[kid7st]

running(pid) decided whether a ticket's step is alive with kill(pid, 0), which also succeeds for a zombie: a process that has exited and that its parent has not reaped.

When it matters

A step is started detached by the command that starts it, which then exits, so the step's parent becomes PID 1. Where PID 1 does not reap orphans (a Docker container without --init, a coding agent's sandbox that runs its shell as PID 1) a step that has exited stays a zombie for good. Then:

• --check never sees the step gone, so its ticket stays working or need_user_action until its deadline instead of failing with "it stopped";
• stop (a stop on request, a newer attempt, a start given up on) waits its full 5 s for a process that is already dead.

Measured in a Linux container whose PID 1 was sleep: after the step's group was sent SIGTERM, the step was in state Z and kill(pid, 0) still answered. CI's runner reaps, so it does not fail there. macOS's launchd reaps an orphan at once, so a step is a zombie there only for an instant.

The fix

Where /proc says what a process is (Linux), a process kill(pid, 0) answers for is taken for stopped when its state is Z or X, or when it is gone since. The state is read after the last ) of /proc/<pid>/stat, since the command name before it is in parentheses and may hold spaces and parentheses. macOS is not asked, for the reason above, and the comment at running says so.

Found in review: a /proc that is not this process's

A process in a PID namespace of its own can be left the host's /proc (unshare --pid --fork without --mount-proc, and sandboxes set up that way). There /proc/<pid> is not the process kill(pid, 0) names: measured in a Linux container, a live child (pid 8) answered kill(8, 0) while /proc/8/stat was ENOENT, and /proc/self/stat said 18 for a process whose process.pid was 1. The first version of this PR took that ENOENT for gone, so a running step was reported stopped by --check, and stop, a newer attempt and join all passed it by (stop returns at once for a process that is not running, so it was never signalled).

/proc is now read only when /proc/self/stat begins with process.pid (procfsIsOurs); otherwise running is kill(pid, 0) alone, as before this PR. Measured in the same container, a live step's ticket under unshare --pid --fork: failed before this fix, need_user_action after; in the same namespace need_user_action both times, and a zombie's ticket still failed.

Tests

A step whose process is a zombie (sh starts a child that exits, then becomes a sleep that never waits for it), named x) S y so its name holds both, is failed by checkTicket with "it stopped". It runs where /proc/self/stat exists (CI, Linux), and is skipped on macOS. On Linux it failed before the fix (need_user_action). procfsIsOurs is tested on the stat of this process, of another namespace's, and of a pid that begins with this one's digits (12 for 1); mutations dropping the space after the pid, or trusting any /proc, fail it. Mutations of the zombie check, each failing it on Linux: zombies taken for running again, the state read by splitting on spaces (survived at first, with a sleep named child; the name now holds a space and a parenthesis), the state read after the first ). The whole suite passed in a Linux container (Node 26.13.0) and on macOS.