AgentSkillsCN

refactoring-03-config-reproducibility

在改善Python研究代码的运行配置、种子与可重复性时使用。

SKILL.md
--- frontmatter
name: refactoring-03-config-reproducibility
description: Use when improving run configuration, seeding, and reproducibility for Python research code.

Refactoring 03: Config and Reproducibility

Goal

Make experiments reproducible by centralizing configuration, controlling randomness, and recording run metadata.

Sequence

  • •Order: 03
  • •Previous: refactoring-02-dependencies-env
  • •Next: refactoring-04-data-io-validation

Workflow

  • •Centralize runtime parameters into a config object or file (YAML/TOML/argparse).
    • •Success: All run parameters are set through a single config surface.
  • •Seed all RNGs (Python, NumPy, framework) and make the seed a first class parameter.
    • •Success: Runs are repeatable with the same seed.
  • •Record metadata: config snapshot, git commit hash, and environment info with outputs.
    • •Success: Each run output includes config and environment metadata.
  • •Create a consistent output directory layout for artifacts and metrics.
    • •Success: Outputs follow a documented directory structure.
  • •Remove hidden global state and implicit defaults where possible.
    • •Success: Behavior is driven by explicit parameters.

Guardrails

  • •Keep config changes backward compatible when possible.
  • •Do not add heavy config frameworks unless required.
  • •Favor explicit parameters over environment variables.