OpenClawDocs
On this page

On this page

View as MarkdownView this page as plain textOpen in ChatGPTAsk questions about this pageOpen in ClaudeAsk questions about this page

Release notes

OpenClaw v2026.10.7

v2026.10.7

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

This patch release fixes problems with updates, repairs, lost replies and conversation context, along with smaller app and provider issues.

Release scale: 61 pull requests, 26 direct commits, and 54 contributors.

Installation and Onboarding

Web UI

  • A /steer message that the agent has consumed now appears in the Control UI conversation without a stale queue chip.

    More details

    PR #180024. Inputs still waiting for the Gateway remain queued. Thanks to @RomneyDa.

  • Supported Control UI clients can now open the current agent's GitHub settings from the screen tool.

    More details

    Commit a16ff7f. Thanks to @RomneyDa.

Updates and Maintenance

  • A suspiciously truncated configuration no longer replaces the healthy recovery copy, so Doctor can restore your settings even if another edit follows the truncation.

    More details

    PR #152565. This covers suspicious external reductions of openclaw.json. Thanks to @LiuwqGit and @RomneyDa for the fix, and @yetval for the report.

  • Doctor avoids unnecessary archive work when repairing many empty sessions and protects a transcript that received new content after the repair began.

    More details

    PR #170122. Thanks to @steipete.

  • On Linux and FreeBSD, updates now identify a missing pacman or pkg executable and can continue the npm update path without telling you to repair a package database that is not there.

    More details

    PR #173293. Thanks to @steipete, @wangmiao0668000666, @sunlit-deng, and @DonnieFi.

  • Completed updates no longer repeat settled plugin migration warnings or tell you to wait for an update that has already finished.

    More details

    PR #173395. Current warnings still appear, and database path advice points to openclaw doctor --fix. Thanks to @steipete, @DonnieFi, and @bhedanikhilkumar-code.

  • An update rehearsal no longer tells you a LAN-bound Gateway is limited to loopback because of the rehearsal's temporary settings.

    More details

    PR #178155. Other Doctor findings remain visible. Thanks to @steipete and @DonnieFi.

  • Doctor can finish a verified session import and plugin repair when an older retained import file is missing. It still warns you that the missing history cannot be recovered.

    More details

    PR #178779. Thanks to @steipete, @jayzhou2309, @lostsock1, and @IliyaNexus.

  • Package updates made with this repaired updater keep database snapshots promised for manual recovery when Doctor later refuses automatic restoration.

    More details

    PR #180010. The fix applies to updates run by the repaired installed updater; it cannot change cleanup already running in an older version. Thanks to @RomneyDa and @DonnieFi.

  • A settled Doctor repair no longer repeats the same legacy checks and notices after its initial inspection.

    More details

    PR #180576. Doctor checks again when settings or plugins change or repair work remains. Thanks to @steipete and @MertBasar0.

  • Package updates now stop before staging a release that requires a newer Node than the Gateway service uses, leaving the current Gateway running. Check the Node executable used by that service, which may differ from your shell, and follow the required version range in the update message.

    More details

    PR #180769. This applies to registry package targets. Thanks to @steipete, @sglion777, @Almost-Certified-ops, @billpang1-sketch, @rayluo88, and @jcheng510.

  • Restoring a backup now retains supported older exec approval rules instead of replacing them with deny-all rules. Malformed rules remain denied.

    More details

    Stable backport PR #181465, from PR #164297. Thanks to @wangmiao0668000666 for the fix, @cristianoc for the report, and @RomneyDa for the backport.

  • Package update reports now separate a completed activation from later maintenance warnings, and an eligible interrupted publication can finish before another update starts.

    More details

    Stable backport PR #181553, from PR #181166 and PR #181307. Thanks to @steipete for the source fixes, @najef1979-code and @agroman1993 for the reports, and @RomneyDa for the backport.

  • Windows updates now handle both long and 8.3 short paths for plugins, so plugin capture works when Windows reports either form.

    More details

    PR #172768 and PR #172769, adapted in stable backport PR #181553. Thanks to @RomneyDa.

  • R2 backup verification and restore now resume downloads after a retryable interruption instead of starting that download over.

    More details

    Stable backport PR #181553, from PR #171292. If the R2 object changes during download, restart the command. Thanks to @steipete for the fix, @davidcittadini for the report, and @RomneyDa for the backport.

  • Post-activation Doctor now stops at its supplied time allowance instead of leaving an update waiting indefinitely.

    More details

    PR #173862, in stable backport PR #181553. On expiry, the updater stops the Doctor worker tree and follows its maintenance-warning or recovery path. Thanks to @steipete for the fix and @RomneyDa for the backport.

Messaging

Memory

  • Re-importing a Memory Wiki OKF bundle now preserves handwritten Notes inside its marked human Notes block. Handwritten text on older pages outside that block can still be replaced.

    More details

    PR #175707. The import adds a marked Notes slot to generated pages; put notes inside that block before re-importing. Thanks to @YangManBOBO, @RomneyDa, and @steipete.

Native Apps

Models and Providers

  • Model catalog refreshes no longer redo work for unchanged plugins when an agent adds a provider.

    More details

    PR #162226. Thanks to @netcrazed, @RomneyDa, and @kiralee.

  • With a Copilot model selected, setup verification and Ask OpenClaw replies now complete while ordinary Ask conversations keep their history.

    More details

    PR #163578. Setup uses the packaged runtime; replacing the SDK requires a restart and another verification. Thanks to @lifuyue and @RomneyDa.

  • Managed Claude CLI conversations now keep the parent's working context when child agents finish, with OpenClaw-configured MCP tools still available.

    More details

    PR #172026 and completion follow-up PR #176892. Restricted actions remain gated on completion turns. Thanks to @Marvinthebored, @VACInc, and @steipete.

  • Managed Claude CLI turns now use OpenClaw's tool surface. After upgrading, they stop loading Claude user settings, CLAUDE.md and rules, auto-memory, user hooks and plugins, Chrome, and Claude-only MCP servers. Existing conversations start one fresh native Claude session on their first turn after upgrading.

    More details

    PR #172026. OpenClaw-configured MCP servers remain available through its Gateway. Thanks to @Marvinthebored, @Peetiegonzalez, and @VACInc.

  • When the available tools really change in a CLI conversation, its next turn starts a fresh native session so the current tools appear. The OpenClaw conversation remains, although the native session's old context may need to be rebuilt.

    More details

    PR #176577. Account authentication can limit how much saved native history is restored. Thanks to @steipete and @jackylecalvez2.

  • A usable account profile pinned to an agent's primary model no longer makes clients block sending because of a false missing-auth result.

    More details

    PR #178522. The fix applies to the matching effective primary model. Thanks to @ericcaiwx-star and @steipete for the fix, and @eduardonunesfreire-rgb for the report.

  • Claude CLI conversations now keep their native context across Control UI, channel, and child-agent completion turns when only caller-specific tools change.

    More details

    PR #178955, commit 4fccf31, and commit f5083ae, which backports the transcripts-tool portion of PR #175309. The transcript tool stays listed on caller-less turns when enabled, but use without an authorizing identity is refused. Thanks to @Marvinthebored, @holgergruenhagen, @joey-at-jack, @obviyus, and @RomneyDa.

  • ClawRouter model lists can now discover available models after settings change while reusing provider policy data.

    More details

    Commit d4a081a. Thanks to @RomneyDa.

Automations and Scheduling

Browser and Computer Use

  • A multi-file apply_patch now catches missing targets, mismatched context, and invalid update text before changing earlier files in the patch.

    More details

    PR #160409. This protection covers problems caught before writing begins; errors during writing can still leave partial changes. Thanks to @masatohoshino and @obviyus.

  • Native computer control now recovers when a control session expires during use.

    More details

    PR #169385, in stable backport PR #181553. Thanks to @stackingrockss and @steipete for the fix, and @RomneyDa for the backport.

Plugins and Integrations

  • Claude CLI now lists and calls MCP tools under the current access rules, and refuses calls made with an outdated grant.

    More details

    Commit eb7c93b. Thanks to @RomneyDa.

  • External TypeScript plugins now start without repeatedly resolving local .js imports that point to .ts helpers.

    More details

    Stable backport PR #181553, from PR #168391. Thanks to @steipete for the fix, @jakeaadams for the report, and @RomneyDa for the backport.

  • An authorized operator can publish a GitHub pull request when their profile attaches after connecting, instead of receiving a false loss-of-authorization error.

    More details

    PR #169400, in stable backport PR #181553. Requests from another person's session or after access is revoked remain denied. Thanks to @Gigaxel and @steipete for the fix, and @RomneyDa for the backport.

Quality-of-Life Improvements

  • Interrupted or compacted agent turns now keep the request they were working on and more complete tool results. After a tool timeout, a main session may resume, but check for uncertain effects before retrying the tool.

    More details

    Commit a16ff7f. Unsafe tools remain restricted during recovery. Thanks to @RomneyDa.

  • Large batches of child-agent results now fit within a bounded completion message, shortening long answers and summarizing overflow before it reaches the parent.

    More details

    PR #172068, in stable backport PR #181553. The notice points to sessions_history for inspection, whose text fields can also be truncated. Thanks to @steipete for the fix and @RomneyDa for the backport.

Other Bug Fixes

  • An authorized reset can restore a Main session archived by restart recovery while keeping its session identity and transcript. The interrupted task is not replayed.

    More details

    PR #176675. This applies to the configured canonical Main session. Thanks to @steipete and @enki.

  • After a delegated turn fails to recover from a restart, its conversation can accept a fresh authenticated message once no pending reply or run still owns the work.

    More details

    PR #176682. The interrupted task is reported locally and is not replayed. Thanks to @steipete and @enki.

  • Independent database readers and borrowers now continue when another caller's database access is refused.

    More details

    PR #175987 and PR #179320, adapted in stable backport PR #181465. Thanks to @steipete for the source fixes and @RomneyDa for the backport.

  • An incompatible secondary-agent database now refuses that agent's sessions without blocking healthy agents or restarting the Gateway.

    More details

    Stable backport PR #181553, from PR #171372. Thanks to @steipete for the fix, @LifeViwer for the report, and @RomneyDa for the backport.

  • Gateway shutdown now drains retained worker tunnels and workspace transfers even when the worker inventory has already retired.

    More details

    Stable backport PR #181553, from PR #168862. Other cleanup failures still surface. Thanks to @steipete for the fix, @Captain69G for the report, and @RomneyDa for the backport.

  • After you stop a conversation, a later message can continue in the same session even if old recovery records remain.

    More details

    Stable backport PR #181867, from PR #181801. Thanks to @steipete for the source fix, @najef1979-code for the report, and @RomneyDa for the backport.

  • An agent can continue after its provider shortens a long conversation, without a false truncated-tool warning.

    More details

    Commit ee878eb. This covers the compaction replay checkpoint.

Was this useful?