The Agent Skills Directory

[COMMAND_EXECUTION]: Performs discovery using shell commands such as ls, grep, and find to detect project files and key names. It also executes package manager commands to install the reeval tool and initialize its configuration.
[EXTERNAL_DOWNLOADS]: Installs the reeval package and zod library from public registries using standard tools (npm, pnpm, yarn, or bun).
[INDIRECT_PROMPT_INJECTION]: Analyzes project files and fetches external research data to generate evaluation tasks, which exposes the agent to untrusted data that could influence generated scripts.
Ingestion points: Project source code (src/), documentation (README.md), package.json, and web-fetched research content (changelogs, docs).
Boundary markers: None explicitly defined; the skill relies on instructional context to manage data ingestion.
Capability inventory: Includes shell command execution in a Docker sandbox, file system access (c.readFile, c.writeFile), and network-based research operations.
Sanitization: No specific sanitization or escaping protocols are provided for handling external project content.
[CREDENTIALS_UNSAFE]: Reads .env and .env.local files to identify required API key names. The skill includes a high-priority safety rule that strictly prohibits the agent from reading, printing, or requesting the values of these keys.

setup