capture

Capture — Browser Automation Skill

Browser automation via CDP for UI validation. For the full command reference, see cli-reference.md.

Session Basics

1. capture session start --url <url> — open a tab and start recording
2. Interact: screenshot, click, type, a11y, exec, navigate
3. capture session stop <id> — bundle artifacts; capture session view <id> to inspect

Session context auto-fills --target and --har — no manual flag threading needed.

Target selection: If the user hasn't specified which app or browser to target, ask before starting. Don't assume Chrome, a specific port, or a running tab — run capture detect to see what's available and confirm with the user when ambiguous.

Validation Pattern

When validating a UI feature:

1. Start a session targeting the page under test
2. Use a11y --interactive to understand what's on the page
3. Interact via click and type using accessible names (not selectors)
4. Take screenshots at key states
5. Use exec to check DOM or JS state when a11y isn't sufficient
6. Use har read to verify network requests if relevant (supports --filter-url, --filter-status, --filter-method, --limit; ID optional in an active session)
7. Stop the session and report pass/fail with evidence

HAR caveat: the session HAR is populated by the commands that run — most reliably navigate. click/type capture the traffic that fires inside their settle window, but cross-navigation traffic after the frame changes is lossy. For continuous "click around" capture, run capture record --duration N in parallel.

Interaction Tips

• Always prefer click "Name" over exec with selectors — a11y names are stable
• If click reports multiple matches, add --role to disambiguate
• Run a11y --interactive before interacting to see available elements
• type --into "Field Name" clicks the field first, then types — no separate click needed
• After interactions, screenshot to capture the resulting state