diff --git a/src/api/transform/__tests__/openai-format.spec.ts b/src/api/transform/__tests__/openai-format.spec.ts
index d5d4404837..111f79101f 100644
--- a/src/api/transform/__tests__/openai-format.spec.ts
+++ b/src/api/transform/__tests__/openai-format.spec.ts
@@ -319,7 +319,55 @@ describe("convertToOpenAiMessages", () => {
 			)
 		})
 
-		it("should NOT merge text when images are present (fall back to user message)", () => {
+		it("should merge text and send images separately when mergeToolResultText is true", () => {
+			const anthropicMessages: Anthropic.Messages.MessageParam[] = [
+				{
+					role: "user",
+					content: [
+						{
+							type: "tool_result",
+							tool_use_id: "tool-123",
+							content: "Tool result content",
+						},
+						{
+							type: "text",
+							text: "<environment_details>Context info</environment_details>",
+						},
+						{
+							type: "image",
+							source: {
+								type: "base64",
+								media_type: "image/png",
+								data: "base64data",
+							},
+						},
+					],
+				},
+			]
+
+			const openAiMessages = convertToOpenAiMessages(anthropicMessages, { mergeToolResultText: true })
+
+			// Should produce a tool message with merged text AND a user message for the image
+			expect(openAiMessages).toHaveLength(2)
+
+			// First message: tool message with merged text
+			const toolMessage = openAiMessages[0] as OpenAI.Chat.ChatCompletionToolMessageParam
+			expect(toolMessage.role).toBe("tool")
+			expect(toolMessage.tool_call_id).toBe("tool-123")
+			expect(toolMessage.content).toBe(
+				"Tool result content\n\n<environment_details>Context info</environment_details>",
+			)
+
+			// Second message: user message with only the image
+			expect(openAiMessages[1].role).toBe("user")
+			const userContent = openAiMessages[1].content as Array<{ type: string; image_url?: { url: string } }>
+			expect(Array.isArray(userContent)).toBe(true)
+			expect(userContent).toHaveLength(1)
+			expect(userContent[0].type).toBe("image_url")
+			expect(userContent[0].image_url?.url).toBe("data:image/png;base64,base64data")
+		})
+
+		it("should send only images as user message when no text content exists with mergeToolResultText", () => {
 			const anthropicMessages: Anthropic.Messages.MessageParam[] = [
 				{
 					role: "user",
@@ -343,9 +391,13 @@ describe("convertToOpenAiMessages", () => {
 
 			const openAiMessages = convertToOpenAiMessages(anthropicMessages, { mergeToolResultText: true })
 
-			// Should produce a tool message AND a user message (because image is present)
+			// Should produce a tool message AND a user message (only image, no text to merge)
 			expect(openAiMessages).toHaveLength(2)
 			expect((openAiMessages[0] as OpenAI.Chat.ChatCompletionToolMessageParam).role).toBe("tool")
+			// Tool message content should NOT be modified since there's no text to merge
+			expect((openAiMessages[0] as OpenAI.Chat.ChatCompletionToolMessageParam).content).toBe(
+				"Tool result content",
+			)
 			expect(openAiMessages[1].role).toBe("user")
 		})
 
diff --git a/src/api/transform/openai-format.ts b/src/api/transform/openai-format.ts
index ad02be5541..f182a2aac1 100644
--- a/src/api/transform/openai-format.ts
+++ b/src/api/transform/openai-format.ts
@@ -138,24 +138,40 @@ export function convertToOpenAiMessages(
 
 				// Process non-tool messages
 				if (nonToolMessages.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 hasToolMessages = toolMessages.length > 0
-					const shouldMergeIntoToolMessage =
-						options?.mergeToolResultText && hasToolMessages && hasOnlyTextContent
 
-					if (shouldMergeIntoToolMessage) {
+					if (options?.mergeToolResultText && hasToolMessages) {
+						// When mergeToolResultText is enabled, separate text and images
+						// Merge text into the last tool message, and send images separately
+						// This is critical for providers like NVIDIA NIM that don't allow user messages after tool messages
+						const textMessages = nonToolMessages.filter(
+							(part) => part.type === "text",
+						) as Anthropic.TextBlockParam[]
+						const imageMessages = nonToolMessages.filter(
+							(part) => part.type === "image",
+						) as Anthropic.ImageBlockParam[]
+
 						// Merge text content into the last tool message
-						const lastToolMessage = openAiMessages[
-							openAiMessages.length - 1
-						] as OpenAI.Chat.ChatCompletionToolMessageParam
-						if (lastToolMessage?.role === "tool") {
-							const additionalText = nonToolMessages
-								.map((part) => (part as Anthropic.TextBlockParam).text)
-								.join("\n")
-							lastToolMessage.content = `${lastToolMessage.content}\n\n${additionalText}`
+						if (textMessages.length > 0) {
+							const lastToolMessage = openAiMessages[
+								openAiMessages.length - 1
+							] as OpenAI.Chat.ChatCompletionToolMessageParam
+							if (lastToolMessage?.role === "tool") {
+								const additionalText = textMessages.map((part) => part.text).join("\n")
+								lastToolMessage.content = `${lastToolMessage.content}\n\n${additionalText}`
+							}
+						}
+
+						// Send images as a separate user message if any
+						// Note: Images must still be sent as user messages since tool messages don't support images
+						if (imageMessages.length > 0) {
+							openAiMessages.push({
+								role: "user",
+								content: imageMessages.map((part) => ({
+									type: "image_url",
+									image_url: { url: `data:${part.source.media_type};base64,${part.source.data}` },
+								})),
+							})
 						}
 					} else {
 						// Standard behavior: add user message with text/image content