# Rollback and recovery > Downgrading, automatic schema-neutral rollback, verified pre-update backups, and triage when an update leaves you stuck - 网址：https://funcoding.ai/agents/openclaw/install/updating/rollback-and-recovery/ - 来源：OpenClaw 官方文档原文（英文），MIT 许可，同步于 2026-10-11 - 官方原文：https://docs.openclaw.ai/zh-CN/install/updating/rollback-and-recovery --- Downgrades, automatic rollback, verified pre-update backups, and triage when an update leaves you stuck. Part of the [Updating](https://funcoding.ai/agents/openclaw/install/updating/) guide. ## Before you upgrade - Create a [verified backup](#before-updating-create-a-verified-backup) and keep the same service account, profile, and state/configuration paths. - Compare `openclaw --version` with `openclaw status` and `openclaw gateway status --deep`. The first reports the CLI version; status reports a Gateway service targeting a different installation. In `openclaw status --json`, compare `update.root` with `gatewayService.layout.packageRootReal`; `gatewayService.layout.entrypointReal` identifies the service binary, and `gatewayService.installationDrift` explains a mismatch. - Check the actual Node executable and version with `node -p "process.execPath + ' ' + process.version"`. Compare them with the updater's detected runtime and required range, since the service can select a different Node than your shell. See [supported Node versions](https://funcoding.ai/agents/openclaw/install/node/). - Run `openclaw update --dry-run` and inspect any named preflight check. Restore service-manager access as the owning account when requested. An unavailable manager produces a warning and skips automatic service restart; an ownership refusal requires resolving the reported mismatch before retrying. - If installation paths differ, align entry points before considering a reinstall. On Windows, use `Get-Command openclaw -All` and `Get-Command node -All`; on macOS/Linux, use `type -a openclaw node`. Retry through the intended installation's absolute launcher and follow the updater's recovery command. Changing the shell's Node alone does not update the service definition. Preflight diagnostics identify the update installation and binary, the detected runtime, and the service installation when inspected. An unresolved service path means it could not be verified, not that no Gateway is running. Older installed updaters keep their original messages for the first upgrade hop. ## Downgrade Verify the upgrade and your session history before retiring recovery originals with `openclaw update cleanup`. Downgrading the package does not reverse config or database migrations. Once state has migrated beyond the older release's supported format, the supported recovery is to restore a verified pre-update backup with its matching OpenClaw release.
Fresh shared-state databases created after tenant-container management was retired omit `fleet_cells`. Older revisions that require this table can reject them even when the numeric schema version matches. Existing databases retain the old table and rows, but that alone does not establish downgrade compatibility. Do not reconstruct the retired table or alter schema markers to bypass validation. Continue with the current release, or restore a complete verified backup with its matching older release using the recovery procedure below.
Prefer `openclaw update` for upgrades and recovery. It validates the target, runs required Doctor migrations, and verifies the activated Gateway. A raw `npm i -g` replacement does not retain the previous package or run this recovery workflow; use `openclaw update` or [create a backup first](#before-updating-create-a-verified-backup). The updater retains the previous package during activation and keeps it when failed recovery cannot prove a working installation. Migration recovery originals and original-state update captures remain until explicit [update cleanup](https://funcoding.ai/agents/openclaw/cli/update/#update-cleanup). These are separate recovery mechanisms: cleanup does not manage package or Git runtime backups, and retained migration originals are not a full pre-update backup. Preserve every recovery location named in the update report until you have verified the installation. Launcher backups compare the link type and target, plus ownership when it can be preserved. Symlink permission bits do not block an update; macOS link modes are copied when supported. Regular-file launchers still require matching modes and contents. If backup verification fails, the report names the differing fields and the retained failed copy for inspection before retrying. Package rollback checks file hashes, inode identity, permissions, ownership, and symlink targets. It ignores regular `node_modules/.package-lock.json` files, which npm documents as a disposable cache, and link counts and change times that can change without altering the retained bytes. Executable `.bin` entries remain part of verification. During the original update process, a mismatch records up to five differing relative paths and field names (including `sha256` for changed contents) in the failure facts and report. Recovery after a process restart retains the saved digest check but cannot reconstruct that entry list. This behavior belongs to the installed updater. An older updater, including 2026.9.4, can refuse package or launcher verification before the target version runs and cannot gain these diagnostics from the candidate. Use the installation's [manual package-manager update procedure](https://funcoding.ai/agents/openclaw/install/updating/update-methods/#alternative-manual-npm-pnpm-or-bun) if that first update is blocked. For a target that can read the current state, preview and use the managed rollback path: ```bash openclaw update --tag --dry-run openclaw update --tag ``` The updater checks compatibility and asks for downgrade confirmation. If the saved channel is `extended-stable`, add `--channel stable` for an exact one-off tag. Supported targets finalize the config writer stamp, restart the service, and verify the running version. Older targets may lack that finalization or migration-continuation contract; follow the printed recovery guidance if activation is refused. Do not bypass a newer-schema or newer-config refusal. When the update report identifies retained originals, use the corresponding [Doctor recovery command](https://funcoding.ai/agents/openclaw/cli/doctor/#session-sqlite-migration) before cleanup. Restoring legacy session artifacts does not reverse SQLite schemas or restore sessions created only in SQLite. If the older release cannot read the current state, restore the pre-update backup using [Restore a full archive](https://funcoding.ai/agents/openclaw/install/backups/#restore-a-full-archive). Keep the Gateway and other writers stopped throughout activation of the restored state, and preserve the current state separately first: restoration discards changes made since the backup. Reinstall the matching package through the installation's package manager; a backup archive does not contain the package. A complete recovery point must cover these together: - The matching OpenClaw package version or source revision and built runtime. - `openclaw.json`, including `meta.lastTouchedVersion`. - `state/openclaw.sqlite` and every `agents//agent/openclaw-agent.sqlite`, including databases at configured paths outside the default layout. - The workspaces, credentials, and retained originals needed by that installation. Restore that set from one backup generation. Do not combine a pre-update config or shared database with a post-update per-agent database, or the reverse. Model auth profiles and their state are authoritative in each `openclaw-agent.sqlite`. Files named `auth-profiles.json.sqlite-import..bak` or `auth-state.json.sqlite-import..bak` are preserved migration inputs or recovery artifacts, not the live credential store. Restoring those JSON files without the matching per-agent database does not restore that generation's auth state. Use `openclaw backup create --verify` for a verified, WAL-aware archive. Never copy only the main `.sqlite` file from a live WAL database: committed data can still be in `-wal`. Restore the verified consolidated database offline; do not mix it with `-wal` or `-shm` files from another database generation. See [Backup](https://funcoding.ai/agents/openclaw/cli/backup/) for archive coverage and omissions. Versions with the [startup preflight repair](https://github.com/openclaw/openclaw/pull/141451) leave configuration, databases, and migration inputs unchanged when preflight refuses startup. A successful start can migrate state forward. An older binary may then refuse both the database schema and the config's `meta.lastTouchedVersion`; changing either version marker does not undo the migration. Repair the installed version with `openclaw doctor --fix --non-interactive`, or use the backup recovery above. During recovery, prevent an enabled [auto-updater](https://funcoding.ai/agents/openclaw/install/updating/automatic-updates/#auto-updater) from immediately reapplying the newer release by setting `OPENCLAW_NO_AUTO_UPDATE=1` in the Gateway environment. After recovery, verify the running installation before cleanup: ```bash openclaw --version openclaw health openclaw gateway status --deep --json openclaw doctor --lint --json openclaw update cleanup --dry-run ``` ### Full-state recovery requires a backup Package updates keep pre-migration SQLite snapshots alongside the package backup until verified successful activation removes them with that backup. Rollback, failed or unverified completion, and refused restoration retain them. If cleanup cannot finish, the update reports a maintenance warning with the retained path. Older snapshot directories are not automatically removed: their ownership and successful outcome cannot be proven from existing receipts. Doctor reports older npm snapshot directories with their size and removal command. Confirm no update is in progress and inspect the corresponding update report and recovery state before manual cleanup. If the Gateway was confirmed stopped during capture, a failed candidate that was never allowed to start can restore those databases before package rollback only when database write fingerprints remain unchanged through Doctor and restoration. Maintenance ownership cannot identify independent SQLite writers, so any change during Doctor, including Doctor's own writes or a newly created database, makes these snapshots available for manual recovery only. A change between capture and Doctor admission, during Doctor, or after Doctor finishes preserves the current databases and reports `state-migrated-no-rollback` with the snapshot location and Doctor recovery guidance. Without Doctor write evidence, rollback requires the last verified database generations to remain unchanged. Snapshot capture first settles local SQLite writers under maintenance ownership, so later writer shutdown is not mistaken for intervening writes. The installed updater owns this ordering; staging a newer candidate cannot change an already-running older updater. On Windows, an eligible capture first runs the same native SQLite exclusion check used by rollback. This settles any retained WAL before recording write fingerprints, so later check cleanup is not mistaken for another writer. If another connection prevents exclusion, snapshots remain available for manual recovery. Snapshots taken while a Gateway may still be writing are available for manual recovery only until verified successful activation, even if it exits later. Migrated files are kept as `.migrated-` for inspection, and the report names the snapshots and displaced files. See [Recovery limits](https://funcoding.ai/agents/openclaw/cli/update/how-updates-run/#recovery-limits) for disk requirements and the lifecycle checks. Database restoration preserves current update history, including failure details recorded after capture. The retained original snapshots remain unchanged. This requires the repaired updater to drive the update; already-running older drivers cannot gain database rollback from the candidate. A candidate that may have served keeps the existing refusal to avoid discarding newer writes. Replacing its package alone cannot reverse migration. Use a verified pre-update backup with its matching release for an intentional downgrade. The updater does not create or replay a full-state checkpoint. An existing pending checkpoint-recovery record blocks further mutable updates. The updater reports that it is unsupported and leaves its records, backups, and state unchanged. Do not remove or alter retained artifacts to force a clean status, and do not use `update finalize` to bypass the refusal. Preserve the reported locations for a compatible recovery implementation or an independent verified backup. An interrupted or refused restore is not a successful rollback. ### Automatic schema-neutral rollback If a newly activated package fails verification, `openclaw update` compares the shared and affected per-agent SQLite `user_version` values with their pre-activation values and checks that the config file still matches the content reported by the new version’s activation Doctor writer. Databases first created during activation or verification are schema-neutral when their version matches the new version's supported version for that database kind. A changed schema version or missing pre-existing database, or a new database at a foreign version, still blocks rollback. Before restoring code, the updater also checks that the previous package supports any new database; unknown or incompatible support refuses rollback with `rollback-state-unverified`. When both checks pass and the retained previous package was verified before the update, it stops the new version and restores the previous generation: package, command shim, service definition, and exact pre-activation config bytes, including the previous writer stamp. Config replacements use owner-only permissions (`0600`); unchanged config needs no write. Owned, writable service metadata is refreshed; protected service definitions are preserved. The CLI verifies the restarted previous Gateway's service health, version/build identity, plugins, channels, and `/readyz` again. Update verification does not use model inference: the managed service must be running and own its port, and the Gateway hello handshake must match the expected artifact. The new version’s Doctor migrations in the main config file and its `$include` files do not block rollback, including on a fresh install’s first update. The updater retains the config immediately before Doctor and verifies that Doctor consumed those captured bytes before making changes. It also checks each current file against the output hash reported by Doctor’s writer. Rollback restores the original bytes only while both hashes match. Restoration holds the normal config writer locks and rechecks the files after acquiring them. Operator edits made after activation block restoration, including edits before Doctor reads the config and between Doctor’s last write and the updater’s capture. Changed include paths also block restoration. Older updater handoffs that retain only the root file still require includes to remain unchanged. The existing intentional-recovery allowance applies only to service commands, so the older-binary guard does not block recovery; it is never saved in config or the service environment. Missing or malformed includes do not prevent Doctor from running. When the updater cannot capture the complete include graph, it warns that automatic config rollback is unavailable and leaves any Doctor repairs in place if the update later fails. Successful recovery leaves the previous Gateway running and finishes the run as `rolled-back`, with `after.version` set to the previous version and downtime measured from service stop through verified recovery. The headline is `↩️ OpenClaw update rolled back to : `, retaining the original verification failure. The command still exits nonzero; recovery does not turn a rejected version into a successful update. Recovery reports distinguish restored package files from a healthy Gateway. A verified rollback names the version serving after recovery, including when an additional repair was needed. If the restored service fails its health check, the result records `recovery.service: "failed"`; the report says health failed and includes the recorded recovery reason. Health is reported as unverified only when verification could not run or complete, such as a readiness timeout. Both outcomes direct you to `openclaw gateway status --deep` to check the serving version and readiness. Rollback uses the same startup allowance as the update's activation check. Restart notifications retain the recovery fields understood by the restored runtime. Detailed recovery reasons remain in the update result, status diagnostics, and failure report. Use `openclaw update status` for the recorded reason and `openclaw triage` to diagnose a failed check. Recovery guidance reports whether the Gateway is running or stopped from the latest service observation, even when the new version is running but did not pass verification. A restored Gateway must pass its own verification checks before the run can finish as `rolled-back`. In the Control UI, open **Settings → Updates** and choose **Diagnose update** to ask OpenClaw to investigate the recorded failure. Loading the dashboard, reconnecting, or receiving an update failure does not start diagnosis. Each diagnostic request requires a button press; it does not retry the update. Automatic CLI triage never follows a verified rollback; it runs only when the update ends failed. In an interactive terminal, you can choose **Diagnose update failure**, **Report update failure**, or **Exit**, which is selected by default. Reporting shows the sanitized preview and requires separate confirmation before issue creation. Skipping or cancelling does not start diagnosis or submit a report. JSON, `--yes`, non-interactive, and managed-service handoff invocations do not show this menu after rollback. If the config file changed after the activation Doctor pass or the databases remain incompatible after eligible pre-start restoration, rollback is refused with `state-migrated-no-rollback`. For config edits, the next action names the file whose changes blocked restoration. The updater preserves the failed outcome and migrated state. Optional [post-failure triage](https://funcoding.ai/agents/openclaw/install/updating/#unattended-repair-on-your-own-inference) can run after update ownership and service compensation settle, including after failed rollback. Use the printed diagnostics and installation-specific repair command before considering an older version. Triage does not rewrite that failed update as successful. On macOS and Linux, if Doctor times out after migration, the updater stops its tracked process groups and waits for each group to disappear before attempting recovery. Once settlement is proven, it can start the installed candidate on the preserved migrated state and verify Gateway health. The report retains the Doctor failure, records a maintenance warning, and recommends `openclaw update repair`; recovery does not claim that unfinished Doctor repairs completed. If any writer remains or cannot be accounted for, the report names the known PIDs and keeps the Gateway stopped because concurrent writes put data at risk. Later update attempts retain this block even if the original updater has exited. Preserve the recovery snapshots and follow the reported process-inspection guidance before retrying repair. The installed updater owns this process supervision: a first update driven by 2026.9.5 remains limited by that older parent's settlement checks. Before starting a preserved candidate, the updater checks shared and existing agent database schemas against the candidate's contract. If migrations remain pending after the failed Doctor has stopped, it runs the candidate Doctor again under Doctor's maintenance and backup safeguards, then checks completion before restarting the Gateway. A refused or incomplete repair stays in the update report with the reason and the commands `openclaw doctor --fix` followed by `openclaw gateway start`. Newer database schemas are never downgraded by this repair. The original update failure remains recorded even when recovery restores service. Its durable restart notice is delivered when the Gateway starts again. Automatic rollback restores code and captured config, and restores pre-migration database snapshots only when the Gateway was confirmed stopped during capture and the candidate was never allowed to start, with matching write fingerprints through Doctor and rollback. The temporary snapshots used to check migrations are removed after validation and do not replace your backup. If the schema comparison cannot be completed, automatic rollback is refused (`rollback-state-unverified`). The newly installed version owns final verification and reporting after migration, preserving the same run ID and recorded activation steps. For pnpm and Bun, changes to sibling global packages after staging refuse automatic rollback (`rollback-project-changed`) without restoring the shared project; keep a reachable version installed, otherwise keep the Gateway stopped and follow the report’s repair command. A refusal before the live swap restarts the unchanged Gateway and preserves the sibling changes. ### Before updating: create a verified backup `openclaw update` preserves an automatic pre-update config copy, not a full-state recovery point. Before a significant update, create an independent verified backup explicitly: ```bash mkdir -p ~/Backups/openclaw openclaw backup create --output ~/Backups/openclaw --verify ``` The archive manifest records the OpenClaw version and the source paths included in the backup. The archive can contain credentials, auth profiles, and channel state, so store it with owner-only permissions and the same protection as the live state directory. See [Backup](https://funcoding.ai/agents/openclaw/cli/backup/) for included and intentionally omitted files. For a byte-for-byte recovery point that includes volatile artifacts omitted by the portable archive, stop the Gateway and use a filesystem, volume, or VM snapshot provided by your platform. This matters for older file-backed installs: the portable archive omits matching JSONL transcripts and logs even when they are no longer being written. When migrating large legacy histories, leave room for the original files, a temporary SQLite spool, and the destination database/WAL simultaneously. SQLite can be larger than the original JSONL; streaming import does not imply a fixed RAM requirement or migration time. Check free space on both the system temporary volume and the state volume. See [Session SQLite migration](https://funcoding.ai/agents/openclaw/cli/doctor/#session-sqlite-migration) for staging and memory details. ## If you are stuck Run `openclaw triage` in a terminal on the Gateway host, using the printed installation-specific command or keeping the same profile and state/config overrides. It opens the first directly launchable coding agent in this order: Claude Code, Codex, OpenCode, then Pi. The agent receives local diagnostics and any recorded failed-update outcome so it can repair the installation and verify Gateway health, using its normal authentication, sandbox, and approval settings. Use `openclaw triage --agent codex` to select a particular agent. Failed interactive updates offer triage after updater cleanup and pass the captured failure to the agent before fresh diagnostics can delay the handoff. Before launch, OpenClaw shows the agent, saved prompt path when available, and use of your own account/tokens. Only an affirmative Yes proceeds. Enter, `n`, cancellation, or 30 seconds without an answer skips the launch, prints a manual recovery command, and preserves diagnostics and the failed update's exit status. Explicit `openclaw triage` does not ask for this confirmation. JSON, `--yes`, and non-interactive update invocations can start one owned automatic repair after an eligible failure; other failures retain diagnostics and handoff commands. For diagnostic collection alone, use `openclaw triage --non-interactive`; add `--update-result ` to include a saved update-failure artifact. See [Triage](https://funcoding.ai/agents/openclaw/cli/triage/) for command formatting and installation targeting. Triage keeps the failed update's report intact. An update started during repair creates its own history entry. After package replacement, restart commands run from the updated installation. A restart accepted by the service owner can still fail readiness checks; inspect `openclaw gateway status --deep` before retrying. Keep a stopped, unverified Gateway stopped and preserve migrated state during repair. A reachable version retained after a schema migration can continue serving while you diagnose it. The failed update retains its nonzero exit code even if the agent repairs it. - For `openclaw update --channel dev` on source checkouts, the updater auto-bootstraps `pnpm` when needed. If you see a pnpm/corepack bootstrap error, install `pnpm` manually (or re-enable `corepack`) and rerun the update. - Check: [Troubleshooting](https://funcoding.ai/agents/openclaw/gateway/troubleshooting/) - Ask in Discord: [https://discord.gg/clawd](https://discord.gg/clawd) ### Unattended repair on your own inference Updates, validation, verification, and rollback do not require inference or model authentication. An unavailable model route cannot block those operations. Validation failures discard the staged candidate while the old Gateway keeps serving. After activation, the updater first completes its existing compatibility-checked rollback and service-compensation flow. Eligible failed updates can then start one owned triage repair after their update ownership has been released. Triage targets the installation that remains, preserves migrated state, and keeps the original failed update's outcome and nonzero exit code. Successful repair does not retrospectively publish a successful update or verified rollback. Reports from older updaters can still contain a `repairing` phase and its attempt summaries. Published 2026.9.4 updaters may invoke the candidate's repair-worker entry before the update settles. New candidates retain its response protocol but report inference repair unavailable without loading a model or changing operator state. The published driver still owns that first update's control flow and budgets. For an explicit repair using configured inference, run `openclaw triage --run` in a terminal on the Gateway host. Triage runs Doctor health checks, attempts up to one embedded repair turn with time and tool-call limits, and runs Doctor again. It uses the normal runtime credential resolver, including inherited profiles and OAuth refresh; unavailable inference produces an external handoff rather than a login prompt. The repair keeps the existing installation scope and tool-policy restrictions. A saved activation or recovery failure additionally requires recorded updater completion and current installation and Gateway verification. An attributed Doctor/config blocker can be resolved by fresh Doctor checks and the recorded installed identity, while its historical failed run remains unchanged. Native service activation and recovery retain their existing authority checks. See [Triage](https://funcoding.ai/agents/openclaw/cli/triage/#installation-target-and-embedded-handoff) for installation targeting, repair limits, and validation results.