The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external source code files to perform structural analysis and generate new project files. This creates a potential surface for indirect prompt injection if the input files contain malicious instructions aimed at the agent.
Ingestion points: scripts/main.py reads user-provided source files from a directory using pathlib.Path.read_text.
Boundary markers: The SKILL.md instructions include a fixed response template to ensure the agent's output is structured and isolated from input data.
Capability inventory: The skill performs file writing (Path.write_text), file permission changes (os.chmod), and defines functions that use system commands (pip freeze).
Sanitization: Input code is analyzed using the ast module for static structural parsing rather than being directly executed, reducing the risk of malicious code triggering during analysis.
[COMMAND_EXECUTION]: The script scripts/main.py generates utility code that uses subprocess.run to execute pip freeze. This is used to capture the software environment for reproducibility purposes. The skill generates this code rather than executing it directly during the refactoring process.
[PRIVILEGE_ESCALATION]: The generation logic in scripts/main.py uses os.chmod to set executable permissions (mode 0o755) on the entry point script of the newly generated project. This is a functional requirement for the generated artifact and does not involve attempts to elevate system privileges or bypass security controls.

code-refactor-for-reproducibility