Command Safety Analysis
Domain: Security
Motivation
When approvals.mode is smart, an auxiliary LLM analyzes each command for risk level before deciding to auto-approve or prompt.
Concrete Example
Command: ls -la. Analysis: low risk (read-only). Auto-approved. Command: rm -rf /. Analysis: critical risk. Prompt user. The auxiliary LLM acts as a safety inspector.
Analogy
A food safety inspector - they check each dish before it leaves the kitchen. Safe dishes go out; suspicious ones are held back. Command safety analysis inspects each command.