Web Inspect

Live point-and-comment QA for frontend apps. The user runs their dev server, opens the app in a browser, pins elements with bugs, types comments, and sends batches to the agent. The agent fixes each bug sequentially in the source.

How it works

1. Boot — node {{scripts_path}}/start.mjs starts a small HTTP+SSE helper server, injects the overlay <script> into the project's HTML entries, prints a JSON blob with { port, token, devUrl, pageFiles }.
2. User tests — User opens their dev URL in a browser. Overlay floats over the page. They click elements, leave comments, and click Send to agent when ready.
3. Poll — Agent runs node {{scripts_path}}/poll.mjs in a loop. The helper server holds the request open until the browser sends a batch (or 10 min timeout).
4. Fix — Each batch contains N annotations. Agent triages, then fixes one at a time in the main thread, proposing each change before applying (auto mode applies directly).
5. Repeat — After fixing, agent polls again. The user keeps testing and sends more batches.
6. Exit — User says stop, or the helper server is killed. Skill runs node {{scripts_path}}/start.mjs stop to restore the HTML files.

The helper runs on a random localhost port with token auth. The PID file lives inside the skill directory at .runtime/state.json, so nothing is written to the project root.

Setup gates

Before any other action, pass these checks. The first failure stops the skill.