Hover is now a VS Code extension — the chat moved into my editor

For most of its life, Hover shipped as a bundler plugin. You added vite-plugin-hover (or the Astro / Nuxt / Next / webpack equivalent) to your config, and at dev time a chat widget was injected into your page in a Shadow DOM. It worked — but it meant touching your app's build config, maintaining a per-framework integration for every bundler, and living with a floating panel on top of the UI you were trying to test.

The going-forward surface is a VS Code extension, and it's live on the Marketplace. Same engine, same thesis — AI explores once, then crystallizes a plain Playwright spec — but the friction is gone.

What changed

Nothing goes into your app anymore. You install the extension; you don't add a plugin, edit vite.config.ts, or take on a dependency. The extension spawns a local engine that connects to your existing dev server over the Chrome DevTools Protocol in an isolated debug Chrome. Whatever you build with — Vite, Next.js, Nuxt, Astro, webpack — it drives it unchanged, because it's not integrating with your bundler at all.

The chat lives in the editor. It's a webview in its own Activity Bar container, next to native tree views for Specs, Sessions, and Environments. The flow that used to be a widget hovering over your page is now a panel beside your code, which is where I wanted it the whole time.

A first session

1. Install Hover — AI E2E Testing & Security from the VS Code Marketplace.
2. Make sure one coding-agent CLI is on your PATH — Claude Code or OpenAI Codex — signed in with the subscription you already pay for. That's the only thing to configure; Hover bundles no model SDK and no keys of its own.
3. Start your dev server as usual.
4. Open the Hover chat and describe a flow: "log in, add a todo named 'verify hover', and confirm it persists after reload."

The agent drives your real Chrome, you watch the steps fold under the narration, and when the run is clean you hit Save as Spec. Out comes __vibe_tests__/<slug>.spec.ts — semantic getByRole / getByLabel selectors, every interaction guarded by a visibility check, plain @playwright/test. It runs in CI with no agent, no model, no token.

What you get beyond authoring

• Specs / Sessions / Environments views — browse crystallized specs (with inline Run and ✨ Optimize), the agent-run ledger with model and cost, and your per-environment targets.
• @account credentials — define test accounts per environment; mention @account in chat and the agent logs in, with secrets parameterized into process.env refs, never written into the spec.
• 🟠 Security and 🔴 Pentest modes — flip the same chat into authorization probing or offensive testing of your own dev app.
• Element → source — Alt-click an element to jump to where it's defined; CodeLens on specs shows the stamped prompt and any optimization candidate.

What about the bundler plugins?

They're frozen, not deleted. If you depend on the in-page widget today, it still installs and still works — but no new versions are planned, and the docs now point you at the extension first. The engine package, @hover-dev/core, is very much alive: it's what the extension ships inside itself.

The thesis hasn't moved an inch. AI is an authoring-time aid; the artifact is deterministic Playwright you own. The extension just makes getting to that artifact feel like part of editing code, instead of a thing bolted onto your dev server.