The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute system commands (e.g., uname -s) to identify the host environment (macOS, Linux, or Windows) before generating scripts. Furthermore, the provided reference scripts (e.g., assets/prepare_environment_java.sh and assets/prepare_environment_python.sh) demonstrate unsafe shell patterns where positional arguments and derived variables are used without proper quoting in commands like rm -rf $JAVA_BUILD_SUBFOLDER and cp -R $1/*, which could lead to command injection if a user or orchestrator provides a maliciously crafted build folder name.
[EXTERNAL_DOWNLOADS]: The scripts produced by this skill are designed to fetch software dependencies from official package registries (such as PyPI, npm, and Maven Central) using standard build tools.
[PROMPT_INJECTION]: The skill defines a process that ingests and acts upon untrusted data from project folders, establishing a surface for indirect prompt injection.
Ingestion points: The generated scripts recursively copy project files from a source directory into a temporary working folder for build staging.
Boundary markers: There are no requirements or instructions within the skill to implement boundary markers or "ignore embedded instructions" guards when processing project files.
Capability inventory: The automation performs powerful actions such as package installation and code compilation based on the content of manifest files (e.g., requirements.txt, pom.xml, package.json).
Sanitization: The skill lacks logic to sanitize project file content or validate input parameters before they are interpolated into shell commands.

implement-prepare-environment-script