skill-refactor

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.75). The runtime gate ingests OUTSIDER-authored free text via test-prompts.json prompts (e.g., user-typed requests like “shrink this skill” / “Refactor my-creative-skill…”) into the LLM-judge ensemble prompts during Q1, so outsider text can enter the LLM context indirectly through the evaluation workflow.