AI failure explanations
A short root-cause summary and a suggested fix for a failed run, written by AI from its log.
Availability
When a pipeline run fails, Prodgator writes a short explanation of why it failed and suggests a fix. The explanation appears on the run's page and in failure notifications.
How failure explanations work
An explanation is written when a run changes to failed, if AI features are on for your organization. Each one costs 1 AI credit.
Data used
The explanation is built from:
- Failing job logs: the part of the job's log around the last error message.
- Commit information: for GitHub, a summary of what changed in the commit. Other providers get an explanation without the diff.
- Similar past failures: earlier explanations for the same workflow, so recurring causes are named as such.
Prodgator removes any secrets it recognizes from logs before sending them. See AI data and privacy.
Viewing explanations
Explanations show in three places:
- The run's page: the full explanation, with the root cause and the suggested fix.
- The Failures page: the full explanation in the failed run's expanded row.
- Failure notifications: the one-line headline is added to email and Slack notifications.
Where no explanation exists yet, Explain failure writes one. The button shows what it costs.
Caching and reuse
Prodgator reuses an explanation for the same failure pattern within 24 hours instead of writing a new one. If the same workflow fails with the same root cause twice in one day, the second failure gets the first explanation, at no cost. Regenerating always writes a new one.
Regenerating explanations
Failed runs show a Regenerate button. Clicking it asks the model again and replaces the explanation.
What explanations cannot do
AI explanations never:
- dismiss or acknowledge a failure,
- change your pipeline or settings,
- read data from the build beyond what the app shows.
Privacy and storage
Explanations are kept for as long as your plan keeps pipeline data, then deleted. Prodgator does not share explanations with outside services or use them to train models.
A run first gets its category (such as ) from failure classification. When an explanation is written while the run is still failed, the run takes the explanation's category, unless the explanation says and the run already has another one.