On this page

On this page

Release notes

v2026.10.4

v2026.10.4

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

OpenClaw v2026.10.4 fixes update and recovery problems, including interrupted work after sleep or clock changes and chats stuck after an update, alongside repairs for messaging, model tools and sign-in.

Release scale: 55 pull requests, 24 direct commits, and 49 contributors.

Web UI

  • Chat refresh after updates. Chats stuck on “Server updated” can refresh back to the same conversation link. If your tab was open before updating, save unsaved work and reload it once to load this fix.

    More details

    #168941, thanks to @steipete.

  • Dashboard connection. Open dashboards can stay connected when a busy Gateway delays its check-in messages, avoiding unnecessary reconnects. Requests still have their own timeouts.

    More details

    #171253, thanks to @steipete.

Updates and Maintenance

  • Sleep and clock changes. Fixed unnecessary interruptions after computer sleep or backward clock changes, with clearer restart guidance when OpenClaw loses access to its saved state.

    More details

    #170790, thanks to @steipete, @jjjhenriksen, @AnvinaLI, @sgh6688, @waynegault, @andersgreg and @shawnholt.

  • Gateway shutdown. Stopping or restarting OpenClaw no longer hangs on stalled conversation indexing, and pending session updates can finish before connected clients close. Saved transcripts remain available after a restart.

    More details

    #170493, #170555, thanks to @steipete, @Oldrich333 and @O96a.

  • Old login files. Unused retired OAuth files no longer block Doctor, updates or Docker startup, and remain preserved. Referenced files still need the documented 2026.9.7 recovery on the original host.

    More details

    #165420, thanks to @obviyus.

  • Update snapshots. After installing 10.4, later upgrades avoid false warnings that an unchanged database changed while its recovery copy was made. Real later changes can still prevent automatic restoration.

    More details

    #170351, thanks to @DonnieFi and @RomneyDa.

  • Updates after data migration. Update checks avoid false verification failures after data migration. Status also stops repeating a saved restart warning once the running Gateway reports the expected version. See rollback and recovery.

    More details

    #170130, thanks to @steipete and @angeliti999.

  • Temporary update files. Doctor and updates can remove abandoned temporary update directories despite unrelated system processes, including ordinary Linux systemd user sessions. Directories still used by live OpenClaw processes remain protected.

    More details

    #166335, #171488, thanks to @vincentkoc, @RomneyDa and @DonnieFi.

  • Temporary disk use. Update checks reuse eligible prepared database copies instead of duplicating them again, and give clearer retry guidance for space failures.

    More details

    #167451, thanks to @steipete, @iFadi, @dankarization, @azuretek and @davidcittadini.

  • Update progress. Temporary congestion while printing progress no longer aborts update inspection, and terminal progress waits through Doctor maintenance before resuming.

    More details

    #168493, #171454, thanks to @vincentkoc, @creanet-64, @RomneyDa and @DonnieFi.

  • Update failure guidance. Snapshot-preparation failures now point to the snapshot step that failed, instead of runtime verification that never ran.

    More details

    #171457, thanks to @RomneyDa and @DonnieFi.

  • Maintenance warnings. Completed plugin upgrades stop leaving stale “unfinished” warnings in update history. Expected database maintenance also stops producing false checkpoint-failure warnings.

    More details

    #170413, #170783, thanks to @RomneyDa, @DonnieFi and @steipete.

  • Scheduled work during updates. After installing 10.4, later update checks give scheduled jobs and cleanup a chance to save while the updater waits to record its progress.

    More details

    #171493, thanks to @RomneyDa and @DonnieFi.

  • Schedule inspection during recovery. Updates and Doctor can continue when optional memory dreaming schedule checks fail, preserving scheduled jobs. Retry pending dreaming repairs with openclaw doctor --fix. Successful job-list reads also survive temporary snapshot cleanup failures.

    More details

    #170754, thanks to @RomneyDa, @aniruddhaadak80 and @anhnt3.

  • Delegated update repair. Delegated repair work spends less time rechecking permissions and can reach repair tools before timing out.

    More details

    #166847, thanks to @steipete.

  • Mac service restart and repair. Fixed a false refusal when restarting OpenClaw from an external Mac terminal. If another start has already registered the background service, Doctor starts it if stopped and leaves it alone if running.

    More details

    #170354, #170740, thanks to @RomneyDa.

  • Offline app-owned repair. Standalone Doctor can repair the same account's configuration and state when an app-owned Gateway is verified offline. Restart or update through OpenClaw.app afterward.

    More details

    #170768, thanks to @RomneyDa and @shanselman.

  • Tailscale diagnosis in Doctor. When Tailscale blocks startup, openclaw doctor tells you whether to reconnect, sign in or request administrator approval.

    More details

    #166947, thanks to @RomneyDa and @odrobnik.

Messaging

  • Telegram polling. Fixed a polling-worker startup failure that prevented Telegram messages from arriving in affected installations.

    More details

    #154286, thanks to @xydt-juyaohui, @steipete and @markfietje.

  • IRC replies. Long replies containing Chinese, Japanese or Korean text and emoji arrive without losing forwarded chunks.

    More details

    #159730, thanks to @rodricksz4h5 and @obviyus.

  • Replies on macOS. Fixed a macOS reply failure that restarted the channel and repeated the incoming message.

    More details

    #166462, thanks to @ericcaiwx-star.

  • Feishu settings reload. Feishu DMs keep receiving replies when a settings reload overlaps a reply already underway, without needing an OpenClaw restart.

    More details

    #169226, thanks to @ruel225, @obviyus, @junniar and @PratikRai0101.

  • Discord voice changes. You can change voices and continue asking for agent checks in the same call without leaving and rejoining.

    More details

    #171029, thanks to @ericcaiwx-star, @steipete and @brainmonger.

  • Long streamed replies. Long replies that use OpenClaw's reply-to-message instructions can arrive with less delay, while keeping their text and intended reply target.

    More details

    #166782, thanks to @boeingchoco and @obviyus.

Memory

  • Manual compaction. /compact gives an active answer up to 60 seconds to finish. If it cancels the turn, its status tells you to resend the original request.

    More details

    #166163, thanks to @holny, @steipete and @alexph-dev.

Native Apps

  • Mac app recovery. The updated Mac app preserves an already-newer managed Gateway during setup and recovery, and can restore its missing Node background service using that installation.

    More details

    #170766, thanks to @RomneyDa.

Models and Providers

  • OpenAI reasoning with tools. Fixed tool-call failures and reasoning being switched off for supported reasoning models using the official OpenAI API.

    More details

    #166182, thanks to @obviyus and @AsTutoAI.

  • Claude CLI conversation history. Switching between a messaging channel and web chat keeps the latest history in the same Claude CLI conversation, preventing new branches that miss intervening actions.

    More details

    #167241, thanks to @steipete and @cmatrx0.

Browser and Computer Use

Plugins and Integrations

Security and Privacy

  • Tlon group-DM permissions. Unverified group-DM author claims no longer gain owner commands, approval/admin actions or elevated tool permissions. A group member can still pose as an allowed sender to chat with the agent, and replies go to the impersonated sender's private conversation. After approving a pending DM saved before updating, ask the sender to resend it.

    More details

    #165801, thanks to @eleqtrizit.

Quality-of-Life Improvements

  • Agent messages and follow-ups. Fixed a launch failure that could stop a message to another agent or a subagent follow-up before it produced a reply.

    More details

    #170077, thanks to @steipete, @mariorivas1 and @ghbalf.

Other Bug Fixes

  • Node worker memory limits. Node workers can honor their configured JavaScript heap limits when OpenClaw starts with a larger heap setting, while the main process keeps its own heap size.

    More details

    b77060f, thanks to @xydigitLybnnnn.

  • Streamed code examples. Long streamed answers that quote XML tool instructions in a code block avoid repeated processing that could delay delivery. The example stays visible, while actual leaked tool instructions are still filtered out.

    More details

    #171573, thanks to @steipete and @boeingchoco.

  • Repeated historical warnings. Later runs of a conversation stop repeating warnings for work already marked as blocked, making old failures less likely to look new. The blocked work is not repaired, and its records remain.

    More details

    #167595, thanks to @jalehman, @Clode2026 and @steipete.

Was this useful?