Custodian Skills in OpenClaw

Custodian skills are operational runbooks that ship with OpenClaw and are versioned per release. In the package, they reside under custodian-skills/, loading at the bundled-skill precedence tier, though only for the agent selected by agents.defaults.systemAgent.agentId.

If that setting is missing, OpenClaw defaults to a retained legacy owner, the single configured agent, or legacy main when no explicit agent list exists. When multiple agents are set up but neither a system agent nor a retained legacy owner is present, the library goes to no one. For any other agent, Custodian skills stay out of discovery, snapshots, slash-command catalogs, sandbox sync, and the skills prompt shown to the model.

Standard skill controls remain in force. A single Custodian skill can be turned off with skills.entries.<name>.enabled: false, and agent skill allowlists can further restrict the final set. Refer to Skills config.

Workflow contract

Every bundled Custodian skill follows the same five sections, in this exact sequence:

Gather pulls redacted current config and checks live state.
Mutate relies on validated non-interactive writes, either openclaw config set / openclaw config patch from a trusted shell or the in-session Custodian tool actions when policy permits, never a direct file edit.
Repair invokes openclaw doctor and keeps diagnosis separate from any sanctioned repair.
Prove runs one live end-to-end outcome.
Report logs what changed, what was seen, and what is still pending.

All five-section playbooks keep secrets out of prompts, logs, and files. Credentials are held in SecretRefs or credential stores. A workflow never claims success without its Prove result; when live proof is unavailable, it states the exact blocker.

First wave

Skill	Outcome
`configure-channel`	Set up and deliver a confirmed test message via a channel family like Discord, Slack, Telegram, or WhatsApp.
`add-model-provider`	Set up API-key or subscription/OAuth provider access and execute one live Gateway inference.
`diagnose-gateway`	Carry out read-only triage on Gateway, config, SecretRef, channel-auth, log, and port.
`cloud-image-bake`	Build a Cloud Worker image, verify it with a timed dispatch, and cleanly retire the superseded snapshot.

Roadmap catalog

The catalog below documents tiers planned for later. These names are roadmap items, not bundled skills or assurances of current behavior.

Tier 2: common operations

configure-search: set up and live-verify a search provider.
create-agent: spin up an agent, confirm its workspace, and verify one turn.
manage-plugin: install, configure, verify, or remove an approved plugin.
rotate-credential: rotate one supported credential through its owning store and verify the consumer.
upgrade-openclaw: stage an upgrade, run health checks, and confirm rollback readiness.

Tier 3: advanced operations

fleet-rollout: deploy one verified config or release across managed Gateways.
incident-response: gather redacted evidence, contain an incident, and confirm recovery.
migrate-gateway: relocate a Gateway while honoring explicit state and identity contracts.
release-validation: run release-track package, install, and live behavior proof.
restore-backup: restore into an isolated target, validate state, and cut over deliberately.

Add an operator skill

Place local additions in the configured Custodian agent's workspace, not in the release-owned package directory:

<custodian-workspace>/skills/<skill-name>/SKILL.md

Workspace skills already outrank the bundled tier and are scoped to that agent's workspace. Stick to the same Gather → Mutate → Repair → Prove → Report contract, keep the description brief, and start a new session after modifying the skill. See Creating skills for the complete format.