What's new in the Hover VS Code extension

The extension is how you use Hover now. It ships on the VS Code Marketplace as hyperyond.hover-dev, and it carries a new name: Hover - AI Vibe Testing. The pitch is short. Vibe-test your app, ship real Playwright specs.

The loop underneath hasn't changed. You type what you want tested in plain language, the agent drives your real Chrome over CDP, and the verified run crystallizes into a plain @playwright/test file. You bring your own CLI (claude or codex), and no AI runs in CI. What's changed is the surface around that loop.

Conversations sidebar

Chat sessions now live in a sidebar, in the style of Claude Code. Each session is its own thread with its own browser, so you can have one driving a checkout flow while another works through a signup form. Click between them to switch context. The browser state follows the session, not the editor, which means parallel work no longer steps on itself.

A reasoning-effort dial

Different tasks want different amounts of thinking. A two-step smoke check doesn't need what a fifty-field form does. The chat toolbar now has a reasoning-effort picker, so you set the per-run effort level before you send. Turn it down for quick, cheap runs. Turn it up when the agent needs to reason through conditional UI.

Environments and @account

Most apps need a logged-in user before any test does anything useful, and that user differs across local, staging, and prod. You can now define test accounts per environment. Mention @account in chat and the agent logs in for you using the account tied to the active environment.

The handling of the credentials is the part worth reading carefully. Passwords live in VS Code SecretStorage and are never written into the spec. When Hover crystallizes the run, the credentials get parameterized into process.env references, and a single click exports them to CI secrets. The spec you commit holds variable names, not passwords. There's a full write-up in multi-environment test logins without leaking secrets.

A pass/fail dashboard

Once you have a handful of specs, you want to see them at a glance. The Dashboard gives you a spec-by-run pass/fail matrix. It reads from a local run ledger, which the ▶ Run actions populate as you go. Run a spec, run a folder, and the grid fills in. It's a quick read on what's green and what regressed.

Grounded actuation

This one is invisible until you read the spec it produces, and it's the difference between a test that replays and a test that drifts. In normal mode, interactions go through Hover's control MCP using grounded targets: a role plus a name read straight off the page snapshot. The selector that drives the click during recording is the exact selector written into the spec.

Record equals replay. No confabulated getByText that looked right during exploration and breaks the next time the copy shifts. We wrote about why this matters and how AI authoring stays deterministic in AI-authored Playwright tests.

Security and pentest modes

The same chat does more than functional tests. Flip it to 🟠 security and the agent hunts IDOR and authorization gaps, then crystallizes confirmed findings into a .api-test.spec.ts that runs as a CI gate. Flip it to 🔴 pentest and it goes offensive against your own app only, writing up what it finds as a findings report rather than a spec.

These are real modes over the same engine, not a separate tool. The full picture is in security and pentest in the same editor chat.

Add CI Workflow

Crystallized specs are plain Playwright, so they run anywhere Playwright runs. To skip the boilerplate, there's an Add CI Workflow command that generates a GitHub Actions workflow to run your specs on pull requests. Combined with the CI-secrets export from environments, you go from a vibe-tested run to a PR gate without hand-writing YAML.

The parts that didn't change

Worth repeating, because they're the point. Hover drives your actual Chrome over CDP, not a fresh headless context. The artifact is standard @playwright/test, no proprietary format. You bring the CLI agent you already have. And once a spec is crystallized, there's no AI in the loop at CI time. The agent authors the test. Playwright runs it.

If you haven't installed the extension yet, the get-started install page walks through it.