obey model features when sending messages

Author: Mickey Knox

app/modules/serviceInterfaces/LocalServerServiceInterface/Sources/sendMessageSchema.generated.swift

@@ -10,7 +10,7 @@ extension Schema {
     public let messages: [Message]
     public let system: String?
     public let projectRoot: String?
-    public let tools: [Tool]
+    public let tools: [Tool]?
     public let model: String
     public let enableReasoning: Bool
     public let provider: APIProvider
@@ -31,7 +31,7 @@ extension Schema {
         messages: [Message],
         system: String? = nil,
         projectRoot: String? = nil,
-        tools: [Tool],
+        tools: [Tool]? = nil,
         model: String,
         enableReasoning: Bool,
         provider: APIProvider,
@@ -52,7 +52,7 @@ extension Schema {
       messages = try container.decode([Message].self, forKey: .messages)
       system = try container.decodeIfPresent(String?.self, forKey: .system)
       projectRoot = try container.decodeIfPresent(String?.self, forKey: .projectRoot)
-      tools = try container.decode([Tool].self, forKey: .tools)
+      tools = try container.decodeIfPresent([Tool]?.self, forKey: .tools)
       model = try container.decode(String.self, forKey: .model)
       enableReasoning = try container.decode(Bool.self, forKey: .enableReasoning)
       provider = try container.decode(APIProvider.self, forKey: .provider)
@@ -64,7 +64,7 @@ extension Schema {
       try container.encode(messages, forKey: .messages)
       try container.encodeIfPresent(system, forKey: .system)
       try container.encodeIfPresent(projectRoot, forKey: .projectRoot)
-      try container.encode(tools, forKey: .tools)
+      try container.encodeIfPresent(tools, forKey: .tools)
       try container.encode(model, forKey: .model)
       try container.encode(enableReasoning, forKey: .enableReasoning)
       try container.encode(provider, forKey: .provider)

app/modules/services/LLMService/Sources/DefaultLLMService.swift

@@ -267,12 +267,14 @@ final class DefaultLLMService: LLMService {
       messages: messageHistory,
       system: system,
       projectRoot: context?.projectRoot?.path,
-      tools: tools
+      tools: model.supportsTools
+        ? tools
         // Unless we are using an external agent, only send to the AI tools that are internal.
         .filter { ($0.canBeExecuted && $0.id == $0.referenceId) || provider.isExternalAgent }
-        .map { .init(name: $0.name, description: $0.description, inputSchema: $0.inputSchema) },
+        .map { .init(name: $0.name, description: $0.description, inputSchema: $0.inputSchema) }
+        : nil,
       model: providerModel.id,
-      enableReasoning: enableReasoning,
+      enableReasoning: model.canReason && enableReasoning,
       provider: .init(
         provider: provider,
         settings: providerSettings,

app/modules/services/LLMService/Tests/LLMModelManagerTests.swift

@@ -1094,15 +1094,15 @@ private func makeSchemaModel(
     maxCompletionTokens: 8_192,
     inputModalities: [.text],
     outputModalities: [.text],
-    supportsChat: true,
-    supportsTools: true,
-    supportsReasoning: false,
-    supportsCompletion: true,
     pricing: pricing ?? Schema.ModelPricing(
       prompt: 1.0,
       completion: 2.0),
     createdAt: Date().timeIntervalSince1970,
-    rankForProgramming: 1)
+    rankForProgramming: 1,
+    supportsChat: true,
+    supportsTools: true,
+    supportsReasoning: false,
+    supportsCompletion: true)
 }
 
 private func makeListModelsOutput(models: [Schema.Model]) -> Schema.ListModelsOutput {

local-server/build.sha256

@@ -1 +1 @@
-e73e92e58623897e84bfc944cf385d6089ea727d20db1a5ad64c5af486c6653b
+a8b504499e156b2f32b035c4cd88fb7acf4f9ac82231cfaf1576032a2a02b3bc

local-server/src/server/endpoints/sendMessage/sendMessage.ts

@@ -103,7 +103,7 @@ export const registerEndpoint = (router: Router, aiProviders: AIProvider[]) => {
 						})
 				}
 
-				await sendMessageImpl({ messages, localExecutable, threadId, tools }, res)
+				await sendMessageImpl({ messages, localExecutable, threadId, tools: tools ?? [] }, res)
 				return
 			}
 
@@ -147,8 +147,8 @@ export const registerEndpoint = (router: Router, aiProviders: AIProvider[]) => {
 				model,
 				abortSignal: abortController.signal,
 				tools: (addProviderOptionsToTools
-					? addProviderOptionsToTools(tools.map(mapTool))
-					: tools.map(mapTool)
+					? addProviderOptionsToTools(tools?.map(mapTool) ?? [])
+					: tools?.map(mapTool)
 				)?.reduce(
 					(acc, tool) => {
 						acc[tool.name] = tool

local-server/src/server/providers/__tests__/ollama.test.ts

@@ -30,7 +30,7 @@ describe("OllamaAIProvider", () => {
 			})
 
 			expect(result.model).toBeDefined()
-			expect(result.generalProviderOptions).toEqual({})
+			expect(result.generalProviderOptions).toBeUndefined()
 		})
 
 		it("should create provider with custom baseURL", () => {
@@ -73,6 +73,17 @@ describe("OllamaAIProvider", () => {
 				delete process.env.OLLAMA_LOCAL_SERVER_PROXY
 			}
 		})
+
+		it("should create provider with thinking option when reasoningBudget is set", () => {
+			const result = provider.build({
+				provider: {},
+				modelName: "llama2",
+				reasoningBudget: 100,
+			})
+
+			expect(result.model).toBeDefined()
+			expect(result.generalProviderOptions).toEqual({ ollama: { think: true } })
+		})
 	})
 
 	describe("listModels", () => {

local-server/src/server/providers/ollama.ts

@@ -1,6 +1,6 @@
 import { AIProvider, AIProviderInput, AIProviderOutput, ProviderModel, ProviderConfig, ModelModality } from "./provider"
 import { APIProviderName } from "@/server/schemas/sendMessageSchema"
-import { createOllama } from "ollama-ai-provider-v2"
+import { createOllama, OllamaCompletionProviderOptions } from "ollama-ai-provider-v2"
 import { LanguageModel } from "ai"
 import { UserFacingError } from "../errors"
 import { ProviderModelFullInfo } from "./provider"
@@ -88,6 +88,7 @@ export class OllamaAIProvider implements AIProvider {
 		const {
 			provider: { baseUrl },
 			modelName,
+			reasoningBudget,
 		} = params
 
 		let finalBaseUrl = process.env["OLLAMA_LOCAL_SERVER_PROXY"] ?? baseUrl ?? "http://localhost:11434/api"
@@ -99,9 +100,16 @@ export class OllamaAIProvider implements AIProvider {
 			baseURL: finalBaseUrl,
 		})
 
+		const providerOptions: OllamaCompletionProviderOptions = {}
+		if (reasoningBudget) {
+			providerOptions.think = true
+		}
+
+		const generalProviderOptions = Object.keys(providerOptions).length > 0 ? { ollama: providerOptions } : undefined
+
 		return {
 			model: provider(modelName) as unknown as LanguageModel,
-			generalProviderOptions: {},
+			generalProviderOptions,
 		}
 	}
 

local-server/src/server/schemas/sendMessageSchema.ts

@@ -2,7 +2,7 @@ export interface SendMessageRequestParams {
 	messages: Message[]
 	system?: string
 	projectRoot: string | undefined
-	tools: Tool[]
+	tools?: Tool[]
 	model: string
 	enableReasoning: boolean
 	provider: APIProvider