Skip to main content

​
v2026.9.4

OpenClaw 2026.9.4 makes plugins and skills easier to find, and lets you turn past conversations into reusable skills through a chat you can steer. It also adds GPT Image 2.5 support, more control over cloud sessions, and interactive questions in the terminal, alongside fixes for updates, memory, and messaging. Release scale: 1,558 pull requests, 20 direct commits, and 293 contributors.

​
Installation and Onboarding

OpenClaw can help you get past a Node version problem without replacing your computer’s existing Node installation. New-agent setup also works with securely saved API keys, and the instructions for connecting a phone or removing OpenClaw now fill in steps that were missing.
If OpenClaw will not start because of Node, the software it runs on, it can now find a compatible copy already on your computer or offer to install one just for OpenClaw. You approve the first installation; after that, OpenClaw can reuse it automatically. This gives you a way past the startup problem without changing Node for your other apps. Some systems, including Alpine Linux, still need manual installation. If you cannot start normally, basic help and troubleshooting commands remain available on some older Node versions.If the background service is still using an old Node after an upgrade, openclaw gateway install can now switch it to an available supported version. This is a separate repair from getting the command line working.
Mac installation no longer gets stuck in the affected Homebrew Bash setups, and people building their own Docker image can get past the missing-file error that stopped installation. Linux instructions now explain which user account should run OpenClaw so it finds the right settings, while installation help makes clear when it will start or restart the background service. For people who manage their own terminal settings, supported command-completion entries are also preserved instead of being added again.
A new agent can now use an API key you have already saved in OpenClaw’s protected storage, fixing a setup failure that left ordinary chats working but new-agent setup stuck. Moonshot’s China API-key setup also stops asking you to reinstall a provider that is already installed, and browser sign-in explains when you have pasted a phone setup code into the wrong place.An empty background memory check no longer makes a fresh Claw think setup is already finished, so you still get its first naming conversation.
If you have several agents, guided messaging setup now lets you choose whose workspace to configure before separately choosing who should receive the messages. You can also choose an agent when managing setup recommendations, so you work with the intended workspace. Setup scripts still need an explicit --agent choice when no default is configured. The device setup instructions also explain that connecting a device and approving the commands it can run are separate steps.
Phone setup help now points you to openclaw qr, and the Android instructions explain how to connect to the computer running OpenClaw and set up the login your phone needs. They cover both a trusted home network and Tailscale. Use HTTPS for browser access when possible, because HTTP does not encrypt your login details or messages.
An error at the very end of an import no longer has to hide work that already finished. OpenClaw preserves completed memory-import results when the final cleanup fails, and can recognize a retry of the same request without copying everything again. If an error says the import completed, check the report and imported files before starting over. That retry protection is temporary and does not survive restarting OpenClaw.
Removing OpenClaw now comes with instructions that match how you installed it, including Git and installations kept in their own folder. The uninstall guide explains which files and launchers to remove and which may be shared with other tools. Remove the background service first, and save any settings or workspaces you want to keep before deleting files.
The first-run help separates getting started from choosing a provider or hosting service, so you can find the answer for the step you are on. Setup examples also fill in missing tools, credentials, and commands, including how to match the Mac app with the right command-line version. When you ask the setup agent to inspect a setting, it can now find names containing dots or brackets that it previously reported as missing, while keeping secret values hidden.

​
The New Web UI

In the web app, you can preview an earlier message before jumping back to it, bring a quote into your next reply, and see what Codex’s helper agents did while working on your request. There are also fixes for the interruptions that make a conversation harder to follow, from disappearing image previews to finished tasks that still look busy.
Your conversation list now keeps the names you choose and avoids changing them back when an older update arrives. Chats can get a useful title even when automatic naming fails, making them easier to recognize later. If the app tries to reopen a chat you deleted, it takes you to a usable conversation instead of leaving you stuck; it does not recover the deleted messages.
A slim message navigator on the left lets you preview a message before jumping to it, so you can find a particular answer without scrolling through the whole conversation. In shared chats, previews also show who wrote the message when that information is available. It appears on desktop when there is enough room and covers the messages already loaded.
When you reload the app, it can show a saved copy of your conversation and sidebar while it reconnects, so you have something to read during the wait. This works with remembered Gateway tokens or paired-device sign-ins; other sign-in methods still show the connection screen. Sending messages and other actions wait for the connection.If you use the same web address to connect to several Gateways, a conversation saved from another Gateway can briefly appear. Wait for the live conversation to load before relying on that view.
Highlight a passage and choose Add to chat to put an editable quote into your draft without losing what you have already written. Quotes include up to 300 characters, and nothing is sent until you send the message yourself. Long drafts are also easier to edit because the line you are typing no longer fades out at the edge of the message box.
Images you send stay visible while the conversation updates, and scrolling back to a recently loaded image avoids another loading placeholder when its preview is still available. Documents sit above the text that goes with them, making files easier to spot in a busy conversation. Opening a file now shows that it is loading instead of briefly claiming the preview is unavailable.
Selected videos can now show a still image in the message box, so you can tell similar recordings apart before sending one. The preview is made on your device without uploading the video. Sent videos also have previews in the conversation that open in Files when selected; if your browser cannot make a preview, the usual file card or icon remains.
Files keeps search and filters within reach while you scroll through long lists. You can find files the agent changed, read, or created, and a failed preview explains why the file could not open. Phone previews also leave more room for the document while keeping its name visible.If an agent file changes after you open it for editing, Save now warns you and keeps your draft. Choose Reload to use the newer file, or Overwrite to replace it with your version. Overwrite discards the newer changes rather than combining them, and the warning cannot catch every simultaneous edit made by another program.
Open a Codex helper agent in Tasks to read its conversation, see its work, and load earlier messages. You can look at a file and return without making the conversation load again. Newly created tasks also stay readable as the main chat continues, although some older tasks still lack readable history, and resetting the chat or changing Codex accounts can prevent access.
Finished and cancelled helper tasks stop showing their old working messages, while their summaries of changed files remain available. Long reports stay readable without hiding the button back to the conversation that started the task, and short progress updates no longer show stray formatting symbols. Agents also get clearer instructions to skip progress cards for quick questions and save them for substantial work.
Click the Dashboard tab to give it the full task area, then click again to return to your previous layout. Your draft and anything entered into the dashboard stay in place. Dashboard progress also follows the conversation it belongs to, so selecting a different agent elsewhere does not make running work look paused.
When you use Claude Code, Codex, OpenCode, or Pi in a terminal on a paired computer, choosing or dropping a file now puts its location into the terminal for you to use. You can edit it before pressing Enter. Reloading while several terminal tabs are reconnecting also keeps tabs that previously disappeared from the list.If uploads remain blocked after a crash, follow the terminal recovery steps. Stop all OpenClaw host processes sharing the upload folder on the computer running that terminal before removing the lock named in the error.
In Settings, removing an earlier list item no longer throws away unfinished text in another item. Correct and save that text before leaving the page, since unfinished drafts are not kept after a reload. Number buttons also stop skipping the first allowed value, and Web search clearly stays off when the person managing OpenClaw has disabled it.
The mention picker keeps suggestions available while you refine a search, making it easier to choose someone for a shared chat. Agent pictures now load in the picker and Settings on token-protected connections, and changing an agent’s name or emoji updates an open chat without losing your draft.
Usage keeps conversation details with the correct agent and counts the conversations actually shown in the list, with the total displayed separately. If you run OpenClaw on Linux with Bun, System busyness can show CPU and delay readings again instead of leaving permanent dashes. A dash just after starting still means the first reading is not ready.
If a reply has finished but the screen still shows Stop, clicking it can bring back the completed reply and Send button without a page reload. On phones, failed messages keep Retry or Discard visible when those actions are available, and a request that times out before replying can leave an explanation that remains after reloading. Before sending again, check whether the first request already did the work, since a connection problem can leave delivery uncertain.
Opening the command palette no longer blurs the conversation behind it, so you can refer to a message while choosing an action. Copying a reply gives a compact green check, GitHub links avoid empty preview popups, and the first reply no longer triggers confetti. On Mac, use Command+B, Command+F, and Command+K for the sidebar, conversation search, and command palette; the Control versions are left for text editing. Windows and Linux keep Control for these shortcuts.
OpenClaw can catch some coding mistakes in widgets, the small interactive apps an agent can place in chat, before they appear. If a recent card fails while you use it, the web app can show a Script error notice and tell the agent what happened, giving it a chance to help you recover.
More labels and help text are translated in Settings, plugin browsing, Files, and cloud-worker controls. Settings also uses available translations when you search for a field, making it easier to find an option in your chosen language.

​
Updates and Maintenance

Updating OpenClaw gets fixes for stalled upgrades, plugins left behind, and confusing recovery messages. There are also improvements to bringing older conversations and settings forward, backing up the files you need, and cleaning up history without getting in the way of other work.
Updates from recent stable releases can get past failures that prevented them from finishing or left OpenClaw stopped afterward. Updating from the browser also gets a startup fix, and active conversations are less likely to hold up an update. Some fixes only help after the newer updater is installed, so if your current version is stuck, follow the update guidance rather than repeatedly trying the same failing command.
If you run OpenClaw from a Git checkout, updates can recover when an upstream release tag changes, and Windows gets a fix for updates failing before they begin. Status now warns when a failed update check left its information out of date. That warning clears after the updater next fetches successfully. If your checkout has several possible update sources, the development-channel guide explains how to choose one.
When an update fails, OpenClaw can put the previous app version back if it still works with your current settings and data. If that cannot be checked, recovery stops and leaves the available recovery files in place for inspection. An interrupted update may still need manual help. Rolling back the app does not restore your data, so make and verify a backup before updating.
An old “update in progress” message can stop blocking settings changes and account sign-in even when the update itself is no longer running. openclaw update repair can clear eligible stuck updates once it confirms the updater has stopped and OpenClaw is working. Some older requests that never got started clear automatically after a day, provided nothing still needs recovery. They remain in history as failed attempts, with guidance to retry.
Updates on larger installations now give Doctor, OpenClaw’s repair tool, time to finish work that used to be cut short. Some failures to remove disposable files become warnings with cleanup advice, while problems that prevent the updated app from working still stop the update. A stuck repair can keep waiting unless you set a limit, and stopping the update may leave repair work running, so use the repair guidance if progress stops.
If an update fails but the previous installation is restored and checked, the menu defaults to Exit. You can choose to investigate or preview a report before sending it. For other interactive update failures, the repair prompt tells you which AI agent will run and that it uses your account allowance or paid tokens. You can decline, but pressing Enter accepts and leaving the prompt unanswered starts the agent after 30 seconds.Claude-assisted repairs start without project customizations that could delay them. If your repair needs those custom tools or instructions, use the manual command provided instead. The repair-agent guide explains the choices and Claude requirements.
Update progress and the final report now agree about what happened, including when an update failed but the previous version was successfully restored. Warnings keep useful repair advice and identify leftover backup files, and a problem reading history no longer hides other update guidance. If an update stalls, more of its troubleshooting information is kept. You can inspect the update report, or run openclaw doctor afterward for checks that are not needed during the update itself.
Running openclaw update now checks eligible plugins even if the main app is already up to date, including after you updated it separately. When nothing changes, it skips unnecessary follow-up work. If you deliberately kept a plugin at a particular version, that choice stays in place and notices of newer official releases explain how to change it. Plugin updates may need a restart, and plugin downloads during an app update can extend the time OpenClaw is unavailable.
Doctor can bring more older settings and saved accounts forward without losing your model choices or asking you to edit configuration files by hand. It can also recognize a moved workspace and preserve its setup history, while leaving uncertain matches for review.If an older Claude conversation has not been converted to the current format, run openclaw doctor --fix before resuming it. The Doctor checks guide explains that step and the cases that still need manual help.
Importing older Codex conversations now keeps assistant replies that could previously go missing, in their original order. If you already imported an affected conversation, recovery can add missing replies, but it adds them at the end rather than rebuilding the original conversation order. Repairs to attachment information also survive when saved conversation archives are recreated. If a backup-linked file blocks an import, the error points to the file and the safe recovery steps.
If you split your settings across several files, a full backup now includes the files your main configuration refers to, even outside the usual backup folders. It stops with repair guidance if those files cannot be captured completely. The --only-config option still saves just the main file.Backups also handle a disappearing temporary database file when its contents are already safely captured. A backup does not freeze the whole installation at one instant, so changes made while it runs may not all be included. Treat the archive as sensitive because settings files can contain secrets, and follow the backup guide when checking or restoring it.
Installations with a lot of saved data get more time to finish database checks, and some checks need less temporary disk space and avoid repeated work. This helps when maintenance previously stopped before it could finish, while keeping the checks for damaged data. If something goes wrong, errors give more useful clues about where the check stopped and when Doctor may be able to help.
Deleting sessions and cleaning up old conversations are less likely to pause other work while OpenClaw checks its saved data. Cleanup also checks again before deleting more history, so it can stop when enough space has been freed and leave newly protected conversations alone. Your retention settings still decide what can be removed.
Linux users get fixes for restarts that stayed stuck and for conflicting settings that made OpenClaw restart over and over. Those conflicts now leave an error to fix before starting again. Shutdown also gives plugins and voice connections time to finish their cleanup and final writes, though a stuck plugin can still delay it.Reconnecting avoids repeating unnecessary health checks, while fresh checks remain available when needed.
If another tool manages your OpenClaw settings, a new read-only configuration option keeps OpenClaw from overwriting them, including during automatic recovery. Set OPENCLAW_CONFIG_READONLY=1 for both the running service and terminal commands. Conversations and other working data still need somewhere writable, and model changes can apply to the current session without changing the saved default.Settings edits after creating or changing an agent also avoid false “changed since last load” errors. Saving and applying settings remain separate, so if you turned automatic reload off, restart OpenClaw to use the changes.
For Windows installations using OpenClaw’s standard launcher, stopping the scheduled Gateway task now stops the work it started too, allowing a fresh start without a leftover copy getting in the way. Early startup failures also reach the log. Custom executable wrappers or extra command-shell layers can still prevent startup. Slow Windows service checks get another chance during updates, but a check that keeps hanging can make the wait much longer; if the service cannot be checked, the update stops before replacing the app.
The monitoring guide makes it easier to find help setting up monitoring, understanding what information it captures, or using that information in dashboards. The instructions are organized around those tasks, with existing links still leading to the right place.

​
Messaging

This update fixes several ways a conversation could lose its words, its reply, or its place. Long answers keep their formatting, Feishu document tools work again in affected chats, and Talk can bring an answer back after handing work to another agent.
Messages arriving while OpenClaw is paused now wait until it is ready again, and unfinished replies are no longer mistaken for completed sends. If you show detailed progress, newer work stays visible instead of being crowded out by old command failures. Missing-reply notices also give a clearer explanation and a reference your administrator can use to investigate.
Long replies break at more natural places, and code keeps its formatting when it continues into another message. Code that arrived successfully no longer causes the whole answer to be repeated at the end. Quoted examples also keep their spacing, making them easier to read and copy.
If you enable Telegram rich messages, links inside lists, quotes, and expandable sections stay clickable and keep their styling. Code examples keep their spacing in messages and photo captions too. Sticker descriptions now use your chosen image model and reach your Claw in time for its first reply.
Telegram recognizes mentions that tag the bot’s account in messages and photo captions. If you have native commands turned off, you can write /think high on one line and your request below it without losing the request. The /think menu also shows the setting for the agent handling that chat, and canceled attempts no longer leave a misleading “stalled” reaction afterward.
If you connect Telegram through a proxy and it refuses the connection before a reply is sent, OpenClaw can keep that reply for another attempt when the connection returns. It still avoids resending messages whose delivery is uncertain, so this does not automatically recover every previously failed reply.
In Discord channels where your Claw can already read messages without being mentioned, voice notes now keep their spoken words in the conversation. Attachment captions also keep repeated instructions, and progress updates remain visible if the final answer cannot be sent.Requests such as checking OpenClaw’s system status can work again from server messages when you have permission, fixing a case where the request was refused despite the tool being available.
Slack can show its configured reaction when it picks up your mention even if you have ongoing status reactions turned off. Messages containing Windows paths also keep their formatting without accidentally turning the text that follows into code.
If a separately installed Feishu plugin stopped letting your Claw work with documents, wikis, Drive, or Bitable from chat, those tools are available again with your existing permissions. Changes to tool and account settings now apply to new requests without restarting OpenClaw, while work already underway keeps its earlier settings. Your Claw also receives the emphasis you put in formatted messages, links, and mentions.
Matrix keeps message text that could previously be mistaken for an attachment and disappear from quoted replies or conversation history. A mention marked as plain text also stays plain when earlier code formatting is unfinished. In the terminal, reaction lists show which emoji people used instead of leaving the label blank.
Your Claw now reads punctuation and symbols correctly in Teams messages supplied as HTML, instead of receiving formatting codes in their place. For example, a copyright symbol arrives as © rather than the code used to represent it.
If you set Signal to quote your message only on the first reply, that quote now stays available when an earlier send fails. The first reply that actually arrives shows what it is answering.
If you use a custom label at the start of your Claw’s WhatsApp replies, it no longer gets added to messages coming from other people. Your outgoing labels still work, and incoming messages keep what the sender wrote.
iMessage conversations can show familiar contact and group names, such as Alice or Family, when that information is available. Names you assigned yourself stay in place, making it easier to find the right chat without sorting through phone numbers or technical IDs.
For Mattermost setups that store login tokens separately from account settings, editing one account no longer unnecessarily reconnects the others; settings shared by all accounts can still restart them together. A replacement progress message also stays visible while the old one is being removed, even when both contain the same plan.
Replies from a resumed Google Chat conversation can use its saved destination correctly again. This fixes a case where changing capital letters in the destination caused a permission error, even though your Claw was allowed to send there.
When diagnostic recording is enabled, a successful Nextcloud Talk reaction no longer has to wait for that recording to finish before reporting success. The recording may be incomplete, but it no longer holds up the result.
If a Tlon message cannot send because setup is incomplete, the error now tells you which account needs attention and which settings are missing. You can go straight to the relevant part of Tlon setup without guessing, and the error does not reveal your login code.
If you use multiple Zalo Personal accounts with an older setup, openclaw doctor --fix --non-interactive can now complete a settings repair it previously skipped. Your account profiles and explicit choices about who may message the bot stay intact.
Account lists and health checks more accurately show which accounts your agents use. Removing an account now reports when nothing was removed, and message-command errors explain how to choose an agent when one is needed.If you use scripts, leave out --account and the directory option --limit when you want their defaults. Passing an empty value now causes an error; a supplied limit must be a positive whole number. Removal scripts must also handle an error when nothing was removed.
Talk can keep listening while another agent works on your request, then bring the finished answer back to the same voice conversation as well as chat. This also works when the task needs more than one round of help. Keep that voice conversation open to receive the spoken result; ending it or restarting OpenClaw can leave the answer available only through chat.
When you interrupt your Claw on a phone call, it now uses the phone service’s playback confirmation to avoid treating an answer you never heard as already spoken. GPT-Live calls through OpenClaw can also continue after an occasional recoverable audio error, though you may hear a brief gap. Repeated failures can still end the call; see the Voice Call guide for setup and troubleshooting.

​
Memory

Memory fixes help your Claw find remembered details and return to older conversations. There is also clearer guidance when memories cannot be saved, or when deleting a conversation leaves its memories behind.
Asking for a short list of memories no longer drops a better match just because you requested fewer results. Your Claw can also keep looking when an optional search gets stuck, and a local memory-search service gets time to start before OpenClaw gives up on it. These fixes help remembered information reach your conversation without unnecessary repair steps.
OpenClaw can wait through temporary limits from your memory-search provider and use the right model when it switches to your configured backup. It also stops telling you to rebuild working memory search just because the main provider has recovered. If you change to an incompatible model, you still need to rebuild with openclaw memory index --force.Mac users running OpenClaw with Bun can enable native memory search with brew install sqlite. If you previously used a custom SQLite startup script, replace it with the OPENCLAW_SQLITE_LIBRARY setting to avoid a startup failure.
With the default settings, a note no longer becomes long-term memory simply because OpenClaw repeatedly encounters it during background scans. It must also turn up in different memory searches, helping keep incidental notes from being saved just because they keep appearing. After updating, some older notes may take longer to qualify, while memories already saved stay in place.
Session search can now open saved messages from before a conversation reset, taking you back to the original discussion instead of an empty page. Fixes also help queued replies continue after OpenClaw shortens a long conversation, so you can pick up where you left off.
Deleting a conversation now tells you when its saved history can still appear in memory search and shows the command for forgetting those memories. Forgetting remains a separate step, which you need to run on the computer hosting OpenClaw, even if you deleted the conversation remotely.When an automatic memory save lacks permission to write, the logs now name the file it could not save. They also explain more reasons why Active Memory skipped a search, making missing memories easier to investigate. These are clearer warnings, so a reported save problem still needs attention.
If you use a plugin to manage conversation history, summaries are less likely to fail because a connection closed too soon. OpenClaw also avoids unexpectedly switching to a backup plugin while a turn starts. Cancelling can still stop further changes to the conversation while the plugin finishes closing its connections.
If another tool adds memories to your LanceDB memory extension, your Claw can find them on its next search without restarting OpenClaw. You can also list or delete those newly added memories without first having to make a change from inside OpenClaw.

​
Skills

Skills give your Claw reusable instructions for the work you do together. This update makes them easier to find and set up, and lets you watch and guide your Claw as it looks for ways to improve skills using past conversations.
The Skills page now searches your installed skills and ClawHub together, so you can see what you already have and what you can add without checking separate places. Cards show which skills are ready or need setup, and switching agents shows the skills available to that agent. You can also browse popular skills without typing a search, while clearer titles and readable descriptions make it easier to choose.
When importing a skill, you can now see which files you selected and clear them before continuing. Skill windows fit the screen and keep Close within reach, making longer instructions easier to read on a phone. Installations that need more time to download or set up supporting software can also finish without being cut off as early.
Choose “Learn from past conversations” in Skill Workshop to open a chat where you can follow the work, add direction, or stop it. Auto lets your Claw apply improvements based on what it finds, while Propose leaves suggestions for you to approve. Starting this yourself does not turn on automatic self-learning, and normal model charges and access permissions still apply.Old unfinished history scans cannot resume in the new flow, so start a new learning chat instead. Your installed skills and existing proposals stay available. Workshop also explains when automatic weekly reviews are paused because scheduling is off and takes you to the setting.
After restarting OpenClaw, you can continue an existing conversation with updated local skill files on its next turn. Skill lists also pick up newly installed or repaired files more reliably. Skills selected from a managed library keep their chosen version until you refresh them.New Workshop changes preserve the full descriptions that help your Claw recognize when to use a skill, instead of replacing them with short change labels. This protects future edits, but does not restore descriptions already lost or repair older pending proposals.
If your Claw works on a cloud or SSH-connected computer, it now spends less time copying skills before starting, especially when you have lots of small skill files. Cloud workers can also use supported links between instruction files within the same local skill, so you do not have to duplicate those files just to get the skill working there.
Doctor can repair several known Workshop problems that stopped upgrades or prevented OpenClaw from starting. Run openclaw doctor --fix when startup asks for it. Proposals and backups that need you to identify their agent remain available for review while other repairs continue. Some older review history may be discarded if its agent cannot be identified, and affected proposals may need reviewing again.
The bundled video-frames skill and its helper script have been removed. If you use them to save still images from videos, switch to ffmpeg directly or install a separate frame-extraction skill. OpenClaw’s video understanding can analyze a video, but does not produce those image files for you, and existing workflows are not converted automatically.

​
Native Apps

The Android app helps you keep your place while reading, see what your Claw has done, and reply from notifications. There are also clearer ways to reconnect the mobile apps, fixes for Mac setup, and changes that keep a sleeping Linux desktop from interrupting OpenClaw on another computer.
The Android app keeps your place while a reply arrives or you resize the window. Read earlier messages without being pulled away, then use Jump to latest to catch up. You can also open tool details to see what your Claw did, while dashboard chats tuck completed steps away and leave the answer visible.Search and the sidebar show which conversation is busy, and a failed refresh keeps the conversations you already loaded. On folding phones, model settings and background tasks stay clear of the fold.
Android notification replies reconnect to the saved computer for that conversation when needed. The notification now tells you whether a reply is queued, failed, or has an unknown status. Queued confirms that the app saved your reply for sending; it does not confirm delivery. If the result is unknown, open the conversation before sending again. You may still need to approve a connection-security prompt before the reply can proceed.
Android waits for your selected files to finish loading before enabling Send, so your message cannot go out with its caption but without the attachment. If you cancel a file or it fails to load, the caption stays in place for you to keep working on.
When Android or iPhone cannot reach the computer running OpenClaw, the app now explains the problem and offers Retry. iPhone retries the connection you were setting up, and connections that appear to use Tailscale show relevant setup help. Android may need a moment to finish closing the previous connection before trying again.Android setup also makes the selected security option available to accessibility tools. When your Claw checks a phone or Apple Watch battery, the result now includes a clear percentage, helping avoid confusion between a full charge and one percent.
You can now reach the official App Store listing straight from the iPhone setup guide. If you install Android outside Google Play, the Android guide helps you find the app download and check that the file is intact. Downloads are attached to selected releases and may appear later than the release itself.
The iPhone app now makes clear how many conversations are showing and how many there are in total, including in its translated messages. If some have not loaded yet, use Refresh to load the rest.
Choosing Set up later or clearing a connection on Mac now stays saved instead of bringing back the old connection when you reopen the app. If saving fails, the app tells you. Password-protected connections also show where to enter the password and try again. Older OpenClaw installations on the other computer need updating to accept a password in the field labeled Gateway token.If you prepare Mac setup from the command line, you can now read a password or token from a private file without putting it in the command itself. Setup also respects the profile you chose instead of changing the default profile. The remote setup guide explains these options and the separate command to use when the app is already running.
When you are using a side-panel tab in the Mac app, Command-W now closes that tab and leaves the main window open. You can press it again to close the next tab. When you are working in the main chat instead, the shortcut still closes the window as usual.
If OpenClaw fails to start on Mac, its service log can now include errors that used to disappear before normal logging began. Existing installations need their background service setup refreshed before these errors appear in the log. The Mac troubleshooting guide explains the setup and where to look.Status and Doctor can also point out background jobs that may be repeatedly restarting OpenClaw. These checks can flag a job incorrectly, so review it before removing anything.
Putting your Linux desktop to sleep no longer tells OpenClaw on another computer to pause, including when you connect over SSH. Saved SSH connections also use their configured port. If you keep a password or token in a separate file or environment setting, reconnecting preserves that setup so later launches can read the updated secret. Settings overwritten by an older version still need to be restored manually.The Linux companion also keeps update notices and tray actions available while you are using the dashboard. An update that is ready to install stays available for another try if a later check or download fails.
An explanation with a little code in it can now be read aloud instead of being skipped or replaced with a message to look at the screen. This applies to spoken replies and Talk when speech is enabled. Answers that are mostly code still follow the existing rules for keeping that content on screen.

​
Models and Providers

You have more choices for making images and transcribing voice notes, and choosing the model behind your Claw is easier too. Model menus are searchable, saved choices keep the right settings, and the controls stop offering options a model cannot use. For people working with Codex, Claude, or their own AI server, there are also fixes for missing instructions, interrupted work, and confusing sign-in failures.
If you connect to more than one Gateway, the model list now follows the Gateway and agent you selected, so you see the choices for the place where your work will run. Browsing the list no longer starts another search for models in the background. Refresh checks for new choices, and a failed refresh shows a warning while keeping compatible saved choices available. Update the Gateway before updating a separate command-line client. Updates to OpenClaw’s model-and-price reference still need a Gateway restart before they take effect.
Saved model choices are less likely to change unexpectedly, particularly when custom models have almost identical names. Each keeps its own settings for images, tools, and the amount of text it can handle. Chats started from another chat also show the model they actually inherit, and choosing Default stays in place after a restart. Default follows the configured model and its backups; explicitly choosing a particular model keeps the chat pinned to that choice.
Searchable model menus make long lists easier to use without changing your choice until you select a result. Chat menus follow the account saved for that conversation, while New Session lets you preview another account you own. If a list fails to load, you can see the problem and try again.Editing backup models now preserves an inherited primary model, and removing every backup stays that way after a reload. Unavailable saved choices remain visible so you can remove them. Global defaults labels help distinguish shared settings from an individual agent’s choices.
The Thinking menu now follows the effort levels offered by the selected model, so it no longer invents choices and correctly shows your saved Off setting. When OpenClaw knows that Fast cannot change a request, it disables that choice while still letting you clear an old preference. These controls make it clearer what you can adjust; actual speed, access, and charges still depend on the provider.
People using custom AI servers keep the connection settings they entered when model lists are refreshed, instead of having another model’s settings take over. Manually added models also stay in the list. If you use a custom server in place of a provider’s usual service, you may need to list the models it supports explicitly. Integrations must use current sign-in details or explicitly configured credentials; old credentials saved only with a downloaded model list no longer authorize requests.
Supported OpenAI sign-in methods can reuse an account or key already saved by Codex, saving you another login or paste. Agents sharing a sign-in also handle renewal more reliably, and a renewal that finished too late for one request can help the next request succeed.If sign-in still needs attention, OpenClaw better distinguishes saved credentials from a connection that has not picked them up and tells you which computer to check. Older credentials awaiting repair no longer disable unrelated providers; use openclaw doctor --fix for the affected accounts.
When a model sends broken instructions for a tool, OpenClaw can try again instead of immediately abandoning the task. With Anthropic, it can also repair certain formatting mistakes. Retries are limited to cases where repeating the request is safe and may add cost. If earlier actions already happened, check their results before asking your Claw to continue.Errors also better distinguish connection problems, rejected sign-ins, and requests a provider cannot accept. Explanations stay in chat history, making it easier to understand why work stopped or a backup model was not tried.
Anthropic conversations that use tools can reuse more of the instructions and conversation text already processed, instead of repeatedly saving the same growing history again. Compatible local models also get a more reusable starting prompt when you begin another conversation with the same instructions and tools. This can avoid repeated processing, though any effect on waiting time or charges depends on the provider and workload.
Codex conversations managed by OpenClaw regain workspace personality, preferences, skills, and memory guidance, including later edits. This applies to supported managed connections; other connection types keep their existing handling and may show a warning.Ordinary Codex chats can also recover a missing final answer after finishing work in a long conversation, without repeating those actions. Follow-ups keep fresh corrections without resending history already covered by a successful final reply.
Returning to a Claude conversation keeps its original chat connection and avoids a repeated user message in the affected resume cases. Claude can also stay running between turns when a skill check finds that nothing changed. In Codex chats with a fixed model choice, pasted text now reaches the agent instead of being lost.
If you run models on your own server, OpenClaw now uses the conversation size reported by vLLM, SGLang, and compatible services instead of falling back to a potentially wrong limit. That helps it budget how much text to send to the model, while any limit you entered yourself still takes priority. Opening vLLM model information for the first time also avoids some unnecessary loading work.
Tencent users can choose Hy4 preview through TokenHub or TokenPlan, and new setups select it by default when the account has access. Existing model choices stay in place, so an upgrade does not silently switch you to the preview. Its thinking options are Off or High; selecting an intermediate level such as Low uses High, which can affect waiting time and usage.
New xAI setups use Grok 4.6, as do xAI web search, X search, and code-execution tools when you have not chosen a model for them. Your explicit choices stay in place, but the new tool default can change cost or availability for your account. If an older subscription setup still uses the retired Auto choice, run openclaw doctor --fix or select an available model yourself.
OpenCode Go and Zen requests made outside an ongoing conversation no longer fail just because the service needs session information that was missing. OpenClaw supplies it automatically for the supported direct connections, including requests from the command line and supported integrations. Custom server connections keep their existing requirements.
If you use Arcee through OpenRouter, OpenClaw now uses your matching saved OpenRouter account instead of asking for a separate Arcee credential. Your model settings stay attached to that choice, and different spellings of the same Arcee model stop looking like an unexpected model switch.
When Kimi explicitly reports that you have used up a weekly allowance, OpenClaw no longer treats your valid key as broken. It can try a configured backup without unnecessarily blocking the provider’s other models. Changing the key will not restore the allowance, though; wait for Kimi’s quota window or use another provider with available capacity.
You can generate and edit images with GPT Image 2.5 Flare or Sunburst through OpenAI or fal, including custom sizes, higher quality settings, and transparent PNG or WebP output. Edits can use up to five reference images through OpenAI or sixteen through fal, and either route can return up to four images.Direct OpenAI access needs an API key, selected explicitly if you already use a ChatGPT sign-in; fal needs its own credentials. Your default image model stays unchanged, so choose one of the new variants when you want to use it.
You can now choose Deepgram Flux to turn incoming voice notes into text your Claw can use, with English and multilingual options. Install ffmpeg on the computer running OpenClaw before selecting Flux. This is an optional choice in the existing audio settings; Nova remains the default.
Gemini web search now uses Gemini 3.6 Flash unless you chose a specific search model, helping accounts that could no longer use the old default. Your main chat model stays unchanged. Check the billing difference when upgrading, because Gemini 3 charges for each search query rather than once per prompt. There is also clearer help for setting up Kimi search and using Ollama cloud search while keeping model conversations local.
Cost reports now warn when totals may still be incomplete, so a temporary zero is less likely to look like no usage. They also distinguish the tokens used to save reusable prompts from ordinary input and fill gaps in some OpenRouter estimates. These are reporting fixes; the provider’s bill remains the amount you actually owe.
With diagnostic recording enabled, OpenRouter can return finished music or a usage-check error without waiting for the recording to finish. You get the result sooner even if that optional recording is incomplete. Related fixes let stopped requests finish closing their connections without cutting off cleanup.If you run openclaw models status --probe directly from the terminal, stop the Gateway first and wait for the command to exit before starting it again. The result can appear while the check is still closing its files and connections.
Setting up an account or checking a provider’s options is easier in the reorganized OpenAI and model-provider guides. Shorter pages take you straight to the relevant instructions, and old bookmarks still point you toward the material that moved.

​
Automations and Scheduling

Automations are easier to manage, with direct links to edit them and names you can recognize. Scheduled tasks also get fixes for getting stuck, returning results to your chat, and staying quiet when there is nothing to report.
Scheduled commands run by OpenClaw can work without turning off secret protection, and Codex chats can set up automations using the tools already available in the conversation. Tasks also avoid startup failures caused by model settings refreshing or an unrelated plugin being disabled. Your existing permissions still apply, and disabled plugins stay off.
A clash between scheduled tasks could make OpenClaw keep retrying until it stopped responding. That loop is fixed. Tasks that hand work to another agent also avoid a pause they could never resume, leaving OpenClaw to deliver the result through the task’s existing settings.
Background commands can bring their results back to the chat after your Claw’s earlier reply has finished. Routine checks with nothing to report can stay quiet instead of sending an unnecessary message.If an automation result is waiting while your Claw answers you, canceling that result can now unblock actions such as pinning the conversation. Your Claw can finish its reply without adding the canceled result to the chat.
You can jump straight to an automation’s editor from Settings → Agents → Automations, even when the task is paused. Lists now show the friendly names you saved, and you can group scheduled conversations together so they are easier to find. Large automation lists also spend less time loading where results will be sent.If you manage automations from the command line, missing jobs and invalid options come with more useful guidance. Warnings also explain what to check when a saved task cannot run because scheduling is turned off.
If you have set conversations to reset each day, they now avoid losing context too early around spring clock changes. The reset still follows the timezone of the computer running OpenClaw.Guidance is easier to find when you want to start a task at a set time or when a command finishes, or set up an automatic response to an event with a hook. These guides explain existing features, with clearer instructions for choosing and setting them up.

​
Browser and Computer Use

You can keep using your browser while your Claw works in another tab, and joining it on a remote desktop takes fewer steps. This release fixes interruptions when opening windows, answering website pop-ups, and following along with work on another computer.
The Browser panel can show what your Claw is doing without switching you away from your own tab. You can still choose Open to bring its page forward. If the live view disconnects after it has started, OpenClaw refreshes the picture and tries to reconnect automatically, while keeping any image you are marking up in place. Blank pages also stop appearing as misleading preview cards in chat.
When a website pop-up stops a browser command, OpenClaw tells you which prompt needs an answer and which steps remain unfinished. Answering one pop-up also no longer prevents the next one from being answered. Longer tasks in an already open browser avoid the premature connection failures covered by these fixes.Cancelling cannot undo changes already made to a page. If you automate browser commands, check their reported results before continuing, because a blocked task does not always return an error status.
This release fixes computer-use errors that could stop your Claw before it had even found the right window. Supported desktop controls are available from the start, and bringing a window forward no longer looks like a failure when it worked. Your Claw still needs permission to use the computer, and the computer must support the controls it requests.If Codex Computer Use cannot start because native plugins are disabled, it now explains that promptly and tells you how to enable them and retry installation.
Opening Desktop from a chat now takes you to the computer running that chat, and the view stays connected while you type in another pane alongside it. Desktop support must be enabled, and you start by watching until you choose Take control.To paste text from your own computer, choose Take control, open Keyboard, and paste there once the connection is ready. Paste shortcuts used directly on the desktop view are sent to the remote computer. If someone else takes control, the notice now shows their signed-in name or user ID when available.
Stopping OpenClaw on a connected computer now lets its active computer-use session finish closing before another one starts. If closing fails, the error stays visible so you know why a new session is blocked and what needs attention.
You can now enable desktop access for Azure cloud workers, whether OpenClaw manages them directly or through a coordinator. Turn on the Cloud Worker Desktop lab and enable desktop access in the worker profile. Existing workers created without a desktop need to be stopped and set up again with that setting enabled. This adds Azure to the supported desktop options; GCP desktops remain unsupported.

​
Plugins and Integrations

Finding something new for your Claw is easier, whether you browse plugins or ask for a recommendation in chat. You can review and set up plugins in the browser, while cloud workers gain ways to reuse project setup, keep a computer ready for your next session, and choose Windows or macOS when your provider supports them.
You can now search installed and available plugins together, browse by purpose, and keep exploring without clicking through pages. In chat, your Claw can recommend up to three official plugins or skills and give you a card that opens their details. Recommendations need the message tool enabled, and choosing Install opens a review before anything is added. An installed plugin may still need settings or an account connection before you can use it.
You can read what a plugin does, check what it needs, and follow installation through to setup without leaving the browser. Plugin settings puts its configuration, permissions, and troubleshooting information together, with clearer messages when setup is unfinished. The install dialog stays with your latest choice and waits through the required restart before confirming completion. Installing or changing plugins still requires the appropriate permissions and your consent.
If you manage plugins from the terminal, error messages now point to recovery commands for your profile or container and explain whether you need to restart OpenClaw. Plugin checks also report cleanup failures instead of claiming success, and slow catalog responses no longer keep commands waiting indefinitely.
People building plugins get fixes for type errors and missing supporting files that could stop a build. You can also opt into live previews in the browser, seeing changes to your plugin against a real OpenClaw installation. If you wrap the multi-account configuration helper, check its updated types when rebuilding; the SDK guide covers development setup.
Requests using model plugins are less likely to fail because a connection closed while work was still finishing. OpenClaw keeps those connections available through cancellation, plugin reloads, and shutdown. A plugin that cannot finish closing can still delay shutdown.
Image, music, and video jobs can finish and save their results even when the plugin setup that started them is replaced. Related fixes keep connections available for speech and document analysis and prevent the last part of streamed audio from being lost. Paid generation that has already started may continue after you cancel.
Plugins with lots of saved data use less extra memory when listing it, helping reduce the strain of a large collection. The first use after the storage update may take extra time and disk space.
If you connect external tools through MCP, large text replies from connected devices no longer include an unnecessary duplicate copy. Stopping a standalone MCP service gives ongoing tool work time to finish cleaning up, and an accepted disconnect no longer gets stuck waiting on response cleanup. Failed disconnects still report a problem that may need attention.
Commands on connected devices no longer time out early or wait longer just because the system clock changed. Time spent waking the device counts toward the same limit as running the command. If you use scripts, omit numeric options when you want their defaults—passing an empty value now gives a clear error before OpenClaw contacts the device.
Google Meet and voice sessions can now ask a separately configured agent for help even when the original conversation has a locked model choice. That agent uses its own conversation without copying your earlier chat, and its own model restrictions still apply. Failed Chrome meeting starts also try to stop the audio they opened and close only newly created tabs, leaving your existing tabs alone.
If you use A2A to connect agents, incoming requests now reach the agent assigned to that connection even when a new conversation begins. You can still choose a different agent for a particular conversation, and separate conversations keep their own history.
Apps using OpenClaw through Chat Completions or the Responses API can now tell when an answer stopped at its output limit. They keep the partial answer and receive a clear signal that it was cut short, including during streaming, so they can decide what to show or do next. Connected apps also get a more accurate report of which settings changed after a configuration update.
Reports opens at the short /reports address inside OpenClaw, and embedded plugin pages use more of the window. Plugin panels also load in affected containers where file permissions previously got in the way.If you use separate Team Reports report or export links, change /reports/... to /plugins/team-reports/...; old links do not redirect automatically. The guide explains how to keep a custom address if you need one.
New cloud sessions can reuse a project that is already set up, saving repeated installation and preparation work while keeping your session’s changes. You can use an eligible local project or start directly from a public GitHub repository. Starting from only a private repository URL still uses a fresh checkout; private projects already checked out locally can use the local-project route. A saved setup still needs a machine to start unless a running spare is available.Back up OpenClaw’s data before upgrading, following the backup and rollback instructions. Older versions cannot read the updated data format, so rolling back the app alone is not enough. Restoring the earlier backup also discards changes made after it.
Ready workers let your next matching session use an already running computer while OpenClaw prepares a replacement. Eligible Linux projects get one spare per project and profile by default, with a shared limit of four. These computers cost money until deletion is confirmed. Set a profile’s Ready workers value to zero to disable its spares, or the shared Prepared pool limit to zero to remove unused spares without stopping active sessions.If you already use saved worker images, stop OpenClaw and any image captures it started, then follow the Doctor upgrade step before starting more workers.
You can choose a cloud computer’s operating system for a profile or an individual session when your provider offers it. Native Windows runs Windows commands, while WSL2 gives you a Linux environment on Windows. Mac workers on AWS require a Dedicated Host and On-Demand capacity; desktop access and reusable worker images remain Linux-only. The operating-system guide covers availability, with separate setup requirements for native Windows.
Cloud workers settings now lets administrators change advanced options, set spare-computer limits, and choose a default profile for each repository. Save your changes and restart OpenClaw to apply them. You can also hover over a cloud session to see its known operating system, CPU, and memory capacity.OpenClaw can automatically download a supported Crabbox tool when yours is missing or outdated. If your computer blocks those downloads or installations, prepare Crabbox 0.55.0 or newer before upgrading, because you need it even to inspect or stop existing cloud workers.
Administrators can save a prepared project in Settings, watch its progress, cancel it, or pin a useful snapshot to keep it. You can choose a previous snapshot for new workers without changing computers already running. Builds use the project’s committed files and setup instructions, and may reuse a suitable existing copy.Before using Recover after an interrupted capture, stop the affected worker and capture and check the saved images with your provider. Recover only clears OpenClaw’s record of the interruption; it does not stop a computer or delete an image for you.
You can send a message while a cloud session is setting up or syncing files and see what it is waiting for. OpenClaw keeps the message and starts it when the intended computer is ready, saving you from copying and resending it.Supported cloud computers can also keep their files, installed tools, and desktop through an OpenClaw update. A message that has not started can continue once the worker is updated, but work already underway is not automatically repeated. The computer must support the update and reconnect; this cannot recover a machine lost by the provider.
Saving a remote workspace with many small files can finish sooner, and its file list needs fewer bytes to transfer. Cancelling preparation stops new scanning and closes the upload connection instead of leaving it hanging, though file operations already underway may need time to finish. Oversized or changing files are also rejected earlier, without leaving a partial workspace result.
When a cloud worker fails, the error now keeps the original problem and the latest cleanup failure visible without piling up repeated messages. Stopping a computer can succeed even if saving an optional image failed, with a warning explaining what remains. Failed image deletion stays visible for retry, and charges may continue until your provider confirms deletion.
Owners and administrators can search Codex plugins, open their setup links, and see installation, permission, and account-connection information together. Opening a setup link does not finish the step—return and submit the completion response when you are done. After connecting, use /codex plugins refresh to update the app list, then /new or /reset where required; refreshing does not sign in or enable an app for you.

​
Security and Privacy

Your Claw can make more decisions about routine commands without interrupting you, while still asking for approval when needed. There are also fixes for keeping file access within the permissions you chose, protecting shared conversation history when deleting an agent, and keeping private update records out of backups and support bundles.
Automatic command review can now let suitable routine commands run, refuse a command with an explanation, or ask you to decide. This takes effect in existing automatic and workspace modes. In OpenClaw’s built-in agent, the reviewer can use part of your conversation to check whether a command fits your request. High-risk actions still need human approval, and settings that always ask remain in force.
Turning tools off now carries through to supported connected agents, including Copilot, even when you return to a conversation that previously allowed them. If an integration cannot honor that restriction, OpenClaw stops rather than continuing with tools enabled. Update the plugin or choose a supported agent to continue.Your Claw’s instructions also allow it to use existing login details for tasks you authorize, such as signing into a website, while keeping your other permissions in place. When it cannot offer secure setup in chat, it directs you to terminal prompts that hide what you type.
Password login now rejects blank passwords and known example values. If one prevents your Gateway from starting, replace it, restart OpenClaw, and update the apps that connect to it. Short real passwords produce a warning rather than a new startup block.More model connections can use protected API keys without turning protection off. Doctor also gives administrators a working repair for certain older device sign-in failures, making it easier to reconnect affected devices.
For Slack connections using HTTP webhooks, OpenClaw now enforces limits on oversized, slow, or excessive incoming requests before processing them. Feishu and Lark webhooks also reject signed requests dated more than an hour before or after the computer’s clock. Keep that clock synchronized, and check Feishu troubleshooting if delayed deliveries are rejected. Feishu’s WebSocket connection mode is unaffected.If you use IPv4-mapped network rules for a reverse proxy or device pairing, review the addresses you allow. Those rules now cover their full stated range, fixing connections that were wrongly rejected. Broad rules may therefore trust more addresses than before, including the computer itself; keep the allowed range as narrow as your setup needs. Other login and approval checks still apply.
When an agent is limited to its workspace, it can now send files from the current chat’s sandbox without being able to attach files from another chat’s sandbox. Plugins that send these attachments must identify the correct workspace, so some third-party plugins may need an update.Changing a shared conversation to draft now also blocks new file requests from readers who can no longer see it, even without separately configured user roles. Download links issued earlier remain valid until they expire, and people sharing one OpenClaw installation still need to trust one another.
If you have read-only access, opening an MCP App in a separate view no longer gives you permission to run its tools. You can still view the App and use the resources you are allowed to read. For MCP tools running on the same computer, rejecting an oversized upload also stops a tool command queued behind it from running on that connection.
Deleting an agent now stops if another agent needs it to read shared history. Moving that history is not yet supported, so keep the agent configured. If deletion fails partway through, fix the reported storage problem and retry. Before deleting a workspace that links to a neighboring folder, check what will be removed. Deletion through the Gateway may move that linked folder to Trash too; the offline command removes only the link.Backups and support bundles leave out recognized private update records. If you move a protected folder, keep its privacy-marker file with it. Files copied elsewhere without that marker are not protected from export.
Running openclaw security audit --fix now keeps the WhatsApp group access list you already chose, instead of adding people just because they had previously paired with your Claw. When changing an open group policy to restricted access, it can still use approved senders if you have not supplied a list. People added by an earlier repair are not removed automatically, so review your list if you were affected.
The telemetry guide now explains what an update check can reveal about your approximate location, what the local preview can show, and the documented three-month retention period. Turning off optional statistics or setting DO_NOT_TRACK does not stop automatic update checks. If you want to stop those requests too, use the separate controls explained in the guide. The location may be that of a VPN, proxy, or remote computer, rather than where you are sitting.

​
Quality-of-Life Improvements

You can now answer your Claw’s questions and enter secrets from the terminal, with clearer output to read and copy afterward. There are also fixes for starting and restoring coding workspaces, giving delegated tasks enough time, and finding the help you need when something goes wrong.
When your Claw needs an answer, you can now choose an option or type a response directly in the terminal, including while a task is waiting for your reply. Press Esc to put the question aside and /question to bring it back.Secret requests get a masked entry box that keeps what you type out of the conversation and input history. Saving a secret requires a Gateway connection, and terminal entry accepts one line at a time. Use the web app if you need to change which sites can use it. Local sessions support ordinary questions, which you should answer before closing the terminal.
Long addresses and other text now wrap in terminal chat without gaining extra spaces, so you can copy them as written. Tables keep related values lined up, and progress indicators stop leaving repeated lines behind in narrow windows. If you use scripts to read command results, approvals JSON and transcript output also keep the unrelated announcements out of the way.
Command errors do a better job of telling you what to do next, such as supplying login details or choosing which agent to use, without making a missing option look like a broken installation. If openclaw agent loses its connection or times out after a task was accepted, the error gives you the task’s ID so you can check what happened before trying again. Check its conversation history and Gateway status first, because the work may have continued after the connection was lost.
Code Mode uses less of its result space for formatting, leaving more room for the information your Claw asked for. The terminal also shows that data as written, without treating parts of it as message formatting. Code that calls itself too deeply now returns an error the Claw can work from, making it possible to correct that code and continue.
A worktree gives your Claw a separate working copy of a code project. Setup retries now keep the version you originally selected, and projects with many branches keep the option to start a worktree available. You can type a branch or commit that is missing from the suggestions, with an error if it cannot be found.Saved worktrees also restore changes that replaced a file with a folder, or a folder with a file, keeping the replacement contents intact.
When a long AGENTS.md instruction file has to be shortened, an example keeps the line directly above it that explains what it is, helping keep a quotation from looking like a new instruction. The shortened instructions also recognize more Traditional Chinese rules, although the available space still limits what can be included.If you track model usage, completed /btw side answers can now include their reported usage in your configured diagnostics. Those side questions still sit outside the conversation’s /usage cost total.
Subagents are helpers your Claw can send to work on part of a task. Helpers with visible conversations now get the time selected for their first task, so longer work is not cut short by the ordinary deadline. Later messages use the normal time limit. A separate fix helps finished results return to the right agent after a restart when several agents share one Gateway.
The FAQ is organized around the question you are trying to answer, and the documentation home gives you direct paths to releases, skills, automations, and building plugins. Setup examples are clearer, and repairs to older bookmarks and help links make the relevant instructions easier to reach. Simplified Chinese navigation also gains translated labels, including separate Releases and Contributing tabs.

​
Other Bug Fixes

OpenClaw now handles more of the interruptions that can leave you without an answer, stop a conversation from loading, or make a successful task look like it failed. Error messages and troubleshooting records also explain more of what happened.
Your Claw can recover a missing final reply in more cases where the work finished but the answer never arrived, including after a lost model connection or an earlier tool error. It asks the model for an explanation without repeating completed actions, giving you another chance to see what was done.
Starting over with /new or /reset avoids unnecessary waits, especially in conversations where your Claw has handed work to other agents, and canceled tasks stay stopped when late results arrive. Your saved history and other conversations remain available.Old plugin messages outside the current conversation no longer stop its history from loading, and long conversations keep recent messages available when your Claw pauses to wait for other work. If a text-only reply fails partway through, the model also receives a short note about that failure when you continue, so the attempt no longer disappears from its context.
When your Claw cannot save a file in its workspace, it now explains whether access was denied, the disk is full, or the location is read-only. That gives you a useful starting point for fixing the problem. Tool Search also fixes a case where a supplied file path was lost before the file could be read.On Linux, affected SSH commands can finish instead of being stopped too early. Commands on macOS and Linux also avoid a false cleanup error after successful work. Restart OpenClaw if you updated with --no-restart to get that cleanup fix.
OpenClaw can now wait through a brief database lock instead of immediately failing to read its saved settings. When a database problem does cause a failure or shutdown, error messages preserve more of the information needed to investigate it. Logs also give operators a clearer view of which work is waiting or taking time, with extra detail available when diagnostics are enabled.Exported conversations keep file-read results that were incorrectly removed, including instructions for reading the next part of a file. With the relevant diagnostic recording enabled, records also include actions taken after switching to a backup model and partial responses received before an interruption. Sensitive information is still removed, and recording size limits still apply.

​
Maintainer and Internal Changes

This release includes 692 maintainer-only units covering code cleanup, testing, release preparation, and contributor tools. Expand the source lists for the full change history.
Shared code and type definitions replace duplicated logic across OpenClaw. This gives maintainers fewer places to update when changing behavior and removes unused code.
Tests are less prone to failures caused by timing, leftover data, or the machine running them. Reusing setup and removing duplicate checks also reduces unnecessary work for contributors.
Automated checks avoid unrelated work when a change affects only tests. Performance reports also measure CPU use more accurately, helping maintainers distinguish product regressions from measurement errors.
Build and package checks better recognize which dependencies and files belong in each release package. This reduces false failures when maintainers prepare plugins and check their contents.
Release tools make interrupted publishing easier to investigate and recover from. Checks for older releases match the features those versions contain, while mobile release jobs get fixes for build preparation and troubleshooting. The checks that read older releases directly from Git require Git 2.45 or newer.
Contributors get less repeated setup during local checks and clearer testing guides. Documentation tools preserve code examples and point errors to the right line. Normal pnpm installs now reject links to another checkout’s dependency folders. Give each source checkout its own dependencies or explicitly configure the dependency location.
A change to conversation-marker spacing was undone. Both changes are listed below; the existing compact left rail remains.