The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides the agent with the ability to terminate any system process using the terminate_process.py script, which utilizes os.kill() with SIGTERM and SIGKILL. While this is the intended function, the instructions in SKILL.md encourage the agent to "bonk known bad creature automatic," potentially leading to the termination of critical development servers (e.g., Next.js, Webpack, Vite) without immediate user confirmation, resulting in data loss.
[COMMAND_EXECUTION]: The AUTO_KILL_PATTERNS in scripts/hunt_processes.py include a regex for claude. If the agent's own process or related sidecar processes contain this string, the skill instructions may lead the agent to terminate its own execution environment.
[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it ingests untrusted data from the system's process list.
Ingestion points: scripts/hunt_processes.py reads the output of ps -eo pid,pcpu,rss,comm,args, which contains arbitrary strings from command-line arguments of all running processes.
Boundary markers: None. The raw process names and commands are processed and eventually displayed in the "Victory Report" and "Impact Report."
Capability inventory: The agent can terminate processes (scripts/terminate_process.py) and write files to the temporary directory (scripts/measure_power.py writes to /tmp/caveman_juice_memory.json).
Sanitization: The skill performs no sanitization or escaping of the process names/commands before including them in reports that the agent is instructed to read and show to the user.

process-hunter