Summary

Running openclaw update from an in-band OpenClaw chat/tool session on a macOS LaunchAgent-managed gateway can stop the same gateway that is carrying the update command, dropping the session before the updater can supervise completion.

This is related to the broader durable-update handoff issues on Linux/systemd, but this report is the macOS/manual invocation variant: a user/agent starts openclaw update from inside the live OpenClaw runtime, the command reaches the gateway stop/restart phase, the webchat/tool session disconnects, and the update no longer has a durable owner.

Environment

• OpenClaw upgraded to 2026.5.20
• macOS, Apple Silicon
• Gateway managed by LaunchAgent: ai.openclaw.gateway
• Install/update path: global npm package
• Invocation path: openclaw update initiated from a live OpenClaw webchat/tool session

What happened

During the first update attempt, the attached updater got as far as the gateway stop/restart phase. The gateway received a clean SIGTERM and webchat disconnected with service-restart semantics:

20:02:30 NDT gateway received SIGTERM
20:02:30 NDT webchat disconnected code=1012 reason=service restart
20:03:47 NDT gateway process present again
20:03:51 NDT gateway ready

A second restart window showed the same shape:

20:11:53 NDT gateway received SIGTERM
20:11:53 NDT webchat disconnected code=1012 reason=service restart
20:12:33 NDT gateway ready

A later supervised/manual package-manager path succeeded cleanly:

20:15:46 NDT final restart after manual install/plugin sync
20:15:57 NDT gateway ready

The key failure was not that the new release was broken. Once the update was run out-of-band, the system was healthy:

• core OpenClaw 2026.5.20
• gateway 2026.5.20
• LaunchAgent loaded/running
• dashboard reachable
• BlueBubbles OK
• Codex OAuth present
• postflight healthy

Why this matters

For an operator using OpenClaw itself as the control surface, openclaw update is self-referential: it can stop the gateway/session transport that owns the command. A plain background child is not enough either; in our local experiment, a quick background-to-log attempt only wrote started and then died, because it was still owned by the dying process tree/session context.

The update needs a durable supervisor boundary before it stops/restarts the gateway.

Expected behavior

When openclaw update is invoked from a live OpenClaw/gateway-owned session on macOS LaunchAgent installs, it should either:

1. hand off to a LaunchAgent/launchd-owned updater before stopping the gateway, then write a durable result file/log that survives the gateway restart, or
2. detect that it is running from an in-band gateway/chat/tool context and refuse with a clear warning, e.g. "Run this from an external shell or submit an out-of-band update job."

Actual behavior

The update can proceed into gateway stop/restart while the command is still supervised by the live OpenClaw session. The chat/tool session drops, and the operator has to manually restart/check gateway state afterward.

Workaround that succeeded locally

We replaced the in-band update with an out-of-band wrapper:

1. preflight backup
2. submit a launchd-owned worker
3. worker runs npm install -g openclaw@<target>
4. plugin/peer sync
5. exactly one openclaw gateway restart
6. wait for gateway/version match
7. run postflight
8. write a timestamped log and JSON result file for the recovered chat session to read

This worked in dry/no-op validation and the direct package-manager/manual flow completed the 2026.5.20 update cleanly.

Related issues

• [Bug]: auto-update can never succeed under systemd — updater is spawned as a child of the gateway it needs to restart #83360: systemd auto-update spawned as a child of the gateway it needs to restart
• Update button no-op on systemd-supervised installs: handoff helper killed by unit cgroup before it runs #84068: Control UI update handoff helper killed by systemd cgroup
• [Bug]: managed macOS update handoff can fail with uv_cwd after LaunchAgent stop #83808: macOS managed update handoff can fail after LaunchAgent stop via uv_cwd
• Gateway restart can leave Codex custom tool call without output #84727: restart can abort active Codex custom tool calls and leave transcript state invalid

This issue is specifically about the macOS LaunchAgent/manual openclaw update path when invoked from inside the live OpenClaw chat/tool runtime.

Suggested acceptance criteria

• macOS LaunchAgent repro where openclaw update is invoked from an OpenClaw webchat/tool session and the update still completes with a durable result after gateway restart, or refuses before stopping the gateway.
• A detached updater must be owned outside the dying gateway/session process tree. For macOS, that likely means launchd/LaunchAgent ownership rather than a plain background child.
• The result should be visible after recovery through a status command, result file, or update status surface.
• Existing systemd handoff fixes should not regress while adding the macOS LaunchAgent/manual path.