tests/test_interactive_experiment.py

import unittest

from agent_harness.interactive_experiment import (
    InteractiveExperimentError,
    parse_query,
    score_selection_response,
)


class InteractiveExperimentTests(unittest.TestCase):
    def test_invalid_query_response_is_rejected_for_runner_fallback(self) -> None:
        response = {"choices": [{"message": {"content": "unfinished reasoning"}}]}

        with self.assertRaises(InteractiveExperimentError):
            parse_query(response)

    def test_invalid_candidate_selection_is_scored_as_protocol_violation(self) -> None:
        response = {
            "choices": [
                {
                    "message": {
                        "content": '{"files":["not-offered.go"],"reasoning":"guess"}'
                    }
                }
            ]
        }

        selection, violation = score_selection_response(response, {"offered.go"})

        self.assertEqual(selection, {"files": [], "reasoning": ""})
        self.assertIsNotNone(violation)
        self.assertIn("outside the candidates", violation or "")

    def test_valid_candidate_selection_has_no_protocol_violation(self) -> None:
        response = {
            "choices": [
                {
                    "message": {
                        "content": '{"files":["offered.go"],"reasoning":"direct match"}'
                    }
                }
            ]
        }

        selection, violation = score_selection_response(response, {"offered.go"})

        self.assertEqual(selection["files"], ["offered.go"])
        self.assertIsNone(violation)