arch-skill
Pass
Audited by Gen Agent Trust Hub on Mar 31, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill is designed to execute various development workflow commands, including Git operations (commit, push, and opening PRs via
/prompts:arch-open-pr) and running automation test harnesses. - [EXTERNAL_DOWNLOADS]: Includes routing to research-focused agents (e.g.,
/prompts:arch-external-research-agent) that utilize web search capabilities to gather technical best practices and documentation. - [DATA_EXFILTRATION]: Facilitates network operations through Git integrations and PR management. These operations are restricted to the repository being worked on and are a primary function of the skill.
- [PROMPT_INJECTION]: Contains an indirect prompt injection surface as the skill ingests and processes untrusted data such as user-provided requirements and external documentation files to generate implementation plans. The skill includes a "Question Policy" to minimize unnecessary information gathering.
- [REMOTE_CODE_EXECUTION]: Implements a dynamic instruction-loading mechanism. It locates Markdown procedures in local configuration directories (such as
~/.claude/commands/prompts/and~/.codex/prompts/) and uses a "Binding Contract" to interpolate user arguments into these templates at runtime. This behavior is standard for prompt-based workflow suites.
Audit Metadata