fix: enforce strict mypy on test files

pyproject.toml

@@ -150,10 +150,6 @@ plugins = ["pydantic.mypy"]
 module = "litellm.*"
 ignore_missing_imports = true
 
-[[tool.mypy.overrides]]
-module = "tests.*"
-disallow_untyped_defs = false
-
 [tool.pydantic-mypy]
 init_forbid_extra = true
 init_typed = true

tests/unit/budget/conftest.py

@@ -27,20 +27,20 @@
 # ── Factories ──────────────────────────────────────────────────────
 
 
-class BudgetAlertConfigFactory(ModelFactory):
+class BudgetAlertConfigFactory(ModelFactory[BudgetAlertConfig]):
     __model__ = BudgetAlertConfig
     warn_at = 75
     critical_at = 90
     hard_stop_at = 100
 
 
-class AutoDowngradeConfigFactory(ModelFactory):
+class AutoDowngradeConfigFactory(ModelFactory[AutoDowngradeConfig]):
     __model__ = AutoDowngradeConfig
     enabled = False
     downgrade_map = ()
 
 
-class BudgetConfigFactory(ModelFactory):
+class BudgetConfigFactory(ModelFactory[BudgetConfig]):
     __model__ = BudgetConfig
     total_monthly = 100.0
     per_task_limit = 5.0
@@ -49,44 +49,44 @@ class BudgetConfigFactory(ModelFactory):
     auto_downgrade = AutoDowngradeConfigFactory
 
 
-class TeamBudgetFactory(ModelFactory):
+class TeamBudgetFactory(ModelFactory[TeamBudget]):
     __model__ = TeamBudget
     budget_percent = 10.0
 
 
-class DepartmentBudgetFactory(ModelFactory):
+class DepartmentBudgetFactory(ModelFactory[DepartmentBudget]):
     __model__ = DepartmentBudget
     budget_percent = 25.0
     teams = ()
 
 
-class BudgetHierarchyFactory(ModelFactory):
+class BudgetHierarchyFactory(ModelFactory[BudgetHierarchy]):
     __model__ = BudgetHierarchy
     departments = ()
 
 
-class CostRecordFactory(ModelFactory):
+class CostRecordFactory(ModelFactory[CostRecord]):
     __model__ = CostRecord
     input_tokens = 1000
     output_tokens = 500
     cost_usd = 0.05
 
 
-class PeriodSpendingFactory(ModelFactory):
+class PeriodSpendingFactory(ModelFactory[PeriodSpending]):
     __model__ = PeriodSpending
     start = datetime(2026, 2, 1, tzinfo=UTC)
     end = datetime(2026, 3, 1, tzinfo=UTC)
 
 
-class AgentSpendingFactory(ModelFactory):
+class AgentSpendingFactory(ModelFactory[AgentSpending]):
     __model__ = AgentSpending
 
 
-class DepartmentSpendingFactory(ModelFactory):
+class DepartmentSpendingFactory(ModelFactory[DepartmentSpending]):
     __model__ = DepartmentSpending
 
 
-class SpendingSummaryFactory(ModelFactory):
+class SpendingSummaryFactory(ModelFactory[SpendingSummary]):
     __model__ = SpendingSummary
     period = PeriodSpendingFactory
     by_agent = ()

tests/unit/budget/test_config.py

@@ -51,7 +51,7 @@ def test_hard_stop_at_boundary_100(self) -> None:
     def test_float_threshold_rejected(self) -> None:
         """Reject float value for threshold (strict int)."""
         with pytest.raises(ValidationError):
-            BudgetAlertConfig(warn_at=75.5, critical_at=90, hard_stop_at=100)
+            BudgetAlertConfig(warn_at=75.5, critical_at=90, hard_stop_at=100)  # type: ignore[arg-type]
 
     def test_negative_threshold_rejected(self) -> None:
         """Reject negative threshold values."""
@@ -153,7 +153,7 @@ def test_duplicate_source_alias_rejected(self) -> None:
     def test_float_threshold_rejected(self) -> None:
         """Reject float value for threshold (strict int)."""
         with pytest.raises(ValidationError):
-            AutoDowngradeConfig(threshold=85.5)
+            AutoDowngradeConfig(threshold=85.5)  # type: ignore[arg-type]
 
     def test_threshold_boundary_0(self) -> None:
         """Accept threshold at lower boundary (0)."""

tests/unit/budget/test_enums.py

@@ -22,12 +22,12 @@ def test_all_members_exist(self) -> None:
 
     def test_values_are_strings(self) -> None:
         """Verify StrEnum produces string values."""
-        assert BudgetAlertLevel.NORMAL == "normal"
-        assert BudgetAlertLevel.WARNING == "warning"
-        assert BudgetAlertLevel.CRITICAL == "critical"
-        assert BudgetAlertLevel.HARD_STOP == "hard_stop"
+        assert BudgetAlertLevel.NORMAL.value == "normal"
+        assert BudgetAlertLevel.WARNING.value == "warning"
+        assert BudgetAlertLevel.CRITICAL.value == "critical"
+        assert BudgetAlertLevel.HARD_STOP.value == "hard_stop"
 
     def test_membership(self) -> None:
         """Verify string-based membership check works."""
-        assert "normal" in BudgetAlertLevel.__members__.values()
-        assert "warning" in BudgetAlertLevel.__members__.values()
+        assert "normal" in [m.value for m in BudgetAlertLevel]
+        assert "warning" in [m.value for m in BudgetAlertLevel]

tests/unit/communication/conftest.py

@@ -17,6 +17,7 @@
     RateLimitConfig,
 )
 from ai_company.communication.enums import (
+    AttachmentType,
     ChannelType,
     MessagePriority,
     MessageType,
@@ -26,11 +27,11 @@
 # ── Factories ──────────────────────────────────────────────────────
 
 
-class AttachmentFactory(ModelFactory):
+class AttachmentFactory(ModelFactory[Attachment]):
     __model__ = Attachment
 
 
-class MessageMetadataFactory(ModelFactory):
+class MessageMetadataFactory(ModelFactory[MessageMetadata]):
     __model__ = MessageMetadata
     task_id = None
     project_id = None
@@ -39,52 +40,52 @@ class MessageMetadataFactory(ModelFactory):
     extra = ()
 
 
-class MessageFactory(ModelFactory):
+class MessageFactory(ModelFactory[Message]):
     __model__ = Message
     priority = MessagePriority.NORMAL
     attachments = ()
     metadata = MessageMetadataFactory
 
 
-class ChannelFactory(ModelFactory):
+class ChannelFactory(ModelFactory[Channel]):
     __model__ = Channel
     type = ChannelType.TOPIC
     subscribers = ()
 
 
-class MessageBusConfigFactory(ModelFactory):
+class MessageBusConfigFactory(ModelFactory[MessageBusConfig]):
     __model__ = MessageBusConfig
 
 
-class MeetingTypeConfigFactory(ModelFactory):
+class MeetingTypeConfigFactory(ModelFactory[MeetingTypeConfig]):
     __model__ = MeetingTypeConfig
     frequency = "daily"
     trigger = None
 
 
-class MeetingsConfigFactory(ModelFactory):
+class MeetingsConfigFactory(ModelFactory[MeetingsConfig]):
     __model__ = MeetingsConfig
     types = ()
 
 
-class HierarchyConfigFactory(ModelFactory):
+class HierarchyConfigFactory(ModelFactory[HierarchyConfig]):
     __model__ = HierarchyConfig
 
 
-class RateLimitConfigFactory(ModelFactory):
+class RateLimitConfigFactory(ModelFactory[RateLimitConfig]):
     __model__ = RateLimitConfig
 
 
-class CircuitBreakerConfigFactory(ModelFactory):
+class CircuitBreakerConfigFactory(ModelFactory[CircuitBreakerConfig]):
     __model__ = CircuitBreakerConfig
 
 
-class LoopPreventionConfigFactory(ModelFactory):
+class LoopPreventionConfigFactory(ModelFactory[LoopPreventionConfig]):
     __model__ = LoopPreventionConfig
     ancestry_tracking = True
 
 
-class CommunicationConfigFactory(ModelFactory):
+class CommunicationConfigFactory(ModelFactory[CommunicationConfig]):
     __model__ = CommunicationConfig
     meetings = MeetingsConfigFactory
     loop_prevention = LoopPreventionConfigFactory
@@ -95,7 +96,7 @@ class CommunicationConfigFactory(ModelFactory):
 
 @pytest.fixture
 def sample_attachment() -> Attachment:
-    return Attachment(type="artifact", ref="pr-42")
+    return Attachment(type=AttachmentType.ARTIFACT, ref="pr-42")
 
 
 @pytest.fixture
@@ -118,7 +119,7 @@ def sample_message(sample_metadata: MessageMetadata) -> Message:
         priority=MessagePriority.NORMAL,
         channel="#backend",
         content="Completed API endpoint for user authentication.",
-        attachments=(Attachment(type="artifact", ref="pr-42"),),
+        attachments=(Attachment(type=AttachmentType.ARTIFACT, ref="pr-42"),),
         metadata=sample_metadata,
     )
 

tests/unit/communication/test_config.py

@@ -435,7 +435,7 @@ def test_ancestry_tracking_false_rejected(self) -> None:
             ValidationError,
             match="Input should be True",
         ):
-            LoopPreventionConfig(ancestry_tracking=False)
+            LoopPreventionConfig(ancestry_tracking=False)  # type: ignore[arg-type]
 
     def test_zero_delegation_depth_rejected(self) -> None:
         with pytest.raises(ValidationError):

tests/unit/communication/test_enums.py

@@ -20,14 +20,14 @@ def test_member_count(self) -> None:
         assert len(MessageType) == 8
 
     def test_values(self) -> None:
-        assert MessageType.TASK_UPDATE == "task_update"
-        assert MessageType.QUESTION == "question"
-        assert MessageType.ANNOUNCEMENT == "announcement"
-        assert MessageType.REVIEW_REQUEST == "review_request"
-        assert MessageType.APPROVAL == "approval"
-        assert MessageType.DELEGATION == "delegation"
-        assert MessageType.STATUS_REPORT == "status_report"
-        assert MessageType.ESCALATION == "escalation"
+        assert MessageType.TASK_UPDATE.value == "task_update"
+        assert MessageType.QUESTION.value == "question"
+        assert MessageType.ANNOUNCEMENT.value == "announcement"
+        assert MessageType.REVIEW_REQUEST.value == "review_request"
+        assert MessageType.APPROVAL.value == "approval"
+        assert MessageType.DELEGATION.value == "delegation"
+        assert MessageType.STATUS_REPORT.value == "status_report"
+        assert MessageType.ESCALATION.value == "escalation"
 
     def test_string_identity(self) -> None:
         assert str(MessageType.TASK_UPDATE) == "task_update"
@@ -39,10 +39,10 @@ def test_member_count(self) -> None:
         assert len(MessagePriority) == 4
 
     def test_values(self) -> None:
-        assert MessagePriority.LOW == "low"
-        assert MessagePriority.NORMAL == "normal"
-        assert MessagePriority.HIGH == "high"
-        assert MessagePriority.URGENT == "urgent"
+        assert MessagePriority.LOW.value == "low"
+        assert MessagePriority.NORMAL.value == "normal"
+        assert MessagePriority.HIGH.value == "high"
+        assert MessagePriority.URGENT.value == "urgent"
 
     def test_normal_not_medium(self) -> None:
         """Message priority uses 'normal', not 'medium' like task Priority."""
@@ -57,9 +57,9 @@ def test_member_count(self) -> None:
         assert len(ChannelType) == 3
 
     def test_values(self) -> None:
-        assert ChannelType.TOPIC == "topic"
-        assert ChannelType.DIRECT == "direct"
-        assert ChannelType.BROADCAST == "broadcast"
+        assert ChannelType.TOPIC.value == "topic"
+        assert ChannelType.DIRECT.value == "direct"
+        assert ChannelType.BROADCAST.value == "broadcast"
 
 
 @pytest.mark.unit
@@ -68,9 +68,9 @@ def test_member_count(self) -> None:
         assert len(AttachmentType) == 3
 
     def test_values(self) -> None:
-        assert AttachmentType.ARTIFACT == "artifact"
-        assert AttachmentType.FILE == "file"
-        assert AttachmentType.LINK == "link"
+        assert AttachmentType.ARTIFACT.value == "artifact"
+        assert AttachmentType.FILE.value == "file"
+        assert AttachmentType.LINK.value == "link"
 
 
 @pytest.mark.unit
@@ -79,10 +79,10 @@ def test_member_count(self) -> None:
         assert len(CommunicationPattern) == 4
 
     def test_values(self) -> None:
-        assert CommunicationPattern.EVENT_DRIVEN == "event_driven"
-        assert CommunicationPattern.HIERARCHICAL == "hierarchical"
-        assert CommunicationPattern.MEETING_BASED == "meeting_based"
-        assert CommunicationPattern.HYBRID == "hybrid"
+        assert CommunicationPattern.EVENT_DRIVEN.value == "event_driven"
+        assert CommunicationPattern.HIERARCHICAL.value == "hierarchical"
+        assert CommunicationPattern.MEETING_BASED.value == "meeting_based"
+        assert CommunicationPattern.HYBRID.value == "hybrid"
 
 
 @pytest.mark.unit
@@ -100,7 +100,7 @@ def test_member_count(self) -> None:
         assert len(MessageBusBackend) == 4
 
     def test_values(self) -> None:
-        assert MessageBusBackend.INTERNAL == "internal"
-        assert MessageBusBackend.REDIS == "redis"
-        assert MessageBusBackend.RABBITMQ == "rabbitmq"
-        assert MessageBusBackend.KAFKA == "kafka"
+        assert MessageBusBackend.INTERNAL.value == "internal"
+        assert MessageBusBackend.REDIS.value == "redis"
+        assert MessageBusBackend.RABBITMQ.value == "rabbitmq"
+        assert MessageBusBackend.KAFKA.value == "kafka"

tests/unit/config/conftest.py

@@ -24,35 +24,35 @@
 # ── Factories ──────────────────────────────────────────────────────
 
 
-class ProviderModelConfigFactory(ModelFactory):
+class ProviderModelConfigFactory(ModelFactory[ProviderModelConfig]):
     __model__ = ProviderModelConfig
 
 
-class ProviderConfigFactory(ModelFactory):
+class ProviderConfigFactory(ModelFactory[ProviderConfig]):
     __model__ = ProviderConfig
     models = ()
 
 
-class RoutingRuleConfigFactory(ModelFactory):
+class RoutingRuleConfigFactory(ModelFactory[RoutingRuleConfig]):
     __model__ = RoutingRuleConfig
 
 
-class RoutingConfigFactory(ModelFactory):
+class RoutingConfigFactory(ModelFactory[RoutingConfig]):
     __model__ = RoutingConfig
     rules = ()
     fallback_chain = ()
 
 
-class AgentConfigFactory(ModelFactory):
+class AgentConfigFactory(ModelFactory[AgentConfig]):
     __model__ = AgentConfig
 
 
-class RootConfigFactory(ModelFactory):
+class RootConfigFactory(ModelFactory[RootConfig]):
     __model__ = RootConfig
     departments = ()
     agents = ()
     custom_roles = ()
-    providers: dict[str, ProviderConfig] = {}  # type: ignore[assignment]  # noqa: RUF012
+    providers: dict[str, ProviderConfig] = {}  # noqa: RUF012
     config = CompanyConfig()
     budget = BudgetConfig()
     communication = CommunicationConfig()
@@ -139,7 +139,7 @@ def sample_root_config() -> RootConfig:
 
 
 @pytest.fixture
-def tmp_config_file(tmp_path: Path) -> Callable[[str, str], Path]:
+def tmp_config_file(tmp_path: Path) -> Callable[..., Path]:
     def _create(content: str, name: str = "config.yaml") -> Path:
         path = tmp_path / name
         path.write_text(content, encoding="utf-8")