The skill promises a fully read-only review but instructs agents to create report files without disclosing an exception.
- Category
- Output Handling
- Confidence
- 98% confidence
- Finding
The read-only assurance conflicts with mandatory report generation and checkpoint advancement elsewhere in the skill. This overlaps with SDI-4 rather than demonstrating a separate exploit, but it can mislead users about authorized side effects and encourage agents to disregard the stated boundary.
- Content
