feat: enable native tool calling for gemini provider

Author: hannesrudolph

packages/types/src/providers/gemini.ts

@@ -35,6 +35,7 @@ export const geminiModels = {
 		maxTokens: 64_000,
 		contextWindow: 1_048_576,
 		supportsImages: true,
+		supportsNativeTools: true,
 		supportsPromptCache: true,
 		inputPrice: 2.5, // This is the pricing for prompts above 200k tokens.
 		outputPrice: 15,
@@ -62,6 +63,7 @@ export const geminiModels = {
 		maxTokens: 65_535,
 		contextWindow: 1_048_576,
 		supportsImages: true,
+		supportsNativeTools: true,
 		supportsPromptCache: true,
 		inputPrice: 2.5, // This is the pricing for prompts above 200k tokens.
 		outputPrice: 15,
@@ -88,6 +90,7 @@ export const geminiModels = {
 		maxTokens: 65_535,
 		contextWindow: 1_048_576,
 		supportsImages: true,
+		supportsNativeTools: true,
 		supportsPromptCache: true,
 		inputPrice: 2.5, // This is the pricing for prompts above 200k tokens.
 		outputPrice: 15,
@@ -112,6 +115,7 @@ export const geminiModels = {
 		maxTokens: 65_535,
 		contextWindow: 1_048_576,
 		supportsImages: true,
+		supportsNativeTools: true,
 		supportsPromptCache: true,
 		inputPrice: 2.5, // This is the pricing for prompts above 200k tokens.
 		outputPrice: 15,
@@ -140,6 +144,7 @@ export const geminiModels = {
 		maxTokens: 65_536,
 		contextWindow: 1_048_576,
 		supportsImages: true,
+		supportsNativeTools: true,
 		supportsPromptCache: true,
 		inputPrice: 0.3,
 		outputPrice: 2.5,
@@ -152,6 +157,7 @@ export const geminiModels = {
 		maxTokens: 65_536,
 		contextWindow: 1_048_576,
 		supportsImages: true,
+		supportsNativeTools: true,
 		supportsPromptCache: true,
 		inputPrice: 0.3,
 		outputPrice: 2.5,
@@ -164,6 +170,7 @@ export const geminiModels = {
 		maxTokens: 64_000,
 		contextWindow: 1_048_576,
 		supportsImages: true,
+		supportsNativeTools: true,
 		supportsPromptCache: true,
 		inputPrice: 0.3,
 		outputPrice: 2.5,
@@ -178,6 +185,7 @@ export const geminiModels = {
 		maxTokens: 65_536,
 		contextWindow: 1_048_576,
 		supportsImages: true,
+		supportsNativeTools: true,
 		supportsPromptCache: true,
 		inputPrice: 0.1,
 		outputPrice: 0.4,
@@ -190,6 +198,7 @@ export const geminiModels = {
 		maxTokens: 65_536,
 		contextWindow: 1_048_576,
 		supportsImages: true,
+		supportsNativeTools: true,
 		supportsPromptCache: true,
 		inputPrice: 0.1,
 		outputPrice: 0.4,

src/api/providers/gemini.ts

@@ -5,6 +5,8 @@ import {
 	type GenerateContentParameters,
 	type GenerateContentConfig,
 	type GroundingMetadata,
+	FunctionCallingConfigMode,
+	Content,
 } from "@google/genai"
 import type { JWTInput } from "google-auth-library"
 
@@ -101,9 +103,9 @@ export class GeminiHandler extends BaseProvider implements SingleCompletionHandl
 			return true
 		})
 
-		const contents = geminiMessages.map((message) =>
-			convertAnthropicMessageToGemini(message, { includeThoughtSignatures }),
-		)
+		const contents: Content[] = geminiMessages
+			.map((message) => convertAnthropicMessageToGemini(message, { includeThoughtSignatures }))
+			.flat()
 
 		const tools: GenerateContentConfig["tools"] = []
 		if (this.options.enableUrlContext) {
@@ -114,6 +116,16 @@ export class GeminiHandler extends BaseProvider implements SingleCompletionHandl
 			tools.push({ googleSearch: {} })
 		}
 
+		if (metadata?.tools && metadata.tools.length > 0) {
+			tools.push({
+				functionDeclarations: metadata.tools.map((tool) => ({
+					name: (tool as any).function.name,
+					description: (tool as any).function.description,
+					parametersJsonSchema: (tool as any).function.parameters,
+				})),
+			})
+		}
+
 		// Determine temperature respecting model capabilities and defaults:
 		// - If supportsTemperature is explicitly false, ignore user overrides
 		//   and pin to the model's defaultTemperature (or omit if undefined).
@@ -133,6 +145,25 @@ export class GeminiHandler extends BaseProvider implements SingleCompletionHandl
 			...(tools.length > 0 ? { tools } : {}),
 		}
 
+		if (metadata?.tool_choice) {
+			config.toolConfig = {
+				functionCallingConfig: {
+					mode:
+						metadata.tool_choice === "auto"
+							? FunctionCallingConfigMode.AUTO
+							: metadata.tool_choice === "required"
+								? FunctionCallingConfigMode.ANY
+								: metadata.tool_choice === "none"
+									? FunctionCallingConfigMode.NONE
+									: FunctionCallingConfigMode.ANY,
+					allowedFunctionNames:
+						typeof metadata.tool_choice === "object" && "function" in metadata.tool_choice
+							? [metadata.tool_choice.function.name]
+							: undefined,
+				},
+			}
+		}
+
 		const params: GenerateContentParameters = { model, contents, config }
 		try {
 			const result = await this.client.models.generateContentStream(params)
@@ -159,6 +190,7 @@ export class GeminiHandler extends BaseProvider implements SingleCompletionHandl
 							thought?: boolean
 							text?: string
 							thoughtSignature?: string
+							functionCall?: { name: string; args: Record<string, unknown> }
 						}>) {
 							// Capture thought signatures so they can be persisted into API history.
 							const thoughtSignature = part.thoughtSignature
@@ -174,6 +206,13 @@ export class GeminiHandler extends BaseProvider implements SingleCompletionHandl
 								if (part.text) {
 									yield { type: "reasoning", text: part.text }
 								}
+							} else if (part.functionCall) {
+								yield {
+									type: "tool_call",
+									id: part.functionCall.name, // Gemini doesn't provide call IDs, so we use the function name
+									name: part.functionCall.name,
+									arguments: JSON.stringify(part.functionCall.args),
+								}
 							} else {
 								// This is regular content
 								if (part.text) {
@@ -351,7 +390,10 @@ export class GeminiHandler extends BaseProvider implements SingleCompletionHandl
 			const countTokensRequest = {
 				model,
 				// Token counting does not need encrypted continuation; always drop thoughtSignature.
-				contents: convertAnthropicContentToGemini(content, { includeThoughtSignatures: false }),
+				contents: convertAnthropicMessageToGemini(
+					{ role: "user", content },
+					{ includeThoughtSignatures: false },
+				),
 			}
 
 			const response = await this.client.models.countTokens(countTokensRequest)

src/api/transform/gemini-format.ts

@@ -105,9 +105,37 @@ export function convertAnthropicContentToGemini(
 export function convertAnthropicMessageToGemini(
 	message: Anthropic.Messages.MessageParam,
 	options?: { includeThoughtSignatures?: boolean },
-): Content {
-	return {
-		role: message.role === "assistant" ? "model" : "user",
-		parts: convertAnthropicContentToGemini(message.content, options),
+): Content | Content[] {
+	const content = Array.isArray(message.content) ? message.content : [{ type: "text", text: message.content ?? "" }]
+	const toolUseParts = content.filter((block) => block.type === "tool_use") as Anthropic.ToolUseBlock[]
+	const toolResultParts = content.filter((block) => block.type === "tool_result") as Anthropic.ToolResultBlockParam[]
+	const otherParts = content.filter((block) => block.type !== "tool_use" && block.type !== "tool_result") as (
+		| Anthropic.TextBlockParam
+		| Anthropic.ImageBlockParam
+	)[]
+
+	const contents: Content[] = []
+
+	if (otherParts.length > 0) {
+		contents.push({
+			role: message.role === "assistant" ? "model" : "user",
+			parts: convertAnthropicContentToGemini(otherParts, options),
+		})
 	}
+
+	if (toolUseParts.length > 0) {
+		contents.push({
+			role: "model",
+			parts: convertAnthropicContentToGemini(toolUseParts, options),
+		})
+	}
+
+	if (toolResultParts.length > 0) {
+		contents.push({
+			role: "user",
+			parts: convertAnthropicContentToGemini(toolResultParts, options),
+		})
+	}
+
+	return contents
 }