benchmark

Capture a baseline with /benchmark <url> --baseline before making changes, then run /benchmark <url> on your branch. The skill compares TTFB, FCP, LCP, bundle sizes, and request counts against the baseline and flags regressions using relative thresholds.

It measures TTFB, First Contentful Paint, Largest Contentful Paint, DOM Interactive, DOM Complete, full load time, total requests, transfer size, and JS/CSS bundle sizes. Data comes from the browser's performance.getEntries() API, not estimates.

Use /benchmark --diff mode. It runs git diff against the base branch to identify changed files and benchmarks only the pages impacted by those changes, rather than auditing every page.

Yes, use /benchmark --trend to load historical baseline files and display a table of FCP, LCP, bundle size, request count, and grade across past benchmarks. This reveals gradual degradation that single runs miss.

Timing metrics regress at over 50% increase or 500ms absolute increase, with warnings at 20%. Bundle sizes regress at 25% increase with warnings at 10%, and request counts warn at 30% increase.

Without a baseline, the skill can only report absolute numbers, which are meaningless without context since acceptable load times vary by page type. Relative comparison against your own baseline is what identifies actual regressions.