Run a bounded job

A background command that hangs never exits, so the notification that waits for its exit never arrives. scripts/run_job.py runs a command in its own process group. It stops the whole group:

• at the deadline,
• when the helper receives SIGINT, SIGTERM, or SIGHUP,
• when the command exits and leaves other processes running in its group.

Run a finite job with a deadline well past its normal duration:

python3 <skill-directory>/scripts/run_job.py --timeout 1800 -- make test

Run a long-lived process, such as a dev server, with --no-timeout. The helper still exits when the process exits or crashes:

python3 <skill-directory>/scripts/run_job.py --no-timeout -- npm run dev

Pass the command as separate arguments. For pipes, redirections, or several commands, use sh -c '...'. Run the work in the command's foreground. If a launcher detaches its worker and exits, the helper stops the worker as a leftover process. Keep the project's test locks, worker limits, and execution wrappers.

The helper needs Python 3.9 or later. It inherits the current directory and environment, reads stdin from /dev/null, and passes stdout and stderr through. Its own status lines go to stderr and start with run-job:.

A command can exit with one of these codes itself. The helper prints a run-job: status line with codes 124 through 127 and a usage: run_job.py line with code 2.

Bound a watcher and confirm its first check

A watcher waits for one outcome, such as a ready line or a CI result. Give it a deadline:

python3 <skill-directory>/scripts/run_job.py --timeout 5400 -- gh pr checks 42 --watch

• Make the watcher exit on every outcome: success, failure, an error from the check itself, and the watched process ending. If the watched process runs through its own run_job.py call, that call's exit notification reports a crash. Stop the watcher when it arrives.
• Make it print its first check before its first sleep. Keep the check's stderr in the watcher's output, since a check can print an error and still exit 0.
• Before you end the turn, read that first check in the output and confirm it ran without errors.

Processes the helper cannot stop

A process that calls setsid leaves the group, and the helper does not stop it. The helper cannot handle SIGKILL, so killing the helper that way leaves the group running.