agents md for using vercel agent browser

AGENTS.md

AGENTS.md

Browser Testing Engine

I am equipped with agent-browser. When asked to "run a test case," I follow this standard operating procedure:

1. Locate: Find the test file in tests/browser/.
2. Setup:
   • Extract the URL and Session name.
   • Run agent-browser --session <name> open <URL>.
3. Execution Loop:
   • For each step in the Markdown file, I will first run agent-browser --session <name> snapshot -i to get element refs (@e1, @e2, etc.).
   • I will map the natural language description (e.g., "Email input") to the correct @ref.
   • I will execute the corresponding command (fill, click, type).
4. Verification:
   • I will check the "Expected Result" section.
   • Run agent-browser --session <name> errors to ensure 0 errors.
   • Run agent-browser --session <name> screenshot <test-name>.png.
5. Report: Summarize the pass/fail status and mention the screenshot location.

Execution Rules

1. Always Use Refs: I prefer element refs (e.g., @e1) over CSS selectors. They are deterministic and faster.
2. Snapshot First: Before any interaction, I run agent-browser snapshot -i to discover current refs.
3. Re-Snapshot: After any click or navigation that changes the DOM, I MUST run a new snapshot -i to refresh the refs.
4. Network Idle: I use wait --load networkidle after open or click to ensure the page is fully hydrated.
5. Session Isolation: I use the --session flag to maintain state across multiple commands.

Key Command Reference:

• agent-browser --help for all commands.
• open <url>: Navigate.
• snapshot -i: Get interactive elements + refs (Best for AI).
• click @ref: Click by ref.
• fill @ref "text": Input text by ref.
• errors: Check console logs for failures.