[teamleaderleo]

When acpmux.service restarts, an acpmux command that runs while the old daemon is still stopping no longer starts its own detached daemon. It waits for the unit's next daemon instead.

The managed-unit wait (from #35174) asked systemctl is-active and fell back to a detached start on any non-zero answer. A unit that is deactivating gives a non-zero answer, and a stop takes as long as the old daemon takes to end its agents (17 s in the case below). A daemon started then runs in the calling command's cgroup, takes the lock as the old one exits, and the unit's own daemon waits behind it.

On a fleet lane host on 11-06, a review box's check (a scope with TasksMax=256 and CPUQuota=400%) hit that window at 21:34:34Z. The detached daemon served the host until 22:28. Its agent hosts and background runners never moved to their own scopes, because only a daemon in a service unit moves them. They piled into the review box's 256-task scope (pids.peak 256, 198 rejected forks), so harness starts (spawn harness process /usr/bin/systemd-run: Resource temporarily unavailable (os error 11)) and bg_start runner spawns failed. Peers also read the host as 4 CPUs, because available_parallelism honors the scope's CPU quota. The host has 32 CPUs.

The wait now reads ActiveState. It keeps waiting while the unit is active, activating, deactivating, reloading or refreshing, and takes the detached path only for inactive, failed or a probe that doesn't answer. Stops and starts are already bounded by the unit's own timeouts.

Tests: managed_unit_still_stopping_keeps_the_wait (a deactivating unit keeps the wait until its daemon serves) and managed_unit_gone_takes_the_detached_path (inactive and failed still fall back), with fake systemctl scripts like the ones the managed-unit tests already use. cargo fmt --check and cargo test -p acpmux --lib managed_unit passed on the Linux pool at 0312ccd, six runs (three parallel, three with --test-threads=1), 4/4 each. One earlier parallel run on a cold worker failed the existing managed_unit_start_avoids_a_symlinked_default_home once (its systemctl start never landed, so it took the detached path). It passed in all six reruns, and that test doesn't reach the changed probe on its success path.

Changelog

none

Agent session: agent session

🤖 Generated with Claude Code

[github-actions]

cmux status _{(one comment, edited in place; each line is updated by the job that owns it)}

Review 0312ccd: HQ feed card · gallery before/after runs · checks

App of this head (app.zip, kept 1 day)

gh workflow run reload-build.yml -R manaflow-ai/cmux --ref feat-cmux-next -f tag=pr37323 -f ref=fm3h--acpmux-unit-stopping-no-detached -f platform=macos -f backend_mode=local

[teamleaderleo]

Owned by an agent session · session

[teamleaderleo]

Review (codex, router-review, started by aid-8707001042ea634e8771): SHIP at 0312ccd

Reviewer: codex gpt-6-astra high, fast tier.

SHIP at 0312ccd

Visual verdict: no images; Leo tries the build

I found no correctness issues. Focused tests could not run because cargo is unavailable in the review box.

Follow-ups: none

[cmux-bot]

cmux-bot approves 0312ccd: an independent review passed this exact head (router-review receipt by teamleaderleo). A push needs a new review; a later BLOCK at this head dismisses this approval.