Self-learning and approval settings

Self-learning

After substantial work, a detached background review can turn corrections and successful procedures into reusable Workshop skills; see Self-learning. Set skills.workshop.autonomous.mode to propose to create pending proposals, or to auto to maintain complete skills with normal agent tools. The Control UI Workshop tab shows whether self-learning is on; use the config setting to choose all three modes.

Scan past sessions

The Control UI can review older work without enabling autonomous self-learning. Open Plugins → Workshop and select Learn from past conversations. A normal session opens with the mining instructions, using the selected agent's configured model, permitted tools, existing skills, and accessible conversation history.

The agent decides what to read and whether the evidence warrants a skill change. It follows the current Workshop mode: auto permits direct improvements and propose leaves suggestions for approval. No separate scanner selects transcripts or limits the run to a fixed number of conversations or proposals.

Watch, steer, or stop the work in chat. The session keeps its instructions and outcome instead of separate coverage counters. A manual request does not change the self-learning setting. Normal session capacity, permissions, provider pricing, and data-handling terms apply.

In propose and auto modes, OpenClaw can review one finished substantial turn after the agent system becomes idle. It records the finished turn's boundary and reads that turn's model context asynchronously with the same provider and model. Review transcript and session metadata stay detached from foreground work. In propose mode, only skill_workshop executes and the reviewer can stage one pending mutation. In auto mode, ordinary file tools can inspect, edit, and verify several connected files in the Workshop directory. The review inherits source permissions and shell approvals. Its process tool cannot control foreground jobs; the Workshop file root is not a shell sandbox. A failed review is recorded after one attempt; completed direct edits remain.

See Self-learning for enablement, eligibility, privacy and cost details, the proposal threshold, and troubleshooting.

Approval and autonomy

{  skills: {    workshop: {      autonomous: {        mode: "auto",      },      approvalPolicy: "auto",      maxPending: 50,      maxSkillBytes: 40000,    },  },}

The selected model reviews retained evidence before deciding whether a durable procedure needs an update. Foreground work does not wait for that review. It starts only when the foreground runtime reports its resolved model and actual skill_workshop availability; restrictive or unknown tool policy fails closed.

In auto mode, the reviewer uses the same direct-maintenance guidance as weekly review. File tools stay rooted at Workshop; shell commands retain the source session's execution policy. Source deletion, replacement, or permission changes invalidate retained review authority. Direct maintenance does not run a post-turn proposal scanner or create rollback snapshots. Use backups for unwanted edits.

In propose mode, the reviewer can read or prepare an exact span before staging one pending mutation. Existing-skill proposals retain read receipts, content-hash binding, size validation, and normal apply-time scanning and rollback metadata. Immediate foreground repair also retains the normal proposal apply path in auto mode; it is separate from direct background maintenance.

See Self-learning for the complete autonomous review behavior and safety model.

Proposal descriptions are always capped at 160 bytes, independent of maxSkillBytes.