repo-native-refactor
Pass
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill implements a sophisticated risk-assessment framework (R0-R4), requiring higher levels of justification and verification for changes to critical boundaries such as authentication, persistence, and concurrency.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external, potentially untrusted data from repository diffs and source code, creating an attack surface for indirect prompt injection.
- Ingestion points: Repository diffs, branches, and PR content (SKILL.md, references/repository-rehabilitation.md).
- Boundary markers: Instructions to treat prose and patterns as 'findings' or 'evidence' rather than authoritative commands (references/finding-taxonomy.md).
- Capability inventory: File system read and write access using the 'codex' tool.
- Sanitization: Employs an 'evidence hierarchy' and a 'minimal intervention' policy to validate the necessity of every change (references/shared-contracts.md).
- [SAFE]: Extensive guidelines on 'hard stops' are provided, directing the agent to halt execution if ownership, intent, or contract consequences cannot be clearly established from the repository evidence.
Audit Metadata