Skip to main content
By the end of this guide, each new production alert or error becomes one GitHub issue, investigated read-only and written up with a severity, evidence, a suspected cause and a proposed fix. A repeat of the same alert comments on that issue instead of opening another. There is no alert route or token to set up: on Reliant hosting the on-call trigger is declared webhook: {source: alerting}, so turning it on connects it to your organization’s alerts (Webhooks).

Before you begin

1

Put the repository's checkout in a Reliant project

The investigator reads the checkout at <project>/<repo name>, or the project folder itself, whose origin is the alert’s repository. Without one, the issue is filed but not investigated.
terminal
The project also scopes the alerts: those of the Forge project its forge.yaml names, and of environments deployed from repositories it holds.
2

Check the workflow's behaviour before it runs for real

On-call ships with scenarios that mock GitHub and the investigator, so you see what it files, comments on and skips without spending tokens:
terminal
All 22 pass. same_fingerprint_dedupes shows a re-sent alert getting a comment, and search_failure_is_not_treated_as_new that it files nothing when it cannot rule out a duplicate (scenario format).
3

Turn on-call on

In the app, open Workflows → Library → Autopilot · on-call and choose Set up. From the terminal, inside the project folder:
terminal
service_repos maps an alert’s service label to a repository; default_repo takes the rest, or leave it empty to ignore unmapped alerts. Other settings are in Autopilot on-call settings.
4

Confirm it is connected to your alerts

Turning it on creates an alert receiver named after the activation. From the Forge project:
terminal
It takes critical and error alerts, plus the error and release alerts at any severity, from every environment of this project that raises alerts (persistent ones Reliant hosts and self_managed ones such as an environment forge obs init registered), not previews. Your other receivers are unchanged.
5

Make an alert fire

Any real error will do. Here a deliberately broken function threw a TypeError on every page view; within minutes it was an issue with a firing NewIssue alert:
terminal
6

Check the issue it filed

Each notification is a firing of the activation. A new one starts a run under Workflows → Runs; a repeat within four hours starts nothing.
terminal
For a new alert or error, the run files an issue in the mapped repository and labels it agent:triage:
  • Title: [error] <service>: <error title> or [alert] <alertname>: <summary>, then the prefix plus the investigator’s one-line summary.
  • Body: severity, confidence, the key it deduplicates on (an error’s issue id or an alert’s fingerprint), then Summary, Evidence, Suspected cause and Proposed fix.
If the investigator ends without an answer, a stub says so and the run is marked failed.
7

Decide what happens next

The issue stays agent:triage until a person acts on it:
  • Have an agent fix it: replace agent:triage with agent:ready. If the issue worker is on for that repository, it writes a failing test and the fix, and opens a pull request for you to review and merge.
  • Fix it yourself, or close it as noise. A later alert with the same key opens a new issue only once this one is closed.
8

Turn it off when you are done

To stop on-call for this project, deactivate it in Workflows → Automations, or:
terminal
Without --project-id it pauses every project’s on-call. Pausing or deleting removes the alert receiver.

What this does not do

  • It does not fix or merge anything. On-call only files and comments; its investigator has no edit tools, no GitHub write tools and a shell with no credentials. Alert to fix is two linked steps with a person at each gate: a person labels the issue agent:ready, the issue worker opens a pull request, and a person merges it. Autopilot never adds agent:ready and never merges.
  • It does not read your logs or traces. For an error it gets the service, title, culprit, first build and a console link, works from those and the code, and says what it could not reach.
  • It does not see every alert. Other warnings, preview environments and silenced alerts never reach it. By default it investigates five new alerts per notification (max_investigations) and files the rest without a write-up.
  • It does not page anyone. It runs beside your Slack, PagerDuty or email receivers.

Troubleshooting

A built-in workflow is named with its scheme on the command line: --workflow builtin://autopilot-oncall.
No clone of that repository at <project>/<repo name> or the project root, or its origin points elsewhere. Clone it there.
No service_repos key matched and default_repo is empty. Run reliant trigger activate again with new --params-json.
obs status shows an alert firing but reliant trigger events lists nothing. Check the alert passes the filter (step 4) and that obs receivers list shows the autopilot-<activation id> receiver. If its env_kinds is persistent alone, it was created before self-managed environments were included: pause and resume the activation to recreate it with the current filter. Pausing and resuming the activation connects its receiver again. NewIssue fires once per issue, so test again with a new error.

Next steps

Autopilot: alerts and errors

Deduplication and what the investigator sees.

Go from a production error to a reviewed fix

From an on-call issue to a merged fix.

Dashboards and alerts

Default alert rules and error issues.

Webhooks

The source: alerting trigger, and self-hosted Alertmanager.