The Agent Skills Directory

[COMMAND_EXECUTION]: The orchestration script scripts/compress.py utilizes subprocess.run() to invoke the claude command-line tool as a fallback authentication method when an API key is not explicitly provided in the environment. This execution is performed using a static argument list and does not invoke a shell, which mitigates the risk of command injection from external input.
[DATA_EXFILTRATION]: The skill's core functionality involves reading local files and transmitting their content to the Anthropic API for compression. While this is the intended behavior, the skill includes a significant security measure in scripts/compress.py via the is_sensitive_path function. This function uses a comprehensive list of regex patterns and keyword heuristics to detect and block the processing of files likely to contain secrets, such as .env files, SSH private keys, AWS credentials, and configuration files in sensitive directories like .ssh or .aws.
[PROMPT_INJECTION]: An indirect prompt injection surface exists in scripts/compress.py where the content of a user-provided file is interpolated directly into the compression prompt sent to the LLM. This could allow a maliciously crafted file to influence the LLM's behavior.
Ingestion points: External file content is read into the original_text variable in scripts/compress.py using Path.read_text().
Boundary markers: The content is presented to the LLM after a TEXT: label without the use of robust delimiters or escape sequences.
Capability inventory: The skill possesses the capability to read and overwrite files within the user's project directory.
Sanitization: The risk is mitigated by the is_sensitive_path check which prevents processing of high-risk files, and a validation step in scripts/validate.py that checks the compressed output against the original to ensure that critical structural elements like headings, code blocks, and URLs remain intact.

caveman-compress