sweeper-fix
Pass
Audited by Gen Agent Trust Hub on Aug 10, 2026
Risk Level: SAFE
Full Analysis
- Input Handling and Safety Boundaries: The skill implements a defensive approach to external data by explicitly instructing the agent to treat GitHub issue content and review records as data rather than instructions. This directly addresses potential risks from indirect prompt injection contained within issue descriptions.
- Repository and Environment Verification: A series of preconditions are enforced to ensure the skill only operates within the official VS Code repository and starting from a clean git state. This prevents accidental modifications to the wrong project or unintended overwriting of a user's local work.
- Mandatory Human-in-the-Loop Review: The skill requires the agent to present a full diff and obtain explicit user confirmation before pushing any code to remote branches or creating pull requests. This ensures that a human developer validates the correctness and safety of the generated fix.
- Scoped Permission Usage: Operations involving the GitHub CLI and Git are restricted to the minimum necessary actions (fetching records, creating branches, and opening draft PRs) and are pinned to specific repositories to avoid ambiguity or redirect attacks.
Audit Metadata