fix: handle special case when VS Code LM API returns tokenCount of 4

- Add special handling in internalCountTokens method for LanguageModelChatMessage
- When tokenCount equals 4, convert message to string and recalculate
- Add comprehensive test coverage for the special case
- Fixes #6290

Author: roomote

src/api/providers/__tests__/vscode-lm.spec.ts

@@ -300,4 +300,75 @@ describe("VsCodeLmHandler", () => {
 			await expect(promise).rejects.toThrow("VSCode LM completion error: Completion failed")
 		})
 	})
+
+	describe("countTokens", () => {
+		beforeEach(() => {
+			const mockModel = { ...mockLanguageModelChat }
+			;(vscode.lm.selectChatModels as Mock).mockResolvedValueOnce([mockModel])
+
+			// Override the default client with our test client
+			handler["client"] = mockLanguageModelChat
+			// Set up cancellation token
+			handler["currentRequestCancellation"] = new vscode.CancellationTokenSource()
+		})
+
+		it("should count tokens for string input", async () => {
+			mockLanguageModelChat.countTokens.mockResolvedValue(10)
+
+			const result = await handler.countTokens([{ type: "text", text: "Hello world" }])
+
+			expect(result).toBe(10)
+			expect(mockLanguageModelChat.countTokens).toHaveBeenCalledWith("Hello world", expect.any(Object))
+		})
+
+		it("should handle special case when LanguageModelChatMessage returns tokenCount of 4", async () => {
+			// First call returns 4 (triggering the special case)
+			// Second call returns the actual count after string conversion
+			mockLanguageModelChat.countTokens.mockResolvedValueOnce(4).mockResolvedValueOnce(25)
+
+			// Use the mocked vscode.LanguageModelChatMessage.User to create a proper message
+			const mockMessage = vscode.LanguageModelChatMessage.User("This is a test message")
+
+			const result = await handler["internalCountTokens"](mockMessage)
+
+			expect(result).toBe(25)
+			expect(mockLanguageModelChat.countTokens).toHaveBeenCalledTimes(2)
+			// First call with the message object
+			expect(mockLanguageModelChat.countTokens).toHaveBeenNthCalledWith(1, mockMessage, expect.any(Object))
+			// Second call with the extracted string
+			expect(mockLanguageModelChat.countTokens).toHaveBeenNthCalledWith(
+				2,
+				"This is a test message",
+				expect.any(Object),
+			)
+		})
+
+		it("should not recalculate when tokenCount is not 4", async () => {
+			mockLanguageModelChat.countTokens.mockResolvedValue(10)
+
+			// Use the mocked vscode.LanguageModelChatMessage.User to create a proper message
+			const mockMessage = vscode.LanguageModelChatMessage.User("This is a test message")
+
+			const result = await handler["internalCountTokens"](mockMessage)
+
+			expect(result).toBe(10)
+			expect(mockLanguageModelChat.countTokens).toHaveBeenCalledTimes(1)
+		})
+
+		it("should handle image blocks", async () => {
+			// The countTokens method converts to string, so it won't trigger the special case
+			mockLanguageModelChat.countTokens.mockResolvedValue(7)
+
+			const result = await handler.countTokens([
+				{ type: "text", text: "Hello" },
+				{ type: "image", source: { type: "base64", media_type: "image/png", data: "base64data" } },
+				{ type: "text", text: " world" },
+			])
+
+			expect(result).toBe(7)
+			// Should only be called once since it's a string, not a LanguageModelChatMessage
+			expect(mockLanguageModelChat.countTokens).toHaveBeenCalledTimes(1)
+			expect(mockLanguageModelChat.countTokens).toHaveBeenCalledWith("Hello[IMAGE] world", expect.any(Object))
+		})
+	})
 })

src/api/providers/vscode-lm.ts

@@ -225,13 +225,40 @@ export class VsCodeLmHandler extends BaseProvider implements SingleCompletionHan
 
 			if (typeof text === "string") {
 				tokenCount = await this.client.countTokens(text, this.currentRequestCancellation.token)
-			} else if (text instanceof vscode.LanguageModelChatMessage) {
+			} else if (text && typeof text === "object" && "content" in text) {
+				// Handle LanguageModelChatMessage-like objects
 				// For chat messages, ensure we have content
 				if (!text.content || (Array.isArray(text.content) && text.content.length === 0)) {
 					console.debug("Roo Code <Language Model API>: Empty chat message content")
 					return 0
 				}
 				tokenCount = await this.client.countTokens(text, this.currentRequestCancellation.token)
+
+				// Special handling: if tokenCount is exactly 4 for a LanguageModelChatMessage,
+				// convert to string and recalculate
+				if (tokenCount === 4) {
+					console.debug(
+						"Roo Code <Language Model API>: Token count is 4 for LanguageModelChatMessage, converting to string and recalculating",
+					)
+
+					// Convert message content to string
+					let messageText = ""
+					if (Array.isArray(text.content)) {
+						for (const part of text.content) {
+							if (part && typeof part === "object" && "value" in part && typeof part.value === "string") {
+								messageText += part.value
+							}
+						}
+					} else if (typeof text.content === "string") {
+						messageText = text.content
+					}
+
+					// Recalculate using string
+					if (messageText) {
+						tokenCount = await this.client.countTokens(messageText, this.currentRequestCancellation.token)
+						console.debug(`Roo Code <Language Model API>: Recalculated token count: ${tokenCount}`)
+					}
+				}
 			} else {
 				console.warn("Roo Code <Language Model API>: Invalid input type for token counting")
 				return 0