Skip to main content

On-Call

On-Call is Autoheal's native paging system. It covers who is on duty right now, who gets woken when an alert fires, and how that page escalates until someone acknowledges it. It replaces a standalone paging tool, and because it runs inside Autoheal, a page arrives with the investigation already underway.

On-Call is a separate add-on module. If you don't see On-Call in the sidebar, it isn't enabled for your organization, so contact support@autoheal.ai.

Everything below lives under On-Call in the sidebar, which has five tabs: My On-Call, Active Pages, All Schedules, All Escalation Rules, and Status Pages.

The four building blocks​

ObjectWhat it is
TeamA group of people who own services and share schedules. Members have roles. Teams and people are the same entities as in the Context Graph.
ScheduleWho is on-call for a team, over time. Owns a timezone.
RotationA layer within a schedule: an ordered list of people and a shift length.
Escalation ruleWhich alerts page, whom to page, and what happens when nobody answers.

A schedule answers who is on duty. An escalation rule answers who gets paged and in what order. They are separate on purpose, so one escalation rule can page a schedule, a specific person, or a Slack channel.

Schedules and rotations​

A schedule is built from one or more rotations, each a layer with its own participants and shift length.

  • Participants are ordered. The rotation cycles through them in order, one shift each.
  • Shift length is arbitrary: a week, a day, or twelve hours.
  • Layers stack by priority. Each rotation has a layer order, and the lowest-numbered active layer wins at any instant. Higher-numbered layers act as fallback coverage. Layer 0 is the top layer.

This is how you express "primary and secondary": a weekday-business-hours rotation on layer 0, and a broader always-on rotation on layer 1 to catch nights and weekends.

Each schedule carries a timezone, so shift boundaries land at sensible local hours for the people on it.

Overrides​

An override replaces whoever the rotation would have picked, for a fixed window such as a doctor's appointment or a swapped weekend. You give it a person, a start and end time, and optionally a reason. Overrides win over rotations, and the timeline shows whether the current on-call came from a rotation or an override.

Previewing before you commit​

Rotation changes are easy to get wrong, so you can preview a timeline before saving. Autoheal renders the resulting shift-by-shift schedule from a proposed configuration. It does not persist anything. You can then check that the handoffs land where you expect.

Escalation rules​

An escalation rule has two halves: which alerts it applies to (matchers), and who it pages (nodes).

Matchers decide what pages​

Matchers are label conditions evaluated against the alert's own labels. Each matcher is a label, an operator, and a value. The operators are =, !=, =~, and !~. The last two are regex operators.

Matchers are grouped as OR-of-AND: within a group every matcher must match, and the rule fires if any group matches. So "severity is critical AND service is checkout, OR team is payments" is one rule with two groups.

Two behaviors worth internalizing:

  • A rule with no matchers never pages. There is deliberately no catch-all. Not every alert should wake someone, so an unmatched alert creates no page. It does not go to a default. An escalation rule with an empty matcher set is inert until you add one.
  • First match wins. When several rules match, Autoheal tries them in their configured order. The first rule that matches and has at least one delivery target creates the page. Autoheal skips a matching rule with no targets, so that rule does not silently swallow the page.

Because matching runs against the alert's own labels, you can route on any label without it being part of your alert grouping configuration.

Nodes decide who gets paged​

A rule's nodes are its escalation levels, in order. Each level has:

  • Targets: a person, a schedule (which resolves to whoever is on-call at that moment), or a Slack channel. A level can have several.
  • A timeout: how long to wait for an acknowledgement before moving to the next level.

Autoheal pages a person by Slack direct message, SMS or voice call. A Slack channel target posts to that channel. Email and mobile push are not used for pages.

You can also set a repeat count to cycle through the levels again if the page is still unanswered. You can also set working hours (timezone, days, start and end) so a rule behaves differently outside business hours.

Pages​

A page is one paging event. It appears under Active Pages and moves through four states:

StatusMeaning
TriggeredEscalation is running. Nobody has acknowledged yet.
AcknowledgedSomeone took it. Escalation stops.
ResolvedTerminal. The underlying issue is done.
Timed outEscalation ran through every level (and repeats) with no acknowledgement.

Pages are either triggered by an alert or created manually, and carry an urgency of high or low.

Three actions are available on a page:

  • Acknowledge: stop the escalation. You have the page. Autoheal records who acknowledged it and when.
  • Resolve: close the page. This action is always available on a page that isn't already resolved, so you can still clear a timed-out page from the queue.
  • Re-page: restart escalation on a page whose escalation is no longer running. This is the recovery path for a page that timed out because nobody was reachable.

A page detail view shows every notification attempt:

  • The level it came from.
  • The target.
  • The channel.
  • The person it resolved to, for schedule targets.
  • Whether delivery succeeded, failed, or was skipped.

When a page escalates through a schedule, this is how you see who was actually reached.

When new alerts join an existing issue​

If more alerts arrive for something already paged, Autoheal does not create a duplicate page. If the existing page was acknowledged or timed out, Autoheal re-escalates it. If escalation is still running, Autoheal does nothing. A resolved page is terminal, so a new occurrence gets a new page.

Paging from alerts is fully deterministic. It runs before, and independently of, any AI triage, so the matcher evaluation above is the whole decision, with no model in the loop.

Paging from an agent​

An agent can also page people, with two system tools. list_escalation_paths lists the escalation rules in the agent's project. create_oncall_page creates a page on one of them, with a title, a description and an urgency. The page is the same as a page you create by hand. It shows as created manually, it escalates through the rule's levels, and you acknowledge or resolve it in the same way. An agent can page only through an escalation rule in its own project. An agent's page does not change which rule an alert pages.

When On-Call is on for your organization, an agent with no system_tools filter can call create_oncall_page. A page from an agent does not wait for approval. To stop an agent from paging, leave the tool out, for example system_tools: name != "create_oncall_page". To let an agent page but take no other write action, write system_tools: operation_type == "read" || name == "create_oncall_page". In the agent's instructions, say when it may page. See Build Your Own Agents for how tool filters work.

My On-Call​

My On-Call is the personal view: whether you are on-call right now, which schedules you are on, and a calendar of your upcoming shifts. It's the tab to check before going into a weekend.

Shift handoff notifications​

Each schedule can notify people when a shift changes hands. Handoff notifications are off by default. You configure them per schedule:

  • A Slack post to a channel you choose, announcing the handoff.
  • SMS to both the outgoing and incoming person.

Notifications fire when the rotation actually advances in real time, not when you edit a schedule. Editing a rotation does not spuriously announce a handoff.

Importing from PagerDuty​

If you already run on-call in PagerDuty, you can import schedules and escalation policies through the PagerDuty integration. You do not need to rebuild them by hand. Imported objects keep a link back to their source, so you can tell an imported schedule from one authored in Autoheal.

Status pages​

Where the rest of On-Call is about telling your team that something is wrong, a status page is about telling your users. It's a public page that shows the current health of your services, active incidents in real time, past incident history, and an email subscription. It's the page you link to from your app, docs, or support channels.

Autoheal status pages follow the pattern your users already know from Atlassian Statuspage. A page lists services. Incidents describe what's happening and which services are affected. Each service shows an at-a-glance status.

The pieces​

  • Status page: one public page (e.g. "Acme US"). You can have several, such as one per region or product. Each has its own public web address.
  • Index page: one public page that lists several of your status pages on a single address (e.g. status.yourcompany.com listing Acme US and Acme EU), each with its current status and services. It has no services or incidents of its own; visitors open a listed page for details and to subscribe.
  • Services: the components shown on the page (e.g. API, Dashboard, Webhooks). Each has a status: Operational, Degraded performance, Partial outage, Major outage, or Under maintenance.
  • Incidents: a post with a lifecycle (Investigating → Identified → Monitoring → Resolved) and a running timeline of updates. An incident names the services it affects and how badly. Their status updates automatically.
  • Updates: a free-form single message for anything that isn't a lifecycle incident: a heads-up, a scheduled-maintenance note, or an announcement. It has no status and no effect on service health.
  • Subscribers: end users who opt in by email (double opt-in) to be notified when you publish or update a post.

How service status is decided​

You don't set a service's overall status by hand. Autoheal derives it from your open incidents: a service shows the worst impact among all unresolved incidents that list it. When you resolve the last incident that affects a service, the service returns to Operational automatically.

Public vs. private​

Every page has a single visibility switch:

  • Private (default): the page exists and you can build it, but it is not served publicly.
  • Public: the page is live at its web address for anyone to see.

Flipping a page to Public is how you publish it. Posts are live the moment you create them. There is no separate draft step, so a private page lets you prepare everything before anyone can see it.

A public page can also be served on your own subdomain (e.g. status.yourcompany.com) instead of the default address.

What your users see​

A published page shows two sections: Active incidents (unresolved, with their latest updates) and Recent updates (resolved incidents and update posts from the last 15 days, newest first). Visitors can subscribe by email to be notified of new and updated posts.

See Status Pages for creating a page, posting incidents, custom domains, and how subscriptions work.

Next steps​