On this page
On this page
Release notes
v2026.11.6
OpenClaw v2026.11.6
AI agents and tools can read these release notes as plain Markdown.
OpenClaw 2026.11.6 is a patch release with bug fixes and quality-of-life improvements throughout OpenClaw, from installation and updates to messaging, model providers, plugins and the native apps.
Release scale: 386 pull requests, 48 direct commits, and 157 contributors.
Installation and Onboarding
- Fixes missing-file errors in new shells after uninstalling OpenClaw state by removing matching stale completion registrations. #186080, #112625. Thanks @MoerAI, @obviyus, and @kangarooo.
- Fixes state cleanup removing a workspace selected for preservation when its path uses a directory alias. Reinstalling a plugin through an alias also no longer removes the newly installed copy as an old replacement. 29b9961. Thanks @obviyus and @RomneyDa.
- Fixes fresh macOS and Linux installs silently choosing stable when you requested beta and npm was not yet installed. If beta cannot be found after setup, the installer warns before falling back to latest. Windows PowerShell installs also recheck the selected package's Node requirement after provisioning. #188005, issue #187982. Thanks @steipete and @mcgatom, who reported the issue.
- Fixes the Linux installer continuing after it fails to set up a user-local npm prefix. It now stops with an error before changing shell profiles or installing the package. #188292. Thanks @McgAtom and @steipete.
- Fixes private npm and Git installations failing their version check or producing unusable launchers when their paths contain shell characters. Quote paths passed to --prefix or --git-dir, and rerun the installer to regenerate an existing launcher. #189541. Thanks @mcgatom and @obviyus.
- Fixes workspace preparation deleting custom onboarding instructions before they are used when a workspace already has profiles or skills. #196411. Thanks @Patrick-Erichsen.
- Fixes local non-interactive onboarding leaving legacy credentials unavailable after you configure authentication. It imports existing profiles first and archives the original files. Commit 6e17e50. Thanks @RomneyDa.
- Fixes onboarding memory imports failing when a compatible Gateway is already running. A configured onboarding rerun also connects its hatch to that Gateway. Full migration discovery still requires the Gateway to be stopped. Commit d7e9ff0. Thanks @sallyom and @RomneyDa.
- Fixes first-time systemd unit installation being refused when local drop-in files already exist. Existing drop-ins remain untouched. 67c6124. Thanks @holgergruenhagen, @sallyom, and @RomneyDa.
Web UI
- Fixes the New Session project picker failing when project data changes while it loads. OpenClaw retries the listing and, if it keeps changing, asks you to try again. #187451, #186844. Thanks @RomneyDa. Thanks @rtfcv for reporting the issue.
- Fixes unnecessary usage-history rebuilds after pricing changes that could leave the Usage page slow or showing partial totals. Sessions with recorded costs can keep cached totals, while those needing estimates still rebuild when prices change. #188117. Thanks @obviyus.
- Chat uploads can finish sending without waiting for the sandbox workspace copy. If that copy fails later, the run reports the error and the saved attachment remains in the conversation. #188728. Thanks @steipete.
- Fixes media captions in webchat being replaced by ClawHub recommendation text when a reply has no separate message text. 471cbb0. Thanks @RomneyDa.
- Fixes false plugin configuration errors that could appear after you save a provider key in the Control UI. #189192. Thanks @obviyus.
- Board metadata can now appear without waiting for unrelated database writes. It reflects a committed snapshot that a concurrent change may supersede. Commit 2d443cc. Thanks @romneyda and @steipete.
- Fixes file and media previews showing old content after a successful Control UI save in a local sandbox. Host-only sessions also read the current host file instead of a retained sandbox copy. #194957. Thanks @steipete.
- Fixes private completions appearing as Control UI activity or session messages during session preparation. 065049d. Thanks @RomneyDa and @fuller-stack-dev.
- Fixes ignored phone notifications causing false WebChat errors and unnecessary model fallback when a CLI-backed agent intentionally stays silent. #200741, issue #200710. Thanks @vaibhav375 and @obviyus, and thanks @dawidkulpa for reporting the issue.
- Fixes searches for sharing identities, model providers and new-session environments missing accented names that look the same but use different Unicode character forms. Commit 45748de. Thanks coauthor @RomneyDa.
- Fixes chat search missing phrases that span visible formatting, including in the loaded full text of a shortened message. Markdown source and link destinations remain searchable. Commit b3fec4b. Thanks @holgergruenhagen and @romneyda.
- Fixes earlier messages disappearing from an already loaded chat after compaction when the history can be reconnected. It also fixes a first jump to the top failing to load older messages. Commit 6aa9e6f. Thanks @vincentkoc and @RomneyDa.
- Fixes text sometimes missing from the chat composer when the clipboard changes during a native paste. Recovered text is outside the native undo stack. Commit d92cdb6. Thanks @RomneyDa.
- Fixes pasted text returning after you immediately delete it from the chat composer. Recovery instructions also tell the agent to respect your constraints, use recorded results and avoid repeating completed actions. Commit 06afb10. Thanks coauthor @RomneyDa and coauthor @vincentkoc.
- Fixes a child session approval request disappearing from its parent session’s sidebar badge when the parent has no attention of its own. Commit 59af80d. Thanks coauthor @RomneyDa.
- Fixes pending questions from a non-archived direct child session appearing only in the child chat. You can see and answer them from the parent chat without leaving your draft. Commit cfabc5c. Thanks @holgergruenhagen and coauthor @RomneyDa.
- Slow requests for conversation branches now return a retryable error after ten seconds, so clients can try again after preparation finishes. Commit a81b987. Thanks @romneyda.
- Fixes the iPad hardware-keyboard accessory covering Control UI chat composer controls. The reserved space remains even when the accessory is absent. 25276a3. Thanks to contributors @jacobtomlinson and @romneyda.
Updates and Maintenance
- Fixes agent configuration being saved but not applied, or background startup work being interrupted, on an isolated Gateway when an older, unrelated Gateway service uses a different database. New agents can appear without a recovery restart while scheduled work is active. #187966, issue #187937. Thanks @shakkernerd.
- On Linux, agent sessions can become available without a full database scan after a Gateway owner dies when recovery can verify matching database provenance. A successful checkpoint and clean close can also avoid a full scan on the next ordinary restart. #188409. Thanks @steipete for requesting the change.
- Reduces delays for eligible agents on Gateway restart when cleanly closed databases were queued behind slow unsealed databases. Cleanly closed databases are checked first, but still need to pass normal validation. #195291, issue #195269. Thanks @steipete and @609NFT.
- Fixes Windows Gateway launcher arguments ending in a backslash consuming the next argument and preventing service command inspection. #189813. Thanks @sebtardif.
- Stopping the Gateway during a long startup database check no longer has to wait for the full scan to finish. #188413. Thanks @steipete.
- Fixes Gateway stops and restarts that could hang on unfinished database snapshot reads or disrupt cleanup by retiring state workers too early. #188496. Thanks @steipete.
- Protects session completion and pending-input state during a slow Gateway stop or restart by allowing required cleanup to continue beyond the diagnostic grace period. The existing shutdown deadline still applies. #188920. Thanks @steipete.
- Fixes an orderly Gateway stop being reported as a failure when a retained state lease needs final cleanup. #200330. Thanks @steipete.
- Stopping the Gateway with SIGTERM or SIGINT while it starts now releases acquired state locks, so those locks do not block a replacement. #200504, issue #132317. Thanks @freeqaz.
- Gateway restart now releases agent databases and their leases after a supervisor cleanup failure has settled, while still reporting the error. 5f854d6. Thanks @romneyda.
- Multi-agent PostgreSQL Gateways retain fewer idle connections, reducing connection-limit failures that can prevent startup, history reads, or maintenance. #188614, issue #184027. Thanks @steipete. Thanks @etzelm for reporting and testing.
- Agent creation protects retained PostgreSQL stores from being replaced when their history cannot be verified. A failed first creation can also be retried without leaving a new incomplete local anchor. ee818d6. Thanks @RomneyDa.
- Fixes updates or repairs aborting plugin finalization after unrelated configuration edits. If the same setting conflicts or edits keep racing, OpenClaw rolls back the pending plugin work and suggests a repair. The plugin may remain unavailable until repaired. #189819, issue #189778. Thanks @steipete and @akifodabas.
- Fixes a Git checkout update or repair rejecting Doctor's configuration repair when pending plugin migration guidance changes, which could leave the Gateway stopped. #190665, issue #190658. Thanks @steipete and @saariuslystoned.
- Fixes a Gateway remaining stopped after Doctor runs when queued command-line maintenance competes with its restart. Doctor and the competing processes must both use the repaired runtime. #189246, issue #189225. Thanks @steipete and @jeffuli755.
- Fixes chat updates requested by configured owners failing when a channel plugin is waiting for a state migration. #190433, issue #178291. Thanks @steipete, @reldubnix, @farzy, and @dankit-dk.
- Doctor now uses held maintenance custody when checking requester policy during delegated update repairs, while continuing to reject revoked requesters. Commit bf3e4be. Thanks @RomneyDa.
- Fixes interrupted updates leaving repair or a later update blocked after Doctor's processes have exited on Linux and macOS. OpenClaw retains control when it cannot establish that the command is gone. Commit 02b3fe8. Thanks coauthor @RomneyDa.
- Fixes Gateway restart and Doctor restoration stopping their wait too early when a slow startup is still making progress. #189364, issue #161158. Thanks @steipete, @angelojiang11, @tosko4, and @hannesrudolph.
- Fixes an update with
--no-restartfailing and rolling back when a running Gateway restarts before the update finishes. A later installation check can still restart the Gateway after the update releases its lease. #200389. Thanks @obviyus. - Fixes failed-update recovery on Linux when a foreground Gateway prevents the installed systemd service from taking over. Recovery now distinguishes unverified readiness from a confirmed stopped Gateway. Commit fc47b8c. Thanks coauthor @steipete and coauthor @RomneyDa.
- Fixes a Linux systemd update stopping when the Gateway restarts during maintenance after an installation replacement. The update can continue when that restart is recorded and verified. Commit d97c2d8. Thanks @fuller-stack-dev and @romneyda.
- Git updates no longer build unnecessary TypeScript declarations by default. Set
OPENCLAW_RUN_NODE_SKIP_DTS_BUILD=0if you need them. If an npm-to-Git conversion was interrupted,openclaw update repaircan recover it using the same state and configuration. #189463. Thanks @vincentkoc. - Git updates now retry interrupted downloads and clones automatically. An older installed updater may still need a manual retry to reach this fix. #189818, with earlier work in #184452. Thanks @steipete. Thanks @duyufansh2008 for the earlier work.
- Fixes Git-source updates failing on older Git during the local-history check. It also restores local commit preservation, cached-target checks, repository diagnostics and immutable update preparation on older Git. The installed updater must already contain this fix to use the compatibility fallback. #194198. Thanks @steipete.
- Fixes a switch from a Git checkout to a package install rolling back when the old command wrapper cannot be removed. The update warns with the affected path. Remove a leftover wrapper if it appears earlier on
PATHand still launches the old checkout. A switch started by an older updater may still roll back. #200579. Thanks @fuller-stack-dev. - Fixes interrupted pre-update configuration snapshot writes damaging the previous rollback copy when the updater taking the snapshot contains this fix. #189946. Thanks @kaguranananana, @steipete and @syxscott.
- Fixes an update being blocked when temporary space or quota runs out during Memory Core inspection. Free temporary space and rerun the Doctor check named in the warning. #192699. Thanks @steipete and @nikolay-bratanov.
- Fixes update activation stopping when temporary space or quota runs out during session inspection. Free temporary snapshot space and run
openclaw doctor --fixafterward. Some older entries may remain unreadable until then. #196370. Thanks @fuller-stack-dev. - Fixes updates using a private Doctor rehearsal rolling back because intact archived transcripts were missing from the rehearsal copy. #200659. Thanks @steipete and @SimonAIbot.
- Updates using a private Doctor rehearsal can continue with a warning when an archived transcript file is already missing. A Doctor repair that reads the missing file will still fail. On Windows, rehearsal cleanup also protects live archived files’ read-only attributes. #200733. Thanks @steipete.
- Fixes large state inventories and backup results hitting an output limit that could prevent update recovery capture. If capture still fails, the update now warns clearly that no complete pre-update recovery copy exists. Commit dcc48a4. Thanks @ourcoms2018, @SunnyShu, @steipete and @RomneyDa.
- Fixes update planning and database inspection failing after repeated copy attempts when a busy SQLite database keeps changing. Recovery requires a complete WAL and shared-memory index with no rollback journal. Commit 50fde57. Thanks @romneyda and @vincentkoc.
- Fixes Gateways with separate state directories mixing their default logs. The default log moves into the selected state directory, so update external log collectors to follow the new path or use
openclaw logs. Existing log files are not moved, and explicitly configured log paths remain unchanged. #197978. Thanks @steipete. - Fixes silent file logging creating an otherwise absent state directory when settings are read or a log is emitted. Commit e349c8c. Thanks @RomneyDa.
- Fixes nonlogging commands and filtered log messages creating the default log directory just by checking logging settings. Commit a3be2ef. Thanks @RomneyDa and @vincentkoc.
- Fixes early CLI diagnostics and certain update previews creating log files in the profile being inspected or rejected when the original console output is available. Commit 60d9a07. Thanks @RomneyDa and @vincentkoc.
- Fixes Gateway maintenance refusals creating a log file in a rejected profile. The console message also reports whether a managed LaunchAgent was parked. Commit 9e18f81. Thanks @RomneyDa and @vincentkoc.
- Fixes a registry timeout during
openclaw update status --jsoncreating a log file during the status check. The timeout still appears in the result. Commit e2427d8. Thanks @RomneyDa and @vincentkoc. - Fixes Doctor and update migration collapsing intentionally separate API and Claude CLI or Gemini CLI routes for the same model, which could change billing or remove a fallback. If a previous migration already collapsed your configuration, restore a pre-update backup or re-add the intended CLI fallback. #193179, #124952, #191149. Thanks @obviyus.
- Fixes timed-out or cancelled CLI updates leaving package-manager processes running after the update step ends. #190674. Thanks @MertBasar0 for the earlier work in #180977.
- Reduces the chance that an interrupted update leaves a build lock behind or the Gateway stopped by giving cleanup more time to finish. #191181. Thanks @steipete for the fix and @fuller-stack-dev for the earlier launcher work in #188859.
- Fixes Gateway restart checks continuing to wait after an update is cancelled. Updates with large SQLite state also have more time for the new version to check the data before falling back to older checks. #193263. Thanks @steipete.
- Fixes some interrupted updates leaving the Gateway stopped during activation. If the updater is killed,
openclaw update repaircan also restart a Gateway whose stop was recorded. If Doctor is still running, wait for it to finish and retry. #193760. Thanks @steipete. - Fixes update cancellation being ignored during some database migrations. Commit 6f404580. Thanks @RomneyDa for coauthoring.
- Fixes timed-out update preparation leaving an update marked as running after its work has stopped. Expired preparation also no longer starts later update steps. Commit 14f4fb34. Thanks @fuller-stack-dev and coauthor @RomneyDa.
- Fixes Doctor refusing to migrate agent databases created by 2026.7.2-beta.6, which could prevent the Gateway from starting. #190676. Thanks @steipete for the fix and @katie-hemus for reporting #190666.
- Fixes an update from 2026.9.2 getting stuck in recovery and requiring repair when the Gateway and update process change state at the same time. #192140. Thanks @steipete.
- Fixes Doctor running out of memory while repairing large compressed SQLite conversation histories. If repair is interrupted, rerun it to continue the remaining work. #195459. Thanks @steipete.
- Fixes updates failing when credentials change after the database backup is captured. The update continues with current data, but automatic restoration of the older backup is unavailable if a later step fails. #200328. Thanks @RomneyDa.
- Fixes some Linux Control UI updates failing when the old Gateway writes to its database while shutting down. The update can continue with current data, but automatic restoration of a stale backup is unavailable if a later step fails. #200569. Thanks @steipete, @zzs12345-web, @damzaru and @Lifaguang.
- Fixes updates stopping during database migration after SQLite rebuilds a nonempty write-ahead log index without changing the database. Automatic restoration of the older backup is unavailable in this case. Commit 8567c081. Thanks @RomneyDa.
- Fixes updates from affected older versions missing shared and agent database schema checks. Commit 7c1b551d. Thanks @RomneyDa.
- Fixes offline database ports rejecting supported older agent databases and preserves recovery images during migration. If your cache is on a network filesystem or cannot be identified as local, set
XDG_CACHE_HOMEto a local filesystem. #190828. Thanks @galiniliev for requesting related portability work. - Updates now check PostgreSQL backup ownership before replacing the installation. If your installed updater lacks this check and an unmanaged Gateway still owns the database, stop that Gateway through its service owner before updating. #193190. Thanks @etzelm and @steipete.
- Fixes configuration reads and Doctor failing after some 2026.11.1 Workboard-only PostgreSQL ports. Workboard stays on PostgreSQL while shared state and other data remain on SQLite, and later saves and ports preserve those choices. #194648. Thanks @ylcn91 for reporting #194626.
- Fixes updates from 2026.11.3 failing when Workboard uses SQLite. It also fixes update failures involving older Workboard-only PostgreSQL settings. #197526. Thanks @daniel-lxs.
- Uses the configured database engine when finalizing migrated updates and publishing their results. Commit a0695db5. Thanks coauthors @etzelm, @steipete and @romneyda.
- Fixes macOS restart checks losing track of parent processes behind a root-owned intermediary. Gateway process checks also retain readable details and continue checking other processes when one cannot be identified. Commit e93ef03e. Thanks @RomneyDa for coauthoring.
- Adds a compatibility record for the updater in OpenClaw 2026.11.4, including its failure-recovery module, to support generating recovery bridges after an update replaces the installed files. #199956. Thanks @RomneyDa and @vincentkoc.
- Adds a compatibility record for the updater in OpenClaw 2026.9.37 to support generating recovery bridges after an update replaces the installed files. Its historical managed-service handoff check returns false, and a legacy pnpm check handles a missing retired generation only when the active installation is verified in the same global root. #200577. Thanks @RomneyDa, and @vincentkoc for coauthoring the release adaptation.
- OpenClaw delays automatic update and model catalog checks, retention cleanup, and optional startup maintenance when available process memory is low. You can still check for updates yourself, and scheduled jobs still run. Commit 5036764. Thanks @RomneyDa.
- Fixes update repair incorrectly reporting that another update owns the installation when its current and legacy paths point to the same database. #193431. Thanks @steipete and @svenmy76.
- Fixes update repair reporting success while a dead closing update lease continues to block managed updates. Repair now either clears an eligible lease or explains why it must be retained. If an older updater is affected, first update the package manually with your package manager to obtain this fix. #194503. Thanks @kuanjames.
- Fixes update repair refusing to recover an eligible interrupted update when its recovery record points to an older store. If recovery cannot safely proceed, repair reports the specific reason. #200532. Thanks @steipete for the fix and @garrietv for reporting and diagnosing the issue.
- Fixes updates and Doctor sometimes reporting completed maintenance as failed when another process briefly holds the state database. If the lock lasts longer, cleanup can continue in the background so later maintenance can proceed. #200742. Thanks @steipete for the fix and @Staruy for reporting the issue.
- Fixes update repair getting stuck after a cancelled update leaves an unused package preparation, so an eligible later update can proceed. Commit 7bf28a9. Thanks @romneyda and @sukhdeepjohar.
- Fixes updates removing files needed by an unfinished managed-service update simply because the files are old. Commit 791b01e. Thanks @RomneyDa.
- Adds compatibility records for the updaters in 2026.11.5-beta.2, beta.3 and beta.4. These records support generating bridges that help those older updaters find recovery code after replacing the package during an update. #197844, #199653, #200152. Thanks @steipete.
- Adds a compatibility record for the 2026.11.5 stable updater to support generating bridges for its recovery imports after a package replacement. #200662. Thanks @steipete.
- Fixes update failure reports dropping the first failed step when later recovery steps also fail, where the report size limit permits it to be retained. #198968. Thanks @steipete and @aribarambo.
- Fixes an obsolete update remaining scheduled after a development Git checkout is found to be current or ahead. The obsolete update is also cancelled if a notification fails. Commit ed28251. Thanks @RomneyDa.
- Fixes early update failures leaving an accepted run without its failure result or identifier. The error now includes the run ID, and readable update state records the failure. Commit a2747ba. Thanks @RomneyDa.
- Fixes extended-stable notices showing the publisher tooling version instead of the latest regular stable release when the notice was prepared. #199923. Thanks @RomneyDa and @vincentkoc.
- Fixes update checks rejecting eligible private plugin dependencies in the staged OpenClaw package when the update candidate contains this repair. #200595, issue #200573. Thanks @steipete and @desksk.
- Fixes update checks reporting an implicitly discovered global memory plugin as missing in the tested ordinary, linked and compatibility-skipped layouts. 5f8fef1. Thanks @RomneyDa.
- Fixes a validated downgrade rehearsal being blocked by valid settings written by a newer version when the candidate contains this repair. ae8d8e8. Thanks @RomneyDa.
- Fixes package updates continuing with later guarded changes after their update authority is withdrawn. Completed package or backup copies are also recorded so an authorized rollback can restore the previous installation. An in-flight publication can still finish. 5e1567c. Thanks @RomneyDa.
- Fixes Windows plugin and pnpm or Bun package updates failing on brief directory locks. Restoration of a native package also retries a temporarily held backup. 0953bb1. Thanks @romneyda.
- Fixes eligible Linux and macOS updates leaving the package path missing if the updater is interrupted at publication. This protection applies once the installed updater contains the fix, and unsupported layouts still use the previous update method. 8e08059. Thanks @fuller-stack-dev and @RomneyDa.
- Prevents a Gateway running as a Windows Scheduled Task from starting an update that could leave the Gateway stopped. Open a PowerShell or Command Prompt window yourself, outside the Gateway, and run the update command printed in the refusal. 87c5419. Thanks @RomneyDa.
- Fixes a managed Linux systemd Gateway failing to start after a package replacement when its supported older databases need migration. Activation now backs up and migrates that state before startup. deeed6c. Thanks @romneyda and @steipete.
- Fixes managed Gateway updates being refused despite a compatible service Node, or another compatible Node when the service can be refreshed. The installed updater must contain the fix. If a 2026.10.7 updater refuses before staging, update once with your package manager while preserving the installation owner, service account and state settings. 83cd1bf. Thanks @steipete and @RomneyDa.
- Fixes updates refusing to convert an interrupted legacy heartbeat monitor when the Gateway is confirmed stopped. Pending recovery notifications are preserved for Gateway startup. Commit 5680455. Thanks @RomneyDa.
- Fixes Doctor failing to report a legacy shared authentication store and recommend
openclaw doctor --fix. Commit 7741a7f. Thanks @RomneyDa. - Fixes agents becoming unavailable when a configuration reload interrupts startup. OpenClaw retries the affected startup work and, if repeated reloads exhaust the retries, advises restarting the Gateway instead of repairing a healthy database. #193049. Thanks @steipete.
- Fixes ordinary configuration edits silently replacing malformed UTF-8 and damaging recovery copies. OpenClaw now refuses the edit and preserves the affected source and backups. #194181. Thanks @alix-007 and contributor @obviyus.
- Gives larger Git backups more time to create, restore and verify, reducing the chance that local content operations time out solely because of their size. Deadlines remain bounded. e23c8fb. Thanks @romneyda and coauthor @wangmiao0668000666.
- Fixes Gateway startup failing after a volume remount broadens eligible private-store permissions. On Linux, it also fixes first-time private-store creation under a setgid parent. Paths that cannot be safely restored may still need attention from the storage owner. 4c6328f. Thanks @tssala23 and contributors @sallyom, @joshavant and @romneyda.
- Lets external update supervisors check whether a candidate can start against copied profile state before stopping the serving Gateway. The check covers isolated core startup and does not guarantee that the update or production services will be ready. cfe8722. Thanks to contributors @steipete and @romneyda.
Messaging
- Fixes messages from identifiable Telegram peer bots bypassing the configured bot-loop limit. Skipped turns clear their thinking status, and loop-protection drops appear in Gateway logs. #151924. Thanks @amittell and @obviyus.
- Fixes Discord presence greetings ignoring sender allowlists, so excluded members coming online no longer trigger agent turns or consume a greeting cooldown. #198331. Thanks @mmaps.
- Fixes Feishu card commands and confirmations using direct-message rules when the chat type cannot be verified. They now refuse to run until the lookup recovers. f4d3d5a. Thanks coauthor @RomneyDa.
- If the runtime confirms an active run did not use your message and its source is still valid, OpenClaw can queue it once as a follow-up in its original position. It does not replay messages whose acceptance is uncertain. #186619, #184909. Thanks @sukhdeepjohar and @steipete. Thanks @hannesrudolph for reporting the issue.
- Fixes queued messages silently disappearing when OpenClaw catches a refusal after their operator authorization is lost. The source now receives a failure and reason. #200213. Thanks @obviyus.
- Fixes channel messages losing a delivery retry when a stale queue owner is recovered without a delivery attempt. 99d6ceb. Thanks @RomneyDa.
- Fixes durable messages queued as followups being discarded as duplicates if preparation is abandoned or fails before the turn starts. Retries wait for the original attempt to decide whether it used the message. 56b00f6. Thanks @RomneyDa.
- Fixes Microsoft Teams outbound messages stalling during formatting when they contain many unfinished mention markers. #198351. Thanks @pgondhi987.
- Fixes a first chat message being rejected while its agent is starting. The message can wait for startup, while recognized cancellation commands such as /stop can still proceed. #200510. Thanks @RomneyDa.
- Fixes Discord thread replies silently dropping attachments supplied as base64 or data URLs while reporting that the message was sent. #196707, issue #196689. Thanks @obviyus. Thanks @abacha for reporting.
- Fixes Slack Socket Mode account stops or reloads waiting on a connection retry that can no longer finish. #188745. Thanks @steipete.
- Fixes WhatsApp reconnects and shutdowns waiting indefinitely for a stuck route write. #189709. Thanks @SebTardif for the superseded fix.
- Fixes queued Telegram replies, including subagent completion replies, aborting when an earlier turn ends before they are sent. #189457. Thanks @vincentkoc.
- Fixes newly queued messages being retried from a different account after the channel default changes. Older queued messages without a saved account cannot recover their original default. #191819. Thanks @vincentkoc.
- Recovers eligible final text replies interrupted during delivery by sending the saved reply once when the channel cannot determine whether it arrived. The original may already have arrived, so a duplicate is possible. Commit 9c20ca8. Thanks @romneyda for coauthoring.
- Discord attachment and sticker downloads no longer inherit broad browser access to private networks. If your CDN hostname resolves to a private address, add it explicitly to
browser.ssrfPolicy.allowedHostnames. #191458, with an earlier fix in #144259. Thanks @SebTardif for the earlier fix. - Fixes multiline slash commands losing line breaks and indentation in Slack, Feishu and bot-mentioned Zalo Personal messages. #190730, #185266. Thanks @masatohoshino for an earlier contribution.
- Fixes Discord user mentions inside code being replaced with display names before the agent receives them. #200053, #195585. Thanks @obviyus. Thanks @saariuslystoned for the report and prepared correction.
- Fixes sent messages changing literal newline escapes inside Markdown code into line breaks. Commit e2b5aa7. Thanks @RomneyDa and @Sedrak-Hovhannisyan.
- Fixes Slack channel and group thread follow-ups losing the original subject after earlier context leaves the conversation history. #190175. Thanks @joetam for an earlier contribution.
- Fixes Slack native streaming duplicating or losing final answers when progress and final delivery overlap. If Slack rejects a final answer because the stream is full, OpenClaw posts it normally and closes the stream. Oversized closing chunks no longer leave the stream running. #199849. Thanks @n1hility and @obviyus.
- Fixes media and link rules applying to unrelated conversations or failing to block the intended conversation when session prefixes are configured. #191016. Thanks @zhangguiping-xydt and @obviyus for the fix and @yetval for the earlier repair in #190008.
- Fixes accepted chat messages disappearing from history when their turn has not started and recovery runs. They remain available for later requests and retries, subject to history limits. Side questions also keep their visible context when sent to providers. #193381. Thanks @liuwqgit and @steipete.
- Fixes a mid-turn context overflow repeatedly retrying an oversized request until the turn fails. When trimming cannot reduce saved tool results further, OpenClaw can try compaction and retain accepted steering and completed tool work on eligible retries. #199636. Thanks @steipete.
- Fixes a follow-up message losing all prior context after an oversized turn when the previous assistant reply still fits. Commit 26bd958. Thanks coauthor @RomneyDa.
- Fixes Matrix recovery commands being unavailable when an unsafe crypto store blocks a pending migration. With the updated host and Matrix plugin, stop running Gateway and CLI processes, then use
openclaw matrix doctor inspectand recover the affected account withopenclaw matrix doctor recover --account <id> --accept-snapshot-rollback. Recovery requires a valid snapshot and explicit acceptance of rollback, which can lose keys newer than the snapshot. Runopenclaw update repairoropenclaw doctor --fixafter recovery before restarting the Gateway. #200805, issue #200790. Thanks @steipete, and thanks @nikolay-bratanov for reporting the issue. - Matrix now waits for its crypto backup enablement check to finish before closing the client. c317a25. Thanks @romneyda and @vincentkoc.
- Fixes Discord channel allowlists resolving non-Latin or mixed-script names to the wrong channel, and searches for non-Latin names returning unrelated channels. Use channel IDs for names made only of emoji. Commit 8851c5c. Thanks coauthor @RomneyDa.
- Fixes Teams accounts failing to start when the default account uses a custom webhook path and a named account uses
/api/messages. The default account can use that compatibility path again after the named account moves away on reload. #186709, #186708. Thanks @nxmxbbd and @steipete.
Memory
- Fixes concurrent sandbox memory flushes silently overwriting each other when they append to the same file in one Gateway process. #188872, issue #117741. Thanks @michielhdoteth for contributing to a superseded fix and @yetval for reporting.
- Doctor now keeps recoverable copies when merging legacy workspace memory into an existing MEMORY.md. If the merge fails before replacement, restore the archived legacy file before retrying. 0ea89fc. Thanks @romneyda, @syxscott, and @steipete.
- Fixes Knowledge store creation and note and attachment publishing on affected Windows installations, where the first attachment could fail. Keep complete store backups because a completed save does not guarantee survival after a power failure. #189088, issue #189087. Thanks @HermanZeng.
- Fixes Knowledge calls failing after the first turn in a warm Claude Code session. #193687. Thanks @VACInc.
- Knowledge transcript imports now reject malformed UTF-8 and unparseable timestamps instead of accepting damaged input. Commit 95a2c35. Thanks @RomneyDa.
- Fixes a Knowledge scan recording a false deletion and extra revision when a page is published at the same time. The affected read now asks you to retry. Commit 0108136. Thanks @RomneyDa and @vincentkoc.
- Fixes broad agent searches stalling on large conversation histories. #200355. Thanks @steipete.
- Fixes a cold conversation search getting stuck while checking access to results. It also preserves device revocation for retained in-process continuations from tokenless trusted-proxy sessions. bae0ea9. Thanks @RomneyDa and @vincentkoc.
- Fixes ready memory turns remaining uncommitted when older unresolved turns in other sessions filled the retry queue. Turns still proceed in order within each session. #190165, #135628. Thanks @Leon-SK668 for earlier work and @abacha for reporting.
- Fixes a memory-wiki synthesis refresh resetting its status to active when no new status is supplied. 6b5afec. Thanks @RomneyDa.
Skills
- Fixes direct skill updates silently replacing a referenced API key with plaintext. Requests to replace a reference with plaintext now fail. You can still clear the key, replace it with another reference, or deliberately edit the full configuration. #200539, issue #193149. Thanks @obviyus and @tommyjoseph.
- ClawHub downloads now block redirects and archive URLs to other private network origins. If your self-hosted registry serves archives from a different private origin, move them to the configured registry origin. #190939. Thanks @pgondhi987.
- Fixes failed skill packaging corrupting an existing archive or leaving an incomplete archive at the destination. #194488. Thanks @zhangguiping-xydt and @obviyus.
- Fixes skill packaging silently leaving out resources when an included directory cannot be read. Packaging now reports the failure and preserves any existing archive. #194635. Thanks @zhangguiping-xydt and @obviyus.
- Fixes malformed meme-maker commands treating the next option as a missing value, which could overwrite a file instead of reporting an error. #196933. Thanks @zhangguiping-xydt for the fix and @obviyus for the runtime proof.
- Fixes Sherpa TTS commands silently using an unintended output or resource path when an explicit path value is empty or missing. Malformed values now stop the command before synthesis. #197103. Thanks @obviyus for the fix and @shuisadev for reporting the issue.
Native Apps
- Fixes iOS Talk aborting slower agent turns after 30 seconds when the Gateway allows them more time. #163474, #162784. Thanks @olli0103 and @obviyus.
- Fixes iOS Talk giving up on replies after 120 seconds while the Gateway is still working. It also recovers completed replies from history when their run and message identifiers differ. If Talk stays on ‘thinking’ after a missed completion, Stop ends the local wait, not the Gateway run. Check the conversation before retrying. #173035. Thanks @ben-milanko and @obviyus.
- Fixes iOS Talk history sometimes saving an assistant reply before the spoken question that prompted it, including a consult preamble. #197136, #197129. Thanks @marvinthebored, @peetiegonzalez, and @vyctorbrzezowski.
- Fixes xAI Talk disconnecting when you interrupt an answer and a replacement starts before the interrupted response finishes. #190308, #184207. Thanks @Marvinthebored for an earlier proposal.
- Keeps later consults eligible to run after a voice change in a channel-owned call. Commit 37b5018. Thanks @romneyda.
- Fixes Talk consults being rejected when a machine-selected thinking preference is unsupported by the chosen model. Commit a2ee26b. Thanks @sallyom and @RomneyDa.
- Fixes iOS reminder responses omitting notes. Agents can now read notes when listing reminders and see saved notes in add responses. #167665, #165668. Thanks @Anurag-5ingh, @obviyus, and @brianmcguire, and thanks @lucasprocoio for reporting the issue.
Models and Providers
- Fixes Codex tool screenshots missing from replies and from Control UI history after a reload. #154125, #153949. Thanks @sobechestnut-dev and @obviyus.
- Fixes Codex final-answer recovery when a conversation contains a media-only message without a caption. #185186, #185079. Thanks @kaguranananana, @obviyus, @uhdpass, and @hellastechgroup.
- Fixes provider deadline errors showing misleading sign-in guidance instead of a timeout message. #169188, #169186. Thanks @xpeng5278-web and @obviyus.
- Fixes an unexpected model switch when a local session read fails because the session changed during the read. Check the conversation for saved work before retrying. #188080, #188064. Thanks @shakkernerd.
- Fixes Claude CLI appearing to succeed without Gateway tools when the OpenClaw MCP server fails to connect. The turn now reports a failure, and retrying starts a fresh process. #176154, #176138. Thanks @xyi74976-del and @obviyus. Thanks @Oldrich333 for reporting the issue.
- Fixes native Claude CLI conversations failing to resume when only the OpenClaw tool-loading hint changes. abb3431. Thanks @RomneyDa and @boeingchoco.
- Fixes completed answers from long, quiet native Codex turns being discarded when the turn finishes. #185180, #185121. Thanks @stronggenetics and @steipete.
- Fixes some local model endpoints receiving hosted watchdog timing when their loopback address is represented as IPv6. 947301f. Thanks @RomneyDa.
- Fixes agent runs waiting indefinitely when provider authentication stalls. The run now stops at its deadline or when cancelled, with an error that identifies the authentication step. #189144. Thanks @amittell.
- Fixes cancelled provider response reads waiting for a timeout when the response stalls. #189147, issue #188252. Thanks @vincentkoc.
- Fixes a slow local Codex app-server launch freezing unrelated Gateway requests. Other requests can continue while the app-server starts. #192543. Thanks @steipete.
- Fixes model requests being sent to another server when a different provider’s transport hook changes the request origin. Same-origin changes, the provider’s own routing and explicit configuration still work. #189423, issue #185272. Thanks @obviyus. Thanks @vishnusenthil-16 for reporting the incident.
- Fixes Codex-backed session assessments and Activity recaps failing when model discovery takes just over five seconds. #189512. Thanks @Marvinthebored, @Peetiegonzalez and @obviyus.
- Fixes repeated model catalog refresh errors in native Codex sessions using the inference relay. #197768. Thanks @bdjben.
- Fixes native model lookup timeouts stopping reply preparation before the selected model or configured fallbacks can be tried. #199039. Thanks @MoerAI, @obviyus, @sangwf and @kqtxmaster-boop.
- Updates bundled Codex to 0.169.0, which can help native Codex discover newer models. Availability still depends on your account and backend. It also fixes failed completed turns with uploaded audio losing their final error snapshot and identifies permanent launch failures for the managed Codex executable. #200784. Thanks @vacinc.
- Fixes missing or non-executable managed Codex files producing an unhelpful startup error. Repair the file or reinstall OpenClaw, then restart the Gateway. Commit fd2763c. Thanks @RomneyDa for the fix and @vincentkoc for coauthoring.
- Fixes discovered models losing automatic Code Mode after a catalog refresh or Gateway restart when an output limit is configured. This also helps preserve prompt-cache reuse. #190326. Thanks @VACInc.
- Fixes eligible live-listed models showing default limits and input types when catalog details are available. Their catalog limits and supported inputs can now be used in model lists and requests. #200403. Thanks @obviyus.
- Fixes model descriptions showing reasoning effort controls that the selected account does not offer when catalog details are added. Commit 5152253. Thanks @RomneyDa and @vincentkoc.
- OpenClaw can close idle Claude CLI sessions before starting a new warm session when the computer is low on memory. If there is still too little memory, the new turn can fail over while active sessions continue. #192534. Thanks @obviyus.
- An eligible retry can resume a Claude CLI session after its first turn exceeds an output limit, instead of starting over without the work and conversation already recorded. #192838. Thanks @yousan, @obviyus, and @slatem.
- Fixes Claude CLI turns with many small tool calls failing early and losing their final reply because discarded stream data counted against output limits. #200501. Thanks @obviyus, @trinity-ai-coder, @terwox, and @slatem.
- OpenClaw stops a Claude CLI turn that exceeds its output limit after a tool call or accepted session spawn, instead of replaying the request on a fallback model and potentially repeating the work. #200801. Thanks @fede-kamel, @obviyus, and @jzmudzinski.
- Fixes two changes to conversation history that could break prompt-cache reuse between runs. Calls to unavailable tools keep their results in history, and GitHub publication receipts arriving during a turn stay in the order the provider saw. #197874. Thanks @obviyus.
- Fixes blank tool results being replaced with placeholders when conversation history is replayed to a provider. Commit c823541. Thanks coauthor @RomneyDa.
- Fixes image attachments changing in replayed requests after a retry or session reopen, which could prevent a provider from reusing a matching prompt cache. Commit 0df0aa8. Thanks coauthor @RomneyDa.
- Fixes an OpenAI Responses replay conflict that could interrupt a chat or prevent a subagent answer from arriving after an unstored turn. #200163. Thanks @obviyus.
- Fixes unchanged OpenAI Responses tool arguments changing when a conversation is replayed, including after it is reloaded. b4c4dda. Thanks co-author @RomneyDa.
- Fixes some Responses conversation item conflicts interrupting a turn or a subagent answer. Before output starts, OpenClaw can retry with repaired history. 2f51297. Thanks co-author @romneyda.
- Fixes Codex quota fallback failing to carry your current request into the fallback run. #200363. Thanks @RomneyDa.
- Fixes oversized Codex requests losing their opening instructions by retaining the beginning and end when the request must be shortened. The middle is still omitted. #200690. Thanks @obviyus.
- Oversized Codex requests can keep their full text in an attachment when the execution environment can read it, so the model can retrieve content otherwise cut from the middle. If staging or access fails, the middle is still cut. #200735. Thanks @obviyus.
- Reduces the risk of slow credential and database checks holding up the Gateway during startup or preparation. Pending legacy credentials still block the affected provider until they are migrated. Commit 5d86961. Thanks coauthor @RomneyDa.
- Prepares selected file identity and descriptor checks asynchronously for plugin model catalog credential work and ordinary SQLite worker admission. Catalog database operations are coordinated within the writer turn, while service settlement remains synchronous. Commit f689592. Thanks coauthor @RomneyDa.
- Fixes fal image and music downloads returning empty files or error documents as generated assets. Affected requests now report a fal download error instead. #189781. Thanks @Yigtwxx for the superseded proposal #139191.
- Removes the Beta label from Sign in with ChatGPT and explains ChatGPT plan usage and where Business members and administrators manage app limits. #196330. Thanks @stevenlee-oai and contributor @hannesrudolph.
- Updates Grok OAuth proxy requests to use Grok client version 1.0.13 and routes simple completions through the OAuth wrapper. 791c00a. Thanks @DonnieFi and contributors @steipete, @vyctorbrzezowski and @RomneyDa.
Automations and Scheduling
- Fixes a quiet manual automation run delaying its caller when it targets the same agent's global conversation. The caller gets the accepted run ID promptly and the job runs after the current turn ends. #188250, issue #188222. Thanks @shakkernerd.
- Fixes an unhandled storage error when scheduled work resumes. The service stays running, logs the error and retains the queued request for a later retry. Commit d0cc893. Thanks coauthor @RomneyDa.
- Fixes recovery of scheduled runs failing when a job's combined retained history exceeds the worker's memory. The individual data needed for recovery must still fit. Commit 9600d0f. Thanks @RomneyDa and @fuller-stack-dev.
- Fixes scheduled script triggers repeatedly starting over when initial preparation outlasts an evaluation timeout. A later tick can use the ongoing or completed preparation, although the timed-out evaluation still reports a timeout. Commit 3c5da8a. Thanks @RomneyDa and @guhyun9454.
- Fixes scheduled jobs stopping after the scheduler loses its state worker. The scheduler can reconnect and resume work even if its first replacement is also lost. Commit a77c494. Thanks coauthor @RomneyDa.
- Fixes isolated scheduled announcements omitting the channel’s named default account from their delivery and account-scoped session route. Commit af9f0fe. Thanks @RomneyDa for coauthoring.
- Fixes heartbeat alerts losing line breaks and indentation when an acknowledgement is removed. #189994. Thanks @yetval and @obviyus.
- Fixes isolated scheduled agent turns failing with no callable tools when
tools.allowis empty. Job tool limits and deny rules still apply. #197590, #197552. Thanks @wr-web. Thanks @primepal2 for reporting the issue. - Fixes an accepted scheduled turn in another session stopping when its sending occurrence ends or its sender session is cleaned up. 83fb9fb. Thanks co-author @RomneyDa.
- Fixes restored automations losing authority after a chat edit, reset or compaction. Interrupted condition checks can also be retried without counting as job failures. 6cf56c5. Thanks co-author @RomneyDa.
- Fixes cron retries running alongside timed-out work that has not finished. Restart recovery also avoids recording the same timeout twice. Commit 331eb67. Thanks @xydigitLybnnnn and @RomneyDa.
- Fixes automation retry decisions that could lose why a run ended. Invalid session targets now produce guidance to update the agent or session target, and a failed one-shot automation can report that it was disabled even when an earlier alert was suppressed. Commit 5442a0c. Thanks @vincentkoc and @romneyda.
- Fixes custom webhook deliveries being rejected above 256 KiB by raising the normal body limit to 1 MiB. The Gateway now logs authenticated body rejections so operators can identify the path and limit. #200514. Thanks @garrytan and contributor @obviyus.
Browser and Computer Use
- Fixes Gateway startup freezing while registering the CUA Computer plugin on Linux. Restricted or unrecognized system mappings can still produce an unsupported-platform message. #188521. Thanks @steipete.
- Fixes browser snapshots exposing filled passwords and marked one-time code, card number and card security code fields to the model. Continue to use a dedicated browser profile for sensitive work. Commit e48a7a3. Thanks @RomneyDa.
- Fixes Linux and macOS host exec sessions staying open after the command exits because a background process holds its output open. Keep work in the foreground if you need its later output. #191632. Thanks @Decent9967 for the earlier repair in #147335.
- While the Gateway is connected, the browser continues checking cross-origin document navigations in a guarded tab between actions, including manual navigation. This does not filter all network traffic. After reconnecting, coverage resumes at the next guarded action or navigation. #198452. Thanks @daniel-lxs.
- Enforces strict browser navigation policy for existing-session profiles by refusing HTTP and HTTPS navigation and tab opens when
dangerouslyAllowPrivateNetworkis explicitly false. Use theopenclawdriver when you need strict network navigation. Commit 269123d. Thanks @RomneyDa. - Fixes OpenShell mirror-mode file tools selecting the wrong path when a filename or directory begins with whitespace. 8e1e778. Thanks @RomneyDa.
- Fixes garbled text in link understanding when a page declares a supported non-UTF-8 encoding. 1342876. Thanks @RomneyDa.
- Fixes garbled fetched pages and feeds when a commented-out charset declaration would otherwise select the wrong encoding. 8782699. Thanks @RomneyDa.
- Fixes the first browser operation failing while an externally managed attach-only CDP browser is starting. OpenClaw can wait about 30 seconds for it without using the action timeout. #188616, #187814. Thanks @daniel-lxs and @hannesrudolph.
- Fixes computer-use field and browser text losing leading or trailing spaces, tabs and newlines, including text made entirely of whitespace. fb5a76a. Thanks @RomneyDa.
Plugins and Integrations
- Fixes queued replies repeatedly retrying while a plugin replacement is unavailable. Eligible replies can resume after replacement or rollback, and unexecuted durable input can return to the channel queue if shutdown interrupts the wait. a9aa37b. Thanks to coauthor @romneyda.
- Fixes plugin tool discovery holding up unrelated Gateway requests, including when you open the tool catalog in the Control UI. Individual synchronous plugin callbacks can still block while they run. #188033, issue #188031. Thanks @ctbritt and @obviyus.
- Fixes bundled MCP server setup on slow storage holding up unrelated Gateway requests. Setup still waits for its data directory and reports a server that cannot be prepared. #190233. Thanks @steipete.
- Compiles captured plugin source in memory when loading through Node module hooks, without reopening the source or writing a temporary compiler tree. Commit 51e94c1. Thanks @RomneyDa.
- GitHub publication has more time to stage large or slow workspaces, and staging failures now explain what to check. Across tools, failure logs can now include error messages containing private tool or response content. Review these logs before sharing them. #188355. Thanks @steipete.
- Cloud sessions can continue after workspace settlement without waiting for GitHub publication. Eligible queued requests can reuse an already published exact snapshot. #188817, issue #188796. Thanks @galiniliev for the source contribution.
- GitHub publication readiness checks avoid a database copy that could hold up other work on large agent databases. 6c8c7a5. Thanks co-author @RomneyDa.
- Personal publication restart bookkeeping uses a dedicated route while preserving a publication in progress for the current instance. 200fa7d. Thanks co-author @RomneyDa.
- Fixes preparation of older owned PRs being rejected when a clean managed workspace contains newer published project history. #188790, issue #188751. Thanks @shakkernerd.
- Preparing a cold PR workspace and fetching a newer publication base can transfer less data by avoiding unneeded historical file contents when the remote supports filtering. 75763e1. Thanks @romneyda and @shakkernerd.
- You can refresh the current target branch of a prepared PR in a writable managed sandbox while keeping your accepted PR head and work. Repeated preparation also keeps integrated history in the review comparison, and a retargeted session-owned PR can continue from its previously published head. Refresh does not import a newer remote PR head or publish. 929f746. Thanks @shakkernerd and co-author @romneyda.
- Fixes Azure Sandbox heartbeats failing to refresh idle activity when the provider rejects an idle-policy change. Warnings now tell you when activity is not being refreshed and whether OpenClaw will retry. b93ab7d. Thanks @romneyda and co-author @galiniliev.
- Fixes plugin updates replacing environment references with their resolved values when a setting moves. References already saved as literal values cannot be recovered automatically. Commit 507b9fc. Thanks @RomneyDa.
- ClickClack HTTP requests now block redirects to other private network origins, and retired callers cannot send further requests. If your self-hosted API redirects to another private origin, configure the final API origin directly. #190944. Thanks @pgondhi987.
- Fixes plugin-started GPT-Live voice sessions failing with an authentication error despite valid ChatGPT OAuth credentials. #190812, #187772. Thanks @hannnnn-l and @obviyus. Thanks @scarabone for reporting the issue.
- Fixes meeting audio capture stopping after the join request finishes when an owning plugin instance is available. Commit cb07044. Thanks @RomneyDa.
- Fixes Gateway startup failing after a channel plugin is removed while its exact
{ enabled: false }channel entry remains. Entries with additional settings remain subject to validation. #190375. Thanks @ericcaiwx-star for an earlier contribution. - OpenClaw delays pending GitHub publication recovery when process memory is low and can resume it on a later recovery pass when memory is available. Commit 143a5b2. Thanks coauthor @RomneyDa.
- Fixes isolated plugin packages losing TypeScript declaration files from patched dependencies when those files are hardlinked. 3f50b17. Thanks co-authors @romneyda and @vincentkoc.
- Fixes managed plugin installs and Doctor repairs failing when an inherited npm allow-scripts setting conflicts with the install. Lifecycle scripts remain disabled. #200786, issue #200773. Thanks @steipete, and thanks @cbhawthorne84 for reporting the issue, and @nasrsuresh for advice on script blocking and plugin projects.
- Fixes valid native plugin files being rejected on filesystems that turn a requested link into a copy. Damaged companion files are still rejected during capture and recovery. Commit 43350c5. Thanks @RomneyDa.
- Reuses inspected package metadata when reconstructing a plugin and preserves discovery of legacy main or index entries in a relocated snapshot. Commit 6db4325. Thanks @RomneyDa.
- Reduces interruptions to other Gateway work while a plugin reload copies the previous version for recovery. Commit 5f61e87. Thanks @vincentkoc and @romneyda.
- Retains native roots for asynchronously prepared plugins through cache and recovery leases after a preparation worker exits or the Gateway restarts. Accepted installed-index publication waits for native commits and can surface their failure. Commit 291274e. Thanks coauthor @RomneyDa.
- Reduces potential Gateway stalls during plugin reload recovery by moving the remaining recovery file work off the Gateway thread. Commit f03b886. Thanks @RomneyDa.
- Detaches long-lived plugin recovery capture from the predecessor invocation context, so recovery resources do not retain that context. Commit 696f1b3. Thanks coauthor @RomneyDa and coauthor @vincentkoc.
Security and Privacy
- Gateway shell commands that need approval are checked more consistently on macOS and Linux. A command that previously ran automatically may now prompt for approval or be denied under your existing policy. #187884. Thanks @art22s.
- Fixes automatic command review being silently disabled when you reapply an unchanged exec policy, including the cautious preset when it matches your current policy. #190024. Thanks @yetval, author of the superseded #168178.
- Fixes the Gateway allowing some shell commands to run without the approval required by your policy. Commit 266a72e. Thanks @RomneyDa.
- On Windows, approvals for inline POSIX shell commands remain one-time approvals. Running the command again requires another approval, even if you previously selected allow always. Commit f1a1157. Thanks @RomneyDa.
- Fixes command approval checks when a command changes its environment. Some commands that previously ran automatically now require a prompt. Commit 58997e1. Thanks coauthor @RomneyDa.
- PostgreSQL-backed writes now honor authority checks before committing. After a profile role change with an uncertain commit result, affected actions are blocked until authority is recovered. d388d9f. Thanks co-author @RomneyDa.
- Queued cloud-worker deletion and cancellation now check that you still have permission before the action is accepted. Cleanup already accepted continues even if your access later changes. 16a293f. Thanks @joshavant and co-author @RomneyDa.
- Prevents plugin-owned credentials from being included in general model credential discovery. A conflicting provider ownership declaration also no longer prevents discovery for a healthy provider. Commit c9418d7. Thanks @RomneyDa.
- Fixes Doctor removing empty restrictions when it removes the last retired bundled skill or invalid secrets egress proxy host. Commit a2e3b38. Thanks @RomneyDa, @yetval and @steipete.
- Fixes configuration reads exposing inline environment values whose names were not recognized as sensitive. Escaped values are also protected, and saving redacted values preserves the originals. #190618. Thanks @yetval.
- Fixes new Doctor SQLite migration failure reports exposing your home directory and username in cases where those paths should be redacted. Previously saved reports are not changed. #200457. Thanks @steipete. Thanks @LolPanda2004 for reporting #200451.
- Fixes sensitive
config setvalues appearing in audit records when a separate--expect-current-jsonvalue is used among the arguments. Commit 9a3a5e2. Thanks @RomneyDa and @wangmiao0668000666. - Fixes environment-sourced PostgreSQL credentials being inherited by host commands and sandbox environments. Restart the Gateway if you change which environment variable supplies the connection. Commit 2bf4762. Thanks @romneyda and @galiniliev.
- Fixes sandboxed web search using providers outside its existing trust policy. If your sandbox uses an excluded third-party provider, enable or select a bundled or verified official provider to restore search. #191748. Thanks @finn763 for the earlier repair in #144520 and @yetval for reporting #144164.
- Prevents OpenClaw from automatically adding workspace-controlled mise or XDG executable directories to its command search path. Move tools outside the workspace if you rely on those directories for automatic discovery. #191102. Thanks @pgondhi987.
- Blocks
PERLLIBandPERL5SHELLin workspace environment files and host command environments, preventing those variables from redirecting Perl module or shell selection. Workflows using them need another supported runtime configuration. #198353. Thanks @pgondhi987. - Fixes session search returning results after access is revoked during the search. It also excludes incognito actor-memory results and their page information. 118c569. Thanks @RomneyDa.
- Fixes a replacement session inheriting authority for a reply that belonged to the previous session. f6e236c. Thanks @RomneyDa and @vincentkoc.
- Uses the admitted agent identity when selecting a session authentication source, avoiding a reverse owner lookup while retaining database path and state-root checks. Mismatched readers are not used as the shared source. Commit fb82e04. Thanks @RomneyDa.
- Auth profile updates now refuse to write to a replacement database if an already selected database file changes while the update is being prepared. Commit eabd92a. Thanks @RomneyDa and @vincentkoc.
- Worker workspace uploads and downloads now verify HTTPS certificates when the configured Gateway fingerprint is blank. Invalid explicit fingerprints are rejected before credentials or workspace data are sent. #190002, #167094, #167198. Thanks coauthor @avelev99 and @aniruddhaadak80 for reporting.
- Fixes unreadable native hook input leaving policy-bearing Codex tool requests without a decision. Those requests are now denied with guidance to shorten or split the input and retry. #194638. Thanks @mmaps.
Other QOL Improvements
- Supports Goal Start and Resume in an eligible, idle local Codex session. Resume keeps the Goal and conversation history and directs the model to read the current objective, including edits made while paused. #186524. Thanks @jjjhenriksen and @brokemac79.
- Reduces processing work for messages with very long punctuation runs followed by text, while keeping stop commands with trailing punctuation recognizable. 77b412c. Thanks @LeizhenZhang and @RomneyDa.
- Latest-session usage transcript lookup now performs filesystem checks asynchronously while retaining its lookup and containment policy. Commit 2e75930. Thanks coauthor @RomneyDa.
- Cancelling workspace preparation now removes a queued final session binding without waiting for an unrelated session writer. #188887, issue #188771, issue #188923. Thanks @shakkernerd.
- Managed worktree creation now reports insufficient disk space promptly when another operation cannot free the needed space, instead of waiting for it unnecessarily. 1920551. Thanks co-author @RomneyDa.
- Managed worktree creation can proceed during slow maintenance when capacity is available. If a checkout cannot be safely snapshotted for eviction, its unsaved work is preserved and the failure is reported. db3cded. Thanks co-author @RomneyDa.
- Reduces Gateway pauses on Linux when monitoring background commands, restarting scheduled command watchers, or opening terminals. #189277. Thanks @steipete.
- Reduces contention with other session writes when deleting SQLite sessions. Slow deletion warnings also give more useful timing information. #189287. Thanks @steipete.
- Reduces database write contention when trajectory capture continues during retention cleanup. #189338. Thanks @steipete.
- Token-only Gateway clients now load local credential and app runtimes only when needed, and asynchronous device-identity imports no longer load the synchronous store. Commit 69b6229. Thanks @RomneyDa.
- OpenClaw checks for memory pressure from startup and can release idle workers sooner, including on memory-limited Linux systems. During shutdown, it lets dependent readers finish before closing idle agent workers. Commit 35b3617. Thanks @RomneyDa.
- OpenClaw avoids starting an optional helper process on Linux computers with at most 1 GiB of physical or process-limited memory, conserving memory. Starting child processes directly may take longer. Commit 7768afc. Thanks @RomneyDa.
- OpenClaw can release an idle session discovery worker, or release it under critical memory pressure after its current preparation finishes. Commit faa2691. Thanks @RomneyDa.
- Reduces the risk of slow repository access holding up the Gateway while an agent turn is prepared. Repository changes can be picked up on a later turn. Commit 670cf66. Thanks @RomneyDa.
- Reduces the risk of slow storage holding up the Gateway while state and agent database reads are prepared. Reads still refuse changed ownership or a replaced database when its expected identity is available. Commit 309067b. Thanks @RomneyDa and @vincentkoc.
- OpenClaw does less work when discovering session tools without spawning a session. A spawn request cancelled or no longer authorized while its runtime loads is stopped before it runs. Commit 12937fe. Thanks coauthor @RomneyDa.
Other Bug Fixes
- Fixes a Gateway crash that could interrupt active work when a steered message is not saved to the conversation. #200156. Thanks @steipete.
- Fixes large text files using excessive memory or failing when an agent requests only a page. Truncated results now omit total line and byte counts until the end of the file is reached. #186992, #186565. Thanks @MoerAI and @steipete. Thanks @JasonBuildAI for reporting the issue.
- Fixes very large assistant messages stalling the Gateway while it prepares ordinary chat previews. Full replies and the saved transcript remain intact. #189804. Thanks @steipete.
- Fixes empty session catalog discovery waiting for session storage to become ready. Catalog reads can also proceed while a reconnect subscription is waiting. 67cea57. Thanks @RomneyDa and @steipete.
- Fixes first-session creation failing after a read found no database for a new agent. #200306. Thanks @vincentkoc.
- Fresh-session input now resolves the selected store to its SQLite database. Commit cc6c06d. Thanks @joshavant and @RomneyDa.
- Fixes transcript writes waiting behind slow session history or model-context reads. A slow read can still take time, but it no longer normally holds up those writes. #187989. Thanks @RomneyDa, @jalehman, @fuller-stack-dev and @vincentkoc.
- Fixes session history and model-context reads being retried or rejected unnecessarily when another conversation in the same database changes. Commit ae14052. Thanks @RomneyDa and @vincentkoc.
- Fixes restart recovery failing before a reply when competing transcript writes repeatedly interrupt a history read. Recovery gets another chance to read current history while preserving stale-data checks. Commit 516f039, backported from #188073. Thanks @vincentkoc and @RomneyDa.
- Reduces delays to foreground database writes during large usage-cache cleanups. Usage updates also become eligible for processing after five minutes of continuous activity, though completion can take longer. #188611. Thanks @steipete.
- Fixes session-cost reports waiting indefinitely behind another usage refresh. After up to ten seconds, /usage cost shows the last known total as stale, or Session n/a when no cached total is available. #188909. Thanks @sebtardif.
- Fixes interrupted cloud-worker Stops that could leave workspace changes unsaved until a Gateway restart or another Stop attempt. If the worker is offline, the save waits for reconnection. #188805, issue #188796. Thanks @galiniliev for the source contribution.
- Fixes sandbox tools waiting behind later checkpoints for the same workspace in one Gateway. If the workspace stays busy, the operation now reports an error after 60 seconds and suggests retrying. #200643. Thanks @steipete.
- Fixes slow filesystem operations during native resource startup or cleanup pausing unrelated Gateway requests. #200216. Thanks @steipete.
- Fixes cold Node discovery blocking the main thread when starting a terminal under Bun. Plugin developers can also use an asynchronous Node resolver. Commit c443580. Thanks coauthor @RomneyDa.
- Fixes Gateway health and readiness reporting a usable child runtime when Node exists but cannot start. Readiness can now report the failure before a child launch. Commit ec31f39. Thanks coauthor @RomneyDa.
- Fixes Talk persistence work stalling Gateway shutdown while it waits for cancellation. #189662. Thanks @steipete.
- Fixes stalled Talk cleanup holding a process-managed Gateway shutdown before its cleanup deadline. Unfinished cleanup is reported, but pending transcripts may not have been saved. #189684. Thanks @steipete.
- Fixes
openclaw path set --value-jsonsilently ignoring the flag for Markdown or YAML, which could write a string instead of the requested typed value. Remove the flag to write a plain value to those formats. #189941. Thanks @obviyus and @jerome-queck. - Fixes invalid Markdown frontmatter paths changing the wrong key when you use
openclaw path set. The command now rejects those writes without changing the file. #193878. Thanks @zhangguiping-xydt and @obviyus. - Fixes Markdown field edits changing a matching key inside an earlier fenced code example. Fields that cannot be located safely in a table are now rejected instead of being edited. Commit 3ee63f8. Thanks @RomneyDa.
- Fixes JSON searches and edits using an earlier value when an object contains duplicate keys. If JSONC insertion encounters an ambiguous duplicate-key path, merge the duplicates and retry. Commit cc97737, related to commit 7f593b6. Thanks @RomneyDa.
- Fixes subsystem warnings corrupting the terminal interface, including when connected to a remote Gateway. Console logging also resumes after a failed local startup. #190163, #172119. Thanks @mike1122real-hub for an earlier contribution.
- Fixes stale Windows locks remaining in place when the operating system can confirm that their owner has exited. Commit e501b891. Thanks @romneyda and coauthor @barry-886.
- Fixes Gateway process cleanup failing on Linux hosts that block process-group signals. It also lets command cleanup finish when the affected process group has exited. Commit c4d49407. Thanks coauthors @sallyom and @RomneyDa.
- Fixes a process abort on non-Windows Node 26.12.1 when an internal process anchor with no standard input is retired before a synchronous child process starts. Commit 45e96e42. Thanks @RomneyDa for coauthoring.
- Addresses a reported assembled-build import conflict while preserving direct process spawning on memory-limited hosts. Commit 3fb0050. Thanks coauthor @RomneyDa.
- Fixes validated OpenClaw tool calls being rejected when an agent supplies supported numbers or booleans as quoted values through Tool Search, Code Mode, Codex dynamic tools, or the Agents API. #193200. Thanks @ayaangazali and @obviyus.
- Fixes compact tool input hints omitting required arguments or showing unknown when a tool has a wide schema. #200064. Thanks @jalehman.
- Fixes Code Mode scripts receiving substituted text instead of supplied structured data when large injected values encounter memory pressure. Commit 72dde26. Thanks @RomneyDa.
- Code Mode stops a cell when it cannot decode a tool reply, instead of continuing with raw text as the result and potentially calling later tools with the wrong data. Commit 6c2bd97. Thanks coauthor @RomneyDa.
- Preserves host-owned tool-call context during validation and keeps versioned session reads within their lifecycle revision. Commit 5297f3f. Thanks @RomneyDa and @vincentkoc.
- Fixes interrupted sandbox workspace base refreshes losing progress repeatedly, so a later retry can continue and finish. Commit 0c62f16. Thanks @vyctorbrzezowski and @RomneyDa.
- Fixes managed local sandbox workspaces becoming unavailable after you rewind a session or switch conversation branches. Accepted edits carry forward, while conflicting edits remain saved for you to resolve before retrying. Commit 0948e8f. Thanks @shakkernerd and @RomneyDa.
- Fixes an interrupted TTS conversion creating or replacing the requested output after the interruption has been accepted, when it occurs before the output copy begins. Commit 8f22f6e. Thanks @romneyda.
- When a session changes before held subagent results can be delivered, the new conversation explains what happened and points you to Tasks to recover them. The notice appears when the batch is evaluated. #200005. Thanks @steipete.
- Fixes a lost answer when the Gateway restarts after a parent has accepted a subagent result but before it finishes replying. Results already marked permanently failed without a pending retry still need explicit recovery. #200279. Thanks @steipete.
- Fixes finished subagent results getting stuck when the requester session store path changes. #200760. Thanks @obviyus.
- Fixes session maintenance removing a completed subagent session while its result is still being settled. 3d2ebc2. Thanks co-authors @sallyom and @romneyda.
- Fixes interrupted subagent result delivery missing an automatic retry when the interrupted run is proven safe to replay. 8bd36aa. Thanks co-author @RomneyDa.
- Fixes already consumed subagent results being replayed after a Gateway restart. It also keeps failed or uncertain answer delivery from being reported as successful merely because the requester received the results. 1d09030. Thanks @RomneyDa.
- Fixes commands on macOS and Linux selecting the wrong file or directory when a path passes through a symlinked directory followed by a parent segment. #200031. Thanks @daniel-lxs.
- Fixes resolved SSH identity filenames being changed when they contain consecutive spaces or tabs. c2689ce. Thanks co-authors @sallyom and @RomneyDa.
- Fixes node file transfers selecting a different filename when the requested path has surrounding whitespace. 7d5bb48. Thanks @RomneyDa.
- Fixes configuration edits interrupting already-started subagent resumes, background follow-ups and scheduled session turns. New settings apply to subsequent turns. #200516. Thanks @obviyus and @romneyda.
- Fixes accepted sibling work losing authority after an edit to the parent conversation when its saved source and Gateway context are available. 4b9c3d8. Thanks co-author @romneyda.
- Fixes interrupted signed-in sessions failing to resume after a Gateway restart when roles are not configured. Queued input and later host-triggered turns can also continue under the appropriate session authority. 88f5cdc. Thanks co-author @RomneyDa.
- Fixes background results and their history being lost after a conversation edit, rewind or reset. An unchanged eligible goal can also continue after the conversation identity changes. c3543b2. Thanks co-author @RomneyDa.
- Fixes instructions from a parent agent arriving only after its busy child finishes, so the child can receive a change of plan during its current run. #200574. Thanks @obviyus.
- Fixes retained guidance and paused resumes being rejected or sent to an older run after the same subagent task advances. It also fixes chat messages without a session ID being rejected from an active earlier view after another message is added. 939f81b. Thanks @RomneyDa.
- Fixes an authorized
/stopfailing to stop current chat work from a stale or empty view. Retained sends now report when a target reset or deletion prevents delivery, while eligible sends can continue after a transcript change. 12f1021. Thanks @romneyda. - Fixes worker-hosted Code Mode tools continuing without a cancellation request after their Gateway response wait fails. Tools must honor the request, and completed effects are not undone. #200647. Thanks @steipete.
- Fixes ACP tasks potentially running twice when a reply handler errors or times out after starting work. The turn now reports a task error instead of trying another runtime. 9eeb788. Thanks @RomneyDa and @xenaocx-dev.
- SQLite-backed session rewinds and forks now refuse histories that exceed the read limits before changing the session, and explain why the request was refused. Commit f2b7305. Thanks coauthor @RomneyDa.
- Preserves the live SQLite admission and its receipt when an asynchronous admission lookup overlaps a native database open. Both descriptors remain in custody until retirement, and a write still invalidates the receipt. Commit bf13058. Thanks @RomneyDa.
- State reads and agent database operations now refuse a changed owner or database instead of adopting it during preparation. When storage prevents the Gateway from renewing ownership in time, it returns an error asking you to retry after storage recovers. Commit fa61397. Thanks @romneyda and @vincentkoc.
- Fixes Gateway startup reads being refused when another process briefly owns the state database schema. Startup now waits up to the configured SQLite busy timeout, while offline maintenance still produces an access error. Gateway startup and Doctor also retain the underlying cause of state registry inspection failures. Settings switches no longer lose clicks during unrelated refreshes, repeated Chat Access selections no longer race a save, and transcript right-click menus work when a floating control appears under the pointer. Commit 8142a61. Thanks @romneyda and @vincentkoc.
Was this useful?