fix: filter out empty text blocks from user messages for Gemini compatibility (#10728)

src/api/transform/__tests__/openai-format.spec.ts

@@ -327,6 +327,98 @@ describe("convertToOpenAiMessages", () => {
 		expect(toolMessage.content).toBe("(empty)")
 	})
 
+	describe("empty text block filtering", () => {
+		it("should filter out empty text blocks from user messages (Gemini compatibility)", () => {
+			// This test ensures that user messages with empty text blocks are filtered out
+			// to prevent "must include at least one parts field" error from Gemini (via OpenRouter).
+			// Empty text blocks can occur in edge cases during message construction.
+			const anthropicMessages: Anthropic.Messages.MessageParam[] = [
+				{
+					role: "user",
+					content: [
+						{
+							type: "text",
+							text: "", // Empty text block should be filtered out
+						},
+						{
+							type: "text",
+							text: "Hello, how are you?",
+						},
+					],
+				},
+			]
+
+			const openAiMessages = convertToOpenAiMessages(anthropicMessages)
+			expect(openAiMessages).toHaveLength(1)
+			expect(openAiMessages[0].role).toBe("user")
+
+			const content = openAiMessages[0].content as Array<{ type: string; text?: string }>
+			// Should only have the non-empty text block
+			expect(content).toHaveLength(1)
+			expect(content[0]).toEqual({ type: "text", text: "Hello, how are you?" })
+		})
+
+		it("should not create user message when all text blocks are empty (Gemini compatibility)", () => {
+			// If all text blocks are empty, no user message should be created
+			const anthropicMessages: Anthropic.Messages.MessageParam[] = [
+				{
+					role: "user",
+					content: [
+						{
+							type: "text",
+							text: "", // Empty
+						},
+						{
+							type: "text",
+							text: "", // Also empty
+						},
+					],
+				},
+			]
+
+			const openAiMessages = convertToOpenAiMessages(anthropicMessages)
+			// No messages should be created since all content is empty
+			expect(openAiMessages).toHaveLength(0)
+		})
+
+		it("should preserve image blocks when filtering empty text blocks", () => {
+			const anthropicMessages: Anthropic.Messages.MessageParam[] = [
+				{
+					role: "user",
+					content: [
+						{
+							type: "text",
+							text: "", // Empty text block should be filtered out
+						},
+						{
+							type: "image",
+							source: {
+								type: "base64",
+								media_type: "image/png",
+								data: "base64data",
+							},
+						},
+					],
+				},
+			]
+
+			const openAiMessages = convertToOpenAiMessages(anthropicMessages)
+			expect(openAiMessages).toHaveLength(1)
+			expect(openAiMessages[0].role).toBe("user")
+
+			const content = openAiMessages[0].content as Array<{
+				type: string
+				image_url?: { url: string }
+			}>
+			// Should only have the image block
+			expect(content).toHaveLength(1)
+			expect(content[0]).toEqual({
+				type: "image_url",
+				image_url: { url: "data:image/png;base64,base64data" },
+			})
+		})
+	})
+
 	describe("mergeToolResultText option", () => {
 		it("should merge text content into last tool message when mergeToolResultText is true", () => {
 			const anthropicMessages: Anthropic.Messages.MessageParam[] = [

src/api/transform/openai-format.ts

@@ -138,11 +138,17 @@ export function convertToOpenAiMessages(
 				// }
 
 				// Process non-tool messages
-				if (nonToolMessages.length > 0) {
+				// Filter out empty text blocks to prevent "must include at least one parts field" error
+				// from Gemini (via OpenRouter). Images always have content (base64 data).
+				const filteredNonToolMessages = nonToolMessages.filter(
+					(part) => part.type === "image" || (part.type === "text" && part.text),
+				)
+
+				if (filteredNonToolMessages.length > 0) {
 					// Check if we should merge text into the last tool message
 					// This is critical for reasoning/thinking models where a user message
 					// after tool results causes the model to drop all previous reasoning_content
-					const hasOnlyTextContent = nonToolMessages.every((part) => part.type === "text")
+					const hasOnlyTextContent = filteredNonToolMessages.every((part) => part.type === "text")
 					const hasToolMessages = toolMessages.length > 0
 					const shouldMergeIntoToolMessage =
 						options?.mergeToolResultText && hasToolMessages && hasOnlyTextContent
@@ -153,7 +159,7 @@ export function convertToOpenAiMessages(
 							openAiMessages.length - 1
 						] as OpenAI.Chat.ChatCompletionToolMessageParam
 						if (lastToolMessage?.role === "tool") {
-							const additionalText = nonToolMessages
+							const additionalText = filteredNonToolMessages
 								.map((part) => (part as Anthropic.TextBlockParam).text)
 								.join("\n")
 							lastToolMessage.content = `${lastToolMessage.content}\n\n${additionalText}`
@@ -162,7 +168,7 @@ export function convertToOpenAiMessages(
 						// Standard behavior: add user message with text/image content
 						openAiMessages.push({
 							role: "user",
-							content: nonToolMessages.map((part) => {
+							content: filteredNonToolMessages.map((part) => {
 								if (part.type === "image") {
 									return {
 										type: "image_url",