The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes user-provided natural language files (such as CLAUDE.md) by interpolating their content into a prompt for LLM-based compression. This creates an attack surface where instructions embedded within those files could attempt to influence the agent's behavior.
Ingestion points: The read_source function in scripts/compress.py reads file content into memory.
Boundary markers: The build_compress_prompt function uses a TEXT: header as a delimiter for the user content.
Capability inventory: The skill possesses file read/write capabilities and the ability to execute the claude CLI via subprocess.run (found in scripts/compress.py).
Sanitization: The skill implements a mask_code_blocks function to protect technical content from modification and an is_sensitive_path check to prevent processing files that likely contain credentials or secrets (e.g., .env, .ssh folder, or files containing 'secret' in the name).
[COMMAND_EXECUTION]: The script scripts/compress.py uses subprocess.run to call the claude CLI as a fallback for compression when no ANTHROPIC_API_KEY is provided.
Evidence: The call in scripts/compress.py uses shutil.which to locate the binary and passes fixed arguments (--print, --strict-mcp-config), which significantly mitigates common command injection vectors. The content is passed via stdin (input=prompt), preventing shell meta-character exploitation in arguments.

caveman-compress