Saint Patrick

The brief was “just automate some clicking.” It became three cooperating services with Postgres as the only scheduler — this recording walks the whole system, then the live dashboard.

Bird's eye: a promo monitor, twelve parallel workers and a Next.js dashboard — every arrow is one service talking to the Supabase spine.

The life of a claimed account: atomic claim, human-paced farming, tasks, channel cleanup, safe release — one MTProto connection, always destroyed in a finally block.

The adversarial part: fruit captchas named in Russian, math challenges, and ban detection that triggers a fleet-wide PANIC — every wrapper re-checks after every tap.

Sponsor gates demand a human. A six-way URL classifier handles what it can; the Cloudflare wall gets relayed to an operator once — then every account executes it forever.

Nine tables, one spine. The constraints aren't decoration — SKIP LOCKED claiming and pre-committed schedules live in the database, not the workers.

The promo pipeline: detect in under five seconds, fan out through a 15-way concurrency pool, skip the gated — flip is_active fleet-wide the moment a code exhausts.

Nothing here was designed up-front — session guards, the PANIC kill-switch, the fail-closed janitor: every card is an incident that already happened.

What it sounds like at 4am — real production excerpts: rewards confirmed, cap limits respected, promo activations landing across the pool.

Now the client's window in: a passkey-gated dashboard with a live account grid — status, region and next slot for every account, straight from Postgres.

Metrics and balances as live charts — click counts, daily limits, error rates and star balances without a single manual refresh.

The promo monitor and action center: from code detection to fleet-wide redemption, with manual overrides when the operator wants a say.

That's the whole machine — a fleet that farms while my client sleeps, and a dashboard that proves it's working.