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):

1. 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.
2. 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).
3. 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.
4. 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).
5. 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.

Verification

• Field (Tier 1): node "${CLAUDE_PLUGIN_ROOT}/scripts/psi-client.mjs" --url <final_url> [--strategy mobile|desktop] [--timeout 60000] calls the PageSpeed Insights API (loadingExperience, method psi_api). It takes a URL only — there is no --snapshot mode, because field data is measured on the live origin, not on a saved DOM. The key comes from the environment (CLAUDE_PLUGIN_OPTION_PSI_API_KEY, else PSI_API_KEY); there is deliberately no --key flag. Without a key the script returns status: "needs_api" at exit 0 — branch on status, never on the exit code. When PSI sets origin_fallback, the numbers are origin-level and the finding must say so. CrUX History (crux_api) is not implemented by the bundled client (references/cwv-thresholds.md): never claim a p75 trend from this tool.
• Tier 0 (no key): emit status: needs_api for every field-CWV finding, plus clearly-labeled lab heuristics (render-blocking count, missing image dimensions, bundle weight). Label lab signals "lab data — not what Google ranks on"; never present a heuristic as a measured field value.
• When the required data tier is unavailable, status is needs_api — never a false pass.

Findings

Emit findings per schema/finding.schema.json. Severity policy: 5 is reserved for catastrophic, eligibility-killing facts at site/template scope; 4 major · 3 moderate · 2 minor · 1 cosmetic · 0 informational — and only an established severity-5 fail in an active category can cap a score (references/scoring-model.md). Field CWV never reaches 5: it is a tie-breaker, not an eligibility gate. Examples:

• M15.lcp.exceeds_p75 — field LCP > 2.5 s at p75 (status fail, severity 4, fixable: advisory, axis search, confidence established). evidence.observed quotes the measured p75 value and the LCP element; verification.reproduce runs the psi-client command.
• M15.cls.unsized_image — image with no width/height shifting layout (status warn, severity 4, fixable: advisory, axis search, confidence directional). evidence.observed quotes the <img> tag.
• M15.field.needs_api — no PSI/CrUX key, field CWV unverifiable (status needs_api, severity 4, fixable: advisory, axis search, confidence established). Each finding: evidence.observed quotes the page/measurement; verification.reproduce is the runnable command; expected_impact is banded + confidence-tagged (no naked %).

Honesty

• Lab ≠ field. A single local Lighthouse run is diagnostic only; Google ranks on field (CrUX) p75. Never let a lab number drive the Search score or masquerade as a field result.
• CWV is a real but modest tie-breaker signal — it does not override relevance/content quality. Don't promise ranking jumps from a green score; report it as banded expected_impact, not a predicted percentage gain.
• A "100" lab performance score is not a pass if field p75 fails; only field data settles a CWV finding.