Finding / openclaw

2026-10-24-openclaw-workshop-replaces-proposal-review-with-direct-edits-and-bounded-undo-history

Workshop replaces proposal review with direct edits and bounded Undo history. Workshop now maintains learned procedures through skill_workshop, applies changes immediately, and saves previous versions. The code stages a version before mutation; restoration also snapshots the current version and checks live authority before replacement. History retains ten versions per skill and the change feed retains 500 entries. auto remains the default; off disables background reviews and cleanup while manual learning remains available. Migration maps legacy propose to off, removes approvalPolicy and maxPending, and deliberately does not turn legacy auto plus pending into off. Background review uses the foreground model route with a restricted tool set; it is an additional model run. Learned Workshop skills bypass the ordinary per-agent skills allowlist. Unused skills can be archived after 30 days, with runtime-specific attribution limits documented.

Channel: tagged-release. Half: both. Date: 2026-10-18.

Operator consequence: Replace the profile’s old propose/approvalPolicy: pending recipe. Use skills.workshop.autonomous.mode: "off" when changes must be initiated manually. If keeping auto, review the change feed and exercise Undo, while retaining backups beyond the bounded history. The proposal queue disappears, but judgment moves to evaluating reusable procedures and reversing bad learning. Reversibility is a real capability addition; it is not evidence that every learned edit improves future work.

Receipt

Finding metadata

Run: 2026-10-24-weekly-digest-2026-10-13_2026-10-24-frontier-v0

Finding ID: 2026-10-24-openclaw-workshop-replaces-proposal-review-with-direct-edits-and-bounded-undo-history

Profile citations

Source links

Primary links, including exact changelog lines when available.