Update OpenClaw

Use for general requests to update an existing Claw/Gateway. Select the existing update workflow; do not create another deployer or scheduler. Release validation, publication, new installations, and unrelated hosts are outside this workflow.

Resolve the owner

Identify the host, runtime user, profile/state root, service, running executable/release, and update owner from operator configuration and live inspection. The current repository or shell CLI may belong to another install. Keep private access details and state out of public artifacts.

Read the owner runbook and choose the matching workflow:

Follow the selected workflow's commands, locks, backup/migration requirements, recovery, and cleanup. For the operator-configured Team server, only its designated Team Server · Night Watch owner executes live deployment, restart, cutover, rollback, and recovery through the canonical controller. Do not delegate those actions or designate an external executor through coordination/admission. Other workers may investigate, test, review, and land fixes, then hand results to Night Watch; an already-running external operation requires the linked Team skill's safe exact-transaction handback, never a duplicate controller or blind kill. The linked Team skill also owns notification timing and content; report non-action outcomes and blockers privately to the requesting user, not to Team. Do not apply standard CLI updates inside a separately managed release or running container, mutate files used by a live Gateway, or overwrite dirty checkouts.

“Latest” preserves the configured channel; explicitly requested “latest main” selects the development source flow (--channel dev for the standard CLI). Preview the target and report the actual serving SHA: development preflight can select an older buildable commit. Do not auto-accept unapproved downgrades or plugin capability changes.

Interruption and recovery

An update request includes routine restart and bounded interruption, subject to deployment policy. Eligible sessions resume automatically. Do not ask again solely because agents, replies, or tasks are active. Use the owner's graceful drain and recovery-preserving interruption.

Task cancellation, session aborts, and queue clearing can discard work; do not use them to manufacture idleness or manually replay recovered turns. PTYs do not survive. Restart recovery owns recovery behavior and limits.

If the owner lacks controlled interruption, report or repair that gap within scope. Do not invent force flags, bypass persistence/compatibility gates, or repeat the interruption approval question.

Follow through

Retain ownership of an explicit update request across busy deferral, failure, rollback, and recovery until the selected owner's full acceptance or a concrete external/access blocker requiring outside action. Respect user pauses and cancellations. Before ending with work pending, establish an active completion/observation path or supported continuation through that owner; return without waiting for another prompt. If no continuation path exists, state that blocker instead of promising one. An existing cadence or restored old process alone does not complete the request.

Keep deployment-specific recovery and target selection with the linked owner. Qualify suspected issues at the actual failing path, improve updater guidance for confirmed durable gaps, and repair confirmed defects within the authorized scope. Use isolated worktrees and subagents where useful; test and review repairs, then land them through normal PR review, CI, and scripts/pr landing. Continue through verified update acceptance after landing; persistence does not authorize bypassing locks, safeguards, or branch protection.

Verify

Use the owner's verification to prove the intended version/commit is serving, RPC/health and configured channels work, and UI/assets load through actual ingress when enabled. Inspect startup/recovery errors without exposing secrets. Read the native deployment result separately from an observer or wrapper exit status: a successful observer can report a deferred or failed update. A build, handoff acknowledgement, or healthy old process is not completion. Keep detailed receipts private; report the result, next action or exact blocker in one to three short, friendly lines, with a short SHA only when useful and no repeated logs or full hashes.