Evals Ui Laravel Package

Packaging only. No code changed, and nothing needs updating for it.

Changed

• Widened the Packagist keywords, adding evaluation, llm, llm-evaluation, ai-agents, pest and laravel-ai-sdk. Packagist search matches on name, description and keywords only — never the README — and keywords come from the tagged release rather than the default branch, which is the only reason this is a release at all.

Changed

• Breaking: vizra/evals must now be ^0.3. As in 0.2.0, the constraint is narrowed to what CI actually installs: the matrix runs composer update and always resolves the highest allowed version, so ^0.2|^0.3 would advertise a combination that is never built. Upgrade vizra/evals first and run php artisan migrate.
• The dashboard no longer reports a first run as fine. "At or above baseline" is vacuously true when there is no baseline, so a suite scoring 37% with a 12.5% pass rate was greeted with "Nothing needs you" — while the suite page one click away correctly said "No baseline set". With runs recorded and no baseline anywhere, the screen now says nothing is being compared yet and points at promoting a run.
• .claude is kept out of the release tarball. It holds this package's own maintainer skills; shipped, a consumer's coding agent loaded them on startup and offered to cut our releases.

Changed

• Breaking: vizra/evals must now be ^0.2. 0.1.1 allowed ^0.1|^0.2, which advertised a combination CI never installed — the matrix always resolved the highest version, so the dashboard was never once built against 0.1.x. It also let a dashboard read runs priced by 0.1.0, where claude-sonnet-5 was 50% high, and show those figures as fact. Nothing here needs an API that 0.2 added; this narrows what is claimed to what is tested. Upgrade vizra/evals first and run php artisan migrate.

Packaging and CI fixes. The dashboard itself is unchanged.

Added

• CONTRIBUTING.md and SECURITY.md.

Changed

• vizra/evals may now be ^0.1|^0.2, so the dashboard installs alongside vizra/evals 0.2.0.
• The Tailwind sources under resources/css are kept out of the release tarball. package.json was already excluded, so a consumer could not rebuild from them anyway, and fonts.css alone was roughly a third of the download. The built resources/dist still ships.
• CI varies Livewire 3.7 and 4.x as well as Laravel 12 and 13. composer.json advertises both Livewire majors and only one had ever been built against.
• README states PHP 8.4+, which is what composer.json has always required, and links to the vizra/evals repository rather than a relative path that only resolved in a side-by-side checkout.

Fixed

• Removed the repositories path block pointing at ../vizra-evals. Composer ignores repositories in a non-root package, so an installed consumer was never affected — but a missing path URL is fatal rather than a warning, so every CI job and every standalone clone died at composer update before installing anything. CONTRIBUTING.md documents the global path repository that replaces it.

First public release.

Added

• A Livewire dashboard for vizra/evals: overview, suites, runs, run detail with per-sample drill-down, and run comparison.
• Attention-first overview — suites furthest below their own baseline first, rather than lowest absolute score.
• Run detail with row filters, search, model filter, sorting and pagination, plus the gate's verdict and the commit that produced it.
• Trigger a run from the browser, dispatched to your queue.
• Promote and clear baselines, and mark a stale run failed, each behind a dialog that states what the action will do.
• ⌘K search across suites and runs.
• Horizon-style access control: the viewEvalsDashboard gate, or EvalsUi::auth(). Local environment is allowed when neither is defined.
• Self-contained styling — fonts and CSS are bundled, so the dashboard makes no third-party request and works offline.