E2e verification

E2E Verification (Playwright + local login)

Overview

Drive the running Chorus dev server in a real browser via the Playwright MCP to verify UI changes end-to-end. Core loop: log in with local credentials → navigate → snapshot to read state / act → screenshot for the human.

This is manual acceptance, not the automated Vitest suite (pnpm test). Use it when an AC says "the user sees X" and only a real browser can confirm it.

Prerequisites

• Dev server healthy on port 8637. Always check first:

  curl -s -o /dev/null -w "%{http_code}" http://localhost:8637/login   # expect 200
  

  If it's not 200 (connection refused, 500, or anything else), the server isn't running or is broken — start it with pnpm dev:local and re-check until it returns 200 before driving the browser. (Plain pnpm dev works too, but pnpm dev:local brings up the full local stack.)
• Playwright MCP tools available. They are deferred — load them first: ToolSearch("select:mcp__playwright__browser_navigate,mcp__playwright__browser_snapshot,mcp__playwright__browser_take_screenshot,mcp__playwright__browser_click,mcp__playwright__browser_fill_form")

Step 1 — Find credentials

Local dev uses the default-auth path (plain email + password on /login). Read it from the .env file at the repo root:

grep -E "DEFAULT_USER|DEFAULT_PASSWORD" .env
# DEFAULT_USER="[email protected]"
# DEFAULT_PASSWORD="chorus"

DEFAULT_USER logs in as a regular workspace user (e.g. [email protected]). This is NOT the same as SUPER_ADMIN_EMAIL (the /login/admin panel) or the Cognito OIDC flow described in the older .claude/skills/oidc-login.md — for normal page verification you want default-auth.

Step 2 — Log in (verified flow)

When default-auth is enabled, /login shows the email+password form directly (no SSO redirect):

1. browser_navigate → http://localhost:8637/login
2. browser_snapshot → grab the ref of the Email box, Password box, and "Sign In" button.
3. browser_fill_form with both fields:

   [{ name: "Email",    target: "<email-ref>",    type: "textbox", value: "[email protected]" },
    { name: "Password", target: "<password-ref>", type: "textbox", value: "chorus" }]
   

4. browser_click the "Sign In" button.
5. Success = URL becomes /projects. The tool result echoes Page URL — confirm it changed. If it stays on /login with an error banner, the credentials or the server's default-auth config are wrong.

Already authenticated from a prior session? The browser keeps the session, so you can skip straight to Step 3. To test the login itself, sign out first: on any dashboard page click the "Sign out" button (bottom-left of the sidebar), which returns you to /login. (If /login already shows the email+password form, you're logged out — there's no "Sign out" button to find; just proceed with Step 2.)

Step 3 — Navigate & observe

• browser_navigate to the target, e.g. a project dashboard: http://localhost:8637/projects/<project-uuid>/dashboard
• browser_snapshot is the workhorse — it returns an accessibility tree with stable ref ids. Use it to read page state and to get the ref/target you pass to browser_click, browser_fill_form, etc. Prefer it over screenshots for acting.
• Toggle UI and re-snapshot to compare states (e.g. the dashboard's Ideas/Stats and Flat/Lineage segmented controls each re-render the tree).
• If a navigate returns a near-empty tree, the page is still hydrating — call browser_snapshot again.

Step 4 — Screenshot for the human

Use browser_take_screenshot when the human needs to see the result (visual layout, spacing, color).

Always prefix the filename with .playwright-mcp/ — that directory is gitignored. A bare filename: "foo.png" saves to the repo root and pollutes git status.

browser_take_screenshot({ type: "png", filename: ".playwright-mcp/dashboard-lineage.png" })

Then Read it to view it inline. The screenshot tool result echoes a repo-relative path; Read needs an absolute one, so read it as <repo-root>/.playwright-mcp/<name>.png (prefix the path you passed with the absolute repo root). To capture a single element instead of the viewport, pass its target ref (from a snapshot). The MCP also auto-saves snapshot .yml and console .log files under .playwright-mcp/.

Quick reference

Need	Tool
Go to a URL	browser_navigate
Read page state / get element refs	browser_snapshot
Click / type / fill	browser_click, browser_type, browser_fill_form
Visual capture for the human	browser_take_screenshot → .playwright-mcp/<name>.png
Check console errors	browser_console_messages

Common mistakes

Mistake	Fix
Calling browser_* before loading schemas	ToolSearch("select:mcp__playwright__browser_...") first; bare calls fail with InputValidationError.
Screenshot with bare filename	Prefix .playwright-mcp/ or it lands in repo root (not gitignored).
Following the old oidc-login.md (Cognito) for local dev	Local dev = default-auth (.env DEFAULT_USER/DEFAULT_PASSWORD) → /api/auth/default-login → /projects.
Acting on coordinates from a screenshot	Screenshots aren't actionable; get refs from browser_snapshot.
"Login failed" but credentials look right	Confirm dev server is up on 8637 and Page URL actually moved off /login.
find-ing for a saved screenshot across the FS	It's already at the .playwright-mcp/<name>.png you passed; just Read it.