The Agent Skills Directory

[COMMAND_EXECUTION]: The skill orchestrates calls to various external AI provider CLIs (such as claude, codex, grok, and opencode) using timeout and env. It performs strict validation on arguments like model names and effort levels using the is_safe_identifier function, which enforces an alphanumeric allowlist to prevent command injection.
[INDIRECT_PROMPT_INJECTION]: As the skill's primary purpose is to relay data from external AI providers, it is exposed to indirect prompt injection. It mitigates this risk by wrapping all provider output in a clearly labeled envelope (# External Provider Output) and explicitly instructing the agent to evaluate the content as data, not instructions. Additionally, scripts/sanitize.sh strips potentially malicious zero-width Unicode characters and terminal escape sequences from the provider responses.
[CREDENTIALS_UNSAFE]: The skill includes proactive security checks to prevent the accidental exposure of secrets. scripts/_common.sh contains a preflight_check that scans prompt files for API keys and tokens, and a _cleanup function that uses regex to redact authentication tokens (e.g., sk-, ghp_, xoxb-) from log files before they are finalized.

spine-envoy