The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates multiple executable scripts at runtime and immediately executes them. Specifically, it writes a Node.js spawn script to /tmp/vibehq-loop-spawn.js and a polling script to /tmp/vibehq-loop-poll.js, both of which are executed using the node command. On Windows, it additionally generates and launches .cmd batch files for each agent.
[PRIVILEGE_ESCALATION]: The instructions command the agent to use the --skip-permissions flag when spawning child agents. This flag is explicitly intended to bypass human-in-the-loop approval and safety filters for those agents, granting them full autonomy without user oversight.
[COMMAND_EXECUTION]: The skill performs extensive process management and system interaction, including terminating processes via wmic on Windows and pkill or tmux kill-session on Unix-like systems. It also uses wt (Windows Terminal) and tmux to manage agent environments.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant vulnerability surface by ingesting untrusted data from external sources.
Ingestion points: The skill reads raw agent interaction logs (.jsonl files) from ~/.claude/projects/ and task state from hub-state.json to perform analysis and optimization.
Boundary markers: There are no instructions for using delimiters or boundary markers when reading these logs.
Capability inventory: The skill possesses capabilities for dynamic script generation, file system writes, and shell command execution.
Sanitization: There is no mention of sanitizing or validating the content of the agent logs before processing them for "optimization" tasks.

benchmark-loop