arch-step
Pass
Audited by Gen Agent Trust Hub on Oct 12, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses bundled Python scripts, such as
arch_stage_gate.py, to manage the architecture planning workflow. These scripts use SHA-256 hashing to create 'receipts' in the documentation, ensuring that each planning stage is verified and that the agent has actually modified the artifact before proceeding. - [EXTERNAL_DOWNLOADS]: Through the
external-researchcommand, the agent can perform web searches to anchor plans in industry best practices. The skill provides strict guidelines for this research, focusing on synthesizing authoritative information and explicitly forbidding the execution of remote scripts. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external research and the existing codebase, creating a surface for indirect prompt injection. It manages this risk by using structured comment-based boundary markers (e.g.,
<!-- arch_skill:block:research_grounding:start -->) to isolate processed data within the main architecture document. - [SAFE]: The skill includes maintenance scripts designed to clean up or verify the absence of legacy session hooks in platform configuration files (such as Claude or Codex settings). These scripts are restricted to removing specific vendor-managed hook entries and use standard file-locking mechanisms to prevent corruption.
Audit Metadata