Check accessibility of a UI from the real render: theme contrast, labeling of every interactive control, and per-text-node APCA + WCAG 2 legibility. text_contrast_failures lists nodes failing the strict body-text floor even when the theme verdict is legible (catching authored low-contrast text). Returns a structured report; this is a normal result whether or not it passes.

Check layout geometry from the real frame: interactive targets below the 24px minimum hit size, and signal-bearing nodes that fall outside the window (clipped / off-screen). Returns a structured report — a normal result whether or not it finds problems.

Return the formal JSON Schema for a fenestra/1 description — a machine-checkable input grammar to validate or autocomplete against before rendering, the structured complement to describe_vocabulary's prose grammar. Structural only (shape, required/optional fields, unions); still call validate for semantic checks like colour roles, numeric ranges, and enum-like fields.

Return the description grammar: every node type with a minimal example, and the theme color roles a color may name. Call this first to learn how to author a fenestra/1 description.

Render a filmstrip: drive optional interaction steps (applied first, so a click can trigger the transition to watch), then capture frames with real motion turned on — every other tool stays reduced-motion for deterministic pixels; this is the one place the point is watching motion play — and compose them into one captioned strip image. Returns the actual frame count/interval/scale used (each is clamped to a documented ceiling, so a hostile request degrades instead of hanging) plus the strip's pixel dimensions, a downscaled inline preview, and a resource_link to the full-resolution strip.

Return the keyboard focus order: the refs a Tab cycle visits, in order, honoring a modal focus trap (disabled controls excluded). Verifies reachability and tab sequence as typed data, without driving the UI.

Drive a UI through scripted interactions (click, type, key, tab, hover, wheel, drag — by semantic selector, never coordinates). Returns the emitted intent messages and the access tree afterwards; set screenshot=true for a preview image.

Match an expected accessibility snapshot (Playwright - role "name" [attr] grammar) against a UI. mode: partial (subset, default) | strict (exact) | regex (each expected line is a pattern). Returns a pass/fail diff — a normal result.

Compare a UI's render against a baseline PNG (path on disk), pixel by pixel, with an optional per-channel tolerance, differing-pixel budget, and mask rectangles to ignore (volatile regions). Returns the diff stats and a diff-image preview on mismatch — a normal result.

Find nodes in a UI by a semantic selector (role, name, value, or id). Returns matches with stable refs; on a miss, returns the nearest candidates to guide a retry.

Render a fenestra/1 UI description to a typed accessibility tree, a downscaled preview image, and automatic accessibility warnings (contrast, labeling, per-text-node legibility). Read the access tree first; the full-resolution PNG comes back as a resource_link.

Run a verification scenario in one pass: a fenestra/1 description, optional interaction steps, and a bundle of expectations (emitted intents, a11y, an aria snapshot, a screenshot baseline, query match-counts). Drives the steps, then asserts every expectation against the resulting frame — the screenshot check compares the POST-interaction pixels. Returns a unified report (one ok plus a per-check breakdown) and a preview: the diff image on a screenshot mismatch, else the final render. A failed check is a normal result, not an error.

Validate a fenestra/1 description without rendering. Structural problems (unknown fields, bad node tags) and semantic ones (an unknown color role) come back path-pointed. Returns isError when the description is invalid.