code-refactor

Code Refactor Skill

Systematically refactor Python code to match this repo's standards.

Workflow

1. Read Target Code

Read every file in the target module. Never propose changes to code you haven't read.

2. Audit Against Checklist

Run through references/checklist.md and note every violation. Group them by category so the user can see the full picture before any edits are made.

3. Run Static Analysis

uv run ruff check <target>
uv run mypy <target>

Report all errors as part of the audit — these are additional violations on top of the checklist.

4. Apply Fixes

Apply all fixes. Prefer the Edit tool over Write for existing files (smaller diffs, easier review). Make all independent edits in parallel.

Priority order (highest impact first):

1. Type safety — enums over bare strings, from __future__ import annotations, proper return types
2. Pydantic — snake_case fields with aliases, model_config, field validators (see references/pydantic-patterns.md)
3. Error handling — replace assert with if/raise, add context to re-raised exceptions
4. Python best practices — frozen dataclasses, Protocol types, avoid mutable defaults
5. PyTorch/Lightning — only apply if the module contains torch or nn.Module code (see references/pytorch-patterns.md)
6. Docs — public functions must have docstrings (see references/docs-patterns.md)

5. Verify

uv run ruff check <target>
uv run ruff format <target>
uv run mypy <target>
uv run pytest tests/unit/<target-name>/ -v

All checks must pass before declaring done. If tests fail, fix them — don't skip.

6. Summarise

Report a table of every change made, grouped by category. Include file and line references.

---

Repo-Specific Conventions

• Python: 3.12 — use StrEnum, match, X | Y unions, from __future__ import annotations in every file
• Linter/Formatter: Ruff (88 char line length)
• Type checker: mypy (strict — all errors must be resolved)
• Package manager: uv (uv run <cmd>)
• Tests: pytest, mirror source structure under tests/unit/
• Logging: from logging_utils import get_logger then logger = get_logger(__name__) — never logging.getLogger(__name__) or print()
• Imports: stdlib → third-party → local, each group separated by blank line
• Config: TOML or YAML only — never JSON for config files