verify-playground

Verify the playground

Read .agents/skills/e2e/SKILL.md first. This skill only says what is true of this app. The config is e2e.config.ts (target web, tests tests/**/*.e2e.ts); the bug bash uses e2e.bugbash.config.ts.

Launch

The target's app.command runs node app/server.mjs with PORT=4271. The runner starts it, waits for http://127.0.0.1:4271, and stops it after the run. Nothing to start by hand. The app needs no env, seeds, or database; todos live in the browser's localStorage, the session in a cookie. To watch the app by hand, pnpm app serves it on the same port, and a run then stops with APP_ALREADY_RUNNING until it is stopped. The port is fixed, not reused.

Doctor

npx e2e list tests/smoke.e2e.ts       # loads the config, starts nothing; exit 0 and four pairs
npx e2e run tests/smoke.e2e.ts        # starts the app, opens it, pins the Playground heading; exit 0

list failing is the config or a dependency. run failing with APP_ALREADY_RUNNING is a server of yours on the port, and LOCATOR_NOT_FOUND on the heading is the app itself. A signed-in check is npx e2e run tests/auth.setup.e2e.ts, which saves the admin and admin-cookie sessions. Run Doctor first whenever anything looks off.

Drive

Live, over the registered e2e mcp server: open_session, then observe, the verbs, and locate before writing any locator (topic mcp). The catalog offers type_secret for the admin credential and no project tools. Scripted: a test under tests/, screen for every control in the feature map (all have accessible names), agent.act for a flow that varies. Signed-in tests start from the admin session that tests/auth.setup.e2e.ts saves ({ session: 'admin' }); admin-cookie is the same account signed in without a fill, so screenshots stay.

Evidence

Every run writes .e2e/report.json. --trace on writes .e2e/results/<test>/trace.md for passing tests too; --video records each attempt under the test's attempt-<n>/; app.screenshot('<name>') saves a named frame. The next run clears .e2e/results/, Doctor included, so a proof that must outlive it runs with --output .e2e/proof/<name>. A proof drives the page the way a user does, the nav link or the route, the labelled control, the button, and captures the action and the state after it, not only the final screen. Side effects: todos are localStorage.todos (read with browser.evaluate), the session is the session cookie (browser.cookies()); the server keeps nothing.

Bug bash

e2e.bugbash.config.ts (untracked) spreads e2e.config.ts, adds tests/bugbash/** to the glob, a model, and the skeptic and fuzzer personas; its context says what is not a bug here (seeded data, the admin account, no email). The charters are in features/, one H2 per feature, already in the slug|target|agent|charter line format the fan-out in topic bug-bash step 3 reads. .e2e/bugbash/ is in .gitignore, with .e2e/proof/. One explorer at a time needs no isolation change; several need pnpm app started first and reuseExisting: true on the command, since the port is fixed. A finding is a bug once tests/bugbash/<slug>.e2e.ts fails with ASSERTION_FAILED on the assertion that encodes it; anything else is the explorer's claim.

Cleanup

The runner stops the app it started; close_session stops the one a session started; a pnpm app you started is yours to stop (its pid, never pkill node). Remove .e2e/bugbash/<slug>/ for a charter whose findings were all rejected; a charter the report cites stays until the user has the report. Keep .e2e/results/, .e2e/proof/, .e2e/report.json, and the videos under each attempt. That is the proof, and all of it is gitignored.

Helpers

None. The e2e CLI is the helper.