diff --git a/src/controllers/feedback.ts b/src/controllers/feedback.ts
index 8061d3a..a36f254 100644
--- a/src/controllers/feedback.ts
+++ b/src/controllers/feedback.ts
@@ -1,6 +1,5 @@
 import { zValidator } from '@hono/zod-validator';
 import { Hono } from 'hono';
-import OpenAI from 'openai';
 import { z } from 'zod';
 import config from '../config';
 import { LearningSystem } from '../review/learning/learning-system';
@@ -16,13 +15,13 @@ let learningSystem: LearningSystem | null = null;
 let reviewStore: FileReviewStore | null = null;
 
 // 初始化反馈系统（记忆系统可选）
-export function initializeFeedbackSystem(openaiClient: OpenAI, store: FileReviewStore): void {
+export function initializeFeedbackSystem(store: FileReviewStore): void {
   // 保存store实例以供handlers重用，避免多实例状态不同步
   reviewStore = store;
 
   // 记忆系统为可选功能
   if (config.review.qdrantUrl && config.review.enableMemory) {
-    memoryStore = new VectorMemoryStore(config.review.qdrantUrl, openaiClient);
+    memoryStore = new VectorMemoryStore(config.review.qdrantUrl);
     learningSystem = new LearningSystem(memoryStore, reviewStore);
 
     memoryStore.initialize().catch((err) => {
diff --git a/src/review/agents/correctness-agent.ts b/src/review/agents/correctness-agent.ts
index ab4dc75..8d26c52 100644
--- a/src/review/agents/correctness-agent.ts
+++ b/src/review/agents/correctness-agent.ts
@@ -1,18 +1,12 @@
-import OpenAI from 'openai';
+import type { LLMGateway } from '../../llm/gateway';
 import type { LearningSystem } from '../learning/learning-system';
 import { ToolRegistry } from '../tools/registry';
 import { SpecialistAgent } from './specialist-agent';
 
 export class CorrectnessAgent extends SpecialistAgent {
-  constructor(
-    openai: OpenAI,
-    model: string,
-    toolRegistry?: ToolRegistry,
-    learningSystem?: LearningSystem
-  ) {
+  constructor(gateway: LLMGateway, toolRegistry?: ToolRegistry, learningSystem?: LearningSystem) {
     super(
-      openai,
-      model,
+      gateway,
       'correctness',
       'Correctness Agent',
       '业务逻辑正确性、边界条件、空值处理和明显bug',
diff --git a/src/review/agents/critic-agent.ts b/src/review/agents/critic-agent.ts
index 3be61ee..db2220d 100644
--- a/src/review/agents/critic-agent.ts
+++ b/src/review/agents/critic-agent.ts
@@ -1,7 +1,8 @@
-import OpenAI from 'openai';
-import { logger } from '../../utils/logger';
-import { withGlobalPrompt } from '../../utils/global-prompt';
 import config from '../../config';
+import type { LLMGateway } from '../../llm/gateway';
+import type { LLMMessage } from '../../llm/types';
+import { withGlobalPrompt } from '../../utils/global-prompt';
+import { logger } from '../../utils/logger';
 import { Finding, ReviewContext } from '../types';
 
 export interface CritiqueResult {
@@ -19,10 +20,7 @@ export interface CritiqueIssue {
 }
 
 export class CriticAgent {
-  constructor(
-    private openai: OpenAI,
-    private model: string
-  ) {}
+  constructor(private gateway: LLMGateway) {}
 
   async critique(
     findings: Omit<Finding, 'id' | 'runId' | 'published'>[],
@@ -71,20 +69,24 @@ ${context.diff.slice(0, 3000)}
 }`;
 
     try {
-      const response = await this.openai.chat.completions.create({
-        model: this.model,
+      const messages: LLMMessage[] = [
+        {
+          role: 'system',
+          content: withGlobalPrompt(
+            '你是严格的代码审查质量评估专家，以高标准评估findings的质量。',
+            config.review.globalPrompt
+          ),
+        },
+        { role: 'user', content: prompt },
+      ];
+
+      const response = await this.gateway.chatForRole('specialist', {
+        messages,
         temperature: 0.1, // 略高于0以允许批判性思考
-        response_format: { type: 'json_object' },
-        messages: [
-          {
-            role: 'system',
-            content: withGlobalPrompt('你是严格的代码审查质量评估专家，以高标准评估findings的质量。', config.openai.globalPrompt),
-          },
-          { role: 'user', content: prompt },
-        ],
+        responseFormat: 'json',
       });
 
-      const content = response.choices[0]?.message.content;
+      const content = response.content;
       if (!content) {
         throw new Error('Critic Agent返回空结果');
       }
@@ -160,20 +162,21 @@ ${context.diff.slice(0, 2000)}
 }`;
 
     try {
-      const response = await this.openai.chat.completions.create({
-        model: this.model,
+      const messages: LLMMessage[] = [
+        {
+          role: 'system',
+          content: withGlobalPrompt('你是代码审查质量评估专家。', config.review.globalPrompt),
+        },
+        { role: 'user', content: prompt },
+      ];
+
+      const response = await this.gateway.chatForRole('specialist', {
+        messages,
         temperature: 0,
-        response_format: { type: 'json_object' },
-        messages: [
-          {
-            role: 'system',
-            content: withGlobalPrompt('你是代码审查质量评估专家。', config.openai.globalPrompt),
-          },
-          { role: 'user', content: prompt },
-        ],
+        responseFormat: 'json',
       });
 
-      const content = response.choices[0]?.message.content;
+      const content = response.content;
       if (!content) {
         throw new Error('评估失败');
       }
diff --git a/src/review/agents/debate-orchestrator.ts b/src/review/agents/debate-orchestrator.ts
index d861d08..3bb75f6 100644
--- a/src/review/agents/debate-orchestrator.ts
+++ b/src/review/agents/debate-orchestrator.ts
@@ -1,7 +1,8 @@
-import OpenAI from 'openai';
-import { logger } from '../../utils/logger';
-import { withGlobalPrompt } from '../../utils/global-prompt';
 import config from '../../config';
+import type { LLMGateway } from '../../llm/gateway';
+import type { LLMMessage } from '../../llm/types';
+import { withGlobalPrompt } from '../../utils/global-prompt';
+import { logger } from '../../utils/logger';
 import { Finding, FindingSeverity } from '../types';
 import { SpecialistAgent } from './specialist-agent';
 
@@ -14,12 +15,10 @@ interface AgentOpinion {
 }
 
 export class DebateOrchestrator {
-  private openai: OpenAI;
-  private model: string;
+  private gateway: LLMGateway;
 
-  constructor(openai: OpenAI, model: string) {
-    this.openai = openai;
-    this.model = model;
+  constructor(gateway: LLMGateway) {
+    this.gateway = gateway;
   }
 
   async conductDebate(
@@ -103,20 +102,24 @@ export class DebateOrchestrator {
 }`;
 
     try {
-      const response = await this.openai.chat.completions.create({
-        model: this.model,
+      const messages: LLMMessage[] = [
+        {
+          role: 'system',
+          content: withGlobalPrompt(
+            `你是${agentName}，从你的专业角度独立评估代码问题。`,
+            config.review.globalPrompt
+          ),
+        },
+        { role: 'user', content: prompt },
+      ];
+
+      const response = await this.gateway.chatForRole('specialist', {
+        messages,
         temperature: 0.2,
-        response_format: { type: 'json_object' },
-        messages: [
-          {
-            role: 'system',
-            content: withGlobalPrompt(`你是${agentName}，从你的专业角度独立评估代码问题。`, config.openai.globalPrompt),
-          },
-          { role: 'user', content: prompt },
-        ],
+        responseFormat: 'json',
       });
 
-      const content = response.choices[0]?.message.content;
+      const content = response.content;
       if (!content) {
         throw new Error('Agent opinion返回空');
       }
@@ -180,20 +183,24 @@ ${otherOpinions
 }`;
 
     try {
-      const response = await this.openai.chat.completions.create({
-        model: this.model,
+      const messages: LLMMessage[] = [
+        {
+          role: 'system',
+          content: withGlobalPrompt(
+            `你是${agentName}，根据同行意见重新评估，但也要坚持你的专业判断。`,
+            config.review.globalPrompt
+          ),
+        },
+        { role: 'user', content: prompt },
+      ];
+
+      const response = await this.gateway.chatForRole('specialist', {
+        messages,
         temperature: 0.3, // 允许一定灵活性
-        response_format: { type: 'json_object' },
-        messages: [
-          {
-            role: 'system',
-            content: withGlobalPrompt(`你是${agentName}，根据同行意见重新评估，但也要坚持你的专业判断。`, config.openai.globalPrompt),
-          },
-          { role: 'user', content: prompt },
-        ],
+        responseFormat: 'json',
       });
 
-      const content = response.choices[0]?.message.content;
+      const content = response.content;
       if (!content) {
         throw new Error('Revised opinion返回空');
       }
diff --git a/src/review/agents/maintainability-agent.ts b/src/review/agents/maintainability-agent.ts
index d5990f2..f06a449 100644
--- a/src/review/agents/maintainability-agent.ts
+++ b/src/review/agents/maintainability-agent.ts
@@ -1,18 +1,12 @@
-import OpenAI from 'openai';
+import type { LLMGateway } from '../../llm/gateway';
 import type { LearningSystem } from '../learning/learning-system';
 import { ToolRegistry } from '../tools/registry';
 import { SpecialistAgent } from './specialist-agent';
 
 export class MaintainabilityAgent extends SpecialistAgent {
-  constructor(
-    openai: OpenAI,
-    model: string,
-    toolRegistry?: ToolRegistry,
-    learningSystem?: LearningSystem
-  ) {
+  constructor(gateway: LLMGateway, toolRegistry?: ToolRegistry, learningSystem?: LearningSystem) {
     super(
-      openai,
-      model,
+      gateway,
       'maintainability',
       'Maintainability Agent',
       '可维护性、复杂度、接口破坏风险和可测试性不足',
diff --git a/src/review/agents/reflexion-agent.ts b/src/review/agents/reflexion-agent.ts
index 72b1cd4..9445c7f 100644
--- a/src/review/agents/reflexion-agent.ts
+++ b/src/review/agents/reflexion-agent.ts
@@ -1,8 +1,9 @@
 import { createHash } from 'node:crypto';
-import OpenAI from 'openai';
-import { logger } from '../../utils/logger';
-import { withGlobalPrompt } from '../../utils/global-prompt';
 import config from '../../config';
+import type { LLMGateway } from '../../llm/gateway';
+import type { LLMMessage } from '../../llm/types';
+import { withGlobalPrompt } from '../../utils/global-prompt';
+import { logger } from '../../utils/logger';
 import { LearningSystem } from '../learning/learning-system';
 import { findingResponseSchema } from '../schema/finding-schema';
 import { ToolRegistry } from '../tools/registry';
@@ -21,16 +22,15 @@ export class ReflexionAgent extends SpecialistAgent {
   private criticAgent: CriticAgent;
 
   constructor(
-    openai: OpenAI,
-    model: string,
+    gateway: LLMGateway,
     category: FindingCategory,
     agentName: string,
     focusPrompt: string,
     toolRegistry?: ToolRegistry,
     learningSystem?: LearningSystem
   ) {
-    super(openai, model, category, agentName, focusPrompt, toolRegistry, learningSystem);
-    this.criticAgent = new CriticAgent(openai, model);
+    super(gateway, category, agentName, focusPrompt, toolRegistry, learningSystem);
+    this.criticAgent = new CriticAgent(gateway);
   }
 
   async reviewWithReflection(
@@ -142,20 +142,24 @@ ${context.diff.slice(0, 3000)}
 }`;
 
     try {
-      const response = await this.openai.chat.completions.create({
-        model: this.model,
+      const messages: LLMMessage[] = [
+        {
+          role: 'system',
+          content: withGlobalPrompt(
+            `你是${this.agentName}，根据批评反馈改进审查结果。`,
+            config.review.globalPrompt
+          ),
+        },
+        { role: 'user', content: prompt },
+      ];
+
+      const response = await this.gateway.chatForRole('specialist', {
+        messages,
         temperature: 0.1,
-        response_format: { type: 'json_object' },
-        messages: [
-          {
-            role: 'system',
-            content: withGlobalPrompt(`你是${this.agentName}，根据批评反馈改进审查结果。`, config.openai.globalPrompt),
-          },
-          { role: 'user', content: prompt },
-        ],
+        responseFormat: 'json',
       });
 
-      const content = response.choices[0]?.message.content;
+      const content = response.content;
       if (!content) {
         logger.warn(`${this.agentName} Refine返回空结果，使用原findings`);
         return draft;
diff --git a/src/review/agents/reliability-agent.ts b/src/review/agents/reliability-agent.ts
index cb3f715..827ee64 100644
--- a/src/review/agents/reliability-agent.ts
+++ b/src/review/agents/reliability-agent.ts
@@ -1,18 +1,12 @@
-import OpenAI from 'openai';
+import type { LLMGateway } from '../../llm/gateway';
 import type { LearningSystem } from '../learning/learning-system';
 import { ToolRegistry } from '../tools/registry';
 import { SpecialistAgent } from './specialist-agent';
 
 export class ReliabilityAgent extends SpecialistAgent {
-  constructor(
-    openai: OpenAI,
-    model: string,
-    toolRegistry?: ToolRegistry,
-    learningSystem?: LearningSystem
-  ) {
+  constructor(gateway: LLMGateway, toolRegistry?: ToolRegistry, learningSystem?: LearningSystem) {
     super(
-      openai,
-      model,
+      gateway,
       'reliability',
       'Reliability Agent',
       '错误处理、重试策略、幂等性、并发一致性和资源释放',
diff --git a/src/review/agents/security-agent.ts b/src/review/agents/security-agent.ts
index fe4ceca..dfd57dc 100644
--- a/src/review/agents/security-agent.ts
+++ b/src/review/agents/security-agent.ts
@@ -1,18 +1,12 @@
-import OpenAI from 'openai';
+import type { LLMGateway } from '../../llm/gateway';
 import type { LearningSystem } from '../learning/learning-system';
 import { ToolRegistry } from '../tools/registry';
 import { SpecialistAgent } from './specialist-agent';
 
 export class SecurityAgent extends SpecialistAgent {
-  constructor(
-    openai: OpenAI,
-    model: string,
-    toolRegistry?: ToolRegistry,
-    learningSystem?: LearningSystem
-  ) {
+  constructor(gateway: LLMGateway, toolRegistry?: ToolRegistry, learningSystem?: LearningSystem) {
     super(
-      openai,
-      model,
+      gateway,
       'security',
       'Security Agent',
       '注入漏洞、权限绕过、敏感信息泄露、反序列化和输入校验缺失',
diff --git a/src/review/agents/specialist-agent.ts b/src/review/agents/specialist-agent.ts
index f979a08..6df7500 100644
--- a/src/review/agents/specialist-agent.ts
+++ b/src/review/agents/specialist-agent.ts
@@ -1,8 +1,9 @@
 import { createHash } from 'node:crypto';
-import OpenAI from 'openai';
-import { logger } from '../../utils/logger';
-import { withGlobalPrompt } from '../../utils/global-prompt';
 import config from '../../config';
+import type { LLMGateway } from '../../llm/gateway';
+import type { LLMMessage, LLMToolCall } from '../../llm/types';
+import { withGlobalPrompt } from '../../utils/global-prompt';
+import { logger } from '../../utils/logger';
 import type { LearningSystem } from '../learning/learning-system';
 import { findingResponseSchema } from '../schema/finding-schema';
 import { ToolRegistry } from '../tools/registry';
@@ -90,8 +91,7 @@ function toCompactContext(context: ReviewContext): string {
 
 export class SpecialistAgent {
   constructor(
-    protected readonly openai: OpenAI,
-    protected readonly model: string,
+    protected readonly gateway: LLMGateway,
     protected readonly category: FindingCategory,
     protected readonly agentName: string,
     protected readonly focusPrompt: string,
@@ -123,21 +123,24 @@ export class SpecialistAgent {
 ${toCompactContext(context)}`;
 
     try {
-      const response = await this.openai.chat.completions.create({
-        model: this.model,
+      const messages: LLMMessage[] = [
+        {
+          role: 'system',
+          content: withGlobalPrompt(
+            '你是严格的代码审查专家。返回结构化JSON，不输出额外文字。confidence取值范围0到1。line必须是正整数且引用新增行。',
+            config.review.globalPrompt
+          ),
+        },
+        { role: 'user', content: prompt },
+      ];
+
+      const response = await this.gateway.chatForRole('specialist', {
+        messages,
         temperature: 0,
-        response_format: { type: 'json_object' },
-        messages: [
-          {
-            role: 'system',
-            content:
-              withGlobalPrompt('你是严格的代码审查专家。返回结构化JSON，不输出额外文字。confidence取值范围0到1。line必须是正整数且引用新增行。', config.openai.globalPrompt),
-          },
-          { role: 'user', content: prompt },
-        ],
+        responseFormat: 'json',
       });
 
-      const content = response.choices[0]?.message.content;
+      const content = response.content;
       if (!content) {
         return { agentName: this.agentName, findings: [] };
       }
@@ -166,10 +169,11 @@ ${toCompactContext(context)}`;
   private async reviewWithReAct(run: ReviewRun, context: ReviewContext): Promise<AgentResult> {
     const maxIterations = 5;
     const findingsMap = new Map<string, Omit<Finding, 'id' | 'runId' | 'published'>>();
-    const messages: OpenAI.Chat.ChatCompletionMessageParam[] = [
+    const messages: LLMMessage[] = [
       {
         role: 'system',
-        content: withGlobalPrompt(`你是${this.agentName}，专注于${this.focusPrompt}。
+        content: withGlobalPrompt(
+          `你是${this.agentName}，专注于${this.focusPrompt}。
 
 你可以使用以下工具进行深入调查：
 ${this.toolRegistry!.getAll()
@@ -198,7 +202,9 @@ ${this.toolRegistry!.getAll()
   ],
   "need_more_investigation": false
 }
-每个 finding 对象的所有字段都是必填的。无问题时返回空数组 {"findings": [], "need_more_investigation": false}。`, config.openai.globalPrompt),
+每个 finding 对象的所有字段都是必填的。无问题时返回空数组 {"findings": [], "need_more_investigation": false}。`,
+          config.review.globalPrompt
+        ),
       },
     ];
 
@@ -211,7 +217,22 @@ ${this.toolRegistry!.getAll()
           run.repo
         );
         if (fewShotExamples.length > 0) {
-          messages.push(...fewShotExamples);
+          const llmFewShotExamples = fewShotExamples
+            .map((msg) => {
+              if (
+                (msg.role === 'system' || msg.role === 'user' || msg.role === 'assistant') &&
+                typeof msg.content === 'string'
+              ) {
+                return { role: msg.role, content: msg.content } as const;
+              }
+              return null;
+            })
+            .filter(
+              (msg): msg is { role: 'system' | 'user' | 'assistant'; content: string } =>
+                msg !== null
+            );
+
+          messages.push(...llmFewShotExamples);
           logger.debug(`${this.agentName} 添加了 ${fewShotExamples.length} 条Few-shot示例`, {
             runId: run.id,
           });
@@ -239,24 +260,24 @@ ${this.toolRegistry!.getAll()
         // 仅在最后一轮迭代强制 JSON 输出（无工具调用时解析结果）
         // 避免 response_format: json_object 与 tools 参数冲突导致工具不被调用
         const isLastIteration = iteration === maxIterations - 1;
-        const response = await this.openai.chat.completions.create({
-          model: this.model,
-          temperature: 0,
-          ...(isLastIteration ? { response_format: { type: 'json_object' as const } } : {}),
+        const response = await this.gateway.chatForRole('specialist', {
           messages,
-          tools: this.toolRegistry!.toOpenAIFunctions(),
-          tool_choice: isLastIteration ? 'none' : 'auto',
+          temperature: 0,
+          tools: this.toolRegistry!.toToolDefinitions(),
+          providerOptions: { tool_choice: isLastIteration ? 'none' : 'auto' },
+          responseFormat: isLastIteration ? 'json' : undefined,
         });
 
-        const choice = response.choices[0];
-        if (!choice) break;
-
         // 处理工具调用
-        if (choice.message.tool_calls && choice.message.tool_calls.length > 0) {
-          messages.push(choice.message as OpenAI.Chat.ChatCompletionMessageParam);
+        if (response.toolCalls.length > 0) {
+          messages.push({
+            role: 'assistant',
+            content: response.content || '',
+            toolCalls: response.toolCalls,
+          });
 
           // 执行所有工具调用
-          const toolResults = await this.executeTools(choice.message.tool_calls, {
+          const toolResults = await this.executeTools(response.toolCalls, {
             workspacePath: context.workspacePath,
             mirrorPath: context.mirrorPath,
             runId: run.id,
@@ -266,7 +287,7 @@ ${this.toolRegistry!.getAll()
           for (const toolResult of toolResults) {
             messages.push({
               role: 'tool',
-              tool_call_id: toolResult.toolCallId,
+              toolCallId: toolResult.toolCallId,
               content: JSON.stringify(toolResult.result || { error: toolResult.error }),
             });
           }
@@ -275,9 +296,9 @@ ${this.toolRegistry!.getAll()
         }
 
         // 解析findings（模型选择返回内容而非调用工具）
-        if (choice.message.content) {
+        if (response.content) {
           try {
-            const parsed = JSON.parse(choice.message.content);
+            const parsed = JSON.parse(response.content);
 
             if (parsed.findings && parsed.findings.length > 0) {
               // 使用schema验证findings，防止畸形数据流入发布系统
@@ -301,7 +322,10 @@ ${this.toolRegistry!.getAll()
             }
 
             // 模型要求继续调查但没有调用工具：注入 user 消息打破潜在的自我重复
-            messages.push(choice.message as OpenAI.Chat.ChatCompletionMessageParam);
+            messages.push({
+              role: 'assistant',
+              content: response.content,
+            });
             messages.push({
               role: 'user',
               content:
@@ -314,7 +338,10 @@ ${this.toolRegistry!.getAll()
               runId: run.id,
               error: parseError instanceof Error ? parseError.message : String(parseError),
             });
-            messages.push(choice.message as OpenAI.Chat.ChatCompletionMessageParam);
+            messages.push({
+              role: 'assistant',
+              content: response.content,
+            });
             messages.push({
               role: 'user',
               content:
@@ -338,28 +365,28 @@ ${this.toolRegistry!.getAll()
   }
 
   private async executeTools(
-    toolCalls: OpenAI.Chat.ChatCompletionMessageToolCall[],
+    toolCalls: LLMToolCall[],
     context: ToolExecutionContext
   ): Promise<ToolResult[]> {
     const results: ToolResult[] = [];
 
     for (const toolCall of toolCalls) {
-      const tool = this.toolRegistry!.get(toolCall.function.name);
+      const tool = this.toolRegistry!.get(toolCall.name);
 
       if (!tool) {
         results.push({
           toolCallId: toolCall.id,
           success: false,
-          error: `工具 ${toolCall.function.name} 未找到`,
+          error: `工具 ${toolCall.name} 未找到`,
         });
         continue;
       }
 
       try {
-        const params = JSON.parse(toolCall.function.arguments);
+        const params = JSON.parse(toolCall.arguments);
         const result = await tool.execute(params, context);
 
-        logger.info(`工具调用成功: ${toolCall.function.name}`, {
+        logger.info(`工具调用成功: ${toolCall.name}`, {
           runId: context.runId,
           params,
         });
@@ -370,7 +397,7 @@ ${this.toolRegistry!.getAll()
           result,
         });
       } catch (error) {
-        logger.error(`工具调用失败: ${toolCall.function.name}`, {
+        logger.error(`工具调用失败: ${toolCall.name}`, {
           runId: context.runId,
           error: error instanceof Error ? error.message : String(error),
         });
diff --git a/src/review/learning/learning-system.ts b/src/review/learning/learning-system.ts
index 5be7120..164f9da 100644
--- a/src/review/learning/learning-system.ts
+++ b/src/review/learning/learning-system.ts
@@ -1,4 +1,4 @@
-import OpenAI from 'openai';
+import type { LLMMessage } from '../../llm/types';
 import config from '../../config';
 import { logger } from '../../utils/logger';
 import { VectorMemoryStore } from '../memory/vector-store';
@@ -59,7 +59,7 @@ export class LearningSystem {
     category: FindingCategory,
     owner?: string,
     repo?: string
-  ): Promise<OpenAI.Chat.ChatCompletionMessageParam[]> {
+  ): Promise<LLMMessage[]> {
     const targetCount = config.review.fewShotExamplesCount;
 
     // 提前检查：如果few-shot被禁用（targetCount=0），直接返回，避免无意义的向量查询
@@ -105,7 +105,7 @@ export class LearningSystem {
       );
     }
 
-    const examples: OpenAI.Chat.ChatCompletionMessageParam[] = [];
+    const examples: LLMMessage[] = [];
 
     const negativeCount = Math.floor(targetCount * 0.4);
 
diff --git a/src/review/memory/vector-store.ts b/src/review/memory/vector-store.ts
index d44ef10..90bcc0f 100644
--- a/src/review/memory/vector-store.ts
+++ b/src/review/memory/vector-store.ts
@@ -1,18 +1,16 @@
 import { QdrantClient } from '@qdrant/js-client-rest';
-import OpenAI from 'openai';
+import { llmGateway } from '../../llm/gateway';
 import { logger } from '../../utils/logger';
 import { Finding } from '../types';
 import { MemoryEntry, MemorySearchResult } from './types';
 
 export class VectorMemoryStore {
   private client: QdrantClient;
-  private openai: OpenAI;
   private collectionName = 'code_review_memory';
   private initialized = false;
 
-  constructor(qdrantUrl: string, openaiClient: OpenAI) {
+  constructor(qdrantUrl: string) {
     this.client = new QdrantClient({ url: qdrantUrl });
-    this.openai = openaiClient;
   }
 
   async initialize(): Promise<void> {
@@ -47,7 +45,7 @@ export class VectorMemoryStore {
   async storeMemory(entry: MemoryEntry): Promise<void> {
     await this.initialize();
 
-    const embedding = await this.getEmbedding(entry.content);
+    const [embedding] = await this.getEmbedding([entry.content]);
 
     await this.client.upsert(this.collectionName, {
       points: [
@@ -73,7 +71,7 @@ export class VectorMemoryStore {
   async searchSimilar(query: string, limit = 5, filter?: any): Promise<MemorySearchResult[]> {
     await this.initialize();
 
-    const queryEmbedding = await this.getEmbedding(query);
+    const [queryEmbedding] = await this.getEmbedding([query]);
 
     const results = await this.client.search(this.collectionName, {
       vector: queryEmbedding,
@@ -101,14 +99,9 @@ export class VectorMemoryStore {
     }));
   }
 
-  private async getEmbedding(text: string): Promise<number[]> {
+  private async getEmbedding(texts: string[]): Promise<number[][]> {
     try {
-      const response = await this.openai.embeddings.create({
-        model: 'text-embedding-3-small',
-        input: text.slice(0, 8000), // 限制长度防止超出token限制
-      });
-
-      return response.data[0].embedding;
+      return llmGateway.embedForRole(texts.map((text) => text.slice(0, 8000))); // 限制长度防止超出token限制
     } catch (error) {
       logger.error('生成embedding失败', {
         error: error instanceof Error ? error.message : String(error),
diff --git a/src/review/orchestrator.ts b/src/review/orchestrator.ts
index fddfd53..913e7df 100644
--- a/src/review/orchestrator.ts
+++ b/src/review/orchestrator.ts
@@ -1,6 +1,6 @@
 import { randomUUID } from 'node:crypto';
-import OpenAI from 'openai';
 import config from '../config';
+import { LLMGateway, llmGateway } from '../llm/gateway';
 import { giteaService } from '../services/gitea';
 import { logger } from '../utils/logger';
 import { DebateOrchestrator } from './agents/debate-orchestrator';
@@ -42,7 +42,7 @@ function summarizeGatedCount(gatedCount: number): string {
 }
 
 export class ReviewOrchestrator {
-  private readonly openai: OpenAI;
+  private readonly gateway: LLMGateway;
   private readonly toolRegistry: ToolRegistry;
   private readonly correctnessAgent: ReflexionAgent;
   private readonly securityAgent: ReflexionAgent;
@@ -58,10 +58,7 @@ export class ReviewOrchestrator {
     private readonly localRepoManager: LocalRepoManager,
     private readonly diffExtractor: DiffExtractor
   ) {
-    this.openai = new OpenAI({
-      baseURL: config.openai.baseUrl,
-      apiKey: config.openai.apiKey,
-    });
+    this.gateway = llmGateway;
 
     // 初始化工具注册表
     this.toolRegistry = new ToolRegistry();
@@ -75,7 +72,7 @@ export class ReviewOrchestrator {
 
     // 初始化记忆和学习系统（可选）
     if (config.review.qdrantUrl && config.review.enableMemory) {
-      this.memoryStore = new VectorMemoryStore(config.review.qdrantUrl, this.openai);
+      this.memoryStore = new VectorMemoryStore(config.review.qdrantUrl);
       this.learningSystem = new LearningSystem(this.memoryStore, this.store);
 
       this.memoryStore.initialize().catch((err) => {
@@ -87,8 +84,7 @@ export class ReviewOrchestrator {
 
     // 创建Reflexion-wrapped agents并传递工具注册表和学习系统
     this.correctnessAgent = new ReflexionAgent(
-      this.openai,
-      config.review.modelSpecialist,
+      this.gateway,
       'correctness',
       'Correctness Agent',
       '业务逻辑正确性、边界条件、空值处理和明显bug',
@@ -97,8 +93,7 @@ export class ReviewOrchestrator {
     );
 
     this.securityAgent = new ReflexionAgent(
-      this.openai,
-      config.review.modelSpecialist,
+      this.gateway,
       'security',
       'Security Agent',
       '注入漏洞、权限绕过、敏感信息泄露、反序列化和输入校验缺失',
@@ -107,8 +102,7 @@ export class ReviewOrchestrator {
     );
 
     this.reliabilityAgent = new ReflexionAgent(
-      this.openai,
-      config.review.modelSpecialist,
+      this.gateway,
       'reliability',
       'Reliability Agent',
       '错误处理、重试策略、幂等性、并发一致性和资源释放',
@@ -117,8 +111,7 @@ export class ReviewOrchestrator {
     );
 
     this.maintainabilityAgent = new ReflexionAgent(
-      this.openai,
-      config.review.modelSpecialist,
+      this.gateway,
       'maintainability',
       'Maintainability Agent',
       '可维护性、复杂度、接口破坏风险和可测试性不足',
@@ -127,7 +120,7 @@ export class ReviewOrchestrator {
     );
 
     this.judgeAgent = new JudgeAgent();
-    this.debateOrchestrator = new DebateOrchestrator(this.openai, config.review.modelSpecialist);
+    this.debateOrchestrator = new DebateOrchestrator(this.gateway);
   }
 
   async execute(run: ReviewRun): Promise<void> {
diff --git a/src/review/tools/registry.ts b/src/review/tools/registry.ts
index 066baca..373cdaf 100644
--- a/src/review/tools/registry.ts
+++ b/src/review/tools/registry.ts
@@ -1,6 +1,7 @@
 import { z } from 'zod';
 import zodToJsonSchema from 'zod-to-json-schema';
 import type { JsonSchema7Type } from 'zod-to-json-schema';
+import type { LLMToolDefinition } from '../../llm/types';
 import { Tool } from './types';
 
 export class ToolRegistry {
@@ -30,6 +31,14 @@ export class ToolRegistry {
     }));
   }
 
+  toToolDefinitions(): LLMToolDefinition[] {
+    return this.getAll().map((tool) => ({
+      name: tool.name,
+      description: tool.description,
+      parameters: this.zodToJsonSchema(tool.parameters) as Record<string, unknown>,
+    }));
+  }
+
   private zodToJsonSchema(schema: z.ZodTypeAny): JsonSchema7Type {
     /**
      * 使用zod-to-json-schema库转换Zod schema为JSON Schema。
diff --git a/src/services/ai-review.ts b/src/services/ai-review.ts
index b9f2fb0..b5fd680 100644
--- a/src/services/ai-review.ts
+++ b/src/services/ai-review.ts
@@ -1,15 +1,10 @@
-import OpenAI from 'openai';
 import config from '../config';
-import { logger } from '../utils/logger';
+import { llmGateway } from '../llm/gateway';
+import { LLMMessage } from '../llm/types';
 import { withGlobalPrompt } from '../utils/global-prompt';
+import { logger } from '../utils/logger';
 import { PullRequestFile, giteaService } from './gitea';
 
-// 创建OpenAI客户端
-const openai = new OpenAI({
-  baseURL: config.openai.baseUrl,
-  apiKey: config.openai.apiKey,
-});
-
 // 代码审查结果接口
 export interface CodeReviewResult {
   summary: string;
@@ -198,23 +193,26 @@ export const aiReviewService = {
 
       请根据以上信息，特别是考虑每个文件的完整内容和上下文，提供代码审查评价。如果没有发现明显问题，请简短说明代码质量良好即可。`;
 
-      const summaryPrompt = config.openai.customSummaryPrompt || defaultSummaryPrompt;
+      const summaryPrompt = config.review.customSummaryPrompt || defaultSummaryPrompt;
 
       // 获取总体评价
-      const summaryResponse = await openai.chat.completions.create({
-        model: config.openai.model,
-        messages: [
-          {
-            role: 'system',
-            content:
-              withGlobalPrompt('你是一个专业的代码审查助手，擅长识别代码中的严重问题和bug。你会查看代码的完整上下文，而不是为了评论而评论。如无明显问题，应给予简短肯定。', config.openai.globalPrompt),
-          },
-          { role: 'user', content: summaryPrompt },
-        ],
+      const messages: LLMMessage[] = [
+        {
+          role: 'system',
+          content: withGlobalPrompt(
+            '你是一个专业的代码审查助手，擅长识别代码中的严重问题和bug。你会查看代码的完整上下文，而不是为了评论而评论。如无明显问题，应给予简短肯定。',
+            config.review.globalPrompt
+          ),
+        },
+        { role: 'user', content: summaryPrompt },
+      ];
+
+      const summaryResponse = await llmGateway.chatForRole('legacy', {
+        messages,
         temperature: 0.1,
       });
 
-      const summary = summaryResponse.choices[0]?.message.content || '无法生成代码审查摘要';
+      const summary = summaryResponse.content || '无法生成代码审查摘要';
       return summary;
     } catch (error: any) {
       logger.error('生成总体评价失败:', error);
@@ -270,24 +268,27 @@ export const aiReviewService = {
         ]
         只返回JSON数组，不要有其他文本。`;
 
-        const filePrompt = config.openai.customLineCommentPrompt || defaultFilePrompt;
+        const filePrompt = config.review.customLineCommentPrompt || defaultFilePrompt;
 
         // 获取行级评论
-        const lineResponse = await openai.chat.completions.create({
-          model: config.openai.model,
-          messages: [
-            {
-              role: 'system',
-              content:
-                withGlobalPrompt('你是一个谨慎的代码审查助手，只对有明显bug或严重问题的代码行提供评论。大多数情况下，如果代码没有严重问题，你应该返回空数组。请以JSON格式返回结果。', config.openai.globalPrompt),
-            },
-            { role: 'user', content: filePrompt },
-          ],
+        const messages: LLMMessage[] = [
+          {
+            role: 'system',
+            content: withGlobalPrompt(
+              '你是一个谨慎的代码审查助手，只对有明显bug或严重问题的代码行提供评论。大多数情况下，如果代码没有严重问题，你应该返回空数组。请以JSON格式返回结果。',
+              config.review.globalPrompt
+            ),
+          },
+          { role: 'user', content: filePrompt },
+        ];
+
+        const lineResponse = await llmGateway.chatForRole('legacy', {
+          messages,
           temperature: 0.1,
-          response_format: { type: 'json_object' },
+          responseFormat: 'json',
         });
 
-        const content = lineResponse.choices[0]?.message.content;
+        const content = lineResponse.content;
         if (!content) continue;
 
         try {