seo-core-web-vitals (M15)

Core Web Vitals are a confirmed Google ranking signal and a proxy for page quality. Thresholds, field-vs-lab honesty rules, and LCP decomposition: references/cwv-thresholds.md (follow it exactly).

Inputs

Work from the PageSnapshot named in your dispatch envelope: read parsed from <run_dir>/pages/<slug>.json (scripts[], stylesheets[], images[]) plus timing.ttfb_ms, body.bytes, and render — field data still needs the PSI tier; Grep pages/<slug>.html for verbatim evidence; site artifacts live in <run_dir>/site/{robots.json,sitemaps.json,discovery.json}. Deterministic findings already emitted by audit.mjs are listed in <run_dir>/findings.deterministic.json — do not re-emit those ids; add model-judged findings only. If invoked directly with a URL/path and no snapshot exists, first run node "${CLAUDE_PLUGIN_ROOT}/scripts/snapshot.mjs" <target> --out "${CLAUDE_PLUGIN_DATA}/runs" and use the printed snapshot path.

Audits

Working from the PageSnapshot (parsed_rendered when render.used is not none, else parsed):

Field CWV at p75 — LCP ≤ 2.5 s, INP ≤ 200 ms (INP replaced FID), CLS ≤ 0.1, all at the 75th percentile of real-user data. All three must pass for a "good" rating. Use the proactive-warning thresholds (LCP > 2.0 s, INP > 160 ms, CLS > 0.08) to warn before failing.
Render-blocking resources — synchronous <script> in <head> (no defer/async/type=module), blocking <link rel="stylesheet">, and @import chains that delay first render (LCP/INP risk).
Unoptimized images — missing width/height (or aspect-ratio), no loading="lazy" below the fold, no responsive srcset/sizes, legacy formats where AVIF/WebP would serve, and a non-preloaded LCP image.
Layout-shift sources — images/iframes/ads/embeds without reserved dimensions, web-font swap without font-display/size-adjust, and content injected above existing content (CLS risk).
LCP decomposition — attribute LCP to TTFB vs resource load delay vs load duration vs render delay (see references/cwv-thresholds.md) so each finding targets the dominant subpart.

Fixes (fixable: advisory)

Advisory only. Performance fixes touch build config, server, CDN, and runtime JavaScript — high breakage risk — so the tool diagnoses and prioritizes but never auto-edits performance code. For each failing metric, name the specific resource/element causing it and rank fixes by expected leverage (e.g. preload the LCP image, defer non-critical JS, reserve image dimensions). Give framework-specific direction where the stack is known. Never fabricate measured values — if field data is absent, ask the user to supply a PSI/CrUX key or leave a clearly-marked TODO placeholder rather than guessing a metric.