AGENTS.md

This file is the agent router for GitHub Copilot and Codex. It provides high-level rules and directs agents to the appropriate skill files for detailed instructions.

CLAUDE.md carries the same content for Claude Code. Everything from ## Hard Rules onward must match between the two files; scripts/sync_agent_docs.py --check enforces this.

This is a Python 3.11 data science / analysis project.

Detailed task-specific procedures are in .claude/skills/*/SKILL.md. Project-specific context is in docs/agent/*.

Hard Rules (Always Apply)

• Never commit raw data, credentials, API keys, tokens, or customer-level records.
• Never modify, overwrite, delete, or regenerate raw data directly.
• Prefer small, reviewable changes.
• Explain assumptions before non-trivial analytical decisions.
• Ask for clarification when data semantics are unclear.
• Use uv exclusively for Python dependency management. Never use pip, conda, poetry, or pipenv.
• When applying a statistical, ML, causal, simulation, or optimization method, always apply the matching *-diagnostics skill (図と値をセットで出す).

Package Management

• Use uv exclusively for all dependency management.
• Never use pip, pip3, python -m pip, poetry, conda, pipenv, or easy_install.

Data Safety

• Never commit raw data, credentials, API keys, tokens, or customer-level records.
• Never modify or delete raw data directly.
• Treat data/raw and data/external as immutable.

Common Commands

uv sync
uv run pytest
uv run ruff check .
uv run ruff format .          # ローカル整形
uv run ruff format --check .  # CI用（差分チェックのみ）
uv run mypy src
uv run python scripts/check_no_raw_data_commit.py
uv run python scripts/check_no_sensitive_patterns.py

Key Conventions

• DataFrame operations: prefer polars over pandas.
• Visualization: use fig, ax = plt.subplots(...), not plt.figure(...).
• File paths: use pathlib.Path, no absolute local paths.
• Docstrings: Google-style.
• Inline comments: Japanese.
• Reports and documentation: Japanese.

File-Specific Guidelines

Python files (**/*.py)

• Follow python-style skill: type hints, Google-style docstrings, Japanese comments.
• Follow dataframe-polars skill: prefer polars over pandas.
• Follow path-and-io skill: use pathlib.Path, no absolute paths.

Data files (data/**)

• Do not create, modify, or commit real data files under data/raw/ or data/external/ — these are immutable.
• Writing derived data to data/interim/ or data/processed/ is allowed in code, but do not commit large data files.
• Only .gitkeep files and documentation (README, markdown) are acceptable to commit.
• See safe-data-handling skill.

SQL files (**/*.sql)

• Use explicit column names; avoid SELECT *.
• Use CTEs for readability.
• Add date filters for large tables.
• Validate join cardinality and row counts.
• Never run DROP, TRUNCATE, DELETE, or UPDATE unless explicitly requested.
• See sql-analysis skill.

Notebooks (**/*.ipynb, notebooks/**/*.md)

• Keep restartable from a clean kernel; no secrets in outputs.
• See notebook-workflow and safe-data-handling skills.

Documentation (README.md, docs/**/*.md)

• Write documentation in clear Japanese.
• Keep code snippets, commands, file paths, and config values in English as-is.
• Do not translate technical terms unnecessarily.

Skill Routing

Project Context (docs/agent)

このリポジトリでしか通用しない知識だけを置く。作業手順・規約は skill 側が正本であり、 ここには重複させない（持ち運ぶ単位は .claude/skills/<name>/ ディレクトリ）。 ディレクトリ構成は README.md を参照する。