vlmkit

Reference for the `vrt` CLI — Visual Regression Testing combined with accessibility (a11y) semantic verification. Use when running `vrt-test` / `vrt` (Playwright VRT) / `vrt-update` / `vrt compare` / `vrt snapshot`, configuring the fix-loop CSS challenge benchmark, or picking a VLM model for diff analysis.

Turn a natural-language spec into Playwright tests with deterministic, reproducible VRT and a self-healing loop. Uses the official init-agents planner/generator to explore the app and generate tests, a determinism layer to keep screenshots byte-stable, and @mizchi/vlmkit-heal to auto-fix failures. Use when authoring Playwright tests from a spec, making visual regression tests deterministic, getting generated tests to reproduce in CI, or healing tests that drift from the UI.

Closed-loop CSS auto-repair. Given a fixture with a known regression (one CSS property or one selector block removed), iterate with a VLM that proposes the missing fix from the diff screenshot, apply it, and re-run until the diff falls below a threshold. Currently scoped to the CSS-challenge fixture set in `src/experiments/css-challenge/`; adapting to an arbitrary repo requires writing a fixture entry. Use when measuring whether a VLM model can recover a known regression, not for production self-healing.

Run vrt diff in a stateful loop where each run is compared against the previous run's persisted summary, surfacing a `⚠ REGRESSION` banner when the majority of viewports get worse. Designed for periodic CI gates (per-PR or scheduled) where you want "did this change make things worse" as a binary signal, not a one-shot snapshot. Stores summary at `.vrt/last-diff-for-agent.json` by default.

Compare two rendered pages (URL pairs or local HTML) and produce a structured Markdown report tailored for coding agents — pixel diff per viewport, per-section diffRatio, computed-style diff split into universal vs. breakpoint-gated, and worst-viewport screenshot paths. Use when the agent just made a UI change and needs to know whether it altered visible output, where it altered it, and which CSS properties drove the change. Works on a single HTML/URL pair without baseline state.

Evaluate whether a framework / CSS-library / build-system swap (Tailwind → vanilla CSS, reset-css switch, bundler swap, etc.) produced a visually-equivalent result. Three modes — compare (deterministic pixel + CSD), blind (agent runs without baseline reference), subagent (dispatched-agent verification) — let the caller pick the rigour level. Use when the diff is large by construction (the markup was rewritten) and a flat pixel diff would drown the actual regressions in noise.

Five DOM/pixel-based signal tools for markup work — turn a screenshot + HTML scaffold into a converging pixel diff (build component), crop a page screenshot into per-component PNGs (scan component), audit hard-coded values against a design-token scale (check tokens), verify theme parity (check theme — light vs dark render), and stress-test layout under inflated text (stress i18n). All five are pure DOM + Playwright + pixel processing — **no VLM / no API key required**. The agent supplies the markup reasoning; the tool surfaces the signal. Use when you've authored HTML/CSS and want signal back without paying VLM latency or cost.

Improve a developer-facing tool (CLI, library, agent harness) by running closed-loop validation with disposable subagents, treating their friction as the deliverable, fixing the friction, and re-running. Use when the user wants to evaluate or harden a tool whose value depends on whether an agent can use it — vrt-like signal loops, agent SDKs, CLI ergonomics, prompt scaffolding, IDE integrations. The loop produces both a stronger tool and a written record (per-run reports + tracked issues) of what was learned and why.