| """Structure is part of the grade: malformed CSV fails even when the values are right, while |
| harmless representation differences (quoting, a trailing blank line, CRLF) still pass.""" |
|
|
| import csv |
| import io |
| import os |
| import tempfile |
|
|
| import pytest |
|
|
| os.environ.setdefault("TERMINAL_DESK_APP", tempfile.mkdtemp(prefix="td-struct-")) |
|
|
| from terminal_desk.core.grading import render_artifact |
| from terminal_desk.families import all_task_ids, reserved_task_ids |
| from terminal_desk.models import Tbench2Action |
| from terminal_desk.server.environment import TerminalDeskEnvironment |
|
|
|
|
| def csv_tasks(): |
| out = [] |
| for t in all_task_ids() + reserved_task_ids(): |
| env = TerminalDeskEnvironment() |
| env.reset(task_id=t) |
| if env._ep["built"].artifact.path.endswith(".csv"): |
| out.append(t) |
| return out |
|
|
|
|
| CSV_TASKS = csv_tasks() |
|
|
|
|
| def grade(task_id: str, mutate) -> float: |
| env = TerminalDeskEnvironment() |
| env.reset(task_id=task_id) |
| built = env._ep["built"] |
| art = built.artifact |
| for path, text in built.notes.get("oracle_files", {}).items(): |
| env.step(Tbench2Action(action_type="write_file", file_path=path, content=text)) |
| content = mutate(render_artifact(art)) |
| env.step(Tbench2Action(action_type="write_file", file_path=art.path, content=content)) |
| return env.step(Tbench2Action(action_type="evaluate")).reward |
|
|
|
|
| def extra_field_every_row(text): |
| head, *rows = text.rstrip("\n").split("\n") |
| return "\n".join([head] + [r + ",extra" for r in rows]) + "\n" |
|
|
|
|
| def extra_field_one_row(text): |
| head, first, *rest = text.rstrip("\n").split("\n") |
| return "\n".join([head, first + ",extra"] + rest) + "\n" |
|
|
|
|
| def missing_field_one_row(text): |
| head, first, *rest = text.rstrip("\n").split("\n") |
| return "\n".join([head, first.rsplit(",", 1)[0]] + rest) + "\n" |
|
|
|
|
| def requote_all(text): |
| out = io.StringIO() |
| csv.writer(out, quoting=csv.QUOTE_ALL, lineterminator="\r\n").writerows(csv.reader(io.StringIO(text))) |
| return out.getvalue() + "\r\n" |
|
|
|
|
| def test_csv_tasks_found(): |
| assert len(CSV_TASKS) >= 4, CSV_TASKS |
|
|
|
|
| @pytest.mark.parametrize("task_id", CSV_TASKS) |
| @pytest.mark.parametrize("mutate", [extra_field_every_row, extra_field_one_row, missing_field_one_row]) |
| def test_malformed_csv_scores_zero(task_id, mutate): |
| assert grade(task_id, mutate) == 0.0 |
|
|
|
|
| @pytest.mark.parametrize("task_id", CSV_TASKS) |
| def test_harmless_csv_representation_still_scores_one(task_id): |
| assert grade(task_id, requote_all) == 1.0 |
|
|
