The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied natural language files and interpolates their content into a prompt for the Claude LLM. This creates a surface for indirect prompt injection where malicious instructions inside a project file could attempt to control the compression process or the agent's behavior.
Ingestion points: The target file content is read and processed in scripts/compress.py.
Boundary markers: The prompt uses a TEXT: header to delimit instructions from user data.
Capability inventory: The skill has file-write capabilities (os.replace) and can execute the claude CLI via subprocess.
Sanitization: The skill masks code blocks before sending text to the LLM to prevent modification of code, providing partial mitigation for injection vectors within code sections.
[COMMAND_EXECUTION]: The skill executes the claude CLI tool using the subprocess module to perform compression when the Anthropic SDK is not available or configured.
Evidence: Found in scripts/compress.py, where subprocess.run calls the claude binary. The execution uses a fixed list of arguments and passes user data via standard input rather than shell arguments.
[DATA_EXFILTRATION]: The skill sends the content of the target file to Anthropic's API for processing.
Evidence: The skill uses the Anthropic Python SDK or the claude CLI to transmit data.
Mitigation: The skill implements a robust sensitive path detection mechanism in is_sensitive_path to prevent the transmission of credentials, keys, and other sensitive system files by checking filenames, path components, and common secret tokens.

caveman-compress