Skip to content

AGENTS.md

Latest commit

 

History

History
90 lines (66 loc) · 5.06 KB

AGENTS.md

File metadata and controls

90 lines (66 loc) · 5.06 KB
Raw

Agents Guide

This document provides guidance for AI coding agents working with the kaggle-environments repository. It covers project structure, commands, code style, and architecture.

For detailed how-to guides, see the skills in .agents/skills/:

  • create-environment -- step-by-step guide for building a new game environment (Python backend)
  • create-harness -- step-by-step guide for writing an LLM harness (connecting a language model to a game)
  • review-harness -- methodology for auditing an existing LLM harness for gameplay-impacting bugs
  • review-docs -- methodology for auditing an environment's README.md and AGENTS.md against its engine implementation
  • create-visualizer -- step-by-step guide for building a web visualizer for any game (regular or OpenSpiel)
  • onboard-open-spiel-game -- step-by-step guide for adding an OpenSpiel game (Python backend)

Project Overview

Kaggle Environments is a Python framework for evaluating episodes in competitive multi-agent environments. It provides a plugin-based system where game environments (e.g., ConnectX, RPS, Chess, Lux AI) are registered and run with configurable agents. The emphasis is on episode evaluation rather than agent training.

Common Commands

Testing

./run_tests.sh                          # Run all tests locally with uv
./run_tests.sh -k "rps"                 # Run tests matching a pattern
uv sync && uv run pytest tests/envs/rps/test_rps.py  # Run a single test file
./run_tests.sh --docker                 # Run tests in Docker container
./run_tests.sh --multicontainer         # Run multi-container integration tests

Linting & Formatting

uv run ruff check --fix .              # Lint with auto-fix
uv run ruff format .                   # Format code
pnpm format                            # Format TS/JS with prettier

Pre-commit hooks run ruff-check --fix and ruff-format automatically.

Package Management

uv sync                                # Install/sync Python dependencies
pnpm install                           # Install frontend dependencies

Frontend (from repo root)

pnpm dev                                # Run a visualizer dev server (interactive game picker)
pnpm build                              # Build a single visualizer (interactive picker)
pnpm build-all                          # Build all visualizers
pnpm test:e2e                           # Run Playwright end-to-end tests

Code Style

  • Python: 3.11+, line length 120, double quotes, space indentation. Import sorting via ruff (extend-select = ["I"]). Build backend: flit.
  • TypeScript/JS: Formatted with prettier, linted with eslint.
  • Git hooks: lefthook runs linters/formatters before committing.

Architecture

Core Framework (kaggle_environments/)

  • __init__.py -- Auto-discovers and registers all environments from envs/ at import time. Exports make, evaluate, register, Agent.
  • core.py -- Environment class: the main runtime that manages specification validation, state machine (ACTIVE -> DONE/ERROR/INVALID/TIMEOUT), interpreter execution, and agent coordination.
  • agent.py -- Agent class: wraps agent functions with timeout handling and error capture. Agents can be Python functions, file paths, URLs, inline strings, or fixed actions.
  • utils.py -- Schema validation (via jsonschema), Struct (dot-access dicts), file utilities.
  • main.py -- CLI entrypoint and Flask HTTP server.
  • errors.py -- Exception hierarchy based on canonical error codes.

Environment Plugins (kaggle_environments/envs/<name>/)

Each environment is a self-contained directory. The main module (<name>.py) must export: specification, interpreter, renderer, html_renderer, and optionally agents. Discovery is automatic at import time.

The open_spiel_env is a special case: it wraps the OpenSpiel library and registers multiple game environments from an ENV_REGISTRY dict.

Tests (tests/envs/)

Test files mirror the environment structure. Tests use make() to create an environment, run() with agents, and assert on env.toJSON() (statuses, rewards, steps).

Web Visualizers

Each environment can have a Vite + TypeScript visualizer at envs/<name>/visualizer/default/. The pnpm workspace (pnpm-workspace.yaml) links web/* and all visualizer/* directories. web/core/ (@kaggle-environments/core) provides shared UI components, replay adapters, and playback controls used by all visualizers.

Key Reference Files

When building environments or visualizers, these files are useful references:

  • kaggle_environments/envs/rps/ -- simplest complete environment (good starting template)
  • kaggle_environments/envs/connectx/ -- board game with shared observations and per-agent defaults
  • kaggle_environments/schemas.json -- framework-level defaults for configuration, observation, status
  • web/vite.config.base.ts -- shared Vite build config for all visualizers
  • web/tsconfig.base.json -- shared TypeScript config
  • web/core/src/index.ts -- all exports from @kaggle-environments/core