Auditing VS Code 1.147's Scheduled AI Agents: Routers, Permissions, Runs
VS Code 1.147 reportedly adds scheduled AI tasks, browser control, and model routers. Taken together they describe an unattended process that holds your credentials, drives a browser, and resolves model routing as policy rather than a per-request choice. This video is the audit plan for that combination: why a report is a weak basis for a security posture, which questions stay confirmed versus unstated, and how to instrument before you enable anything. It walks through finding the shipped settings schema, the shape of a per-run record you can diff, why output should land as a patch instead of a side effect on your working tree, and the blast radius of an editor agent compared with a CI runner. Key points covered: - Confirmed facts versus inference, with every inference labelled - Trigger model, permission model, and run logging as open questions - What changes if task definitions live in workspace files, or in user settings - Indirect prompt injection as the default failure mode for browser control - Why local execution is a deployment property, not containment - The rule that a capability you cannot log should not be scheduled Read the full post: https://hackyjs.com/posts/auditing-vs-code-1147s-scheduled-ai-agents-model-routers-permissions-and-reproducible-runs The code shown is portable scaffolding, not a reverse-engineered integration. All identifiers and values are placeholders.