dynamic-markup

Run check animation on the attempt page and compare its evaluated animations list against the brief line by line: count, target selector, duration, iterations, and leg time for oscillating animations. It flags no-visible-effect animations, missing reduced-motion overrides, and uncontrolled motion from rAF scripts or videos.

Use check breakpoints to render at B-1, B, and B+1 for every discovered breakpoint, catching boundary spikes, gaps, and overflow-at-boundary defects. Add --sweep to fuzz the full width range for fixed-width-child bugs that boundary checks never render.

No API key is required. All gates are deterministic Playwright plus pixel and DOM math; the agent's own vision reads the target screenshots. Keys are only needed for optional vlmkit features like heal markup or check copy --vlm.

Yes, check interactions Tab-walks the page, fires each role's canonical keys, and records ARIA transitions, focus traps, and popup behavior. The --handlers flag also finds pointer-only controls that have click handlers but no role or keyboard support.

Pixel diffs cannot see behavior bugs: a panel that grew instead of scrolling, an animation that never fires, or a dead aria-expanded attribute. The dynamic gates (scan scroll, check animation, check interactions) detect exactly these classes, which is why the done condition requires both converged composition and clean gates.

Use dynamic-markup when the task specifies how the page behaves — resizes, scrolls, animates, or responds to keyboard input — not just how one screenshot looks. Static pixel convergence itself is delegated to the auto-markup pipeline as Phase A.