OpenClaw Pitches 2026.10.5 on Skill Workshop Undo and Public Chat Links as Manual Upgrades Trip on Skipped Doctor Runs
The official 2026.10.5 announcement leads with the sidebar, versioned Skill Workshop skills and no-sign-in conversation links. Meanwhile, upgrades done outside openclaw update are exiting 78 until Doctor runs. Plus: what a public chat link really exposes, fresh memory reports on 10.5, and the ClawSec suite.
The Official 2026.10.5 Pitch Is About Everyday Use, Not Recovery
When 2026.10.5 went stable on Sunday, its GitHub notes led with update recovery and Doctor fixes. The project's own announcement tells a different story. The official @openclaw announcement of v2026.10.5 on X, posted at 12:11 PM Pacific on Monday, picked three highlights: a sidebar that "keeps your agents within reach," Skill Workshop 2.0 with Undo, and public conversation links that need no sign-in. It credits 2,250 pull requests and 637 contributors.
The formatted OpenClaw 2026.10.5 release notes on docs.openclaw.ai are now live and open the same way: the release "refreshes the sidebar, adds saved versions and Undo to Skill Workshop, and lets people read published conversations at their normal chat links." The GitHub release page now carries a short summary pointing to that page and to a plain Markdown changelog. In the sidebar, long agent lists become rows with search once more than six agents exist, and new Subagents and Processes panels beside the chat let you read recent command output and stop one task without leaving the conversation.
The project choosing these three features for its public pitch says who 2026.10.5 is for: people living in the Control UI every day, not only operators. For teams, the Processes panel is the practical win, because stopping one runaway command without abandoning the whole conversation used to mean opening a terminal.
Skill Workshop 2.0 Adds Saved Versions, Undo and a Cost Warning
Skill Workshop holds the skills an agent writes for itself. According to the OpenClaw Skill Workshop documentation, every change now saves the previous version first and can be undone, and the newest 10 versions of each skill are kept. You can press Undo on the "Learned" notice, say "undo" in the chat, send /learn undo <id> to revert a whole background review, or run openclaw skills workshop restore <name>. A restore saves the current copy first, so an undo can itself be undone.
A background review runs after 10 accumulated model iterations in a conversation, and /learn asks for a save on demand. Learned skills unused for 30 days are archived, not deleted, though that cleanup stays off for Codex agents. Administrators should know two more details: learned skills bypass per-agent skill allowlists, so archiving is the way to hide one, and a write is refused when the security scanner finds a critical issue such as a literal secret. The release notes add a cost warning: background learning "sends conversation and tool content to your configured model provider and can add usage charges." Mode Off stops it while /learn keeps working. On upgrade, Auto and unset settings are preserved and legacy proposal modes move to Off.
Versioned learned skills remove the biggest objection to letting agents edit their own procedures, because there is now a clean way back. Still, decide on purpose. If conversations carry client data or your provider bills per token, start with learning Off and use /learn, then review the change feed before switching to Auto.
Security Practice: Treat a Public Conversation Link as Publishing
Published chats are readable by anyone with the link, including future messages
Public links are the 2026.10.5 feature most likely to leak something. The release notes say "the link exposes existing and future user messages and final replies to anyone who has it," that redaction is best effort, and that turning sharing off cannot recall saved copies. The OpenClaw guide to Control UI URLs and public session transcripts lists what readers see: every user and assistant message, progress updates, images, and collapsed tool steps with tool names and short command, path or query summaries. Tool results and reasoning stay private. The guide warns that "Included text is not secret-masked" and that the thread URL "is not a secret capability."
- Only session creators and Gateway admins can publish. Read the whole thread for keys, hostnames and paths first, since command summaries are public too.
- Publish a fresh, purpose-made session rather than a working one, because later messages become public as they arrive.
- Behind a login proxy, allow anonymous access only to
/chat/*and/share/*. Keep the app root, WebSocket, APIs, media, dashboards, settings and the/__openclaw__/session-entryhandoff behind authentication, and never bypass auth for the whole host. - After revoking, assume the content is out: downloaded copies remain, and preview images may stay cached by recipients for five minutes.
Manual 2026.10.5 Upgrades Are Tripping Over a Skipped Doctor Run
The newest upgrade reports share one cause: something other than openclaw update replaced the package. In issue #178177 on an unattended 2026.9.9 to 2026.10.5 upgrade, filed at 11:46 PM Pacific on Monday, an npm install on a systemd host exited with status 78 twice: first on a removed skills.workshop setting from the old proposal system, then on an agent database still at schema 24. Recovery then left a dead update_runs row that blocked plugin convergence until the reporter edited SQLite by hand.
OpenClaw maintainer steipete replied on GitHub at 2:25 AM Pacific Tuesday that three of those problems are fixed on main: config commands report a retired key instead of hanging (#178289), a killed repair no longer leaves a zombie run under pull request #178292 reconciling dead update repairs, and Codex and Voice Call migrations stop repeating warnings (#178286). He said the crashes came from the manual npm path, which skips the Gateway stop and Doctor run that openclaw update performs, and gave the manual order: openclaw gateway stop, npm i -g openclaw@latest, openclaw doctor --fix, openclaw gateway restart. None of the three fixes is in a release tag yet.
Hosts where the operating system owns the package hit the same gap. Issue #178346 on systemd installs lacking a pre-start Doctor step, opened at 3:50 AM Pacific today, describes a Fedora bootc upgrade to 2026.10.5 whose startup refused the schema-24 database and exited 78. The generated unit sets RestartPreventExitStatus=78, so systemd never retried and the Gateway stayed down for about 27 hours. The OpenClaw updating guide notes that the container image entrypoint, by contrast, runs Doctor before starting the Gateway. Triage flagged the issue for a maintainer decision; the reporter's interim workaround is openclaw doctor --fix after each image upgrade plus an OnFailure= alert on the unit.
If Ansible, a distro package, Nix or an image pipeline upgrades OpenClaw for you, that pipeline now owns the Doctor step. Put the stop, install, doctor --fix, restart order in the playbook and add an OnFailure alert to the unit today. Exit 78 is deliberate, and systemd will wait silently until a person notices.
Release Watch: No New Tag, and Memory Reports Move to 2026.10.5
No release has appeared since the OpenClaw 2026.10.10-beta.2 prerelease notes at 5:23 AM Pacific on Monday, so the update-hang fixes on main and today's upgrade fixes are not in any shipped build. Both P1 hang reports, #173764 and #174842, were still open on Tuesday morning, with no new comments since Friday.
Memory is the other thread to watch. Issue #169797 on Gateway memory use since 2026.9.9 showed peak memory on a single-CPU host rising from 706 MB on 2026.9.4 to 1,938 MB on 2026.10.1, with startup and first replies three to five times slower. Two field reports on final 2026.10.5 arrived in the last day: a Raspberry Pi 5 running five agents measured 3.20 GiB of proportional memory with 20 worker isolates, and a four-CPU Linux host measured a 2.24 GiB Gateway process. Both reporters say these are production systems, not benchmarks or proven leaks. The issue remains open at P1, and the helper-startup pull request #170187 is not merged.
Small hosts deserve a fresh measurement before they take 2026.10.5. On a Pi or a one-CPU VPS, check memory with your real agent and channel count after a clean restart, leave headroom for transcript workers, and keep the previous version ready to restore if the box starts swapping.
Skill Card: ClawSec Suite Puts a Signed Advisory Check in Front of Installs
clawsec-suite, signed advisories and approval-gated installs from Prompt Security
The ClawSec security skill suite repository from Prompt Security is an AGPL-3.0 collection for OpenClaw, NanoClaw, Hermes and Picoclaw. Its OpenClaw entry point, clawsec-suite 0.1.16, installs with npx skills add prompt-security/clawsec --skill clawsec-suite -a openclaw --global -y. Enabling its advisory hook is a separate step, and the setup script prints a preflight before changing OpenClaw configuration.
The suite matches a signed advisory feed against installed skills, recommends removing malicious ones only with explicit approval, and makes risky installs stop for a second confirmation. Signatures are required by default, using a detached Ed25519 signature and a pinned key. The ClawSec signed advisory feed held 835 advisories when we fetched it, 32 critical and 320 high, last updated October 13. Listed side effects: a hook under ~/.openclaw/hooks, an optional unattended openclaw cron job, and network fetches unless you pin local paths.
ClawSec gets the order right: verify the feed signature first, then gate installs and removals behind a person. Read the hook preflight before enabling it and skip the cron job until you know what it touches. Note that the suite's own last release was in July, and a signed feed is only as current as its last update.
Upgrading OpenClaw outside openclaw update?
SEN-X helps teams wire Doctor into package and image pipelines, set safe defaults for Skill Workshop learning, review conversations before public sharing, and size small hosts for 2026.10.5.
Contact SEN-X →