Alerts

Backed by: our own hourly schedule, reading your Worker. Scopes: payments:purchases:read, payments:reconcile:read and testers:roster:read, plus manifest:read to find the routes. Plan: Team.

What it answers

Would I have heard about this before a customer told me?

Every other pane knows what it knows only while you are looking at it. A renewal that failed overnight, a reconcile that stopped running, a cohort about to break its streak: each one sits in a pane and reaches nobody. Alerts is the schedule that asks for you, once an hour, and mails you once per occurrence.

Reconciliation, purchases and cohorts are where the facts come from, notifications is who hears them, and what it never stores is what the schedule keeps.

The four conditions

ConditionWhat it looks forDefault threshold
Failed renewalsPurchases newly in_grace or on_holdOne. Several in one hour are one mail with a count
Webhook driftA finished reconciliation pass that repaired purchases its webhooks left out of step. Never a dry run, and never a pass older than 48 hoursOne repaired purchase
Reconcile stopped or failingNo finished pass inside twice the interval, a pass that could not read some purchases, or no pass at all. And once when it resumes24 hours between passes, so overdue at 48
Tester streak at riskAn open cohort still holding its run, with fewer active testers than its targetHeadroom of 0: the next lapse resets or pauses the run. Once per cohort per UTC day

Every condition has its own switch and its own threshold, per connection. A dry run keeps nothing alive: only a real pass counts as reconciliation running.

When it runs

Hourly, at the same minute each hour. Each connection is given its own second in the hour, worked out from its id, so its check time is stable from run to run and the fleet does not all arrive at once.

Production is evaluated without asking. Any other environment is off until somebody adds it with Evaluate this environment.

A Worker that does not answer is not polled forever. Each read gives up after eight seconds. From the third failure in a row the checks back off, two hours, then four, up to a day, and you are told once that it stopped answering and once when it reads again.

Once per occurrence, not once per run

Each finding has a key: a purchase and its status, a reconciliation pass, a cohort and a day. The same failed renewal seen again next hour sends nothing.

What was already said is remembered by a SHA-256 of that key, for 180 days. The hash answers was this said? and cannot answer what was it? No purchase, subject or amount is stored.

A condition that cannot evaluate says so

A withheld scope, or a route your Worker does not declare, is never silence. The condition shows as Not evaluating with the scope or route named, and you get one mail saying which. Grant the scope, or upgrade the capability that declares the route.

Who hears it

Each condition is an event in the notification catalog, so it follows the same switches as everything else: Notify me per person, Alert the team per organization. All four are on by default.

Plus up to ten extra addresses per organization: an on-call alias, a compliance inbox. Each one follows Alert everyone, and gets nothing until it confirms. The confirmation link opens a page with one button, so a mail scanner fetching the link confirms nothing. Changing an address tells the previous one.

A failed delivery is on the screen. Each condition’s record shows how its last mail went: sent, on its way, or failed.

What a mail carries

Counts, the names of capabilities, environments and cohorts, and a link into the pane. Never a subject id, an address or an amount.

Who may change it

Owners and admins set the switches, thresholds, environments and extra addresses. Every one of those changes is audited. Members see the same record, read-only.

On Free and Solo

Every row is drawn locked, and says what Team buys. Nothing is read: the schedule does not run for an account without it. Every read it would make stays open to you in its own pane, through your CLI, and through your Worker’s API.

A billing lapse stops it. The schedule checks the plan before it dispatches a connection, and each check checks again before it signs anything, so a lapse stops the next read.

What the schedule’s reads look like on your side

They are ordinary control-plane calls: signed, single-scope, single-use, sixty seconds, under the subject system. Each lands in your own audit trail like any other dashboard read, including the ones that found nothing wrong. Response bodies are not logged, cached or persisted; only the occurrence hashes are kept.

ESC