How are you handling CLI permissions for Claude Code, Codex, or OpenCode right now?

Curious how everyone here runs terminal coding agents day to day:

1. Default manual prompts (approving every bash command one by one)

2. Native Auto Mode

3. Full --yolo / --dangerously-skip-permissions inside a container or on the host

We ran a 456-command benchmark across all three setups while building yolonot (launching tomorrow!). Happy to share what we found on false negatives and prompt fatigue if anyone is interested.