The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses Bash, Task, and Agent tools to coordinate multi-agent workflows. Execution is managed via clear phase gates and deterministic Python utility scripts (e.g., build_dag.py, assemble-plan.py) that handle mechanical tasks like dependency mapping and JSON assembly without using unsafe eval() patterns.
[DATA_EXPOSURE]: The framework includes scripts for registering new tools in user configuration files (~/.claude.json). These operations are atomic, well-documented, and follow standard tool-registration patterns. Explicit safety checks are implemented to prevent the accidental inclusion of sensitive files like .env or credentials in commits.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external repository data and user instructions. It mitigates injection risks through several layers: 'Quarantine' agents for untrusted content, mandatory schema validation of tool outputs, and specialized review agents (including contrarian and skeptical perspectives) that audit all proposed changes.
[EXTERNAL_DOWNLOADS]: The skill interacts with well-known services and registries (GitHub API, NPM, PyPI) to fetch metadata and documentation. It specifically mandates API-based sampling for GitHub interactions to avoid the risks associated with cloning entire remote repositories.
[SAFE]: Security is integrated into the core methodology, with a dedicated Security Auditor agent defined in the toolkit's own improvement roster to scan for injection vectors and permission issues. Handoff mechanisms include validation logic that rejects path traversal, shell metacharacters, and symlink escapes.

workflow