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.

68.6 KB
Raw Blame
"""Comprehensive coverage tests for benchmarks/web_api/benchmark_service.py.

Focuses on uncovered code paths: complex branching in _run_benchmark_thread,
_process_benchmark_task, _sync_results_to_database, get_benchmark_status,
sync_pending_results, cancel_benchmark, update_benchmark_status, and more.
"""

import json
from datetime import UTC, datetime, timedelta
from unittest.mock import MagicMock, Mock, patch

import pytest


# ---------------------------------------------------------------------------
# Helpers to import from the module under test
# ---------------------------------------------------------------------------
MODULE = "local_deep_research.benchmarks.web_api.benchmark_service"
SETTINGS_CTX_MODULE = "local_deep_research.config.thread_settings"


def _import_service():
    """Import BenchmarkService with SocketIOService mocked."""
    with patch(f"{MODULE}.SocketIOService"):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkService,
        )

        return BenchmarkService


def _make_service(socket=None):
    """Create a BenchmarkService with a mock socket service."""
    cls = _import_service()
    svc = cls(socket_service=socket or MagicMock())
    return svc


# ============================================================
# BenchmarkQueueTracker – deeper coverage
# ============================================================


class TestQueueTrackerCleanup:
    """Cover cleanup_completed_tasks edge cases."""

    def test_cleanup_removes_old_completed(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkQueueTracker,
            BenchmarkTaskStatus,
        )

        tracker = BenchmarkQueueTracker()
        tracker.add_task("old1", "user")
        tracker.update_task_status("old1", BenchmarkTaskStatus.COMPLETED)
        # Manually set updated_at to the past
        tracker.tasks["old1"]["updated_at"] = datetime.now(UTC) - timedelta(
            seconds=7200
        )

        tracker.cleanup_completed_tasks(max_age_seconds=3600)
        assert "old1" not in tracker.tasks

    def test_cleanup_keeps_recent_completed(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkQueueTracker,
            BenchmarkTaskStatus,
        )

        tracker = BenchmarkQueueTracker()
        tracker.add_task("new1", "user")
        tracker.update_task_status("new1", BenchmarkTaskStatus.COMPLETED)

        tracker.cleanup_completed_tasks(max_age_seconds=3600)
        assert "new1" in tracker.tasks

    def test_cleanup_removes_old_failed(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkQueueTracker,
            BenchmarkTaskStatus,
        )

        tracker = BenchmarkQueueTracker()
        tracker.add_task("fail1", "user")
        tracker.update_task_status("fail1", BenchmarkTaskStatus.FAILED)
        tracker.tasks["fail1"]["updated_at"] = datetime.now(UTC) - timedelta(
            seconds=7200
        )

        tracker.cleanup_completed_tasks(max_age_seconds=3600)
        assert "fail1" not in tracker.tasks

    def test_cleanup_removes_old_cancelled(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkQueueTracker,
            BenchmarkTaskStatus,
        )

        tracker = BenchmarkQueueTracker()
        tracker.add_task("canc1", "user")
        tracker.update_task_status("canc1", BenchmarkTaskStatus.CANCELLED)
        tracker.tasks["canc1"]["updated_at"] = datetime.now(UTC) - timedelta(
            seconds=7200
        )

        tracker.cleanup_completed_tasks(max_age_seconds=3600)
        assert "canc1" not in tracker.tasks

    def test_cleanup_keeps_processing_tasks(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkQueueTracker,
            BenchmarkTaskStatus,
        )

        tracker = BenchmarkQueueTracker()
        tracker.add_task("proc1", "user")
        tracker.update_task_status("proc1", BenchmarkTaskStatus.PROCESSING)
        tracker.tasks["proc1"]["updated_at"] = datetime.now(UTC) - timedelta(
            seconds=7200
        )

        tracker.cleanup_completed_tasks(max_age_seconds=3600)
        assert "proc1" in tracker.tasks

    def test_cleanup_uses_created_at_when_no_updated_at(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkQueueTracker,
            BenchmarkTaskStatus,
        )

        tracker = BenchmarkQueueTracker()
        tracker.add_task("t1", "user")
        # Set status directly without going through update (no updated_at)
        tracker.tasks["t1"]["status"] = BenchmarkTaskStatus.COMPLETED.value
        tracker.tasks["t1"]["created_at"] = datetime.now(UTC) - timedelta(
            seconds=7200
        )
        if "updated_at" in tracker.tasks["t1"]:
            del tracker.tasks["t1"]["updated_at"]

        tracker.cleanup_completed_tasks(max_age_seconds=3600)
        assert "t1" not in tracker.tasks

    def test_update_nonexistent_task_logs_warning(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkQueueTracker,
            BenchmarkTaskStatus,
        )

        tracker = BenchmarkQueueTracker()
        # Should not raise; logs a warning
        tracker.update_task_status("nonexistent", BenchmarkTaskStatus.COMPLETED)

    def test_remove_task(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkQueueTracker,
        )

        tracker = BenchmarkQueueTracker()
        tracker.add_task("rm1", "user")
        tracker.remove_task("rm1")
        assert tracker.get_task_status("rm1") is None

    def test_remove_nonexistent_task(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkQueueTracker,
        )

        tracker = BenchmarkQueueTracker()
        # Should not raise
        tracker.remove_task("doesnotexist")


# ============================================================
# BenchmarkService – initialization
# ============================================================


class TestBenchmarkServiceInit:
    def test_init_with_provided_socket(self):
        mock_socket = MagicMock()
        svc = _make_service(socket=mock_socket)
        assert svc.socket_service is mock_socket

    def test_init_default_socket_emit_to_room_delegates(self):
        """With no socket provided, __init__ constructs a real
        ``SocketIOService`` adapter whose ``emit_to_room`` forwards to the
        FastAPI-native ``socketio_asgi`` emit path.

        Replaces the old ``test_init_fallback_socket``: pre-FastAPI the
        constructor wrapped ``SocketIOService()`` in a try/except that fell
        back to a no-op shim when the Flask app was unavailable. The
        migration (commit dec8e1452) dropped that fallback because the new
        ``SocketIOService`` is a default-constructible adapter that cannot
        fail at construction — so the obsolete fallback assertion is replaced
        with the live delegation contract.
        """
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkService,
        )

        svc = BenchmarkService()
        with patch(
            "local_deep_research.web.services.socketio_asgi.emit_socket_event",
            return_value=True,
        ) as emit:
            result = svc.socket_service.emit_to_room("event", {}, room="room")

        assert result is True
        emit.assert_called_once_with("event", {}, room="room")


# ============================================================
# generate_config_hash / generate_query_hash
# ============================================================


class TestHashGeneration:
    def test_config_hash_deterministic(self):
        svc = _make_service()
        cfg = {"iterations": 5, "search_tool": "searxng", "model_name": "gpt-4"}
        h1 = svc.generate_config_hash(cfg)
        h2 = svc.generate_config_hash(cfg)
        assert h1 == h2
        assert len(h1) == 8

    def test_config_hash_ignores_none_values(self):
        svc = _make_service()
        cfg1 = {"iterations": 5, "search_tool": None}
        cfg2 = {"iterations": 5}
        assert svc.generate_config_hash(cfg1) == svc.generate_config_hash(cfg2)

    def test_config_hash_differs_for_different_configs(self):
        svc = _make_service()
        h1 = svc.generate_config_hash({"iterations": 5})
        h2 = svc.generate_config_hash({"iterations": 10})
        assert h1 != h2

    def test_query_hash_deterministic(self):
        svc = _make_service()
        h1 = svc.generate_query_hash("What is AI?", "simpleqa", "ex1")
        h2 = svc.generate_query_hash("What is AI?", "simpleqa", "ex1")
        assert h1 == h2

    def test_query_hash_strips_whitespace(self):
        svc = _make_service()
        h1 = svc.generate_query_hash("  What is AI?  ", "simpleqa", "ex1")
        h2 = svc.generate_query_hash("What is AI?", "simpleqa", "ex1")
        assert h1 == h2

    def test_query_hash_case_insensitive_dataset(self):
        svc = _make_service()
        h1 = svc.generate_query_hash("What is AI?", "SimpleQA", "ex1")
        h2 = svc.generate_query_hash("What is AI?", "simpleqa", "ex1")
        assert h1 == h2

    def test_query_hash_keys_on_example_id(self):
        # #4860: distinct examples, identical question -> distinct hashes.
        svc = _make_service()
        h1 = svc.generate_query_hash("What is AI?", "simpleqa", "ex1")
        h2 = svc.generate_query_hash("What is AI?", "simpleqa", "ex2")
        assert h1 != h2


# ============================================================
# create_benchmark_run
# ============================================================


class TestCreateBenchmarkRun:
    @patch(f"{MODULE}.BenchmarkRun")
    def test_create_benchmark_run_success(self, mock_run_cls):
        svc = _make_service()
        mock_session = MagicMock()
        mock_run_instance = MagicMock()
        mock_run_instance.id = 42
        mock_run_cls.return_value = mock_run_instance

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            result = svc.create_benchmark_run(
                run_name="test",
                search_config={"iterations": 5},
                evaluation_config={},
                datasets_config={"simpleqa": {"count": 10}},
                username="user1",
            )

        assert result == 42
        mock_session.add.assert_called_once_with(mock_run_instance)
        mock_session.commit.assert_called_once()

    @patch(f"{MODULE}.BenchmarkRun")
    def test_create_benchmark_run_db_error(self, mock_run_cls):
        svc = _make_service()
        mock_session = MagicMock()
        mock_session.commit.side_effect = RuntimeError("db error")
        mock_run_cls.return_value = MagicMock()

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            with pytest.raises(RuntimeError, match="db error"):
                svc.create_benchmark_run(
                    run_name="test",
                    search_config={},
                    evaluation_config={},
                    datasets_config={},
                )

        mock_session.rollback.assert_called_once()


# ============================================================
# _create_task_queue
# ============================================================


class TestCreateTaskQueue:
    @patch(f"{MODULE}.load_dataset")
    def test_creates_tasks_from_dataset(self, mock_load):
        svc = _make_service()
        mock_load.return_value = [
            {"id": "ex1", "problem": "Q1?", "answer": "A1"},
            {"id": "ex2", "problem": "Q2?", "answer": "A2"},
        ]

        tasks = svc._create_task_queue(
            {"simpleqa": {"count": 2}},
            benchmark_run_id=1,
        )

        assert len(tasks) == 2
        assert tasks[0]["question"] == "Q1?"
        assert tasks[0]["dataset_type"] == "simpleqa"

    @patch(f"{MODULE}.load_dataset")
    def test_passes_seed_from_dataset_config(self, mock_load):
        """A "seed" in the dataset config drives reproducible sampling."""
        svc = _make_service()
        mock_load.return_value = [
            {"id": "ex1", "problem": "Q1?", "answer": "A1"},
        ]

        svc._create_task_queue(
            {"simpleqa": {"count": 1, "seed": 7}},
            benchmark_run_id=1,
        )

        mock_load.assert_called_once_with(
            dataset_type="simpleqa", num_examples=1, seed=7
        )

    @patch(f"{MODULE}.load_dataset")
    def test_no_seed_means_random_sampling(self, mock_load):
        """Without a configured seed the sample stays random (seed=None)."""
        svc = _make_service()
        mock_load.return_value = [
            {"id": "ex1", "problem": "Q1?", "answer": "A1"},
        ]

        svc._create_task_queue(
            {"simpleqa": {"count": 1}},
            benchmark_run_id=1,
        )

        mock_load.assert_called_once_with(
            dataset_type="simpleqa", num_examples=1, seed=None
        )

    @patch(f"{MODULE}.load_dataset")
    def test_queues_every_sampled_question(self, mock_load):
        """Regression test for #4498: every sampled question becomes a task.

        The removed cross-run reuse feature silently skipped questions that
        had results in previous compatible runs, which broke the
        completed/total accounting (#4451). A run must always process its
        full sample, even if identical questions were answered before."""
        svc = _make_service()
        examples = [
            {"id": f"ex{i}", "problem": f"Q{i}?", "answer": f"A{i}"}
            for i in range(5)
        ]
        mock_load.return_value = examples

        tasks = svc._create_task_queue(
            {"simpleqa": {"count": 5}},
            benchmark_run_id=1,
        )

        assert len(tasks) == len(examples)
        assert [t["question"] for t in tasks] == [
            e["problem"] for e in examples
        ]
        assert [t["task_index"] for t in tasks] == list(range(5))

    @patch(f"{MODULE}.load_dataset")
    def test_skips_zero_count_datasets(self, mock_load):
        svc = _make_service()
        tasks = svc._create_task_queue(
            {"simpleqa": {"count": 0}},
            benchmark_run_id=1,
        )
        assert len(tasks) == 0
        mock_load.assert_not_called()

    @patch(f"{MODULE}.load_dataset")
    def test_browsecomp_dataset_type(self, mock_load):
        svc = _make_service()
        mock_load.return_value = [
            {"id": "b1", "problem": "Browse Q?", "answer": "Browse A"},
        ]

        tasks = svc._create_task_queue(
            {"browsecomp": {"count": 1}},
            benchmark_run_id=1,
        )

        assert len(tasks) == 1
        assert tasks[0]["dataset_type"] == "browsecomp"

    @patch(f"{MODULE}.load_dataset")
    def test_example_without_id_gets_default(self, mock_load):
        svc = _make_service()
        mock_load.return_value = [
            {"problem": "Q?", "answer": "A"},  # no id
        ]

        tasks = svc._create_task_queue(
            {"simpleqa": {"count": 1}},
            benchmark_run_id=1,
        )

        assert tasks[0]["example_id"] == "example_0"


# ============================================================
# _process_benchmark_task
# ============================================================


class TestProcessBenchmarkTask:
    def _make_task(self, **overrides):
        task = {
            "benchmark_run_id": 1,
            "example_id": "ex1",
            "dataset_type": "simpleqa",
            "question": "What is 2+2?",
            "correct_answer": "4",
            "query_hash": "abc123",
            "task_index": 0,
            "username": "user1",
            "user_password": None,
        }
        task.update(overrides)
        return task

    @pytest.mark.parametrize(
        ("findings", "is_failure"),
        [
            ([{"phase": "Error", "content": "Error: synthesis detail"}], True),
            (
                [
                    {
                        "phase": "Final synthesis",
                        "content": "Error: synthesis detail",
                    }
                ],
                False,
            ),
            (
                [
                    {"phase": "Error", "content": "earlier step"},
                    {
                        "phase": "Final synthesis",
                        "content": "Error: synthesis detail",
                    },
                ],
                False,
            ),
            ([{"phase": "Error", "content": "different content"}], False),
            (None, False),
        ],
    )
    @patch(f"{MODULE}.grade_single_result")
    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.format_query")
    def test_terminal_research_finding_controls_result_state(
        self,
        mock_format,
        mock_summary,
        mock_extract,
        mock_grade,
        findings,
        is_failure,
    ):
        svc = _make_service()
        mock_format.return_value = "q"
        summary = "Error: synthesis detail"
        mock_summary.return_value = {
            "summary": summary,
            "findings": findings,
            "sources": [],
        }
        mock_extract.return_value = {"extracted_answer": "answer"}
        mock_grade.return_value = {"is_correct": True}
        context = MagicMock()
        context.snapshot = {}
        with patch(
            f"{SETTINGS_CTX_MODULE}.get_settings_context", return_value=context
        ):
            result = svc._process_benchmark_task(self._make_task(), {}, {})
        if is_failure:
            assert result["research_error"] == (
                "Research failed. Check the server logs for details."
            )
            assert "response" not in result
            assert "extracted_answer" not in result
            mock_extract.assert_not_called()
            mock_grade.assert_not_called()
        else:
            assert result["response"] == summary
            mock_extract.assert_called_once()
            mock_grade.assert_called_once()

    @patch(f"{MODULE}.grade_single_result")
    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.format_query")
    def test_successful_processing_with_grading(
        self, mock_format, mock_summary, mock_extract, mock_grade
    ):
        svc = _make_service()
        mock_format.return_value = "formatted query"
        mock_summary.return_value = {
            "summary": "The answer is 4",
            "sources": [{"url": "http://example.com"}],
        }
        mock_extract.return_value = {
            "extracted_answer": "4",
            "confidence": "95",
        }
        mock_grade.return_value = {
            "is_correct": True,
            "graded_confidence": "98",
            "grader_response": "Correct answer",
        }

        mock_settings = MagicMock()
        mock_settings.snapshot = {}

        with patch(
            f"{SETTINGS_CTX_MODULE}.get_settings_context",
            return_value=mock_settings,
        ):
            result = svc._process_benchmark_task(
                self._make_task(),
                {"iterations": 5},
                {},
            )

        assert result["is_correct"] is True
        assert result["extracted_answer"] == "4"
        assert result["response"] == "The answer is 4"

    @patch(f"{MODULE}.grade_single_result")
    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.format_query")
    def test_grading_error_in_result(
        self, mock_format, mock_summary, mock_extract, mock_grade
    ):
        svc = _make_service()
        mock_format.return_value = "q"
        mock_summary.return_value = {"summary": "resp", "sources": []}
        mock_extract.return_value = {
            "extracted_answer": "ans",
            "confidence": "50",
        }
        mock_grade.return_value = {"grading_error": "model unavailable"}

        mock_settings = MagicMock()
        mock_settings.snapshot = {}

        with patch(
            f"{SETTINGS_CTX_MODULE}.get_settings_context",
            return_value=mock_settings,
        ):
            result = svc._process_benchmark_task(self._make_task(), {}, {})

        assert result["is_correct"] is None
        assert "model unavailable" in result["grader_response"]

    @patch(f"{MODULE}.grade_single_result")
    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.format_query")
    def test_grading_returns_none(
        self, mock_format, mock_summary, mock_extract, mock_grade
    ):
        svc = _make_service()
        mock_format.return_value = "q"
        mock_summary.return_value = {"summary": "resp", "sources": []}
        mock_extract.return_value = {"extracted_answer": "ans"}
        mock_grade.return_value = None

        mock_settings = MagicMock()
        mock_settings.snapshot = {}

        with patch(
            f"{SETTINGS_CTX_MODULE}.get_settings_context",
            return_value=mock_settings,
        ):
            result = svc._process_benchmark_task(self._make_task(), {}, {})

        assert result["is_correct"] is None
        assert "No evaluation results returned" in result["grader_response"]

    @patch(f"{MODULE}.grade_single_result")
    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.format_query")
    def test_grading_exception(
        self, mock_format, mock_summary, mock_extract, mock_grade
    ):
        svc = _make_service()
        mock_format.return_value = "q"
        mock_summary.return_value = {"summary": "resp", "sources": []}
        mock_extract.return_value = {"extracted_answer": "ans"}
        mock_grade.side_effect = ValueError("grade fail")

        mock_settings = MagicMock()
        mock_settings.snapshot = {}

        with patch(
            f"{SETTINGS_CTX_MODULE}.get_settings_context",
            return_value=mock_settings,
        ):
            result = svc._process_benchmark_task(self._make_task(), {}, {})

        assert result["is_correct"] is None
        assert "grade fail" in result["grader_response"]

    @patch(f"{MODULE}.format_query")
    def test_research_error(self, mock_format):
        svc = _make_service()
        mock_format.side_effect = RuntimeError("research crashed")

        mock_settings = MagicMock()
        mock_settings.snapshot = {}

        with patch(
            f"{SETTINGS_CTX_MODULE}.get_settings_context",
            return_value=mock_settings,
        ):
            result = svc._process_benchmark_task(self._make_task(), {}, {})

        assert "research_error" in result
        assert "research crashed" in result["research_error"]

    @patch(f"{MODULE}.grade_single_result")
    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.format_query")
    def test_extract_returns_string(
        self, mock_format, mock_summary, mock_extract, mock_grade
    ):
        """When extract_answer_from_response returns a string instead of dict."""
        svc = _make_service()
        mock_format.return_value = "q"
        mock_summary.return_value = {"summary": "resp", "sources": []}
        mock_extract.return_value = "plain string answer"
        mock_grade.return_value = {
            "is_correct": False,
            "graded_confidence": "10",
            "grader_response": "Wrong",
        }

        mock_settings = MagicMock()
        mock_settings.snapshot = {}

        with patch(
            f"{SETTINGS_CTX_MODULE}.get_settings_context",
            return_value=mock_settings,
        ):
            result = svc._process_benchmark_task(self._make_task(), {}, {})

        assert result["extracted_answer"] == "plain string answer"
        assert result["confidence"] == "100"

    @patch(f"{MODULE}.grade_single_result")
    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.format_query")
    def test_sources_from_all_links(
        self, mock_format, mock_summary, mock_extract, mock_grade
    ):
        """When sources is empty but all_links_of_system is present."""
        svc = _make_service()
        mock_format.return_value = "q"
        mock_summary.return_value = {
            "summary": "resp",
            "sources": [],
            "all_links_of_system": ["http://link1.com"],
        }
        mock_extract.return_value = {"extracted_answer": "ans"}
        mock_grade.return_value = {
            "is_correct": True,
            "graded_confidence": "90",
            "grader_response": "ok",
        }

        mock_settings = MagicMock()
        mock_settings.snapshot = {}

        with patch(
            f"{SETTINGS_CTX_MODULE}.get_settings_context",
            return_value=mock_settings,
        ):
            result = svc._process_benchmark_task(self._make_task(), {}, {})

        sources = json.loads(result["sources"])
        assert "http://link1.com" in sources


# ============================================================
# _send_progress_update
# ============================================================


class TestSendProgressUpdate:
    def test_sends_progress_via_socket(self):
        mock_socket = MagicMock()
        svc = _make_service(socket=mock_socket)

        svc._send_progress_update(1, 5, 10, "user1")

        mock_socket.emit_to_subscribers.assert_called_once()
        call_args = mock_socket.emit_to_subscribers.call_args
        assert call_args[0][0] == "research_progress"
        assert call_args[0][2]["progress"] == 50.0

    def test_sends_zero_progress_when_total_zero(self):
        mock_socket = MagicMock()
        svc = _make_service(socket=mock_socket)

        svc._send_progress_update(1, 0, 0, "user1")

        call_args = mock_socket.emit_to_subscribers.call_args
        assert call_args[0][2]["progress"] == 0

    def test_exception_in_send_is_caught(self):
        mock_socket = MagicMock()
        mock_socket.emit_to_subscribers.side_effect = RuntimeError(
            "socket error"
        )
        svc = _make_service(socket=mock_socket)

        # Should not raise
        svc._send_progress_update(1, 5, 10, "user1")


# ============================================================
# sync_pending_results
# ============================================================


class TestSyncPendingResults:
    def test_returns_zero_when_no_active_run(self):
        svc = _make_service()
        assert svc.sync_pending_results(999, "user1") == 0

    def test_returns_zero_when_entry_deleted_during_toctou_window(self):
        """The worker thread's ``del self.active_runs[run_key]`` can land between
        the membership check and the subscript in ``sync_pending_results``.
        The read must not raise KeyError (which surfaces as a spurious 500 on
        /api/results); a run that vanished has nothing pending. See #4859.
        """

        class _VanishingRuns(dict):
            # Present at the ``in`` check, gone at the read: models the
            # concurrent delete firing inside the TOCTOU window.
            def __contains__(self, key):
                return True

        svc = _make_service()
        # dict.get on an empty mapping returns None while __contains__ lies True.
        svc.active_runs = _VanishingRuns()

        assert svc.sync_pending_results(123, "user1") == 0

    def test_saves_new_results(self):
        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "data": {"username": "user1", "user_password": None},
            "results": [
                {
                    "example_id": "ex1",
                    "query_hash": "h1",
                    "dataset_type": "simpleqa",
                    "question": "Q?",
                    "correct_answer": "A",
                    "task_index": 0,
                },
            ],
            "result_persistence_failed": True,
        }

        mock_session = MagicMock()
        # No rows persisted for this run yet
        mock_session.query.return_value.filter.return_value.all.return_value = []

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            count = svc.sync_pending_results(1, "user1")

        assert count == 1
        mock_session.add.assert_called_once()
        mock_session.commit.assert_called_once()
        assert svc.get_result_persistence_error(1, "user1") is None

    def test_skips_already_saved_indices(self):
        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "data": {"username": "user1", "user_password": None},
            "results": [
                {
                    "example_id": "ex1",
                    "query_hash": "h1",
                    "dataset_type": "simpleqa",
                    "question": "Q?",
                    "correct_answer": "A",
                    "task_index": 0,
                },
            ],
            "saved_indices": {0},
        }

        mock_session = MagicMock()
        # No rows persisted for this run yet; idx 0 is skipped via saved_indices.
        mock_session.query.return_value.filter.return_value.all.return_value = []

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            count = svc.sync_pending_results(1, "user1")

        assert count == 0

    def test_skips_existing_db_result(self):
        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "data": {"username": "user1", "user_password": None},
            "results": [
                {
                    "example_id": "ex1",
                    "query_hash": "h1",
                    "dataset_type": "simpleqa",
                    "question": "Q?",
                    "correct_answer": "A",
                    "task_index": 0,
                },
            ],
        }

        mock_session = MagicMock()
        # query_hash "h1" already persisted for this run
        mock_session.query.return_value.filter.return_value.all.return_value = [
            ("h1",)
        ]

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            count = svc.sync_pending_results(1, "user1")

        assert count == 0
        mock_session.add.assert_not_called()

    def test_handles_db_error(self):
        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "data": {"username": "user1", "user_password": None},
            "results": [
                {
                    "example_id": "ex1",
                    "query_hash": "h1",
                    "dataset_type": "simpleqa",
                    "question": "Q?",
                    "correct_answer": "A",
                    "task_index": 0,
                },
            ],
        }

        mock_session = MagicMock()
        mock_session.query.side_effect = RuntimeError("db error")

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            count = svc.sync_pending_results(1, "user1")

        assert count == 0
        error = svc.get_result_persistence_error(1, "user1")
        assert error is not None
        assert error["code"] == "database_write_failed"
        assert "db error" not in error["message"]

    def test_opens_the_db_for_the_caller_supplied_user(self):
        """``username`` is now an argument, not something recovered from the
        run entry: ``active_runs`` is keyed by ``(username, run_id)``, so it
        is needed to find the entry at all. Recovering it from the entry was
        never safe anyway -- the entry a bare id found could belong to
        another user (ADR-0009)."""
        svc = _make_service()
        svc.active_runs[("fromdata", 1)] = {
            "data": {"username": "fromdata", "user_password": "pw"},
            "results": [],
        }

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_session = MagicMock()
            mock_session.query.return_value.filter.return_value.all.return_value = []
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            # active_runs is keyed by (username, id): an internal caller must
            # name the owner to reach the run; the stored password still comes
            # from run_data.
            svc.sync_pending_results(1, "fromdata")

            mock_get_session.assert_called_once_with("fromdata", "pw")

    def test_another_users_run_with_the_same_id_is_not_found(self):
        """Two users' first benchmark are both id 1. Asking as one user must
        not reach the other's entry."""
        svc = _make_service()
        svc.active_runs[("alice", 1)] = {
            "data": {"username": "alice", "user_password": "pw"},
            "results": [{"id": "r1"}],
        }

        assert svc.sync_pending_results(1, "bob") == 0

    def test_anonymous_caller_falls_back_to_recorded_owner(self):
        """When no username is passed (anonymous/internal caller), the run stored
        under the normalized ``benchmark_user`` key is found and its recorded
        owner drives the DB session -- exercising the ``if not username`` owner
        fallback that the composite key otherwise makes unnecessary."""
        svc = _make_service()
        # An anonymous run is stored under the normalized owner "benchmark_user".
        svc.active_runs[("benchmark_user", 1)] = {
            "data": {"username": "benchmark_user", "user_password": "anon-pw"},
            "results": [],
        }

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_session = MagicMock()
            mock_session.query.return_value.filter.return_value.all.return_value = []
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            # No username -> _run_key(None, 1) == ("benchmark_user", 1) hits the
            # anonymous run, then the fallback derives the owner for the session.
            svc.sync_pending_results(1)

            mock_get_session.assert_called_once_with(
                "benchmark_user", "anon-pw"
            )


# ============================================================
# _sync_results_to_database
# ============================================================


class TestSyncResultsToDatabase:
    def test_returns_early_if_no_active_run(self):
        svc = _make_service()
        # Should not raise
        svc._sync_results_to_database("user1", 999)

    def test_returns_early_if_thread_not_complete(self):
        svc = _make_service()
        svc.active_runs[("user1", 1)] = {"thread_complete": False}
        svc._sync_results_to_database("user1", 1)
        # No DB calls expected

    def test_syncs_completed_run(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "thread_complete": True,
            "data": {"username": "user1", "user_password": None},
            "completion_info": {
                "status": BenchmarkStatus.COMPLETED,
                "end_time": datetime.now(UTC),
                "completed_examples": 2,
                "failed_examples": 0,
            },
            "results": [
                {
                    "example_id": "ex1",
                    "query_hash": "h1",
                    "dataset_type": "simpleqa",
                    "question": "Q?",
                    "correct_answer": "A",
                    "is_correct": True,
                    "processing_time": 10.0,
                    "task_index": 0,
                },
                {
                    "example_id": "ex2",
                    "query_hash": "h2",
                    "dataset_type": "simpleqa",
                    "question": "Q2?",
                    "correct_answer": "A2",
                    "is_correct": False,
                    "processing_time": 20.0,
                    "task_index": 1,
                },
            ],
        }

        mock_session = MagicMock()
        mock_benchmark_run = MagicMock()
        mock_benchmark_run.status = BenchmarkStatus.COMPLETED
        mock_session.query.return_value.filter.return_value.first.return_value = mock_benchmark_run
        # _persist_unsaved_results reads existing query_hashes for the run.
        mock_session.query.return_value.filter.return_value.all.return_value = []

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            svc._sync_results_to_database("user1", 1)

        mock_session.commit.assert_called_once()
        # Active run should be cleaned up
        assert ("user1", 1) not in svc.active_runs

    def test_calculates_accuracy_correctly(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "thread_complete": True,
            "data": {"username": "user1", "user_password": None},
            "completion_info": {
                "status": BenchmarkStatus.COMPLETED,
                "completed_examples": 2,
                "failed_examples": 0,
            },
            "results": [
                {
                    "example_id": "ex1",
                    "query_hash": "h1",
                    "dataset_type": "simpleqa",
                    "question": "Q?",
                    "correct_answer": "A",
                    "is_correct": True,
                    "processing_time": 10.0,
                    "task_index": 0,
                },
                {
                    "example_id": "ex2",
                    "query_hash": "h2",
                    "dataset_type": "simpleqa",
                    "question": "Q2?",
                    "correct_answer": "A2",
                    "is_correct": True,
                    "processing_time": 20.0,
                    "task_index": 1,
                },
            ],
        }

        mock_session = MagicMock()
        mock_benchmark_run = MagicMock()
        mock_benchmark_run.status = BenchmarkStatus.COMPLETED
        mock_session.query.return_value.filter.return_value.first.return_value = mock_benchmark_run
        # _persist_unsaved_results reads existing query_hashes for the run.
        mock_session.query.return_value.filter.return_value.all.return_value = []

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            svc._sync_results_to_database("user1", 1)

        # 2 correct out of 2 = 100%
        assert mock_benchmark_run.overall_accuracy == 100.0

    def test_skips_already_saved_results(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "thread_complete": True,
            "data": {"username": "user1", "user_password": None},
            "completion_info": {
                "status": BenchmarkStatus.FAILED,
                "error_message": "something failed",
            },
            "results": [
                {
                    "example_id": "ex1",
                    "query_hash": "h1",
                    "dataset_type": "simpleqa",
                    "question": "Q?",
                    "correct_answer": "A",
                    "task_index": 0,
                },
            ],
            "saved_indices": {0},
        }

        mock_session = MagicMock()
        mock_benchmark_run = MagicMock()
        mock_session.query.return_value.filter.return_value.first.return_value = mock_benchmark_run
        # _persist_unsaved_results reads existing query_hashes for the run.
        mock_session.query.return_value.filter.return_value.all.return_value = []

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            svc._sync_results_to_database("user1", 1)

        # Only the run status update, no result adds
        assert mock_session.add.call_count == 0

    def test_handles_db_exception(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "thread_complete": True,
            "data": {"username": "user1", "user_password": None},
            "completion_info": {
                "status": BenchmarkStatus.COMPLETED,
            },
            "results": [],
        }

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                side_effect=RuntimeError("db down")
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            # Should not raise
            svc._sync_results_to_database("user1", 1)

        error = svc.get_result_persistence_error(1, "user1")
        assert error is not None
        assert error["code"] == "database_write_failed"

    def test_no_accuracy_for_non_completed_status(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "thread_complete": True,
            "data": {"username": "user1", "user_password": None},
            "completion_info": {
                "status": BenchmarkStatus.FAILED,
                "error_message": "crashed",
            },
            "results": [
                {
                    "example_id": "ex1",
                    "query_hash": "h1",
                    "dataset_type": "simpleqa",
                    "question": "Q?",
                    "correct_answer": "A",
                    "is_correct": True,
                    "processing_time": 10.0,
                    "task_index": 0,
                },
            ],
        }

        mock_session = MagicMock()
        mock_benchmark_run = MagicMock()
        mock_benchmark_run.status = BenchmarkStatus.FAILED
        mock_session.query.return_value.filter.return_value.first.return_value = mock_benchmark_run
        # _persist_unsaved_results reads existing query_hashes for the run.
        mock_session.query.return_value.filter.return_value.all.return_value = []

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            svc._sync_results_to_database("user1", 1)

        # Should NOT set overall_accuracy for failed runs
        # (the code only calculates accuracy when status == COMPLETED)
        # The mock starts with whatever default it has; we just verify
        # the condition was checked by checking the status was set to FAILED
        assert mock_benchmark_run.status == BenchmarkStatus.FAILED


# ============================================================
# _calculate_final_accuracy
# ============================================================


class TestCalculateFinalAccuracy:
    def test_calculates_accuracy(self):
        svc = _make_service()
        mock_session = MagicMock()

        mock_r1 = MagicMock()
        mock_r1.is_correct = True
        mock_r1.processing_time = 30.0

        mock_r2 = MagicMock()
        mock_r2.is_correct = False
        mock_r2.processing_time = 60.0

        mock_session.query.return_value.filter.return_value.filter.return_value.all.return_value = [
            mock_r1,
            mock_r2,
        ]

        mock_run = MagicMock()
        mock_session.query.return_value.filter.return_value.first.return_value = mock_run

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            svc._calculate_final_accuracy(1, "user1")

        assert mock_run.overall_accuracy == 50.0

    def test_no_results(self):
        svc = _make_service()
        mock_session = MagicMock()
        mock_session.query.return_value.filter.return_value.filter.return_value.all.return_value = []

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            # Should not raise
            svc._calculate_final_accuracy(1)

    def test_handles_exception(self):
        svc = _make_service()
        mock_session = MagicMock()
        mock_session.query.side_effect = RuntimeError("db error")

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            # Should not raise
            svc._calculate_final_accuracy(1)


# ============================================================
# update_benchmark_status
# ============================================================


class TestUpdateBenchmarkStatus:
    def test_update_status_in_progress(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        svc = _make_service()
        mock_session = MagicMock()
        mock_run = MagicMock()
        mock_run.start_time = None
        mock_run.end_time = None
        mock_session.query.return_value.filter.return_value.first.return_value = mock_run

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            svc.update_benchmark_status(1, BenchmarkStatus.IN_PROGRESS)

        assert mock_run.status == BenchmarkStatus.IN_PROGRESS
        assert mock_run.start_time is not None

    def test_update_status_completed(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        svc = _make_service()
        mock_session = MagicMock()
        mock_run = MagicMock()
        mock_run.start_time = datetime.now(UTC)
        mock_run.end_time = None
        mock_session.query.return_value.filter.return_value.first.return_value = mock_run

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            svc.update_benchmark_status(1, BenchmarkStatus.COMPLETED)

        assert mock_run.end_time is not None

    def test_update_status_failed_sets_end_time(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        svc = _make_service()
        mock_session = MagicMock()
        mock_run = MagicMock()
        mock_run.start_time = datetime.now(UTC)
        mock_run.end_time = None
        mock_session.query.return_value.filter.return_value.first.return_value = mock_run

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            svc.update_benchmark_status(
                1, BenchmarkStatus.FAILED, error_message="oops"
            )

        assert mock_run.end_time is not None
        assert mock_run.error_message == "oops"

    def test_update_status_not_found(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        svc = _make_service()
        mock_session = MagicMock()
        mock_session.query.return_value.filter.return_value.first.return_value = None

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            # Should not raise
            svc.update_benchmark_status(1, BenchmarkStatus.COMPLETED)

    def test_update_status_db_error(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        svc = _make_service()
        mock_session = MagicMock()
        mock_session.query.side_effect = RuntimeError("db error")

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            # Should not raise
            svc.update_benchmark_status(1, BenchmarkStatus.COMPLETED)

        mock_session.rollback.assert_called_once()


# ============================================================
# cancel_benchmark
# ============================================================


class TestCancelBenchmark:
    def test_cancel_active_run(self):
        svc = _make_service()
        # Owned by the cancelling user: active_runs is keyed by (username, id),
        # so cancel only ever finds — and stops — the in-memory run for the
        # requesting owner.
        svc.active_runs[("user1", 1)] = {
            "status": "running",
            "data": {"username": "user1"},
        }

        with patch.object(svc, "update_benchmark_status"):
            result = svc.cancel_benchmark(1, "user1")

        assert result is True
        assert svc.active_runs[("user1", 1)]["status"] == "cancelled"

    def test_cancel_nonexistent_run(self):
        svc = _make_service()

        with patch.object(svc, "update_benchmark_status"):
            result = svc.cancel_benchmark(999, "user1")

        assert result is True

    def test_cancel_exception(self):
        svc = _make_service()

        with patch.object(
            svc, "update_benchmark_status", side_effect=RuntimeError("err")
        ):
            result = svc.cancel_benchmark(1)

        assert result is False


# ============================================================
# get_benchmark_status
# ============================================================


class TestGetBenchmarkStatus:
    def test_returns_none_when_not_found(self):
        svc = _make_service()
        mock_session = MagicMock()
        mock_session.query.return_value.filter.return_value.first.return_value = None

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            result = svc.get_benchmark_status(999)

        assert result is None

    def test_returns_status_with_running_accuracy(self):
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        svc = _make_service()
        mock_session = MagicMock()

        mock_run = MagicMock()
        mock_run.id = 1
        mock_run.run_name = "test run"
        mock_run.status = BenchmarkStatus.IN_PROGRESS
        mock_run.completed_examples = 2
        mock_run.total_examples = 10
        mock_run.failed_examples = 0
        mock_run.overall_accuracy = None
        mock_run.processing_rate = None
        mock_run.config_hash = "abc123"
        mock_run.created_at = datetime.now(UTC)
        mock_run.start_time = datetime.now(UTC) - timedelta(seconds=60)
        mock_run.end_time = None
        mock_run.error_message = None

        mock_r1 = MagicMock()
        mock_r1.is_correct = True
        mock_r1.dataset_type = MagicMock()
        mock_r1.dataset_type.value = "simpleqa"

        mock_r2 = MagicMock()
        mock_r2.is_correct = False
        mock_r2.dataset_type = MagicMock()
        mock_r2.dataset_type.value = "simpleqa"

        # Mock the chain of queries
        # We need to handle multiple query() calls differently
        call_count = {"n": 0}

        def side_effect_query(*args):
            call_count["n"] += 1
            m = MagicMock()
            if call_count["n"] == 1:
                # benchmark_run query
                m.filter.return_value.first.return_value = mock_run
            elif call_count["n"] == 2:
                # current_results query
                m.filter.return_value.filter.return_value.all.return_value = [
                    mock_r1,
                    mock_r2,
                ]
            elif call_count["n"] == 3:
                # all_results_for_timing query
                m.filter.return_value.all.return_value = [mock_r1, mock_r2]
            return m

        mock_session.query.side_effect = side_effect_query

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            result = svc.get_benchmark_status(1, "user1")

        assert result is not None
        assert result["running_accuracy"] == 50.0
        assert "simpleqa_accuracy" in result

    def test_handles_exception(self):
        svc = _make_service()
        mock_session = MagicMock()
        mock_session.query.side_effect = RuntimeError("db error")

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            result = svc.get_benchmark_status(1)

        assert result is None


# ============================================================
# _run_benchmark_thread
# ============================================================


class TestRunBenchmarkThread:
    @patch(f"{MODULE}._global_research_semaphore")
    @patch(f"{SETTINGS_CTX_MODULE}.set_settings_context")
    def test_thread_runs_tasks(self, mock_set_ctx, mock_semaphore):
        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "data": {
                "username": "user1",
                "user_password": None,
                "config_hash": "abc",
                "datasets_config": {"simpleqa": {"count": 1}},
                "search_config": {"iterations": 1},
                "evaluation_config": {},
                "settings_snapshot": {},
            },
            "results": [],
        }

        mock_task_result = {"example_id": "ex1", "is_correct": True}

        with (
            patch.object(
                svc,
                "_create_task_queue",
                return_value=[
                    {
                        "benchmark_run_id": 1,
                        "example_id": "ex1",
                        "task_index": 0,
                    }
                ],
            ),
            patch.object(
                svc, "_process_benchmark_task", return_value=mock_task_result
            ),
            patch.object(svc, "_send_progress_update"),
            patch.object(svc, "_sync_results_to_database"),
        ):
            svc._run_benchmark_thread("user1", 1)

        assert len(svc.active_runs[("user1", 1)]["results"]) == 1
        assert svc.active_runs[("user1", 1)]["thread_complete"] is True

    @patch(f"{MODULE}._global_research_semaphore")
    @patch(f"{SETTINGS_CTX_MODULE}.set_settings_context")
    def test_thread_handles_cancelled_run(self, mock_set_ctx, mock_semaphore):
        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "data": {
                "username": "user1",
                "user_password": None,
                "config_hash": "abc",
                "datasets_config": {},
                "search_config": {},
                "evaluation_config": {},
                "settings_snapshot": {},
            },
            "status": "cancelled",
            "results": [],
        }

        with (
            patch.object(
                svc,
                "_create_task_queue",
                return_value=[
                    {
                        "benchmark_run_id": 1,
                        "example_id": "ex1",
                        "task_index": 0,
                    }
                ],
            ),
            patch.object(svc, "_sync_results_to_database"),
        ):
            svc._run_benchmark_thread("user1", 1)

        info = svc.active_runs[("user1", 1)]["completion_info"]
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        assert info["status"] == BenchmarkStatus.CANCELLED

    @patch(f"{MODULE}._global_research_semaphore")
    @patch(f"{SETTINGS_CTX_MODULE}.set_settings_context")
    def test_thread_handles_task_error_rate_limit(
        self, mock_set_ctx, mock_semaphore
    ):
        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "data": {
                "username": "user1",
                "user_password": None,
                "config_hash": "abc",
                "datasets_config": {},
                "search_config": {},
                "evaluation_config": {},
                "settings_snapshot": {},
            },
            "results": [],
        }

        with (
            patch.object(
                svc,
                "_create_task_queue",
                return_value=[
                    {
                        "benchmark_run_id": 1,
                        "example_id": "ex1",
                        "task_index": 0,
                    }
                ],
            ),
            patch.object(
                svc,
                "_process_benchmark_task",
                side_effect=RuntimeError("403 Forbidden rate limit"),
            ),
            patch.object(svc, "_sync_results_to_database"),
        ):
            svc._run_benchmark_thread("user1", 1)

        assert svc.rate_limit_detected == {("user1", 1): True}

    @patch(f"{MODULE}._global_research_semaphore")
    @patch(f"{SETTINGS_CTX_MODULE}.set_settings_context")
    def test_thread_handles_missing_data(self, mock_set_ctx, mock_semaphore):
        svc = _make_service()
        svc.active_runs[("user1", 1)] = {}  # No "data" key

        with patch.object(svc, "_sync_results_to_database"):
            svc._run_benchmark_thread("user1", 1)

        info = svc.active_runs[("user1", 1)]["completion_info"]
        from local_deep_research.benchmarks.web_api.benchmark_service import (
            BenchmarkStatus,
        )

        assert info["status"] == BenchmarkStatus.FAILED

    @patch(f"{MODULE}._global_research_semaphore")
    @patch(f"{SETTINGS_CTX_MODULE}.set_settings_context")
    def test_thread_settings_context_values(self, mock_set_ctx, mock_semaphore):
        """Test that SettingsContext correctly extracts values from setting objects."""
        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "data": {
                "username": "user1",
                "user_password": None,
                "config_hash": "abc",
                "datasets_config": {},
                "search_config": {},
                "evaluation_config": {},
                "settings_snapshot": {
                    "key1": {"value": "val1"},
                    "key2": "direct_val",
                },
            },
            "results": [],
        }

        captured_ctx = {}

        def capture_ctx(ctx):
            captured_ctx["ctx"] = ctx

        mock_set_ctx.side_effect = capture_ctx

        with (
            patch.object(svc, "_create_task_queue", return_value=[]),
            patch.object(svc, "_sync_results_to_database"),
        ):
            svc._run_benchmark_thread("user1", 1)

        ctx = captured_ctx["ctx"]
        assert ctx.get_setting("key1") == "val1"
        assert ctx.get_setting("key2") == "direct_val"
        assert ctx.get_setting("nonexistent", "default") == "default"

    @patch(f"{MODULE}._global_research_semaphore")
    @patch(f"{SETTINGS_CTX_MODULE}.set_settings_context")
    def test_thread_total_zero_examples_progress(
        self, mock_set_ctx, mock_semaphore
    ):
        """Cover branch where total_examples is 0 for progress calculation."""
        svc = _make_service()
        svc.active_runs[("user1", 1)] = {
            "data": {
                "username": "user1",
                "user_password": None,
                "config_hash": "abc",
                "datasets_config": {},
                "search_config": {},
                "evaluation_config": {},
                "settings_snapshot": {},
            },
            "results": [],
        }

        with (
            patch.object(svc, "_create_task_queue", return_value=[]),
            patch.object(svc, "_sync_results_to_database"),
        ):
            svc._run_benchmark_thread("user1", 1)

        # Completion should still happen (progress = 0 since total = 0)
        assert svc.active_runs[("user1", 1)]["thread_complete"] is True


# ============================================================
# start_benchmark
# ============================================================


class TestStartBenchmark:
    def test_start_benchmark_run_not_found(self):
        svc = _make_service()
        mock_session = MagicMock()
        mock_session.query.return_value.filter.return_value.first.return_value = None

        with patch(
            "local_deep_research.database.session_context.get_user_db_session"
        ) as mock_get_session:
            mock_get_session.return_value.__enter__ = Mock(
                return_value=mock_session
            )
            mock_get_session.return_value.__exit__ = Mock(return_value=False)

            result = svc.start_benchmark(999, "user1")

        assert result is False


# ============================================================
# Progress callback in _process_benchmark_task
# ============================================================


class TestProgressCallback:
    @patch(f"{MODULE}.grade_single_result")
    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.format_query")
    def test_callback_milestone_types(
        self, mock_format, mock_summary, mock_extract, mock_grade
    ):
        """Test that the progress callback categorizes log types correctly."""
        mock_socket = MagicMock()
        svc = _make_service(socket=mock_socket)

        def capture_quick_summary(**kwargs):
            cb = kwargs.get("progress_callback")
            if cb:
                # Test various status strings
                cb("Starting phase", 10, {"phase": "init"})
                cb("Completed search", 50, {})
                cb("Error occurred", 80, {})
                cb("Processing data", 60, {})
            return {"summary": "result", "sources": []}

        mock_format.return_value = "q"
        mock_summary.side_effect = capture_quick_summary
        mock_extract.return_value = {"extracted_answer": "ans"}
        mock_grade.return_value = {
            "is_correct": True,
            "graded_confidence": "90",
            "grader_response": "ok",
        }

        mock_settings = MagicMock()
        mock_settings.snapshot = {}

        task = {
            "benchmark_run_id": 1,
            "example_id": "ex1",
            "dataset_type": "simpleqa",
            "question": "Q?",
            "correct_answer": "A",
            "query_hash": "h1",
            "task_index": 0,
            "username": "user1",
            "user_password": None,
        }

        with patch(
            f"{SETTINGS_CTX_MODULE}.get_settings_context",
            return_value=mock_settings,
        ):
            svc._process_benchmark_task(task, {}, {})

        # Verify socket was called multiple times (from callback + regular calls)
        assert mock_socket.emit_to_subscribers.call_count >= 4

    @patch(f"{MODULE}.grade_single_result")
    @patch(f"{MODULE}.extract_answer_from_response")
    @patch(f"{MODULE}.quick_summary")
    @patch(f"{MODULE}.format_query")
    def test_callback_handles_exception(
        self, mock_format, mock_summary, mock_extract, mock_grade
    ):
        """Test that exceptions in the progress callback are caught."""
        mock_socket = MagicMock()
        mock_socket.emit_to_subscribers.side_effect = RuntimeError(
            "socket dead"
        )
        svc = _make_service(socket=mock_socket)

        def trigger_callback(**kwargs):
            cb = kwargs.get("progress_callback")
            if cb:
                cb("test status", 50, {})  # Should not raise
            return {"summary": "result", "sources": []}

        mock_format.return_value = "q"
        mock_summary.side_effect = trigger_callback
        mock_extract.return_value = {"extracted_answer": "ans"}
        mock_grade.return_value = {
            "is_correct": True,
            "graded_confidence": "90",
            "grader_response": "ok",
        }

        mock_settings = MagicMock()
        mock_settings.snapshot = {}

        task = {
            "benchmark_run_id": 1,
            "example_id": "ex1",
            "dataset_type": "simpleqa",
            "question": "Q?",
            "correct_answer": "A",
            "query_hash": "h1",
            "task_index": 0,
            "username": "user1",
            "user_password": None,
        }

        with patch(
            f"{SETTINGS_CTX_MODULE}.get_settings_context",
            return_value=mock_settings,
        ):
            # Should not raise despite socket error in callback
            result = svc._process_benchmark_task(task, {}, {})

        assert result is not None