The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to execute npx -y jev-use install, which fetches and runs a package from the npm registry. The source repository is a community-owned project (shitianfang/jev-use) and is not affiliated with a trusted or verified organization.
[REMOTE_CODE_EXECUTION]: Running the npx installation command allows the execution of remote scripts to configure local environments and wire MCP servers. This process lacks automated integrity verification for the downloaded content.
[COMMAND_EXECUTION]: The installation command modifies the local configuration of agent harnesses (such as Claude Code and Codex) to integrate the tool server, which changes the execution environment of the host system.
[DATA_EXFILTRATION]: The jev_judge and jev_gate tools transmit the content of the state parameter to external AI judgment backends (TypeSafe, OpenRouter, or AI Gateway). Users are required to ensure that no sensitive data, credentials, or private information are included in the context strings sent to these external services.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted context (e.g., tool outputs, build results) within the state field to drive agent orchestration, creating a surface where malicious external input could manipulate the judgment outcome.
Ingestion points: The state parameter in the jev_judge and jev_gate MCP tools (SKILL.md).
Boundary markers: None identified. The instructions encourage users to include raw facts and tool output without explicit delimiters or safety prompts.
Capability inventory: The skill performs external network requests and acts as a gating mechanism for agent actions through jev_gate.
Sanitization: No input sanitization or validation is described for the content placed into the state parameter.

jev-use