The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from requirement files and code history to generate technical documentation. While this creates a surface for indirect instructions, the risk is mitigated by the implementation of defensive programming practices:
Sanitization: The skill explicitly requires slug validation using a strict regular expression (/^[a-z0-9][a-z0-9._-]*$/i) for any inputs derived from the environment (branch names, diff paths) before using them in filesystem operations, preventing path traversal attacks.
Least Privilege: The skill configuration limits tool access to a scoped Bash(git:*) permission and excludes permissions like Bash(node:*) to ensure it cannot execute complex build scripts or arbitrary code during its resolution process.
Human-in-the-loop: The discovery logic for technical specifications includes an explicit "Need Human" exit for ambiguous matches, preventing the agent from making autonomous decisions based on conflicting or potentially poisoned data names.

tech-spec