shipyard
Pass
Audited by Gen Agent Trust Hub on Nov 19, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill includes several Python utility scripts (
cleanup_merged.py,model_settings.py,shipyard.py) that execute system commands such asgit,gh(GitHub CLI), andcodex. A detailed review of these calls shows they are well-sanitized; for example, repository names are validated against strict regular expressions, and PR numbers are cast to integers. The skill specifically implements a 'Safe host bridge' that passes user arguments via structured JSON files to avoid shell interpolation vulnerabilities. - [DYNAMIC_EXECUTION]: The Python scripts use
importlibandrunpyto load internal modules likestage_registry.pyandmodel_policy.py. This is a legitimate use of dynamic loading to share configuration and logic between the skill's own components and does not involve the execution of untrusted or remote code. - [FILE_SYSTEM_OPERATIONS]: The
install_clients.pyscript installs configuration templates for Claude and Codex into the user's home directory. The script implements defensive programming practices to prevent directory traversal and symlink-following attacks, including the use ofos.O_NOFOLLOWand explicit validation of parent directories to ensure operations remain within the intended configuration scope. - [DATA_EXPOSURE_EXFILTRATION]: While the skill interacts with the GitHub API via the
ghtool, these operations are limited to fetching PR metadata for cleanup purposes. No sensitive data or credentials are exfiltrated, and the script explicitly verifies that remote URLs belong togithub.combefore proceeding. - [INDIRECT_PROMPT_INJECTION]: The skill instructions in
SKILL.mdproactively warn against the risk of interpolating raw arguments into shell code or dynamic context expressions. It provides clear guidance for host environments on how to handle user-originated text safely using its bundled helper scripts.
Audit Metadata