web-perf

Run a performance trace with page reload to capture cold-load metrics, then analyze insights like LCPBreakdown and CLSCulprits. Compare results against thresholds such as LCP under 2.5s and CLS under 0.1 to determine ratings.

List network requests filtered by Script and Stylesheet types, then check for JS or CSS in the document head without async, defer, or media attributes. The RenderBlocking insight from a trace also flags these directly.

It uses Chrome DevTools MCP tools such as navigate_page, performance_start_trace, performance_analyze_insight, and list_network_requests. If trace tools are unavailable, it falls back to source and network analysis and notes what could not be measured.

Yes, when codebase access exists it detects the bundler from config files like vite.config.ts or next.config.js, then checks tree-shaking settings, barrel files, polyfills, minification, and compression. This phase is skipped for third-party sites.

Empty traces usually mean the page did not load correctly before recording started. Verify the page loads with navigate_page first, and check the trace response for the correct insightSetId and available insight names, which vary across Chrome DevTools versions.