The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs potentially destructive file system operations, including find -delete and rm -rf, to remove identified junk files and empty directories.
[COMMAND_EXECUTION]: The workflow involves executing external binaries and project-specific scripts such as npm test, pytest, and npm run build. This presents a risk of Indirect Remote Code Execution if the repository being cleaned contains malicious configurations in package.json or test files.
[DATA_EXFILTRATION]: The included Python script scripts/analyze_codebase.py reads the contents of all source files to build a dependency graph, providing the agent with access to the project's source code, which may contain sensitive information.
[PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection because it processes untrusted data from the codebase to drive cleanup decisions.
Ingestion points: All source code files within the project directory are read and analyzed via the scripts/analyze_codebase.py script and shell commands like grep and find.
Boundary markers: The skill lacks explicit boundary markers or instructions for the agent to ignore malicious patterns embedded within the analyzed code comments or file metadata.
Capability inventory: The agent has the ability to delete files, modify the git history, and execute arbitrary shell commands defined in the project's build system (e.g., package scripts).
Sanitization: No sanitization or validation of the analyzed file content is performed to prevent instructions embedded in the code from influencing the agent's behavior.

eng-codebase-cleanup