plan-loop
Pass
Audited by Gen Agent Trust Hub on Jul 5, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: No malicious patterns or behaviors such as credential exfiltration, unauthorized network access, or persistence mechanisms were detected. The skill's functionality is limited to coordinating file-based reviews between the user and established AI tools.
- [PROMPT_INJECTION]: The skill processes external design plans and source code files, which constitutes a potential surface for indirect prompt injection. However, this is mitigated by the workflow's structure.
- Ingestion points: Local design plan files and related source code files identified in Step 1 and Step 2 of the SKILL.md workflow.
- Boundary markers: Prompts for the subagents use Markdown headers (e.g., '## プランファイル') to separate external file content from instructions.
- Capability inventory: The skill reads local files, invokes subagents (Codex/Copilot), and modifies local files based on user-approved feedback.
- Sanitization: No automated content filtering is described, but the workflow requires the user to explicitly approve specific feedback points before the 'apply_feedback' function modifies any files, providing a critical human verification layer.
Audit Metadata