add comprehensive unit test suite for all arenas (#84)

Author: Muhtasham

tests/arenas/__init__.py

@@ -0,0 +1 @@
+# Arena unit tests

tests/arenas/conftest.py

@@ -0,0 +1,118 @@
+"""
+Shared fixtures and mocks for arena unit tests.
+
+These fixtures allow testing arena logic (validation, result parsing)
+without requiring Docker containers or actual game execution.
+"""
+
+from pathlib import Path
+from typing import Any
+from unittest.mock import MagicMock
+
+import pytest
+
+
+class MockEnvironment:
+    """Mock environment that simulates container file system and command execution."""
+
+    def __init__(self, files: dict[str, str] | None = None, command_outputs: dict[str, dict] | None = None):
+        """
+        Args:
+            files: Dict mapping file paths to their contents
+            command_outputs: Dict mapping command prefixes to their outputs
+                             Format: {"ls": {"output": "file1.py\nfile2.py", "returncode": 0}}
+        """
+        self.files = files or {}
+        self.command_outputs = command_outputs or {}
+        self.config = MagicMock()
+        self.config.cwd = "/workspace"
+        self._executed_commands: list[str] = []
+
+    def execute(self, cmd: str, cwd: str | None = None, timeout: int | None = None) -> dict[str, Any]:
+        """Simulate command execution based on configured outputs."""
+        self._executed_commands.append(cmd)
+
+        # Check for exact matches first
+        if cmd in self.command_outputs:
+            return self.command_outputs[cmd]
+
+        # Check for prefix matches
+        for prefix, output in self.command_outputs.items():
+            if cmd.startswith(prefix):
+                return output
+
+        # Default behavior for common commands
+        if cmd.startswith("ls"):
+            # Extract path from command
+            parts = cmd.split()
+            path = parts[1] if len(parts) > 1 else "."
+            matching_files = [Path(f).name for f in self.files.keys() if f.startswith(path) or path == "."]
+            return {"output": "\n".join(matching_files), "returncode": 0}
+
+        if cmd.startswith("cat "):
+            file_path = cmd.split("cat ", 1)[1].strip()
+            if file_path in self.files:
+                return {"output": self.files[file_path], "returncode": 0}
+            return {"output": f"cat: {file_path}: No such file or directory", "returncode": 1}
+
+        if cmd.startswith("test -f ") and "echo" in cmd:
+            file_path = cmd.split("test -f ")[1].split(" &&")[0].strip()
+            exists = file_path in self.files
+            return {"output": "exists" if exists else "", "returncode": 0 if exists else 1}
+
+        if cmd.startswith("test -d ") and "echo" in cmd:
+            dir_path = cmd.split("test -d ")[1].split(" &&")[0].strip()
+            # Check if any file path starts with this directory
+            exists = any(f.startswith(dir_path + "/") or f == dir_path for f in self.files.keys())
+            return {"output": "exists" if exists else "", "returncode": 0 if exists else 1}
+
+        # Default: command succeeded with no output
+        return {"output": "", "returncode": 0}
+
+
+class MockPlayer:
+    """Mock player for testing arena validation and result parsing."""
+
+    def __init__(self, name: str, environment: MockEnvironment | None = None):
+        self.name = name
+        self.environment = environment or MockEnvironment()
+
+
+def create_mock_player(name: str, files: dict[str, str] | None = None, **kwargs) -> MockPlayer:
+    """Create a mock player with specified file system contents."""
+    env = MockEnvironment(files=files, **kwargs)
+    return MockPlayer(name=name, environment=env)
+
+
+@pytest.fixture
+def mock_player_factory():
+    """Factory fixture for creating mock players."""
+    return create_mock_player
+
+
+@pytest.fixture
+def minimal_config():
+    """Minimal config dict for arena initialization."""
+    return {
+        "game": {
+            "name": "Test",
+            "sims_per_round": 10,
+        },
+        "tournament": {
+            "rounds": 3,
+        },
+        "players": [
+            {"name": "p1", "agent": "dummy"},
+            {"name": "p2", "agent": "dummy"},
+        ],
+    }
+
+
+@pytest.fixture
+def tmp_log_dir(tmp_path):
+    """Create a temporary log directory structure."""
+    log_dir = tmp_path / "logs"
+    log_dir.mkdir()
+    rounds_dir = log_dir / "rounds"
+    rounds_dir.mkdir()
+    return log_dir

tests/arenas/test_battlecode.py

@@ -0,0 +1,224 @@
+"""
+Unit tests for BattleCodeArena.
+
+Tests validate_code() and get_results() methods without requiring Docker.
+"""
+
+import pytest
+
+from codeclash.arenas.arena import RoundStats
+from codeclash.arenas.battlecode.battlecode import BC_FOLDER, BC_LOG, BC_TIE, BattleCodeArena
+
+from .conftest import MockPlayer
+
+VALID_BOT_PY = """
+from battlecode25.stubs import *
+
+def turn():
+    # Simple bot that does nothing
+    pass
+"""
+
+
+class TestBattleCodeValidation:
+    """Tests for BattleCodeArena.validate_code()"""
+
+    @pytest.fixture
+    def arena(self, tmp_log_dir, minimal_config):
+        """Create BattleCodeArena instance with mocked environment."""
+        arena = BattleCodeArena.__new__(BattleCodeArena)
+        arena.submission = f"src/{BC_FOLDER}"
+        arena.log_local = tmp_log_dir
+        arena.run_cmd_round = "python run.py run"
+        return arena
+
+    def test_valid_submission(self, arena, mock_player_factory):
+        """Test that a valid bot.py passes validation."""
+        player = mock_player_factory(
+            name="test_player",
+            files={
+                f"src/{BC_FOLDER}/bot.py": VALID_BOT_PY,
+            },
+            command_outputs={
+                "ls src": {"output": f"{BC_FOLDER}\n", "returncode": 0},
+                f"ls src/{BC_FOLDER}": {"output": "bot.py\n__init__.py\n", "returncode": 0},
+                f"cat src/{BC_FOLDER}/bot.py": {"output": VALID_BOT_PY, "returncode": 0},
+            },
+        )
+        is_valid, error = arena.validate_code(player)
+        assert is_valid is True
+        assert error is None
+
+    def test_missing_mysubmission_directory(self, arena, mock_player_factory):
+        """Test that missing src/mysubmission/ fails validation."""
+        player = mock_player_factory(
+            name="test_player",
+            files={},
+            command_outputs={
+                "ls src": {"output": "otherpackage\n", "returncode": 0},
+            },
+        )
+        is_valid, error = arena.validate_code(player)
+        assert is_valid is False
+        assert BC_FOLDER in error
+
+    def test_missing_bot_file(self, arena, mock_player_factory):
+        """Test that missing bot.py fails validation."""
+        player = mock_player_factory(
+            name="test_player",
+            files={
+                f"src/{BC_FOLDER}/__init__.py": "",
+            },
+            command_outputs={
+                "ls src": {"output": f"{BC_FOLDER}\n", "returncode": 0},
+                f"ls src/{BC_FOLDER}": {"output": "__init__.py\n", "returncode": 0},
+            },
+        )
+        is_valid, error = arena.validate_code(player)
+        assert is_valid is False
+        assert "bot.py" in error
+
+    def test_missing_turn_function(self, arena, mock_player_factory):
+        """Test that bot.py without turn() function fails validation."""
+        invalid_bot = """
+from battlecode25.stubs import *
+
+def setup():
+    pass
+
+def run():
+    pass
+"""
+        player = mock_player_factory(
+            name="test_player",
+            files={
+                f"src/{BC_FOLDER}/bot.py": invalid_bot,
+            },
+            command_outputs={
+                "ls src": {"output": f"{BC_FOLDER}\n", "returncode": 0},
+                f"ls src/{BC_FOLDER}": {"output": "bot.py\n", "returncode": 0},
+                f"cat src/{BC_FOLDER}/bot.py": {"output": invalid_bot, "returncode": 0},
+            },
+        )
+        is_valid, error = arena.validate_code(player)
+        assert is_valid is False
+        assert "turn()" in error
+
+
+class TestBattleCodeResults:
+    """Tests for BattleCodeArena.get_results()"""
+
+    @pytest.fixture
+    def arena(self, tmp_log_dir, minimal_config):
+        """Create BattleCodeArena instance."""
+        config = minimal_config.copy()
+        config["game"]["name"] = "BattleCode"
+        config["game"]["sims_per_round"] = 3
+        arena = BattleCodeArena.__new__(BattleCodeArena)
+        arena.submission = f"src/{BC_FOLDER}"
+        arena.log_local = tmp_log_dir
+        arena.config = config
+        arena.logger = type("Logger", (), {"debug": lambda self, msg: None, "info": lambda self, msg: None})()
+        return arena
+
+    def _create_sim_log(self, round_dir, idx: int, winner_key: str, is_coin_flip: bool = False):
+        """
+        Create a simulation log file.
+
+        Args:
+            winner_key: "A" or "B" to indicate which player won
+            is_coin_flip: If True, sets reason to coin flip (arbitrary win)
+        """
+        log_file = round_dir / BC_LOG.format(idx=idx)
+        reason = BC_TIE if is_coin_flip else "Reason: Team won by controlling more territory."
+        # The log format has winner info in third-to-last line
+        log_file.write_text(
+            f"""Round starting...
+Turn 100...
+Turn 200...
+Winner: Team ({winner_key}) wins (game over)
+{reason}
+Final stats
+"""
+        )
+
+    def test_parse_results_player_a_wins(self, arena, tmp_log_dir):
+        """Test parsing results when player A (first player) wins."""
+        round_dir = tmp_log_dir / "rounds" / "1"
+        round_dir.mkdir(parents=True)
+
+        # A wins 2 games, B wins 1
+        self._create_sim_log(round_dir, 0, "A")
+        self._create_sim_log(round_dir, 1, "A")
+        self._create_sim_log(round_dir, 2, "B")
+
+        agents = [MockPlayer("Alice"), MockPlayer("Bob")]
+        stats = RoundStats(round_num=1, agents=agents)
+
+        arena.get_results(agents, round_num=1, stats=stats)
+
+        assert stats.winner == "Alice"
+        assert stats.scores["Alice"] == 2
+        assert stats.scores["Bob"] == 1
+
+    def test_parse_results_player_b_wins(self, arena, tmp_log_dir):
+        """Test parsing results when player B (second player) wins."""
+        round_dir = tmp_log_dir / "rounds" / "1"
+        round_dir.mkdir(parents=True)
+
+        # A wins 1 game, B wins 2
+        self._create_sim_log(round_dir, 0, "B")
+        self._create_sim_log(round_dir, 1, "B")
+        self._create_sim_log(round_dir, 2, "A")
+
+        agents = [MockPlayer("Alice"), MockPlayer("Bob")]
+        stats = RoundStats(round_num=1, agents=agents)
+
+        arena.get_results(agents, round_num=1, stats=stats)
+
+        assert stats.winner == "Bob"
+        assert stats.scores["Alice"] == 1
+        assert stats.scores["Bob"] == 2
+
+    def test_parse_results_with_coin_flips(self, arena, tmp_log_dir):
+        """Test parsing results where some wins are coin flips (don't count)."""
+        round_dir = tmp_log_dir / "rounds" / "1"
+        round_dir.mkdir(parents=True)
+
+        # Coin flip wins should be treated as ties
+        self._create_sim_log(round_dir, 0, "A")
+        self._create_sim_log(round_dir, 1, "A", is_coin_flip=True)  # Doesn't count
+        self._create_sim_log(round_dir, 2, "B")
+
+        agents = [MockPlayer("Alice"), MockPlayer("Bob")]
+        stats = RoundStats(round_num=1, agents=agents)
+
+        arena.get_results(agents, round_num=1, stats=stats)
+
+        # Only non-coin-flip wins count
+        assert stats.scores["Alice"] == 1
+        assert stats.scores["Bob"] == 1
+
+
+class TestBattleCodeConfig:
+    """Tests for BattleCodeArena configuration and properties."""
+
+    def test_arena_name(self):
+        """Test that arena has correct name."""
+        assert BattleCodeArena.name == "BattleCode"
+
+    def test_submission_path(self):
+        """Test that submission path is correct."""
+        assert BattleCodeArena.submission == f"src/{BC_FOLDER}"
+
+    def test_bc_folder_name(self):
+        """Test that BC folder name is mysubmission."""
+        assert BC_FOLDER == "mysubmission"
+
+    def test_default_args(self):
+        """Test default arguments."""
+        assert BattleCodeArena.default_args.get("maps") == "quack"
+
+    def test_description_mentions_python(self):
+        """Test that description mentions Python as the language."""
+        assert "python" in BattleCodeArena.description.lower()