refacto(admin): Remove ModelProviderGateway

.github/badges/coverage.json

@@ -1 +1 @@
-{"schemaVersion": 1, "label": "coverage", "message": "59.82%", "color": "red"}
+{"schemaVersion": 1, "label": "coverage", "message": "59.78%", "color": "red"}

adr/2026-05-28-refactoring-model-forwarding.md

@@ -109,6 +109,7 @@ The new architecture follows the principles of clean architecture: the model for
 * **Domain contracts for forwarding:** provider client, provider load balancer, router rate limiter, model tokenizer and environmental impact computer are exposed as domain abstractions. The use case depends on these contracts, not on Redis, HTTP, Ecologit or Tiktoken directly.
 * **HTTP client simplified:** `HttpProviderClient` only sends an already formatted request to the selected provider and returns the raw provider response or a model error. It no longer owns endpoint selection, usage computation, metrics or rate limiting.
 * **Endpoint adapters extracted:** provider-specific adapters convert OpenGate requests and responses to each provider format. `build_adapter` selects the right adapter from the source endpoint and provider type, while common usage computation stays in the base adapter.
+* **Provider gateway replaced by a capabilities probe:** the `ProviderGateway` contract and its `ModelProviderGateway` implementation are replaced by a single `ProviderCapabilitiesProbe` class in the use case layer. Probing a provider for its capabilities (max context length, vector size) is an acquisition concern, not a repository. The class holds no technology of its own: it composes the `ProviderClient` and `ProviderAdapterBuilder` domain contracts and sequences two calls, which makes it an application service shared by `CreateProviderUseCase` and `BootstrapModelsUseCase` rather than an adapter. No abstraction is declared for it, since it is not a boundary the domain needs to invert. It does not appear in the diagram below: the model forwarding use case does not probe capabilities.
 * **Redis responsibilities isolated:** Redis implementations handle provider load balancing, provider metrics and router rate limits behind dedicated contracts. The use case decides when those operations happen.
 * **Usage and impacts made explicit:** prompt tokens are computed before rate limiting, response usage is computed after provider response formatting, and environmental impacts are delegated to the Ecologit implementation through a domain contract.
 * **FastAPI endpoint thinned:** the HTTP endpoint builds the command, calls the use case and maps domain errors to HTTP exceptions. It no longer contains forwarding logic.
@@ -134,7 +135,6 @@ subgraph DL[**Domain layer**]
         provider_adapter_builder[ProviderAdapterBuilder]
         provider_adapter[ProviderAdapter]
         provider_repository[ProviderRepository]
-        provider_gateway[ProviderGateway]
         provider_load_balancer[ProviderLoadBalancer]
         provider_client[ProviderClient]
         provider_metrics_logger[ProviderMetricsLogger]
@@ -187,7 +187,6 @@ use_case --> router_rate_limiter
 use_case --> provider_repository
 use_case --> provider_adapter_builder
 use_case --> provider_adapter
-use_case --> provider_gateway
 use_case --> provider_load_balancer
 use_case --> provider_client
 use_case --> provider_metrics_logger
@@ -198,8 +197,6 @@ use_case --> usage_computer
 usage_computer --> model_environmental_impacts_computer
 usage_computer --> model_tokenizer
 
-
-
 provider_adapter_builder --> http_provider_adapter_builder
 provider_client --> http_provider_client
 http_provider_adapter_builder --VLLM, Mistral, TEI...<br>Chat completions, OCR, Rerank...--> http_provider_adapter

api/dependencies.py

@@ -12,7 +12,6 @@
 from api.domain.provider import (
     ProviderAdapterBuilder,
     ProviderClient,
-    ProviderGateway,
     ProviderLoadBalancer,
     ProviderMetricsLogger,
     ProviderRepository,
@@ -26,7 +25,6 @@
 from api.infrastructure.fastapi.context import request_context
 from api.infrastructure.http import HttpProviderAdapterBuilder, HttpProviderClient
 from api.infrastructure.jwt import JwtKeyEncoder
-from api.infrastructure.model import ModelProviderGateway
 from api.infrastructure.postgres import (
     PostgresAuthenticatedUserQuery,
     PostgresKeyRepository,
@@ -56,6 +54,7 @@
 from api.use_cases.health import GetHealthModelsUseCase
 from api.use_cases.models import GetModelsUseCase, GetModelUseCase
 from api.use_cases.reranks import CreateRerankUseCase
+from api.use_cases.services import ProviderCapabilitiesProbe
 from api.utils.configuration import configuration
 from api.utils.context import global_context
 
@@ -122,13 +121,6 @@ def _provider_client() -> ProviderClient:
     return HttpProviderClient()
 
 
-def _provider_gateway(
-    provider_client: ProviderClient = Depends(_provider_client),
-    provider_adapter_builder: ProviderAdapterBuilder = Depends(_provider_adapter_builder),
-) -> ProviderGateway:
-    return ModelProviderGateway(provider_client=provider_client, provider_adapter_builder=provider_adapter_builder)
-
-
 def _provider_load_balancer(redis_client: Redis = Depends(get_redis_client)) -> ProviderLoadBalancer:
     return RedisProviderLoadBalancer(redis_client=redis_client)
 
@@ -170,6 +162,13 @@ def _provider_repository(session: AsyncSession) -> ProviderRepository:
     return PostgresProviderRepository(postgres_session=session)
 
 
+def _provider_capabilities_probe(
+    provider_client: ProviderClient = Depends(_provider_client),
+    provider_adapter_builder: ProviderAdapterBuilder = Depends(_provider_adapter_builder),
+) -> ProviderCapabilitiesProbe:
+    return ProviderCapabilitiesProbe(provider_client=provider_client, provider_adapter_builder=provider_adapter_builder)
+
+
 # health use cases
 def get_health_models_use_case_factory(
     postgres_session: AsyncSession = Depends(get_postgres_session),
@@ -362,12 +361,12 @@ def update_router_use_case_factory(postgres_session: AsyncSession = Depends(get_
 # provider use cases
 def create_provider_use_case_factory(
     postgres_session: AsyncSession = Depends(get_postgres_session),
-    provider_client: ProviderClient = Depends(_provider_client),
+    provider_capabilities_probe: ProviderCapabilitiesProbe = Depends(_provider_capabilities_probe),
 ) -> CreateProviderUseCase:
     return CreateProviderUseCase(
         router_repository=_router_repository(postgres_session),
         provider_repository=_provider_repository(postgres_session),
-        provider_gateway=_provider_gateway(provider_client=provider_client, provider_adapter_builder=HttpProviderAdapterBuilder()),
+        provider_capabilities_probe=provider_capabilities_probe,
     )
 
 

api/domain/provider/__init__.py

@@ -1,7 +1,6 @@
 from api.domain.provider._provideradapter import ProviderAdapter
 from api.domain.provider._provideradapterbuilder import ProviderAdapterBuilder
 from api.domain.provider._providerclient import ProviderClient, ProviderClientResponse
-from api.domain.provider._providergateway import ProviderCapabilities, ProviderGateway
 from api.domain.provider._providerloadbalancer import ProviderLoadBalancer
 from api.domain.provider._providermetricslogger import ProviderMetricsLogger
 from api.domain.provider._providerrepository import ProviderRepository
@@ -11,8 +10,6 @@
     "ProviderAdapterBuilder",
     "ProviderClient",
     "ProviderClientResponse",
-    "ProviderCapabilities",
-    "ProviderGateway",
     "ProviderLoadBalancer",
     "ProviderMetricsLogger",
     "ProviderRepository",

api/domain/provider/entities.py

@@ -146,6 +146,11 @@ def is_compatible_with(self, router: Router) -> bool:
         return self.type.is_compatible_with(router.type)
 
 
+class ProviderCapabilities(BaseModel):
+    max_context_length: int | None
+    vector_size: int | None = None
+
+
 class ProviderOriginalRequest(BaseModel):
     endpoint: Annotated[EndpointRoute, Field(description="The source endpoint (at the user side) of the request.")]
     body: Annotated[CreateEmbeddingsBody | CreateRerankBody | None, Field(default=None, description="The JSON body to use for the request.")]

api/infrastructure/fastapi/endpoints/admin/providers.py

@@ -12,7 +12,7 @@
     update_provider_use_case_factory,
 )
 from api.domain import SortOrder
-from api.domain.model.errors import InconsistentModelMaxContextLengthError, InconsistentModelVectorSizeError
+from api.domain.model.errors import InconsistentModelMaxContextLengthError, InconsistentModelVectorSizeError, ModelNotFoundError
 from api.domain.provider.entities import ProviderSortField
 from api.domain.provider.errors import InvalidProviderTypeError, ProviderAlreadyExistsError, ProviderNotFoundError, ProviderNotReachableError
 from api.domain.router.errors import RouterNotFoundError
@@ -25,6 +25,7 @@
     InconsistentModelVectorSizeHTTPException,
     InternalServerHTTPException,
     InvalidProviderTypeHTTPException,
+    ModelNotFoundHTTPException,
     NotAdminUserHTTPException,
     ProviderAlreadyExistsHTTPException,
     ProviderNotFoundHTTPException,
@@ -70,6 +71,7 @@
             InconsistentModelVectorSizeHTTPException,
             InvalidProviderTypeHTTPException,
             ProviderNotReachableHTTPException,
+            ModelNotFoundHTTPException,
             ProviderAlreadyExistsHTTPException,
             RouterNotFoundHTTPException,
             NotAdminUserHTTPException,
@@ -127,6 +129,9 @@ async def create_provider(
         case ProviderNotReachableError() as error:
             raise ProviderNotReachableHTTPException(name=error.model_name, status_code=error.status_code, detail=error.detail)
 
+        case ModelNotFoundError(name=model_name):
+            raise ModelNotFoundHTTPException(name=model_name)
+
         case ProviderAlreadyExistsError(model_name=model_name, url=url, router_id=router_id):
             raise ProviderAlreadyExistsHTTPException(model_name=model_name, url=url, router_id=router_id)
 

api/tests/integration/endpoints/admin/provider/test_create_provider.py

@@ -7,7 +7,7 @@
 
 from api.dependencies import create_provider_use_case_factory
 from api.domain.model.entities import ModelType as RouterType
-from api.domain.model.errors import InconsistentModelMaxContextLengthError, InconsistentModelVectorSizeError
+from api.domain.model.errors import InconsistentModelMaxContextLengthError, InconsistentModelVectorSizeError, ModelNotFoundError
 from api.domain.provider.entities import ProviderType
 from api.domain.provider.errors import InvalidProviderTypeError, ProviderAlreadyExistsError, ProviderNotReachableError
 from api.domain.router.errors import RouterNotFoundError
@@ -84,6 +84,11 @@ async def test_happy_path(self, client: AsyncClient, db_session):
                 424,
                 "Model provider my-model not reachable (500): error_detail",
             ),
+            (
+                ModelNotFoundError(name="my-model"),
+                404,
+                "Model my-model not found.",
+            ),
             (
                 ProviderAlreadyExistsError(model_name="my-model", url=DEFAULT_PROVIDER_URL, router_id=1),
                 409,

api/tests/integration/models/test_modelprovidergateway.py → api/tests/integration/use_case/test_providercapabilitiesprobe.py

@@ -5,12 +5,11 @@
 import respx
 
 from api.domain.model.entities import ModelType as RouterType
-from api.domain.provider import ProviderCapabilities
-from api.domain.provider.entities import ProviderType
+from api.domain.provider.entities import ProviderCapabilities, ProviderType
 from api.infrastructure.http import HttpProviderAdapterBuilder, HttpProviderClient
-from api.infrastructure.model import ModelProviderGateway
 from api.tests.integration.factories.tei import TeiEmbeddingsResponseFactory, TeiModelsResponseFactory
 from api.tests.integration.factories.vllm import VllmModelsResponseFactory
+from api.use_cases.services import ProviderCapabilitiesProbe
 
 DEFAULT_PROVIDER_URL = "http://my-test-provider/"
 DEFAULT_MODEL_ID = "test/my-model"
@@ -32,22 +31,22 @@ def _mock_embeddings_response(respx_mock, body: dict, status_code: int) -> None:
 
 
 @pytest.fixture
-def gateway() -> ModelProviderGateway:
-    return ModelProviderGateway(provider_client=HttpProviderClient(), provider_adapter_builder=HttpProviderAdapterBuilder())
+def probe() -> ProviderCapabilitiesProbe:
+    return ProviderCapabilitiesProbe(provider_client=HttpProviderClient(), provider_adapter_builder=HttpProviderAdapterBuilder())
 
 
 @pytest.mark.asyncio(loop_scope="session")
-class TestModelProviderGateway:
+class TestProviderCapabilitiesProbe:
     @respx.mock
-    async def test_get_capabilities_of_non_embeddings_providers(self, gateway: ModelProviderGateway):
+    async def test_get_capabilities_of_non_embeddings_providers(self, probe: ProviderCapabilitiesProbe):
         _mock_models_response(
             respx_mock=respx,
             provider_type=ProviderType.VLLM,
             body=VllmModelsResponseFactory(model_id=DEFAULT_MODEL_ID, max_context_length=DEFAULT_MAX_CONTEXT_LENGTH),
             status_code=VllmModelsResponseFactory._status_code,
         )
 
-        result = await gateway.get_capabilities(
+        result = await probe.get_capabilities(
             router_type=RouterType.TEXT_GENERATION,
             provider_type=ProviderType.VLLM,
             url=DEFAULT_PROVIDER_URL,
@@ -61,7 +60,7 @@ async def test_get_capabilities_of_non_embeddings_providers(self, gateway: Model
     @respx.mock
     async def test_get_capabilities_of_embeddings_providers(
         self,
-        gateway: ModelProviderGateway,
+        probe: ProviderCapabilitiesProbe,
     ):
         _mock_models_response(
             respx_mock=respx,
@@ -75,7 +74,7 @@ async def test_get_capabilities_of_embeddings_providers(
             status_code=TeiEmbeddingsResponseFactory._status_code,
         )
 
-        result = await gateway.get_capabilities(
+        result = await probe.get_capabilities(
             router_type=RouterType.TEXT_EMBEDDINGS_INFERENCE,
             provider_type=ProviderType.TEI,
             url=DEFAULT_PROVIDER_URL,