Capstone: run, inspect, stop, and remove the first safe agent

Verifiedon 2026-07-302026.7.1-2

Capstone: First Safe Agent

Run exactly one controlled task from the prior lesson. The agent reads only the test input, produces a proposal, and writes no external side effect. The trusted operator reviews the proposal and records either approved for a future, separate action or not approved. For this capstone, do not perform the future action.

Evidence checklist

Your redacted evidence record must contain:

• platform and OpenClaw version;
• provider live-check result without a credential;
• CLI/Gateway verification results;
• task contract and controlled-input label;
• proposal outcome and human approval decision;
• a statement that no real channel or external side effect was connected;
• credential rotation/revocation result; and
• gateway stop plus workspace/install teardown result.

Secret rotation drill

Assume the test credential was copied into an unsafe note. Revoke it with the provider, replace it if a repeat test is necessary, and verify that the old credential cannot be used. The OpenClaw secrets guidance is clear that plaintext residue and backups remain secrets until removed; a new key does not erase the old one.

Teardown

Use the primary uninstall documentation for your exact install route. Preview the built-in removal scope first with openclaw uninstall --dry-run --all; inspect the listed service, state, workspace, and app targets before approving any removal. On a VPS, retain only the redacted operational evidence and any backup required by your policy; remove test state/workspace only after confirming it is not shared. On a local host, stop the gateway first, then run the approved uninstall scope. Do not improvise deletion commands from memory.

Rubric

A pass means the exercise is reversible and inspectable—not that the agent is safe for production.

Check your understanding