OpenClaw merged a supervised Gateway reliability fix today in PR #146189: restarts now wait for lifecycle ownership when another OpenClaw process still holds the coordinator lock.
Before this change, a supervised Gateway restart could fail immediately if a predecessor process had not yet released lifecycle ownership. Repeated failures could burn through systemd's start limit and leave the Gateway down even though the old owner might have released the lock moments later.
The bug was reported in issue #146142, and the merged fix is aimed at that startup race.
What Changed
Gateway acquisition now waits for lifecycle-coordinator contention under a five-minute monotonic deadline. The retry loop backs off from 250 milliseconds to two seconds, logs the initial wait, logs successful acquisition, and includes elapsed wait time in terminal ownership errors.
Supervised startup also recognizes lifecycle contention and passes the same absolute deadline into acquisition. Time spent loading startup code counts against that budget. A healthy port cannot turn coordinator contention into a false successful startup.
OpenClaw's generated systemd units now allow ten starts per 300 seconds while keeping the five-second restart delay. The PR notes that launchd already has throttling behavior without the same finite start-limit latch.
User Impact
The visible behavior should be calmer during restart handoff. Instead of failing right away because another OpenClaw process still owns the lifecycle lock, the replacement Gateway waits for a bounded period and proceeds when the predecessor releases ownership.
This does not steal locks from live owners. A foreign process that never releases ownership still causes startup to fail. The coordinator also remains an anonymous SQLite lock; holder identity is documented as a follow-up design decision.
For operators, the important changes are:
- Restart handoff has a bounded wait instead of an immediate failure.
- Logs explain when OpenClaw is waiting for lifecycle ownership.
- systemd units get a larger burst allowance for restart attempts.
- No schema, dependency, configuration, or retained-state change is required.
Validation
The PR reports focused tests across lifecycle and role locks, migration leases, supervised startup, run-loop shutdown, restart sentinels, systemd units, and launchd plists.
It also includes Linux Testbox proof with real systemd restart behavior. In that test, an independent synthetic holder retained the coordinator, the replacement logged that it was waiting, ownership was acquired after a short delay, and the service became active with no failed-start result.
A separate published-driver test used openclaw@2026.9.3 to run gateway restart --preserve-definition --json against the candidate under the same injected contention. That path exited successfully and preserved the unit definition while waiting for ownership.
The takeaway from PR #146189 is practical: supervised OpenClaw Gateways should now ride out short lifecycle handoff contention instead of turning it into an avoidable outage.