v2026.10.2

AI agents and tools can read these release notes as plain Markdown.

OpenClaw 2026.10.2 fixes errors during upgrades from 2026.10.1 and Doctor repairs, addresses missing or repeated replies, and helps long conversations continue, with fixes for credential masking in large tool output.

Release scale: 80 pull requests, 12 direct commits, and 107 contributors.

Installation and Onboarding

• Node during installation. Recovery installs use the compatible Node runtime OpenClaw selected even when an older system Node comes first in PATH, avoiding a misleading package-script version error. See the Node setup guide. #164677, thanks to @steipete and @jarvisnextgeneration20-lang.

Web UI

• Dashboard replies. Chats can answer after a daily or idle reset when background title work previously blocked them. Background commands that delegate and wait also avoid producing the final answer twice. #162689, #166362, thanks to @ericcaiwx-star, @dimonnld, @OmniScaleMatt, @gloria0021, @obviyus and @shakkernerd.
• Dictation. Holding the microphone and releasing it no longer stops recording prematurely. You can keep speaking, then tap Stop to insert the transcript. #166843, thanks to @alfredjbclaw and @obviyus.
• Opening and reconnecting chats. Discovering available tools can reuse compatible large plugins already running, avoiding a duplicate load that could freeze the chat and other requests. #167374, thanks to @jalehman.
• Typed stop commands. In a connected non-global chat, /stop clears queued follow-ups as well as active work, so waiting messages do not immediately restart the agent. The Stop button and offline replay still stop only the particular run. See the chat cancellation guide. #167860, thanks to @masatohoshino and @steipete.

Updates and Maintenance

• Upgrading from 2026.10.1. The first upgrade from published stable 2026.10.1 avoids a cleanup error after replacing the package. Windows upgrades also get past the final plugin check that the older updater could no longer call. #168849, #169031, thanks to @RomneyDa.
• Windows update checks. Windows updates no longer reject a valid update solely because of how its launcher starts it. Backup checks also recognize short names and different letter casing for the same folder. #166530, #168968, thanks to @steipete, @wqj11a, @H0Nm4M1suZu and @RomneyDa.
• Update cleanup. A fixed CLI no longer reports a missing-module error while cleaning up an update that already succeeded. An older CLI can still show that error during its first upgrade; use the newly installed openclaw update status to check the result. See the update guide. #165233, thanks to @steipete.
• Update and repair reports. An update confirmed successful in saved history no longer appears to fail just because OpenClaw cannot display its result. Optional configuration fallback notices now appear as warnings. Repair reports also call out configured chat accounts that remain stopped. Use openclaw health --json to investigate those accounts; see the repair guide. #169037, #165539, thanks to @steipete, @kvnloo, @RomneyDa and @Patrick-Erichsen.
• Container file copying. Updates can copy file bytes when a container blocks file cloning. Affected 2026.9.7 installations need one manual npm installation of a fixed release before later updates can use the repair. Keep native filesystem safeguards enabled and follow the recovery steps. #165257, thanks to @steipete, @MrJanvar, @rybing7 and @Mohl.
• Workspace avatars. Update checks no longer reject valid absolute avatar paths inside your workspace. An older updater may still need a simple workspace-relative path such as avatar.png until it has been updated. #164638, thanks to @nitinjwadhawan, @vincentkoc and @BottaniCals.
• Doctor after container replacement. After replacing an exited container, Doctor can wait for the previous Gateway's lock to expire and continue repairs. Updates also avoid a false “undergoing offline maintenance” refusal caused by Doctor's own work. If Doctor reports a live Gateway, follow the stop instructions in the Doctor recovery guide. #165355, #165417, thanks to @steipete.
• Linux repair checks. Timestamp changes on an unchanged database no longer cause false update or recovery refusals. Slow local database checks also avoid a false service-manager timeout before Doctor can stop the Gateway for maintenance. #165775, #168161, thanks to @steipete, @donquiche61, @naturecurly, @dreamkeeper, @wangmiao0668000666, @clockthunder, @honghongxinxi-zhb, @jayzhou2309, @obviyus and @Tosko4.
• Older session history. Doctor can continue repairing leftover session files while keeping current conversation labels and active chats intact. Disputed originals are preserved for recovery; a warning or protected archive does not mean that history was imported or is safe to delete. 329857b, 80117e8, thanks to @steipete and @coolcoolwcr.
• Leftover migration files. Doctor can archive old sender lists for removed channel accounts and continue repairs. Memory-indexing lock files also stop triggering repeated recovery warnings for a healthy database. Damaged files for configured accounts and genuine recovery holds still need attention; see the migration guide. #167651, #166517, thanks to @steipete, @oshunter and @noelff.
• Restart after refused maintenance. If an update finds OpenClaw already current but stops the Gateway for maintenance that cannot proceed, it can restart the service once recovery is safe. Maintenance still reports failure. Services you stopped yourself, or those kept stopped because data is at risk, remain stopped. #166548, thanks to @steipete and @Shirleyyyhk.

Messaging

• Replies during startup. Messages can wait for their own agent to finish starting, while agents already ready keep answering independently. #168826, thanks to @steipete, @xiaozishan and @obviyus.
• Telegram previews. A partial reply stays visible while its replacement is delayed or rejected. The old preview is removed after the replacement is accepted, preserving useful text even when the final send fails. See the streaming guide. #164649, thanks to @jayzhou2309, @obviyus and @jmar5439.
• Telegram send failures. A reply blocked locally by unresolved account ownership shows an accurate failure notice instead of an uncertain-delivery warning. The ownership setting may still need correction; see the channel ownership guide. #167470, thanks to @obviyus and @mgr-nick.
• Feishu streaming. Outdated waiting previews can be skipped so the completed answer does not wait behind every earlier version. #166610, thanks to @rudycelekli, @feie22, @cas1per, @L-Trump, @xiaobao-k8s and @csxjihk.
• Messages between agents. Replayed messages retain the original sender label, including when one agent forwards another sender's text, so the receiving agent keeps that context on later turns. #169130, thanks to @RomneyDa.

Memory

• Memory recall through Tool Search. Supported memory searches through Tool Search no longer fail when an agent asks for a score cutoff or a limit on results. #158632, thanks to @masatohoshino, @obviyus and @he-yufeng.
• Summaries before completion replies. When a long conversation needs a summary, OpenClaw summarizes its history before answering a helper result or following up on an approval, without changing the selected model. #163117, thanks to @SunnyShu0925, @steipete, @AXEG0 and @obviyus.
• Corrections during an answer. With a durable context engine, sending a correction while an answer is running no longer causes that completed exchange to be left out of the agent's stored conversation context. #168394, thanks to @steipete and @robertoronderosjr.

Models and Providers

• Provider compaction. Eligible direct OpenAI Responses and Anthropic API-key chats now shorten long conversations with provider-managed checkpoints by default. Anthropic chats can recover on the next turn after the provider rejects an earlier checkpoint. OpenAI manual /compact can cost more than a local summary because it processes much more uncached input; see compaction and OpenAI options for route limits, fallback and opt-outs. #168164, thanks to @obviyus.
• Compaction during model cooldowns. A working configured fallback can summarize a long conversation while the primary credentials cool down after a quota failure, without changing the selected session model. Explicit compaction-model overrides stay exact; see the compaction guide. #166459, thanks to @steipete and @hawkit.
• Claude CLI compaction. Automatic post-turn compaction stays with Claude Code, avoiding an unnecessary Anthropic API-key error and preserving the native conversation connection. Manual /compact is a separate path; see CLI backend guidance. #162066, thanks to @rogeriochaves, @llmsommelier, @RomneyDa, @ajmtrz, @MaShAn1027, @Jeehut and @Oldrich333.
• Long Claude CLI replies. Fixed missing final reports from completed Claude CLI tasks with many streaming updates. #153545, thanks to @zhangguiping-xydt, @methylester, @obviyus, @VACInc, @aaajiao, @telbus, @aoclaw-glitch, @aikohoshinoai-prog and @RTapiador.
• Claude CLI history. Imported history avoids duplicating your messages and forwarded input, and preserves who sent them. Recognized background instructions no longer look like messages you wrote, while large tool results appear as tool activity. The updated CLI guide explains these display changes and confirms that stored transcripts remain intact. #163521, #166080, #168851, thanks to @rogeriochaves, @llmsommelier, @RomneyDa, @Marvinthebored, @Peetiegonzalez, @IndySweeney, @agierror and @Heighte.
• Long ChatGPT subscription replies. Long ChatGPT subscription replies over HTTP can finish without being cut off by the total amount of streamed data. Individual streaming updates remain size-limited. In native interactive chat over Responses WebSockets, OpenClaw can resume from the saved conversation and continue the answer if a reused connection expires after a reply has begun. #166853, #167063, thanks to @jackten and @steipete.
• Codex session startup. Closing idle threads with background terminals no longer blocks later turns. Stalled startup waits can be cancelled and respect their request budget. #166693, thanks to @jalehman.
• Codex provider errors. A provider failure after tool work remains visible with recovery advice instead of looking merely stopped. Completed actions are not repeated, and an explicit Stop still means cancellation. #165488, thanks to @steipete.
• Code Mode tasks. Long Code Mode tasks can keep using tools when a model reuses an earlier request number. Node-command errors also explain how to run a shell command on the intended computer. #165210, thanks to @steipete.
• OpenAI tool compatibility. OpenAI replies no longer fail merely because an enabled plugin or MCP tool contains an affected pattern in its input rules, even when that tool is not used. See advanced provider guidance. #166999, thanks to @scs0209, @adamjgmiller and @obviyus.
• Local model replies. Models with strict Chat Completions templates can keep replying after background notices that previously caused a message-format error. #167309, thanks to @jlapenna and @steipete.
• Custom Bedrock routes. Requests using a custom Bedrock provider ID no longer fail before contacting AWS, and use that route's configured region. See the Bedrock guide and custom-provider setup. #164406, thanks to @Ayushdevo, @obviyus and @kai-ocean-kurt.
• Local llama.cpp setup. Setup checks whether the server can run before downloading model files and explains alternatives when it cannot. Managed macOS binaries require 13.3 or later; older Macs can use a working existing or compatible custom server through local-model service configuration. e0de414, thanks to @NovaUnboundAi, @xXG0DLessXx and @VACInc.
• Loading models for saved chats. Continuing a conversation from the saved-session catalog or changing chat settings no longer hangs indefinitely while loading the model list. A timeout lets you retry without leaving a partially created chat or applying a rejected settings change later. Bulk settings changes still report each chat's result separately. See session-control guidance. #166089, #166376, thanks to @ericcaiwx-star, @steipete and @dprev.
• Model-list changes during startup. A model-list update between startup checks and execution no longer causes the reported refusal to start a chat or scheduled job. #168410, thanks to @RomneyDa.
• Fallback and saved messages. Automatic fallback can reach another provider after the first attempt was sent, instead of failing before the retry begins. Your saved message stays unchanged. Messages between agents also keep the same text and sender context when first delivered and replayed later. 9def9c1, 36f6305.
• Claude history caching. After a large document, image or tool exchange, Claude can still reuse eligible cached earlier conversation history. Bedrock also chooses a caching duration supported by the selected model and profile. See prompt-caching guidance. #168042, thanks to @steipete.
• Conversation history and caching. Reopened chats keep the same shortened tool-result text the model saw originally. Browser chats can keep reusing cached conversation text when a helper reports back privately, and Claude caches earlier history without temporary instructions that disappear later. See prompt-caching guidance. #168392, thanks to @steipete.

Automations and Scheduling

• Codex max thinking. Scheduled and webhook turns honor the existing max-thinking setting when the selected model supports it, instead of silently using its default effort. #165444, thanks to @Oldrich333.
• Condition checks after restart. The first scheduled script check after a restart no longer reloads plugins already running, avoiding the resulting timeout and stalls in other Gateway work. See schedule guidance. #166991, thanks to @67-luck, @steipete and @lucianosmori.
• Manually started jobs. An accepted automation can keep using its allowed tools after the chat turn that started it ends. It uses the job's saved permissions, and disabling or removing the job still stops its work. #168037, thanks to @obviyus.
• Home notifications. Creating many chats no longer fills Home's notification queue with a separate notice for each one. Recent new chats share one summary, leaving room for reminders and other notices on Home's next turn or heartbeat. See Home guidance. #168239, thanks to @steipete.

Browser and Computer Use

• Managed browser tabs. At the shared tab limit, locally managed browser cleanup favors untracked and older pages before recently used session tabs. The oldest eligible session tab can still be removed. #166666, thanks to @ruel225, @obviyus and @vokaplok.

Plugins and Integrations

• Plugin approval waits. Affected plugin tools can wait for your approval for the existing response window instead of being cancelled early as stalled. Unanswered requests still expire; see permission-request guidance. #161821, thanks to @fransqaas and @obviyus.
• Long tool conversations. Affected context-engine plugins keep older history within the model's available space as tools run, avoiding the reported context-limit failure. Oversized new tool output can still exceed the limit. #166120, thanks to @ruel225, @steipete and @igs-rogenlo.
• Codex model discovery cleanup. Timed-out model discovery releases abandoned background processes while active sibling work can finish, preventing that source of accumulating memory use. Later work can start a replacement; see Codex recovery guidance. #168365, thanks to @steipete and @609NFT.

Security and Privacy

• Credential masking. Built-in masking no longer misses recognized credentials split across long tool output. Long or repetitive text also avoids the reported masking freezes and stack overflows, with less processing work under the existing rules. #161480, #166701, thanks to @eleqtrizit and @steipete.
• HTTPS commands through the secret proxy. HTTPS commands using the secret egress proxy no longer fail strict certificate checks with newly generated certificates, including affected Python commands on macOS. After updating, restart the Gateway and start a new command. Keep certificate verification enabled. Retained older debug-proxy certificates are unchanged and can still fail strict checks. f1cd78c, thanks to @steipete, @vortexopenclaw and @A1fred-AI.

Quality-of-Life Improvements

• Configuration errors. When a helper cannot launch, configuration checks report the actual launch error and affected file instead of an unrelated trim error. #165373, thanks to @steipete and @nnljl2003.
• File and console logs. Masking secrets in file logs and JSON console output uses less temporary memory, with the same masking rules and output. #167943, thanks to @steipete.

Other Bug Fixes

• Quoted Windows commands. Commands on Windows nodes preserve multiword arguments, URLs and executable paths with spaces through the supported command wrapper. #158160, thanks to @Dimitri-lcht, @igs-rogenlo, @steipete and @kassol.
• Rosetta worktree and process checks. Intel OpenClaw running under Rosetta on Apple silicon uses Git for worktrees and avoids the crashing native process check. Maintenance that needs that process inspection still requires native arm64 Node. #165088, thanks to @steipete.
• Helper answers. Helpers using ACP or the shared CLI fallback pass their complete new answers back to the main agent, including the end of long reports and the answer for the right task. Delayed older results also stop replacing newer accepted answers. See announcement guidance. #165800, #166139, thanks to @yetval, @steipete, @holgergruenhagen, @potterdigital, @Leli1024, @zyz619963502zyz, @Yun-0000 and @michaeljgood71.
• Replies after helper results. The main agent can keep working on a helper's result without losing its final reply when the helper's delivery deadline expires. Normal task time limits and Stop still apply; see handoff guidance. #164282, thanks to @Oldrich333 and @obviyus.
• Background tools. Tasks no longer get cancelled as stalled after background tools have successfully finished, even if the next model response reaches its output limit. See queue guidance. #166997, thanks to @SunnyShu0925, @steipete, @hannesrudolph and @hippoley.
• Earlier images. With an image decoder available, corrupt images are omitted from requests so text follow-ups and reopened conversations can continue. Missing older images also receive a neutral note instead of instructions to deny the assistant's earlier inspection. #166564, #166165, thanks to @jayzhou2309, @steipete, @mirasrael and @nogikeyaksh.
• Conversation context on later turns. Context added to a message by prompt hooks or replacement prompts now stays with that message on later turns and after reopening the chat, while the visible chat keeps your original text. Replies after a helper finishes also keep the main conversation's instructions, reducing repeated prompt processing for affected local models. #167938, #167898, thanks to @steipete, @Eragim and @apalmiere.
• Stopped node workers. Force-stopping a node worker can stop surviving child processes and free its slot for later work once cleanup is verified. See background-process guidance. #168746, thanks to @steipete, @aniruddhaadak80 and @linhongyu510.