v2026.6.11

We heard the feedback. v2026.6.11 focuses on the rough edges that make OpenClaw feel less dependable, with fixes for misplaced replies, stuck sends, reconnects, model setup failures, and safer admin defaults.

Full release notes

Highlights

#### Channel delivery reliability

Delivery and reconnect fixes span Telegram, WhatsApp, Matrix, Google Chat, iMessage, Feishu, Mattermost, WebChat, the Control UI, and the terminal UI.

• Fixes newer Google Chat direct messages sometimes being treated like group conversations, so they reach the correct one-to-one chat while Space and group-chat messages keep their existing routing. #58993 Thanks @starhappysh, @vincentkoc.
• Feishu voice replies from OpenClaw now show their duration in the chat bubble, so recipients can see how long the audio is before playing it. #89172 Related #53798. Thanks @areslp, @fxz26284407, @kinrocw.
• Discord and Telegram replies and mirrored chat history stay tied to the intended conversation more consistently, including across repeated Telegram replies and session changes. #89911 Thanks @jalehman.
• Background image, video, and music results now return to the chat that requested them when the task starts without a full conversation target, instead of appearing to fail after creation or being sent to the wrong peer as the session moves. #89949 Related #86034. Thanks @tianxiaochannel-oss88, @wangwllu.
• Telegram answers now stay attached to the user's current question when they quote an earlier bot message, while quotes of other people's messages still reply to the selected quote. #90475 Thanks @moeedahmed.
• QQBot group admins can choose how broadly slash commands are available, and private-only commands now direct users to a private chat instead of being exposed or silently ignored in groups. #92154 Thanks @sliverp.
• Heartbeat checks using reasoning-capable models now show the assistant's intended reply instead of exposing internal reasoning in Telegram, WhatsApp, and other channels, while opt-in Thinking messages still work. #92356 Related #92260. Thanks @jmpei, @tangtaizong666, @vincentkoc.
• Telegram progress-mode chats now clear an old progress bubble before newer tool output or artifacts appear, keeping the conversation in a clean, readable order. #93002 Related #90753. Thanks @shadow-enthusiast, @zhangguiping-xydt.
• iMessage command-and-link messages now stay together as one OpenClaw turn when delayed link previews arrive, while unrelated quick messages remain separate for users who enabled same-sender DM coalescing. #93143 Thanks @omarshahine.
• Successful Discord replies sent through the message tool no longer trigger a misleading failure warning in affected message_tool_only source-channel turns. #94072 Related #93875. Thanks @chenyangjun-xy, @hoyanhan, @vincentkoc.
• WhatsApp group conversations now preserve the right message and group context more reliably during retries, reconnects, and group changes. #94338 Related #7433. Thanks @mcaxtr, @octopuslabs-fl, @xialonglee.
• Fixes OpenClaw sometimes replying to its own delayed iMessage echoes when stray leading characters keep the sent message from being recognized. #94442 Thanks @ly-wang19.
• Telegram webhook users can keep receiving DMs and group messages through brief channel restarts, configuration reloads, and recovery cycles without temporary message blackouts. #94506 Related #90254. Thanks @obviyus, @travellingsoldier85, @xialonglee.
• Matrix E2EE gateways can stay online during long-running use instead of gradually consuming memory until a crash takes channels and in-flight work down. #94942 Related #90455. Thanks @xzh-icenter, @yar-sh.
• Telegram users now see the intended native reaction instead of leaked instructions or a dropped reaction-only reply, with success recorded only after Telegram accepts it. #94977 Related #71140. Thanks @cuttingwater, @hugenshen.
• Telegram progress updates for commands, searches, updates, and API activity now stay readable instead of exposing noisy HTML or code-style rows, with plain-text fallback when Telegram cannot parse the formatting. #95007 Related #95002.
• Telegram conversations continued in WebChat now show one assistant reply per turn and keep later replies with the active conversation instead of duplicating answers or sending them back to Telegram. #95069 Related #94930. Thanks @heichaowo.
• Google Chat now hides misleading internal failure banners when a tool result is harmless, leaving users with the completed answer while normal assistant text remains unchanged. #95084 Related #90684. Thanks @jailbirt, @studentzhou-svg.
• Bound multi-agent channel conversations now load the workspace files for the configured agent instead of the default agent, though previously misfiled conversations may start fresh in the corrected agent store. #95118 Related #92903. Thanks @849261680, @axjing.
• People sharing an OpenClaw gateway can now assign different models to individual direct-message contacts across supported chat channels, while existing group and wildcard model choices keep working as before. #95120 Related #53638. Thanks @gandalf-at-lerian, @thomaszta, @xydigit-zt.
• Telegram now shows that OpenClaw is still working during short initial previews or progress-mode replies instead of leaving the chat silent until the final message arrives. #95183 Related #95004. Thanks @obviyus.
• Matrix users and operators now get a clear failure when a homeserver sends an oversized or stalled response, instead of OpenClaw continuing to buffer it and risking unbounded memory use. #95240 Thanks @alix-007.
• Fixes delayed or missing Telegram and other queued channel replies in Kubernetes-style deployments with many injected environment variables, where opening the queue database could stall the gateway. #95278 Related #94571. Thanks @kaka-srp.
• Telegram chats recover after one stuck message times out, allowing later messages in the same chat or topic to reach the agent without restarting the gateway. #95299 Related #95248. Thanks @kriegerbangerz-ship-it, @mikasa0818, @obviyus.
• When people switch between Telegram and another OpenClaw client in a shared direct conversation, short Telegram replies now follow the latest conversation instead of responding to an older, unrelated Telegram proposal. #95390 Related #95378. Thanks @maiduy708, @mikasa0818, @obviyus.
• Fixes completed assistant messages appearing twice in Telegram, Discord, Slack, and other streamed chats after a multi-message reply. #95432 Thanks @vincentkoc, @yetval.
• WhatsApp replies now stay attached to the direct or group message being answered instead of appearing as a separate message that loses the conversation context. #95483 Thanks @mcaxtr.
• Telegram rich-message replies now keep paragraphs, bullets, and status lines separated instead of collapsing multi-line content into one run-on block, with no configuration change required. #95532 Related #95409. Thanks @amknight.
• Mattermost operators who enable native slash commands can now use /oc_queue directly in Mattermost to tune active-run queuing, including its mode, debounce timing, cap, and drop handling. #95546 Thanks @amknight.
• Previously allowed messages keep reaching named accounts after legacy multi-account channel upgrades, with inherited DM and group access rules preserved across Mattermost, Discord, Slack, Telegram, Signal, WhatsApp, iMessage, and IRC. #95550 Thanks @amknight.
• Mattermost users can keep talking in a thread without mentioning the bot again after it replies, and that participation survives gateway restarts until the thread has been idle for seven days. #95552 Thanks @amknight.
• Inbound Telegram messages now reach the configured OpenClaw session promptly instead of sitting unanswered until the next polling interval, a gateway restart, or manual intervention. #95577 Related #86957. Thanks @freidrich-goldenflow, @liuwqgit.
• QQBot users now receive complete markdown tables when valid separators use one or two dashes per column, instead of losing the header and all but the final row. #95637 Thanks @ly-wang19.
• Synology Chat users can now receive agent replies that take more than 120 seconds when the configured core timeout allows it, instead of having the channel reject them early. #95707 Thanks @sahibzada-allahyar, @vincentkoc.
• Telegram forum-topic cron jobs now keep separately configured failure alerts going to their intended destination, even when the main announcement uses a topic in the same chat. #95794 Thanks @vincentkoc.
• Fixes WhatsApp group replies that could quote an older OpenClaw message instead of the user's triggering message, so final answers stay attached to the intended message when a reply target is available and avoid pointing back to stale bot context when it is not. #95914 Thanks @mcaxtr.
• WhatsApp users can approve or deny prompts by reaction without the prompt staying stuck when WhatsApp identifies the same direct chat differently, while group approvals remain tied to the correct group and person. #95935 Thanks @mcaxtr.
• Final reply processing now uses less CPU when OpenClaw checks whether block text was already sent, without changing which reply reaches the chat or how duplicate text is suppressed. #96087 Thanks @vincentkoc.
• Exec approval results from external channel plugins now return to the channel or DM where the command started instead of falling back to WebChat or seeming to disappear after approval. #96140 Related #96103. Thanks @lansenger-pm, @vincentkoc, @yetval.
• WhatsApp's final answer now stays quoted to the follow-up message a user just sent when replying to an older OpenClaw message, instead of arriving unquoted or pointing back to the older bot reply. #96220 Thanks @mcaxtr.
• Nextcloud Talk bots now ignore ordinary file-share and lifecycle events without logging them as bot errors or risking disabled delivery, while malformed chat payloads still return an error. #96243 Related #81566. Thanks @arkyu2077, @rafaelmgbh, @vincentkoc.
• Replies and message-tool delivery in Mattermost channels now use channel and thread guidance because the agent identifies those conversations as channels rather than group chats, while existing group-chat behavior remains unchanged. #96244 Related #95645. Thanks @arkyu2077, @iloveleon19, @vincentkoc.
• MCP channel integrations now keep conversation lists, message reads, event polls, and waits within predictable bounds even when a client requests excessive limits or timeouts. a39e548 Thanks @vincentkoc.
• Long-running streamed auto-replies are less likely to stop early or abort inconsistently when an unusually large timeout is configured. 6c85b90 Thanks @vincentkoc.
• Channel progress now shows a repeated status when work genuinely returns to it after another update, instead of hiding useful context as a duplicate. 8a75c4d
• Completed channel replies no longer gain late progress notices, preventing stale status text from appearing after the answer is finished. a594d2c Thanks @vincentkoc.
• During streaming channel replies, progress messages now keep showing the latest state instead of getting stuck on an older update. e114001 Thanks @vincentkoc.
• Matrix forced resets now handle unavailable secret storage without a runtime error, treating recovery access as unavailable so the reset path can continue safely. 5c5a8a4 Thanks @vincentkoc.
• Configured channels now remain visible in openclaw channels status --json, while scheduled announcements reject stale entries that have no active plugin to deliver them. a641c0d
• Discord voice conversations now keep back-to-back assistant responses moving, so a queued reply plays after the previous audio stream closes instead of remaining stuck. 88b64e4 Thanks @vincentkoc.
• Discord progress previews are less likely to stop before the final edits when an agent response has already started arriving. 86ea382
• Chats no longer show stray NO_REPLY text when the assistant means to stay silent, while legitimate media responses still arrive without the placeholder. 96c6f80
• Telegram streaming replies now show each progress heading once, keeping tool and search updates easier to scan. 013e33c Thanks @vincentkoc.
• Telegram messages that get stuck after a long-running task, crash, or gateway restart now resume processing automatically, so later messages no longer wait silently or require operators to repair the queue by hand. #97543 Thanks @romneyda, @vincentkoc.

#### Provider and model recovery

Model setup, OpenAI, OpenRouter, opencode-go, and fast-mode follow-up behavior now recover more clearly in affected configurations.

• MiniMax text-to-speech and voice notes are less likely to fail because OpenClaw now explicitly requests the audio format it can decode instead of relying on provider defaults. #73079 Thanks @efe-arv.
• Gateway operators can again see provider, model, request status, and timing details in normal logs, making model-routing and transport problems easier to diagnose without enabling extra debug logging. #89648 Related #89300. Thanks @enominera, @xiaobao-k8s.
• Models reached through Google, Mistral, OpenAI Responses, Azure OpenAI Responses, and ChatGPT/Codex Responses now receive clean system instructions without OpenClaw's internal cache-boundary marker leaking into the prompt. #89716 Thanks @enominera, @masatohoshino.
• Cron tool calls using Gemini models through OpenAI-compatible providers now run without nullable fields triggering provider schema rejections. #91559 Related #91542. Thanks @pick-cat, @qiukui666.
• Provider-qualified model IDs now honor their configured agent runtime policies and CLI aliases instead of unexpectedly falling back to OpenClaw's default runtime. #91724 Thanks @vincentkoc, @yu-xin-c.
• The chat /models list and other plugin-aware model or provider selection paths now respond quickly instead of stalling for seconds and consuming a CPU core through repeated setup scans, while plugin changes still refresh normally. #93356 Thanks @obuchowski.
• Hosted Ollama Cloud users can keep only the models they explicitly configured after a restart, without the full shared catalog being added back, while automatic discovery continues for local and self-hosted Ollama servers. #93956 Thanks @jason-allen-oneal.
• Cron jobs can now retry or switch to a configured fallback model when a local provider returns the generic LLM request failed. error, instead of failing with the fallback unused. #94062 Related #93931. Thanks @hugenshen.
• Expired provider tokens no longer bury useful operator logs under repeated fallback warnings, while the first warning and later duplicate summaries remain available for diagnosis. #94233 Related #56979. Thanks @goutamadwant, @yanan1991.
• Google Gemini 3.5 Flash can now be selected with its full 1,048,576-token context window, avoiding missing-model errors and needless prompt-size rejections. #94726 Related #94723. Thanks @ajwan8998, @anguslogan01, @kevinat.
• Dashboard child sessions now handle allowed provider-qualified model choices consistently and give accurate recovery guidance when saved model state is stale. #94752 Related #94713. Thanks @gr4via.
• Claude CLI users no longer get promises of completion updates that may never arrive, because OpenClaw now blocks unsupported native background work before it can strand progress. #95008 Thanks @anagnorisis2peripeteia.
• OpenClaw now rejects oversized provider catalog or JSON responses with a clear error before buffering the entire response in memory. #95218 Thanks @alix-007.
• OpenRouter users can now select and run the advertised short DeepSeek V4 model IDs without requests failing with model_not_found because OpenClaw sent a duplicated provider prefix. #95268 Related #95198. Thanks @daniel-alejandro-t, @darren2030.
• With /reasoning on, DeepSeek-style OpenAI-compatible models now show the final answer separately from their reasoning instead of folding it into the reasoning block, with no configuration change required. #95283 Related #95280. Thanks @marvinthebored, @vincentkoc, @zengwen-dt.
• When a Codex subscription reaches its usage limit, OpenClaw now moves to configured fallback models instead of stopping on the failed result, and it does not retry runs that already produced visible output. #95400 Thanks @jason-allen-oneal, @sallyom.
• LM Studio users can now run quantized or multi-variant local models without false assistant-turn failures or phantom suffixed model entries caused by mismatched model keys. #95401 Thanks @monkeyleet.
• Google-backed embedded-agent runs now stop reading oversized or never-ending prompt-cache responses before they can exhaust memory or leave the run stalled. #95417 Thanks @alix-007.
• OpenRouter model scans fail safely on oversized or malformed catalogs instead of risking excessive memory use that can destabilize OpenClaw. #95418 Thanks @alix-007.
• OpenRouter setups now reject oversized model catalogs before they can exhaust OpenClaw's memory, without caching or immediately refetching the failed response. #95420 Thanks @alix-007, @sallyom.
• Configured fallback models can now answer when Claude CLI runs out of credits or hits a generic runner failure, instead of leaving users with the failure message as the final response. #95508 Related #95489. Thanks @mikasa0818, @riazrahaman, @sallyom.
• Gemini-backed web searches using freshness: "day" or pd now complete instead of failing with a provider 400 error, while broader freshness choices and explicit date ranges retain stricter filtering. #95682 Thanks @sunjae-k, @vincentkoc.
• Follow-up answers from xAI reasoning models such as Grok Composer now preserve earlier reasoning context more reliably, even when configurable reasoning effort is unsupported. #95686 Thanks @fuller-stack-dev, @geraint0923.
• Vercel AI Gateway users can now run models chosen from the live catalog, including live-only model IDs that are absent from OpenClaw's bundled list. #95710 Thanks @vincentkoc.
• Fixes manifest-defined providers turning valid model IDs into broken ones when stripPrefixes entries have stray spaces or different casing, so operators and plugin authors get the intended provider model. #95744 Related #95743. Thanks @parveshsaini.
• First-run setup now opens the credential prompt for a newly installed external provider instead of appearing to loop and leaving OpenAI selected. #95792 Related #95765.
• Oversized or stalled provider catalogs now fail quickly with a clear error instead of hanging OpenClaw or consuming unbounded memory, while normal catalogs continue to load. #95827 Thanks @alix-007.
• Xiaomi Token Plan users can now use up to 128K output tokens with mimo-v2.5 and mimo-v2.5-pro instead of being stopped at the outdated 32,000-token limit. #95934 Thanks @idootop.
• Tool-heavy model responses can stream with less overhead while repeated tool-call IDs and encrypted reasoning details stay matched to the correct call across Google and OpenAI-compatible providers. #95957 Thanks @vincentkoc.
• Token-usage accounting is more reliable for bundled ACPX users because OpenClaw now includes ACPX 0.11.2's persistence fix by default, without a separate package override or manual client update. #96124 Thanks @vincentkoc.
• Ollama Cloud users can now find and select glm-5.2:cloud with its 1,000,000-token context window, reasoning, and tool support even when it is absent from the public model list. 11484f8
• MiniMax image-understanding requests no longer fail before reaching the provider when a timeout is zero, negative, or extremely large; invalid values now use a normal or safe maximum wait. 4b6182e Thanks @vincentkoc.
• Codex runs now follow the current fast-mode choice instead of carrying over an old speed tier, and the status line clearly shows when fast mode is automatic. 77012f9 Thanks @vincentkoc.
• Codex-backed conversations now return to normal routing after automatic fast mode is cleared, preventing later turns or model changes from reusing a stale priority tier. 8afc1f7
• Fallback agent runs now honor each model's configured automatic fast-mode cutoff even when fast mode is overridden for the run, keeping fallback behavior aligned with the selected model policy. efd3172
• Live model-switch retries now preserve the original fast-mode cutoff for long-running sessions, while explicit fast mode avoids misleading automatic-cutoff progress messages. d990115 Thanks @vincentkoc.
• Embedded agent runs now keep automatic fast mode working consistently through retries and progress updates without confusing it with a manually selected fast-mode setting. cf1b6fe Thanks @vincentkoc.
• Fast-mode runs now keep their speed setting through model fallback retries and show the configured automatic threshold in status, avoiding inconsistent retry behavior and an unhelpful generic label. aa3797c Thanks @vincentkoc.
• Agent replies and scheduled cron runs now handle fast-mode fallback retries more reliably, keeping the state needed for the final attempt to finish or report progress correctly. 14e448e Thanks @vincentkoc.
• Users no longer see a fast-mode reset notice while model fallback attempts are still running; it appears only when the run reaches its final fallback attempt. 6eb72a8
• Users and operators now get clearer handling when a configured live model becomes unavailable because OpenClaw recognizes the provider's "selected model was not found" response as a model-not-found failure instead of a generic error. 2405d02 Thanks @vincentkoc.
• Qwen and vLLM now preserve existing chat-template settings consistently when thinking is switched on or off, and provider plugins can use the same tested helper. 2ba9d6e Thanks @vincentkoc.
• OpenAI-compatible proxy providers can handle thinking levels and legacy reasoning_effort fields more consistently, with plugin developers and provider maintainers using one documented normalization helper across OpenRouter, Kilocode, and the SDK. 35bafea
• Browser and Vite builds can now load the OpenAI ChatGPT Responses provider without a server-only dependency breaking the bundle, while WebSocket failures still appear normally. 8c8eb86 Thanks @vincentkoc.
• OpenRouter model scans now accept the same larger valid catalogs as runtime discovery while still rejecting oversized responses before they can consume unbounded memory. ad3b2f4 Thanks @vincentkoc.
• OpenAI Responses users, including affected Bedrock Mantle GPT-5.x reasoning setups, now get one clean final answer with aligned saved transcripts and replay context instead of dozens of repeated cumulative copies. #92399 Related #91959. Thanks @amersheeny, @daimingnj, @phoenixyy, @pigfoot.
• Scheduled jobs and isolated sessions using opencode-go models now move stalled requests into configured timeout or fallback handling instead of hanging for minutes before ending with a generic LLM request failed error. #93965 Related #93610. Thanks @forceconstant, @zhangguiping-xydt.

#### Session, memory, and trust continuity

Sessions, compaction, memory, QMD-backed memory, and Tool Search retain useful state more consistently, while Matrix recovery, tool policies, and approvals stay attached to the intended trust boundary.

• Affected agent conversations using OpenAI Responses can now recover and keep replying after a visible channel response leaves their saved history incomplete, instead of every later turn failing before a reply appears. #84708 Thanks @anyech.
• When a Codex-backed agent produces unusually large tool output, saved and replayed conversations now keep its text within the usual size limit while leaving non-text content unchanged. #87912 Thanks @adrianip0204.
• Control UI conversations now stay visible and continue in the same session after a sleep, network drop, or Gateway reconnect instead of disappearing when the next message is sent. #89017 Related #87700. Thanks @zhangguiping-xydt, @asicoe.
• Bundled Codex and Copilot integrations now keep mirrored chat history and transcript updates tied to the correct OpenClaw session as storage evolves, while existing file-backed active transcripts continue working during the migration. #89518 Thanks @jalehman.
• WebChat's current-session status now matches the conversation you are actually using, so the session identity, thinking level, token context, and cost details no longer come from the fallback main session. #89800 Related #89773. Thanks @killo3967, @sweetcornna.
• Your conversation is less likely to lose its context after you press stop during automatic compaction because the compaction request is now cancelled too. #89886 Related #89868. Thanks @lykeion-dev, @openperf, @vincentkoc.
• When cross-agent session access is blocked, OpenClaw now lists all required visibility, agent-to-agent, and allow-list settings, helping operators correct policy configuration instead of chasing a nonexistent agent failure. #90489 Related #90443. Thanks @ramitrkar-hash, @sahibzada-allahyar, @vincentkoc.
• openclaw memory status now shows an active light or REM dreaming phase instead of incorrectly reporting Dreaming: off, so operators can see that valid memory configurations are enabled. #93113 Related #67868. Thanks @agentarclab, @mrossit.
• Timed-out QMD memory searches now stop their background work when the agent moves on, preventing abandoned processes from continuing to consume CPU and memory. #93394 Thanks @alix-007.
• Repeated instructions sent after compaction now remain in the conversation, preventing lost turns, orphaned replies, and malformed history that some providers reject. #94328 Thanks @vincentkoc, @yetval.
• Memory Wiki's Stale Pages report now leaves durable concept and synthesis pages out of freshness warnings, keeping attention on source and entity pages that may actually need review. #94369 Thanks @sunnyshu0925, @vincentkoc.
• Long embedded runs with recent progress are now less likely to be interrupted by stale-session recovery, while genuinely stalled runs can still be cleared so queued work continues. #94701 Thanks @imadal1n, @mrclawfield.
• Ollama memory search now respects a configured smaller embedding dimension and keeps indexes for different dimensions separate, avoiding incompatible vectors being mixed together. #94811 Thanks @mushuiyu886.
• Memory searches and targeted refreshes now stay connected to the correct OpenClaw session even when transcript filenames change or QMD exports use a different name. #95087 Thanks @jalehman.
• Long-running conversations with screenshots or other images now keep their continuity more consistently when OpenClaw makes room for new messages, instead of repeatedly filling up without moving the retained conversation forward. #95128 Thanks @yetval.
• Windows users can now run QMD-backed memory indexing and search through configured absolute memory.qmd.command paths, including drive-letter and UNC locations, without OpenClaw stripping the path separators before launch. #95274 Related #92302. Thanks @ardooken, @ly85206559.
• Usage footers selected with /usage full or /usage tokens now remain visible after daily or idle session rollover, so users do not have to turn them on again. #95322 Thanks @litang9.
• Follow-up replies, reactions, threaded messages, and status checks stay with the chat they belong to after webchat or system activity, while real channel switches still clear outdated routing details. #95467 Thanks @yetval.
• Long-running main conversations now keep their prior context when users return after an overnight or delayed follow-up, rather than silently starting over after an otherwise normal completion. #95472 Thanks @xydt-tanshanshan.
• People with large session histories can list, preview, and find sessions without multi-second freezes, while older mixed-case session keys are still migrated at startup. #95699 Thanks @jalehman, @jzakirov.
• Fixes delivered replies sometimes being saved to the wrong conversation history, or omitted from it, when operators use a custom or per-agent session.store, improving continuity and auditability for the intended session. #95782 Related #95781. Thanks @youngting520.
• Saved session-memory summaries now leave out raw model tokens, tool-call blocks, media placeholders, role tags, and stale NO_REPLY markers so future conversations keep useful context. #95791 Thanks @sweetsophia, @vincentkoc, @yb0y.
• Long-running OpenAI sessions using Codex/ChatGPT OAuth can now compact without a separate API key, whether /compact is run manually or triggered automatically. #95831 Related #95693. Thanks @sallyom, @yui-tien.
• Long, tool-heavy sessions now compact oversized conversations instead of getting stuck when a large tool result appears at the end. #95860 Related #78478. Thanks @jw8957, @wzhgba, @yetval.
• When memory_search is unavailable because the Node runtime lacks node:sqlite, OpenClaw now points users to a compatible runtime instead of sending them through unrelated embedding-provider troubleshooting. #95916 Thanks @rrrrrredy, @vincentkoc.
• Developers and operators inspecting a compacted Copilot session now get its summary, before-and-after token counts, and session details instead of an incomplete result. #96049 Thanks @vincentkoc.
• The /stop and abort commands now keep stopping active runs, clearing queued followups, and ending related subagents promptly even when session keys need canonicalizing or abort metadata cannot be saved. #96201 Thanks @jalehman.
• Voice Wake upgrades now keep existing trigger phrases and routing rules working as OpenClaw moves them from retired settings files into the shared state database. bdf81a8
• Upgrades from older OpenClaw state layouts now preserve update notifications, check throttling, available-version records, and automatic-update attempt history as that state moves into SQLite. eb00d49 Thanks @vincentkoc.
• Plugin-channel conversations keep their intended session more reliably through startup, doctor checks, and state repairs, with older binding records migrated into OpenClaw's shared database. 9f888d9
• Windows memory-backed session syncing now keeps using the intended transcript file even when path formatting differs. b3b5b08 Thanks @vincentkoc.
• Embedded agent runs with a missing or blank session key now stay attached to the intended session instead of being sent through inconsistent session routing. 911f853 Thanks @vincentkoc.
• When a model guesses the wrong tool name, Tool Search and Code Mode now show how to find and retry the correct tool, reducing the risk that long-running sessions get stuck or lose durable memory during compaction. #93374 Related #92273. Thanks @mushuiyu886, @poison, @vincentkoc.
• Fixes assistant replies disappearing from webchat, Control UI, Feishu, and other embedded conversations after compaction, keeping refreshed chats readable and follow-up requests separate. #95484 Related #76729. Thanks @maweibin, @njuboy11, @vincentkoc.
• OpenClaw memory features now keep active, reset, and deleted transcript coverage aligned with configured session stores and agent ownership, making dreaming, QMD exports, indexing, and sync less likely to miss or misattribute conversation history. #96162 Thanks @jalehman.
• Gateway TLS setup now rejects blank certificate or key paths clearly or uses OpenClaw's defaults, avoiding confusing startup and certificate-generation failures while preserving valid paths. #94054 Thanks @miorbnli.
• Configured plugin policies keep blocking or rewriting sensitive tool calls after Gateway registry changes, reloads, or later hook initialization instead of being silently skipped. #94545 Thanks @jesse-merhi.
• Mobile operators with operator.approvals can now see and resolve chat-triggered exec approvals on the iOS device that started the request, including while the app is open, without relying only on push notifications. #95175 Thanks @joshavant.
• Control UI users now get the patched DOMPurify release, reducing exposure to the GHSA-cmwh-pvxp-8882 sanitizer vulnerability without changing how the interface behaves. #95691 Thanks @vincentkoc.
• "Always allow" approvals for plugin conversation bindings now carry over from the old settings file and are less likely to be lost or overwritten when multiple OpenClaw processes are running. ae41b00 Thanks @vincentkoc.
• Matrix users now see that the active recovery key is required before a forced cross-signing reset can proceed, preventing a second reset from leaving encryption recovery and room-key backups unusable. #95720 Related #78396. Thanks @jteddy, @vincentkoc, @xialonglee.

#### Slack router relay mode

Slack router relay mode gives managed and multi-gateway deployments a supported way to centralize incoming Slack traffic while preserving gateway ownership of mentions, threads, and replies.

• Managed Slack deployments can now use a central router to send mentions and ongoing threads to the right OpenClaw gateway while replies still appear through Slack. #94707 Thanks @pash-openai, @sjf-oa.

#### Raft External Agent wake bridge

The Raft channel and Raft plugin now support the local CLI wake path for External Agents.

• Raft External Agent operators can now wake an OpenClaw agent when a workspace has pending work through the supported local CLI bridge, with named profiles and checks for missing CLI prerequisites. #95497 Thanks @vincentkoc.

#### Official plugin installation and repair

Plugin management and the plugin inventory now cover more official integrations through normal installation, update, and repair workflows.

• When plugins.allow uses a channel or package name instead of the real plugin id, startup guidance now identifies the unmatched entry and shows the discovered plugin ids needed to correct the configuration. #68389 Related #68352. Thanks @aym9999, @jirboy, @lyfuci, @pahuchi-joe, @zmxccxy.
• Plugin trust warnings for first-time or fresh installs now include a ready-to-copy plugins.allow example and commands to list or inspect plugin ids, so users can resolve the warning before trusting or reinstalling plugin code. #78105 Related #68780. Thanks @jirboy, @pahuchi-joe.
• Codex migrations now work with standard global plugin installs because openclaw migrate can find the installed provider instead of failing with Unknown migration provider. #89612 Related #89609. Thanks @mugabuga, @zerone0x.
• Plugin installs and updates recover from stale OpenClaw-managed dependency pins instead of failing with npm EOVERRIDE, without later synchronization downgrading or removing packages users installed explicitly. #91786 Related #91772. Thanks @amknight, @mkdelta221.
• Channel plugin developers can now carry native sender and conversation identifiers through hooks and selected exec workflows, giving integrations more precise routing without breaking existing sender and chat fields. #91903 Thanks @lanzhi-lee, @vincentkoc.
• Plugin discovery now repeats fewer blocking filesystem checks during startup, reducing avoidable cold-start work for bundled plugin trees, especially on slower Windows filesystems, without changing bundle discovery behavior. #93919 Related #76209. Thanks @ml12580, @shenhonglong456-ai.
• Plugin Gateway methods now work through openclaw gateway call after registration, so plugin authors can use them from scripts and cron jobs instead of hitting an unknown method error. #94154 Related #94127. Thanks @brycemurray, @pick-cat, @vincentkoc.
• ClawHub skill discovery and install checks are less likely to stall or crash OpenClaw because oversized or stalled marketplace responses are now stopped before they can exhaust memory. #95226 Thanks @alix-007.
• Pinned official plugins no longer stay on an old release when operators follow the repair advice from openclaw doctor or deep gateway status after an upgrade. #95541 Thanks @ooiuuii, @vincentkoc.
• Managed npm plugin updates are less likely to break work on a running gateway with missing-module errors, because the older plugin files remain available until a later gateway start cleans them up. #95589 Thanks @ooiuuii, @vincentkoc.
• Official plugin cards for supported brands now show recognizable icons in ClawHub and other catalogs, and plugin authors can provide marketplace artwork through the documented manifest field. #95845 Thanks @patrick-erichsen.
• Official plugin icons in ClawHub and other catalogs are no longer forced into the same hard-coded color, allowing Simple Icons to use its default artwork instead. #95987 Thanks @patrick-erichsen.
• Docker users now have an official openclaw/openclaw Docker Hub mirror alongside GHCR, with versioned beta releases kept from moving the stable latest and main aliases. #97122 Thanks @