add: builder skill

[blizhan]

No description provided.

[gemini-code-assist]

Code Review

This pull request adds the aimx-hydra-lightning-builder skill, enabling users to scaffold, audit, and migrate machine learning repositories using Hydra, PyTorch Lightning, and Aim. It includes a project template, architectural guides, and scripts for repository analysis. The review feedback identified several optimization opportunities in the audit_repo.py script, including improving memory efficiency for file reading, optimizing directory traversal by pruning ignored paths, and reducing redundant I/O through content caching.

[gemini-code-assist]

The current implementation of read_text reads the entire file into memory before slicing it. This can lead to high memory usage or even OOM errors if the repository contains very large text files (e.g., logs or datasets). It is more efficient to read only up to the specified limit directly from the file stream.

Suggested change

	def read_text(path: Path, limit: int = 300_000) -> str:
	try:
	return path.read_text(encoding="utf-8", errors="ignore")[:limit]
	except OSError:
	return ""
	def read_text(path: Path, limit: int = 300_000) -> str:
	try:
	with path.open("r", encoding="utf-8", errors="ignore") as f:
	return f.read(limit)
	except OSError:
	return ""

[gemini-code-assist]

Using repo.rglob("*") and then checking path.parts for every file is inefficient because it traverses into ignored directories (like .git or .venv) before filtering them out. For large repositories, this can significantly slow down the audit. Using os.walk allows you to prune these directories and avoid unnecessary traversal.

Suggested change

	def iter_text_files(repo: Path) -> list[Path]:
	files: list[Path] = []
	for path in repo.rglob("*"):
	if any(part in IGNORED_PARTS for part in path.parts):
	continue
	if path.is_file() and path.suffix in TEXT_EXTENSIONS:
	files.append(path)
	return files
	def iter_text_files(repo: Path) -> list[Path]:
	import os
	files: list[Path] = []
	for root, dirs, filenames in os.walk(repo):
	dirs[:] = [d for d in dirs if d not in IGNORED_PARTS]
	for filename in filenames:
	path = Path(root) / filename
	if path.suffix in TEXT_EXTENSIONS:
	files.append(path)
	return files

[gemini-code-assist]

Using configs_dir.exists() is not sufficient because iterdir() will raise a NotADirectoryError if configs_dir exists but is a file. Using is_dir() is safer.

Suggested change

	config_groups = sorted(path.name for path in configs_dir.iterdir() if path.is_dir()) if configs_dir.exists() else []
	config_groups = sorted(path.name for path in configs_dir.iterdir() if path.is_dir()) if configs_dir.is_dir() else []

[gemini-code-assist]

The audit function calls grep multiple times, and each grep call re-reads the files from disk. This results in redundant I/O operations. It is more efficient to read the file contents once and then perform the regex searches on the cached strings.

    py_contents = {p: read_text(p) for p in py_files}
    yaml_contents = {p: read_text(p) for p in yaml_files}

    def find(contents: dict[Path, str], pattern: str) -> list[Path]:
        regex = re.compile(pattern)
        return [p for p, text in contents.items() if regex.search(text)]

    hydra_entrypoints = find(py_contents, r"@hydra\.main|hydra\.main\(")
    lightning_modules = find(py_contents, r"LightningModule|lightning\.LightningModule|L\.LightningModule")
    lightning_datamodules = find(py_contents, r"LightningDataModule|lightning\.LightningDataModule")
    trainer_usage = find(py_contents, r"\bTrainer\b|trainer\.fit|trainer\.validate|trainer\.test")
    aim_logger_configs = find(yaml_contents, r"aim\.pytorch_lightning\.AimLogger")
    direct_aim_tracks = find(py_contents, r"experiment\.track|aim_run\.track|from aim import")
    scalar_logs = find(py_contents, r"\bself\.log\(")
    hparam_logs = find(py_contents, r"log_hyperparams|log_hyperparameters")