OpenBot v0.0.4 fixes a policy fail-open where the guard judged browser clicks with the element half empty
CopilotKit's OpenBot gives each AI coworker its own computer and judges every action before it happens. In v0.0.4, published 22 August, a Bot cites a ref from a page snapshot and the server resolves it back to an element before the boundary decides; when the lookup failed the action proceeded with the element half of the decision blank, so a rule like 'never click anything named submit' was never shown the element and the shipped default permitted the click. The staleness check did not catch it because it compares a citation against the computer's own counter, and the failing case is the one where the computer is content and only the deployment is out of step, which is exactly what restarting a computer under a stored snapshot leaves behind. The same release lets a tenant package ship `skills.yaml` so tool-narrowing works on a fresh clone instead of requiring a hand-built skill-to-tool map per deployment.
↳ Follow the thread