Nightly Lighthouse (PageSpeed Insights), weekly Chrome UX Report field evidence, and agent-readiness monitoring for a watchlist of priority Webflow.com pages. For each page it tracks per-strategy (mobile + desktop) scores over time, classifies status, surfaces recommendations, lets you triage them into tasks, log change markers, posts one daily regression digest to a workspace webhook, and sends 2/7/30-day follow-up comparisons to Slack.
Built with Next.js (App Router) + React. TypeScript throughout.
npm install
npm run dev # http://localhost:3000 (the bundled launch config uses 3100)Other scripts:
npm run build # production build
npm start # serve the production build
npm run lint # eslint
npm test # vitest (unit + integration-focused concurrency tests)
npm run collector:check # type-check the durable Workflow worker
npm run collector:dry-run # bundle/validate the Workflow without deployingAll are optional for local development — the app runs without them.
| Variable | Purpose |
|---|---|
PAGESPEED_API_KEY |
PSI credential. Configure it on the collector Worker; it is also used by the local runner. |
CRUX_API_KEY |
Google Cloud key with the Chrome UX Report API enabled; required by the collector Worker. |
CRON_SECRET |
Existing shared bearer secret for FDE state/report access, Workflow dispatch, and manual nightly requests. |
COLLECTOR_URL |
Production Workflow endpoint, ending in /jobs. |
FDE_DATA_URL |
Optional FDE Worker base URL. If omitted, it is derived from COLLECTOR_URL. |
STORAGE_DRIVER |
Set to remote only after the one-time FDE copy verifies; unset keeps Webflow storage as the source. |
DATASET_MODE |
demo uses the existing sample namespace; live uses an isolated, initially empty namespace. |
BASE_URL |
Webflow Cloud mount path (for example /page-watch); client routes and APIs are prefixed automatically. |
SLACK_WEBHOOK_URL |
Incoming Slack webhook for 2/7/30-day follow-up reports. |
PSI_MOCK |
Local-only deterministic scores instead of PSI. |
PSI_RUNS |
Samples per strategy (1–5, default 5). |
ANTHROPIC_API_KEY |
Enables post-commit recommendation explanations and Watcher narratives on the app. |
ANTHROPIC_MOCK |
Uses deterministic placeholder AI text without making Anthropic requests. |
WEBFLOW_TOKEN_ENCRYPTION_KEY |
Collector-only base64 AES-256 key used to encrypt tenant Webflow site tokens before D1 persistence. |
AUTH_BROKER_URL |
Access-protected gateway origin used only for the email-code identity step. |
AUTH_PUBLIC_ORIGIN |
Fixed public app origin used as the signed handoff audience and callback origin. |
AUTH_HANDOFF_SECRET |
HMAC key shared only by the gateway and Webflow Cloud; use at least 32 random characters. |
AUTH_SESSION_SECRET |
Webflow-only HMAC key for host-only login sessions; use an independent random value of at least 32 characters. |
Put these in .env.local.
- Collection — baseline, on-demand, and nightly actions reserve a durable
job in FDE D1 and return
202. A Cloudflare Workflow performs up to five staggered PSI samples per strategy and stores every successful raw response, Lighthouse warning, and normalized failing audit in FDE R2. Cached/replayed responses are retained for diagnosis but do not count as independent samples. Warned runs are also excluded from trusted medians. A collection needs at least three unique, warning-free measurements before it can commit; otherwise the Workflow sleeps durably and retries up to twice at one-hour intervals. Recommendations require a strict-majority Lighthouse finding across those eligible runs. The Workflow also scans agent readiness and runs one fail-open Kitesurf rendered-page probe. Kitesurf retains compact DOM, accessibility, network, runtime-error, and diagnostic navigation evidence while staging the rendered HTML and accessibility snapshot in R2. Its non-Chromium timings never affect Lighthouse/CrUX scores, baselines, status, or collection success. The rendered HTML improves native-element detection when available; the existing HTTP scan remains the fallback. The Workflow commits the completed result directly into FDE storage. The SSO-protected Webflow app only makes authenticated outbound requests; the Worker never has to call into the Webflow Access tenant. Retries are durable, duplicates coalesce, stale jobs become visible failures, and a run ID can append history only once. - Baselines — ordinary on-demand/nightly runs may store snapshots, history, recommendations, and scan results, but a page stays Pending until the user explicitly captures a baseline. Zero placeholders are never treated or shown as real baselines.
- Scheduled collection — every 15 minutes, the FDE Worker reads the live
watchlist from its own D1 and dispatches only pages due in the workspace's
saved local-time/timezone window. Active pages receive stable 15-minute
offsets, and individual PSI samples are spaced one minute apart.
POST /nightlystill provides a force-all operation for authenticated manual tests. There is no Webflow callback and no Cloudflare Access service token. - Weekly data audit — each Monday at 05:30 UTC, the Worker reconciles the prior seven days of stored scores against their raw PSI responses. It checks raw-report coverage, usable sample counts, Lighthouse warnings/runtime errors, score medians/ranges, and finding quorum. The saved health report has hashed page references and no customer names, URLs, raw payloads, or raw errors. See audits/README.md.
- Weekly field evidence — each Tuesday at 06:15 UTC, the Worker queries the
CrUX History API for phone and desktop evidence on every active page. It
prefers exact-URL data, falls back to origin data only when URL evidence is
unavailable, stores rolling 28-day p75 metrics and histograms in dedicated
D1 tables, and retains the complete provider response in R2. An authenticated
POST /crux/collectruns the same collection manually. - Storage — the production source of truth is the FDE-owned
page-watcher-fdeD1 database plus thepage-watcher-reportsR2 bucket. The Webflow app uses a tenant-scoped remoteDataStore; D1 state updates use version-guarded compare-and-swap retries. Local development continues to use.data/. The Webflow-provisioned D1/R2 bindings remain intact as a reversible migration source and are not deleted by the migration. - State mutations go through targeted server-side domain endpoints
(
/api/pages,/api/recs,/api/pages/[id]/*) and the store-level atomic update primitive. External PSI/webhook/Slack work happens outside that critical section; result commits re-read authoritative state. - Background execution — Cloudflare Workflows own production execution.
The local runner uses Next.js
after()only for development. - Post-commit enrichment — after the scores/raw reports are safely stored, optional Anthropic recommendation explanations, Watcher narrative refreshes, webhook alerts, and due Slack follow-ups cannot roll back or mislabel a successful collection.
- Webflow Enterprise activity — Settings can connect one Enterprise site
per tenant with a site token scoped to
sites:read,site_activity:read,pages:read,assets:read, andcms:read. The collector validates every capability, encrypts the token with AES-GCM before D1 persistence, imports normalized actor/resource activity into D1, retains bounded raw events in R2, and refreshes activity on the existing 15-minute scheduler. Webflow failures are recorded on the connection and never block scheduled PSI. See docs/webflow-activity.md for the product contract and remaining publish-verification phases.
The interactive app is served directly from https://page-watcher.webflow.io.
Its login button briefly sends the browser to the existing Access-protected
gateway for Cloudflare's email-code authentication. The gateway validates the
Access JWT, signs a one-minute state-bound handoff, and returns the browser to
the Webflow hostname, which creates a signed host-only session cookie. No
custom email sender or DNS access is required. The app's role registry remains
authoritative for app-admin and project-level access, including after login.
Local storage keeps only the last opened project ID. See
docs/native-auth.md for deployment and cutover steps.
The immutable bootstrap app administrators are matthew@webflow.com,
ben@webflow.com, and diego.rangel@webflow.com. Additional app admins must
also use @webflow.com. Non-interactive nightly and collector result endpoints
remain protected separately by CRON_SECRET.
-
Apply
migrations/to the FDE-ownedpage-watcher-fdedatabase, then deploycollector-worker/wrangler.jsonc. The Worker uses itsBROWSERbinding for Kitesurf and needsPAGESPEED_API_KEY,CRUX_API_KEY,CRON_SECRET, and a base64-encoded 32-byteWEBFLOW_TOKEN_ENCRYPTION_KEYgenerated withopenssl rand -base64 32. Its D1, R2, Workflow, 15-minute due-page/Webflow-activity scheduler, Monday 05:30 UTC audit, and Tuesday 06:15 UTC CrUX Cron bindings are declared in that config. Each schedule resolves the shared project registry and runs every active tenant independently; one project's failure does not block the remaining projects. -
Deploy the Webflow app code with
STORAGE_DRIVERstill unset. The app keeps reading and writing its existing Webflow-provisioned D1/R2 bindings at this stage. -
While signed into the app as an app administrator, make this same-origin request from the browser console:
fetch("/api/admin/migrate-fde", { method: "POST", headers: { "content-type": "application/json", "x-page-watcher-migration": "copy-to-fde", }, body: "{}", }).then((response) => response.json()).then(console.log)
The endpoint copies the tenant selected by the request's
projectquery parameter (or the first active project when omitted) and all of its raw reports outward, verifies a SHA-256 state checksum, and never mutates the source. It refuses to overwrite differing destination state unless the body explicitly contains{ "replace": true }. -
Only after the response reports
ok: true, setSTORAGE_DRIVER=remoteon the Webflow app and redeploy.FDE_DATA_URLis optional whenCOLLECTOR_URLalready points to the same Worker and ends in/jobs. -
Call
/api/health;storage.drivershould beremote. Run one page or send an authenticatedPOST /nightlyto the FDE Worker, then verify the new history entry before relying on the next scheduled run. After the first weekly audit, an authenticatedGET /audits/weekly/latest?tenant=<tenant>on the collector returns that project's privacy-safe health summary; the public collector/healthresponse exposes only aggregate status and project counts. Before relying on the CrUX Cron, send one authenticatedPOST /crux/collectand verify thecrux_snapshotsandcrux_statusrows plus the public/healthsummary.
Rollback is just as deliberate: remove STORAGE_DRIVER=remote and redeploy to
return to the preserved Webflow bindings. Do not write to both stores after
cutover and expect automatic bidirectional merging.
Use DATASET_MODE=demo for the existing sample state and DATASET_MODE=live
for real URLs. The two modes use separate tenant keys in the same D1 database,
so switching modes is reversible and never overwrites the demo dataset.
The bundled demo is date-relative and intentionally scenario-rich. It includes stable, improving, regressing, paused, failed, partial, agent-only, and pending pages; custom and task-linked chart markers; a verified PSI provider incident with excluded anomaly bands; Lighthouse quality/provenance, native Webflow elements, Kitesurf evidence, agent-ignore history, and recommendation/task lifecycles. Separate CrUX fixtures cover URL and origin scope, stable/improving/ worsening trends, partial metrics, insufficient traffic, and provider errors. When no collector URL is configured, Settings also shows a local Webflow activity/publish example.
The disposable brand-studio demo namespace carries a fixture version. Local
filesystem and demo D1/FDE stores replace an older bundled fixture once when
that version changes; live namespaces and arbitrary project tenants are never
refreshed this way.
- Automated page remediation
The four fixed scoring/scheduling choices for this phase are documented in DECISIONS.md.