OpenClaw 2026.10.5 Beta Makes Conversations Restart-Safe, Adds an X Channel, and Brings ClawHub Into the Gateway
A one-day-old prerelease treats Gateway restarts and interrupted updates as routine events, adds a public X channel with hard spend limits, and fixes the plugin compatibility trap reported last week.
2026.10.5-beta.1 Makes Interruption a Normal Event
The OpenClaw 2026.10.5-beta.1 prerelease notes landed on GitHub on October 11 at 6:08 AM Pacific, one day after 2026.10.1 became the stable release. It is a pre-release, not a stable channel build, and its audited contribution record covers 1,971 pull requests. The headline theme is continuity: interrupted turns, first messages sent while a worktree is still being prepared, and deferred tasks are meant to resume after a Gateway restart, while user input and committed session data are meant to survive restarts, concurrent writes, and failed publications.
The update path receives the same treatment. According to the notes, updates should now finish across devices and filesystems, through Windows elevation, and after interrupted activations, recovering from stale leases or changed evidence instead of stranding the install. Specific fixes include completing package swaps across devices instead of rolling back on cross-device errors, falling back to byte copies when file cloning is denied, and refusing non-elevated Windows updates immediately with an elevation hint rather than failing midway. Doctor also stops reporting skipped maintenance as repaired, a small change that makes its health output more honest.
Restart safety is the feature that determines whether an agent platform can be trusted with long-running work. A Gateway that loses a half-finished turn during an update forces operators to schedule around the software. Test this beta by deliberately restarting mid-turn in a staging install and confirming that the conversation, its queued input, and any subagent results come back intact.
A Steadier Gateway Under Load
Performance work in this beta is extensive and mostly invisible. Session, transcript, history, and metadata reads move onto worker threads with far less SQLite work per turn, and startup maintenance now runs after the Gateway reports ready so busy installations stay responsive. The notes describe reduced SQLite statements per chat turn, batched snapshot and metadata queries, fixes for writer-reader admission deadlocks, and cron ticks and session scans moved off the main thread. For operators running many agents on one host, these changes target the stalls that make a Gateway feel unreliable even when nothing has technically failed.
Turn reliability gets parallel attention. Turns are designed to recover from stalled model streams, reused tool-call identifiers, and expired ChatGPT websocket sessions, and incomplete turns now end with a visible notice instead of disappearing. Subagent results should keep returning through follow-ups, restarts, and expired delivery windows. A related fix stops transcript-only channel replies from being reported as sent, which matters for anyone auditing whether a message actually left the system.
The X Channel Arrives With a Budget Built In
The most visible new capability is an X mentions channel plugin. The official X channel documentation explains that mentions of a bot account become agent conversations and the answer is posted as a public reply. Only allowlisted authors can trigger a reply by default, and unknown authors are silently dropped unless guest mode is enabled. Direct messages, original posts, likes, follows, and media uploads are not supported. Authorization is based on the numeric X author ID, never on a handle, display name, post text, or model output.
Cost control is part of the design rather than an afterthought. The plugin defaults to estimated limits of $100 per UTC day and $1,000 per billing cycle for each account, covering X API calls only; model tokens are billed separately. Setting a limit to zero blocks paid requests, and there is no unlimited setting. The plugin reserves the worst-case cost before each paid request and settles afterward, and when a budget is exhausted it pauses ingress without advancing its cursor so missed mentions can be recovered after reset. The documentation also warns that OAuth refresh tokens are stored in plugin state without a promise of encryption at rest, so the Gateway's state directory and backups should be protected as credentials.
A public social channel is the riskiest surface an agent can have: anyone can address it, every reply is permanent, and every API call costs money. The X plugin's defaults reflect that, with allowlist-only admission, numeric-ID authorization, and hard spend ceilings. Keep those defaults, start with a very small daily limit, and review a week of replies before widening who can trigger the bot.
ClawHub Moves Into the Gateway, With Version-Aware Plugin Picks
The beta lets operators browse and search ClawHub plugins and skills directly from the Gateway, and plugins now pick the newest official release that the running OpenClaw supports. That second change is the fix for the compatibility failure covered in yesterday's briefing, where a stable 2026.9.7 Gateway selected a Perplexity plugin requiring a newer plugin API. The merged pull request for host-version-aware plugin selection explains that managed stable installs of official ClawHub plugins first try the build matching the running OpenClaw version, then warn and fall back to the original catalog candidate if that build is unpublished.
The pull request is also a useful window into how OpenClaw reviews compatibility-sensitive changes. An automated review of an earlier revision flagged two regressions, one blocking a declared npm fallback and another breaking drift diagnostics in Doctor, before the change was revised and merged on October 8. The final description says default update and drift detection keep their registry-latest behavior, while explicit operator versions, non-latest tags, and integrity pins keep their targets, and no configuration or schema changes are required.
Security Practice: Contain a Public Channel Before You Open It
Lock down the filesystem and queue mode before enabling X guest mode
The X documentation spells out a containment recipe for guest mode, which lets people outside the maintainer allowlist ask repository questions. Guests receive a narrow default tool set: read, ls, sessions_spawn, sessions_yield, and subagents. They cannot edit files, run commands, browse the web, use memory, or send messages. But tool names alone do not confine filesystem reads, so the plugin refuses guests unless the front-door agent meets specific conditions.
- Point the agent's
workspaceandcwdat the repository clone and settools.fs.workspaceOnlyto true. - Disable selected skills with
skills: []and turn sandbox mode off, because skill directories and sandbox mounts are explicit read exceptions. - Set the X queue mode to
followuporcollectso guest messages cannot steer or interrupt an active turn. - Keep the per-author daily guest limit low; the default is five mentions per UTC day, and zero admits no guests.
Run openclaw config validate and openclaw channels status after each change. If guest admission is blocked, channel status reports the required correction as guestModeBlockedReason while maintainer mentions continue normally.
Tool Spotlight: Skill Workshop Shows Its Work
Skill Workshop
Skill Workshop, OpenClaw's surface for skills that agents learn from experience, gets three changes in the beta notes. Each skill now appears with its changes, files, and history in one view. Learned skills are steered toward repeatable procedures rather than facts or codebase knowledge, which keeps them closer to reusable runbooks than to an unreviewed memory dump. And approval-first setups are prevented from writing skills without review, while configurations that rely on an approval policy keep automatic learning.
The combination is the right shape for teams that want agents to improve without letting them silently rewrite their own instructions. A single view of a skill's revision history makes review practical, and the approval-first fix closes a path where a learned procedure could appear without a human ever seeing it.
Community Context: Stable and Beta Are Now Visibly Different Tracks
The beta arrives while most operators are still absorbing the stable release. OpenClaw Chronicles' report on 2026.10.1 reaching npm latest notes that npm metadata changed at 22:45 UTC on October 10, with 2026.10.1 as latest, 2026.10.1-beta.2 then on the beta tag, and 2026.8.35 as extended-stable. Default install commands now resolve to 2026.10.1, so ordinary package consumers received that build, not the new beta.
That separation is worth keeping deliberate. Production installations should track stable or extended-stable and read the 2026.10.5 notes as a preview of what is coming, while a staging Gateway tries the beta's restart recovery, X channel, and in-Gateway ClawHub browsing. The scale of the beta, nearly two thousand pull requests, is a reason for patience rather than haste: broad hardening releases tend to surface their own edge cases before they settle.
Run two lanes on purpose. Keep customer-facing agents on the stable channel with a pinned version and a tested rollback, and give one staging Gateway the beta with real but low-risk workloads. Promote only after the restart, update, and plugin-selection behaviors you depend on have passed there. Channel discipline is cheaper than recovering a production fleet from a pre-release.
Need help operating OpenClaw safely?
SEN-X helps teams plan OpenClaw upgrades, channel guardrails, permission boundaries, and staged beta testing.
Contact SEN-X →