weave-typescript 0.40.0 → 0.42.4

package/dist/weaveapi/{inference/v1/inference.pb.js → gateway/v1/gateway.pb.js}

@@ -3,204 +3,204 @@
 // versions:
 //   protoc-gen-ts_proto  v2.11.8
 //   protoc               unknown
-// source: weaveapi/inference/v1/inference.proto
+// source: weaveapi/gateway/v1/gateway.proto
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.OpenAIImageData = exports.OpenAICompatibilityServiceGenerateImageRequest = exports.OpenAICompatibilityServiceCreateEmbeddingResponse = exports.OpenAIEmbeddingData = exports.OpenAICompatibilityServiceCreateEmbeddingRequest = exports.OpenAICompatibilityServiceStreamChatCompletionResponse = exports.OpenAICompatibilityServiceStreamChatCompletionRequest = exports.OpenAIChatDeltaChoice = exports.OpenAICompatibilityServiceCreateChatCompletionResponse = exports.OpenAIChatChoice = exports.OpenAICompatibilityServiceCreateChatCompletionRequest = exports.OpenAIChatToolCall = exports.OpenAIFunctionCall = exports.OpenAIResponseFormat = exports.OpenAIToolDefinition = exports.OpenAIFunctionDefinition = exports.OpenAIChatMessage = exports.CallMCPResponse = exports.CallMCPRequest = exports.GenerateImagesResponse = exports.GenerateImagesRequest = exports.CreateEmbeddingsResponse = exports.CreateEmbeddingsRequest = exports.CallToolsResponse = exports.CallToolsRequest = exports.StreamChatCompletionResponse = exports.StreamChatCompletionRequest = exports.CreateChatCompletionResponse = exports.CreateChatCompletionRequest = exports.GeneratedImage = exports.EmbeddingVector = exports.EmbeddingInput = exports.ChatMessage = exports.InferenceContentPart = exports.ToolResult = exports.ToolCall = exports.ToolDefinition = exports.StructuredOutputSchema = exports.InferenceGenerationConfig = exports.InferenceUsage = exports.InferenceTokenUsageBreakdown = exports.ResolvedInferenceContext = exports.InferenceRequestContext = exports.InferenceTokenKind = exports.ImageDeliveryMode = exports.ToolChoiceMode = exports.InferenceFinishReason = exports.InferenceMessageRole = exports.InferenceOperationType = exports.protobufPackage = void 0;
-exports.AnthropicCompatibilityServiceStreamMessagesResponse = exports.AnthropicCompatibilityServiceStreamMessagesRequest = exports.AnthropicCompatibilityServiceCreateMessageResponse = exports.AnthropicUsage = exports.AnthropicContentBlock = exports.AnthropicCompatibilityServiceCreateMessageRequest = exports.AnthropicToolDefinition = exports.AnthropicMessage = exports.OpenAICompatibilityServiceGenerateImageResponse = void 0;
-exports.inferenceOperationTypeFromJSON = inferenceOperationTypeFromJSON;
-exports.inferenceOperationTypeToJSON = inferenceOperationTypeToJSON;
-exports.inferenceMessageRoleFromJSON = inferenceMessageRoleFromJSON;
-exports.inferenceMessageRoleToJSON = inferenceMessageRoleToJSON;
-exports.inferenceFinishReasonFromJSON = inferenceFinishReasonFromJSON;
-exports.inferenceFinishReasonToJSON = inferenceFinishReasonToJSON;
+exports.OpenAICompatibilityServiceGenerateImageRequest = exports.OpenAICompatibilityServiceCreateEmbeddingResponse = exports.OpenAIEmbeddingData = exports.OpenAICompatibilityServiceCreateEmbeddingRequest = exports.OpenAICompatibilityServiceStreamChatCompletionResponse = exports.OpenAICompatibilityServiceStreamChatCompletionRequest = exports.OpenAIChatDeltaChoice = exports.OpenAICompatibilityServiceCreateChatCompletionResponse = exports.OpenAIChatChoice = exports.OpenAICompatibilityServiceCreateChatCompletionRequest = exports.OpenAIChatToolCall = exports.OpenAIFunctionCall = exports.OpenAIResponseFormat = exports.OpenAIToolDefinition = exports.OpenAIFunctionDefinition = exports.OpenAIChatMessage = exports.CallMCPResponse = exports.CallMCPRequest = exports.GenerateImagesResponse = exports.GenerateImagesRequest = exports.CreateEmbeddingsResponse = exports.CreateEmbeddingsRequest = exports.CallToolsResponse = exports.CallToolsRequest = exports.StreamChatCompletionResponse = exports.StreamChatCompletionRequest = exports.CreateChatCompletionResponse = exports.CreateChatCompletionRequest = exports.GeneratedImage = exports.EmbeddingVector = exports.EmbeddingInput = exports.ChatMessage = exports.ContentPart = exports.ToolResult = exports.ToolCallDelta = exports.ToolCall = exports.ToolDefinition = exports.StructuredOutputSchema = exports.GenerationConfig = exports.Usage = exports.TokenUsageBreakdown = exports.ResolvedModelRoute = exports.RequestContext = exports.TokenKind = exports.ImageDeliveryMode = exports.ToolChoiceMode = exports.FinishReason = exports.MessageRole = exports.ModelOperation = exports.protobufPackage = void 0;
+exports.AnthropicCompatibilityServiceStreamMessagesResponse = exports.AnthropicCompatibilityServiceStreamMessagesRequest = exports.AnthropicCompatibilityServiceCreateMessageResponse = exports.AnthropicUsage = exports.AnthropicContentBlock = exports.AnthropicCompatibilityServiceCreateMessageRequest = exports.AnthropicToolDefinition = exports.AnthropicMessage = exports.OpenAICompatibilityServiceGenerateImageResponse = exports.OpenAIImageData = void 0;
+exports.modelOperationFromJSON = modelOperationFromJSON;
+exports.modelOperationToJSON = modelOperationToJSON;
+exports.messageRoleFromJSON = messageRoleFromJSON;
+exports.messageRoleToJSON = messageRoleToJSON;
+exports.finishReasonFromJSON = finishReasonFromJSON;
+exports.finishReasonToJSON = finishReasonToJSON;
 exports.toolChoiceModeFromJSON = toolChoiceModeFromJSON;
 exports.toolChoiceModeToJSON = toolChoiceModeToJSON;
 exports.imageDeliveryModeFromJSON = imageDeliveryModeFromJSON;
 exports.imageDeliveryModeToJSON = imageDeliveryModeToJSON;
-exports.inferenceTokenKindFromJSON = inferenceTokenKindFromJSON;
-exports.inferenceTokenKindToJSON = inferenceTokenKindToJSON;
+exports.tokenKindFromJSON = tokenKindFromJSON;
+exports.tokenKindToJSON = tokenKindToJSON;
 /* eslint-disable */
 const wire_1 = require("@bufbuild/protobuf/wire");
 const struct_pb_1 = require("../../../google/protobuf/struct.pb");
 const timestamp_pb_1 = require("../../../google/protobuf/timestamp.pb");
 const provider_pb_1 = require("../../provider/v1/provider.pb");
-exports.protobufPackage = "weaveapi.inference.v1";
-/** InferenceOperationType identifies one canonical execution capability. */
-var InferenceOperationType;
-(function (InferenceOperationType) {
-    InferenceOperationType[InferenceOperationType["INFERENCE_OPERATION_TYPE_UNSPECIFIED"] = 0] = "INFERENCE_OPERATION_TYPE_UNSPECIFIED";
-    InferenceOperationType[InferenceOperationType["INFERENCE_OPERATION_TYPE_CHAT_COMPLETION"] = 1] = "INFERENCE_OPERATION_TYPE_CHAT_COMPLETION";
-    InferenceOperationType[InferenceOperationType["INFERENCE_OPERATION_TYPE_STREAM_CHAT_COMPLETION"] = 2] = "INFERENCE_OPERATION_TYPE_STREAM_CHAT_COMPLETION";
-    InferenceOperationType[InferenceOperationType["INFERENCE_OPERATION_TYPE_STRUCTURED_OUTPUT"] = 3] = "INFERENCE_OPERATION_TYPE_STRUCTURED_OUTPUT";
-    InferenceOperationType[InferenceOperationType["INFERENCE_OPERATION_TYPE_TOOL_CALLING"] = 4] = "INFERENCE_OPERATION_TYPE_TOOL_CALLING";
-    InferenceOperationType[InferenceOperationType["INFERENCE_OPERATION_TYPE_EMBEDDINGS"] = 5] = "INFERENCE_OPERATION_TYPE_EMBEDDINGS";
-    InferenceOperationType[InferenceOperationType["INFERENCE_OPERATION_TYPE_IMAGE_GENERATION"] = 6] = "INFERENCE_OPERATION_TYPE_IMAGE_GENERATION";
-    InferenceOperationType[InferenceOperationType["INFERENCE_OPERATION_TYPE_MCP_CALL"] = 7] = "INFERENCE_OPERATION_TYPE_MCP_CALL";
-    InferenceOperationType[InferenceOperationType["UNRECOGNIZED"] = -1] = "UNRECOGNIZED";
-})(InferenceOperationType || (exports.InferenceOperationType = InferenceOperationType = {}));
-function inferenceOperationTypeFromJSON(object) {
+exports.protobufPackage = "weaveapi.gateway.v1";
+/** ModelOperation identifies one model gateway capability. */
+var ModelOperation;
+(function (ModelOperation) {
+    ModelOperation[ModelOperation["MODEL_OPERATION_UNSPECIFIED"] = 0] = "MODEL_OPERATION_UNSPECIFIED";
+    ModelOperation[ModelOperation["MODEL_OPERATION_CHAT_COMPLETION"] = 1] = "MODEL_OPERATION_CHAT_COMPLETION";
+    ModelOperation[ModelOperation["MODEL_OPERATION_STREAM_CHAT_COMPLETION"] = 2] = "MODEL_OPERATION_STREAM_CHAT_COMPLETION";
+    ModelOperation[ModelOperation["MODEL_OPERATION_STRUCTURED_OUTPUT"] = 3] = "MODEL_OPERATION_STRUCTURED_OUTPUT";
+    ModelOperation[ModelOperation["MODEL_OPERATION_TOOL_CALLING"] = 4] = "MODEL_OPERATION_TOOL_CALLING";
+    ModelOperation[ModelOperation["MODEL_OPERATION_EMBEDDINGS"] = 5] = "MODEL_OPERATION_EMBEDDINGS";
+    ModelOperation[ModelOperation["MODEL_OPERATION_IMAGE_GENERATION"] = 6] = "MODEL_OPERATION_IMAGE_GENERATION";
+    ModelOperation[ModelOperation["MODEL_OPERATION_MCP_CALL"] = 7] = "MODEL_OPERATION_MCP_CALL";
+    ModelOperation[ModelOperation["UNRECOGNIZED"] = -1] = "UNRECOGNIZED";
+})(ModelOperation || (exports.ModelOperation = ModelOperation = {}));
+function modelOperationFromJSON(object) {
     switch (object) {
         case 0:
-        case "INFERENCE_OPERATION_TYPE_UNSPECIFIED":
-            return InferenceOperationType.INFERENCE_OPERATION_TYPE_UNSPECIFIED;
+        case "MODEL_OPERATION_UNSPECIFIED":
+            return ModelOperation.MODEL_OPERATION_UNSPECIFIED;
         case 1:
-        case "INFERENCE_OPERATION_TYPE_CHAT_COMPLETION":
-            return InferenceOperationType.INFERENCE_OPERATION_TYPE_CHAT_COMPLETION;
+        case "MODEL_OPERATION_CHAT_COMPLETION":
+            return ModelOperation.MODEL_OPERATION_CHAT_COMPLETION;
         case 2:
-        case "INFERENCE_OPERATION_TYPE_STREAM_CHAT_COMPLETION":
-            return InferenceOperationType.INFERENCE_OPERATION_TYPE_STREAM_CHAT_COMPLETION;
+        case "MODEL_OPERATION_STREAM_CHAT_COMPLETION":
+            return ModelOperation.MODEL_OPERATION_STREAM_CHAT_COMPLETION;
         case 3:
-        case "INFERENCE_OPERATION_TYPE_STRUCTURED_OUTPUT":
-            return InferenceOperationType.INFERENCE_OPERATION_TYPE_STRUCTURED_OUTPUT;
+        case "MODEL_OPERATION_STRUCTURED_OUTPUT":
+            return ModelOperation.MODEL_OPERATION_STRUCTURED_OUTPUT;
         case 4:
-        case "INFERENCE_OPERATION_TYPE_TOOL_CALLING":
-            return InferenceOperationType.INFERENCE_OPERATION_TYPE_TOOL_CALLING;
+        case "MODEL_OPERATION_TOOL_CALLING":
+            return ModelOperation.MODEL_OPERATION_TOOL_CALLING;
         case 5:
-        case "INFERENCE_OPERATION_TYPE_EMBEDDINGS":
-            return InferenceOperationType.INFERENCE_OPERATION_TYPE_EMBEDDINGS;
+        case "MODEL_OPERATION_EMBEDDINGS":
+            return ModelOperation.MODEL_OPERATION_EMBEDDINGS;
         case 6:
-        case "INFERENCE_OPERATION_TYPE_IMAGE_GENERATION":
-            return InferenceOperationType.INFERENCE_OPERATION_TYPE_IMAGE_GENERATION;
+        case "MODEL_OPERATION_IMAGE_GENERATION":
+            return ModelOperation.MODEL_OPERATION_IMAGE_GENERATION;
         case 7:
-        case "INFERENCE_OPERATION_TYPE_MCP_CALL":
-            return InferenceOperationType.INFERENCE_OPERATION_TYPE_MCP_CALL;
+        case "MODEL_OPERATION_MCP_CALL":
+            return ModelOperation.MODEL_OPERATION_MCP_CALL;
         case -1:
         case "UNRECOGNIZED":
         default:
-            return InferenceOperationType.UNRECOGNIZED;
+            return ModelOperation.UNRECOGNIZED;
     }
 }
-function inferenceOperationTypeToJSON(object) {
+function modelOperationToJSON(object) {
     switch (object) {
-        case InferenceOperationType.INFERENCE_OPERATION_TYPE_UNSPECIFIED:
-            return "INFERENCE_OPERATION_TYPE_UNSPECIFIED";
-        case InferenceOperationType.INFERENCE_OPERATION_TYPE_CHAT_COMPLETION:
-            return "INFERENCE_OPERATION_TYPE_CHAT_COMPLETION";
-        case InferenceOperationType.INFERENCE_OPERATION_TYPE_STREAM_CHAT_COMPLETION:
-            return "INFERENCE_OPERATION_TYPE_STREAM_CHAT_COMPLETION";
-        case InferenceOperationType.INFERENCE_OPERATION_TYPE_STRUCTURED_OUTPUT:
-            return "INFERENCE_OPERATION_TYPE_STRUCTURED_OUTPUT";
-        case InferenceOperationType.INFERENCE_OPERATION_TYPE_TOOL_CALLING:
-            return "INFERENCE_OPERATION_TYPE_TOOL_CALLING";
-        case InferenceOperationType.INFERENCE_OPERATION_TYPE_EMBEDDINGS:
-            return "INFERENCE_OPERATION_TYPE_EMBEDDINGS";
-        case InferenceOperationType.INFERENCE_OPERATION_TYPE_IMAGE_GENERATION:
-            return "INFERENCE_OPERATION_TYPE_IMAGE_GENERATION";
-        case InferenceOperationType.INFERENCE_OPERATION_TYPE_MCP_CALL:
-            return "INFERENCE_OPERATION_TYPE_MCP_CALL";
-        case InferenceOperationType.UNRECOGNIZED:
+        case ModelOperation.MODEL_OPERATION_UNSPECIFIED:
+            return "MODEL_OPERATION_UNSPECIFIED";
+        case ModelOperation.MODEL_OPERATION_CHAT_COMPLETION:
+            return "MODEL_OPERATION_CHAT_COMPLETION";
+        case ModelOperation.MODEL_OPERATION_STREAM_CHAT_COMPLETION:
+            return "MODEL_OPERATION_STREAM_CHAT_COMPLETION";
+        case ModelOperation.MODEL_OPERATION_STRUCTURED_OUTPUT:
+            return "MODEL_OPERATION_STRUCTURED_OUTPUT";
+        case ModelOperation.MODEL_OPERATION_TOOL_CALLING:
+            return "MODEL_OPERATION_TOOL_CALLING";
+        case ModelOperation.MODEL_OPERATION_EMBEDDINGS:
+            return "MODEL_OPERATION_EMBEDDINGS";
+        case ModelOperation.MODEL_OPERATION_IMAGE_GENERATION:
+            return "MODEL_OPERATION_IMAGE_GENERATION";
+        case ModelOperation.MODEL_OPERATION_MCP_CALL:
+            return "MODEL_OPERATION_MCP_CALL";
+        case ModelOperation.UNRECOGNIZED:
         default:
             return "UNRECOGNIZED";
     }
 }
-/** InferenceMessageRole identifies one participant role in a canonical message. */
-var InferenceMessageRole;
-(function (InferenceMessageRole) {
-    InferenceMessageRole[InferenceMessageRole["INFERENCE_MESSAGE_ROLE_UNSPECIFIED"] = 0] = "INFERENCE_MESSAGE_ROLE_UNSPECIFIED";
-    InferenceMessageRole[InferenceMessageRole["INFERENCE_MESSAGE_ROLE_SYSTEM"] = 1] = "INFERENCE_MESSAGE_ROLE_SYSTEM";
-    InferenceMessageRole[InferenceMessageRole["INFERENCE_MESSAGE_ROLE_USER"] = 2] = "INFERENCE_MESSAGE_ROLE_USER";
-    InferenceMessageRole[InferenceMessageRole["INFERENCE_MESSAGE_ROLE_ASSISTANT"] = 3] = "INFERENCE_MESSAGE_ROLE_ASSISTANT";
-    InferenceMessageRole[InferenceMessageRole["INFERENCE_MESSAGE_ROLE_TOOL"] = 4] = "INFERENCE_MESSAGE_ROLE_TOOL";
-    InferenceMessageRole[InferenceMessageRole["INFERENCE_MESSAGE_ROLE_DEVELOPER"] = 5] = "INFERENCE_MESSAGE_ROLE_DEVELOPER";
-    InferenceMessageRole[InferenceMessageRole["UNRECOGNIZED"] = -1] = "UNRECOGNIZED";
-})(InferenceMessageRole || (exports.InferenceMessageRole = InferenceMessageRole = {}));
-function inferenceMessageRoleFromJSON(object) {
+/** MessageRole identifies one participant role in a model gateway message. */
+var MessageRole;
+(function (MessageRole) {
+    MessageRole[MessageRole["MESSAGE_ROLE_UNSPECIFIED"] = 0] = "MESSAGE_ROLE_UNSPECIFIED";
+    MessageRole[MessageRole["MESSAGE_ROLE_SYSTEM"] = 1] = "MESSAGE_ROLE_SYSTEM";
+    MessageRole[MessageRole["MESSAGE_ROLE_USER"] = 2] = "MESSAGE_ROLE_USER";
+    MessageRole[MessageRole["MESSAGE_ROLE_ASSISTANT"] = 3] = "MESSAGE_ROLE_ASSISTANT";
+    MessageRole[MessageRole["MESSAGE_ROLE_TOOL"] = 4] = "MESSAGE_ROLE_TOOL";
+    MessageRole[MessageRole["MESSAGE_ROLE_DEVELOPER"] = 5] = "MESSAGE_ROLE_DEVELOPER";
+    MessageRole[MessageRole["UNRECOGNIZED"] = -1] = "UNRECOGNIZED";
+})(MessageRole || (exports.MessageRole = MessageRole = {}));
+function messageRoleFromJSON(object) {
     switch (object) {
         case 0:
-        case "INFERENCE_MESSAGE_ROLE_UNSPECIFIED":
-            return InferenceMessageRole.INFERENCE_MESSAGE_ROLE_UNSPECIFIED;
+        case "MESSAGE_ROLE_UNSPECIFIED":
+            return MessageRole.MESSAGE_ROLE_UNSPECIFIED;
         case 1:
-        case "INFERENCE_MESSAGE_ROLE_SYSTEM":
-            return InferenceMessageRole.INFERENCE_MESSAGE_ROLE_SYSTEM;
+        case "MESSAGE_ROLE_SYSTEM":
+            return MessageRole.MESSAGE_ROLE_SYSTEM;
         case 2:
-        case "INFERENCE_MESSAGE_ROLE_USER":
-            return InferenceMessageRole.INFERENCE_MESSAGE_ROLE_USER;
+        case "MESSAGE_ROLE_USER":
+            return MessageRole.MESSAGE_ROLE_USER;
         case 3:
-        case "INFERENCE_MESSAGE_ROLE_ASSISTANT":
-            return InferenceMessageRole.INFERENCE_MESSAGE_ROLE_ASSISTANT;
+        case "MESSAGE_ROLE_ASSISTANT":
+            return MessageRole.MESSAGE_ROLE_ASSISTANT;
         case 4:
-        case "INFERENCE_MESSAGE_ROLE_TOOL":
-            return InferenceMessageRole.INFERENCE_MESSAGE_ROLE_TOOL;
+        case "MESSAGE_ROLE_TOOL":
+            return MessageRole.MESSAGE_ROLE_TOOL;
         case 5:
-        case "INFERENCE_MESSAGE_ROLE_DEVELOPER":
-            return InferenceMessageRole.INFERENCE_MESSAGE_ROLE_DEVELOPER;
+        case "MESSAGE_ROLE_DEVELOPER":
+            return MessageRole.MESSAGE_ROLE_DEVELOPER;
         case -1:
         case "UNRECOGNIZED":
         default:
-            return InferenceMessageRole.UNRECOGNIZED;
+            return MessageRole.UNRECOGNIZED;
     }
 }
-function inferenceMessageRoleToJSON(object) {
+function messageRoleToJSON(object) {
     switch (object) {
-        case InferenceMessageRole.INFERENCE_MESSAGE_ROLE_UNSPECIFIED:
-            return "INFERENCE_MESSAGE_ROLE_UNSPECIFIED";
-        case InferenceMessageRole.INFERENCE_MESSAGE_ROLE_SYSTEM:
-            return "INFERENCE_MESSAGE_ROLE_SYSTEM";
-        case InferenceMessageRole.INFERENCE_MESSAGE_ROLE_USER:
-            return "INFERENCE_MESSAGE_ROLE_USER";
-        case InferenceMessageRole.INFERENCE_MESSAGE_ROLE_ASSISTANT:
-            return "INFERENCE_MESSAGE_ROLE_ASSISTANT";
-        case InferenceMessageRole.INFERENCE_MESSAGE_ROLE_TOOL:
-            return "INFERENCE_MESSAGE_ROLE_TOOL";
-        case InferenceMessageRole.INFERENCE_MESSAGE_ROLE_DEVELOPER:
-            return "INFERENCE_MESSAGE_ROLE_DEVELOPER";
-        case InferenceMessageRole.UNRECOGNIZED:
+        case MessageRole.MESSAGE_ROLE_UNSPECIFIED:
+            return "MESSAGE_ROLE_UNSPECIFIED";
+        case MessageRole.MESSAGE_ROLE_SYSTEM:
+            return "MESSAGE_ROLE_SYSTEM";
+        case MessageRole.MESSAGE_ROLE_USER:
+            return "MESSAGE_ROLE_USER";
+        case MessageRole.MESSAGE_ROLE_ASSISTANT:
+            return "MESSAGE_ROLE_ASSISTANT";
+        case MessageRole.MESSAGE_ROLE_TOOL:
+            return "MESSAGE_ROLE_TOOL";
+        case MessageRole.MESSAGE_ROLE_DEVELOPER:
+            return "MESSAGE_ROLE_DEVELOPER";
+        case MessageRole.UNRECOGNIZED:
         default:
             return "UNRECOGNIZED";
     }
 }
-/** InferenceFinishReason describes why one inference operation ended. */
-var InferenceFinishReason;
-(function (InferenceFinishReason) {
-    InferenceFinishReason[InferenceFinishReason["INFERENCE_FINISH_REASON_UNSPECIFIED"] = 0] = "INFERENCE_FINISH_REASON_UNSPECIFIED";
-    InferenceFinishReason[InferenceFinishReason["INFERENCE_FINISH_REASON_STOP"] = 1] = "INFERENCE_FINISH_REASON_STOP";
-    InferenceFinishReason[InferenceFinishReason["INFERENCE_FINISH_REASON_LENGTH"] = 2] = "INFERENCE_FINISH_REASON_LENGTH";
-    InferenceFinishReason[InferenceFinishReason["INFERENCE_FINISH_REASON_TOOL_CALLS"] = 3] = "INFERENCE_FINISH_REASON_TOOL_CALLS";
-    InferenceFinishReason[InferenceFinishReason["INFERENCE_FINISH_REASON_CONTENT_FILTER"] = 4] = "INFERENCE_FINISH_REASON_CONTENT_FILTER";
-    InferenceFinishReason[InferenceFinishReason["INFERENCE_FINISH_REASON_ERROR"] = 5] = "INFERENCE_FINISH_REASON_ERROR";
-    InferenceFinishReason[InferenceFinishReason["UNRECOGNIZED"] = -1] = "UNRECOGNIZED";
-})(InferenceFinishReason || (exports.InferenceFinishReason = InferenceFinishReason = {}));
-function inferenceFinishReasonFromJSON(object) {
+/** FinishReason describes why one model call ended. */
+var FinishReason;
+(function (FinishReason) {
+    FinishReason[FinishReason["FINISH_REASON_UNSPECIFIED"] = 0] = "FINISH_REASON_UNSPECIFIED";
+    FinishReason[FinishReason["FINISH_REASON_STOP"] = 1] = "FINISH_REASON_STOP";
+    FinishReason[FinishReason["FINISH_REASON_LENGTH"] = 2] = "FINISH_REASON_LENGTH";
+    FinishReason[FinishReason["FINISH_REASON_TOOL_CALLS"] = 3] = "FINISH_REASON_TOOL_CALLS";
+    FinishReason[FinishReason["FINISH_REASON_CONTENT_FILTER"] = 4] = "FINISH_REASON_CONTENT_FILTER";
+    FinishReason[FinishReason["FINISH_REASON_ERROR"] = 5] = "FINISH_REASON_ERROR";
+    FinishReason[FinishReason["UNRECOGNIZED"] = -1] = "UNRECOGNIZED";
+})(FinishReason || (exports.FinishReason = FinishReason = {}));
+function finishReasonFromJSON(object) {
     switch (object) {
         case 0:
-        case "INFERENCE_FINISH_REASON_UNSPECIFIED":
-            return InferenceFinishReason.INFERENCE_FINISH_REASON_UNSPECIFIED;
+        case "FINISH_REASON_UNSPECIFIED":
+            return FinishReason.FINISH_REASON_UNSPECIFIED;
         case 1:
-        case "INFERENCE_FINISH_REASON_STOP":
-            return InferenceFinishReason.INFERENCE_FINISH_REASON_STOP;
+        case "FINISH_REASON_STOP":
+            return FinishReason.FINISH_REASON_STOP;
         case 2:
-        case "INFERENCE_FINISH_REASON_LENGTH":
-            return InferenceFinishReason.INFERENCE_FINISH_REASON_LENGTH;
+        case "FINISH_REASON_LENGTH":
+            return FinishReason.FINISH_REASON_LENGTH;
         case 3:
-        case "INFERENCE_FINISH_REASON_TOOL_CALLS":
-            return InferenceFinishReason.INFERENCE_FINISH_REASON_TOOL_CALLS;
+        case "FINISH_REASON_TOOL_CALLS":
+            return FinishReason.FINISH_REASON_TOOL_CALLS;
         case 4:
-        case "INFERENCE_FINISH_REASON_CONTENT_FILTER":
-            return InferenceFinishReason.INFERENCE_FINISH_REASON_CONTENT_FILTER;
+        case "FINISH_REASON_CONTENT_FILTER":
+            return FinishReason.FINISH_REASON_CONTENT_FILTER;
         case 5:
-        case "INFERENCE_FINISH_REASON_ERROR":
-            return InferenceFinishReason.INFERENCE_FINISH_REASON_ERROR;
+        case "FINISH_REASON_ERROR":
+            return FinishReason.FINISH_REASON_ERROR;
         case -1:
         case "UNRECOGNIZED":
         default:
-            return InferenceFinishReason.UNRECOGNIZED;
+            return FinishReason.UNRECOGNIZED;
     }
 }
-function inferenceFinishReasonToJSON(object) {
+function finishReasonToJSON(object) {
     switch (object) {
-        case InferenceFinishReason.INFERENCE_FINISH_REASON_UNSPECIFIED:
-            return "INFERENCE_FINISH_REASON_UNSPECIFIED";
-        case InferenceFinishReason.INFERENCE_FINISH_REASON_STOP:
-            return "INFERENCE_FINISH_REASON_STOP";
-        case InferenceFinishReason.INFERENCE_FINISH_REASON_LENGTH:
-            return "INFERENCE_FINISH_REASON_LENGTH";
-        case InferenceFinishReason.INFERENCE_FINISH_REASON_TOOL_CALLS:
-            return "INFERENCE_FINISH_REASON_TOOL_CALLS";
-        case InferenceFinishReason.INFERENCE_FINISH_REASON_CONTENT_FILTER:
-            return "INFERENCE_FINISH_REASON_CONTENT_FILTER";
-        case InferenceFinishReason.INFERENCE_FINISH_REASON_ERROR:
-            return "INFERENCE_FINISH_REASON_ERROR";
-        case InferenceFinishReason.UNRECOGNIZED:
+        case FinishReason.FINISH_REASON_UNSPECIFIED:
+            return "FINISH_REASON_UNSPECIFIED";
+        case FinishReason.FINISH_REASON_STOP:
+            return "FINISH_REASON_STOP";
+        case FinishReason.FINISH_REASON_LENGTH:
+            return "FINISH_REASON_LENGTH";
+        case FinishReason.FINISH_REASON_TOOL_CALLS:
+            return "FINISH_REASON_TOOL_CALLS";
+        case FinishReason.FINISH_REASON_CONTENT_FILTER:
+            return "FINISH_REASON_CONTENT_FILTER";
+        case FinishReason.FINISH_REASON_ERROR:
+            return "FINISH_REASON_ERROR";
+        case FinishReason.UNRECOGNIZED:
         default:
             return "UNRECOGNIZED";
     }
@@ -293,84 +293,84 @@ function imageDeliveryModeToJSON(object) {
             return "UNRECOGNIZED";
     }
 }
-/** InferenceTokenKind classifies one token bucket for usage accounting. */
-var InferenceTokenKind;
-(function (InferenceTokenKind) {
-    InferenceTokenKind[InferenceTokenKind["INFERENCE_TOKEN_KIND_UNSPECIFIED"] = 0] = "INFERENCE_TOKEN_KIND_UNSPECIFIED";
-    InferenceTokenKind[InferenceTokenKind["INFERENCE_TOKEN_KIND_INPUT"] = 1] = "INFERENCE_TOKEN_KIND_INPUT";
-    InferenceTokenKind[InferenceTokenKind["INFERENCE_TOKEN_KIND_INPUT_CACHE_WRITE"] = 2] = "INFERENCE_TOKEN_KIND_INPUT_CACHE_WRITE";
-    InferenceTokenKind[InferenceTokenKind["INFERENCE_TOKEN_KIND_INPUT_CACHE_READ"] = 3] = "INFERENCE_TOKEN_KIND_INPUT_CACHE_READ";
-    InferenceTokenKind[InferenceTokenKind["INFERENCE_TOKEN_KIND_INPUT_AUDIO"] = 4] = "INFERENCE_TOKEN_KIND_INPUT_AUDIO";
-    InferenceTokenKind[InferenceTokenKind["INFERENCE_TOKEN_KIND_INPUT_IMAGE"] = 5] = "INFERENCE_TOKEN_KIND_INPUT_IMAGE";
-    InferenceTokenKind[InferenceTokenKind["INFERENCE_TOKEN_KIND_OUTPUT"] = 6] = "INFERENCE_TOKEN_KIND_OUTPUT";
-    InferenceTokenKind[InferenceTokenKind["INFERENCE_TOKEN_KIND_OUTPUT_REASONING"] = 7] = "INFERENCE_TOKEN_KIND_OUTPUT_REASONING";
-    InferenceTokenKind[InferenceTokenKind["INFERENCE_TOKEN_KIND_OUTPUT_AUDIO"] = 8] = "INFERENCE_TOKEN_KIND_OUTPUT_AUDIO";
-    InferenceTokenKind[InferenceTokenKind["UNRECOGNIZED"] = -1] = "UNRECOGNIZED";
-})(InferenceTokenKind || (exports.InferenceTokenKind = InferenceTokenKind = {}));
-function inferenceTokenKindFromJSON(object) {
+/** TokenKind classifies one token bucket for usage accounting. */
+var TokenKind;
+(function (TokenKind) {
+    TokenKind[TokenKind["TOKEN_KIND_UNSPECIFIED"] = 0] = "TOKEN_KIND_UNSPECIFIED";
+    TokenKind[TokenKind["TOKEN_KIND_INPUT"] = 1] = "TOKEN_KIND_INPUT";
+    TokenKind[TokenKind["TOKEN_KIND_INPUT_CACHE_WRITE"] = 2] = "TOKEN_KIND_INPUT_CACHE_WRITE";
+    TokenKind[TokenKind["TOKEN_KIND_INPUT_CACHE_READ"] = 3] = "TOKEN_KIND_INPUT_CACHE_READ";
+    TokenKind[TokenKind["TOKEN_KIND_INPUT_AUDIO"] = 4] = "TOKEN_KIND_INPUT_AUDIO";
+    TokenKind[TokenKind["TOKEN_KIND_INPUT_IMAGE"] = 5] = "TOKEN_KIND_INPUT_IMAGE";
+    TokenKind[TokenKind["TOKEN_KIND_OUTPUT"] = 6] = "TOKEN_KIND_OUTPUT";
+    TokenKind[TokenKind["TOKEN_KIND_OUTPUT_REASONING"] = 7] = "TOKEN_KIND_OUTPUT_REASONING";
+    TokenKind[TokenKind["TOKEN_KIND_OUTPUT_AUDIO"] = 8] = "TOKEN_KIND_OUTPUT_AUDIO";
+    TokenKind[TokenKind["UNRECOGNIZED"] = -1] = "UNRECOGNIZED";
+})(TokenKind || (exports.TokenKind = TokenKind = {}));
+function tokenKindFromJSON(object) {
     switch (object) {
         case 0:
-        case "INFERENCE_TOKEN_KIND_UNSPECIFIED":
-            return InferenceTokenKind.INFERENCE_TOKEN_KIND_UNSPECIFIED;
+        case "TOKEN_KIND_UNSPECIFIED":
+            return TokenKind.TOKEN_KIND_UNSPECIFIED;
         case 1:
-        case "INFERENCE_TOKEN_KIND_INPUT":
-            return InferenceTokenKind.INFERENCE_TOKEN_KIND_INPUT;
+        case "TOKEN_KIND_INPUT":
+            return TokenKind.TOKEN_KIND_INPUT;
         case 2:
-        case "INFERENCE_TOKEN_KIND_INPUT_CACHE_WRITE":
-            return InferenceTokenKind.INFERENCE_TOKEN_KIND_INPUT_CACHE_WRITE;
+        case "TOKEN_KIND_INPUT_CACHE_WRITE":
+            return TokenKind.TOKEN_KIND_INPUT_CACHE_WRITE;
         case 3:
-        case "INFERENCE_TOKEN_KIND_INPUT_CACHE_READ":
-            return InferenceTokenKind.INFERENCE_TOKEN_KIND_INPUT_CACHE_READ;
+        case "TOKEN_KIND_INPUT_CACHE_READ":
+            return TokenKind.TOKEN_KIND_INPUT_CACHE_READ;
         case 4:
-        case "INFERENCE_TOKEN_KIND_INPUT_AUDIO":
-            return InferenceTokenKind.INFERENCE_TOKEN_KIND_INPUT_AUDIO;
+        case "TOKEN_KIND_INPUT_AUDIO":
+            return TokenKind.TOKEN_KIND_INPUT_AUDIO;
         case 5:
-        case "INFERENCE_TOKEN_KIND_INPUT_IMAGE":
-            return InferenceTokenKind.INFERENCE_TOKEN_KIND_INPUT_IMAGE;
+        case "TOKEN_KIND_INPUT_IMAGE":
+            return TokenKind.TOKEN_KIND_INPUT_IMAGE;
         case 6:
-        case "INFERENCE_TOKEN_KIND_OUTPUT":
-            return InferenceTokenKind.INFERENCE_TOKEN_KIND_OUTPUT;
+        case "TOKEN_KIND_OUTPUT":
+            return TokenKind.TOKEN_KIND_OUTPUT;
         case 7:
-        case "INFERENCE_TOKEN_KIND_OUTPUT_REASONING":
-            return InferenceTokenKind.INFERENCE_TOKEN_KIND_OUTPUT_REASONING;
+        case "TOKEN_KIND_OUTPUT_REASONING":
+            return TokenKind.TOKEN_KIND_OUTPUT_REASONING;
         case 8:
-        case "INFERENCE_TOKEN_KIND_OUTPUT_AUDIO":
-            return InferenceTokenKind.INFERENCE_TOKEN_KIND_OUTPUT_AUDIO;
+        case "TOKEN_KIND_OUTPUT_AUDIO":
+            return TokenKind.TOKEN_KIND_OUTPUT_AUDIO;
         case -1:
         case "UNRECOGNIZED":
         default:
-            return InferenceTokenKind.UNRECOGNIZED;
+            return TokenKind.UNRECOGNIZED;
     }
 }
-function inferenceTokenKindToJSON(object) {
+function tokenKindToJSON(object) {
     switch (object) {
-        case InferenceTokenKind.INFERENCE_TOKEN_KIND_UNSPECIFIED:
-            return "INFERENCE_TOKEN_KIND_UNSPECIFIED";
-        case InferenceTokenKind.INFERENCE_TOKEN_KIND_INPUT:
-            return "INFERENCE_TOKEN_KIND_INPUT";
-        case InferenceTokenKind.INFERENCE_TOKEN_KIND_INPUT_CACHE_WRITE:
-            return "INFERENCE_TOKEN_KIND_INPUT_CACHE_WRITE";
-        case InferenceTokenKind.INFERENCE_TOKEN_KIND_INPUT_CACHE_READ:
-            return "INFERENCE_TOKEN_KIND_INPUT_CACHE_READ";
-        case InferenceTokenKind.INFERENCE_TOKEN_KIND_INPUT_AUDIO:
-            return "INFERENCE_TOKEN_KIND_INPUT_AUDIO";
-        case InferenceTokenKind.INFERENCE_TOKEN_KIND_INPUT_IMAGE:
-            return "INFERENCE_TOKEN_KIND_INPUT_IMAGE";
-        case InferenceTokenKind.INFERENCE_TOKEN_KIND_OUTPUT:
-            return "INFERENCE_TOKEN_KIND_OUTPUT";
-        case InferenceTokenKind.INFERENCE_TOKEN_KIND_OUTPUT_REASONING:
-            return "INFERENCE_TOKEN_KIND_OUTPUT_REASONING";
-        case InferenceTokenKind.INFERENCE_TOKEN_KIND_OUTPUT_AUDIO:
-            return "INFERENCE_TOKEN_KIND_OUTPUT_AUDIO";
-        case InferenceTokenKind.UNRECOGNIZED:
+        case TokenKind.TOKEN_KIND_UNSPECIFIED:
+            return "TOKEN_KIND_UNSPECIFIED";
+        case TokenKind.TOKEN_KIND_INPUT:
+            return "TOKEN_KIND_INPUT";
+        case TokenKind.TOKEN_KIND_INPUT_CACHE_WRITE:
+            return "TOKEN_KIND_INPUT_CACHE_WRITE";
+        case TokenKind.TOKEN_KIND_INPUT_CACHE_READ:
+            return "TOKEN_KIND_INPUT_CACHE_READ";
+        case TokenKind.TOKEN_KIND_INPUT_AUDIO:
+            return "TOKEN_KIND_INPUT_AUDIO";
+        case TokenKind.TOKEN_KIND_INPUT_IMAGE:
+            return "TOKEN_KIND_INPUT_IMAGE";
+        case TokenKind.TOKEN_KIND_OUTPUT:
+            return "TOKEN_KIND_OUTPUT";
+        case TokenKind.TOKEN_KIND_OUTPUT_REASONING:
+            return "TOKEN_KIND_OUTPUT_REASONING";
+        case TokenKind.TOKEN_KIND_OUTPUT_AUDIO:
+            return "TOKEN_KIND_OUTPUT_AUDIO";
+        case TokenKind.UNRECOGNIZED:
         default:
             return "UNRECOGNIZED";
     }
 }
-function createBaseInferenceRequestContext() {
+function createBaseRequestContext() {
     return { organizationId: "", approvedModelId: "", idempotencyKey: "", metadata: undefined };
 }
-exports.InferenceRequestContext = {
+exports.RequestContext = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.organizationId !== "") {
             writer.uint32(10).string(message.organizationId);
@@ -389,7 +389,7 @@ exports.InferenceRequestContext = {
     decode(input, length) {
         const reader = input instanceof wire_1.BinaryReader ? input : new wire_1.BinaryReader(input);
         const end = length === undefined ? reader.len : reader.pos + length;
-        const message = createBaseInferenceRequestContext();
+        const message = createBaseRequestContext();
         while (reader.pos < end) {
             const tag = reader.uint32();
             switch (tag >>> 3) {
@@ -466,11 +466,11 @@ exports.InferenceRequestContext = {
         return obj;
     },
     create(base) {
-        return exports.InferenceRequestContext.fromPartial(base !== null && base !== void 0 ? base : {});
+        return exports.RequestContext.fromPartial(base !== null && base !== void 0 ? base : {});
     },
     fromPartial(object) {
         var _a, _b, _c, _d;
-        const message = createBaseInferenceRequestContext();
+        const message = createBaseRequestContext();
         message.organizationId = (_a = object.organizationId) !== null && _a !== void 0 ? _a : "";
         message.approvedModelId = (_b = object.approvedModelId) !== null && _b !== void 0 ? _b : "";
         message.idempotencyKey = (_c = object.idempotencyKey) !== null && _c !== void 0 ? _c : "";
@@ -478,20 +478,20 @@ exports.InferenceRequestContext = {
         return message;
     },
 };
-function createBaseResolvedInferenceContext() {
+function createBaseResolvedModelRoute() {
     return {
         organizationId: "",
         approvedModelId: "",
         providerConfigurationId: "",
         providerKind: 0,
         resolvedModel: "",
-        operationType: 0,
+        operation: 0,
         requestId: "",
         upstreamRequestId: "",
         requestedAt: undefined,
     };
 }
-exports.ResolvedInferenceContext = {
+exports.ResolvedModelRoute = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.organizationId !== "") {
             writer.uint32(10).string(message.organizationId);
@@ -508,8 +508,8 @@ exports.ResolvedInferenceContext = {
         if (message.resolvedModel !== "") {
             writer.uint32(42).string(message.resolvedModel);
         }
-        if (message.operationType !== 0) {
-            writer.uint32(48).int32(message.operationType);
+        if (message.operation !== 0) {
+            writer.uint32(48).int32(message.operation);
         }
         if (message.requestId !== "") {
             writer.uint32(58).string(message.requestId);
@@ -525,7 +525,7 @@ exports.ResolvedInferenceContext = {
     decode(input, length) {
         const reader = input instanceof wire_1.BinaryReader ? input : new wire_1.BinaryReader(input);
         const end = length === undefined ? reader.len : reader.pos + length;
-        const message = createBaseResolvedInferenceContext();
+        const message = createBaseResolvedModelRoute();
         while (reader.pos < end) {
             const tag = reader.uint32();
             switch (tag >>> 3) {
@@ -568,7 +568,7 @@ exports.ResolvedInferenceContext = {
                     if (tag !== 48) {
                         break;
                     }
-                    message.operationType = reader.int32();
+                    message.operation = reader.int32();
                     continue;
                 }
                 case 7: {
@@ -627,11 +627,7 @@ exports.ResolvedInferenceContext = {
                 : isSet(object.resolved_model)
                     ? globalThis.String(object.resolved_model)
                     : "",
-            operationType: isSet(object.operationType)
-                ? inferenceOperationTypeFromJSON(object.operationType)
-                : isSet(object.operation_type)
-                    ? inferenceOperationTypeFromJSON(object.operation_type)
-                    : 0,
+            operation: isSet(object.operation) ? modelOperationFromJSON(object.operation) : 0,
             requestId: isSet(object.requestId)
                 ? globalThis.String(object.requestId)
                 : isSet(object.request_id)
@@ -666,8 +662,8 @@ exports.ResolvedInferenceContext = {
         if (message.resolvedModel !== "") {
             obj.resolvedModel = message.resolvedModel;
         }
-        if (message.operationType !== 0) {
-            obj.operationType = inferenceOperationTypeToJSON(message.operationType);
+        if (message.operation !== 0) {
+            obj.operation = modelOperationToJSON(message.operation);
         }
         if (message.requestId !== "") {
             obj.requestId = message.requestId;
@@ -681,27 +677,27 @@ exports.ResolvedInferenceContext = {
         return obj;
     },
     create(base) {
-        return exports.ResolvedInferenceContext.fromPartial(base !== null && base !== void 0 ? base : {});
+        return exports.ResolvedModelRoute.fromPartial(base !== null && base !== void 0 ? base : {});
     },
     fromPartial(object) {
         var _a, _b, _c, _d, _e, _f, _g, _h, _j;
-        const message = createBaseResolvedInferenceContext();
+        const message = createBaseResolvedModelRoute();
         message.organizationId = (_a = object.organizationId) !== null && _a !== void 0 ? _a : "";
         message.approvedModelId = (_b = object.approvedModelId) !== null && _b !== void 0 ? _b : "";
         message.providerConfigurationId = (_c = object.providerConfigurationId) !== null && _c !== void 0 ? _c : "";
         message.providerKind = (_d = object.providerKind) !== null && _d !== void 0 ? _d : 0;
         message.resolvedModel = (_e = object.resolvedModel) !== null && _e !== void 0 ? _e : "";
-        message.operationType = (_f = object.operationType) !== null && _f !== void 0 ? _f : 0;
+        message.operation = (_f = object.operation) !== null && _f !== void 0 ? _f : 0;
         message.requestId = (_g = object.requestId) !== null && _g !== void 0 ? _g : "";
         message.upstreamRequestId = (_h = object.upstreamRequestId) !== null && _h !== void 0 ? _h : "";
         message.requestedAt = (_j = object.requestedAt) !== null && _j !== void 0 ? _j : undefined;
         return message;
     },
 };
-function createBaseInferenceTokenUsageBreakdown() {
+function createBaseTokenUsageBreakdown() {
     return { kind: 0, tokenCount: 0, providerMetricName: "" };
 }
-exports.InferenceTokenUsageBreakdown = {
+exports.TokenUsageBreakdown = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.kind !== 0) {
             writer.uint32(8).int32(message.kind);
@@ -717,7 +713,7 @@ exports.InferenceTokenUsageBreakdown = {
     decode(input, length) {
         const reader = input instanceof wire_1.BinaryReader ? input : new wire_1.BinaryReader(input);
         const end = length === undefined ? reader.len : reader.pos + length;
-        const message = createBaseInferenceTokenUsageBreakdown();
+        const message = createBaseTokenUsageBreakdown();
         while (reader.pos < end) {
             const tag = reader.uint32();
             switch (tag >>> 3) {
@@ -752,7 +748,7 @@ exports.InferenceTokenUsageBreakdown = {
     },
     fromJSON(object) {
         return {
-            kind: isSet(object.kind) ? inferenceTokenKindFromJSON(object.kind) : 0,
+            kind: isSet(object.kind) ? tokenKindFromJSON(object.kind) : 0,
             tokenCount: isSet(object.tokenCount)
                 ? globalThis.Number(object.tokenCount)
                 : isSet(object.token_count)
@@ -768,7 +764,7 @@ exports.InferenceTokenUsageBreakdown = {
     toJSON(message) {
         const obj = {};
         if (message.kind !== 0) {
-            obj.kind = inferenceTokenKindToJSON(message.kind);
+            obj.kind = tokenKindToJSON(message.kind);
         }
         if (message.tokenCount !== 0) {
             obj.tokenCount = Math.round(message.tokenCount);
@@ -779,21 +775,21 @@ exports.InferenceTokenUsageBreakdown = {
         return obj;
     },
     create(base) {
-        return exports.InferenceTokenUsageBreakdown.fromPartial(base !== null && base !== void 0 ? base : {});
+        return exports.TokenUsageBreakdown.fromPartial(base !== null && base !== void 0 ? base : {});
     },
     fromPartial(object) {
         var _a, _b, _c;
-        const message = createBaseInferenceTokenUsageBreakdown();
+        const message = createBaseTokenUsageBreakdown();
         message.kind = (_a = object.kind) !== null && _a !== void 0 ? _a : 0;
         message.tokenCount = (_b = object.tokenCount) !== null && _b !== void 0 ? _b : 0;
         message.providerMetricName = (_c = object.providerMetricName) !== null && _c !== void 0 ? _c : "";
         return message;
     },
 };
-function createBaseInferenceUsage() {
+function createBaseUsage() {
     return { inputTokens: 0, outputTokens: 0, totalTokens: 0, tokenBreakdown: [], providerUsageRaw: undefined };
 }
-exports.InferenceUsage = {
+exports.Usage = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.inputTokens !== 0) {
             writer.uint32(8).int64(message.inputTokens);
@@ -805,7 +801,7 @@ exports.InferenceUsage = {
             writer.uint32(24).int64(message.totalTokens);
         }
         for (const v of message.tokenBreakdown) {
-            exports.InferenceTokenUsageBreakdown.encode(v, writer.uint32(34).fork()).join();
+            exports.TokenUsageBreakdown.encode(v, writer.uint32(34).fork()).join();
         }
         if (message.providerUsageRaw !== undefined) {
             struct_pb_1.Struct.encode(struct_pb_1.Struct.wrap(message.providerUsageRaw), writer.uint32(42).fork()).join();
@@ -815,7 +811,7 @@ exports.InferenceUsage = {
     decode(input, length) {
         const reader = input instanceof wire_1.BinaryReader ? input : new wire_1.BinaryReader(input);
         const end = length === undefined ? reader.len : reader.pos + length;
-        const message = createBaseInferenceUsage();
+        const message = createBaseUsage();
         while (reader.pos < end) {
             const tag = reader.uint32();
             switch (tag >>> 3) {
@@ -844,7 +840,7 @@ exports.InferenceUsage = {
                     if (tag !== 34) {
                         break;
                     }
-                    message.tokenBreakdown.push(exports.InferenceTokenUsageBreakdown.decode(reader, reader.uint32()));
+                    message.tokenBreakdown.push(exports.TokenUsageBreakdown.decode(reader, reader.uint32()));
                     continue;
                 }
                 case 5: {
@@ -880,9 +876,9 @@ exports.InferenceUsage = {
                     ? globalThis.Number(object.total_tokens)
                     : 0,
             tokenBreakdown: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.tokenBreakdown)
-                ? object.tokenBreakdown.map((e) => exports.InferenceTokenUsageBreakdown.fromJSON(e))
+                ? object.tokenBreakdown.map((e) => exports.TokenUsageBreakdown.fromJSON(e))
                 : globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.token_breakdown)
-                    ? object.token_breakdown.map((e) => exports.InferenceTokenUsageBreakdown.fromJSON(e))
+                    ? object.token_breakdown.map((e) => exports.TokenUsageBreakdown.fromJSON(e))
                     : [],
             providerUsageRaw: isObject(object.providerUsageRaw)
                 ? object.providerUsageRaw
@@ -904,7 +900,7 @@ exports.InferenceUsage = {
             obj.totalTokens = Math.round(message.totalTokens);
         }
         if ((_a = message.tokenBreakdown) === null || _a === void 0 ? void 0 : _a.length) {
-            obj.tokenBreakdown = message.tokenBreakdown.map((e) => exports.InferenceTokenUsageBreakdown.toJSON(e));
+            obj.tokenBreakdown = message.tokenBreakdown.map((e) => exports.TokenUsageBreakdown.toJSON(e));
         }
         if (message.providerUsageRaw !== undefined) {
             obj.providerUsageRaw = message.providerUsageRaw;
@@ -912,21 +908,21 @@ exports.InferenceUsage = {
         return obj;
     },
     create(base) {
-        return exports.InferenceUsage.fromPartial(base !== null && base !== void 0 ? base : {});
+        return exports.Usage.fromPartial(base !== null && base !== void 0 ? base : {});
     },
     fromPartial(object) {
         var _a;
         var _b, _c, _d, _e;
-        const message = createBaseInferenceUsage();
+        const message = createBaseUsage();
         message.inputTokens = (_b = object.inputTokens) !== null && _b !== void 0 ? _b : 0;
         message.outputTokens = (_c = object.outputTokens) !== null && _c !== void 0 ? _c : 0;
         message.totalTokens = (_d = object.totalTokens) !== null && _d !== void 0 ? _d : 0;
-        message.tokenBreakdown = ((_a = object.tokenBreakdown) === null || _a === void 0 ? void 0 : _a.map((e) => exports.InferenceTokenUsageBreakdown.fromPartial(e))) || [];
+        message.tokenBreakdown = ((_a = object.tokenBreakdown) === null || _a === void 0 ? void 0 : _a.map((e) => exports.TokenUsageBreakdown.fromPartial(e))) || [];
         message.providerUsageRaw = (_e = object.providerUsageRaw) !== null && _e !== void 0 ? _e : undefined;
         return message;
     },
 };
-function createBaseInferenceGenerationConfig() {
+function createBaseGenerationConfig() {
     return {
         temperature: 0,
         maxOutputTokens: 0,
@@ -940,7 +936,7 @@ function createBaseInferenceGenerationConfig() {
         metadata: undefined,
     };
 }
-exports.InferenceGenerationConfig = {
+exports.GenerationConfig = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.temperature !== 0) {
             writer.uint32(9).double(message.temperature);
@@ -977,7 +973,7 @@ exports.InferenceGenerationConfig = {
     decode(input, length) {
         const reader = input instanceof wire_1.BinaryReader ? input : new wire_1.BinaryReader(input);
         const end = length === undefined ? reader.len : reader.pos + length;
-        const message = createBaseInferenceGenerationConfig();
+        const message = createBaseGenerationConfig();
         while (reader.pos < end) {
             const tag = reader.uint32();
             switch (tag >>> 3) {
@@ -1137,12 +1133,12 @@ exports.InferenceGenerationConfig = {
         return obj;
     },
     create(base) {
-        return exports.InferenceGenerationConfig.fromPartial(base !== null && base !== void 0 ? base : {});
+        return exports.GenerationConfig.fromPartial(base !== null && base !== void 0 ? base : {});
     },
     fromPartial(object) {
         var _a;
         var _b, _c, _d, _e, _f, _g, _h, _j, _k;
-        const message = createBaseInferenceGenerationConfig();
+        const message = createBaseGenerationConfig();
         message.temperature = (_b = object.temperature) !== null && _b !== void 0 ? _b : 0;
         message.maxOutputTokens = (_c = object.maxOutputTokens) !== null && _c !== void 0 ? _c : 0;
         message.stopSequences = ((_a = object.stopSequences) === null || _a === void 0 ? void 0 : _a.map((e) => e)) || [];
@@ -1337,7 +1333,7 @@ exports.ToolDefinition = {
     },
 };
 function createBaseToolCall() {
-    return { id: "", name: "", arguments: undefined };
+    return { id: "", name: "", argumentsJson: "" };
 }
 exports.ToolCall = {
     encode(message, writer = new wire_1.BinaryWriter()) {
@@ -1347,8 +1343,8 @@ exports.ToolCall = {
         if (message.name !== "") {
             writer.uint32(18).string(message.name);
         }
-        if (message.arguments !== undefined) {
-            struct_pb_1.Struct.encode(struct_pb_1.Struct.wrap(message.arguments), writer.uint32(26).fork()).join();
+        if (message.argumentsJson !== "") {
+            writer.uint32(26).string(message.argumentsJson);
         }
         return writer;
     },
@@ -1377,7 +1373,7 @@ exports.ToolCall = {
                     if (tag !== 26) {
                         break;
                     }
-                    message.arguments = struct_pb_1.Struct.unwrap(struct_pb_1.Struct.decode(reader, reader.uint32()));
+                    message.argumentsJson = reader.string();
                     continue;
                 }
             }
@@ -1392,7 +1388,11 @@ exports.ToolCall = {
         return {
             id: isSet(object.id) ? globalThis.String(object.id) : "",
             name: isSet(object.name) ? globalThis.String(object.name) : "",
-            arguments: isObject(object.arguments) ? object.arguments : undefined,
+            argumentsJson: isSet(object.argumentsJson)
+                ? globalThis.String(object.argumentsJson)
+                : isSet(object.arguments_json)
+                    ? globalThis.String(object.arguments_json)
+                    : "",
         };
     },
     toJSON(message) {
@@ -1403,8 +1403,8 @@ exports.ToolCall = {
         if (message.name !== "") {
             obj.name = message.name;
         }
-        if (message.arguments !== undefined) {
-            obj.arguments = message.arguments;
+        if (message.argumentsJson !== "") {
+            obj.argumentsJson = message.argumentsJson;
         }
         return obj;
     },
@@ -1416,7 +1416,114 @@ exports.ToolCall = {
         const message = createBaseToolCall();
         message.id = (_a = object.id) !== null && _a !== void 0 ? _a : "";
         message.name = (_b = object.name) !== null && _b !== void 0 ? _b : "";
-        message.arguments = (_c = object.arguments) !== null && _c !== void 0 ? _c : undefined;
+        message.argumentsJson = (_c = object.argumentsJson) !== null && _c !== void 0 ? _c : "";
+        return message;
+    },
+};
+function createBaseToolCallDelta() {
+    return { id: "", name: "", argumentsDelta: "", argumentsJson: "" };
+}
+exports.ToolCallDelta = {
+    encode(message, writer = new wire_1.BinaryWriter()) {
+        if (message.id !== "") {
+            writer.uint32(10).string(message.id);
+        }
+        if (message.name !== "") {
+            writer.uint32(18).string(message.name);
+        }
+        if (message.argumentsDelta !== "") {
+            writer.uint32(26).string(message.argumentsDelta);
+        }
+        if (message.argumentsJson !== "") {
+            writer.uint32(34).string(message.argumentsJson);
+        }
+        return writer;
+    },
+    decode(input, length) {
+        const reader = input instanceof wire_1.BinaryReader ? input : new wire_1.BinaryReader(input);
+        const end = length === undefined ? reader.len : reader.pos + length;
+        const message = createBaseToolCallDelta();
+        while (reader.pos < end) {
+            const tag = reader.uint32();
+            switch (tag >>> 3) {
+                case 1: {
+                    if (tag !== 10) {
+                        break;
+                    }
+                    message.id = reader.string();
+                    continue;
+                }
+                case 2: {
+                    if (tag !== 18) {
+                        break;
+                    }
+                    message.name = reader.string();
+                    continue;
+                }
+                case 3: {
+                    if (tag !== 26) {
+                        break;
+                    }
+                    message.argumentsDelta = reader.string();
+                    continue;
+                }
+                case 4: {
+                    if (tag !== 34) {
+                        break;
+                    }
+                    message.argumentsJson = reader.string();
+                    continue;
+                }
+            }
+            if ((tag & 7) === 4 || tag === 0) {
+                break;
+            }
+            reader.skip(tag & 7);
+        }
+        return message;
+    },
+    fromJSON(object) {
+        return {
+            id: isSet(object.id) ? globalThis.String(object.id) : "",
+            name: isSet(object.name) ? globalThis.String(object.name) : "",
+            argumentsDelta: isSet(object.argumentsDelta)
+                ? globalThis.String(object.argumentsDelta)
+                : isSet(object.arguments_delta)
+                    ? globalThis.String(object.arguments_delta)
+                    : "",
+            argumentsJson: isSet(object.argumentsJson)
+                ? globalThis.String(object.argumentsJson)
+                : isSet(object.arguments_json)
+                    ? globalThis.String(object.arguments_json)
+                    : "",
+        };
+    },
+    toJSON(message) {
+        const obj = {};
+        if (message.id !== "") {
+            obj.id = message.id;
+        }
+        if (message.name !== "") {
+            obj.name = message.name;
+        }
+        if (message.argumentsDelta !== "") {
+            obj.argumentsDelta = message.argumentsDelta;
+        }
+        if (message.argumentsJson !== "") {
+            obj.argumentsJson = message.argumentsJson;
+        }
+        return obj;
+    },
+    create(base) {
+        return exports.ToolCallDelta.fromPartial(base !== null && base !== void 0 ? base : {});
+    },
+    fromPartial(object) {
+        var _a, _b, _c, _d;
+        const message = createBaseToolCallDelta();
+        message.id = (_a = object.id) !== null && _a !== void 0 ? _a : "";
+        message.name = (_b = object.name) !== null && _b !== void 0 ? _b : "";
+        message.argumentsDelta = (_c = object.argumentsDelta) !== null && _c !== void 0 ? _c : "";
+        message.argumentsJson = (_d = object.argumentsJson) !== null && _d !== void 0 ? _d : "";
         return message;
     },
 };
@@ -1527,7 +1634,7 @@ exports.ToolResult = {
         return message;
     },
 };
-function createBaseInferenceContentPart() {
+function createBaseContentPart() {
     return {
         text: undefined,
         imageUrl: undefined,
@@ -1538,7 +1645,7 @@ function createBaseInferenceContentPart() {
         mimeType: "",
     };
 }
-exports.InferenceContentPart = {
+exports.ContentPart = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.text !== undefined) {
             writer.uint32(10).string(message.text);
@@ -1566,7 +1673,7 @@ exports.InferenceContentPart = {
     decode(input, length) {
         const reader = input instanceof wire_1.BinaryReader ? input : new wire_1.BinaryReader(input);
         const end = length === undefined ? reader.len : reader.pos + length;
-        const message = createBaseInferenceContentPart();
+        const message = createBaseContentPart();
         while (reader.pos < end) {
             const tag = reader.uint32();
             switch (tag >>> 3) {
@@ -1688,11 +1795,11 @@ exports.InferenceContentPart = {
         return obj;
     },
     create(base) {
-        return exports.InferenceContentPart.fromPartial(base !== null && base !== void 0 ? base : {});
+        return exports.ContentPart.fromPartial(base !== null && base !== void 0 ? base : {});
     },
     fromPartial(object) {
         var _a, _b, _c, _d, _e;
-        const message = createBaseInferenceContentPart();
+        const message = createBaseContentPart();
         message.text = (_a = object.text) !== null && _a !== void 0 ? _a : undefined;
         message.imageUrl = (_b = object.imageUrl) !== null && _b !== void 0 ? _b : undefined;
         message.binaryData = (_c = object.binaryData) !== null && _c !== void 0 ? _c : undefined;
@@ -1716,7 +1823,7 @@ exports.ChatMessage = {
             writer.uint32(8).int32(message.role);
         }
         for (const v of message.content) {
-            exports.InferenceContentPart.encode(v, writer.uint32(18).fork()).join();
+            exports.ContentPart.encode(v, writer.uint32(18).fork()).join();
         }
         if (message.participantName !== "") {
             writer.uint32(26).string(message.participantName);
@@ -1744,7 +1851,7 @@ exports.ChatMessage = {
                     if (tag !== 18) {
                         break;
                     }
-                    message.content.push(exports.InferenceContentPart.decode(reader, reader.uint32()));
+                    message.content.push(exports.ContentPart.decode(reader, reader.uint32()));
                     continue;
                 }
                 case 3: {
@@ -1771,10 +1878,8 @@ exports.ChatMessage = {
     },
     fromJSON(object) {
         return {
-            role: isSet(object.role) ? inferenceMessageRoleFromJSON(object.role) : 0,
-            content: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.content)
-                ? object.content.map((e) => exports.InferenceContentPart.fromJSON(e))
-                : [],
+            role: isSet(object.role) ? messageRoleFromJSON(object.role) : 0,
+            content: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.content) ? object.content.map((e) => exports.ContentPart.fromJSON(e)) : [],
             participantName: isSet(object.participantName)
                 ? globalThis.String(object.participantName)
                 : isSet(object.participant_name)
@@ -1791,10 +1896,10 @@ exports.ChatMessage = {
         var _a;
         const obj = {};
         if (message.role !== 0) {
-            obj.role = inferenceMessageRoleToJSON(message.role);
+            obj.role = messageRoleToJSON(message.role);
         }
         if ((_a = message.content) === null || _a === void 0 ? void 0 : _a.length) {
-            obj.content = message.content.map((e) => exports.InferenceContentPart.toJSON(e));
+            obj.content = message.content.map((e) => exports.ContentPart.toJSON(e));
         }
         if (message.participantName !== "") {
             obj.participantName = message.participantName;
@@ -1812,7 +1917,7 @@ exports.ChatMessage = {
         var _b, _c, _d;
         const message = createBaseChatMessage();
         message.role = (_b = object.role) !== null && _b !== void 0 ? _b : 0;
-        message.content = ((_a = object.content) === null || _a === void 0 ? void 0 : _a.map((e) => exports.InferenceContentPart.fromPartial(e))) || [];
+        message.content = ((_a = object.content) === null || _a === void 0 ? void 0 : _a.map((e) => exports.ContentPart.fromPartial(e))) || [];
         message.participantName = (_c = object.participantName) !== null && _c !== void 0 ? _c : "";
         message.toolCallId = (_d = object.toolCallId) !== null && _d !== void 0 ? _d : "";
         return message;
@@ -2154,7 +2259,7 @@ function createBaseCreateChatCompletionRequest() {
 exports.CreateChatCompletionRequest = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.context !== undefined) {
-            exports.InferenceRequestContext.encode(message.context, writer.uint32(10).fork()).join();
+            exports.RequestContext.encode(message.context, writer.uint32(10).fork()).join();
         }
         if (message.model !== "") {
             writer.uint32(18).string(message.model);
@@ -2172,7 +2277,7 @@ exports.CreateChatCompletionRequest = {
             writer.uint32(48).int32(message.toolChoiceMode);
         }
         if (message.generationConfig !== undefined) {
-            exports.InferenceGenerationConfig.encode(message.generationConfig, writer.uint32(58).fork()).join();
+            exports.GenerationConfig.encode(message.generationConfig, writer.uint32(58).fork()).join();
         }
         if (message.responseSchema !== undefined) {
             exports.StructuredOutputSchema.encode(message.responseSchema, writer.uint32(66).fork()).join();
@@ -2190,7 +2295,7 @@ exports.CreateChatCompletionRequest = {
                     if (tag !== 10) {
                         break;
                     }
-                    message.context = exports.InferenceRequestContext.decode(reader, reader.uint32());
+                    message.context = exports.RequestContext.decode(reader, reader.uint32());
                     continue;
                 }
                 case 2: {
@@ -2232,7 +2337,7 @@ exports.CreateChatCompletionRequest = {
                     if (tag !== 58) {
                         break;
                     }
-                    message.generationConfig = exports.InferenceGenerationConfig.decode(reader, reader.uint32());
+                    message.generationConfig = exports.GenerationConfig.decode(reader, reader.uint32());
                     continue;
                 }
                 case 8: {
@@ -2252,7 +2357,7 @@ exports.CreateChatCompletionRequest = {
     },
     fromJSON(object) {
         return {
-            context: isSet(object.context) ? exports.InferenceRequestContext.fromJSON(object.context) : undefined,
+            context: isSet(object.context) ? exports.RequestContext.fromJSON(object.context) : undefined,
             model: isSet(object.model) ? globalThis.String(object.model) : "",
             messages: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.messages)
                 ? object.messages.map((e) => exports.ChatMessage.fromJSON(e))
@@ -2269,9 +2374,9 @@ exports.CreateChatCompletionRequest = {
                     ? toolChoiceModeFromJSON(object.tool_choice_mode)
                     : 0,
             generationConfig: isSet(object.generationConfig)
-                ? exports.InferenceGenerationConfig.fromJSON(object.generationConfig)
+                ? exports.GenerationConfig.fromJSON(object.generationConfig)
                 : isSet(object.generation_config)
-                    ? exports.InferenceGenerationConfig.fromJSON(object.generation_config)
+                    ? exports.GenerationConfig.fromJSON(object.generation_config)
                     : undefined,
             responseSchema: isSet(object.responseSchema)
                 ? exports.StructuredOutputSchema.fromJSON(object.responseSchema)
@@ -2284,7 +2389,7 @@ exports.CreateChatCompletionRequest = {
         var _a, _b;
         const obj = {};
         if (message.context !== undefined) {
-            obj.context = exports.InferenceRequestContext.toJSON(message.context);
+            obj.context = exports.RequestContext.toJSON(message.context);
         }
         if (message.model !== "") {
             obj.model = message.model;
@@ -2302,7 +2407,7 @@ exports.CreateChatCompletionRequest = {
             obj.toolChoiceMode = toolChoiceModeToJSON(message.toolChoiceMode);
         }
         if (message.generationConfig !== undefined) {
-            obj.generationConfig = exports.InferenceGenerationConfig.toJSON(message.generationConfig);
+            obj.generationConfig = exports.GenerationConfig.toJSON(message.generationConfig);
         }
         if (message.responseSchema !== undefined) {
             obj.responseSchema = exports.StructuredOutputSchema.toJSON(message.responseSchema);
@@ -2317,7 +2422,7 @@ exports.CreateChatCompletionRequest = {
         var _c, _d, _e;
         const message = createBaseCreateChatCompletionRequest();
         message.context = (object.context !== undefined && object.context !== null)
-            ? exports.InferenceRequestContext.fromPartial(object.context)
+            ? exports.RequestContext.fromPartial(object.context)
             : undefined;
         message.model = (_c = object.model) !== null && _c !== void 0 ? _c : "";
         message.messages = ((_a = object.messages) === null || _a === void 0 ? void 0 : _a.map((e) => exports.ChatMessage.fromPartial(e))) || [];
@@ -2325,7 +2430,7 @@ exports.CreateChatCompletionRequest = {
         message.tools = ((_b = object.tools) === null || _b === void 0 ? void 0 : _b.map((e) => exports.ToolDefinition.fromPartial(e))) || [];
         message.toolChoiceMode = (_e = object.toolChoiceMode) !== null && _e !== void 0 ? _e : 0;
         message.generationConfig = (object.generationConfig !== undefined && object.generationConfig !== null)
-            ? exports.InferenceGenerationConfig.fromPartial(object.generationConfig)
+            ? exports.GenerationConfig.fromPartial(object.generationConfig)
             : undefined;
         message.responseSchema = (object.responseSchema !== undefined && object.responseSchema !== null)
             ? exports.StructuredOutputSchema.fromPartial(object.responseSchema)
@@ -2346,7 +2451,7 @@ function createBaseCreateChatCompletionResponse() {
 exports.CreateChatCompletionResponse = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.context !== undefined) {
-            exports.ResolvedInferenceContext.encode(message.context, writer.uint32(10).fork()).join();
+            exports.ResolvedModelRoute.encode(message.context, writer.uint32(10).fork()).join();
         }
         if (message.message !== undefined) {
             exports.ChatMessage.encode(message.message, writer.uint32(18).fork()).join();
@@ -2358,7 +2463,7 @@ exports.CreateChatCompletionResponse = {
             writer.uint32(32).int32(message.finishReason);
         }
         if (message.usage !== undefined) {
-            exports.InferenceUsage.encode(message.usage, writer.uint32(42).fork()).join();
+            exports.Usage.encode(message.usage, writer.uint32(42).fork()).join();
         }
         if (message.structuredOutput !== undefined) {
             struct_pb_1.Struct.encode(struct_pb_1.Struct.wrap(message.structuredOutput), writer.uint32(50).fork()).join();
@@ -2376,7 +2481,7 @@ exports.CreateChatCompletionResponse = {
                     if (tag !== 10) {
                         break;
                     }
-                    message.context = exports.ResolvedInferenceContext.decode(reader, reader.uint32());
+                    message.context = exports.ResolvedModelRoute.decode(reader, reader.uint32());
                     continue;
                 }
                 case 2: {
@@ -2404,7 +2509,7 @@ exports.CreateChatCompletionResponse = {
                     if (tag !== 42) {
                         break;
                     }
-                    message.usage = exports.InferenceUsage.decode(reader, reader.uint32());
+                    message.usage = exports.Usage.decode(reader, reader.uint32());
                     continue;
                 }
                 case 6: {
@@ -2424,7 +2529,7 @@ exports.CreateChatCompletionResponse = {
     },
     fromJSON(object) {
         return {
-            context: isSet(object.context) ? exports.ResolvedInferenceContext.fromJSON(object.context) : undefined,
+            context: isSet(object.context) ? exports.ResolvedModelRoute.fromJSON(object.context) : undefined,
             message: isSet(object.message) ? exports.ChatMessage.fromJSON(object.message) : undefined,
             toolCalls: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.toolCalls)
                 ? object.toolCalls.map((e) => exports.ToolCall.fromJSON(e))
@@ -2432,11 +2537,11 @@ exports.CreateChatCompletionResponse = {
                     ? object.tool_calls.map((e) => exports.ToolCall.fromJSON(e))
                     : [],
             finishReason: isSet(object.finishReason)
-                ? inferenceFinishReasonFromJSON(object.finishReason)
+                ? finishReasonFromJSON(object.finishReason)
                 : isSet(object.finish_reason)
-                    ? inferenceFinishReasonFromJSON(object.finish_reason)
+                    ? finishReasonFromJSON(object.finish_reason)
                     : 0,
-            usage: isSet(object.usage) ? exports.InferenceUsage.fromJSON(object.usage) : undefined,
+            usage: isSet(object.usage) ? exports.Usage.fromJSON(object.usage) : undefined,
             structuredOutput: isObject(object.structuredOutput)
                 ? object.structuredOutput
                 : isObject(object.structured_output)
@@ -2448,7 +2553,7 @@ exports.CreateChatCompletionResponse = {
         var _a;
         const obj = {};
         if (message.context !== undefined) {
-            obj.context = exports.ResolvedInferenceContext.toJSON(message.context);
+            obj.context = exports.ResolvedModelRoute.toJSON(message.context);
         }
         if (message.message !== undefined) {
             obj.message = exports.ChatMessage.toJSON(message.message);
@@ -2457,10 +2562,10 @@ exports.CreateChatCompletionResponse = {
             obj.toolCalls = message.toolCalls.map((e) => exports.ToolCall.toJSON(e));
         }
         if (message.finishReason !== 0) {
-            obj.finishReason = inferenceFinishReasonToJSON(message.finishReason);
+            obj.finishReason = finishReasonToJSON(message.finishReason);
         }
         if (message.usage !== undefined) {
-            obj.usage = exports.InferenceUsage.toJSON(message.usage);
+            obj.usage = exports.Usage.toJSON(message.usage);
         }
         if (message.structuredOutput !== undefined) {
             obj.structuredOutput = message.structuredOutput;
@@ -2475,16 +2580,14 @@ exports.CreateChatCompletionResponse = {
         var _b, _c;
         const message = createBaseCreateChatCompletionResponse();
         message.context = (object.context !== undefined && object.context !== null)
-            ? exports.ResolvedInferenceContext.fromPartial(object.context)
+            ? exports.ResolvedModelRoute.fromPartial(object.context)
             : undefined;
         message.message = (object.message !== undefined && object.message !== null)
             ? exports.ChatMessage.fromPartial(object.message)
             : undefined;
         message.toolCalls = ((_a = object.toolCalls) === null || _a === void 0 ? void 0 : _a.map((e) => exports.ToolCall.fromPartial(e))) || [];
         message.finishReason = (_b = object.finishReason) !== null && _b !== void 0 ? _b : 0;
-        message.usage = (object.usage !== undefined && object.usage !== null)
-            ? exports.InferenceUsage.fromPartial(object.usage)
-            : undefined;
+        message.usage = (object.usage !== undefined && object.usage !== null) ? exports.Usage.fromPartial(object.usage) : undefined;
         message.structuredOutput = (_c = object.structuredOutput) !== null && _c !== void 0 ? _c : undefined;
         return message;
     },
@@ -2504,7 +2607,7 @@ function createBaseStreamChatCompletionRequest() {
 exports.StreamChatCompletionRequest = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.context !== undefined) {
-            exports.InferenceRequestContext.encode(message.context, writer.uint32(10).fork()).join();
+            exports.RequestContext.encode(message.context, writer.uint32(10).fork()).join();
         }
         if (message.model !== "") {
             writer.uint32(18).string(message.model);
@@ -2522,7 +2625,7 @@ exports.StreamChatCompletionRequest = {
             writer.uint32(48).int32(message.toolChoiceMode);
         }
         if (message.generationConfig !== undefined) {
-            exports.InferenceGenerationConfig.encode(message.generationConfig, writer.uint32(58).fork()).join();
+            exports.GenerationConfig.encode(message.generationConfig, writer.uint32(58).fork()).join();
         }
         if (message.responseSchema !== undefined) {
             exports.StructuredOutputSchema.encode(message.responseSchema, writer.uint32(66).fork()).join();
@@ -2540,7 +2643,7 @@ exports.StreamChatCompletionRequest = {
                     if (tag !== 10) {
                         break;
                     }
-                    message.context = exports.InferenceRequestContext.decode(reader, reader.uint32());
+                    message.context = exports.RequestContext.decode(reader, reader.uint32());
                     continue;
                 }
                 case 2: {
@@ -2582,7 +2685,7 @@ exports.StreamChatCompletionRequest = {
                     if (tag !== 58) {
                         break;
                     }
-                    message.generationConfig = exports.InferenceGenerationConfig.decode(reader, reader.uint32());
+                    message.generationConfig = exports.GenerationConfig.decode(reader, reader.uint32());
                     continue;
                 }
                 case 8: {
@@ -2602,7 +2705,7 @@ exports.StreamChatCompletionRequest = {
     },
     fromJSON(object) {
         return {
-            context: isSet(object.context) ? exports.InferenceRequestContext.fromJSON(object.context) : undefined,
+            context: isSet(object.context) ? exports.RequestContext.fromJSON(object.context) : undefined,
             model: isSet(object.model) ? globalThis.String(object.model) : "",
             messages: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.messages)
                 ? object.messages.map((e) => exports.ChatMessage.fromJSON(e))
@@ -2619,9 +2722,9 @@ exports.StreamChatCompletionRequest = {
                     ? toolChoiceModeFromJSON(object.tool_choice_mode)
                     : 0,
             generationConfig: isSet(object.generationConfig)
-                ? exports.InferenceGenerationConfig.fromJSON(object.generationConfig)
+                ? exports.GenerationConfig.fromJSON(object.generationConfig)
                 : isSet(object.generation_config)
-                    ? exports.InferenceGenerationConfig.fromJSON(object.generation_config)
+                    ? exports.GenerationConfig.fromJSON(object.generation_config)
                     : undefined,
             responseSchema: isSet(object.responseSchema)
                 ? exports.StructuredOutputSchema.fromJSON(object.responseSchema)
@@ -2634,7 +2737,7 @@ exports.StreamChatCompletionRequest = {
         var _a, _b;
         const obj = {};
         if (message.context !== undefined) {
-            obj.context = exports.InferenceRequestContext.toJSON(message.context);
+            obj.context = exports.RequestContext.toJSON(message.context);
         }
         if (message.model !== "") {
             obj.model = message.model;
@@ -2652,7 +2755,7 @@ exports.StreamChatCompletionRequest = {
             obj.toolChoiceMode = toolChoiceModeToJSON(message.toolChoiceMode);
         }
         if (message.generationConfig !== undefined) {
-            obj.generationConfig = exports.InferenceGenerationConfig.toJSON(message.generationConfig);
+            obj.generationConfig = exports.GenerationConfig.toJSON(message.generationConfig);
         }
         if (message.responseSchema !== undefined) {
             obj.responseSchema = exports.StructuredOutputSchema.toJSON(message.responseSchema);
@@ -2667,7 +2770,7 @@ exports.StreamChatCompletionRequest = {
         var _c, _d, _e;
         const message = createBaseStreamChatCompletionRequest();
         message.context = (object.context !== undefined && object.context !== null)
-            ? exports.InferenceRequestContext.fromPartial(object.context)
+            ? exports.RequestContext.fromPartial(object.context)
             : undefined;
         message.model = (_c = object.model) !== null && _c !== void 0 ? _c : "";
         message.messages = ((_a = object.messages) === null || _a === void 0 ? void 0 : _a.map((e) => exports.ChatMessage.fromPartial(e))) || [];
@@ -2675,7 +2778,7 @@ exports.StreamChatCompletionRequest = {
         message.tools = ((_b = object.tools) === null || _b === void 0 ? void 0 : _b.map((e) => exports.ToolDefinition.fromPartial(e))) || [];
         message.toolChoiceMode = (_e = object.toolChoiceMode) !== null && _e !== void 0 ? _e : 0;
         message.generationConfig = (object.generationConfig !== undefined && object.generationConfig !== null)
-            ? exports.InferenceGenerationConfig.fromPartial(object.generationConfig)
+            ? exports.GenerationConfig.fromPartial(object.generationConfig)
             : undefined;
         message.responseSchema = (object.responseSchema !== undefined && object.responseSchema !== null)
             ? exports.StructuredOutputSchema.fromPartial(object.responseSchema)
@@ -2698,7 +2801,7 @@ function createBaseStreamChatCompletionResponse() {
 exports.StreamChatCompletionResponse = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.context !== undefined) {
-            exports.ResolvedInferenceContext.encode(message.context, writer.uint32(10).fork()).join();
+            exports.ResolvedModelRoute.encode(message.context, writer.uint32(10).fork()).join();
         }
         if (message.deltaText !== "") {
             writer.uint32(18).string(message.deltaText);
@@ -2707,13 +2810,13 @@ exports.StreamChatCompletionResponse = {
             exports.ChatMessage.encode(message.deltaMessage, writer.uint32(26).fork()).join();
         }
         if (message.toolCallDelta !== undefined) {
-            exports.ToolCall.encode(message.toolCallDelta, writer.uint32(34).fork()).join();
+            exports.ToolCallDelta.encode(message.toolCallDelta, writer.uint32(34).fork()).join();
         }
         if (message.finishReason !== 0) {
             writer.uint32(40).int32(message.finishReason);
         }
         if (message.usage !== undefined) {
-            exports.InferenceUsage.encode(message.usage, writer.uint32(50).fork()).join();
+            exports.Usage.encode(message.usage, writer.uint32(50).fork()).join();
         }
         if (message.done !== false) {
             writer.uint32(56).bool(message.done);
@@ -2734,7 +2837,7 @@ exports.StreamChatCompletionResponse = {
                     if (tag !== 10) {
                         break;
                     }
-                    message.context = exports.ResolvedInferenceContext.decode(reader, reader.uint32());
+                    message.context = exports.ResolvedModelRoute.decode(reader, reader.uint32());
                     continue;
                 }
                 case 2: {
@@ -2755,7 +2858,7 @@ exports.StreamChatCompletionResponse = {
                     if (tag !== 34) {
                         break;
                     }
-                    message.toolCallDelta = exports.ToolCall.decode(reader, reader.uint32());
+                    message.toolCallDelta = exports.ToolCallDelta.decode(reader, reader.uint32());
                     continue;
                 }
                 case 5: {
@@ -2769,7 +2872,7 @@ exports.StreamChatCompletionResponse = {
                     if (tag !== 50) {
                         break;
                     }
-                    message.usage = exports.InferenceUsage.decode(reader, reader.uint32());
+                    message.usage = exports.Usage.decode(reader, reader.uint32());
                     continue;
                 }
                 case 7: {
@@ -2796,7 +2899,7 @@ exports.StreamChatCompletionResponse = {
     },
     fromJSON(object) {
         return {
-            context: isSet(object.context) ? exports.ResolvedInferenceContext.fromJSON(object.context) : undefined,
+            context: isSet(object.context) ? exports.ResolvedModelRoute.fromJSON(object.context) : undefined,
             deltaText: isSet(object.deltaText)
                 ? globalThis.String(object.deltaText)
                 : isSet(object.delta_text)
@@ -2808,16 +2911,16 @@ exports.StreamChatCompletionResponse = {
                     ? exports.ChatMessage.fromJSON(object.delta_message)
                     : undefined,
             toolCallDelta: isSet(object.toolCallDelta)
-                ? exports.ToolCall.fromJSON(object.toolCallDelta)
+                ? exports.ToolCallDelta.fromJSON(object.toolCallDelta)
                 : isSet(object.tool_call_delta)
-                    ? exports.ToolCall.fromJSON(object.tool_call_delta)
+                    ? exports.ToolCallDelta.fromJSON(object.tool_call_delta)
                     : undefined,
             finishReason: isSet(object.finishReason)
-                ? inferenceFinishReasonFromJSON(object.finishReason)
+                ? finishReasonFromJSON(object.finishReason)
                 : isSet(object.finish_reason)
-                    ? inferenceFinishReasonFromJSON(object.finish_reason)
+                    ? finishReasonFromJSON(object.finish_reason)
                     : 0,
-            usage: isSet(object.usage) ? exports.InferenceUsage.fromJSON(object.usage) : undefined,
+            usage: isSet(object.usage) ? exports.Usage.fromJSON(object.usage) : undefined,
             done: isSet(object.done) ? globalThis.Boolean(object.done) : false,
             structuredOutput: isObject(object.structuredOutput)
                 ? object.structuredOutput
@@ -2829,7 +2932,7 @@ exports.StreamChatCompletionResponse = {
     toJSON(message) {
         const obj = {};
         if (message.context !== undefined) {
-            obj.context = exports.ResolvedInferenceContext.toJSON(message.context);
+            obj.context = exports.ResolvedModelRoute.toJSON(message.context);
         }
         if (message.deltaText !== "") {
             obj.deltaText = message.deltaText;
@@ -2838,13 +2941,13 @@ exports.StreamChatCompletionResponse = {
             obj.deltaMessage = exports.ChatMessage.toJSON(message.deltaMessage);
         }
         if (message.toolCallDelta !== undefined) {
-            obj.toolCallDelta = exports.ToolCall.toJSON(message.toolCallDelta);
+            obj.toolCallDelta = exports.ToolCallDelta.toJSON(message.toolCallDelta);
         }
         if (message.finishReason !== 0) {
-            obj.finishReason = inferenceFinishReasonToJSON(message.finishReason);
+            obj.finishReason = finishReasonToJSON(message.finishReason);
         }
         if (message.usage !== undefined) {
-            obj.usage = exports.InferenceUsage.toJSON(message.usage);
+            obj.usage = exports.Usage.toJSON(message.usage);
         }
         if (message.done !== false) {
             obj.done = message.done;
@@ -2861,19 +2964,17 @@ exports.StreamChatCompletionResponse = {
         var _a, _b, _c, _d;
         const message = createBaseStreamChatCompletionResponse();
         message.context = (object.context !== undefined && object.context !== null)
-            ? exports.ResolvedInferenceContext.fromPartial(object.context)
+            ? exports.ResolvedModelRoute.fromPartial(object.context)
             : undefined;
         message.deltaText = (_a = object.deltaText) !== null && _a !== void 0 ? _a : "";
         message.deltaMessage = (object.deltaMessage !== undefined && object.deltaMessage !== null)
             ? exports.ChatMessage.fromPartial(object.deltaMessage)
             : undefined;
         message.toolCallDelta = (object.toolCallDelta !== undefined && object.toolCallDelta !== null)
-            ? exports.ToolCall.fromPartial(object.toolCallDelta)
+            ? exports.ToolCallDelta.fromPartial(object.toolCallDelta)
             : undefined;
         message.finishReason = (_b = object.finishReason) !== null && _b !== void 0 ? _b : 0;
-        message.usage = (object.usage !== undefined && object.usage !== null)
-            ? exports.InferenceUsage.fromPartial(object.usage)
-            : undefined;
+        message.usage = (object.usage !== undefined && object.usage !== null) ? exports.Usage.fromPartial(object.usage) : undefined;
         message.done = (_c = object.done) !== null && _c !== void 0 ? _c : false;
         message.structuredOutput = (_d = object.structuredOutput) !== null && _d !== void 0 ? _d : undefined;
         return message;
@@ -2893,7 +2994,7 @@ function createBaseCallToolsRequest() {
 exports.CallToolsRequest = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.context !== undefined) {
-            exports.InferenceRequestContext.encode(message.context, writer.uint32(10).fork()).join();
+            exports.RequestContext.encode(message.context, writer.uint32(10).fork()).join();
         }
         if (message.model !== "") {
             writer.uint32(18).string(message.model);
@@ -2911,7 +3012,7 @@ exports.CallToolsRequest = {
             writer.uint32(48).int32(message.toolChoiceMode);
         }
         if (message.generationConfig !== undefined) {
-            exports.InferenceGenerationConfig.encode(message.generationConfig, writer.uint32(58).fork()).join();
+            exports.GenerationConfig.encode(message.generationConfig, writer.uint32(58).fork()).join();
         }
         return writer;
     },
@@ -2926,7 +3027,7 @@ exports.CallToolsRequest = {
                     if (tag !== 10) {
                         break;
                     }
-                    message.context = exports.InferenceRequestContext.decode(reader, reader.uint32());
+                    message.context = exports.RequestContext.decode(reader, reader.uint32());
                     continue;
                 }
                 case 2: {
@@ -2968,7 +3069,7 @@ exports.CallToolsRequest = {
                     if (tag !== 58) {
                         break;
                     }
-                    message.generationConfig = exports.InferenceGenerationConfig.decode(reader, reader.uint32());
+                    message.generationConfig = exports.GenerationConfig.decode(reader, reader.uint32());
                     continue;
                 }
             }
@@ -2981,7 +3082,7 @@ exports.CallToolsRequest = {
     },
     fromJSON(object) {
         return {
-            context: isSet(object.context) ? exports.InferenceRequestContext.fromJSON(object.context) : undefined,
+            context: isSet(object.context) ? exports.RequestContext.fromJSON(object.context) : undefined,
             model: isSet(object.model) ? globalThis.String(object.model) : "",
             messages: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.messages)
                 ? object.messages.map((e) => exports.ChatMessage.fromJSON(e))
@@ -2998,9 +3099,9 @@ exports.CallToolsRequest = {
                     ? toolChoiceModeFromJSON(object.tool_choice_mode)
                     : 0,
             generationConfig: isSet(object.generationConfig)
-                ? exports.InferenceGenerationConfig.fromJSON(object.generationConfig)
+                ? exports.GenerationConfig.fromJSON(object.generationConfig)
                 : isSet(object.generation_config)
-                    ? exports.InferenceGenerationConfig.fromJSON(object.generation_config)
+                    ? exports.GenerationConfig.fromJSON(object.generation_config)
                     : undefined,
         };
     },
@@ -3008,7 +3109,7 @@ exports.CallToolsRequest = {
         var _a, _b;
         const obj = {};
         if (message.context !== undefined) {
-            obj.context = exports.InferenceRequestContext.toJSON(message.context);
+            obj.context = exports.RequestContext.toJSON(message.context);
         }
         if (message.model !== "") {
             obj.model = message.model;
@@ -3026,7 +3127,7 @@ exports.CallToolsRequest = {
             obj.toolChoiceMode = toolChoiceModeToJSON(message.toolChoiceMode);
         }
         if (message.generationConfig !== undefined) {
-            obj.generationConfig = exports.InferenceGenerationConfig.toJSON(message.generationConfig);
+            obj.generationConfig = exports.GenerationConfig.toJSON(message.generationConfig);
         }
         return obj;
     },
@@ -3038,7 +3139,7 @@ exports.CallToolsRequest = {
         var _c, _d, _e;
         const message = createBaseCallToolsRequest();
         message.context = (object.context !== undefined && object.context !== null)
-            ? exports.InferenceRequestContext.fromPartial(object.context)
+            ? exports.RequestContext.fromPartial(object.context)
             : undefined;
         message.model = (_c = object.model) !== null && _c !== void 0 ? _c : "";
         message.messages = ((_a = object.messages) === null || _a === void 0 ? void 0 : _a.map((e) => exports.ChatMessage.fromPartial(e))) || [];
@@ -3046,7 +3147,7 @@ exports.CallToolsRequest = {
         message.tools = ((_b = object.tools) === null || _b === void 0 ? void 0 : _b.map((e) => exports.ToolDefinition.fromPartial(e))) || [];
         message.toolChoiceMode = (_e = object.toolChoiceMode) !== null && _e !== void 0 ? _e : 0;
         message.generationConfig = (object.generationConfig !== undefined && object.generationConfig !== null)
-            ? exports.InferenceGenerationConfig.fromPartial(object.generationConfig)
+            ? exports.GenerationConfig.fromPartial(object.generationConfig)
             : undefined;
         return message;
     },
@@ -3057,7 +3158,7 @@ function createBaseCallToolsResponse() {
 exports.CallToolsResponse = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.context !== undefined) {
-            exports.ResolvedInferenceContext.encode(message.context, writer.uint32(10).fork()).join();
+            exports.ResolvedModelRoute.encode(message.context, writer.uint32(10).fork()).join();
         }
         if (message.message !== undefined) {
             exports.ChatMessage.encode(message.message, writer.uint32(18).fork()).join();
@@ -3069,7 +3170,7 @@ exports.CallToolsResponse = {
             writer.uint32(32).int32(message.finishReason);
         }
         if (message.usage !== undefined) {
-            exports.InferenceUsage.encode(message.usage, writer.uint32(42).fork()).join();
+            exports.Usage.encode(message.usage, writer.uint32(42).fork()).join();
         }
         return writer;
     },
@@ -3084,7 +3185,7 @@ exports.CallToolsResponse = {
                     if (tag !== 10) {
                         break;
                     }
-                    message.context = exports.ResolvedInferenceContext.decode(reader, reader.uint32());
+                    message.context = exports.ResolvedModelRoute.decode(reader, reader.uint32());
                     continue;
                 }
                 case 2: {
@@ -3112,7 +3213,7 @@ exports.CallToolsResponse = {
                     if (tag !== 42) {
                         break;
                     }
-                    message.usage = exports.InferenceUsage.decode(reader, reader.uint32());
+                    message.usage = exports.Usage.decode(reader, reader.uint32());
                     continue;
                 }
             }
@@ -3125,7 +3226,7 @@ exports.CallToolsResponse = {
     },
     fromJSON(object) {
         return {
-            context: isSet(object.context) ? exports.ResolvedInferenceContext.fromJSON(object.context) : undefined,
+            context: isSet(object.context) ? exports.ResolvedModelRoute.fromJSON(object.context) : undefined,
             message: isSet(object.message) ? exports.ChatMessage.fromJSON(object.message) : undefined,
             toolCalls: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.toolCalls)
                 ? object.toolCalls.map((e) => exports.ToolCall.fromJSON(e))
@@ -3133,18 +3234,18 @@ exports.CallToolsResponse = {
                     ? object.tool_calls.map((e) => exports.ToolCall.fromJSON(e))
                     : [],
             finishReason: isSet(object.finishReason)
-                ? inferenceFinishReasonFromJSON(object.finishReason)
+                ? finishReasonFromJSON(object.finishReason)
                 : isSet(object.finish_reason)
-                    ? inferenceFinishReasonFromJSON(object.finish_reason)
+                    ? finishReasonFromJSON(object.finish_reason)
                     : 0,
-            usage: isSet(object.usage) ? exports.InferenceUsage.fromJSON(object.usage) : undefined,
+            usage: isSet(object.usage) ? exports.Usage.fromJSON(object.usage) : undefined,
         };
     },
     toJSON(message) {
         var _a;
         const obj = {};
         if (message.context !== undefined) {
-            obj.context = exports.ResolvedInferenceContext.toJSON(message.context);
+            obj.context = exports.ResolvedModelRoute.toJSON(message.context);
         }
         if (message.message !== undefined) {
             obj.message = exports.ChatMessage.toJSON(message.message);
@@ -3153,10 +3254,10 @@ exports.CallToolsResponse = {
             obj.toolCalls = message.toolCalls.map((e) => exports.ToolCall.toJSON(e));
         }
         if (message.finishReason !== 0) {
-            obj.finishReason = inferenceFinishReasonToJSON(message.finishReason);
+            obj.finishReason = finishReasonToJSON(message.finishReason);
         }
         if (message.usage !== undefined) {
-            obj.usage = exports.InferenceUsage.toJSON(message.usage);
+            obj.usage = exports.Usage.toJSON(message.usage);
         }
         return obj;
     },
@@ -3168,16 +3269,14 @@ exports.CallToolsResponse = {
         var _b;
         const message = createBaseCallToolsResponse();
         message.context = (object.context !== undefined && object.context !== null)
-            ? exports.ResolvedInferenceContext.fromPartial(object.context)
+            ? exports.ResolvedModelRoute.fromPartial(object.context)
             : undefined;
         message.message = (object.message !== undefined && object.message !== null)
             ? exports.ChatMessage.fromPartial(object.message)
             : undefined;
         message.toolCalls = ((_a = object.toolCalls) === null || _a === void 0 ? void 0 : _a.map((e) => exports.ToolCall.fromPartial(e))) || [];
         message.finishReason = (_b = object.finishReason) !== null && _b !== void 0 ? _b : 0;
-        message.usage = (object.usage !== undefined && object.usage !== null)
-            ? exports.InferenceUsage.fromPartial(object.usage)
-            : undefined;
+        message.usage = (object.usage !== undefined && object.usage !== null) ? exports.Usage.fromPartial(object.usage) : undefined;
         return message;
     },
 };
@@ -3187,7 +3286,7 @@ function createBaseCreateEmbeddingsRequest() {
 exports.CreateEmbeddingsRequest = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.context !== undefined) {
-            exports.InferenceRequestContext.encode(message.context, writer.uint32(10).fork()).join();
+            exports.RequestContext.encode(message.context, writer.uint32(10).fork()).join();
         }
         if (message.model !== "") {
             writer.uint32(18).string(message.model);
@@ -3217,7 +3316,7 @@ exports.CreateEmbeddingsRequest = {
                     if (tag !== 10) {
                         break;
                     }
-                    message.context = exports.InferenceRequestContext.decode(reader, reader.uint32());
+                    message.context = exports.RequestContext.decode(reader, reader.uint32());
                     continue;
                 }
                 case 2: {
@@ -3265,7 +3364,7 @@ exports.CreateEmbeddingsRequest = {
     },
     fromJSON(object) {
         return {
-            context: isSet(object.context) ? exports.InferenceRequestContext.fromJSON(object.context) : undefined,
+            context: isSet(object.context) ? exports.RequestContext.fromJSON(object.context) : undefined,
             model: isSet(object.model) ? globalThis.String(object.model) : "",
             inputs: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.inputs) ? object.inputs.map((e) => exports.EmbeddingInput.fromJSON(e)) : [],
             dimensions: isSet(object.dimensions) ? globalThis.Number(object.dimensions) : 0,
@@ -3281,7 +3380,7 @@ exports.CreateEmbeddingsRequest = {
         var _a;
         const obj = {};
         if (message.context !== undefined) {
-            obj.context = exports.InferenceRequestContext.toJSON(message.context);
+            obj.context = exports.RequestContext.toJSON(message.context);
         }
         if (message.model !== "") {
             obj.model = message.model;
@@ -3308,7 +3407,7 @@ exports.CreateEmbeddingsRequest = {
         var _b, _c, _d, _e;
         const message = createBaseCreateEmbeddingsRequest();
         message.context = (object.context !== undefined && object.context !== null)
-            ? exports.InferenceRequestContext.fromPartial(object.context)
+            ? exports.RequestContext.fromPartial(object.context)
             : undefined;
         message.model = (_b = object.model) !== null && _b !== void 0 ? _b : "";
         message.inputs = ((_a = object.inputs) === null || _a === void 0 ? void 0 : _a.map((e) => exports.EmbeddingInput.fromPartial(e))) || [];
@@ -3324,13 +3423,13 @@ function createBaseCreateEmbeddingsResponse() {
 exports.CreateEmbeddingsResponse = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.context !== undefined) {
-            exports.ResolvedInferenceContext.encode(message.context, writer.uint32(10).fork()).join();
+            exports.ResolvedModelRoute.encode(message.context, writer.uint32(10).fork()).join();
         }
         for (const v of message.embeddings) {
             exports.EmbeddingVector.encode(v, writer.uint32(18).fork()).join();
         }
         if (message.usage !== undefined) {
-            exports.InferenceUsage.encode(message.usage, writer.uint32(26).fork()).join();
+            exports.Usage.encode(message.usage, writer.uint32(26).fork()).join();
         }
         return writer;
     },
@@ -3345,7 +3444,7 @@ exports.CreateEmbeddingsResponse = {
                     if (tag !== 10) {
                         break;
                     }
-                    message.context = exports.ResolvedInferenceContext.decode(reader, reader.uint32());
+                    message.context = exports.ResolvedModelRoute.decode(reader, reader.uint32());
                     continue;
                 }
                 case 2: {
@@ -3359,7 +3458,7 @@ exports.CreateEmbeddingsResponse = {
                     if (tag !== 26) {
                         break;
                     }
-                    message.usage = exports.InferenceUsage.decode(reader, reader.uint32());
+                    message.usage = exports.Usage.decode(reader, reader.uint32());
                     continue;
                 }
             }
@@ -3372,24 +3471,24 @@ exports.CreateEmbeddingsResponse = {
     },
     fromJSON(object) {
         return {
-            context: isSet(object.context) ? exports.ResolvedInferenceContext.fromJSON(object.context) : undefined,
+            context: isSet(object.context) ? exports.ResolvedModelRoute.fromJSON(object.context) : undefined,
             embeddings: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.embeddings)
                 ? object.embeddings.map((e) => exports.EmbeddingVector.fromJSON(e))
                 : [],
-            usage: isSet(object.usage) ? exports.InferenceUsage.fromJSON(object.usage) : undefined,
+            usage: isSet(object.usage) ? exports.Usage.fromJSON(object.usage) : undefined,
         };
     },
     toJSON(message) {
         var _a;
         const obj = {};
         if (message.context !== undefined) {
-            obj.context = exports.ResolvedInferenceContext.toJSON(message.context);
+            obj.context = exports.ResolvedModelRoute.toJSON(message.context);
         }
         if ((_a = message.embeddings) === null || _a === void 0 ? void 0 : _a.length) {
             obj.embeddings = message.embeddings.map((e) => exports.EmbeddingVector.toJSON(e));
         }
         if (message.usage !== undefined) {
-            obj.usage = exports.InferenceUsage.toJSON(message.usage);
+            obj.usage = exports.Usage.toJSON(message.usage);
         }
         return obj;
     },
@@ -3400,12 +3499,10 @@ exports.CreateEmbeddingsResponse = {
         var _a;
         const message = createBaseCreateEmbeddingsResponse();
         message.context = (object.context !== undefined && object.context !== null)
-            ? exports.ResolvedInferenceContext.fromPartial(object.context)
+            ? exports.ResolvedModelRoute.fromPartial(object.context)
             : undefined;
         message.embeddings = ((_a = object.embeddings) === null || _a === void 0 ? void 0 : _a.map((e) => exports.EmbeddingVector.fromPartial(e))) || [];
-        message.usage = (object.usage !== undefined && object.usage !== null)
-            ? exports.InferenceUsage.fromPartial(object.usage)
-            : undefined;
+        message.usage = (object.usage !== undefined && object.usage !== null) ? exports.Usage.fromPartial(object.usage) : undefined;
         return message;
     },
 };
@@ -3426,7 +3523,7 @@ function createBaseGenerateImagesRequest() {
 exports.GenerateImagesRequest = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.context !== undefined) {
-            exports.InferenceRequestContext.encode(message.context, writer.uint32(10).fork()).join();
+            exports.RequestContext.encode(message.context, writer.uint32(10).fork()).join();
         }
         if (message.model !== "") {
             writer.uint32(18).string(message.model);
@@ -3468,7 +3565,7 @@ exports.GenerateImagesRequest = {
                     if (tag !== 10) {
                         break;
                     }
-                    message.context = exports.InferenceRequestContext.decode(reader, reader.uint32());
+                    message.context = exports.RequestContext.decode(reader, reader.uint32());
                     continue;
                 }
                 case 2: {
@@ -3544,7 +3641,7 @@ exports.GenerateImagesRequest = {
     },
     fromJSON(object) {
         return {
-            context: isSet(object.context) ? exports.InferenceRequestContext.fromJSON(object.context) : undefined,
+            context: isSet(object.context) ? exports.RequestContext.fromJSON(object.context) : undefined,
             model: isSet(object.model) ? globalThis.String(object.model) : "",
             prompt: isSet(object.prompt) ? globalThis.String(object.prompt) : "",
             negativePrompt: isSet(object.negativePrompt)
@@ -3571,7 +3668,7 @@ exports.GenerateImagesRequest = {
     toJSON(message) {
         const obj = {};
         if (message.context !== undefined) {
-            obj.context = exports.InferenceRequestContext.toJSON(message.context);
+            obj.context = exports.RequestContext.toJSON(message.context);
         }
         if (message.model !== "") {
             obj.model = message.model;
@@ -3609,7 +3706,7 @@ exports.GenerateImagesRequest = {
         var _a, _b, _c, _d, _e, _f, _g, _h, _j;
         const message = createBaseGenerateImagesRequest();
         message.context = (object.context !== undefined && object.context !== null)
-            ? exports.InferenceRequestContext.fromPartial(object.context)
+            ? exports.RequestContext.fromPartial(object.context)
             : undefined;
         message.model = (_a = object.model) !== null && _a !== void 0 ? _a : "";
         message.prompt = (_b = object.prompt) !== null && _b !== void 0 ? _b : "";
@@ -3629,7 +3726,7 @@ function createBaseGenerateImagesResponse() {
 exports.GenerateImagesResponse = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.context !== undefined) {
-            exports.ResolvedInferenceContext.encode(message.context, writer.uint32(10).fork()).join();
+            exports.ResolvedModelRoute.encode(message.context, writer.uint32(10).fork()).join();
         }
         for (const v of message.images) {
             exports.GeneratedImage.encode(v, writer.uint32(18).fork()).join();
@@ -3647,7 +3744,7 @@ exports.GenerateImagesResponse = {
                     if (tag !== 10) {
                         break;
                     }
-                    message.context = exports.ResolvedInferenceContext.decode(reader, reader.uint32());
+                    message.context = exports.ResolvedModelRoute.decode(reader, reader.uint32());
                     continue;
                 }
                 case 2: {
@@ -3667,7 +3764,7 @@ exports.GenerateImagesResponse = {
     },
     fromJSON(object) {
         return {
-            context: isSet(object.context) ? exports.ResolvedInferenceContext.fromJSON(object.context) : undefined,
+            context: isSet(object.context) ? exports.ResolvedModelRoute.fromJSON(object.context) : undefined,
             images: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.images) ? object.images.map((e) => exports.GeneratedImage.fromJSON(e)) : [],
         };
     },
@@ -3675,7 +3772,7 @@ exports.GenerateImagesResponse = {
         var _a;
         const obj = {};
         if (message.context !== undefined) {
-            obj.context = exports.ResolvedInferenceContext.toJSON(message.context);
+            obj.context = exports.ResolvedModelRoute.toJSON(message.context);
         }
         if ((_a = message.images) === null || _a === void 0 ? void 0 : _a.length) {
             obj.images = message.images.map((e) => exports.GeneratedImage.toJSON(e));
@@ -3689,7 +3786,7 @@ exports.GenerateImagesResponse = {
         var _a;
         const message = createBaseGenerateImagesResponse();
         message.context = (object.context !== undefined && object.context !== null)
-            ? exports.ResolvedInferenceContext.fromPartial(object.context)
+            ? exports.ResolvedModelRoute.fromPartial(object.context)
             : undefined;
         message.images = ((_a = object.images) === null || _a === void 0 ? void 0 : _a.map((e) => exports.GeneratedImage.fromPartial(e))) || [];
         return message;
@@ -3701,7 +3798,7 @@ function createBaseCallMCPRequest() {
 exports.CallMCPRequest = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.context !== undefined) {
-            exports.InferenceRequestContext.encode(message.context, writer.uint32(10).fork()).join();
+            exports.RequestContext.encode(message.context, writer.uint32(10).fork()).join();
         }
         if (message.model !== "") {
             writer.uint32(18).string(message.model);
@@ -3731,7 +3828,7 @@ exports.CallMCPRequest = {
                     if (tag !== 10) {
                         break;
                     }
-                    message.context = exports.InferenceRequestContext.decode(reader, reader.uint32());
+                    message.context = exports.RequestContext.decode(reader, reader.uint32());
                     continue;
                 }
                 case 2: {
@@ -3779,7 +3876,7 @@ exports.CallMCPRequest = {
     },
     fromJSON(object) {
         return {
-            context: isSet(object.context) ? exports.InferenceRequestContext.fromJSON(object.context) : undefined,
+            context: isSet(object.context) ? exports.RequestContext.fromJSON(object.context) : undefined,
             model: isSet(object.model) ? globalThis.String(object.model) : "",
             serverName: isSet(object.serverName)
                 ? globalThis.String(object.serverName)
@@ -3798,7 +3895,7 @@ exports.CallMCPRequest = {
     toJSON(message) {
         const obj = {};
         if (message.context !== undefined) {
-            obj.context = exports.InferenceRequestContext.toJSON(message.context);
+            obj.context = exports.RequestContext.toJSON(message.context);
         }
         if (message.model !== "") {
             obj.model = message.model;
@@ -3824,7 +3921,7 @@ exports.CallMCPRequest = {
         var _a, _b, _c, _d, _e;
         const message = createBaseCallMCPRequest();
         message.context = (object.context !== undefined && object.context !== null)
-            ? exports.InferenceRequestContext.fromPartial(object.context)
+            ? exports.RequestContext.fromPartial(object.context)
             : undefined;
         message.model = (_a = object.model) !== null && _a !== void 0 ? _a : "";
         message.serverName = (_b = object.serverName) !== null && _b !== void 0 ? _b : "";
@@ -3840,7 +3937,7 @@ function createBaseCallMCPResponse() {
 exports.CallMCPResponse = {
     encode(message, writer = new wire_1.BinaryWriter()) {
         if (message.context !== undefined) {
-            exports.ResolvedInferenceContext.encode(message.context, writer.uint32(10).fork()).join();
+            exports.ResolvedModelRoute.encode(message.context, writer.uint32(10).fork()).join();
         }
         if (message.result !== undefined) {
             struct_pb_1.Struct.encode(struct_pb_1.Struct.wrap(message.result), writer.uint32(18).fork()).join();
@@ -3858,7 +3955,7 @@ exports.CallMCPResponse = {
                     if (tag !== 10) {
                         break;
                     }
-                    message.context = exports.ResolvedInferenceContext.decode(reader, reader.uint32());
+                    message.context = exports.ResolvedModelRoute.decode(reader, reader.uint32());
                     continue;
                 }
                 case 2: {
@@ -3878,14 +3975,14 @@ exports.CallMCPResponse = {
     },
     fromJSON(object) {
         return {
-            context: isSet(object.context) ? exports.ResolvedInferenceContext.fromJSON(object.context) : undefined,
+            context: isSet(object.context) ? exports.ResolvedModelRoute.fromJSON(object.context) : undefined,
             result: isObject(object.result) ? object.result : undefined,
         };
     },
     toJSON(message) {
         const obj = {};
         if (message.context !== undefined) {
-            obj.context = exports.ResolvedInferenceContext.toJSON(message.context);
+            obj.context = exports.ResolvedModelRoute.toJSON(message.context);
         }
         if (message.result !== undefined) {
             obj.result = message.result;
@@ -3899,7 +3996,7 @@ exports.CallMCPResponse = {
         var _a;
         const message = createBaseCallMCPResponse();
         message.context = (object.context !== undefined && object.context !== null)
-            ? exports.ResolvedInferenceContext.fromPartial(object.context)
+            ? exports.ResolvedModelRoute.fromPartial(object.context)
             : undefined;
         message.result = (_a = object.result) !== null && _a !== void 0 ? _a : undefined;
         return message;
@@ -4807,7 +4904,7 @@ exports.OpenAICompatibilityServiceCreateChatCompletionResponse = {
             exports.OpenAIChatChoice.encode(v, writer.uint32(34).fork()).join();
         }
         if (message.usage !== undefined) {
-            exports.InferenceUsage.encode(message.usage, writer.uint32(42).fork()).join();
+            exports.Usage.encode(message.usage, writer.uint32(42).fork()).join();
         }
         return writer;
     },
@@ -4850,7 +4947,7 @@ exports.OpenAICompatibilityServiceCreateChatCompletionResponse = {
                     if (tag !== 42) {
                         break;
                     }
-                    message.usage = exports.InferenceUsage.decode(reader, reader.uint32());
+                    message.usage = exports.Usage.decode(reader, reader.uint32());
                     continue;
                 }
             }
@@ -4869,7 +4966,7 @@ exports.OpenAICompatibilityServiceCreateChatCompletionResponse = {
             choices: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.choices)
                 ? object.choices.map((e) => exports.OpenAIChatChoice.fromJSON(e))
                 : [],
-            usage: isSet(object.usage) ? exports.InferenceUsage.fromJSON(object.usage) : undefined,
+            usage: isSet(object.usage) ? exports.Usage.fromJSON(object.usage) : undefined,
         };
     },
     toJSON(message) {
@@ -4888,7 +4985,7 @@ exports.OpenAICompatibilityServiceCreateChatCompletionResponse = {
             obj.choices = message.choices.map((e) => exports.OpenAIChatChoice.toJSON(e));
         }
         if (message.usage !== undefined) {
-            obj.usage = exports.InferenceUsage.toJSON(message.usage);
+            obj.usage = exports.Usage.toJSON(message.usage);
         }
         return obj;
     },
@@ -4903,9 +5000,7 @@ exports.OpenAICompatibilityServiceCreateChatCompletionResponse = {
         message.object = (_c = object.object) !== null && _c !== void 0 ? _c : "";
         message.model = (_d = object.model) !== null && _d !== void 0 ? _d : "";
         message.choices = ((_a = object.choices) === null || _a === void 0 ? void 0 : _a.map((e) => exports.OpenAIChatChoice.fromPartial(e))) || [];
-        message.usage = (object.usage !== undefined && object.usage !== null)
-            ? exports.InferenceUsage.fromPartial(object.usage)
-            : undefined;
+        message.usage = (object.usage !== undefined && object.usage !== null) ? exports.Usage.fromPartial(object.usage) : undefined;
         return message;
     },
 };
@@ -5255,7 +5350,7 @@ exports.OpenAICompatibilityServiceStreamChatCompletionResponse = {
             exports.OpenAIChatDeltaChoice.encode(v, writer.uint32(34).fork()).join();
         }
         if (message.usage !== undefined) {
-            exports.InferenceUsage.encode(message.usage, writer.uint32(42).fork()).join();
+            exports.Usage.encode(message.usage, writer.uint32(42).fork()).join();
         }
         if (message.done !== false) {
             writer.uint32(48).bool(message.done);
@@ -5301,7 +5396,7 @@ exports.OpenAICompatibilityServiceStreamChatCompletionResponse = {
                     if (tag !== 42) {
                         break;
                     }
-                    message.usage = exports.InferenceUsage.decode(reader, reader.uint32());
+                    message.usage = exports.Usage.decode(reader, reader.uint32());
                     continue;
                 }
                 case 6: {
@@ -5327,7 +5422,7 @@ exports.OpenAICompatibilityServiceStreamChatCompletionResponse = {
             choices: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.choices)
                 ? object.choices.map((e) => exports.OpenAIChatDeltaChoice.fromJSON(e))
                 : [],
-            usage: isSet(object.usage) ? exports.InferenceUsage.fromJSON(object.usage) : undefined,
+            usage: isSet(object.usage) ? exports.Usage.fromJSON(object.usage) : undefined,
             done: isSet(object.done) ? globalThis.Boolean(object.done) : false,
         };
     },
@@ -5347,7 +5442,7 @@ exports.OpenAICompatibilityServiceStreamChatCompletionResponse = {
             obj.choices = message.choices.map((e) => exports.OpenAIChatDeltaChoice.toJSON(e));
         }
         if (message.usage !== undefined) {
-            obj.usage = exports.InferenceUsage.toJSON(message.usage);
+            obj.usage = exports.Usage.toJSON(message.usage);
         }
         if (message.done !== false) {
             obj.done = message.done;
@@ -5365,9 +5460,7 @@ exports.OpenAICompatibilityServiceStreamChatCompletionResponse = {
         message.object = (_c = object.object) !== null && _c !== void 0 ? _c : "";
         message.model = (_d = object.model) !== null && _d !== void 0 ? _d : "";
         message.choices = ((_a = object.choices) === null || _a === void 0 ? void 0 : _a.map((e) => exports.OpenAIChatDeltaChoice.fromPartial(e))) || [];
-        message.usage = (object.usage !== undefined && object.usage !== null)
-            ? exports.InferenceUsage.fromPartial(object.usage)
-            : undefined;
+        message.usage = (object.usage !== undefined && object.usage !== null) ? exports.Usage.fromPartial(object.usage) : undefined;
         message.done = (_e = object.done) !== null && _e !== void 0 ? _e : false;
         return message;
     },
@@ -5604,7 +5697,7 @@ exports.OpenAICompatibilityServiceCreateEmbeddingResponse = {
             writer.uint32(26).string(message.model);
         }
         if (message.usage !== undefined) {
-            exports.InferenceUsage.encode(message.usage, writer.uint32(34).fork()).join();
+            exports.Usage.encode(message.usage, writer.uint32(34).fork()).join();
         }
         return writer;
     },
@@ -5640,7 +5733,7 @@ exports.OpenAICompatibilityServiceCreateEmbeddingResponse = {
                     if (tag !== 34) {
                         break;
                     }
-                    message.usage = exports.InferenceUsage.decode(reader, reader.uint32());
+                    message.usage = exports.Usage.decode(reader, reader.uint32());
                     continue;
                 }
             }
@@ -5656,7 +5749,7 @@ exports.OpenAICompatibilityServiceCreateEmbeddingResponse = {
             object: isSet(object.object) ? globalThis.String(object.object) : "",
             data: globalThis.Array.isArray(object === null || object === void 0 ? void 0 : object.data) ? object.data.map((e) => exports.OpenAIEmbeddingData.fromJSON(e)) : [],
             model: isSet(object.model) ? globalThis.String(object.model) : "",
-            usage: isSet(object.usage) ? exports.InferenceUsage.fromJSON(object.usage) : undefined,
+            usage: isSet(object.usage) ? exports.Usage.fromJSON(object.usage) : undefined,
         };
     },
     toJSON(message) {
@@ -5672,7 +5765,7 @@ exports.OpenAICompatibilityServiceCreateEmbeddingResponse = {
             obj.model = message.model;
         }
         if (message.usage !== undefined) {
-            obj.usage = exports.InferenceUsage.toJSON(message.usage);
+            obj.usage = exports.Usage.toJSON(message.usage);
         }
         return obj;
     },
@@ -5686,9 +5779,7 @@ exports.OpenAICompatibilityServiceCreateEmbeddingResponse = {
         message.object = (_b = object.object) !== null && _b !== void 0 ? _b : "";
         message.data = ((_a = object.data) === null || _a === void 0 ? void 0 : _a.map((e) => exports.OpenAIEmbeddingData.fromPartial(e))) || [];
         message.model = (_c = object.model) !== null && _c !== void 0 ? _c : "";
-        message.usage = (object.usage !== undefined && object.usage !== null)
-            ? exports.InferenceUsage.fromPartial(object.usage)
-            : undefined;
+        message.usage = (object.usage !== undefined && object.usage !== null) ? exports.Usage.fromPartial(object.usage) : undefined;
         return message;
     },
 };