Core Web Vitals CrUX Monitor

AGENT ROLE: Autonomous field-performance agent. Query CrUX for each target, classify against the thresholds in references/thresholds.json, compare to any supplied baseline, and emit the JSON in references/output.schema.json. Field data only — do NOT substitute Lighthouse lab numbers.

OBJECTIVE

Report the p75 of LCP, INP, and CLS per URL and form factor, the good/needs-improvement/poor rating for each, the overall CWV pass/fail, and any regression versus the previous run.

INPUTS

• targets (REQUIRED): array of { "url": "..." } OR { "origin": "https://example.com" }. Origin returns aggregated data (useful when a URL lacks enough samples).
• form_factors (OPTIONAL): subset of ["PHONE","DESKTOP"]. Default both.
• baseline (OPTIONAL): prior run's results array for regression comparison.
• regression_delta (OPTIONAL): min change to flag. Default: LCP/INP +50 (ms worse), CLS +0.02.

AUTHENTICATION (CrUX API)

1. REQUIRE env CRUX_API_KEY — a Google Cloud API key with the "Chrome UX Report API" enabled.
   • IF unset THEN STOP error.code="AUTH_MISSING_API_KEY": "Create an API key in Google Cloud, enable the Chrome UX Report API, and set CRUX_API_KEY."
2. Endpoint: POST https://chromeuxreport.googleapis.com/v1/records:queryRecord?key={CRUX_API_KEY}.
3. This is an API key, NOT a service account — no OAuth scope, no property access needed (CrUX is public field data).

EXPECTED TOOL CALLS

• Run scripts/crux_monitor.py --targets targets.json [--baseline baseline.json].
• Per target × form factor: POST { "url"|"origin": ..., "formFactor": ..., "metrics": ["largest_contentful_paint","interaction_to_next_paint","cumulative_layout_shift"] }.

PROCEDURE (deterministic)

STEP 1 — For each target × form_factor call queryRecord. STEP 2 — IF 404 record not found (CrUX has no data for that URL) THEN:

• IF the input was a url THEN retry once as origin; set data_level="origin_fallback".
• IF still 404 THEN emit {status:"no_crux_data"} for that target (common for low-traffic pages). Do NOT error the batch. STEP 3 — Extract each metric's percentiles.p75. Rating: compare p75 to thresholds.json → good | needs_improvement | poor. STEP 4 — cwv_pass = true IFF LCP, INP, and CLS are ALL good at p75. STEP 5 — REGRESSION: IF baseline present, for each metric compute delta = current_p75 − baseline_p75; flag regressed=true IF delta ≥ regression_delta[metric] (worse direction). CLS worse = larger; LCP/INP worse = larger ms. STEP 6 — EMIT results sorted poor-first.

RATE LIMITS & ERROR HANDLING

• CrUX default quota is limited (per-key/minute + per-day). IF 429 RESOURCE_EXHAUSTED THEN backoff 2^attempt (max 5) then STOP error.code="RATE_LIMITED" with the count processed so far in partial.
• Serialize requests or cap concurrency at 2 to stay under the per-minute quota.
• IF 400 invalid url/origin THEN record {status:"bad_target"} for that target and continue.

MISSING / INSUFFICIENT DATA

• CrUX only reports metrics with sufficient real-user samples. IF a metric is absent from the response THEN set its value null, rating insufficient_data, and exclude it from cwv_pass (which then becomes unknown, never a false pass).
• Never fabricate a percentile; absence ≠ good.

OUTPUT

One JSON object per references/output.schema.json. No prose.

FILES

• scripts/crux_monitor.py — CrUX API client with fallback + regression logic.
• references/thresholds.json — Google's official CWV good/poor boundaries.
• references/output.schema.json — output contract.