The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to invoke the claude CLI as a fallback authentication method. This execution is handled safely without a shell environment and with fixed arguments, preventing command injection vulnerabilities.
[DATA_EXFILTRATION]: The skill sends file contents to the Anthropic API to perform compression. To mitigate the risk of accidental exposure of secrets, it implements a robust heuristic check (is_sensitive_path) that blocks the processing of files matching sensitive patterns, such as .env, .ssh/, or files containing 'secret' or 'credential' in their names.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from user-specified files. While this creates a theoretical surface for indirect prompt injection, the risk is localized to the content of the file being compressed. The skill effectively minimizes potential data corruption or malicious command modification by performing post-compression validation, ensuring that all code blocks, headings, URLs, and file paths are preserved exactly.

caveman-compress