OpenClaw 2026.10.8 Ships Update Recovery and Steadier Conversations as GPT-6.1 Terra Arrives for OpenClaw Agents
OpenClaw 2026.10.8 is the new stable release, built around updates that leave the Gateway running and conversations that keep accepting messages, with 2026.9.35 and 2026.10.10-beta.4 out the same night. Also: GPT-6.1 Terra, default-on update reports, Skill Workshop drafts and an exec allowlist fix.
2026.10.8 Is the New Stable Release, Built Around Updates That Finish
OpenClaw's newest stable build arrived before sunrise Pacific time. The GitHub release notes for openclaw 2026.10.8 were published at 4:56 AM Pacific on Tuesday, October 27, about half an hour after the package reached npm, where 2026.10.8 is now the latest tag. The notes describe an audited record of 162 unique pull requests since 2026.10.7. The docs site had no 2026.10.8 page yet, and the project revised its 2026.10.7 count after release, so treat that figure as provisional.
The first highlight is the one most operators will feel. An ordinary Doctor failure after an update now restarts the Gateway instead of leaving it stopped, and an orphaned maintenance worker no longer blocks every CLI command and Gateway start. Updates also stop aborting on ACL filesystems, Windows pnpm optional-dependency links, pnpm 10 global installs, debug console logging and a missing legacy OAuth directory. On macOS, the updater now waits for the serving Gateway to release its databases before Doctor runs.
The second theme is conversations that keep taking messages. Killed runs release their restart-recovery fence, interrupted and just-accepted child sessions recover after a Gateway restart, and inbound messages stop retrying forever when the reply runtime cannot publish. Startup work no longer inherits a closed SQLite reader scope, so models, chat metadata and channel replies stay available. Default Slack, Google Chat, X, Nextcloud Talk and Zalo accounts also keep starting and keep their secrets after a named account is added.
Take this release if your last upgrade left a Gateway stopped after Doctor, or if a named Slack, Google Chat or X account knocked the default account offline. Upgrade one Gateway first, confirm it comes back on its own, then check that every channel account still starts before rolling 2026.10.8 to the rest of the fleet.
Extended-Stable 2026.9.35 and Beta 4 Land the Same Night
Two more releases shipped Monday night. The openclaw 2026.9.35 extended-stable release notes, published at 7:20 PM Pacific, describe a Gateway-only build that the project calls its current equivalent to LTS, adding 78 fixes from 77 numbered pull requests to OpenClaw from the end of September. Long-running Gateways should stop accumulating memory, since worker heap caps now apply under service managers and idle Codex app-server clients retire after 30 seconds. Doctor also repairs stale full-text indexes left by the SQLite 3.54 tokenizer change instead of quarantining agent databases.
The openclaw 2026.10.10-beta.4 prerelease notes, published at 10:39 PM Pacific, cover 60 pull requests on the beta line. The most practical change is in the CLI: openclaw config set, config unset and openclaw setup now send their change through a running Gateway as one config.patch instead of refusing, which the notes say lets Windows companion setup complete again. Scheduled runs now wait for tool results before answering, so cron summaries stop reporting a premature "no results".
GPT-6.1 Terra Is Selectable in OpenClaw
On Monday at 3:44 PM Pacific, the official @openclaw post announcing GPT-6.1 Terra support said OpenAI's new model "is already available in OpenClaw": sign in with ChatGPT or use an OpenAI API key, then select openai/gpt-6.1-terra. It names no minimum OpenClaw version. We found no OpenClaw pull request naming the model ID, and the OpenClaw OpenAI provider guide does not yet mention Terra, so check your own model list rather than assuming every build shows it.
The OpenAI announcement introducing GPT-6.1 Terra positions it as a fast, efficient model for computer use and everyday communication work. OpenAI lists API pricing of $1 per million input tokens and $5 per million output tokens, 50% below GPT-5.6 Terra and GPT-6.1 Sol. It also says Sol and GPT-6 Astra remain the better choice for the most complex coding tasks.
Terra looks like a candidate for the high-volume side of an OpenClaw deployment: utility turns, form-filling and browser work, and scheduled automations. Before switching defaults, run your own evals on a few real automations and compare cost per completed task, not per token, and keep Sol or Astra on agents handling the most complex coding work, which is where OpenAI says they remain the better choice.
On Main: Update-Outcome Reports Become Default-On
One change merged Monday is not in a release yet but deserves attention before it ships. Pull request #154067 reporting update outcomes by default, merged at 10:06 AM Pacific, lets the updater send, by default, a short record of whether an update succeeded, failed or rolled back. The pull request says the record has 18 fields, such as versions, platform, duration and allowlisted failure codes, with no user or install identifiers, logs, paths or config. A maintainer approved sending it even when DO_NOT_TRACK or telemetry.enabled: false is set; the opt-outs are update.checkOnStart: false or OPENCLAW_NO_AUTO_UPDATE=1. Nothing is posted unless the receiving endpoint advertises support.
If your policy treats DO_NOT_TRACK as a blanket off switch, update reporting is the change to flag before your next upgrade cycle. The payload is narrow and documented, but it ignores that variable by design. Decide now whether managed hosts should set update.checkOnStart: false, and remember that also turns off automatic update checks.
Security Practice: Upgrade Allowlist Hosts, and Keep the Hook Token Away From Agents
An allowlisted builtin could run commands without an approval prompt
Both 2026.10.8 and 2026.9.35 close an exec bypass. Pull request #180880 requiring approval for test -v operands in allowlist mode explains that the allowlist classifier trusted test and [ as safe builtins without inspecting their operands, while bash 4.2 and later evaluates the array subscript of a test -v operand even inside single quotes. A prompt-injected model could therefore run an arbitrary command with the Gateway user's privileges, with no allowlist match and no approval prompt, on hosts whose exec shell is bash 4.2 or newer. The OpenClaw exec approvals guide covers the allowlist security mode and ask: "on-miss" prompting that this affected; after the fix those commands trigger the normal approval prompt.
- If any agent runs with allowlist exec approvals, move to 2026.10.8, or to 2026.9.35 on extended-stable.
- If you set
hooks.tokenfrom an environment variable, note that beta.4 now keeps it out of agent exec commands, CLI backends and other host exec children, per pull request #179199 keeping an env-referenced hooks.token away from agent commands. Before, any agent command could read it. The 2026.10.8 notes do not list it, and the OpenClaw inbound webhooks guide says every hook request authenticates with that token, so on stable builds disable unneeded hooks and rotate the token if an agent may have seen it.
Skill Card: Skill Workshop Drafts Bring Back Review Before Learning Applies
Propose mode returns on main
Since an earlier change, OpenClaw's background skill learning either edited skills on its own or did nothing, and Doctor silently switched configs that said propose to off. Pull request #181747 adding Skill Workshop propose mode, merged at 6:05 PM Pacific on Monday, brings the review step back. With skills.workshop.autonomous.mode set to "propose", background reviews save drafts instead of changing skills. Anyone Doctor already moved to "off" must opt in again.
The OpenClaw Skill Workshop guide says each draft is the complete proposed SKILL.md with its reason, never loads as a skill, and appears in a Drafts list in the Control UI, where an update to a live skill opens as a diff. Apply uses the normal versioned write, so Undo still works. In chat, /learn drafts has the agent judge waiting drafts together and ask before saving. Up to 50 drafts wait per agent.
For shared or customer-facing agents, propose mode is the setting we would choose once it ships. Skills change how an agent behaves on every future task, so a human diff review is cheap insurance. Check after upgrading that Doctor has not left you on off, and keep an eye on the 50-draft cap so learning does not quietly stop.
Community: A Voice Bar for Omarchy, and an Installer's Story
The Omarchy OpenClaw Voice plugin repository offers a native Omarchy bar control and speech bubbles for talking to an OpenClaw agent. Its README asks users to run OpenClaw main for a voice-permission fix, pull request #159697, that is in the 2026.10.10 beta line but not listed in the 2026.10.8 notes. The official @openclaw OpenClaw Hero post features @ArkStabler, who did pest control in Ocala, Florida nine months ago and now sets up OpenClaw agents for roofers, pool builders and a contractor.
Rolling 2026.10.8 across a fleet?
SEN-X helps teams stage OpenClaw upgrades, evaluate new models like GPT-6.1 Terra on real automations, and lock down exec approvals and hook tokens.
Contact SEN-X →