On this page
On this page
Release notes
v2026.11.4
v2026.11.4
AI agents and tools can read these release notes as plain Markdown.
This is a patch release with fixes for problems that could skip buffered Signal messages after a reconnect, lose replies after a settings reload, or damage memory and configuration recovery copies. It also adds a delegated change approval setting and lets policy plugins inspect images before they reach a model.
Release scale: 125 source pull requests, 106 direct commits, and 90 contributors.
Installation and Onboarding
-
Fixes private CLI launchers failing when npm install prefixes or Git checkout paths contain spaces or shell characters.
More details
Released change. Thanks to @RomneyDa. Original work by @McgAtom in #189541. Thanks to @obviyus for the documentation in the original PR.
-
Fixes the Windows PowerShell installer handing batch shims to onboarding without the system command interpreter. If launch fails, the installer shows a command to run onboarding manually.
More details
Released change. Thanks to @RomneyDa. Original work by @DonnieFi in #185309.
-
Fixes macOS installation reporting Homebrew setup as successful and continuing provisioning after the Homebrew bootstrap fails. The original error and exit status are preserved.
More details
Released change. Thanks to @RomneyDa. Original work by @steipete in #187688. Thanks to @McgAtom for the report.
Web UI
-
Fixes the Control UI freezing in Safari when a picker replaces an opening menu during a layout change, so clicks and typing keep working.
More details
The stylesheet change applies to document overlays across browsers, but not to overlays inside shadow roots. A closing overlay that had flipped to fit the window may briefly return to its usual side while fading out.
Released change, adapting #198658 by @fuller-stack-dev. Thanks to @RomneyDa for the release adaptation and @vincentkoc for the release coauthorship. Thanks to @erik-toc for the related report and @alfred429 for related WebKit tracing.
-
Fixes recent session titles being difficult to tap on narrow touch screens.
More details
Released change, adapting #194840. Thanks to @RomneyDa, @najef1979-code and @vyctorbrzezowski.
-
Prevents internal wake messages from becoming automatic session titles or Doctor title-repair topics.
More details
Released change, adapting #179219. Thanks to @RomneyDa and @steipete.
-
Recovers space for Team session labels in the touch sidebar.
More details
Released change. Thanks to @RomneyDa and @vincentkoc. See the Control UI guide.
-
Reduces Control UI slowdowns in WebKit after repeatedly switching conversations.
More details
Released change, #194027. Thanks to @steipete, @RomneyDa, @vincentkoc. See the guide.
Updates and Maintenance
-
Fixes later database work being delayed until a lease expires when offline maintenance interrupts lease startup. Once maintenance ends, cleanup can retry, and successful cleanup removes that wait.
More details
Released change. Thanks to @RomneyDa and @vincentkoc.
-
Fixes update finalization, repair, Doctor, and plugin installs and updates failing with a lost plugin lease when long-running work blocks OpenClaw. The process running the operation must already include this fix; a newer update cannot fix an older process already running.
More details
#188404, released change. Thanks to @steipete, @RomneyDa and @vincentkoc, and to @patosullivan and @yyds-xxxx for the related reports #188335 and #182476. See the update guide.
-
Reduces the amount of retained conversation data loaded when checking for interrupted sessions after a Gateway restart.
More details
Released change, adapting #176737. Thanks to @steipete, @vincentkoc and @RomneyDa. See the restart recovery guide.
-
Fixes restart recovery checking interrupted conversations before the plugin services it may need have started. Lets the session database prepare after plugins load, even when worker placement is not configured. Rejects outdated recovery decisions if settings change before they are saved. Channels, startup hooks and startup-gated requests still wait for the recovery check.
More details
Released change, adapting #181289. Thanks to @steipete, @vincentkoc and @RomneyDa, and to @oxen-horse for the related report and investigation. See the restart recovery guide.
-
Reduces memory used by Doctor when repairing large SQLite transcript histories and lets interrupted repairs resume.
More details
Released change. Thanks to @RomneyDa. Original work by @steipete in #195459.
-
Fixes stable updates following saved prerelease tags for trusted official plugins. Previously selected tags are also affected, and the installed plugin is retained if no compatible stable version is available.
-
Fixes configuration recovery losing accepted edits after a large formatting or size reduction, while keeping suspiciously truncated input out of the recovery baseline.
-
Prevents a failed Doctor merge of legacy memory from truncating MEMORY.md and keeps recovery copies of the original files.
More details
#190497, Released change. Thanks to @syxscott, @steipete, @RomneyDa. See the Doctor guide.
-
Fixes Doctor and update finalization stopping when a checkpointed audit archive has lost its sanitized companion but has no records left to import. Doctor preserves the raw archive and warns when its checkpoint still matches. If records remain unimported, Doctor still stops; restore the named companion from a verified backup, then rerun
openclaw doctor --fixagainst the same state directory.More details
#189935, release merge, adapting #186998. Released change. Thanks to @amittell, @steipete, @RomneyDa and @vincentkoc, and to @AlyaKoni for the report. See the repair and recovery guide.
-
Fixes startup and Doctor repair for older profile databases missing the login column, while preserving existing profile data.
More details
Released change. Thanks to @RomneyDa. Original work by @steipete in #186529. Thanks to @ausvald for the report.
-
Prevents an interrupted pre-update configuration snapshot from damaging an existing rollback copy. The updater taking the snapshot must already include this fix, and the retained copy may be older.
More details
Released change. Thanks to @kaguranananana, @steipete, @RomneyDa.
-
Fixes updates failing to finish after an unrelated configuration edit during plugin finalization, while preserving that edit. Resolve conflicts in the same fields before retrying update repair.
More details
Released change. Thanks to @RomneyDa for the release adaptation. Original work by @steipete in #189819.
-
Prevents a newer CLI from migrating a database still used by an identified older Gateway service. If migration is refused, run openclaw update or use the service installation’s CLI. Service detection is best effort.
More details
Released change. Thanks to @shakkernerd, @RomneyDa. See the update guide. Original sources: #186971, #187819, #187870, #187886, #187966, #190194. Thanks to @steipete, @jalehman, @fuller-stack-dev, @vincentkoc and @Satoshi-Kusumoto for the original work. Thanks to @tomdailey55 for the report.
-
Fixes Gateway startup failures when a database policy changes while old cleanup connections are still open. Conflicting policy changes are refused while work is active instead of hanging.
More details
Released change, adapting #187870. Thanks to @shakkernerd, @RomneyDa and @steipete.
-
Fixes manual trimming of large session transcripts exhausting memory when only a small recent tail is kept.
More details
Released change. Thanks to @RomneyDa for the release adaptation. Original work by @vincentkoc in #194066.
-
Fixes updates being blocked by an unsafe temporary-directory ancestor when the installed updater can use a safe private workspace in the state directory. An older updater may still need the temporary hierarchy repaired before its first update.
More details
Released change. Thanks to @RomneyDa. See the update guide. Original work by @steipete in #187737. Thanks to @Yachiyo1680 for the report.
-
Fixes PostgreSQL writes ignoring your configured commit durability. The bridge now inherits your PostgreSQL setting, which can increase write latency; choosing asynchronous commits can risk losing recent acknowledged writes after a crash.
More details
#180821, Released change. Thanks to @steipete, @RomneyDa.
-
Fixes configuration changes getting stuck during a Gateway restart when plugin lifecycle work needs to finish. Restart deferral warnings also identify pending configuration preparation.
More details
Released change. Thanks to @RomneyDa.
-
Fixes Doctor being unable to preserve original state before repair when a volume remount widens permissions on an eligible recovery capture folder.
More details
#194307, #192959, Released change. Thanks to @tssala23, @sallyom, @RomneyDa. See the Kubernetes guide.
-
Fixes resumed database ports reporting success when the staged copy is rejected, while retaining the verified copy and recovery manifest for later recovery.
-
Fixes updates failing on low disk space when Doctor inspects a stopped agent database by avoiding an extra copy for qualifying databases.
More details
Released change, #196346. Thanks to @fuller-stack-dev, @RomneyDa. See the Doctor guide.
-
Allows update and deployment restart handoffs to wait up to 25 seconds for pending final chat writes instead of refusing immediately. If pending writes block the handoff, the refusal reports their count and can include limited redacted run IDs.
More details
Released change. Thanks to @RomneyDa. See the external apps guide.
-
Clarifies when an update advances the Gateway installation but leaves a different shell CLI unchanged, and warns that the CLI may no longer match shared settings or plugins.
More details
Released change, #196342, Source. Thanks to @RomneyDa, @vincentkoc. Thanks to @DonnieFi for the report addressed by #196342. See the update guide.
-
Fixes status and Doctor recommending an ineffective reinstall when a service launcher override still points elsewhere.
More details
Released change, #196342, Source. Thanks to @RomneyDa, @vincentkoc. Thanks to @DonnieFi for the report addressed by #196342. See the update guide.
-
Corrects Doctor messages that could imply all repairs succeeded when a service repair was blocked, and identifies unsafe systemd directories in repair errors.
More details
Released change, #196342, Source. Thanks to @RomneyDa, @vincentkoc. Thanks to @DonnieFi for the report addressed by #196342. See the update guide.
-
Fixes Doctor hitting a missing-column error while preparing older conversation databases for the media migration.
More details
Released change. Thanks to @RomneyDa. See the Doctor migrations guide.
-
Fixes rejected plugin reload retries preventing another attempt after timed-out work and cleanup settle. Fully reverting the failed settings also lets other plugin changes apply.
More details
Released change. Thanks to @RomneyDa, @vincentkoc. See the restart recovery guide.
-
Fixes an update recovery error crashing while reporting that the original service could not be verified after a failed activation.
More details
Released change, #196474, Related issue #196350. Thanks to @RomneyDa.
-
Fixes Doctor rejecting retained Windows update recovery captures when equivalent paths use different spellings.
More details
Released change. Thanks to @RomneyDa. See the repair and recovery guide.
-
Fixes Windows updaters launched through short paths or junctions failing to load required code after a package is replaced. The updater keeps using the code and shared state it started with through final cleanup. The updater running the update must already contain this fix.
More details
Released change, Source. Thanks to @RomneyDa. See the update guide.
-
Fixes an agent remaining unavailable after Gateway startup when a model refresh overtakes its preparation. Recovery retries automatically, though persistent failures can still leave the agent unavailable.
More details
Released change, #172632, Issue #172628. Thanks to @dlmccoy, @steipete, @RomneyDa, @vincentkoc. See the guide.
-
Fixes update reports missing details about why the Gateway failed to start, including a signal or readiness timeout. Preserves warnings when cleanup cannot be confirmed after the startup check.
More details
Released change. Thanks to @RomneyDa, @vincentkoc. See the guide.
-
Fixes failed updates being unable to restore the previous release after Doctor or Gateway startup fails, when the state is safe to roll back. Checks explicitly local updates for Gateway startup failures even when the service is not restarted and no Gateway already owns the state. Remote and TLS Gateways are skipped.
More details
#197851, Released change. Thanks to @obviyus, @RomneyDa, @vincentkoc. See the guide.
-
Fixes update failure reports hiding recognized startup causes and useful configuration fields while keeping private values redacted.
More details
Released change, #178301, Issue #178094, Issue #178244. Thanks to @steipete, @RomneyDa and @vincentkoc, and to @piccolllo-vc and @KenAng-tech for the related reports. See the update guide.
-
Improves update and repair failure reports to distinguish an expired, replaced or missing state lease when those details are available.
More details
#183279, Released change. Thanks to @steipete, @RomneyDa, @vincentkoc, @ooiuuii. See the guide.
-
Prevents agent database cleanup from finishing before an in-progress session disk-usage scan has finished. If a failed scan’s worker cannot be stopped, cleanup fails instead of reporting completion and can be retried.
More details
Released change. Thanks to @RomneyDa, @vincentkoc.
-
Fixes Doctor starting to reclaim unused disk space from the session database while a related disk-usage scan is still running after importing older session history.
More details
Released change. Thanks to @RomneyDa and @vincentkoc.
-
Fixes messages and new sessions failing immediately after a Gateway restart while their agent is still preparing. The Control UI and TUI can wait up to five minutes for the agent. Preserves results for sends and session creation already dispatched when a view closes or is stopped.
More details
#174958, Released change. Thanks to @steipete, @RomneyDa, @vincentkoc. See the guide.
Messaging
-
Fixes Telegram callback buttons sent or edited by the message tool losing your choice or showing “This action is no longer available” when no plugin handles the tap. Unclaimed slash-prefixed choices reach the agent as data, not commands. Resend or replace buttons already delivered or queued to use the fix.
More details
#196217, released change. Thanks to @goslingmanagment, @dmsrg399, @obviyus, @RomneyDa and @vincentkoc, and to @kkernan for the report. See the Telegram rich messages guide.
-
Fixes some SQLite reads and writes getting stuck, and refreshes outdated connections used to read conversation history.
More details
#181661, Released change. Thanks to @steipete, @RomneyDa.
-
Fixes buffered Signal messages being skipped when the native HTTP connection reconnects.
More details
Recovery depends on retained daemon events and a usable cursor. #187667, Released change. Thanks to @hermanzeng, @obviyus, @RomneyDa. See the Signal guide.
-
Fixes new native approval requests using stale forwarding and account settings after a Gateway configuration reload.
More details
Released change. Thanks to @steipete and @RomneyDa.
-
Fixes WhatsApp final replies disappearing when unrelated settings reload during a turn.
More details
Released change. Thanks to @RomneyDa for the release adaptation. Original work by @mcaxtr in #168778. See the WhatsApp guide.
-
Prevents Tlon citations outside the current allowed scope from expanding and keeps discovered channels awaiting approval when their settings are saved.
More details
With auto-discovery enabled, or in the current owner’s one-to-one DM, citations to channels the bot can read can still expand. Group DMs do not receive the owner exception. Released change. Thanks to @RomneyDa, @mmaps, @steipete and @alkor2000 for the released change.
-
Fixes IRC channels with an empty sender list accepting messages through broader permissions. Existing empty lists now deny all senders; remove the override to inherit access or specify the intended senders.
More details
#196293, Released change. Thanks to @mmaps, @RomneyDa. See the IRC guide.
-
Reduces Slack Gateway stalls and delayed replies during long reasoning streams when Slack streaming is off.
More details
#196250, #196249, Released change. Thanks to @ptu-gh and @RomneyDa. Thanks to @obviyus for the Gateway investigation and testing. See the Slack guide.
-
Fixes recognized binary attachments such as SQLite databases being pasted into the model prompt as garbled text. The attachment remains available to tools.
More details
Released change, #198309, Issue #198289, #198309. Thanks to @Yun-0000, @obviyus, @mulderp82, @RomneyDa, @vincentkoc.
-
Fixes Feishu replies losing both the caption and attachment when block streaming is disabled.
More details
#191585, released change, superseding #145487. Thanks to @forrystudio, @RomneyDa, @vincentkoc. See the guide.
-
Fixes Telegram approval cards appearing in the root DM instead of the originating private topic when the requester is also an approver on the same bot account. Prevents duplicate final results in that topic for system-agent change approvals.
More details
#198038, Released change. Thanks to @dankarization, @obviyus, @RomneyDa, @vincentkoc. See the guide.
Memory
-
Reduces memory pressure when indexing and scanning large session transcripts for memory.
More details
Released change, #196480. Thanks to @RomneyDa. See the plugin SDK migration guide. Original work by @obviyus in #196480.
-
Changes memory indexing to prefer supplied content when both supplied and remote content are available.
More details
Released change, #196480. Thanks to @RomneyDa. See the plugin SDK migration guide. Original work by @obviyus in #196480.
-
Fixes some large SQLite memory-index rebuilds running out of worker memory and leaving vector search unavailable, including rebuilds after an index format upgrade.
More details
#196378, Released change. Thanks to @garrytan, @RomneyDa, @vincentkoc, @obviyus. See the guide.
Native Apps
-
Fixes the iOS app crashing when you allow speech recognition for chat dictation or in Device Settings.
More details
Released change, #196822. Thanks to @obviyus, @RomneyDa, @vincentkoc.
Models and Providers
-
Fixes Copilot setup verification accepting a changed SDK or packaged runtime. Restart OpenClaw after an SDK update before repeating verification.
More details
Verification uses the packaged stdio runtime. Released change. Thanks to @lifuyue and @RomneyDa.
-
Fixes detached Copilot one-shot attempts reusing a supplied previous session and removes their newly created session history during cleanup, after tool calls finish and compaction completes or its cleanup wait ends.
More details
Cleanup errors are reported and history may remain if cleanup fails. Released change. Thanks to @lifuyue and @RomneyDa.
-
Fixes retained Z.AI catalog models sending requests to the wrong endpoint when model discovery is unavailable.
More details
Released change. Thanks to @RomneyDa for the release adaptation. Original work by @jason-allen-oneal in #195390.
-
Fixes off-catalog model selection accepting a stale result when the model catalog changes during provider discovery.
More details
Released change. Thanks to @steipete and @RomneyDa.
-
Fixes chat sends asking you to retry when model access settings are migrated without changing access, or when unrestricted settings change form without adding restrictions. If access to the selected model is revoked or an alias changes which model is allowed during preparation, the send still asks you to retry.
More details
Released change. Thanks to @RomneyDa and @vincentkoc.
-
Fixes Claude CLI runs ignoring global execution restrictions when an agent has no explicit override.
More details
Released change. Thanks to @RomneyDa, @eleqtrizit and @pgondhi987 for the released change.
-
Fixes Claude CLI sessions stalling on the first turn when startup hooks emit a burst of events.
More details
Released change, adapting #173011. Thanks to @RomneyDa for the release adaptation, @ericcaiwx-star for the original repair, @steipete for strengthening it, and @sopa-gpters for the report.
-
Fixes a problem that could prevent a subagent’s completion reply from reaching its Claude CLI parent directly.
More details
#197434, Released change, release integration. Thanks to @holgergruenhagen and @slange-a11y, and to @ceo633 for the report. Thanks to @RomneyDa for the release integration.
-
Fixes delayed quota checks putting an auth profile back into cooldown after a newer request succeeded with the same credential.
More details
Released change. Thanks to @RomneyDa. Original work by @steipete in #176602.
-
Fixes Agents API continuations waiting through a long turn deadline when native session startup stays silent. The attempt fails after 120 seconds without a native event, or sooner if its deadline is shorter, while retaining the saved session for recovery.
More details
#196517, Released change. Thanks to @steipete, @RomneyDa.
-
Fixes cancelled Claude login checks leaving refresh locks behind, which could stall a later Claude CLI turn or trigger provider fallback.
More details
Released change, #197032, #189214, Related issue #156442. Thanks to @obviyus, @kir1ll, @RomneyDa.
-
Fixes affected Gemini requests being rejected when the built-in read tool is available.
More details
#196742, #196717, Released change. Thanks to @obviyus, @yixiaoer-520, @RomneyDa.
-
Fixes queued model credential refresh failing when the plugin call that changed the credential has already ended.
More details
Released change. Thanks to @RomneyDa and @vincentkoc. See the models guide.
-
Fixes conversations using the public OpenAI Responses API repeatedly failing after compaction with an item conflict, while preserving saved history. Adds recovery guidance for remaining item conflicts to run
/compactand retry, then choose another model and share Gateway logs with the operator if the error persists.More details
#198357, Released change. Thanks to @steipete, @RomneyDa, @vincentkoc. See the guide.
-
Reduces repeated work when preparing the model list and picker for large configured OpenAI model lists.
More details
Released change, #198837. Thanks to @RomneyDa, @vincentkoc. See the guide.
-
Fixes stopped or yielded Codex parent conversations losing the ability to continue or receive native child results while children are still running. Follow-ups can resume after the parent turn has stopped in the original eligible session.
More details
Released change. Thanks to @RomneyDa, @vincentkoc. See the guide.
-
Fixes model requests losing configured request handling or provider-specific compatibility settings when native plugins load additional transport code.
More details
Released change, Released change, Released change. Thanks to @RomneyDa, @vincentkoc. See the guide.
-
Fixes qualifying Anthropic setup-token requests being rejected when the advertised Claude Code version is too old by learning the required version and retrying once.
More details
Released change, adapting #196693. Thanks to @RomneyDa, @obviyus, @vincentkoc. See the guide.
-
Changes requests for unrecognized Claude models to use adaptive thinking and default sampling without assistant prefill, with newer refusal and replay rules.
More details
Released change. Thanks to @RomneyDa, @obviyus, @vincentkoc. See the guide.
-
Fixes Claude 5 fetch requests being limited to 32,000 output tokens when the catalog permits more.
More details
Released change. Thanks to @RomneyDa, @obviyus, @vincentkoc. See the guide.
-
Fixes Anthropic requests ignoring model-listed server compaction support.
More details
Released change. Thanks to @RomneyDa, @obviyus, @vincentkoc. See the guide.
-
Fixes Claude model switches discarding reusable older thinking while omitting thinking tied to the previous model; omitted thinking is not restored by switching back.
More details
Released change. Thanks to @RomneyDa, @obviyus, @vincentkoc. See the guide.
Automations and Scheduling
-
Fixes scheduled agent replies bound to their creating conversation going missing after message edits, branch switches or resets. It also fixes authenticated follow-ups scheduled into that same conversation being refused after those changes, while keeping them within the original operator’s permissions and current access. If the creating conversation is deleted, those follow-ups cannot resume it and delivery of replies bound to it reports failure.
More details
Explicit delivery destinations continue to use their configured routes.
Released change, adapting #199512 by @obviyus. Thanks to @RomneyDa for the release adaptation and @vincentkoc for the release coauthorship. See the cron delivery guide, payloads guide, and cron CLI reference.
-
Prevents sandboxed cron sessions from selecting a host CLI that cannot honor their isolation.
More details
Released change. Thanks to @RomneyDa, @eleqtrizit and @pgondhi987 for the released change.
-
Fixes replies from session-targeted cron jobs and system events failing to reach their conversations after the agent turn finishes.
More details
#196750, #196749, Released change. Thanks to @masatohoshino, @obviyus, @RomneyDa. See the cron delivery guide.
-
Prevents automatic cleanup from archiving or removing sessions used by enabled cron jobs. Ordinary session archiving waits when the job list is unavailable.
More details
Released change. Thanks to @RomneyDa. See the cron delivery guide.
-
Fixes recurring cron jobs being silently skipped when the scheduler wakes late while the Gateway is running and
cron.skipMissedJobsis enabled. Missed slots after downtime or a clock jump remain skipped.More details
#197484, Released change. Thanks to @Stronggenetics, @obviyus, @RomneyDa, @vincentkoc. See the guide.
-
Fixes running cron jobs being interrupted by a cron settings reload when cron remains enabled on the same store, so they can finish and deliver their results. A Gateway stop or restart still aborts active jobs.
More details
#197606, Released change, Issue #197591. Thanks to @obviyus, @wupingyuan81-oss, @RomneyDa, @vincentkoc. See the guide.
-
Keeps scheduled jobs queued while their agent prepares, leaving ready agents free to run.
More details
#174958, Released change. Thanks to @steipete, @RomneyDa and @vincentkoc. Thanks to @JoPaMu for the historical startup and cron investigation. See the guide.
Plugins and Integrations
-
Fixes a stuck plugin service startup blocking other services during a configuration reload.
More details
Released change. Thanks to @RomneyDa for the release adaptation. Original work by @steipete in #182331.
-
Fixes plugin auto-enable ignoring deny or disabled settings when plugin IDs use different capitalization, while preserving existing credentials.
More details
Released change. Thanks to @RomneyDa. Original work by @masatohoshino in #188846.
-
Fixes MCP tool-policy hooks missing available trusted CLI requester identity.
More details
Released change. Thanks to @RomneyDa, @mmaps, @steipete and @alkor2000 for the released change.
-
Excludes SDK testing and QA-only chunks from cloud node bootstrap packages unless a shipped runtime entry needs them.
More details
Released change. Thanks to @RomneyDa. Original work by @steipete in #181824.
-
Fixes other SQLite database writes being blocked while a plugin waits for permission to update its state.
More details
Released change. Thanks to @RomneyDa. Original work by @vincentkoc in #182117.
-
Fixes wiki ingestion hanging when a source or existing page is a special file, such as a FIFO. The bad path is reported so you can correct it and retry.
More details
#187621, Released change. Thanks to @Leon-SK668, @obviyus, @RomneyDa.
-
Fixes a plugin loading error when a host import and a CommonJS plugin load the same SDK entry at the same time.
More details
Released change. Thanks to @1052326311, @RomneyDa.
-
Adds session exports that plugin authors can process one page at a time, with a reset when a sliced SQLite transcript changes during export.
More details
Released change, #196480. Thanks to @RomneyDa. See the plugin SDK migration guide. Original work by @obviyus in #196480.
-
Fixes automatic plugin settings reloads blocking new work or disturbing work already in progress while affected work or cleanup is still running. The Control UI and configuration commands show when settings are saved but waiting to apply, and setup does not activate credentials until the settings apply.
More details
Released change. Thanks to @steipete and @RomneyDa. See the plugin architecture guide.
-
Stops memory work that has not yet selected a provider when an embedding provider is replaced, preventing it from starting after cleanup has begun. This can also restart work that would have used an unrelated provider.
More details
Released change. Thanks to @RomneyDa, @vincentkoc. See the plugin runtime guide.
-
Fixes plugin services inheriting an already-ended operation when starting or reloading, so they can register HTTP routes during startup and run scheduled callbacks afterward.
More details
Released change. Thanks to @RomneyDa, @vincentkoc.
-
Adds image inspection to policy plugins so they can block a run before those images reach the model.
More details
Released change, adapting #196337. Thanks to @RomneyDa and @joshavant.
-
Adds an API for context engines to read retained session messages and their context beyond recent history. Engines must adopt the API, and reads are limited to eligible messages within size limits.
More details
Released change. Thanks to @RomneyDa. See the plugin runtime guide. Original work by @rabsef-bicrym and @jalehman in #194362.
-
Fixes reconstructed messages from other sessions losing their source marker before reaching the model.
More details
Released change. Thanks to @RomneyDa. See the plugin runtime guide. Original work by @rabsef-bicrym and @jalehman in #194362.
-
Fixes memory export plugins receiving transcript messages from abandoned attempts when an export restarts.
More details
Released change. Thanks to @RomneyDa. See the plugin SDK migration guide.
-
Fixes retained plugin work, such as transcript capture cleanup, failing after the call that started it ends.
More details
This applies only to plugin instances the work has already acquired and only while the retained work remains open. It does not extend access to other instances. Released change, follow-up. Thanks to @RomneyDa and @vincentkoc. See the plugin runtime guide.
-
Fixes Telegram message hooks missing the session key that plugins need to associate sends with the originating session.
More details
Released change, #197519, Issue #197186, Issue #197669. Thanks to @lester076, @rodricksz4h5, @obviyus, @RomneyDa and @vincentkoc, and to @tommyjoseph for the report. See the guide.
-
Fixes Claude CLI commands waiting for approval without prompting the ACP client during an active turn.
More details
Released change, #197853, Issue #197829. Thanks to @gulagala001, @obviyus, @RomneyDa and @vincentkoc, and to @Borekobama for the report and @harjothkhara for testing a related smaller variant.
Security and Privacy
-
Blocks page-initiated document navigation to policy-blocked addresses between actions in connected browser tabs where OpenClaw has performed a guarded navigation or action, subject to the same-origin exception below.
More details
On direct connections, HTTP(S) documents can still navigate to the current frame or page origin between actions, even if a later policy change blocks that origin. This exception does not apply with an explicitly configured browser proxy. Actions in progress use their own checks; fetches and other subresources are outside this document-navigation check.
Released change, adapting #198452 by @daniel-lxs, who also provided the related report. Thanks to @RomneyDa for the release adaptation and @vincentkoc for the release coauthorship.
-
Fixes guarded browser navigation reporting a secondary browser abort instead of the underlying safety-check error that caused it.
More details
Released change. Thanks to @RomneyDa and @vincentkoc.
-
Fixes recognized private identifiers appearing as map keys in shareable triage archives.
More details
Released change. Thanks to @RomneyDa. Original work by @steipete in #175272. Thanks to @thefourcraft for the report.
-
Prevents queued worker credentials from being attached after the requester or worker placement loses permission.
More details
Released change. Thanks to @RomneyDa for the release adaptation. Original work by @sallyom and @vincentkoc in #159106.
-
Fixes whole-conversation Draft copies becoming Shared when no audience was chosen. If Draft creation is disabled, the copy is refused instead.
More details
#186828, Released change. Thanks to @shakkernerd, @RomneyDa. See the operator scopes.
-
Rejects reserved object keys in provider environment and MCP command-line assignments before saving configuration.
More details
Released change. Thanks to @RomneyDa. Original work by @wangmiao0668000666 in #191389.
-
Prevents active PostgreSQL connection variables from being inherited by host commands and sandbox processes, including variables with custom names.
More details
Released change. Thanks to @galiniliev, @RomneyDa.
-
Prevents a worker Stop request from cancelling a run after the requester loses permission.
More details
Released change. Thanks to @steipete and @RomneyDa.
-
Fixes captured model-only prompt guidance appearing in chat history and live message display metadata while preserving it for replay.
More details
Released change. Thanks to @RomneyDa. Original work by @obviyus in #173460.
-
Rejects malformed scoped Gateway paths that could be interpreted differently by access checks and request handlers.
More details
Released change. Thanks to @RomneyDa. Original work by @arpe1618 in #190863.
-
Fixes file and folder containment through nested or dangling symbolic links and rejects redirected parent folders for sibling outputs with the fs-safe 0.26.7 update. If you use
FS_SAFE_PYTHON_MODEorOPENCLAW_FS_SAFE_PYTHON_MODEwithout an explicit native policy, replace each with its correspondingFS_SAFE_NATIVE_MODEorOPENCLAW_FS_SAFE_NATIVE_MODEand keep the same value before updating. See the migration guidance.More details
See the fs-safe 0.26.7 release. Released change. Thanks to @RomneyDa.
-
Fixes Doctor importing stale plaintext credentials from a generated model catalog when the provider already has a stored SecretRef.
More details
#179802, Released change. Thanks to @steipete, @ensignwesley, @RomneyDa.
-
Fixes PostgreSQL writes proceeding after a required permission check refuses them and prevents potentially revoked access from being restored when a commit outcome is uncertain.
More details
#193842, Released change. Thanks to @vincentkoc, @RomneyDa.
-
Fixes Doctor cleanup removing restrictive skill and secrets egress allowlists when their last entry is removed.
More details
Released change. Thanks to @RomneyDa, @steipete. Original work by @yetval in #190068.
-
Fixes recognized secrets appearing in Gateway and Doctor diagnostic output and Gateway client logs.
More details
Released change. Thanks to @RomneyDa, @mmaps, @steipete and @alkor2000 for the released change.
-
Prevents paired-node screenshots and photos used for inspection from being attached to outgoing messages automatically. You can still share their saved paths explicitly.
More details
Released change. Thanks to @RomneyDa for the release adaptation. Original work by @eleqtrizit in #174814.
-
Limits project discovery and new source handoffs for operators without Gateway write scope to repositories confirmed public. Existing checkouts remain in place, and a failed visibility check can also hide a public repository.
More details
#194830, Released change. Thanks to @vyctorbrzezowski, @RomneyDa. See the operator scopes.
-
Fixes recognized multiline secret assignments appearing in exported OpenTelemetry attributes when model and tool content capture is enabled.
More details
Released change. Thanks to @RomneyDa. Original work by @mmaps in #194353.
-
Prevents recursively evaluated standalone assignments and legacy
$[...]arithmetic arguments from passing on allowlist approval alone, prompts for commands starting with aNAME+=valueenvironment assignment, and keeps Windows approvals for POSIX shell payloads one-time.More details
Released change. Thanks to @RomneyDa, @eleqtrizit and @pgondhi987 for the released change.
-
Fixes managed skills missing read-only protection on first sandbox launch and prevents protected nested mounts being moved aside. Eligible existing Docker and Podman containers can be recreated after a Gateway restart to add the protection while keeping host-backed workspace files.
More details
Automatic recreation requires a read-only container root and matching existing mounts. Released change. Thanks to @RomneyDa, @momothemage, @shakkernerd, @eleqtrizit. Original work by @steipete in #182412, #182789. Thanks to @akang1798 for the report.
-
Fixes Side chat questions and answers being shared between different authenticated profiles viewing the same session. Connections using a shared Gateway token or another nonpersonal identity still share a thread.
More details
Released change. Thanks to @RomneyDa for the release adaptation. Original work by @Patrick-Erichsen in #185390.
-
Fixes recognized credentials appearing in quoted compaction context saved and replayed in summaries. The original transcript remains unchanged.
More details
Released change. Thanks to @RomneyDa. See the compaction guide. Original work by @obviyus in #197349.
-
Stops iOS Gateway setup codes from being saved in app preferences and removes the old saved value when the updated app starts.
More details
#197049, Released change. Thanks to @obviyus, @RomneyDa.
-
Requires approval for delegated persistent changes when authority is restricted or unverified and no other grant applies. Adds a standing approval setting for permitted delegated changes. Full Access or standing approval can authorize changes that enable host code execution even when direct execution or editing tools are denied.
More details
Released change, adapting #196984, and backport prerequisite. Thanks to @RomneyDa and @joshavant. See the permission modes guide.
-
Fixes delegated OpenClaw configuration changes relying on an outdated automatic approval setting when settings change during a run or while a tool is being prepared. A revoked
autoApproveChangesgrant is rechecked even when the run keeps earlier model settings.More details
Released change, follow-up. Thanks to @RomneyDa and @vincentkoc. See the permission modes guide.
-
Fixes retained legacy exec approvals blocking a valid current policy and Node host startup. Retained legacy input still blocks approval operations when no current policy exists, and malformed policies fail closed.
More details
Released change, #196401. Thanks to @RomneyDa. See the Doctor guide. Original work by @steipete in #196401.
-
Prevents OpenShell mirror sync from copying nested Git metadata between host and sandbox or deleting it on the host. Sandbox-created Git metadata is removed before the next mirrored command. Download conflicting files with the OpenShell CLI before another mirrored command, and use remote mode for persistent sandbox repositories initialized or cloned there.
More details
Released change, #196747, #171669. Thanks to @mmaps and @RomneyDa, and to @yetval for prior filesystem evidence in unmerged #164778. See the OpenShell guide.
-
Prevents sandbox-written workspace setup files from blocking Gateway startup and configuration reloads for all agents. A sandbox with legacy setup state can still be blocked until Doctor repairs it.
More details
Released change, #197248. Thanks to @mmaps, @RomneyDa. See the Doctor migrations guide.
-
Prevents saving a skill API key from replacing an environment-backed or referenced key with plaintext. Keys with an unknown source remain read-only, and a failed configuration read can be retried without losing pending Settings drafts.
More details
Released change, #197319. Thanks to @RomneyDa. See the skills guide. Original work by @mmaps in #197319.
-
Prevents deferred tools from accepting inputs that their own rules reject when a model provider receives a simplified tool definition.
More details
Released change. Thanks to @RomneyDa, @vincentkoc. See the tool search guide.
-
Strengthens the checks used to clean unusual markup and namespaces in Control UI messages, link-reader documents and Mermaid diagrams before they are displayed.
More details
Released change, adapting the DOMPurify update from #178071. Thanks to @RomneyDa and @vincentkoc, and to @steipete for the original work. See the guide.
Quality-of-Life Improvements
-
Prevents an old Control UI Files draft from overwriting a replacement managed workspace and keeps the unsaved draft when the workspace source changes.
More details
#178336, Released change. Thanks to @shakkernerd, @RomneyDa.
-
Fixes session reads rejecting valid deeply nested transcript data and preserves unusual text characters when sessions are reopened.
More details
Released change. Thanks to @jondihaynes, @ylcn91, @RomneyDa, @steipete and @obviyus.
-
Fixes an accepted prompt disappearing from conversation history when the Gateway restarts before its first response.
More details
#187157, Released change. Thanks to @shakkernerd, @RomneyDa.
-
Changes session search indexing to replace NUL characters with spaces while preserving the original transcript.
More details
Released change. Thanks to @ylcn91, @RomneyDa and @obviyus.
-
Fixes successful local message-tool replies missing from the source conversation history and records the content that was actually delivered.
More details
Released change. Thanks to @RomneyDa for the release adaptation. Original work by @MertBasar0 in #187795.
-
Fixes supervised commands losing their exit result or final output when required Gateway cleanup fails during shutdown.
More details
Released change. Thanks to @RomneyDa. Original work by @steipete in #188891.
-
Fixes some large session exports for memory running out of worker memory while rendering the conversation.
More details
Released change. Thanks to @RomneyDa. Original work by @ericcaiwx-star and @obviyus in #196158. Thanks to @garrytan for the report.
-
Prevents incomplete summaries from replacing conversation history when compaction reaches the model output limit. Summarized branch navigation also leaves the current branch unchanged if its summary is truncated.
More details
Released change. Thanks to @zhangguiping-xydt and @RomneyDa. See the compaction guide.
-
Fixes uncertainty notices waiting for another message after a final reply cannot be confirmed. The notice explains that the earlier reply may not have arrived and will not be resent automatically. The notice can also fail if the channel remains unavailable.
More details
Released change. Thanks to @RomneyDa for the release adaptation. Original work by @jason-allen-oneal in #195400.
-
Fixes completed subagent and follow-up replies being held by a descendant whose own delivery is already suspended.
More details
#197323, #197305, Released change. Thanks to @wr-web, @obviyus, @xRyann2255, @RomneyDa. See the subagents guide.
-
Fixes acknowledged queued inputs being treated as cancelled during a graceful Gateway restart. Their original authenticated source can retry with the same idempotency key; non-main sessions are not replayed automatically.
More details
Released change. Thanks to @RomneyDa. See the restart recovery guide.
-
Fixes Gateway memory remaining allocated after completed subagents that do not send completion messages are cleaned up.
More details
Released change, #196672. Thanks to @steipete, @RomneyDa. See the subagents guide.
-
Prevents automatic idle or daily session resets from rolling over a conversation while its reply is still running. Rejects a request with a retry message if the reset decision changes while the session is being saved.
More details
Released change, Source. Thanks to @RomneyDa. See the session configuration guide. Original work by @Jo2234 and @obviyus in #197345. Thanks to @abacha for the report.
-
Fixes message-only memory export retries missing the reset needed to discard earlier messages when a conversation changes during export.
More details
Released change. Thanks to @RomneyDa. See the memory guide.
-
Fixes agent runs failing with "Plugin invocation scope is closed" when they make plugin calls after dispatch ends.
More details
Released change. Thanks to @RomneyDa and @vincentkoc.
-
Fixes subagent completion replies, announcements, cleanup and requester wake-ups being interrupted by plugin call failures after the call that started the subagent has ended.
More details
Released change. Thanks to @RomneyDa and @vincentkoc. See the subagents guide.
-
Fixes queued follow-up and collected messages failing to use the current provider after the previous plugin call ends and its provider is replaced.
More details
Released change. Thanks to @RomneyDa and @vincentkoc. See the queue guide.
-
Fixes failed subagent results going missing when the parent turn ends before the failure is reported.
More details
Released change, #197658, #197658. Thanks to @obviyus, @RomneyDa and @vincentkoc, and to @dankarization for the related investigation. See the guide.
Other Bug Fixes
-
Fixes false “unsaved work” errors caused by generated skill folders when guests prepare their existing pull request in a fresh workspace.
More details
#185777, Released change. Thanks to @shakkernerd and @RomneyDa.
-
Fixes SQLite writes being reported as ordinary errors when the write may have succeeded but its result could not be returned.
More details
#177321, Released change. Thanks to @steipete, @RomneyDa.
-
Stops Code Mode cells from running later tools when a tool reply cannot be decoded.
More details
Released change. Thanks to @RomneyDa. Original work by @vincentkoc in #188023.
-
Fixes Code Mode failing to collect an existing process result at the suspended-run limit when using
awaitResults: true.More details
Released change. Thanks to @RomneyDa, @vincentkoc. See the Code Mode troubleshooting guide.
-
Fixes failed Code Mode calls blocking an agent and queued messages while another call remains unanswered. Tool effects already performed are not undone. Prevents tool heartbeats from keeping a silent model stream open indefinitely.
More details
#196435, Released change. Thanks to @steipete, @RomneyDa. See the Code Mode guide.
-
Fixes the Gateway freezing while displaying tool progress for a deeply nested path followed by a space.
More details
Released change, #196282. Thanks to @leizhenzhang, @obviyus, @RomneyDa.
-
Fixes completed SQLite reads holding up database checkpoints, which can cause the write-ahead log to grow. Prevents an old read from restarting a newer query.
More details
Released change, #190242, superseding #184076, and issue #143524. Thanks to @ava-perry, @RomneyDa, @vincentkoc. See the guide.