The Agent Skills Directory

[SAFE]: The skill acts as a procedural guide for refactoring tasks, emphasizing human-in-the-loop oversight and environmental isolation.\n- [COMMAND_EXECUTION]: Describes the use of standard search utilities like grep to identify code patterns across a repository. This is a legitimate and expected operation for the described workflow.\n- [REMOTE_CODE_EXECUTION]: Outlines a phase where agents implement code changes and execute tests. While this involves code execution, the risk is mitigated by scoping the work to isolated git worktrees and mandating a manual approval step from the user.\n- [PROMPT_INJECTION]: The skill is subject to indirect prompt injection risks because it reads and processes repository content (Ingestion: Phase 1). It uses human verification of the proposed unit plan as a boundary marker (Boundary: Phase 3). The agents have the capability to write code and run test suites (Capability: Phase 4). Human review of the decomposed tasks serves as the primary sanitization mechanism (Sanitization: Phase 3).

batch-orchestration