mirror of
https://github.com/kevinveenbirkenbach/computer-playbook.git
synced 2026-11-10 20:56:39 +00:00
kevinveenbirkenbachandClaude Opus 5.5
3346049481
91f25a35f2drove the budget-kill drills through `timeout -k GRACE 0` and sent it SIGALRM at the deadline, relying on GNU timeout treating that signal as its own expiry: TERM to the stand-in's process group, KILL GRACE seconds later. The ubuntu-26.04 runner ships coreutils from uutils (`9.5-1ubuntu2+0.0.0~ubuntu25`), whose timeout does not handle SIGALRM that way, so the stand-in's `sleep 1000` outlived the driver, kept the captured pipes open and both drills hit their 120 s subprocess timeout (CI run 40869769501, test_window.py:191 and :204). Production code uses neither `kill -ALRM` nor `timeout … 0`; only the drill did. The driver now starts the stand-in as its own process group (`set -m`), keeps the armed FIFO from91f25a35f2, sends TERM to the group at the deadline and KILL after GRACE seconds if it is still alive, and exits 124 or 137 like GNU timeout did. It no longer depends on any timeout implementation; all eight drills pass. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>