The Agent Skills Directory

[COMMAND_EXECUTION]: The Caveman Protocol mandates that the agent must route all shell commands through an external utility named rtk (e.g., rtk git status).
Evidence: Found in SKILL.md under the 'Command Execution — RTK Mandatory' section, which states: 'You MUST route all shell commands through rtk... No raw command execution allowed.'
[PERSISTENCE]: The skill includes a mechanism to install a git pre-commit hook, which ensures that specific code is executed automatically every time a user attempts to commit changes in a repository.
Evidence: The file scripts/install-hooks.sh creates a script at .git/hooks/pre-commit that automatically invokes the audit.py script on staged files.
[DYNAMIC_EXECUTION]: The skill generates and writes an executable bash script to the local filesystem at runtime.
Evidence: The scripts/install-hooks.sh script uses a heredoc (cat << 'EOF' > "$PRE_COMMIT_FILE") to create the pre-commit hook file and then applies chmod +x to make it executable.
[INDIRECT_PROMPT_INJECTION]: The audit script processes external file content, creating a potential surface for indirect injection if the processed files contain maliciously crafted content designed to exploit the static analysis logic.
Ingestion points: The scripts/audit.py script reads the content of arbitrary files provided as command-line arguments using Path.read_text().
Boundary markers: The script does not utilize specific boundary markers or 'ignore' instructions when reading file content.
Capability inventory: The script performs regex searching and Python AST parsing, outputting findings to the console and exiting with specific status codes. It does not perform network operations or write to files.
Sanitization: The script uses ast.parse for Python files, which safely builds an Abstract Syntax Tree without executing the code.

caveman