The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local implementer CLIs (e.g., claude, codex, opencode) to check version, authentication status, and available models.
[COMMAND_EXECUTION]: Inputs such as model identifiers and configuration dials are validated against strict regular expressions (e.g., MODEL_TOKEN.shellSafe) before being used in shell commands or written to configuration files.
[DATA_EXFILTRATION]: The skill reads directory metadata (file counts and modification times) from local session stores to determine tool usage. It explicitly avoids opening or reading the content of these files to protect user privacy.
[SAFE]: Project-level configuration files require explicit user approval. The skill generates a SHA-256 hash of the approved config and stores it within the repository's .git directory to prevent the execution of untrusted or malicious configurations from cloned repositories.
[SAFE]: File system operations are protected against directory traversal and symlink attacks. The skill ensures that project configuration files resolve within the repository's .delegate directory and refuses to write to symbolic links.

delegate-setup