The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes subprocess.run to call the claude command-line interface as a fallback for the Anthropic SDK. This execution uses a fixed list of arguments and does not invoke a shell, which prevents shell injection vulnerabilities. The purpose is to utilize the user's local authentication for compression tasks.
[INDIRECT_PROMPT_INJECTION]: The skill processes text from local files by including it in a prompt for compression. This creates a potential surface for indirect prompt injection if the source file contains instructions designed to confuse the LLM. The skill mitigates this through specific instruction boundaries and a validation step that checks for the preservation of structure and technical content.
[DATA_EXPOSURE]: The skill reads the content of files specified by the user to send them to an external LLM provider for processing. To prevent the accidental transmission of sensitive data, the implementation includes a heuristic check (is_sensitive_path) that blocks processing for files matching known secret patterns, such as private keys, environment variables, and credential stores.

caveman-compress