OpenClaw 2026.10.10-beta.3 Repairs Windows and Schema Upgrades as the Control UI Moves From Lit to Solid 2.0
The new beta fixes Windows startup and recovers agent stores held after schema-changing updates, while the agent-driven move of the browser dashboard from Lit to Solid 2.0 reaches the beta line. Plus: MiniMax M3.1 Flash Preview without an upgrade, what the resurfaced Canvas CVE record says, a Gateway exposure check and the audit-explainer skill.
2026.10.10-beta.3 Arrives With Windows Repairs and Schema-Upgrade Recovery
The OpenClaw 2026.10.10-beta.3 prerelease notes on GitHub were published at 12:00 AM Pacific on Wednesday, about half an hour after the npm package went out, and npm's beta tag now points at it. Stable is unchanged: latest is still 2026.10.5.
On Windows, beta.3 restores fresh state initialization and Gateway startup, hides SQLite helper windows during app updates, and lets forced Gateway reinstalls tolerate launcher timestamp settling. For upgrade recovery, app-owned plugins now converge to the upgraded core version, migration rollback backups are scoped to the tables actually changing, and pull request #176783 on recovering held system agents after schema-changing updates lets openclaw doctor --fix restore the built-in openclaw and crestodian stores once it has verified ownership, integrity and backups.
Background task completions are no longer dropped when config changes mid-turn, Discord native slash commands work again, long exec commands through the macOS app node no longer fail after 20 seconds, and two agents initializing at once can no longer claim the same agent database.
The three fixes for Monday's manual-upgrade exit-78 reports (#178289, #178292 and #178286) merged to main early Tuesday, but none of them appears in beta.3's contribution record, so hosts upgraded outside openclaw update still need the manual Doctor order. Separately, maintainer steipete closed issue #173764 on update runs hanging in the post-activation Doctor step at 11:18 AM Pacific on Tuesday. He wrote that "an update can no longer hang forever in that step" thanks to #173862 and #174060.
Beta.3 is the build to test if your fleet is Windows-heavy or has ever had a schema-changing update leave an agent store held. Everyone else should stay on 2026.10.5. The exit-78 fixes are not in this beta, so package-managed hosts still need the stop, install, doctor --fix, restart sequence in their playbooks.
The Control UI Leaves Lit for Solid 2.0 After a Two-Week Agent Loop
OpenClaw's browser Control UI has been rewritten from the Lit web-component library to Solid 2.0, mostly by coding agents. OpenClaw maintainer Peter Steinberger said on X on Tuesday that an agent loop he started two weeks ago took about six days to remove Lit, another week to reach idiomatic Solid, and roughly 600 pull requests, with him checking in every other day.
The repository backs that up. Pull request #174213 retiring the Lit migration infrastructure, merged on Friday, October 16, says the Control UI and plugin views "no longer import Lit". It notes that the Canvas renderer still pulls in seven transitive Lit packages. A public OpenClaw migration timeline shared from the team's own Control UI counts 255 pull requests that explicitly name Solid, Lit or Web Awesome, plus 352 later UI refactors listed separately. It reports that 666 manual requestUpdate calls remained when Lit was dropped and none by October 20.
The migration is still moving. Solid 2.0 itself is a release candidate, and pull request #179116 upgrading the Control UI to Solid 2 rc.16 merged at 12:46 AM Pacific on Wednesday. Checking the tagged package manifests, the Control UI in stable 2026.10.5 still depends on Lit only, while the 2026.10.10 betas carry Solid 2 rc.14.
This is a rare public record of agents carrying a framework migration end to end, with a human reviewing every other day. For operators, the practical point is simpler: the dashboard you use daily will change underneath you in 2026.10.10. Test custom plugin views and any browser automation against the beta before it goes stable.
MiniMax M3.1 Flash Preview Shows Up Without an Upgrade
The official @openclaw post announcing MiniMax M3.1 Flash support, published at 8:39 AM Pacific on Tuesday, says the model "is now available in OpenClaw" and that anyone running v2026.10.1 or newer needs no update. The OpenClaw MiniMax provider documentation lists it as MiniMax-M3.1-Flash-Preview, with text and image input and a 1,000,000-token context window.
The docs say MiniMax currently offers the preview only through its Token Plan and MiniMax Code, so availability depends on the account, and MiniMax-M3 remains the default chat model. Thinking cannot be turned off: /think low through /think max select five efforts, the default is max, and MiniMax rejects disabled thinking with HTTP 400. The bundled cost fields read zero only because MiniMax has not published per-token pricing for the preview, "not because the service is free."
Zero-cost catalog entries quietly break budget dashboards. If you trial M3.1 Flash Preview, set real prices in models.json first, and remember the default effort is max, the most expensive setting. A one-million-token window helps long repository work, but measure latency at lower efforts before routing routine chat to it.
A Month-Old Canvas CVE Resurfaces, With Its Fix Status Unclear
CVE-2026-94094 circulated on X early Wednesday with a claim of a published exploit and no patch. The record is real but a month old. The NVD entry for CVE-2026-94094, the OpenClaw Canvas Host denial of service, was published on September 20 from a VulDB-assigned record, and NVD lists its analysis status as Deferred. It scores 4.3 (medium) under CVSS 3.1, with a vector that requires low-privilege access and only affects availability. The record names 2026.9.0 through 2026.9.5 as affected and gives no fixed version.
The original Canvas Host denial-of-service report and proof of concept tested openclaw 2026.6.11. It describes GET /__openclaw__/canvas/<file> reading whole files into memory, so slow clients requesting a large file can pin big buffers in the Gateway. We found no matching OpenClaw GitHub security advisory and no maintainer statement either way, so we can't confirm or rule out the "no patch yet" claim for current releases.
Security Practice: Read Your Own Audit Before Trusting Anyone's Alert
Keep Canvas and the rest of the Gateway's HTTP surface on trusted networks, then let the built-in audit rank what to fix
The OpenClaw network exposure guide explains why a Canvas bug is really a Gateway bug: one port carries the WebSocket, the Control UI, hosted widget documents under /__openclaw__/canvas and A2UI assets. It says widget documents contain agent-authored HTML and JavaScript and should not be exposed "to untrusted networks or users." LAN, tailnet or custom binds need Gateway auth and a real firewall.
- Run
openclaw security auditfor the read-only config check, and add--deepfor best-effort live Gateway checks and plugin-owned collectors per the OpenClaw security CLI reference. - Fix in the order the OpenClaw guide to running the security audit gives: anything open with tools enabled, then public exposure such as LAN binds, Funnel or missing auth, then remote browser control, file permissions and plugins.
--fixtightens file permissions and flips open group policies to allowlists, but it does not rotate secrets, disable tools or change bind and auth choices. Those remain your job.- If a login proxy fronts the Gateway, never bypass authentication for the whole
__openclaw__namespace, as the OpenClaw Control UI URL guide warns.
A hashtag-heavy CVE post is not an incident, but it is a useful prompt. The record's CVSS vector assumes a low-privilege caller who can already reach the Canvas route. If your Gateway stays on loopback or a tailnet with auth, the realistic exposure is small. Run the audit anyway and confirm that it is.
Skill Card: audit-explainer Turns the Audit JSON Into a Plain Briefing
audit-explainer by Is My Agent Safe, a read-only reader for openclaw security audit --json
The audit-explainer skill listing on ClawHub shows version 1.0.1 under the MIT-0 license, installable with openclaw skills install @ismyagentsafe/audit-explainer, and a passing ClawHub security audit badge. It runs one command, openclaw security audit --json. It never runs --fix or --deep, write no files, make no network calls and replace any secret in the output with "[hidden]". It lists findings by severity in OpenClaw's own triage order, quoting OpenClaw's remediation text and flagging suppressed findings.
The publisher is Your Evenings Ltd. It mentions, once, an optional $29 website report and tells the agent not to open that page for the user. The audit-explainer source repository on GitHub was created on Tuesday, so it has no track record yet.
The design is right: a skill that only reads the built-in audit and admits it is not a scanner is easy to review. Read the SKILL.md before installing, since it is short. Treat the upsell as a choice rather than a step, and never paste deep audit output into any website without removing secrets first.
Planning your move to OpenClaw 2026.10.10?
SEN-X helps teams test betas on Windows and package-managed hosts, check custom plugin views against the new Solid Control UI, price new models correctly and keep the Gateway off untrusted networks.
Contact SEN-X →