self-improving-agent

SUSPICIOUS: the core behavior—curating memory into durable project rules—is coherent, but the install footprint is mixed. The OpenClaw/ClawHub path is plausibly same-ecosystem, while the alternative Codex install path relies on a personal-repo script and is not verified against official OpenAI docs. The skill also has transitive trust risk because it can create/install further skills and persists command-error output into memory, which could capture sensitive data. No confirmed malicious exfiltration or hidden behavior is present.