The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes internal scripts via subprocess.run using shell=False. It validates that all script paths are within the skill directory, preventing arbitrary command injection.
[DYNAMIC_EXECUTION]: The skill uses importlib.import_module to load its diagnostic contract. This loading is restricted to internal skill files and does not process external inputs.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill scans repository files for structure and hashing. It uses robust path validation to ensure it only accesses files within the target repository, preventing exposure of sensitive system data.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted Markdown data from the repository via structure_state.py. While no boundary markers are present, the skill's capabilities are limited to structural analysis and writing hardcoded guidance text to AGENTS.md. The absence of sanitization for ingested data is mitigated by the fact that no ingested content is used in the guidance publication process.

atlas