The Agent Skills Directory

[COMMAND_EXECUTION]: Command execution is the core purpose of this skill. It implements robust safety boundaries by defaulting to 'direct' execution mode, which bypasses shell interpolation. It requires explicit user authorization to enable 'native' shell execution, mitigating command injection risks from repository READMEs.
[DATA_EXPOSURE]: The skill implements multiple data protection layers. The orchestrator scrubs the environment of child processes by removing variables matching patterns like 'KEY', 'TOKEN', or 'SECRET'. Additionally, the agentic file tools in run_agent.py explicitly block access to .git and .env files, and the lessons_store.py script uses a regex-based blocklist to prevent secrets from being recorded in local logs.
[INDIRECT_PROMPT_INJECTION]: Because the skill processes untrusted markdown and code from external repositories, it is a surface for indirect injection. It mitigates this by using structural parsing (Python's ast module) for code analysis and maintaining strict separation between untrusted data and the agent's control logic.
[OBFUSCATION]: No obfuscation was detected. All scripts are in plain text and use standard libraries. The skill includes a 'doctor' script that verifies the integrity of bundled components using SHA-256 hashes against a manifest.
[PERSISTENCE]: The skill maintains a persistent 'lesson store' in ~/.rigorpilot/ to track failure fixes and user preferences across sessions. This is implemented with user-auditable plain-text files and automated secret filtering, making it a legitimate productivity feature rather than a malicious persistence mechanism.

ai-research-reproduction