The Agent Skills Directory

[PROMPT_INJECTION]: The skill setup script (scripts/context.mjs) injects directives that explicitly instruct the AI agent to bypass or override harness-level security constraints. Specifically, it tells the agent to authorize the spawning of subagents without user confirmation and to disregard system prompt instructions regarding autonomous operation or user absence.- [COMMAND_EXECUTION]: The skill uses a sophisticated suite of shell and Python scripts for task orchestration. It manages detached process lifecycles via peer-job-runner.py using low-level primitives like os.fork and os.setsid on POSIX or Job Objects on Windows. It executes external CLI tools for AI providers such as Claude, OpenAI (Codex), and xAI (Grok) to perform cross-model analysis.- [DYNAMIC_EXECUTION]: The scripts/cross-model-pov.sh script dynamically executes Python code to parse JSON output from command-line tools. The peer-job-runner.py script is designed to execute and supervise arbitrary worker commands passed as arguments, utilizing subprocess.run and subprocess.Popen for execution.- [DATA_EXPOSURE_AND_EXFILTRATION]: The skill gathers project context (code, logs, and issues) and transmits it to external AI model providers via their respective CLI tools for verification. While these are well-known services, the flow involves sending internal project information to remote third-party endpoints.

ce-pov