Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

40.7 KB
Raw Blame
"""
Coverage tests for benchmarks/runners.py.

Targets uncovered paths: dataset fallback loading, example processing with
dataset class methods, legacy browsecomp field handling, quick_summary calls,
error handling, evaluation paths (human, automated, fallback), metrics/report
generation, progress callbacks at every stage, and convenience wrappers.
"""

import json
import tempfile
from pathlib import Path
from unittest.mock import Mock, patch

import pytest


MODULE = "local_deep_research.benchmarks.runners"


@pytest.fixture(autouse=True)
def _pin_data_directory(tmp_path, monkeypatch):
    """Keep benchmark output out of the real user data directory.

    run_benchmark resolves its output directory inside LDR_DATA_DIR and replaces
    any other location with a default directory there. Pinning the data
    directory to this test's tmp_path keeps that default per test.
    """
    monkeypatch.setenv("LDR_DATA_DIR", str(tmp_path))


# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------


def _make_example(problem="Q1", answer="A1", example_id="ex_1", **extra):
    d = {"id": example_id, "problem": problem, "answer": answer}
    d.update(extra)
    return d


def _quick_summary_response(summary="Response text", sources=None):
    return {"summary": summary, "sources": sources or []}


def _setup_registry_with_fake_class(mock_registry, examples):
    """Set up mock registry so isinstance check fails (legacy path)."""
    FakeClass = type("OtherClass", (), {})
    mock_dataset = Mock()
    mock_dataset.load.return_value = examples
    mock_registry.create_dataset.return_value = mock_dataset
    mock_registry.get_dataset_class.return_value = FakeClass


# ---------------------------------------------------------------------------
# format_query (small gap: default parameter)
# ---------------------------------------------------------------------------


class TestFormatQueryDefaults:
    def test_default_dataset_type_is_simpleqa(self):
        from local_deep_research.benchmarks.runners import format_query

        assert format_query("hello") == "hello"


# ---------------------------------------------------------------------------
# run_benchmark - dataset loading
# ---------------------------------------------------------------------------


class TestDatasetLoadingFallback:
    """Cover lines 97-107: fallback to legacy load_dataset on registry error."""

    @patch(f"{MODULE}.load_dataset")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_fallback_to_legacy_load_dataset(self, mock_registry, mock_load):
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_registry.create_dataset.side_effect = ValueError("no such dataset")
        mock_load.return_value = []

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
            )

        mock_load.assert_called_once()
        assert result["status"] == "complete_no_eval"


# ---------------------------------------------------------------------------
# run_benchmark - example processing with dataset class methods
# ---------------------------------------------------------------------------


class TestExampleProcessingDatasetClass:
    """Cover lines 146-155: using dataset_instance.get_question / get_answer."""

    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_uses_dataset_class_methods(
        self, mock_registry, mock_qs, mock_extract
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        FakeDatasetClass = type(
            "FakeDataset",
            (),
            {
                "load": lambda self: [
                    {"id": "1", "problem": "Q", "answer": "A"}
                ],
                "get_question": lambda self, ex: "class_question",
                "get_answer": lambda self, ex: "class_answer",
            },
        )
        instance = FakeDatasetClass()
        mock_registry.create_dataset.return_value = instance
        mock_registry.get_dataset_class.return_value = FakeDatasetClass

        mock_qs.return_value = _quick_summary_response()
        mock_extract.return_value = {
            "extracted_answer": "ans",
            "confidence": 0.9,
        }

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
            )
            assert result["total_examples"] == 1
            with open(result["results_path"]) as f:
                line = json.loads(f.readline())
            assert line["correct_answer"] == "class_answer"
            assert line["problem"] == "class_question"


# ---------------------------------------------------------------------------
# run_benchmark - legacy fallback for simpleqa and browsecomp
# ---------------------------------------------------------------------------


class TestLegacyFieldExtraction:
    """Cover lines 156-167: legacy approach for extracting question/answer."""

    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_legacy_simpleqa_fields(self, mock_registry, mock_qs, mock_extract):
        from local_deep_research.benchmarks.runners import run_benchmark

        _setup_registry_with_fake_class(
            mock_registry,
            [{"id": "1", "problem": "legacy_q", "answer": "legacy_a"}],
        )
        mock_qs.return_value = _quick_summary_response()
        mock_extract.return_value = {"extracted_answer": "x", "confidence": 0.5}

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
            )
            with open(result["results_path"]) as f:
                line = json.loads(f.readline())
            assert line["problem"] == "legacy_q"
            assert line["correct_answer"] == "legacy_a"

    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_legacy_browsecomp_correct_answer_field(
        self, mock_registry, mock_qs, mock_extract
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        _setup_registry_with_fake_class(
            mock_registry,
            [{"id": "1", "problem": "bc_q", "correct_answer": "bc_correct"}],
        )
        mock_qs.return_value = _quick_summary_response()
        mock_extract.return_value = {"extracted_answer": "x", "confidence": 0.5}

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="browsecomp",
                output_dir=tmpdir,
                run_evaluation=False,
            )
            with open(result["results_path"]) as f:
                line = json.loads(f.readline())
            assert line["correct_answer"] == "bc_correct"

    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_legacy_browsecomp_fallback_to_answer_field(
        self, mock_registry, mock_qs, mock_extract
    ):
        """Cover line 166-167: fallback to 'answer' when 'correct_answer' missing."""
        from local_deep_research.benchmarks.runners import run_benchmark

        _setup_registry_with_fake_class(
            mock_registry,
            [{"id": "1", "problem": "bc_q", "answer": "bc_fallback_answer"}],
        )
        mock_qs.return_value = _quick_summary_response()
        mock_extract.return_value = {"extracted_answer": "x", "confidence": 0.5}

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="browsecomp",
                output_dir=tmpdir,
                run_evaluation=False,
            )
            with open(result["results_path"]) as f:
                line = json.loads(f.readline())
            assert line["correct_answer"] == "bc_fallback_answer"


# ---------------------------------------------------------------------------
# run_benchmark - successful example processing
# ---------------------------------------------------------------------------


class TestSuccessfulExampleProcessing:
    """Cover lines 188-245: quick_summary call, extract, write result."""

    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_result_contains_expected_fields(
        self, mock_registry, mock_qs, mock_extract
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        _setup_registry_with_fake_class(mock_registry, [_make_example()])
        mock_qs.return_value = _quick_summary_response("my summary", ["src1"])
        mock_extract.return_value = {
            "extracted_answer": "extracted",
            "confidence": 0.8,
        }

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
                search_config={
                    "iterations": 2,
                    "questions_per_iteration": 1,
                    "search_tool": "wiki",
                },
            )
            with open(result["results_path"]) as f:
                line = json.loads(f.readline())

        assert line["response"] == "my summary"
        assert line["extracted_answer"] == "extracted"
        assert line["confidence"] == 0.8
        assert line["sources"] == ["src1"]
        assert "processing_time" in line
        assert line["search_config"]["search_tool"] == "wiki"

    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_quick_summary_called_with_search_config(
        self, mock_registry, mock_qs, mock_extract
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        _setup_registry_with_fake_class(mock_registry, [_make_example()])
        mock_qs.return_value = _quick_summary_response()
        mock_extract.return_value = {"extracted_answer": "a", "confidence": 0.5}

        cfg = {
            "iterations": 7,
            "questions_per_iteration": 5,
            "search_tool": "google",
        }

        with tempfile.TemporaryDirectory() as tmpdir:
            run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
                search_config=cfg,
            )

        mock_qs.assert_called_once()
        _, kwargs = mock_qs.call_args
        assert kwargs["iterations"] == 7
        assert kwargs["questions_per_iteration"] == 5
        assert kwargs["search_tool"] == "google"


# ---------------------------------------------------------------------------
# run_benchmark - error handling during processing
# ---------------------------------------------------------------------------


class TestExampleProcessingError:
    """Cover lines 247-280: error during quick_summary."""

    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_error_result_written_on_exception(self, mock_registry, mock_qs):
        from local_deep_research.benchmarks.runners import run_benchmark

        _setup_registry_with_fake_class(mock_registry, [_make_example()])
        mock_qs.side_effect = RuntimeError("search failed")

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
            )
            with open(result["results_path"]) as f:
                line = json.loads(f.readline())

        assert "error" in line
        assert "search failed" in line["error"]
        assert line["id"] == "ex_1"

    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_error_with_progress_callback(self, mock_registry, mock_qs):
        from local_deep_research.benchmarks.runners import run_benchmark

        _setup_registry_with_fake_class(mock_registry, [_make_example()])
        mock_qs.side_effect = RuntimeError("boom")
        callback = Mock()

        with tempfile.TemporaryDirectory() as tmpdir:
            run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
                progress_callback=callback,
            )

        statuses = [
            c[0][2]["status"] for c in callback.call_args_list if len(c[0]) >= 3
        ]
        assert "error" in statuses
        assert "started" in statuses

    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_error_result_uses_fallback_id(self, mock_registry, mock_qs):
        """When example has no 'id', fallback to example_{i}."""
        from local_deep_research.benchmarks.runners import run_benchmark

        _setup_registry_with_fake_class(
            mock_registry, [{"problem": "q", "answer": "a"}]
        )
        mock_qs.side_effect = RuntimeError("fail")

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
            )
            with open(result["results_path"]) as f:
                line = json.loads(f.readline())

        assert line["id"] == "example_0"


# ---------------------------------------------------------------------------
# run_benchmark - progress callbacks throughout
# ---------------------------------------------------------------------------


class TestProgressCallbacks:
    """Cover all progress_callback invocations."""

    @patch(f"{MODULE}.generate_report")
    @patch(f"{MODULE}.calculate_metrics")
    @patch(f"{MODULE}.grade_results")
    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_all_progress_stages_with_evaluation(
        self,
        mock_registry,
        mock_qs,
        mock_extract,
        mock_grade,
        mock_calc,
        mock_report,
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        _setup_registry_with_fake_class(mock_registry, [_make_example()])
        mock_qs.return_value = _quick_summary_response()
        mock_extract.return_value = {"extracted_answer": "a", "confidence": 0.5}
        mock_grade.return_value = []
        mock_calc.return_value = {"accuracy": 0.5}
        mock_report.return_value = "/report.md"

        callback = Mock()

        with tempfile.TemporaryDirectory() as tmpdir:
            run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=True,
                progress_callback=callback,
            )

        statuses = [c[0][2]["status"] for c in callback.call_args_list]
        assert "started" in statuses
        assert "processing" in statuses
        assert "completed_example" in statuses
        assert "evaluating" in statuses
        assert "calculating_metrics" in statuses
        assert "generating_report" in statuses
        assert "complete" in statuses

    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_progress_truncates_long_question(
        self, mock_registry, mock_qs, mock_extract
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        _setup_registry_with_fake_class(
            mock_registry, [{"id": "1", "problem": "x" * 100, "answer": "a"}]
        )
        mock_qs.return_value = _quick_summary_response()
        mock_extract.return_value = {"extracted_answer": "a", "confidence": 0.5}

        callback = Mock()

        with tempfile.TemporaryDirectory() as tmpdir:
            run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
                progress_callback=callback,
            )

        processing_calls = [
            c
            for c in callback.call_args_list
            if len(c[0]) >= 3 and c[0][2].get("status") == "processing"
        ]
        assert len(processing_calls) == 1
        assert processing_calls[0][0][2]["question"].endswith("...")

    @patch(f"{MODULE}.DatasetRegistry")
    def test_no_eval_progress_callback(self, mock_registry):
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        callback = Mock()

        with tempfile.TemporaryDirectory() as tmpdir:
            run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
                progress_callback=callback,
            )

        statuses = [c[0][2]["status"] for c in callback.call_args_list]
        assert "complete_no_eval" in statuses

    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_progress_short_question_no_truncation(
        self, mock_registry, mock_qs, mock_extract
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        short_q = "Short question"
        _setup_registry_with_fake_class(
            mock_registry, [{"id": "1", "problem": short_q, "answer": "a"}]
        )
        mock_qs.return_value = _quick_summary_response()
        mock_extract.return_value = {"extracted_answer": "a", "confidence": 0.5}

        callback = Mock()

        with tempfile.TemporaryDirectory() as tmpdir:
            run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
                progress_callback=callback,
            )

        processing_calls = [
            c
            for c in callback.call_args_list
            if len(c[0]) >= 3 and c[0][2].get("status") == "processing"
        ]
        assert processing_calls[0][0][2]["question"] == short_q


# ---------------------------------------------------------------------------
# run_benchmark - evaluation paths
# ---------------------------------------------------------------------------


class TestEvaluationPaths:
    """Cover lines 285-425: evaluation with grading, metrics, report."""

    @patch(f"{MODULE}.generate_report")
    @patch(f"{MODULE}.calculate_metrics")
    @patch(f"{MODULE}.grade_results")
    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_automated_evaluation_success(
        self,
        mock_registry,
        mock_qs,
        mock_extract,
        mock_grade,
        mock_calc,
        mock_report,
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        _setup_registry_with_fake_class(mock_registry, [_make_example()])
        mock_qs.return_value = _quick_summary_response()
        mock_extract.return_value = {"extracted_answer": "a", "confidence": 0.9}
        mock_grade.return_value = [{"grade": "correct"}]
        mock_calc.return_value = {"accuracy": 1.0}
        mock_report.return_value = "/tmp/report.md"

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=True,
            )

        assert result["status"] == "complete"
        assert result["accuracy"] == 1.0
        assert "report_path" in result
        assert "evaluation_path" in result
        mock_grade.assert_called_once()
        mock_calc.assert_called_once()

    @patch(f"{MODULE}.generate_report")
    @patch(f"{MODULE}.calculate_metrics")
    @patch(f"{MODULE}.grade_results")
    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_automated_evaluation_with_eval_config(
        self,
        mock_registry,
        mock_qs,
        mock_extract,
        mock_grade,
        mock_calc,
        mock_report,
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        _setup_registry_with_fake_class(mock_registry, [_make_example()])
        mock_qs.return_value = _quick_summary_response()
        mock_extract.return_value = {"extracted_answer": "a", "confidence": 0.5}
        mock_grade.return_value = []
        mock_calc.return_value = {"accuracy": 0.0}
        mock_report.return_value = "/report.md"

        eval_config = {"model": "test-model"}

        with tempfile.TemporaryDirectory() as tmpdir:
            run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=True,
                evaluation_config=eval_config,
            )

        _, kwargs = mock_grade.call_args
        assert kwargs["evaluation_config"] == eval_config

    @patch(f"{MODULE}.generate_report")
    @patch(f"{MODULE}.calculate_metrics")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_human_evaluation_path(self, mock_registry, mock_calc, mock_report):
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        mock_calc.return_value = {"accuracy": 0.5}
        mock_report.return_value = "/report.md"

        with (
            patch(f"{MODULE}.grade_results") as mock_grade,
            patch(
                "local_deep_research.benchmarks.graders.human_evaluation"
            ) as mock_human,
        ):
            mock_human.return_value = [{"grade": "correct"}]

            with tempfile.TemporaryDirectory() as tmpdir:
                result = run_benchmark(
                    dataset_type="simpleqa",
                    output_dir=tmpdir,
                    run_evaluation=True,
                    human_evaluation=True,
                )

            mock_human.assert_called_once()
            mock_grade.assert_not_called()
            assert result["status"] == "complete"

    @patch(f"{MODULE}.generate_report")
    @patch(f"{MODULE}.calculate_metrics")
    @patch(f"{MODULE}.grade_results")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_evaluation_report_config_info(
        self, mock_registry, mock_grade, mock_calc, mock_report
    ):
        """Cover the config_info dict passed to generate_report."""
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        mock_grade.return_value = []
        mock_calc.return_value = {"accuracy": 0.0}
        mock_report.return_value = "/report.md"

        with tempfile.TemporaryDirectory() as tmpdir:
            run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=True,
                human_evaluation=False,
                dataset_path="/custom/path.json",
                search_config={
                    "iterations": 10,
                    "questions_per_iteration": 2,
                    "search_tool": "bing",
                },
            )

        _, kwargs = mock_report.call_args
        assert kwargs["dataset_name"] == "Simpleqa"
        config = kwargs["config_info"]
        assert config["Dataset"] == "/custom/path.json"
        assert config["Iterations"] == 10
        assert config["Search tool"] == "bing"
        assert config["Evaluation method"] == "Automated"

    @patch(f"{MODULE}.generate_report")
    @patch(f"{MODULE}.calculate_metrics")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_evaluation_report_human_method_label(
        self, mock_registry, mock_calc, mock_report
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        mock_calc.return_value = {"accuracy": 0.0}
        mock_report.return_value = "/report.md"

        with patch(
            "local_deep_research.benchmarks.graders.human_evaluation"
        ) as mock_human:
            mock_human.return_value = []
            with tempfile.TemporaryDirectory() as tmpdir:
                run_benchmark(
                    dataset_type="simpleqa",
                    output_dir=tmpdir,
                    run_evaluation=True,
                    human_evaluation=True,
                )

        _, kwargs = mock_report.call_args
        assert kwargs["config_info"]["Evaluation method"] == "Human"


# ---------------------------------------------------------------------------
# run_benchmark - automated evaluation failure
# ---------------------------------------------------------------------------


class TestEvaluationFailure:
    """Cover lines 320-367: grade_results raises, fallback decision."""

    @patch("builtins.input", return_value="n")
    @patch("builtins.print")
    @patch(f"{MODULE}.grade_results")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_eval_failure_skip_human_fallback(
        self, mock_registry, mock_grade, mock_print, mock_input
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        mock_grade.side_effect = RuntimeError("eval error")

        with patch(
            "local_deep_research.security.file_write_verifier.write_file_verified"
        ) as mock_write:
            with tempfile.TemporaryDirectory() as tmpdir:
                result = run_benchmark(
                    dataset_type="simpleqa",
                    output_dir=tmpdir,
                    run_evaluation=True,
                )

        assert result["status"] == "evaluation_error"
        assert "eval error" in result["evaluation_error"]
        mock_write.assert_called_once()

    @patch("builtins.input", return_value="y")
    @patch("builtins.print")
    @patch(f"{MODULE}.generate_report")
    @patch(f"{MODULE}.calculate_metrics")
    @patch(f"{MODULE}.grade_results")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_eval_failure_accept_human_fallback(
        self,
        mock_registry,
        mock_grade,
        mock_calc,
        mock_report,
        mock_print,
        mock_input,
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        mock_grade.side_effect = RuntimeError("eval error")
        mock_calc.return_value = {"accuracy": 0.5}
        mock_report.return_value = "/report.md"

        with patch(
            "local_deep_research.benchmarks.graders.human_evaluation"
        ) as mock_human:
            mock_human.return_value = []
            with tempfile.TemporaryDirectory() as tmpdir:
                result = run_benchmark(
                    dataset_type="simpleqa",
                    output_dir=tmpdir,
                    run_evaluation=True,
                )

        mock_human.assert_called_once()
        assert result["status"] == "complete"

    @patch("builtins.input", return_value="n")
    @patch("builtins.print")
    @patch(f"{MODULE}.grade_results")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_eval_failure_with_progress_callback(
        self, mock_registry, mock_grade, mock_print, mock_input
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        mock_grade.side_effect = RuntimeError("eval error")
        callback = Mock()

        with patch(
            "local_deep_research.security.file_write_verifier.write_file_verified"
        ):
            with tempfile.TemporaryDirectory() as tmpdir:
                run_benchmark(
                    dataset_type="simpleqa",
                    output_dir=tmpdir,
                    run_evaluation=True,
                    progress_callback=callback,
                )

        statuses = [
            c[0][2]["status"] for c in callback.call_args_list if len(c[0]) >= 3
        ]
        assert "evaluation_fallback" in statuses


# ---------------------------------------------------------------------------
# run_benchmark - grade_results progress_callback lambda
# ---------------------------------------------------------------------------


class TestGradeProgressCallback:
    """Cover lines 310-318: the lambda passed as progress_callback to grade_results."""

    @patch(f"{MODULE}.generate_report")
    @patch(f"{MODULE}.calculate_metrics")
    @patch(f"{MODULE}.grade_results")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_grade_results_progress_lambda_invoked(
        self, mock_registry, mock_grade, mock_calc, mock_report
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        def fake_grade(**kwargs):
            cb = kwargs.get("progress_callback")
            if cb:
                cb(0, 10, {"detail": "grading"})
                cb(5, 10, {"detail": "grading"})
            return []

        mock_grade.side_effect = fake_grade
        mock_calc.return_value = {"accuracy": 0.0}
        mock_report.return_value = "/report.md"

        outer_callback = Mock()

        with tempfile.TemporaryDirectory() as tmpdir:
            run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=True,
                progress_callback=outer_callback,
            )

        evaluating_calls = [
            c
            for c in outer_callback.call_args_list
            if len(c[0]) >= 3 and c[0][2].get("status") == "evaluating"
        ]
        assert len(evaluating_calls) >= 2

    @patch(f"{MODULE}.generate_report")
    @patch(f"{MODULE}.calculate_metrics")
    @patch(f"{MODULE}.grade_results")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_grade_results_progress_lambda_without_outer_callback(
        self, mock_registry, mock_grade, mock_calc, mock_report
    ):
        """When no outer callback, the lambda should still work (returns None)."""
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        captured_cb = {}

        def fake_grade(**kwargs):
            captured_cb["cb"] = kwargs.get("progress_callback")
            return []

        mock_grade.side_effect = fake_grade
        mock_calc.return_value = {"accuracy": 0.0}
        mock_report.return_value = "/report.md"

        with tempfile.TemporaryDirectory() as tmpdir:
            run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=True,
                progress_callback=None,
            )

        cb = captured_cb["cb"]
        assert cb is not None
        result = cb(0, 10, {"x": 1})
        assert result is None


# ---------------------------------------------------------------------------
# run_benchmark - no evaluation path
# ---------------------------------------------------------------------------


class TestNoEvaluation:
    """Cover lines 427-441: run_evaluation=False."""

    @patch(f"{MODULE}.DatasetRegistry")
    def test_no_eval_returns_correct_status(self, mock_registry):
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
            )

        assert result["status"] == "complete_no_eval"
        assert "evaluation_path" not in result
        assert "metrics" not in result


# ---------------------------------------------------------------------------
# run_benchmark - file cleanup of existing output files
# ---------------------------------------------------------------------------


class TestOutputFileCleanup:
    """Cover lines 122-125: unlinking existing output files."""

    @patch(f"{MODULE}.DatasetRegistry")
    @patch(f"{MODULE}.time")
    def test_existing_files_are_removed(
        self, mock_time, mock_registry, monkeypatch
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_time.strftime.return_value = "20260101_000000"
        mock_time.time.return_value = 0

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        with tempfile.TemporaryDirectory() as tmpdir:
            # run_benchmark output_dir is now confined below the LDR data
            # directory (never its root); point the data dir at tmpdir so
            # cleanup operates on a directory inside it.
            monkeypatch.setenv("LDR_DATA_DIR", tmpdir)
            out = Path(tmpdir) / "results"
            out.mkdir()
            for suffix in ["_results.jsonl", "_evaluation.jsonl", "_report.md"]:
                p = out / f"simpleqa_20260101_000000{suffix}"
                p.write_text("old content")

            run_benchmark(
                dataset_type="simpleqa",
                output_dir=str(out),
                run_evaluation=False,
            )

            results_file = out / "simpleqa_20260101_000000_results.jsonl"
            assert not results_file.exists() or results_file.read_text() == ""


# ---------------------------------------------------------------------------
# Convenience wrapper functions
# ---------------------------------------------------------------------------


class TestConvenienceWrappers:
    """Cover run_simpleqa_benchmark, run_browsecomp_benchmark, run_xbench_deepsearch_benchmark."""

    @patch(f"{MODULE}.run_benchmark")
    def test_run_simpleqa_benchmark(self, mock_rb):
        from local_deep_research.benchmarks.runners import (
            run_simpleqa_benchmark,
        )

        mock_rb.return_value = {"status": "ok"}
        result = run_simpleqa_benchmark(num_examples=50, output_dir="/tmp/test")

        mock_rb.assert_called_once_with(
            dataset_type="simpleqa", num_examples=50, output_dir="/tmp/test"
        )
        assert result == {"status": "ok"}

    @patch(f"{MODULE}.run_benchmark")
    def test_run_browsecomp_benchmark(self, mock_rb):
        from local_deep_research.benchmarks.runners import (
            run_browsecomp_benchmark,
        )

        mock_rb.return_value = {"status": "ok"}
        result = run_browsecomp_benchmark(num_examples=25)

        mock_rb.assert_called_once_with(
            dataset_type="browsecomp", num_examples=25
        )
        assert result == {"status": "ok"}

    @patch(f"{MODULE}.run_benchmark")
    def test_run_xbench_deepsearch_benchmark(self, mock_rb):
        from local_deep_research.benchmarks.runners import (
            run_xbench_deepsearch_benchmark,
        )

        mock_rb.return_value = {"status": "ok"}
        result = run_xbench_deepsearch_benchmark(num_examples=10, seed=99)

        mock_rb.assert_called_once_with(
            dataset_type="xbench_deepsearch", num_examples=10, seed=99
        )
        assert result == {"status": "ok"}


# ---------------------------------------------------------------------------
# run_benchmark - multiple examples
# ---------------------------------------------------------------------------


class TestMultipleExamples:
    """Cover loop iteration with multiple examples."""

    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_multiple_examples_all_written(
        self, mock_registry, mock_qs, mock_extract
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        examples = [
            _make_example(f"Q{i}", f"A{i}", f"id_{i}") for i in range(3)
        ]
        _setup_registry_with_fake_class(mock_registry, examples)
        mock_qs.return_value = _quick_summary_response()
        mock_extract.return_value = {"extracted_answer": "a", "confidence": 0.5}

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
            )
            assert result["total_examples"] == 3
            with open(result["results_path"]) as f:
                lines = f.readlines()
            assert len(lines) == 3

    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_mix_of_success_and_error(
        self, mock_registry, mock_qs, mock_extract
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        examples = [
            _make_example(f"Q{i}", f"A{i}", f"id_{i}") for i in range(3)
        ]
        _setup_registry_with_fake_class(mock_registry, examples)

        mock_qs.side_effect = [
            _quick_summary_response(),
            RuntimeError("fail"),
            _quick_summary_response(),
        ]
        mock_extract.return_value = {"extracted_answer": "a", "confidence": 0.5}

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=False,
            )
            with open(result["results_path"]) as f:
                lines = [json.loads(line) for line in f.readlines()]

        assert len(lines) == 3
        assert "error" not in lines[0]
        assert "error" in lines[1]
        assert "error" not in lines[2]


# ---------------------------------------------------------------------------
# run_benchmark - DEFAULT_DATASET_URLS fallback in report config
# ---------------------------------------------------------------------------


class TestDefaultDatasetUrlInReport:
    """Cover line 391: when dataset_path is None, use DEFAULT_DATASET_URLS."""

    @patch(f"{MODULE}.generate_report")
    @patch(f"{MODULE}.calculate_metrics")
    @patch(f"{MODULE}.grade_results")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_report_uses_default_url_when_no_path(
        self, mock_registry, mock_grade, mock_calc, mock_report
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        mock_grade.return_value = []
        mock_calc.return_value = {"accuracy": 0.0}
        mock_report.return_value = "/report.md"

        with tempfile.TemporaryDirectory() as tmpdir:
            run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=True,
                dataset_path=None,
            )

        _, kwargs = mock_report.call_args
        assert kwargs["config_info"]["Dataset"] is not None


# ---------------------------------------------------------------------------
# run_benchmark - metrics accuracy key missing
# ---------------------------------------------------------------------------


class TestMetricsAccuracyFallback:
    """Cover line 424: metrics.get('accuracy', 0) when key missing."""

    @patch(f"{MODULE}.generate_report")
    @patch(f"{MODULE}.calculate_metrics")
    @patch(f"{MODULE}.grade_results")
    @patch(f"{MODULE}.DatasetRegistry")
    def test_accuracy_defaults_to_zero(
        self, mock_registry, mock_grade, mock_calc, mock_report
    ):
        from local_deep_research.benchmarks.runners import run_benchmark

        mock_dataset = Mock()
        mock_dataset.load.return_value = []
        mock_registry.create_dataset.return_value = mock_dataset

        mock_grade.return_value = []
        mock_calc.return_value = {}  # No accuracy key
        mock_report.return_value = "/report.md"

        with tempfile.TemporaryDirectory() as tmpdir:
            result = run_benchmark(
                dataset_type="simpleqa",
                output_dir=tmpdir,
                run_evaluation=True,
            )

        assert result["accuracy"] == 0