The Agent Skills Directory

[DATA_EXFILTRATION]: The skill documents an integration with external services (TypeSafe and OpenRouter) via the omh_jev_ask tool. This tool is designed to send workspace 'state'—which the documentation explicitly notes can include shell commands, working directory paths, source diffs, and full file contents—to external hosts for critique or scoring. Although the skill mandates explicit user consent ('ask jev') and implements filters to refuse credential-like content, the structured capability to transmit local file data to external services is a notable exfiltration surface.
[PROMPT_INJECTION]: The skill contains instructions in SKILL.md and references/wrapper-routing.md directing the agent to skip user confirmation for certain backend CLI operations, such as omh list and omh recommend. These instructions are intended to prevent the user from being prompted for 'shell approval' during routine metadata lookups for the workflow picker. While optimized for user experience, instructions to suppress output or skip review checkpoints for tool execution are classified as concealment patterns.
[COMMAND_EXECUTION]: The skill acts as a coordinator for a large suite of command-line tools (including omh, claude, codex, ast-grep, and git). It provides detailed protocols for preparing, dispatching, and observing the results of shell commands and delegation prompts. It specifically addresses how to handle permissions and background execution for these tools, emphasizing a security posture that distinguishes between prepared plans and observed runtime evidence.

omh-routing