On an eventNeeds a repository

Fix a bug from an alert

A PR with checks

Your alerting posts to a webhook trigger; an agent reproduces the failure, fixes it and opens a pull request with the checks passing.

View .md

Best for

On-call teams with alerts that point at code
Failures a test can reproduce

Uses

A connected GitHub repository
A Project
An automation with a webhook trigger

Automation · Fix from alert

POST /o/{org}/p/{project}/automations/{automation}/triggers

promptAn alert fired; its payload follows. Reproduce the failure with a test, fix it, and open a pull request. If you cannot reproduce it, say what you tried.

{
  "kind": "webhook"
}

Fig. The automation: its prompt, and the trigger that starts a run. Field names are the API's.

Watch it work, then read what comes back

GET /o/{org}/sessions/{session}/events

event: thinkingThe stack trace points at retry.ts:88.

event: tool_useedit · retry.test.ts (reproduces it)

event: tool_result1 failed

event: tool_useedit · retry.ts

event: tool_result12 passed

event: messageOpened a pull request with the fix and the test.

Fig. The live stream while it works. Event names are the stream's.

GET /o/{org}/sessions/{session}

status: completed
branch: rush/fix-retry-timeout
pr_url: github.com/acme/api/pull/…
total_cost_usd: 0.88

Fig. The session when it ends. Field names are the API's.

Steps

01
Add a webhook trigger
Create an automation with the prompt above and a webhook trigger. Rush shows its URL and signing secret once.
02
Point your alerting at it
Send the alert's payload to the trigger URL. The payload reaches the agent as untrusted data after your prompt, never as instructions.
03
Fire a test alert
A run starts per alert. Duplicate deliveries are deduplicated, and a busy or rate-limited automation reports why it skipped.
04
Review the pull request
The run opens one pull request on its own branch. When the checks do not pass, the run ends needs_review, never a false done.