The Agent Skills Directory

COMMAND_EXECUTION (SAFE): The hooks execute various development tools (e.g., black, pytest, git) via shell commands. All environment-provided variables like $CLAUDE_TOOL_FILE_PATH are properly quoted to prevent command injection vulnerabilities.
DATA_EXFILTRATION (LOW): The prompt-preprocessor hook logs user prompts to a local file at .claude/prompt_log.txt. While this is a form of local data exposure, it is a documented feature for session tracking and user analysis.
PROMPT_INJECTION (LOW): The prompt-preprocessor hook includes functionality to inject project-specific context from .claude/context.txt into the agent's input stream. This is an intended feature and does not attempt to bypass agent safety filters.
DEFENSIVE_CONTROLS (SAFE): The skill includes several security-focused hooks. The security-scan hook detects hardcoded secrets (API keys, tokens) before git pushes, and the pre-tool-validation hook attempts to protect sensitive configuration files from being modified by the agent. Note: The pre-tool-validation hook is incorrectly configured for PostToolUse in its JSON file, which may prevent it from blocking writes before they occur.

hook-factory