Summary

The current evaluation suite lacks coverage for behavioral patterns related to shell command usage and safety.

Problem

There is no explicit validation for:

• Proper use of run_shell_command for shell-related tasks
• Avoidance of destructive commands like rm -rf without safeguards
• Correct tool selection when creating files (prefer write_file / create_file over shell commands)

Proposed Solution

Add a new evaluation suite (evals/shell_command_safety.eval.ts) with behavioral tests that:

• Verify run_shell_command is used for file listing and disk usage queries
• Ensure destructive commands are not executed silently
• Enforce usage of dedicated file-writing tools for file creation

Impact

• Improves safety guarantees
• Strengthens tool selection correctness
• Expands eval coverage