What Mario approved
Mario asked us to connect automatic reading to the private review desk, keep the project on Cloudflare Free and use only the integrated browser for Cloudflare administration. A saved proposal still does nothing by itself. Only an approved website enters the reading queue. This release supersedes the decision-only stage described in note 009.
A deliberately small pace
A scheduled step checks for work every fifteen minutes and processes at most one page. The application allows eight page attempts per UTC day and 240 per UTC calendar month. Failed or excluded attempts count too. Each approved website has a maximum of ten pages and two link levels, or just its submitted page when internal pages are not selected. The collection remains limited to 100 website records and 1,000 indexed pages.
Pause before more work
The database reserves an attempt before any website is fetched. It prevents overlapping processing and reserves sixty browser seconds per attempt against an eight-minute daily application allowance. Reading pauses when an application allowance is reached or the provider reports a relevant quota error. Allowances reopen in the next eligible UTC period; a manual pause remains in place until Mario resumes it. Failed pages need a deliberate retry. The display measures this application’s reservations, not every service or other project in the Cloudflare account. The account dashboard still shows Workers Free; no paid upgrade was enabled.
What gets read
The crawler checks public HTTPS addresses and redirects, reads robots.txt and respects page-level noindex and nofollow. It follows only links on the approved hostname. This first version reads HTML without running website scripts or signing in. Cloudflare Browser Rendering processes stripped HTML with external requests blocked. BGE-M3 embeds a bounded text extract, Vectorize stores the vector, and D1 holds the URL, title, short extract and processing state. JavaScript-only pages, oversized documents and excluded content may be skipped.
What we verified
Mario used Read next page in the protected live desk. The homepage of mario-obst.fyi was fetched, embedded and submitted to the index, and internal links appeared in the bounded queue. A public collaboration query returned real results from the updated collection. The first UI query reported a connection failure; a deliberate repeat succeeded. We did not identify the cause of that single failure and have not claimed it was fixed. Local SQLite tests cover the daily and monthly stops, exclusive processing, period changes, manual pause, retry permissions and exclusion of unapproved pages from results. Local UI checks used clearly labelled test data. That first verification was a manually requested processing step. Follow-up on 14 October: we left the fifteen-minute schedule unchanged and did not request another manual step. At 23:30:08 Europe/Berlin time (21:30:08 UTC), processing started automatically. The approved marioobst.com homepage completed at 23:30:12. D1 held three pages instead of two. The first search was too early for Vectorize; a later deliberate search for ‘strategische Kommunikation’ returned marioobst.com first among three results. This verifies one complete scheduled cycle, not long-term reliability. Cloudflare’s separate cron-history view still showed its stated reporting delay.
What stays separate
The existing 500 daily and 15,000 monthly search allowances remain unchanged. Public website suggestions remain closed and IndexNow remains deferred. Administrator authentication stays private. There are no new visitor identifiers or analytics. The privacy notice now explains the reading queue, public source content and provider processing. Recrawling, removal controls and JavaScript-heavy websites remain later work.