OpenClaw adds Undo for learned skills and simplifies public chat links

OpenClaw’s version 2026.10.5 gives users a way to undo changes to learned skills while making published conversations readable at their ordinary chat addresses. The release notes draw an important distinction: skill changes have saved versions, but public links expose future messages too, and disabling sharing cannot retrieve copies readers have saved.

Learned procedures get a version trail

Skill Workshop 2.0 brings learned procedures, their files and their change history into one library. Users can edit a skill, compare versions, undo changes, and archive or restore it. The library keeps up to ten saved versions, and new sessions pick up skill edits.

The learning guidance also distinguishes procedures from other information. Skill Workshop is directed to learn repeatable processes, while personal facts belong in memory and knowledge specific to a code repository belongs in project documentation. Saved versions show their own descriptions, and comparisons include metadata changes as well.

There is a data and billing consequence to automatic learning. OpenClaw says background learning sends conversation and tool content to the configured model provider and can incur usage charges. Choosing Off stops background learning and unused-skill cleanup, but users can still request learning explicitly with /learn.

The upgrade treats existing learning settings differently, rather than moving everyone to one default:

• Existing Auto settings and unset automatic-learning settings are preserved.
• Legacy proposal modes move to Off. Settings already changed to Off stay off until the user changes them.
• Restoring pre-upgrade proposal history requires its matching backup; pending proposals are exported inactive.

Redaction is best effort, and turning sharing off cannot recall saved copies.

OpenClaw v2026.10.5 release notes

A public conversation is not a frozen snapshot

Published conversations now open at their normal chat address. Anyone holding that link can read without signing in; signed-in readers use the same address with their existing permissions. The anonymous view is read-only, but publication covers existing and future user messages and final replies—not only the exchange visible when sharing begins.

OpenClaw tells users to review a conversation before publishing. Sharing requests also recheck current permissions and the conversation’s identity before returning a public link. For deployments behind a login proxy, administrators need to configure a public-reader route while keeping private application and API routes protected.

Delegated work stays beside the parent chat

The sidebar changes focus on supervising work already underway. Subagents and Processes panels sit beside the parent conversation, letting users inspect delegated tasks and recent command output without leaving the chat. Targeted Stop controls let them halt a task, while finished commands start collapsed and running work remains visible.

Larger agent lists appear in rows, with IDs to distinguish repeated names. Search by name or ID becomes available when more than six agents are present, and users can pin favorites from the agent menu when at least four are available. Session-list updates also reuse unchanged rows while unrelated subagents run, keeping status and permissions current without reloading the whole list.