mbtorch-refactor

What problem does it solve?

Provides a disciplined, test-first approach to refactoring the MbTorch codebase so that internal structure, readability, and layering are improved without changing external behavior or breaking tests. It turns vague impulses to "clean up code" into concrete, auditable steps that respect project layering and TDD requirements.

Core Features & Use Cases

• Refactoring coach: Guides the developer through safety rules, explicit problem statements, and small, test-verified change units.
• Layering & dependency enforcement: Highlights and prevents layer-violation refactors (core, nn, optim, io) and suggests correct extraction locations.
• TDD integration: Requires adding or strengthening tests before structural changes and prescribes atomic steps that keep tests green at every checkpoint.
• Typical refactorings: File/function splitting for large tensor modules, consolidating scattered I/O into io/, and moving pure numeric ops from nn to core.
• Use case: When core/tensor.mbt exceeds maintainable size or duplicated logic appears across modules, the skill produces a test-first plan to split, move, and re-export pieces safely.

Quick Start

Ask the assistant to analyze the target files, list concrete problems in the prescribed "問題/影響/対象/提案" format, and propose the first atomic refactoring step with required tests.