The Agent Skills Directory

[COMMAND_EXECUTION]: The skill analyzes local repository state by executing system commands such as Git, ast-grep, and various language-specific build tools (e.g., Cargo, Go). These operations are performed safely using a centralized subprocess management script that disables shell interpretation, sets a strict timeout (typically 30 seconds), and limits output to 4MB to prevent resource exhaustion.
[INDIRECT_PROMPT_INJECTION]: As an audit utility, the skill ingests content from external repository files. It implements defensive boundary checks, such as a fail-closed scanning limit of 2,000,000 characters and a 100,000 character per-line limit for individual authored source files, to ensure the agent does not process pathologically large or malicious content during the audit phase.
[DYNAMIC_EXECUTION]: Internal regression tests use Python's importlib.import_module to load local skill components. This is restricted to the skill's own package hierarchy and does not process any external or user-provided input, posing no threat to the execution environment.
[SAFE]: A thorough review of all skill instructions, reference materials, and scripts confirmed the absence of any malicious behavior. The skill follows best practices for tool development, including clear provenance rules for generated artifacts and the use of structured JSON for tool interop, with no evidence of credential harvesting or obfuscated payloads.

architecture-boundaries