agent-setup

Universal setup for Claude Code and OpenAI Codex CLI — security hooks, notifications, status line, and skills.

Features

• Security Guard — denies destructive and credential-leaking commands (rm -rf ~, git push --force, private key reads, archiving ~/.ssh) with a reason instead of a prompt, so unattended runs don't stall; also catches a pkill -f that would kill the agent's own shell, and systemctl stop|restart of units listed in GUARD_PROTECTED_UNITS
• TTS Notifications — cached voice alerts when Claude finishes or needs input
• Desktop Notifications — native OS notifications (Linux, macOS, Windows)
• Status Line — project, branch, model, effort, rate limits with time to reset, context tokens, session cost
• Skills — commit («сделай коммит и пуш» chains into a PR), push-and-pr, /publish, /repo-context; for research use Claude Code's built-in /deep-research
• Cross-platform — Linux, macOS, Windows

---

Setup

• Claude Code setup
• Codex CLI setup

---

Recommendations

Tip

Disable desktop notifications — set DESKTOP_NOTIFICATIONS=0 in ~/.claude/settings.json (env section); for Codex, export it in the shell that starts Codex (hooks get Codex's own environment, not [shell_environment_policy]). Audio TTS keeps working.

Tip

Remote monitoring and control — use herdr with herdrgram to bridge your terminal multiplexer to Telegram, so you can watch and steer your AI coding agents from your phone while keeping full terminal access.

License

MIT