Updating OpenClaw without breaking what worked yesterday

Ask the community what breaks OpenClaw and the answer is updates. The upstream project has rebuilt its updater around that complaint: the current command validates the new version while the old Gateway is still serving, boots a canary on a copy of your state, and only then swaps. This is how it works, what still goes wrong, and how a managed host turns it into a button.

What the update command does now

It detects how you installed OpenClaw, whether npm, pnpm, Bun or git, checks that your enabled plugins are available before it stops anything, validates the candidate, then runs a canary boot on copied state with the background listeners suppressed. The stopped interval is only the swap, the required migrations and plugin convergence, and the final report records the downtime and the verification results.

• A version you already have is not reinstalled; the run finishes as skipped.
• A plugin that fails to update does not fail the core update; you get a named next command instead.
• Invalid config or state, ownership errors, and a Gateway that will not start still fail the update.

Back up first, and know what a backup is

Upstream is blunt about this: before a significant update, create a verified backup, because the automatic config copies and the migration recovery originals are not a full-state backup. If you run in Docker, replacing the image runs the same startup-safe migrations, and a Gateway that cannot repair its mounted state exits rather than pretending to be healthy.

When it still goes wrong

A failed update lands in built-in triage. In a terminal, OpenClaw asks before launching an agent to investigate, using your own account and tokens; in the Control UI, a failed attempt opens Ask OpenClaw with the recorded details. A lost connection or a verification timeout is reported as an unknown outcome, not a success, and the original failure stays the authoritative record.

Updates regularly break things that worked last week.

What a managed host changes

On Diali the update is a button. Each OpenClaw release runs on our own agents before it reaches yours, your assistant restarts briefly with its configuration, channels and keys carried across, and the dashboard keeps undo, pause, restart and reinstall as things you control. Hosted OpenClaw on Diali describes the rest. The trade is the same as everywhere else in hosting: you give up choosing the exact hour, and you get back the weekend.

The short version

• Back up, verified, before anything significant.
• Use the updater; it validates before it swaps.
• Treat an unknown outcome as unknown, not as done.
• If you would rather not own any of this, host it where updates are tested before they reach you.

That last point is the whole of the OpenClaw alternatives guide, if you want the longer argument.