The Agent Skills Directory

[DATA_EXFILTRATION]: The skill transmits potentially sensitive context, including terminal commands, directory paths, and user task descriptions, to an external service (OpenRouter or a Jev-specific host) to perform security analysis.
Evidence: The procedure in SKILL.md and the inputs defined in references/preset.md (command, cwd, stated_task) specify data sent off-machine.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data in the form of shell commands and task descriptions which are then interpreted by an external LLM-based tool, creating a surface where malicious instructions could be embedded in the data to be screened.
Ingestion point: The command and stated_task variables provided by the user, as described in SKILL.md and references/full-contract.md.
Capability inventory: The skill calls the omh_jev_ask tool to process this input using the action_check/v1 preset.
Boundary markers: The instructions require the agent to notify the user precisely what data is leaving the machine before the request is made.
Sanitization: Instructions in references/full-contract.md mandate the removal of secret-like content if the tool flags it, ensuring credentials are not inadvertently sent to the external service.

omh-jev-action-check