test: post_test hooks

tests/assets/problems/assess_post_tests.py

@@ -0,0 +1,31 @@
+from gapper import post_test, problem, test_cases
+from gapper.core.result_synthesizer import ResultSynthesizer
+from gapper.core.test_result import TestResult
+
+
+def check_at_least_half_is_correct(
+    synthesizer: ResultSynthesizer, result_proxy: TestResult
+) -> None:
+    if (
+        sum(result.is_passed for result in synthesizer.results)
+        > len(synthesizer.results) // 2
+    ):
+        result_proxy.set_max_score(0)
+        result_proxy.set_pass_status("passed")
+
+        result_proxy.set_extra_points(5)
+
+
+@test_cases.singular_param_iter([i for i in range(10)], gap_max_score=1)
+@post_test(check_at_least_half_is_correct)
+@problem()
+def square(x: int | float) -> int | float:
+    return x**2
+
+
+__problem_config__ = {
+    "is_script": False,
+    "check_stdout": False,
+    "mock_input": False,
+    "captured_context": (),
+}

tests/assets/submissions/assess_post_tests.py

@@ -0,0 +1,2 @@
+def square(x: int | float) -> int | float:
+    return x**2

tests/test_result_synthesizer.py

@@ -0,0 +1,24 @@
+import pytest
+
+from gapper.core.problem import Problem
+from gapper.core.result_synthesizer import ResultSynthesizer
+from gapper.core.test_result import TestResult
+from tests.conftest import _make_problem_name
+
+
+def test_post_test(request: pytest.FixtureRequest) -> None:
+    problem: Problem = request.getfixturevalue(
+        _make_problem_name("assess_post_tests.py")
+    )
+    results = [
+        TestResult("test result", pass_status="passed", max_score=1, score=1)
+        for _ in range(10)
+    ]
+    old_len = len(results)
+    old_score = sum(result.score for result in results)
+    syn = ResultSynthesizer(
+        results=results, post_tests=problem.post_tests, total_score=10
+    ).run_post_tests()
+
+    assert len(results) == old_len + 1
+    assert syn.synthesize_score() == old_score + 5