vasiuuu/DGX_AI

1from __future__ import annotations2 3from dataclasses import dataclass4 5 6@dataclass(frozen=True)7class Task:8    task_id: str9    task_level: str10    brief: str11    initial_files: dict[str, str]12    target_score: float13    max_budget: int14    tools: tuple[str, ...]15    hidden_tests: dict[str, str] = ()  # type: ignore[assignment]16    """Hidden correctness tests injected by the environment during grading.17 18    The agent never sees these. They are written into the sandbox temp dir19    alongside the agent's submitted files so pytest runs them automatically.20    This prevents "clean garbage" exploits where syntactically valid but21    semantically wrong code scores perfectly.22    """23 24 25# -- Hidden test suites (agent never sees these) ----------------------------26 27_HIDDEN_EASY = {28    "test_hidden_greet.py": (29        "from __future__ import annotations\n"30        "from main import greet\n\n"31        "def test_greet_alice() -> None:\n"32        '    assert greet("Alice") == "Hello, Alice!"\n\n'33        "def test_greet_bob() -> None:\n"34        '    assert greet("Bob") == "Hello, Bob!"\n\n'35        "def test_greet_empty() -> None:\n"36        '    assert greet("") == "Hello, !"\n'37    ),38}39 40_HIDDEN_MEDIUM = {41    "test_hidden_greet.py": (42        "from __future__ import annotations\n"43        "import pytest\n"44        "from main import greet\n\n"45        "def test_greet_alice() -> None:\n"46        '    assert greet("Alice") == "Hello, Alice!"\n\n'47        "def test_greet_none_raises() -> None:\n"48        "    with pytest.raises(ValueError):\n"49        "        greet(None)  # type: ignore[arg-type]\n\n"50        "def test_greet_returns_str() -> None:\n"51        '    assert isinstance(greet("X"), str)\n'52    ),53}54 55_HIDDEN_HARD = {56    "test_hidden_core.py": (57        "from __future__ import annotations\n"58        "import pytest\n"59        "from core import greet\n\n"60        "def test_greet_alice() -> None:\n"61        '    assert greet("Alice") == "Hello, Alice!"\n\n'62        "def test_greet_bob() -> None:\n"63        '    assert greet("Bob") == "Hello, Bob!"\n\n'64        "def test_greet_returns_str() -> None:\n"65        '    assert isinstance(greet("X"), str)\n\n'66        "def test_greet_empty() -> None:\n"67        '    assert greet("") == "Hello, !"\n'68    ),69}70 71 72TASKS: tuple[Task, ...] = (73    Task(74        task_id="greet_single_file",75        task_level="easy",76        brief=(77            "Implement `greet(name)` in `main.py` so that `greet(\"Alice\")` returns "78            '`"Hello, Alice!"`. Use type hints. Keep the module under 15 lines.'79        ),80        initial_files={"main.py": "def greet(name):\n    pass\n"},81        target_score=0.90,82        max_budget=4,83        tools=("ruff", "imports", "mypy", "pytest"),84        hidden_tests=_HIDDEN_EASY,85    ),86    Task(87        task_id="greet_with_tests",88        task_level="medium",89        brief=(90            "Extend `main.py` so that `greet(None)` raises `ValueError`, "91            "and add a `test_main.py` with pytest assertions. Keep `ruff` and "92            "`mypy --strict` clean."93        ),94        initial_files={95            "main.py": (96                "from __future__ import annotations\n\n\n"97                "def greet(name: str) -> str:\n"98                '    return f"Hello, {name}!"\n'99            ),100            "test_main.py": "",101        },102        target_score=0.80,103        max_budget=6,104        tools=("ruff", "imports", "mypy", "pytest"),105        hidden_tests=_HIDDEN_MEDIUM,106    ),107    Task(108        task_id="multi_file_module",109        task_level="hard",110        brief=(111            "Split into three files: `main.py` (entry), `core.py` (the greet "112            "function), `test_core.py` (tests). Every function must be type-hinted. "113            "All tests pass. `mypy --strict` clean."114        ),115        initial_files={116            "main.py": (117                "from __future__ import annotations\n\nfrom core import greet\n\n\n"118                'if __name__ == "__main__":\n'119                '    print(greet("World"))\n'120            ),121            "core.py": "",122            "test_core.py": "",123        },124        target_score=0.70,125        max_budget=10,126        tools=("ruff", "imports", "mypy", "pytest"),127        hidden_tests=_HIDDEN_HARD,128    ),129)130 131 132def get_task(task_level: str) -> Task:133    for t in TASKS:134        if t.task_level == task_level:135            return t136    msg = f"unknown task_level: {task_level!r} (expected easy|medium|hard)"137    raise ValueError(msg)138