Work completed 2026-10-15 · written up 2026-10-16.
Two test runs on the tool that scores a property test suite against a bug's before-fix and after-fix code were killed at the time limit after 10m09s each, and both were reported as exit 1 [refusal:1758] [refusal:1769]
When a command is killed for running too long but reported as a failure, the builder looks for a bug that isn't there
The final test check now reports a killed command as a timeout, and a project can give its final command more time [landed:1279]
The after-fix side's tests alone took 1126.1s, nearly twice the 600s limit [refusal:1774]
On this project the random-input testing ran long, not the clone or the compile, so a faster build would not have fixed it
The director ruled that the runner lowers its number of fuzz runs and leaves the results file out of the change; neither was done in that build
The alarm that watches design work used a job name that never runs on the web projects, and it counted plans that were already done [landed:1274] [landed:1275]
An alarm that counts finished work reports a stall that isn't happening
The alarm no longer counts finished plans
The code review alarms for Swift, Kotlin, Python, Ruby, Rust, Go and PHP could fire for any open plan that had moved past review, however long ago that review ran [landed:1277] [landed:1278]
An alarm should look for the event it watches, not for how far a plan has got
The seven alarms now fire only on a review that ran within the period they check