Flow

This package mostly does what it says (parse requests, find skills, scan code, and compose new skills), but take these precautions before installing/running: - Review requirements.txt and install dependencies in an isolated environment (virtualenv/container). The README and SKILL.md disagree about dependencies. - Run Flow in a sandbox or non-sensitive workspace first: by default it will read from ./skills, write composed flows to ./flows, and update ./skill_registry.json. If you have an important registry file, back it up or set registry_path in a config to a different location. - The composer can write/generated Python code and can accept templates — do not point it at untrusted templates or inputs that might cause it to write and later execute malicious code. - The code bundle shows signs of truncation/bugs (incomplete variable names / truncated functions). Expect runtime exceptions; review the full source in your environment before trusting automated composition. - Increase the scanner strictness (set security_level to 'strict') if you plan to auto-compose flows from third-party skills; inspect any skills that the scanner marks as HIGH/CRITICAL before composing. - Because the package author and homepage are unknown, prefer to run it locally in an isolated environment and perform a manual code review on any composed flows before executing them. What would change this assessment: receiving the complete, non-truncated source files (showing fixes), a clear statement of which dependencies are optional, and assurance that composed flows do not automatically execute third-party code (only write code for manual review) would raise confidence and could move the verdict toward benign.