proc, a Free Process Tool, Now Has an Agent Skill That Stops Claude Code and Codex From Piping lsof Into kill -9 on Processes They Didn't Start

Port 3000 is busy, so the agent does what a tired human does at 1 a.m.: lsof, grab a PID, kill -9. The pipe never checks whether that PID was this session’s dev server or the one you’ve had running since Tuesday.

The proc agent skill asks. It installs with npx skills add yazeed/proc-cli (or as a Claude Code plugin from the same repo), and it tells the agent to “only touch processes it started or you named,” to preview with --dry-run, and to ask before reaching for sudo. Every command takes --json and prints one document, so the agent reads a structured answer instead of scraping columns.

The skill landed in a changelog post its maintainer, who goes by yazeed on GitHub, published October 25. “The last post here was in March, when proc was at 1.8,” it opens. Nineteen releases later it’s at 1.14.5, and the skill is there to teach agents “when to reach for proc instead of piping lsof into kill -9,” and “to leave alone the processes they didn’t start.”

The fix list reads like a tour of how process tools lie to you.

On Linux, “proc used to list each thread of a process as if it were a separate process,” so one busy server looked like a crowd. That was fixed in 1.14.4.

On macOS, “processes owned by other users used to report an uptime of about 56 years.” Now “they now show as unknown,” which is more honest than pretending a process started around 1970.

The Linux binaries are static now. “They used to need a recent glibc, so they failed on Debian 12 and older.” They “now run anywhere from Debian 10 and Ubuntu 20.04 to Alpine.”

And the one every agent user should read twice: “proc in lists processes still running in a directory you’ve since removed,” the post says, “such as an old git worktree, which is where forgotten servers tend to live.”

The tool itself is the rest of the pitch. proc on :3000 says what’s on a port. proc why “traces the chain of parent processes behind whatever is listening, with how long each one has been running.” proc free stops the process and waits until the port is actually released before it calls the job done. It runs on macOS, Linux and Windows, from GitHub, crates.io, npm or Docker Hub.

What nobody has measured yet is how many agents actually load the skill before they reach for kill -9.

proc puts the safety in the verb. Nothing gets sandboxed; the agent just gets a command that does less. Expect more small Unix tools to ship a skill file alongside the binary, written to change what the model reaches for, and to compete on how well they say “don’t.” The open question is whether agents follow the polite instruction when the port is busy and the task is due.