feat: support tool calls

- Added StreamingToolCallProcessor to handle tool call data and convert JSON chunks to XML format.
- Introduced ToolCallProcessingState to manage the parsing state and track progress.
- Enhanced applyDiffTool to support new tool call functionality based on provider state.
- Updated generateSystemPrompt to include tool call configuration.
- Added supportToolCall utility to determine if a provider supports tool calls.
- Created ToolCallSettingsControl component for user interface settings.
- Localized new tool call settings in multiple languages.

Author: NaccOll

packages/types/src/provider-settings.ts

@@ -78,6 +78,7 @@ const baseProviderSettingsSchema = z.object({
 	includeMaxTokens: z.boolean().optional(),
 	diffEnabled: z.boolean().optional(),
 	todoListEnabled: z.boolean().optional(),
+	toolCallEnabled: z.boolean().optional(),
 	fuzzyMatchThreshold: z.number().optional(),
 	modelTemperature: z.number().nullish(),
 	rateLimitSeconds: z.number().optional(),

src/api/index.ts

@@ -1,6 +1,6 @@
 import { Anthropic } from "@anthropic-ai/sdk"
 
-import type { ProviderSettings, ModelInfo } from "@roo-code/types"
+import type { ProviderSettings, ModelInfo, ToolName } from "@roo-code/types"
 
 import { ApiStream } from "./transform/stream"
 
@@ -37,6 +37,7 @@ import {
 	FireworksHandler,
 } from "./providers"
 import { NativeOllamaHandler } from "./providers/native-ollama"
+import { ToolArgs } from "../core/prompts/tools/types"
 
 export interface SingleCompletionHandler {
 	completePrompt(prompt: string): Promise<string>
@@ -52,6 +53,14 @@ export interface ApiHandlerCreateMessageMetadata {
 	 * Used to enforce "skip once" after a condense operation.
 	 */
 	suppressPreviousResponseId?: boolean
+	/**
+	 * tool call
+	 */
+	tools?: ToolName[]
+	/**
+	 * tool call args
+	 */
+	toolArgs?: ToolArgs
 }
 
 export interface ApiHandler {

src/api/providers/base-provider.ts

@@ -32,4 +32,34 @@ export abstract class BaseProvider implements ApiHandler {
 
 		return countTokens(content, { useWorker: true })
 	}
+
+	/**
+	 * Convert tool schemas to text format for token counting
+	 */
+	protected convertToolSchemasToText(toolSchemas: Anthropic.ToolUnion[]): string {
+		if (toolSchemas.length === 0) {
+			return ""
+		}
+
+		const toolsDescription = toolSchemas
+			.map((tool) => {
+				// Handle different tool types by accessing properties safely
+				const toolName = tool.name
+				let toolText = `Tool: ${toolName}\n`
+
+				// Try to access description and input_schema properties
+				if ("description" in tool) {
+					toolText += `Description: ${tool.description}\n`
+				}
+
+				if ("input_schema" in tool && tool.input_schema && typeof tool.input_schema === "object") {
+					toolText += `Parameters:\n${JSON.stringify(tool.input_schema, null, 2)}\n`
+				}
+
+				return toolText
+			})
+			.join("\n---\n")
+
+		return `Available Tools:\n${toolsDescription}`
+	}
 }

src/api/providers/lm-studio.ts

@@ -15,6 +15,7 @@ import { BaseProvider } from "./base-provider"
 import type { SingleCompletionHandler, ApiHandlerCreateMessageMetadata } from "../index"
 import { getModels, getModelsFromCache } from "./fetchers/modelCache"
 import { getApiRequestTimeout } from "./utils/timeout-config"
+import { getToolRegistry } from "../../core/prompts/tools/schemas/tool-registry"
 
 export class LmStudioHandler extends BaseProvider implements SingleCompletionHandler {
 	protected options: ApiHandlerOptions
@@ -40,6 +41,8 @@ export class LmStudioHandler extends BaseProvider implements SingleCompletionHan
 			{ role: "system", content: systemPrompt },
 			...convertToOpenAiMessages(messages),
 		]
+		const toolCallEnabled = metadata?.tools && metadata.tools.length > 0
+		const toolRegistry = getToolRegistry()
 
 		// -------------------------
 		// Track token usage
@@ -68,7 +71,17 @@ export class LmStudioHandler extends BaseProvider implements SingleCompletionHan
 
 		let inputTokens = 0
 		try {
-			inputTokens = await this.countTokens([{ type: "text", text: systemPrompt }, ...toContentBlocks(messages)])
+			const inputMessages: Anthropic.Messages.ContentBlockParam[] = [{ type: "text", text: systemPrompt }]
+			if (toolCallEnabled) {
+				const toolSchemas: Anthropic.ToolUnion[] = toolRegistry.generateAnthropicToolSchemas(
+					metadata.tools!,
+					metadata.toolArgs,
+				)
+				const toolsText = this.convertToolSchemasToText(toolSchemas)
+				inputMessages.push({ type: "text", text: toolsText })
+			}
+			inputMessages.push(...toContentBlocks(messages))
+			inputTokens = await this.countTokens(inputMessages)
 		} catch (err) {
 			console.error("[LmStudio] Failed to count input tokens:", err)
 			inputTokens = 0
@@ -83,6 +96,10 @@ export class LmStudioHandler extends BaseProvider implements SingleCompletionHan
 				temperature: this.options.modelTemperature ?? LMSTUDIO_DEFAULT_TEMPERATURE,
 				stream: true,
 			}
+			if (toolCallEnabled) {
+				params.tools = toolRegistry.generateFunctionCallSchemas(metadata.tools!, metadata.toolArgs)
+				params.tool_choice = "auto"
+			}
 
 			if (this.options.lmStudioSpeculativeDecodingEnabled && this.options.lmStudioDraftModelId) {
 				params.draft_model = this.options.lmStudioDraftModelId
@@ -108,6 +125,9 @@ export class LmStudioHandler extends BaseProvider implements SingleCompletionHan
 						yield processedChunk
 					}
 				}
+				if (delta?.tool_calls) {
+					yield { type: "tool_call", toolCalls: delta.tool_calls, toolCallType: "openai" }
+				}
 			}
 
 			for (const processedChunk of matcher.final()) {

src/api/providers/openai.ts

@@ -24,6 +24,7 @@ import { DEFAULT_HEADERS } from "./constants"
 import { BaseProvider } from "./base-provider"
 import type { SingleCompletionHandler, ApiHandlerCreateMessageMetadata } from "../index"
 import { getApiRequestTimeout } from "./utils/timeout-config"
+import { getToolRegistry } from "../../core/prompts/tools/schemas/tool-registry"
 
 // TODO: Rename this to OpenAICompatibleHandler. Also, I think the
 // `OpenAINativeHandler` can subclass from this, since it's obviously
@@ -92,6 +93,9 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 		const deepseekReasoner = modelId.includes("deepseek-reasoner") || enabledR1Format
 		const ark = modelUrl.includes(".volces.com")
 
+		const toolCallEnabled = metadata?.tools && metadata.tools.length > 0
+		const toolRegistry = getToolRegistry()
+
 		if (modelId.includes("o1") || modelId.includes("o3") || modelId.includes("o4")) {
 			yield* this.handleO3FamilyMessage(modelId, systemPrompt, messages)
 			return
@@ -163,6 +167,10 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 				...(isGrokXAI ? {} : { stream_options: { include_usage: true } }),
 				...(reasoning && reasoning),
 			}
+			if (toolCallEnabled) {
+				requestOptions.tools = toolRegistry.generateFunctionCallSchemas(metadata.tools!, metadata.toolArgs)
+				requestOptions.tool_choice = "auto"
+			}
 
 			// Add max_tokens if needed
 			this.addMaxTokensIfNeeded(requestOptions, modelInfo)
@@ -198,6 +206,9 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 						text: (delta.reasoning_content as string | undefined) || "",
 					}
 				}
+				if (delta?.tool_calls) {
+					yield { type: "tool_call", toolCalls: delta.tool_calls, toolCallType: "openai" }
+				}
 				if (chunk.usage) {
 					lastUsage = chunk.usage
 				}

src/api/providers/openrouter.ts

@@ -24,7 +24,8 @@ import { getModelEndpoints } from "./fetchers/modelEndpointCache"
 
 import { DEFAULT_HEADERS } from "./constants"
 import { BaseProvider } from "./base-provider"
-import type { SingleCompletionHandler } from "../index"
+import type { SingleCompletionHandler, ApiHandlerCreateMessageMetadata } from "../index"
+import { getToolRegistry } from "../../core/prompts/tools/schemas/tool-registry"
 
 // Add custom interface for OpenRouter params.
 type OpenRouterChatCompletionParams = OpenAI.Chat.ChatCompletionCreateParams & {
@@ -72,10 +73,13 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
 	override async *createMessage(
 		systemPrompt: string,
 		messages: Anthropic.Messages.MessageParam[],
+		metadata?: ApiHandlerCreateMessageMetadata,
 	): AsyncGenerator<ApiStreamChunk> {
 		const model = await this.fetchModel()
 
 		let { id: modelId, maxTokens, temperature, topP, reasoning } = model
+		const toolCallEnabled = metadata?.tools && metadata.tools.length > 0
+		const toolRegistry = getToolRegistry()
 
 		// OpenRouter sends reasoning tokens by default for Gemini 2.5 Pro
 		// Preview even if you don't request them. This is not the default for
@@ -133,6 +137,10 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
 			...(transforms && { transforms }),
 			...(reasoning && { reasoning }),
 		}
+		if (toolCallEnabled) {
+			completionParams.tools = toolRegistry.generateFunctionCallSchemas(metadata.tools!, metadata.toolArgs!)
+			completionParams.tool_choice = "auto"
+		}
 
 		const stream = await this.client.chat.completions.create(completionParams)
 
@@ -156,6 +164,10 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
 				yield { type: "text", text: delta.content }
 			}
 
+			if (delta?.tool_calls) {
+				yield { type: "tool_call", toolCalls: delta.tool_calls, toolCallType: "openai" }
+			}
+
 			if (chunk.usage) {
 				lastUsage = chunk.usage
 			}

src/api/transform/stream.ts

@@ -1,6 +1,13 @@
+import { ToolCallProviderType } from "../../shared/tools"
+
 export type ApiStream = AsyncGenerator<ApiStreamChunk>
 
-export type ApiStreamChunk = ApiStreamTextChunk | ApiStreamUsageChunk | ApiStreamReasoningChunk | ApiStreamError
+export type ApiStreamChunk =
+	| ApiStreamTextChunk
+	| ApiStreamUsageChunk
+	| ApiStreamReasoningChunk
+	| ApiStreamError
+	| ApiStreamToolCallChunk
 
 export interface ApiStreamError {
 	type: "error"
@@ -27,3 +34,9 @@ export interface ApiStreamUsageChunk {
 	reasoningTokens?: number
 	totalCost?: number
 }
+
+export interface ApiStreamToolCallChunk {
+	type: "tool_call"
+	toolCalls: any
+	toolCallType: ToolCallProviderType
+}

src/core/assistant-message/AssistantMessageParser.ts

@@ -1,6 +1,7 @@
 import { type ToolName, toolNames } from "@roo-code/types"
 import { TextContent, ToolUse, ToolParamName, toolParamNames } from "../../shared/tools"
 import { AssistantMessageContent } from "./parseAssistantMessage"
+import { ToolCallParam } from "../task/tool-call-helper"
 
 /**
  * Parser for assistant messages. Maintains state between chunks
@@ -51,7 +52,7 @@ export class AssistantMessageParser {
 	 * Process a new chunk of text and update the parser state.
 	 * @param chunk The new chunk of text to process.
 	 */
-	public processChunk(chunk: string): AssistantMessageContent[] {
+	public processChunk(chunk: string, toolCallParam?: ToolCallParam): AssistantMessageContent[] {
 		if (this.accumulator.length + chunk.length > this.MAX_ACCUMULATOR_SIZE) {
 			throw new Error("Assistant message exceeds maximum allowed size")
 		}
@@ -174,6 +175,11 @@ export class AssistantMessageParser {
 						name: extractedToolName as ToolName,
 						params: {},
 						partial: true,
+						toolUseId: toolCallParam && toolCallParam.toolUserId ? toolCallParam.toolUserId : undefined,
+						toolUseParam:
+							toolCallParam && toolCallParam?.anthropicContent
+								? toolCallParam?.anthropicContent
+								: undefined,
 					}
 
 					this.currentToolUseStartIndex = this.accumulator.length

src/core/assistant-message/parseAssistantMessage.ts

@@ -1,10 +1,14 @@
 import { type ToolName, toolNames } from "@roo-code/types"
 
 import { TextContent, ToolUse, ToolParamName, toolParamNames } from "../../shared/tools"
+import { ToolCallParam } from "../task/tool-call-helper"
 
 export type AssistantMessageContent = TextContent | ToolUse
 
-export function parseAssistantMessage(assistantMessage: string): AssistantMessageContent[] {
+export function parseAssistantMessage(
+	assistantMessage: string,
+	toolCallParam?: ToolCallParam,
+): AssistantMessageContent[] {
 	let contentBlocks: AssistantMessageContent[] = []
 	let currentTextContent: TextContent | undefined = undefined
 	let currentTextContentStartIndex = 0
@@ -103,6 +107,9 @@ export function parseAssistantMessage(assistantMessage: string): AssistantMessag
 					name: toolUseOpeningTag.slice(1, -1) as ToolName,
 					params: {},
 					partial: true,
+					toolUseId: toolCallParam && toolCallParam.toolUserId ? toolCallParam.toolUserId : undefined,
+					toolUseParam:
+						toolCallParam && toolCallParam?.anthropicContent ? toolCallParam?.anthropicContent : undefined,
 				}
 
 				currentToolUseStartIndex = accumulator.length

src/core/assistant-message/presentAssistantMessage.ts

@@ -33,6 +33,7 @@ import { Task } from "../task/Task"
 import { codebaseSearchTool } from "../tools/codebaseSearchTool"
 import { experiments, EXPERIMENT_IDS } from "../../shared/experiments"
 import { applyDiffToolLegacy } from "../tools/applyDiffTool"
+import Anthropic from "@anthropic-ai/sdk"
 
 /**
  * Processes and presents assistant message content to the user interface.
@@ -61,6 +62,7 @@ export async function presentAssistantMessage(cline: Task) {
 		return
 	}
 
+	const toolCallEnabled = cline.apiConfiguration?.toolCallEnabled
 	cline.presentAssistantMessageLocked = true
 	cline.presentAssistantMessageHasPendingUpdates = false
 
@@ -245,12 +247,28 @@ export async function presentAssistantMessage(cline: Task) {
 			}
 
 			const pushToolResult = (content: ToolResponse) => {
-				cline.userMessageContent.push({ type: "text", text: `${toolDescription()} Result:` })
-
+				const newUserMessages: (Anthropic.TextBlockParam | Anthropic.ImageBlockParam)[] = [
+					{ type: "text", text: `${toolDescription()} Result:` },
+				]
 				if (typeof content === "string") {
-					cline.userMessageContent.push({ type: "text", text: content || "(tool did not return anything)" })
+					newUserMessages.push({ type: "text", text: content || "(tool did not return anything)" })
+				} else {
+					newUserMessages.push(...content)
+				}
+
+				if (toolCallEnabled) {
+					const lastToolUseMessage = cline.assistantMessageContent.find((msg) => msg.type === "tool_use")
+					if (lastToolUseMessage && lastToolUseMessage.toolUseId) {
+						const toolUseId = lastToolUseMessage.toolUseId
+						const toolMessage: Anthropic.ToolResultBlockParam = {
+							tool_use_id: toolUseId,
+							type: "tool_result",
+							content: newUserMessages,
+						}
+						cline.userMessageContent.push(toolMessage)
+					}
 				} else {
-					cline.userMessageContent.push(...content)
+					cline.userMessageContent.push(...newUserMessages)
 				}
 
 				// Once a tool result has been collected, ignore all other tool
@@ -429,7 +447,7 @@ export async function presentAssistantMessage(cline: Task) {
 						)
 					}
 
-					if (isMultiFileApplyDiffEnabled) {
+					if (isMultiFileApplyDiffEnabled || toolCallEnabled) {
 						await checkpointSaveAndMark(cline)
 						await applyDiffTool(cline, block, askApproval, handleError, pushToolResult, removeClosingTag)
 					} else {