pkill -f / pgrep -f kills or matches the calling shell itself (exit 143, command never continues)

问题

A one-liner like pkill -f app.mjs; node app.mjs & run through sh -c, ssh host '...', docker exec ... sh -c '...' or an agent's bash tool terminates immediately with exit 143 and the restart never happens.

报错原文（已脱敏/规范化）

(the sh -c / bash -lc command that contains `pkill -f app.mjs` exits with 143 Terminated before the following commands run)

适用环境

Any Linux/macOS; procps pkill/pgrep; command executed via a wrapper shell whose full command line contains the pattern

如何确认是这个问题

sh -c 'pkill -f app.mjs; echo survived' → expect: no survived printed and exit 143. With app[.]mjs the echo prints.

不是这个问题时 / 边界

If the shell survives but the process does not die, the pattern does not match the target's argv — check pgrep -af output instead.

根因

-f matches against the full command line of every process. The wrapper shell's own command line contains the literal pattern (sh -c 'pkill -f app.mjs; …'), so pkill sends SIGTERM to the shell that is running it.

修复步骤

1. Make the pattern not match its own text with a character class: pkill -f 'node /work/app[.]mjs' (the regex app[.]mjs matches app.mjs but the literal string app[.]mjs does not match itself).
2. Or anchor to the executable: pkill -f '^node /work/app.mjs$'.
3. Or record the PID when starting (node app.mjs & echo $! > /tmp/app.pid) and kill $(cat /tmp/app.pid) instead of pattern matching.
4. Or put the restart logic in a script file and run the file (its command line is sh /tmp/restart.sh, which does not contain the pattern).

如何验证

The one-liner exits 0 and pgrep -af 'node /work/app[.]mjs' shows exactly one new process.

试过没用的办法

Adding sleep; using pkill -x (matches process name only, misses node); setsid nohup (the shell is still killed first).

风险提示

Anchored patterns must match the exact argv used when starting the process.

验证范围

procps 3.x/4.x, 2026-10