Skip to content

Open-source AI browser agent for Chrome and Firefox (monorepo) đź§ 

3.7 KB
Raw Blame

WebBrain Cloud E2E

Persistent authenticated account workflows use the separate opt-in authenticated runner. They do not run in this public/incognito CI suite or inherit its mutation permissions.

ci/ runs catalogued browser-agent scenarios in a fresh WebBrain Cloud incognito browser. Each scenario produces a structured result, exported trace, deterministic rubric, optional .webm recording, and a suite summary.

The suite intentionally limits public-site coverage to read-only tasks. State-changing scenarios use the resettable Gnippets challenge lab owned by this project; they never touch production members or content.

Packs

Pack Coverage
cloud-smoke Ask-mode Wikipedia and NYTimes reads, a short YouTube download, two scheduled Google Form submissions, plus the isolated disposable-email, Turnstile, and OTP signup chain
public-readonly Wikipedia navigation, linked-page reasoning, table extraction
gnippets-readonly Public Gnippets developer/skill discovery
gnippets-spa Hydration, auth, prompt-injection resistance, search races, optimistic rollback, Shadow DOM, iframe, virtual list, modal portal, persistence
gnippets-captcha Standalone authorized Turnstile solve on the owned challenge lab

Run locally

$env:WEBBRAIN_API_KEY = "..."
$env:GNIPPETS_E2E_CONTROL_TOKEN = "..."
$env:CAPSOLVER_API_KEY = "..." # required by cloud-smoke signup and gnippets-captcha
npm run ci:e2e -- --pack cloud-smoke --no-video

Useful commands:

npm run ci:e2e:dry
npm run ci:e2e -- --scenario gnippets-spa-dom-gauntlet --no-video
npm run test:ci

Optional configuration:

  • WEBBRAIN_BASE_URL defaults to https://webbrain.cloud.
  • GNIPPETS_BASE_URL defaults to https://gnippets.com.
  • --concurrency N defaults to 2.

Artifacts are written beneath ci/artifacts/<timestamp>/ and ignored by Git. The GitHub Actions workflow uploads that directory even when a scenario fails.

Scenario session_settings are hard requirements. Provisioning must either report them in an applied/accepted list or attest their exact values in webbrain_config_result.enforced; a managed-but-unattested setting fails the scenario before the browser run starts.

Gnippets deployment

The sibling Gnippets app must explicitly enable its challenge lab:

GNIPPETS_E2E_ENABLED=true
GNIPPETS_E2E_CONTROL_TOKEN=replace-with-a-long-random-secret
GNIPPETS_E2E_TTL_SECONDS=3600

The control token is used only for create/inspect/delete lifecycle calls. The browser receives a high-entropy run capability URL. State is file-backed, expires automatically, and is separate from the application database.

The signup checkpoint sends a real six-digit OTP to the supplied inbox after Turnstile succeeds, but stores only hashes and expiring fixture state. Its CI artifacts are intentionally sanitized: no video, raw trace, disposable inbox credentials, password, message body, or OTP is uploaded. The sanitized trace retains only the Mail.tm request origin, first path segment, method, and matching response success/status needed to prove that the disposable-account and inbox APIs were actually exercised successfully.

Rubric

Scenarios are graded from three independent signals:

  1. WebBrain run status and schema-valid done_json output.
  2. Expected structured-result values and final host.
  3. For Gnippets challenge runs, server-observed events such as login_succeeded, post_created, or captcha_solved.

Failures include a stuck_at stage (setup, planning, navigation, execution, scheduled_execution, user_handoff, verification, artifact_capture, or cleanup) so the summary answers both “what passed?” and “where did it stop?”.