify Atlas Drive a running Atlas desktop build over CDP: inspect the DOM, take screenshots, click, type, read the accessibility tree, evaluate JS, and manage checkout-isolated parallel instances. `node .cursor/skills/verify-atlas/control-atlas.mjs --help` lists commands, flags, and examples. This sample documents usage only; the driver scripts are not included. ## Pick a flow `test -f .git && echo checkout` distinguishes an isolated checkout (`.git` is a file) from the main workspace (`.git` is a directory). - **Main workspace:** the desktop build runs on CDP port 9222 with the shared user-data-dir. Drive it directly without `--checkout`. - **Isolated checkout:** every command takes `--checkout` so the port and user-data-dir are derived for this tree. That keeps the run from fighting the user's main session or polluting auth. From a checkout, a connect/launch command without `--checkout` (and without an explicit `--cdp`) is refused. Pass `--checkout`, or `--cdp ` if you mean another instance on purpose. Run `doctor` first when anything looks off. It reports checkout context, who owns the CDP port, watcher liveness, and build freshness. ## Checkout quick start ```bash # provision an isolated checkout, optionally seed auth, restart, then health-check node .cursor/skills/verify-atlas/control-atlas.mjs checkout-setup --watch node .cursor/skills/verify-atlas/control-atlas.mjs clone-userdata --force --overwrite # optional node .cursor/skills/verify-atlas/control-atlas.mjs --checkout restart node .cursor/skills/verify-atlas/control-atlas.mjs --checkout info node .cursor/skills/verify-atlas/control-atlas.mjs doctor node .cursor/skills/verify-atlas/control-atlas.mjs cleanup --remove-checkout ``` ## Main-workspace quick start ```bash node .cursor/skills/verify-atlas/control-atlas.mjs info node .cursor/skills/verify-atlas/control-atlas.mjs doctor ``` The desktop build must already be running with remote debugging enabled (default port 9222 in the Harbor Labs desktop launcher). ## Command surface `--help` is canonical. Categories: - **Inspection:** `info`, `snapshot`, `screenshot`, `components` - **Navigation:** `home`, `new-session`, `select-project`, `select-runtime`, `scroll` - **Interaction:** `send`, `click`, `click-xy`, `aria-click`, `type`, `press`, `eval`, `upload-image`, `add-context`, `feature-flag` - **Performance:** `trace`, `profile`, `record`, `perf-metrics`, `wait-settle` - **Streaming:** `console`, `network-log`, `network-summary` - **Health & cleanup:** `doctor`, `cleanup`, `watch --restart` Examples: ```bash # health node .cursor/skills/verify-atlas/control-atlas.mjs doctor # open a blank thread and send node .cursor/skills/verify-atlas/control-atlas.mjs new-session node .cursor/skills/verify-atlas/control-atlas.mjs send "list open tasks in this project" # keyboard path node .cursor/skills/verify-atlas/control-atlas.mjs press "Meta+KeyN" # accessibility snapshot of the live UI node .cursor/skills/verify-atlas/control-atlas.mjs snapshot # screenshot for evidence node .cursor/skills/verify-atlas/control-atlas.mjs screenshot /tmp/atlas-proof.png # wait for streaming / layout to settle node .cursor/skills/verify-atlas/control-atlas.mjs wait-settle # flip a gated path for the session node .cursor/skills/verify-atlas/control-atlas.mjs feature-flag rooms_v2 on ``` ## Proof bar Do not submit "look, it opens" captures. A proof must exercise the production user path and show the observable result a skeptical reviewer would accept. - Drive the running build through user-visible UI, registered keybindings, menus, and normal backend plumbing. Use `eval` to inspect state after the user path ran, not to invoke internal handlers as the primary proof. - Run `doctor` first. A video against a stale bundle is not evidence. - Read the relevant file under `references/features/`. Exercise every reachable entry point, mode, enabled gated variant, and the success / cancel / error / empty / persistence paths the change can affect. - For broad regression sweeps, walk `references/features/README.md` top to bottom, then `multi-surface-journeys.md`. - Show the trigger and the stable end state in the same recording. - Verify side effects, not just pixels: DOM attributes, clipboard, network/RPC log, file contents, or a reload / switch-away round trip. ## Driving conventions - Prefer ARIA labels, `data-component`, `data-action-id`, and `data-message-*` selectors. Class selectors are fallbacks. - Streaming is non-deterministic. Wait on observable end states with `wait-settle`, not fixed sleeps. - Prefer `press` with the registered chord over coordinate clicks. - The command service is not on `window`. Do not invoke `atlas.*` through `eval`. - Use `feature-flag` for gated paths. Reload when the feature file says the mounted tree needs it. - Treat real OS file drops, native context menus, and system-browser auth as manual unless the driver has first-class support. ## Feature map The behavior inventory lives in [`references/features/`](references/features/). Each file uses the same four H2s: `Sub-features`, `How to get to it (user POV)`, `Driving it with control-atlas`, `Gotchas`.