Should I use Jev for log and CI failure triage?
Verdict for the sample idea
Run it on
Jev
A bounded judgment over text, made many times. A System One model fits.
Sample idea
When a CI build fails, read the last part of the build log and classify the failure as one of: code signing, dependency resolution, compile error, test failure or flaky infrastructure. The code then picks the retry policy. About 300 failed builds a week.
Explanation
Log triage has two halves, and only one of them is a Jev task.
The first half is labelling one event. What kind of failure is this? That is a judgment on text with a fixed list of answers, and the wording of errors varies too much for a regular expression to cover. Jev fits. The code reads the label and picks the action: retry once for flaky infrastructure, notify the author for a compile error, page nobody.
The second half is everything that needs numbers. How many errors in the last hour? Did the rate double? Did this start after the deploy at 14:05? That is counting and time comparison, and TypeSafe lists both as weaknesses. Code does this better, for free, and without error.
Shape the Jev part as a Choice with one option per cause. Write includes with real fragments from your own logs. "No signing certificate found" and "provisioning profile expired" both belong to code signing, and a model that has never seen your toolchain may not connect them. Add other, and review what lands there each week. That is how you find the sixth category.
Cut the state in code first. A build log has 50,000 lines, and the cause is in 30 of them. Send the first error and its context, not the tail of the file. Large irrelevant state makes the answer worse.
The common trap is to ask Jev for a severity from a raw log stream and to alert on it. Alerts need rates and thresholds, and those are arithmetic. Let Jev label, let code count the labels, and alert on the count.
300 builds a week is low volume. Jev still fits, because the alternative is a person who reads logs.
Related tasks
Verdict produced by Jev on 2026-09-19, rubric version 2. Text written by a person. How the tool works: the report.