safe-kill

Safe process termination for AI agents with ancestry-based access control

Supported Platforms

English | 日本語

---

`safe-kill` is a CLI tool that prevents AI agents from accidentally killing system processes or unrelated applications. It enforces **ancestry-based access control** — only processes that are descendants of the agent's session can be terminated.

It takes the place of `kill` / `pkill` / `killall` for agents such as Claude Code. Every target is checked against the session's process tree and a denylist of system processes, and its identity is checked again right before the signal is sent.

Features

• Ancestry Verification: Only kill processes spawned by your session
• Suicide Prevention: Cannot kill self or parent processes; the current parent PID is re-resolved from the OS immediately before signaling, failing closed even if the process was re-parented between the policy decision and the kill
• PID 1 Protection: PID 1 (init/launchd, or a custom container entrypoint) is never killable — even via allowlist match or --port bypass — so containerized age