fix: improve GLM-4.5 model handling to prevent hallucination and enhance tool understanding

roomote · roomote · commit 667a79bb3704 · 2025-08-11T17:32:00.000Z
- Add GLM-specific system prompt enhancements to prevent file hallucination - Include clear instructions for tool usage protocol and content management - Implement message preprocessing for better GLM model understanding - Add token limit adjustments and model-specific parameters for GLM-4.5 - Enhance completePrompt method with instruction prefix for GLM models - Add comprehensive tests for GLM-specific functionality Fixes #6942
diff --git a/src/api/providers/__tests__/zai.spec.ts b/src/api/providers/__tests__/zai.spec.ts
@@ -193,7 +193,6 @@ describe("ZAiHandler", () => {
 
 		it("createMessage should pass correct parameters to Z AI client", async () => {
 			const modelId: InternationalZAiModelId = "glm-4.5"
-			const modelInfo = internationalZAiModels[modelId]
 			const handlerWithModel = new ZAiHandler({
 				apiModelId: modelId,
 				zaiApiKey: "test-zai-api-key",
@@ -216,14 +215,159 @@ describe("ZAiHandler", () => {
 			const messageGenerator = handlerWithModel.createMessage(systemPrompt, messages)
 			await messageGenerator.next()
 
+			// For GLM-4.5, expect enhanced system prompt and adjusted parameters
 			expect(mockCreate).toHaveBeenCalledWith(
 				expect.objectContaining({
 					model: modelId,
-					max_tokens: modelInfo.maxTokens,
+					max_tokens: 32768, // Adjusted for GLM models
 					temperature: ZAI_DEFAULT_TEMPERATURE,
-					messages: expect.arrayContaining([{ role: "system", content: systemPrompt }]),
+					messages: expect.arrayContaining([
+						{
+							role: "system",
+							content: expect.stringContaining(systemPrompt), // Contains original prompt plus enhancements
+						},
+					]),
 					stream: true,
 					stream_options: { include_usage: true },
+					top_p: 0.95,
+					frequency_penalty: 0.1,
+					presence_penalty: 0.1,
+				}),
+			)
+		})
+
+		it("should enhance system prompt for GLM-4.5 models", async () => {
+			const modelId: InternationalZAiModelId = "glm-4.5"
+			const handlerWithGLM = new ZAiHandler({
+				apiModelId: modelId,
+				zaiApiKey: "test-zai-api-key",
+				zaiApiLine: "international",
+			})
+
+			mockCreate.mockImplementationOnce(() => {
+				return {
+					[Symbol.asyncIterator]: () => ({
+						async next() {
+							return { done: true }
+						},
+					}),
+				}
+			})
+
+			const systemPrompt = "Test system prompt"
+			const messages: Anthropic.Messages.MessageParam[] = [{ role: "user", content: "Test message" }]
+
+			const messageGenerator = handlerWithGLM.createMessage(systemPrompt, messages)
+			await messageGenerator.next()
+
+			// Check that the system prompt was enhanced with GLM-specific instructions
+			expect(mockCreate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					messages: expect.arrayContaining([
+						{
+							role: "system",
+							content: expect.stringContaining("CRITICAL INSTRUCTIONS FOR GLM MODEL"),
+						},
+					]),
+				}),
+			)
+		})
+
+		it("should apply max token adjustment for GLM-4.5 models", async () => {
+			const modelId: InternationalZAiModelId = "glm-4.5"
+			const handlerWithGLM = new ZAiHandler({
+				apiModelId: modelId,
+				zaiApiKey: "test-zai-api-key",
+				zaiApiLine: "international",
+			})
+
+			mockCreate.mockImplementationOnce(() => {
+				return {
+					[Symbol.asyncIterator]: () => ({
+						async next() {
+							return { done: true }
+						},
+					}),
+				}
+			})
+
+			const messageGenerator = handlerWithGLM.createMessage("system", [])
+			await messageGenerator.next()
+
+			// Check that max_tokens is capped at 32768 for GLM models
+			expect(mockCreate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					max_tokens: 32768,
+					top_p: 0.95,
+					frequency_penalty: 0.1,
+					presence_penalty: 0.1,
+				}),
+			)
+		})
+
+		it("should enhance prompt in completePrompt for GLM-4.5 models", async () => {
+			const modelId: InternationalZAiModelId = "glm-4.5"
+			const handlerWithGLM = new ZAiHandler({
+				apiModelId: modelId,
+				zaiApiKey: "test-zai-api-key",
+				zaiApiLine: "international",
+			})
+
+			const expectedResponse = "Test response"
+			mockCreate.mockResolvedValueOnce({ choices: [{ message: { content: expectedResponse } }] })
+
+			const testPrompt = "Test prompt"
+			await handlerWithGLM.completePrompt(testPrompt)
+
+			// Check that the prompt was enhanced with GLM-specific prefix
+			expect(mockCreate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					messages: [
+						{
+							role: "user",
+							content: expect.stringContaining(
+								"[INSTRUCTION] Please provide a direct and accurate response",
+							),
+						},
+					],
+					temperature: ZAI_DEFAULT_TEMPERATURE,
+					max_tokens: 4096,
+				}),
+			)
+		})
+
+		it("should handle GLM-4.5-air model correctly", async () => {
+			const modelId: InternationalZAiModelId = "glm-4.5-air"
+			const handlerWithGLMAir = new ZAiHandler({
+				apiModelId: modelId,
+				zaiApiKey: "test-zai-api-key",
+				zaiApiLine: "international",
+			})
+
+			mockCreate.mockImplementationOnce(() => {
+				return {
+					[Symbol.asyncIterator]: () => ({
+						async next() {
+							return { done: true }
+						},
+					}),
+				}
+			})
+
+			const messageGenerator = handlerWithGLMAir.createMessage("system", [])
+			await messageGenerator.next()
+
+			// Should apply GLM enhancements for glm-4.5-air as well
+			expect(mockCreate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					model: modelId,
+					max_tokens: 32768,
+					messages: expect.arrayContaining([
+						{
+							role: "system",
+							content: expect.stringContaining("CRITICAL INSTRUCTIONS FOR GLM MODEL"),
+						},
+					]),
 				}),
 			)
 		})
diff --git a/src/api/providers/zai.ts b/src/api/providers/zai.ts
@@ -7,12 +7,19 @@ import {
 	type MainlandZAiModelId,
 	ZAI_DEFAULT_TEMPERATURE,
 } from "@roo-code/types"
+import { Anthropic } from "@anthropic-ai/sdk"
+import OpenAI from "openai"
 
 import type { ApiHandlerOptions } from "../../shared/api"
+import { ApiStream } from "../transform/stream"
+import { convertToOpenAiMessages } from "../transform/openai-format"
+import type { ApiHandlerCreateMessageMetadata } from "../index"
 
 import { BaseOpenAiCompatibleProvider } from "./base-openai-compatible-provider"
 
 export class ZAiHandler extends BaseOpenAiCompatibleProvider<InternationalZAiModelId | MainlandZAiModelId> {
+	private readonly isGLM45: boolean
+
 	constructor(options: ApiHandlerOptions) {
 		const isChina = options.zaiApiLine === "china"
 		const models = isChina ? mainlandZAiModels : internationalZAiModels
@@ -27,5 +34,187 @@ export class ZAiHandler extends BaseOpenAiCompatibleProvider<InternationalZAiMod
 			providerModels: models,
 			defaultTemperature: ZAI_DEFAULT_TEMPERATURE,
 		})
+
+		// Check if the model is GLM-4.5 or GLM-4.5-Air
+		const modelId = options.apiModelId || defaultModelId
+		this.isGLM45 = modelId.includes("glm-4.5")
+	}
+
+	/**
+	 * Override createMessage to add GLM-specific handling
+	 */
+	override async *createMessage(
+		systemPrompt: string,
+		messages: Anthropic.Messages.MessageParam[],
+		metadata?: ApiHandlerCreateMessageMetadata,
+	): ApiStream {
+		// For GLM-4.5 models, enhance the system prompt with clearer instructions
+		let enhancedSystemPrompt = systemPrompt
+
+		if (this.isGLM45) {
+			// Add GLM-specific instructions to prevent hallucination and improve tool understanding
+			const glmInstructions = `
+
+# CRITICAL INSTRUCTIONS FOR GLM MODEL
+
+## File and Code Awareness
+- NEVER assume or hallucinate files that don't exist. Always verify file existence using the provided tools.
+- When exploring code, ALWAYS use the available tools (read_file, list_files, search_files) to examine actual files.
+- If you're unsure about a file's existence or location, use list_files to explore the directory structure first.
+- Base all code analysis and modifications on actual file contents retrieved through tools, not assumptions.
+
+## Tool Usage Protocol
+- Tools are invoked using XML-style tags as shown in the examples.
+- Each tool invocation must be properly formatted with the exact tool name as the XML tag.
+- Wait for tool execution results before proceeding to the next step.
+- Never simulate or imagine tool outputs - always use actual results.
+
+## Content Management
+- When working with large files or responses, focus on the specific sections relevant to the task.
+- Use partial reads when available to efficiently handle large files.
+- Condense and summarize appropriately while maintaining accuracy.
+- Keep responses concise and within token limits by focusing on essential information.
+
+## Code Indexing Integration
+- The code index provides semantic understanding of the codebase.
+- Use codebase_search for initial exploration when available.
+- Combine index results with actual file reading for complete understanding.
+- Trust the index for finding relevant code patterns and implementations.`
+
+			enhancedSystemPrompt = systemPrompt + glmInstructions
+		}
+
+		const {
+			id: model,
+			info: { maxTokens: max_tokens },
+		} = this.getModel()
+
+		const temperature = this.options.modelTemperature ?? this.defaultTemperature
+
+		// For GLM models, we may need to adjust the max_tokens to leave room for proper responses
+		// GLM models sometimes struggle with very high token limits
+		const adjustedMaxTokens = this.isGLM45 && max_tokens ? Math.min(max_tokens, 32768) : max_tokens
+
+		const params: OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming = {
+			model,
+			max_tokens: adjustedMaxTokens || 32768,
+			temperature,
+			messages: [
+				{ role: "system", content: enhancedSystemPrompt },
+				...this.preprocessMessages(convertToOpenAiMessages(messages)),
+			],
+			stream: true,
+			stream_options: { include_usage: true },
+		}
+
+		// Add additional parameters for GLM models to improve response quality
+		if (this.isGLM45) {
+			// GLM models benefit from explicit top_p and frequency_penalty settings
+			Object.assign(params, {
+				top_p: 0.95,
+				frequency_penalty: 0.1,
+				presence_penalty: 0.1,
+			})
+		}
+
+		const stream = await this.client.chat.completions.create(params)
+
+		for await (const chunk of stream) {
+			const delta = chunk.choices[0]?.delta
+
+			if (delta?.content) {
+				yield {
+					type: "text",
+					text: delta.content,
+				}
+			}
+
+			if (chunk.usage) {
+				yield {
+					type: "usage",
+					inputTokens: chunk.usage.prompt_tokens || 0,
+					outputTokens: chunk.usage.completion_tokens || 0,
+				}
+			}
+		}
+	}
+
+	/**
+	 * Preprocess messages for GLM models to ensure better understanding
+	 */
+	private preprocessMessages(
+		messages: OpenAI.Chat.ChatCompletionMessageParam[],
+	): OpenAI.Chat.ChatCompletionMessageParam[] {
+		if (!this.isGLM45) {
+			return messages
+		}
+
+		// For GLM models, ensure tool-related messages are clearly formatted
+		return messages.map((msg) => {
+			if (msg.role === "assistant" && typeof msg.content === "string") {
+				// Ensure XML tags in assistant messages are properly formatted
+				// GLM models sometimes struggle with complex XML structures
+				const content = msg.content
+					.replace(/(<\/?[^>]+>)/g, "\n$1\n") // Add newlines around XML tags
+					.replace(/\n\n+/g, "\n") // Remove excessive newlines
+					.trim()
+
+				return { ...msg, content }
+			}
+
+			if (msg.role === "user" && Array.isArray(msg.content)) {
+				// For user messages with multiple content blocks, ensure text is clear
+				const processedContent = msg.content.map((block: any) => {
+					if (block.type === "text") {
+						// Add clear markers for tool results to help GLM understand context
+						if (block.text.includes("[ERROR]") || block.text.includes("Error:")) {
+							return {
+								...block,
+								text: `[TOOL EXECUTION RESULT - ERROR]\n${block.text}\n[END TOOL RESULT]`,
+							}
+						} else if (block.text.includes("Success:") || block.text.includes("successfully")) {
+							return {
+								...block,
+								text: `[TOOL EXECUTION RESULT - SUCCESS]\n${block.text}\n[END TOOL RESULT]`,
+							}
+						}
+					}
+					return block
+				})
+
+				return { ...msg, content: processedContent }
+			}
+
+			return msg
+		})
+	}
+
+	/**
+	 * Override completePrompt for better GLM handling
+	 */
+	override async completePrompt(prompt: string): Promise<string> {
+		const { id: modelId } = this.getModel()
+
+		try {
+			// For GLM models, add a clear instruction prefix
+			const enhancedPrompt = this.isGLM45
+				? `[INSTRUCTION] Please provide a direct and accurate response based on facts. Do not hallucinate or make assumptions.\n\n${prompt}`
+				: prompt
+
+			const response = await this.client.chat.completions.create({
+				model: modelId,
+				messages: [{ role: "user", content: enhancedPrompt }],
+				temperature: this.defaultTemperature,
+				max_tokens: 4096,
+			})
+
+			return response.choices[0]?.message.content || ""
+		} catch (error) {
+			if (error instanceof Error) {
+				throw new Error(`${this.providerName} completion error: ${error.message}`)
+			}
+
+			throw error
+		}
 	}
 }
