[dezween]

Problem

make stop reported "sysmon is not running" while a sysmon process kept running. Observed live: PID 70263, alive since Aug 19, 2:44 of accumulated CPU, with no /tmp/sysmon.pid present. It had to be killed by hand.

stop and status derived liveness solely from $(PIDFILE). The pidfile desyncs from reality in two ways:

1. Double make start — the second start overwrites the pidfile with its own PID, orphaning the first instance permanently.
2. make clean — deleted the pidfile without killing anything.

stop also ran rm -f $(PIDFILE) unconditionally, even when no kill happened.

Fix

Process liveness becomes the source of truth; the pidfile is demoted to a label that is trusted only when a live scan corroborates it. Makefile is the only file changed.

• PGREP helper — pgrep -x -U $(id -u) sysmon, the single liveness source for start/stop/status.
• stop — kills by pidfile when valid, falls back to the live scan when the pidfile is missing or stale, confirms termination (10 polls × 0.2s), and removes the pidfile only after a confirmed stop. A failed kill exits non-zero and leaves the pidfile in place.
• start — refuses with a non-zero exit when any live instance exists, instead of spawning the next orphan.
• clean — declares stop as a prerequisite, so it can never strand a running process.
• status — reports running-and-tracked, running-but-UNTRACKED, stale-pidfile, and not-running as distinct states, and warns when a tracked instance coexists with an orphan.

Match precision (important)

The obvious pkill -f sysmon would have been dangerous: on macOS, pgrep -f sysmon matches the system daemon /usr/libexec/sysmond (verified — PID 92862 on the dev machine). This PR uses -x (exact process name, so the trailing d excludes sysmond) and -U (current user only), and never uses -f or pkill.

Two related hardening points:

• The pidfile PID is honoured only when it is a member of the live-scan list — otherwise a recycled PID in a stale pidfile could send SIGTERM to an unrelated process, a worse bug than the one being fixed.
• stop kills every PID in the scan, not just the tracked one, so a tracked instance plus stray orphans are cleared in one pass.

Verification

Full manual sequence run on macOS 26 / arm64, all passing:

#	Step	Result
1	make start	started (PID 98220)
2	make status	running, tracked (PID 98220)
3	second make start	refused, exit≠0, still exactly one PID
4	make stop	stopping tracked PID → stopped
5	make status	not running, pidfile gone
6	start + rm pidfile	running, UNTRACKED (PID(s) 98367)
7	make stop ← the original bug	pidfile missing or stale -- stopping PID(s) found by live scan: 98367 → stopped
8	make status / pgrep	not running / empty
9	start + make clean	stop output precedes cleaned, nothing left

Edge cases also verified: stale pidfile with nothing running (cleared, exit 0); tracked instance coexisting with an orphan (status warns, stop clears both); sysmond never matched by the Makefile's scan.

Note on project state

This supersedes the recorded decision in .planning/STATE.md — "pidfile-based start/stop in Makefile retained as-is (not touched by this refactor)". The pidfile is retained, but it is no longer the sole liveness signal. STATE.md is updated accordingly.

🤖 Generated with Claude Code