From 4c5a84e0284ae220cbfa7b0b462214e0dcf56997 Mon Sep 17 00:00:00 2001
From: musistudio <m@musiiot.top>
Date: Thu, 1 Jan 2026 22:11:21 +0800
Subject: [PATCH] fix docs

---
 docs/docs/cli/config/basic.md                 |  45 +-
 docs/docs/cli/intro.md                        |  36 +-
 docs/docs/cli/quick-start.md                  |  54 +-
 docs/docs/server/config/routing.md            | 109 +++
 docs/docs/server/config/transformers.md       | 692 +++++++++++++-
 docs/docs/server/intro.md                     |  96 +-
 docs/docusaurus.config.ts                     |  30 +-
 .../current.json                              |  18 +-
 .../current/server/config/routing.md          | 109 +++
 .../current/server/config/transformers.md     | 879 ++++++++++++++----
 .../current/server/intro.md                   |  96 +-
 .../docusaurus-theme-classic/navbar.json      |   4 +-
 docs/sidebars.ts                              |  96 +-
 package.json                                  |   2 +-
 packages/core/src/api/routes.ts               | 173 +++-
 packages/core/src/server.ts                   |   2 +
 packages/core/src/utils/router.ts             |  39 +-
 17 files changed, 2092 insertions(+), 388 deletions(-)

diff --git a/docs/docs/cli/config/basic.md b/docs/docs/cli/config/basic.md
index 187f86a..5f17bee 100644
--- a/docs/docs/cli/config/basic.md
+++ b/docs/docs/cli/config/basic.md
@@ -6,15 +6,29 @@ title: Basic Configuration
 
 CLI uses the same configuration file as Server: `~/.claude-code-router/config.json`
 
-## Configuration File Location
+## Configuration Methods
+
+You can configure Claude Code Router in three ways:
+
+### Option 1: Edit Configuration File Directly
+
+Edit `~/.claude-code-router/config.json` with your favorite editor:
 
 ```bash
-~/.claude-code-router/config.json
+nano ~/.claude-code-router/config.json
 ```
 
-## Quick Configuration
+### Option 2: Use Web UI
 
-Use interactive command to configure:
+Open the web interface and configure visually:
+
+```bash
+ccr ui
+```
+
+### Option 3: Interactive Configuration
+
+Use the interactive command-line configuration:
 
 ```bash
 ccr model
@@ -26,16 +40,23 @@ This will guide you through:
 3. Select model
 4. Set routing rules
 
-## Manual Configuration
+## Restart After Configuration Changes
 
-### Edit Configuration File
+After modifying the configuration file or making changes through the Web UI, you must restart the service:
 
 ```bash
-# Open configuration file
-nano ~/.claude-code-router/config.json
+ccr restart
 ```
 
-### Minimal Configuration Example
+Or restart directly through the Web UI.
+
+## Configuration File Location
+
+```bash
+~/.claude-code-router/config.json
+```
+
+## Minimal Configuration Example
 
 ```json5
 {
@@ -129,14 +150,16 @@ Configuration is automatically backed up on each update:
 ~/.claude-code-router/config.backup.{timestamp}.json
 ```
 
-## Reload Configuration
+## Apply Configuration Changes
 
-Restart service after modifying configuration:
+After modifying the configuration file or making changes through the Web UI, restart the service:
 
 ```bash
 ccr restart
 ```
 
+Or restart directly through the Web UI by clicking the "Save and Restart" button.
+
 ## View Current Configuration
 
 ```bash
diff --git a/docs/docs/cli/intro.md b/docs/docs/cli/intro.md
index 78e41a1..7b1ffb3 100644
--- a/docs/docs/cli/intro.md
+++ b/docs/docs/cli/intro.md
@@ -26,27 +26,49 @@ npm install -g @musistudio/claude-code-router
 
 ## Basic Usage
 
-### Start Service
+### Configuration
+
+Before using Claude Code Router, you need to configure your providers. You can either:
+
+1. **Edit configuration file directly**: Edit `~/.claude-code-router/config.json` manually
+2. **Use Web UI**: Run `ccr ui` to open the web interface and configure visually
+
+After making configuration changes, restart the service:
 
 ```bash
-ccr start
+ccr restart
 ```
 
-### View Status
+Or restart directly through the Web UI.
+
+### Start Claude Code
+
+Once configured, you can start Claude Code with:
 
 ```bash
-ccr status
+ccr code
 ```
 
-### Stop Service
+This will launch Claude Code and route your requests through the configured provider.
+
+### Service Management
 
 ```bash
-ccr stop
+ccr start    # Start the router service
+ccr status   # View service status
+ccr stop     # Stop the router service
+ccr restart  # Restart the router service
+```
+
+### Web UI
+
+```bash
+ccr ui       # Open Web management interface
 ```
 
 ## Configuration File
 
-`ccr` uses the same configuration file as Server: `~/.claude-code-router/config.json`
+`ccr` uses the configuration file at `~/.claude-code-router/config.json`
 
 Configure once, and both CLI and Server will use it.
 
diff --git a/docs/docs/cli/quick-start.md b/docs/docs/cli/quick-start.md
index 76ea0a0..d814d72 100644
--- a/docs/docs/cli/quick-start.md
+++ b/docs/docs/cli/quick-start.md
@@ -6,7 +6,41 @@ sidebar_position: 3
 
 Get up and running with Claude Code Router in 5 minutes.
 
-## 1. Start the Router
+## 1. Configure the Router
+
+Before using Claude Code Router, you need to configure your LLM providers. You can either:
+
+### Option A: Edit Configuration File Directly
+
+Edit `~/.claude-code-router/config.json`:
+
+```json
+{
+  "HOST": "0.0.0.0",
+  "PORT": 8080,
+  "Providers": [
+    {
+      "name": "openai",
+      "api_base_url": "https://api.openai.com/v1",
+      "api_key": "your-api-key-here",
+      "models": ["gpt-4", "gpt-3.5-turbo"]
+    }
+  ],
+  "Router": {
+    "default": "openai,gpt-4"
+  }
+}
+```
+
+### Option B: Use Web UI
+
+```bash
+ccr ui
+```
+
+This will open the web interface where you can configure providers visually.
+
+## 2. Start the Router
 
 ```bash
 ccr start
@@ -14,7 +48,7 @@ ccr start
 
 The router will start on `http://localhost:8080` by default.
 
-## 2. Use Claude Code
+## 3. Use Claude Code
 
 Now you can use Claude Code normally:
 
@@ -24,8 +58,18 @@ ccr code
 
 Your requests will be routed through Claude Code Router to your configured provider.
 
+## Restart After Configuration Changes
+
+If you modify the configuration file or make changes through the Web UI, restart the service:
+
+```bash
+ccr restart
+```
+
+Or restart directly through the Web UI.
+
 ## What's Next?
 
-- [Basic Configuration](/docs/config/basic) - Learn about configuration options
-- [Routing](/docs/config/routing) - Configure smart routing rules
-- [CLI Commands](/docs/cli/start) - Explore all CLI commands
+- [Basic Configuration](/docs/cli/config/basic) - Learn about configuration options
+- [Routing](/docs/cli/config/routing) - Configure smart routing rules
+- [CLI Commands](/docs/category/cli-commands) - Explore all CLI commands
diff --git a/docs/docs/server/config/routing.md b/docs/docs/server/config/routing.md
index f6b5d48..58ffc66 100644
--- a/docs/docs/server/config/routing.md
+++ b/docs/docs/server/config/routing.md
@@ -81,6 +81,115 @@ Route image-related tasks:
 }
 ```
 
+## Fallback
+
+When a request fails, you can configure a list of backup models. The system will try each model in sequence until one succeeds:
+
+### Basic Configuration
+
+```json
+{
+  "Router": {
+    "default": "deepseek,deepseek-chat",
+    "background": "ollama,qwen2.5-coder:latest",
+    "think": "deepseek,deepseek-reasoner",
+    "longContext": "openrouter,google/gemini-2.5-pro-preview",
+    "longContextThreshold": 60000,
+    "webSearch": "gemini,gemini-2.5-flash"
+  },
+  "fallback": {
+    "default": [
+      "aihubmix,Z/glm-4.5",
+      "openrouter,anthropic/claude-sonnet-4"
+    ],
+    "background": [
+      "ollama,qwen2.5-coder:latest"
+    ],
+    "think": [
+      "openrouter,anthropic/claude-3.7-sonnet:thinking"
+    ],
+    "longContext": [
+      "modelscope,Qwen/Qwen3-Coder-480B-A35B-Instruct"
+    ],
+    "webSearch": [
+      "openrouter,anthropic/claude-sonnet-4"
+    ]
+  }
+}
+```
+
+### How It Works
+
+1. **Trigger**: When a model request fails for a routing scenario (HTTP error response)
+2. **Auto-switch**: The system automatically checks the fallback configuration for that scenario
+3. **Sequential retry**: Tries each backup model in order
+4. **Success**: Once a model responds successfully, returns immediately
+5. **All failed**: If all backup models fail, returns the original error
+
+### Configuration Details
+
+- **Format**: Each backup model format is `provider,model`
+- **Validation**: Backup models must exist in the `Providers` configuration
+- **Flexibility**: Different scenarios can have different fallback lists
+- **Optional**: If a scenario doesn't need fallback, omit it or use an empty array
+
+### Use Cases
+
+#### Scenario 1: Primary Model Quota Exhausted
+
+```json
+{
+  "Router": {
+    "default": "openrouter,anthropic/claude-sonnet-4"
+  },
+  "fallback": {
+    "default": [
+      "deepseek,deepseek-chat",
+      "aihubmix,Z/glm-4.5"
+    ]
+  }
+}
+```
+
+Automatically switches to backup models when the primary model quota is exhausted.
+
+#### Scenario 2: Service Reliability
+
+```json
+{
+  "Router": {
+    "background": "volcengine,deepseek-v3-250324"
+  },
+  "fallback": {
+    "background": [
+      "modelscope,Qwen/Qwen3-Coder-480B-A35B-Instruct",
+      "dashscope,qwen3-coder-plus"
+    ]
+  }
+}
+```
+
+Automatically switches to other providers when the primary service fails.
+
+### Log Monitoring
+
+The system logs detailed fallback process:
+
+```
+[warn] Request failed for default, trying 2 fallback models
+[info] Trying fallback model: aihubmix,Z/glm-4.5
+[warn] Fallback model aihubmix,Z/glm-4.5 failed: API rate limit exceeded
+[info] Trying fallback model: openrouter,anthropic/claude-sonnet-4
+[info] Fallback model openrouter,anthropic/claude-sonnet-4 succeeded
+```
+
+### Important Notes
+
+1. **Cost consideration**: Backup models may incur different costs, configure appropriately
+2. **Performance differences**: Different models may have varying response speeds and quality
+3. **Quota management**: Ensure backup models have sufficient quotas
+4. **Testing**: Regularly test the availability of backup models
+
 ## Project-Level Routing
 
 Configure routing per project in `~/.claude/projects/<project-id>/claude-code-router.json`:
diff --git a/docs/docs/server/config/transformers.md b/docs/docs/server/config/transformers.md
index 63806cf..4a17f0f 100644
--- a/docs/docs/server/config/transformers.md
+++ b/docs/docs/server/config/transformers.md
@@ -4,7 +4,152 @@ sidebar_position: 4
 
 # Transformers
 
-Transformers adapt API differences between providers.
+Transformers are the core mechanism for adapting API differences between LLM providers. They convert requests and responses between different formats, handle authentication, and manage provider-specific features.
+
+## Understanding Transformers
+
+### What is a Transformer?
+
+A transformer is a plugin that:
+- **Transforms requests** from the unified format to provider-specific format
+- **Transforms responses** from provider format back to unified format
+- **Handles authentication** for provider APIs
+- **Modifies requests** to add or adjust parameters
+
+### Data Flow
+
+```
+┌─────────────────┐
+│ Incoming Request│ (Anthropic format from Claude Code)
+└────────┬────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  transformRequestOut            │ ← Parse incoming request to unified format
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  UnifiedChatRequest             │
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  transformRequestIn (optional)  │ ← Modify unified request before sending
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  Provider API Call              │
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  transformResponseIn (optional) │ ← Convert provider response to unified format
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  transformResponseOut (optional)│ ← Convert unified response to Anthropic format
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────┐
+│ Outgoing Response│ (Anthropic format to Claude Code)
+└─────────────────┘
+```
+
+### Transformer Interface
+
+All transformers implement the following interface:
+
+```typescript
+interface Transformer {
+  // Convert unified request to provider-specific format
+  transformRequestIn?: (
+    request: UnifiedChatRequest,
+    provider: LLMProvider,
+    context: TransformerContext
+  ) => Promise<Record<string, any>>;
+
+  // Convert provider request to unified format
+  transformRequestOut?: (
+    request: any,
+    context: TransformerContext
+  ) => Promise<UnifiedChatRequest>;
+
+  // Convert provider response to unified format
+  transformResponseIn?: (
+    response: Response,
+    context?: TransformerContext
+  ) => Promise<Response>;
+
+  // Convert unified response to provider format
+  transformResponseOut?: (
+    response: Response,
+    context: TransformerContext
+  ) => Promise<Response>;
+
+  // Custom endpoint path (optional)
+  endPoint?: string;
+
+  // Transformer name (for custom transformers)
+  name?: string;
+
+  // Custom authentication handler (optional)
+  auth?: (
+    request: any,
+    provider: LLMProvider,
+    context: TransformerContext
+  ) => Promise<any>;
+
+  // Logger instance (auto-injected)
+  logger?: any;
+}
+```
+
+### Key Types
+
+#### UnifiedChatRequest
+
+```typescript
+interface UnifiedChatRequest {
+  messages: UnifiedMessage[];
+  model: string;
+  max_tokens?: number;
+  temperature?: number;
+  stream?: boolean;
+  tools?: UnifiedTool[];
+  tool_choice?: any;
+  reasoning?: {
+    effort?: ThinkLevel;  // "none" | "low" | "medium" | "high"
+    max_tokens?: number;
+    enabled?: boolean;
+  };
+}
+```
+
+#### UnifiedMessage
+
+```typescript
+interface UnifiedMessage {
+  role: "user" | "assistant" | "system" | "tool";
+  content: string | null | MessageContent[];
+  tool_calls?: Array<{
+    id: string;
+    type: "function";
+    function: {
+      name: string;
+      arguments: string;
+    };
+  }>;
+  tool_call_id?: string;
+  thinking?: {
+    content: string;
+    signature?: string;
+  };
+}
+```
 
 ## Built-in Transformers
 
@@ -23,6 +168,12 @@ Transforms requests to be compatible with Anthropic-style APIs:
 }
 ```
 
+**Features:**
+- Converts Anthropic message format to/from OpenAI format
+- Handles tool calls and tool results
+- Supports thinking/reasoning content blocks
+- Manages streaming responses
+
 ### deepseek
 
 Specialized transformer for DeepSeek API:
@@ -38,6 +189,11 @@ Specialized transformer for DeepSeek API:
 }
 ```
 
+**Features:**
+- DeepSeek-specific reasoning format
+- Handles `reasoning_content` in responses
+- Supports thinking budget tokens
+
 ### gemini
 
 Transformer for Google Gemini API:
@@ -53,39 +209,381 @@ Transformer for Google Gemini API:
 }
 ```
 
-### groq
+### maxtoken
 
-Transformer for Groq API:
+Limits max_tokens in requests:
 
 ```json
 {
   "transformers": [
     {
-      "name": "groq",
-      "providers": ["groq"]
+      "name": "maxtoken",
+      "options": {
+        "max_tokens": 8192
+      },
+      "models": ["deepseek,deepseek-chat"]
     }
   ]
 }
 ```
 
-### openrouter
+### customparams
 
-Transformer for OpenRouter API:
+Injects custom parameters into requests:
 
 ```json
 {
   "transformers": [
     {
-      "name": "openrouter",
-      "providers": ["openrouter"]
+      "name": "customparams",
+      "options": {
+        "include_reasoning": true,
+        "custom_header": "value"
+      }
     }
   ]
 }
 ```
 
+## Creating Custom Transformers
+
+### Simple Transformer: Modifying Requests
+
+The simplest transformers just modify the request before it's sent to the provider.
+
+**Example: Add a custom header to all requests**
+
+```javascript
+// custom-header-transformer.js
+module.exports = class CustomHeaderTransformer {
+  name = 'custom-header';
+
+  constructor(options) {
+    this.headerName = options?.headerName || 'X-Custom-Header';
+    this.headerValue = options?.headerValue || 'default-value';
+  }
+
+  async transformRequestIn(request, provider, context) {
+    // Add custom header (will be used by auth method)
+    request._customHeaders = {
+      [this.headerName]: this.headerValue
+    };
+    return request;
+  }
+
+  async auth(request, provider) {
+    const headers = {
+      'authorization': `Bearer ${provider.apiKey}`,
+      ...request._customHeaders
+    };
+    return {
+      body: request,
+      config: { headers }
+    };
+  }
+};
+```
+
+**Usage in config:**
+
+```json
+{
+  "transformers": [
+    {
+      "name": "custom-header",
+      "path": "/path/to/custom-header-transformer.js",
+      "options": {
+        "headerName": "X-My-Header",
+        "headerValue": "my-value"
+      }
+    }
+  ]
+}
+```
+
+### Intermediate Transformer: Request/Response Conversion
+
+This example shows how to convert between different API formats.
+
+**Example: Mock API format transformer**
+
+```javascript
+// mockapi-transformer.js
+module.exports = class MockAPITransformer {
+  name = 'mockapi';
+  endPoint = '/v1/chat';  // Custom endpoint
+
+  // Convert from MockAPI format to unified format
+  async transformRequestOut(request, context) {
+    const messages = request.conversation.map(msg => ({
+      role: msg.sender,
+      content: msg.text
+    }));
+
+    return {
+      messages,
+      model: request.model_id,
+      max_tokens: request.max_tokens,
+      temperature: request.temp
+    };
+  }
+
+  // Convert from unified format to MockAPI format
+  async transformRequestIn(request, provider, context) {
+    return {
+      model_id: request.model,
+      conversation: request.messages.map(msg => ({
+        sender: msg.role,
+        text: typeof msg.content === 'string' ? msg.content : JSON.stringify(msg.content)
+      })),
+      max_tokens: request.max_tokens || 4096,
+      temp: request.temperature || 0.7
+    };
+  }
+
+  // Convert MockAPI response to unified format
+  async transformResponseIn(response, context) {
+    const data = await response.json();
+
+    const unifiedResponse = {
+      id: data.request_id,
+      object: 'chat.completion',
+      created: data.timestamp,
+      model: data.model,
+      choices: [{
+        index: 0,
+        message: {
+          role: 'assistant',
+          content: data.reply.text
+        },
+        finish_reason: data.stop_reason
+      }],
+      usage: {
+        prompt_tokens: data.tokens.input,
+        completion_tokens: data.tokens.output,
+        total_tokens: data.tokens.input + data.tokens.output
+      }
+    };
+
+    return new Response(JSON.stringify(unifiedResponse), {
+      status: response.status,
+      statusText: response.statusText,
+      headers: { 'Content-Type': 'application/json' }
+    });
+  }
+};
+```
+
+### Advanced Transformer: Streaming Response Processing
+
+This example shows how to handle streaming responses.
+
+**Example: Add custom metadata to streaming responses**
+
+```javascript
+// streaming-metadata-transformer.js
+module.exports = class StreamingMetadataTransformer {
+  name = 'streaming-metadata';
+
+  constructor(options) {
+    this.metadata = options?.metadata || {};
+    this.logger = null;  // Will be injected by the system
+  }
+
+  async transformResponseOut(response, context) {
+    const contentType = response.headers.get('Content-Type');
+
+    // Handle streaming response
+    if (contentType?.includes('text/event-stream')) {
+      return this.transformStream(response, context);
+    }
+
+    // Handle non-streaming response
+    return response;
+  }
+
+  async transformStream(response, context) {
+    const decoder = new TextDecoder();
+    const encoder = new TextEncoder();
+
+    const transformedStream = new ReadableStream({
+      start: async (controller) => {
+        const reader = response.body.getReader();
+        let buffer = '';
+
+        try {
+          while (true) {
+            const { done, value } = await reader.read();
+            if (done) break;
+
+            buffer += decoder.decode(value, { stream: true });
+            const lines = buffer.split('\n');
+            buffer = lines.pop() || '';
+
+            for (const line of lines) {
+              if (!line.trim() || !line.startsWith('data: ')) {
+                controller.enqueue(encoder.encode(line + '\n'));
+                continue;
+              }
+
+              const data = line.slice(6).trim();
+              if (data === '[DONE]') {
+                controller.enqueue(encoder.encode(line + '\n'));
+                continue;
+              }
+
+              try {
+                const chunk = JSON.parse(data);
+
+                // Add custom metadata
+                if (chunk.choices && chunk.choices[0]) {
+                  chunk.choices[0].metadata = this.metadata;
+                }
+
+                // Log for debugging
+                this.logger?.debug({
+                  chunk,
+                  context: context.req.id
+                }, 'Transformed streaming chunk');
+
+                const modifiedLine = `data: ${JSON.stringify(chunk)}\n\n`;
+                controller.enqueue(encoder.encode(modifiedLine));
+              } catch (parseError) {
+                // If parsing fails, pass through original line
+                controller.enqueue(encoder.encode(line + '\n'));
+              }
+            }
+          }
+        } catch (error) {
+          this.logger?.error({ error }, 'Stream transformation error');
+          controller.error(error);
+        } finally {
+          controller.close();
+          reader.releaseLock();
+        }
+      }
+    });
+
+    return new Response(transformedStream, {
+      status: response.status,
+      statusText: response.statusText,
+      headers: {
+        'Content-Type': 'text/event-stream',
+        'Cache-Control': 'no-cache',
+        'Connection': 'keep-alive'
+      }
+    });
+  }
+};
+```
+
+### Real-World Example: Reasoning Content Transformer
+
+This is based on the actual `reasoning.transformer.ts` from the codebase.
+
+```typescript
+// reasoning-transformer.ts
+import { Transformer, TransformerOptions } from "@musistudio/llms";
+
+export class ReasoningTransformer implements Transformer {
+  static TransformerName = "reasoning";
+  enable: boolean;
+
+  constructor(private readonly options?: TransformerOptions) {
+    this.enable = this.options?.enable ?? true;
+  }
+
+  // Transform request to add reasoning parameters
+  async transformRequestIn(request: UnifiedChatRequest): Promise<UnifiedChatRequest> {
+    if (!this.enable) {
+      request.thinking = {
+        type: "disabled",
+        budget_tokens: -1,
+      };
+      request.enable_thinking = false;
+      return request;
+    }
+
+    if (request.reasoning) {
+      request.thinking = {
+        type: "enabled",
+        budget_tokens: request.reasoning.max_tokens,
+      };
+      request.enable_thinking = true;
+    }
+    return request;
+  }
+
+  // Transform response to convert reasoning_content to thinking format
+  async transformResponseOut(response: Response): Promise<Response> {
+    if (!this.enable) return response;
+
+    const contentType = response.headers.get("Content-Type");
+
+    // Handle non-streaming response
+    if (contentType?.includes("application/json")) {
+      const jsonResponse = await response.json();
+      if (jsonResponse.choices[0]?.message.reasoning_content) {
+        jsonResponse.thinking = {
+          content: jsonResponse.choices[0].message.reasoning_content
+        };
+      }
+      return new Response(JSON.stringify(jsonResponse), {
+        status: response.status,
+        statusText: response.statusText,
+        headers: response.headers,
+      });
+    }
+
+    // Handle streaming response
+    if (contentType?.includes("stream")) {
+      // [Streaming transformation code here]
+      // See the full implementation in the codebase
+    }
+
+    return response;
+  }
+}
+```
+
+## Transformer Registration
+
+### Method 1: Static Name (Class-based)
+
+Use this when creating a transformer in TypeScript/ES6:
+
+```typescript
+export class MyTransformer implements Transformer {
+  static TransformerName = "my-transformer";
+
+  async transformRequestIn(request: UnifiedChatRequest): Promise<any> {
+    // Transformation logic
+    return request;
+  }
+}
+```
+
+### Method 2: Instance Name (Instance-based)
+
+Use this for JavaScript transformers:
+
+```javascript
+module.exports = class MyTransformer {
+  constructor(options) {
+    this.name = 'my-transformer';
+    this.options = options;
+  }
+
+  async transformRequestIn(request, provider, context) {
+    // Transformation logic
+    return request;
+  }
+};
+```
+
 ## Applying Transformers
 
-### Global Application
+### Global Application (Provider Level)
 
 Apply to all requests for a provider:
 
@@ -104,7 +602,7 @@ Apply to all requests for a provider:
 
 ### Model-Specific Application
 
-Apply to specific models:
+Apply to specific models only:
 
 ```json
 {
@@ -120,9 +618,26 @@ Apply to specific models:
 }
 ```
 
+Note: The model format is `provider,model` (e.g., `deepseek,deepseek-chat`).
+
+### Global Transformers (All Providers)
+
+Apply transformers to all providers:
+
+```json
+{
+  "transformers": [
+    {
+      "name": "custom-logger",
+      "path": "/path/to/custom-logger.js"
+    }
+  ]
+}
+```
+
 ### Passing Options
 
-Some transformers accept options:
+Some transformers accept configuration options:
 
 ```json
 {
@@ -132,45 +647,144 @@ Some transformers accept options:
       "options": {
         "max_tokens": 8192
       }
+    },
+    {
+      "name": "customparams",
+      "options": {
+        "custom_param_1": "value1",
+        "custom_param_2": 42
+      }
     }
   ]
 }
 ```
 
-## Custom Transformers
+## Best Practices
 
-Create custom transformer plugins:
+### 1. Immutability
 
-1. Create a transformer file:
+Always create new objects rather than mutating existing ones:
 
 ```javascript
-module.exports = {
-  name: 'my-transformer',
-  transformRequest: async (req, config) => {
-    // Modify request
-    return req;
-  },
-  transformResponse: async (res, config) => {
-    // Modify response
-    return res;
-  }
-};
-```
+// Bad
+async transformRequestIn(request) {
+  request.max_tokens = 4096;
+  return request;
+}
 
-2. Load in configuration:
-
-```json
-{
-  "transformers": [
-    {
-      "name": "my-transformer",
-      "path": "/path/to/transformer.js"
-    }
-  ]
+// Good
+async transformRequestIn(request) {
+  return {
+    ...request,
+    max_tokens: request.max_tokens || 4096
+  };
 }
 ```
 
+### 2. Error Handling
+
+Always handle errors gracefully:
+
+```javascript
+async transformResponseIn(response) {
+  try {
+    const data = await response.json();
+    // Process data
+    return new Response(JSON.stringify(processedData), {
+      status: response.status,
+      headers: response.headers
+    });
+  } catch (error) {
+    this.logger?.error({ error }, 'Transformation failed');
+    // Return original response if transformation fails
+    return response;
+  }
+}
+```
+
+### 3. Logging
+
+Use the injected logger for debugging:
+
+```javascript
+async transformRequestIn(request, provider, context) {
+  this.logger?.debug({
+    model: request.model,
+    provider: provider.name
+  }, 'Transforming request');
+
+  // Your transformation logic
+
+  return modifiedRequest;
+}
+```
+
+### 4. Stream Handling
+
+When handling streams, always:
+- Use a buffer to handle incomplete chunks
+- Properly release the reader lock
+- Handle errors in the stream
+- Close the controller when done
+
+```javascript
+const transformedStream = new ReadableStream({
+  start: async (controller) => {
+    const reader = response.body.getReader();
+    let buffer = '';
+
+    try {
+      while (true) {
+        const { done, value } = await reader.read();
+        if (done) break;
+
+        // Process stream...
+      }
+    } catch (error) {
+      controller.error(error);
+    } finally {
+      controller.close();
+      reader.releaseLock();
+    }
+  }
+});
+```
+
+### 5. Context Usage
+
+The `context` parameter contains useful information:
+
+```javascript
+async transformRequestIn(request, provider, context) {
+  // Access request ID
+  const requestId = context.req.id;
+
+  // Access original request
+  const originalRequest = context.req.original;
+
+  // Your transformation logic
+}
+```
+
+## Testing Your Transformer
+
+### Manual Testing
+
+1. Add your transformer to the config
+2. Start the server: `ccr restart`
+3. Check logs: `tail -f ~/.claude-code-router/logs/ccr-*.log`
+4. Make a test request
+5. Verify the output
+
+### Debug Tips
+
+- Add logging to track transformation steps
+- Test with both streaming and non-streaming requests
+- Verify error handling with invalid inputs
+- Check that original responses are returned on error
+
 ## Next Steps
 
-- [Advanced Topics](/docs/advanced/custom-router) - Advanced routing customization
-- [Agents](/docs/advanced/agents) - Extending with agents
+- [Advanced Topics](/docs/server/advanced/custom-router) - Advanced routing customization
+- [Agents](/docs/server/advanced/agents) - Extending with agents
+- [Core Package](/docs/server/intro) - Learn about @musistudio/llms
diff --git a/docs/docs/server/intro.md b/docs/docs/server/intro.md
index f1ab94e..7518d91 100644
--- a/docs/docs/server/intro.md
+++ b/docs/docs/server/intro.md
@@ -15,17 +15,103 @@ Claude Code Router Server is a core service component responsible for routing Cl
 ## Architecture Overview
 
 ```
-┌─────────────┐     ┌──────────────────┐     ┌──────────────┐
-│ Claude Code │────▶│ CCR Server       │────▶│ LLM Provider │
-│   Client    │     │  (Router +       │     │  (OpenAI/    │
-└─────────────┘     │   Transformer)   │     │   Gemini/etc)│
-                    └──────────────────┘     └──────────────┘
+┌─────────────┐     ┌─────────────────────────────┐     ┌──────────────┐
+│ Claude Code │────▶│ CCR Server                  │────▶│ LLM Provider │
+│   Client    │     │  ┌─────────────────────┐    │     │  (OpenAI/    │
+└─────────────┘     │  │ @musistudio/llms    │    │     │   Gemini/etc)│
+                    │  │ (Core Package)       │    │     └──────────────┘
+                    │  │ - Request Transform  │    │
+                    │  │ - Response Transform │    │
+                    │  │ - Auth Handling      │    │
+                    │  └─────────────────────┘    │
+                    │                             │
+                    │  - Routing Logic            │
+                    │  - Agent System             │
+                    │  - Configuration            │
+                    └─────────────────────────────┘
                            │
                            ├─ Web UI
                            ├─ Config API
                            └─ Logs API
 ```
 
+## Core Package: @musistudio/llms
+
+The server is built on top of **@musistudio/llms**, a universal LLM API transformation library that provides the core request/response transformation capabilities.
+
+### What is @musistudio/llms?
+
+`@musistudio/llms` is a standalone npm package (`@musistudio/llms`) that handles:
+
+- **API Format Conversion**: Transforms between different LLM provider APIs (Anthropic, OpenAI, Gemini, etc.)
+- **Request/Response Transformation**: Converts requests and responses to a unified format
+- **Authentication Handling**: Manages different authentication methods across providers
+- **Streaming Support**: Handles streaming responses from different providers
+- **Transformer System**: Provides an extensible architecture for adding new providers
+
+### Key Concepts
+
+#### 1. Unified Request/Response Format
+
+The core package defines a unified format (`UnifiedChatRequest`, `UnifiedChatResponse`) that abstracts away provider-specific differences:
+
+```typescript
+interface UnifiedChatRequest {
+  messages: UnifiedMessage[];
+  model: string;
+  max_tokens?: number;
+  temperature?: number;
+  stream?: boolean;
+  tools?: UnifiedTool[];
+  tool_choice?: any;
+  reasoning?: {
+    effort?: ThinkLevel;
+    max_tokens?: number;
+    enabled?: boolean;
+  };
+}
+```
+
+#### 2. Transformer Interface
+
+All transformers implement a common interface:
+
+```typescript
+interface Transformer {
+  transformRequestIn?: (request: UnifiedChatRequest, provider: LLMProvider, context: TransformerContext) => Promise<any>;
+  transformRequestOut?: (request: any, context: TransformerContext) => Promise<UnifiedChatRequest>;
+  transformResponseIn?: (response: Response, context?: TransformerContext) => Promise<Response>;
+  transformResponseOut?: (response: Response, context: TransformerContext) => Promise<Response>;
+  endPoint?: string;
+  name?: string;
+  auth?: (request: any, provider: LLMProvider, context: TransformerContext) => Promise<any>;
+}
+```
+
+#### 3. Built-in Transformers
+
+The core package includes transformers for:
+- **anthropic**: Anthropic API format
+- **openai**: OpenAI API format
+- **gemini**: Google Gemini API format
+- **deepseek**: DeepSeek API format
+- **groq**: Groq API format
+- **openrouter**: OpenRouter API format
+- And more...
+
+### Integration with CCR Server
+
+The CCR server integrates `@musistudio/llms` through:
+
+1. **Transformer Service** (`packages/core/src/services/transformer.ts`): Manages transformer registration and instantiation
+2. **Provider Configuration**: Maps provider configs to core package's LLMProvider interface
+3. **Request Pipeline**: Applies transformers in sequence during request processing
+4. **Custom Transformers**: Supports loading external transformer plugins
+
+### Version and Updates
+
+The current version of `@musistudio/llms` is `1.0.51`. It's published as an independent npm package and can be used standalone or as part of CCR Server.
+
 ## Core Features
 
 ### 1. Request Routing
diff --git a/docs/docusaurus.config.ts b/docs/docusaurus.config.ts
index 06bef0d..af177fc 100644
--- a/docs/docusaurus.config.ts
+++ b/docs/docusaurus.config.ts
@@ -81,35 +81,7 @@ const config: Config = {
     },
     footer: {
       style: 'light',
-      links: [
-        {
-          title: 'Docs',
-          items: [
-            {
-              label: 'Tutorial',
-              to: '/docs/intro',
-            },
-          ],
-        },
-        {
-          title: 'Community',
-          items: [
-            {
-              label: 'GitHub',
-              href: 'https://github.com/musistudio/claude-code-router',
-            },
-          ],
-        },
-        {
-          title: 'More',
-          items: [
-            {
-              label: 'Blog',
-              to: '/blog',
-            },
-          ],
-        },
-      ],
+      links: [],
       copyright: `Copyright © ${new Date().getFullYear()} Claude Code Router. Built with Docusaurus.`,
     },
     prism: {
diff --git a/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current.json b/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current.json
index 3bebc6f..fd4a06d 100644
--- a/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current.json
+++ b/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current.json
@@ -27,15 +27,15 @@
     "message": "服务器 API 接口文档",
     "description": "The generated-index page description for category 'API Reference' in sidebar 'tutorialSidebar'"
   },
-  "sidebar.tutorialSidebar.category.Configuration": {
+  "sidebar.tutorialSidebar.category.server-configuration-category": {
     "message": "配置",
     "description": "The label for category 'Configuration' in sidebar 'tutorialSidebar'"
   },
-  "sidebar.tutorialSidebar.category.Configuration.link.generated-index.title": {
+  "sidebar.tutorialSidebar.category.server-configuration-category.link.generated-index.title": {
     "message": "服务器配置",
     "description": "The generated-index page title for category 'Configuration' in sidebar 'tutorialSidebar'"
   },
-  "sidebar.tutorialSidebar.category.Configuration.link.generated-index.description": {
+  "sidebar.tutorialSidebar.category.server-configuration-category.link.generated-index.description": {
     "message": "服务器配置说明",
     "description": "The generated-index page description for category 'Configuration' in sidebar 'tutorialSidebar'"
   },
@@ -74,5 +74,17 @@
   "sidebar.tutorialSidebar.category.Commands.link.generated-index.description": {
     "message": "完整的命令参考",
     "description": "The generated-index page description for category 'Commands' in sidebar 'tutorialSidebar'"
+  },
+  "sidebar.tutorialSidebar.category.cli-configuration-category": {
+    "message": "配置",
+    "description": "The label for category 'Configuration' in sidebar 'tutorialSidebar'"
+  },
+  "sidebar.tutorialSidebar.category.cli-configuration-category.link.generated-index.title": {
+    "message": "CLI 配置",
+    "description": "The generated-index page title for category 'Configuration' in sidebar 'tutorialSidebar'"
+  },
+  "sidebar.tutorialSidebar.category.cli-configuration-category.link.generated-index.description": {
+    "message": "CLI 配置指南",
+    "description": "The generated-index page description for category 'Configuration' in sidebar 'tutorialSidebar'"
   }
 }
diff --git a/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/server/config/routing.md b/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/server/config/routing.md
index e5280ed..31326df 100644
--- a/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/server/config/routing.md
+++ b/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/server/config/routing.md
@@ -82,6 +82,115 @@ sidebar_position: 3
 }
 ```
 
+## 故障转移（Fallback）
+
+当请求失败时，可以配置备用模型列表。系统会按顺序尝试每个模型，直到请求成功：
+
+### 基本配置
+
+```json
+{
+  "Router": {
+    "default": "deepseek,deepseek-chat",
+    "background": "ollama,qwen2.5-coder:latest",
+    "think": "deepseek,deepseek-reasoner",
+    "longContext": "openrouter,google/gemini-2.5-pro-preview",
+    "longContextThreshold": 60000,
+    "webSearch": "gemini,gemini-2.5-flash"
+  },
+  "fallback": {
+    "default": [
+      "aihubmix,Z/glm-4.5",
+      "openrouter,anthropic/claude-sonnet-4"
+    ],
+    "background": [
+      "ollama,qwen2.5-coder:latest"
+    ],
+    "think": [
+      "openrouter,anthropic/claude-3.7-sonnet:thinking"
+    ],
+    "longContext": [
+      "modelscope,Qwen/Qwen3-Coder-480B-A35B-Instruct"
+    ],
+    "webSearch": [
+      "openrouter,anthropic/claude-sonnet-4"
+    ]
+  }
+}
+```
+
+### 工作原理
+
+1. **触发条件**：当某个路由场景的模型请求失败时（HTTP 错误响应）
+2. **自动切换**：系统自动检查该场景的 fallback 配置
+3. **顺序尝试**：按照列表顺序依次尝试每个备用模型
+4. **成功返回**：一旦某个模型成功响应，立即返回结果
+5. **全部失败**：如果所有备用模型都失败，返回原始错误
+
+### 配置说明
+
+- **格式**：每个备用模型格式为 `provider,model`
+- **验证**：备用模型必须在 `Providers` 配置中存在
+- **灵活性**：可以为不同场景配置不同的备用列表
+- **可选性**：如果某个场景不需要备用，可以不配置或使用空数组
+
+### 使用场景
+
+#### 场景一：主模型配额不足
+
+```json
+{
+  "Router": {
+    "default": "openrouter,anthropic/claude-sonnet-4"
+  },
+  "fallback": {
+    "default": [
+      "deepseek,deepseek-chat",
+      "aihubmix,Z/glm-4.5"
+    ]
+  }
+}
+```
+
+当主模型配额用完时，自动切换到备用模型。
+
+#### 场景二：服务稳定性保障
+
+```json
+{
+  "Router": {
+    "background": "volcengine,deepseek-v3-250324"
+  },
+  "fallback": {
+    "background": [
+      "modelscope,Qwen/Qwen3-Coder-480B-A35B-Instruct",
+      "dashscope,qwen3-coder-plus"
+    ]
+  }
+}
+```
+
+当主服务商出现故障时，自动切换到其他服务商。
+
+### 日志监控
+
+系统会记录详细的 fallback 过程：
+
+```
+[warn] Request failed for default, trying 2 fallback models
+[info] Trying fallback model: aihubmix,Z/glm-4.5
+[warn] Fallback model aihubmix,Z/glm-4.5 failed: API rate limit exceeded
+[info] Trying fallback model: openrouter,anthropic/claude-sonnet-4
+[info] Fallback model openrouter,anthropic/claude-sonnet-4 succeeded
+```
+
+### 注意事项
+
+1. **成本考虑**：备用模型可能产生不同的费用，请合理配置
+2. **性能差异**：不同模型的响应速度和质量可能有差异
+3. **配额管理**：确保备用模型有足够的配额
+4. **测试验证**：定期测试备用模型的可用性
+
 ## 项目级路由
 
 在 `~/.claude/projects/<project-id>/claude-code-router.json` 中为每个项目配置路由：
diff --git a/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/server/config/transformers.md b/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/server/config/transformers.md
index 0f08fc1..c861eb4 100644
--- a/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/server/config/transformers.md
+++ b/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/server/config/transformers.md
@@ -5,7 +5,152 @@ sidebar_position: 4
 
 # 转换器
 
-转换器用于适配不同提供商之间的 API 差异。
+转换器是适配不同 LLM 提供商 API 差异的核心机制。它们在不同格式之间转换请求和响应，处理认证，并管理提供商特定的功能。
+
+## 理解转换器
+
+### 什么是转换器？
+
+转换器是一个插件，它可以：
+- **转换请求**：从统一格式转换为提供商特定格式
+- **转换响应**：从提供商格式转换回统一格式
+- **处理认证**：为提供商 API 处理认证
+- **修改请求**：添加或调整参数
+
+### 数据流
+
+```
+┌─────────────────┐
+│ 传入请求        │ (来自 Claude Code 的 Anthropic 格式)
+└────────┬────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  transformRequestOut            │ ← 将传入请求解析为统一格式
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  UnifiedChatRequest             │
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  transformRequestIn (可选)      │ ← 在发送前修改统一请求
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  提供商 API 调用                │
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  transformResponseIn (可选)     │ ← 将提供商响应转换为统一格式
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  transformResponseOut (可选)    │ ← 将统一响应转换为 Anthropic 格式
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────┐
+│ 传出响应        │ (返回给 Claude Code 的 Anthropic 格式)
+└─────────────────┘
+```
+
+### 转换器接口
+
+所有转换器都实现以下接口：
+
+```typescript
+interface Transformer {
+  // 将统一请求转换为提供商特定格式
+  transformRequestIn?: (
+    request: UnifiedChatRequest,
+    provider: LLMProvider,
+    context: TransformerContext
+  ) => Promise<Record<string, any>>;
+
+  // 将提供商请求转换为统一格式
+  transformRequestOut?: (
+    request: any,
+    context: TransformerContext
+  ) => Promise<UnifiedChatRequest>;
+
+  // 将提供商响应转换为统一格式
+  transformResponseIn?: (
+    response: Response,
+    context?: TransformerContext
+  ) => Promise<Response>;
+
+  // 将统一响应转换为提供商格式
+  transformResponseOut?: (
+    response: Response,
+    context: TransformerContext
+  ) => Promise<Response>;
+
+  // 自定义端点路径（可选）
+  endPoint?: string;
+
+  // 转换器名称（用于自定义转换器）
+  name?: string;
+
+  // 自定义认证处理器（可选）
+  auth?: (
+    request: any,
+    provider: LLMProvider,
+    context: TransformerContext
+  ) => Promise<any>;
+
+  // Logger 实例（自动注入）
+  logger?: any;
+}
+```
+
+### 关键类型
+
+#### UnifiedChatRequest
+
+```typescript
+interface UnifiedChatRequest {
+  messages: UnifiedMessage[];
+  model: string;
+  max_tokens?: number;
+  temperature?: number;
+  stream?: boolean;
+  tools?: UnifiedTool[];
+  tool_choice?: any;
+  reasoning?: {
+    effort?: ThinkLevel;  // "none" | "low" | "medium" | "high"
+    max_tokens?: number;
+    enabled?: boolean;
+  };
+}
+```
+
+#### UnifiedMessage
+
+```typescript
+interface UnifiedMessage {
+  role: "user" | "assistant" | "system" | "tool";
+  content: string | null | MessageContent[];
+  tool_calls?: Array<{
+    id: string;
+    type: "function";
+    function: {
+      name: string;
+      arguments: string;
+    };
+  }>;
+  tool_call_id?: string;
+  thinking?: {
+    content: string;
+    signature?: string;
+  };
+}
+```
 
 ## 内置转换器
 
@@ -15,13 +160,20 @@ sidebar_position: 4
 
 ```json
 {
-  "transformer": {
-    "use": ["anthropic"]
-  }
+  "transformers": [
+    {
+      "name": "anthropic",
+      "providers": ["deepseek", "groq"]
+    }
+  ]
 }
 ```
 
-如果只使用这一个转换器，它将直接透传请求和响应（您可以用来接入其他支持 Anthropic 端点的服务商）。
+**功能：**
+- 在 Anthropic 消息格式和 OpenAI 格式之间转换
+- 处理工具调用和工具结果
+- 支持思考/推理内容块
+- 管理流式响应
 
 ### deepseek
 
@@ -29,169 +181,421 @@ sidebar_position: 4
 
 ```json
 {
-  "transformer": {
-    "use": ["deepseek"]
-  }
+  "transformers": [
+    {
+      "name": "deepseek",
+      "providers": ["deepseek"]
+    }
+  ]
 }
 ```
 
+**功能：**
+- DeepSeek 特定的推理格式
+- 处理响应中的 `reasoning_content`
+- 支持思考预算令牌
+
 ### gemini
 
 用于 Google Gemini API 的转换器：
 
 ```json
 {
-  "transformer": {
-    "use": ["gemini"]
-  }
-}
-```
-
-### groq
-
-用于 Groq API 的转换器：
-
-```json
-{
-  "transformer": {
-    "use": ["groq"]
-  }
-}
-```
-
-### openrouter
-
-用于 OpenRouter API 的转换器：
-
-```json
-{
-  "transformer": {
-    "use": ["openrouter"]
-  }
-}
-```
-
-OpenRouter 转换器还支持 `provider` 路由参数，以指定 OpenRouter 应使用哪些底层提供商：
-
-```json
-{
-  "transformer": {
-    "use": ["openrouter"],
-    "moonshotai/kimi-k2": {
-      "use": [
-        ["openrouter", {
-          "provider": {
-            "only": ["moonshotai/fp8"]
-          }
-        }]
-      ]
+  "transformers": [
+    {
+      "name": "gemini",
+      "providers": ["gemini"]
     }
-  }
+  ]
 }
 ```
 
 ### maxtoken
 
-设置特定的 `max_tokens` 值：
+限制请求中的 max_tokens：
 
 ```json
 {
-  "transformer": {
-    "use": [
-      ["maxtoken", { "max_tokens": 65536 }]
-    ]
+  "transformers": [
+    {
+      "name": "maxtoken",
+      "options": {
+        "max_tokens": 8192
+      },
+      "models": ["deepseek,deepseek-chat"]
+    }
+  ]
+}
+```
+
+### customparams
+
+向请求中注入自定义参数：
+
+```json
+{
+  "transformers": [
+    {
+      "name": "customparams",
+      "options": {
+        "include_reasoning": true,
+        "custom_header": "value"
+      }
+    }
+  ]
+}
+```
+
+## 创建自定义转换器
+
+### 简单转换器：修改请求
+
+最简单的转换器只修改发送到提供商之前的请求。
+
+**示例：为所有请求添加自定义头**
+
+```javascript
+// custom-header-transformer.js
+module.exports = class CustomHeaderTransformer {
+  name = 'custom-header';
+
+  constructor(options) {
+    this.headerName = options?.headerName || 'X-Custom-Header';
+    this.headerValue = options?.headerValue || 'default-value';
+  }
+
+  async transformRequestIn(request, provider, context) {
+    // 添加自定义头（将被 auth 方法使用）
+    request._customHeaders = {
+      [this.headerName]: this.headerValue
+    };
+    return request;
+  }
+
+  async auth(request, provider) {
+    const headers = {
+      'authorization': `Bearer ${provider.apiKey}`,
+      ...request._customHeaders
+    };
+    return {
+      body: request,
+      config: { headers }
+    };
+  }
+};
+```
+
+**在配置中使用：**
+
+```json
+{
+  "transformers": [
+    {
+      "name": "custom-header",
+      "path": "/path/to/custom-header-transformer.js",
+      "options": {
+        "headerName": "X-My-Header",
+        "headerValue": "my-value"
+      }
+    }
+  ]
+}
+```
+
+### 中级转换器：请求/响应转换
+
+此示例展示如何在不同 API 格式之间转换。
+
+**示例：Mock API 格式转换器**
+
+```javascript
+// mockapi-transformer.js
+module.exports = class MockAPITransformer {
+  name = 'mockapi';
+  endPoint = '/v1/chat';  // 自定义端点
+
+  // 从 MockAPI 格式转换为统一格式
+  async transformRequestOut(request, context) {
+    const messages = request.conversation.map(msg => ({
+      role: msg.sender,
+      content: msg.text
+    }));
+
+    return {
+      messages,
+      model: request.model_id,
+      max_tokens: request.max_tokens,
+      temperature: request.temp
+    };
+  }
+
+  // 从统一格式转换为 MockAPI 格式
+  async transformRequestIn(request, provider, context) {
+    return {
+      model_id: request.model,
+      conversation: request.messages.map(msg => ({
+        sender: msg.role,
+        text: typeof msg.content === 'string' ? msg.content : JSON.stringify(msg.content)
+      })),
+      max_tokens: request.max_tokens || 4096,
+      temp: request.temperature || 0.7
+    };
+  }
+
+  // 将 MockAPI 响应转换为统一格式
+  async transformResponseIn(response, context) {
+    const data = await response.json();
+
+    const unifiedResponse = {
+      id: data.request_id,
+      object: 'chat.completion',
+      created: data.timestamp,
+      model: data.model,
+      choices: [{
+        index: 0,
+        message: {
+          role: 'assistant',
+          content: data.reply.text
+        },
+        finish_reason: data.stop_reason
+      }],
+      usage: {
+        prompt_tokens: data.tokens.input,
+        completion_tokens: data.tokens.output,
+        total_tokens: data.tokens.input + data.tokens.output
+      }
+    };
+
+    return new Response(JSON.stringify(unifiedResponse), {
+      status: response.status,
+      statusText: response.statusText,
+      headers: { 'Content-Type': 'application/json' }
+    });
+  }
+};
+```
+
+### 高级转换器：流式响应处理
+
+此示例展示如何处理流式响应。
+
+**示例：向流式响应添加自定义元数据**
+
+```javascript
+// streaming-metadata-transformer.js
+module.exports = class StreamingMetadataTransformer {
+  name = 'streaming-metadata';
+
+  constructor(options) {
+    this.metadata = options?.metadata || {};
+    this.logger = null;  // 将由系统注入
+  }
+
+  async transformResponseOut(response, context) {
+    const contentType = response.headers.get('Content-Type');
+
+    // 处理流式响应
+    if (contentType?.includes('text/event-stream')) {
+      return this.transformStream(response, context);
+    }
+
+    // 处理非流式响应
+    return response;
+  }
+
+  async transformStream(response, context) {
+    const decoder = new TextDecoder();
+    const encoder = new TextEncoder();
+
+    const transformedStream = new ReadableStream({
+      start: async (controller) => {
+        const reader = response.body.getReader();
+        let buffer = '';
+
+        try {
+          while (true) {
+            const { done, value } = await reader.read();
+            if (done) break;
+
+            buffer += decoder.decode(value, { stream: true });
+            const lines = buffer.split('\n');
+            buffer = lines.pop() || '';
+
+            for (const line of lines) {
+              if (!line.trim() || !line.startsWith('data: ')) {
+                controller.enqueue(encoder.encode(line + '\n'));
+                continue;
+              }
+
+              const data = line.slice(6).trim();
+              if (data === '[DONE]') {
+                controller.enqueue(encoder.encode(line + '\n'));
+                continue;
+              }
+
+              try {
+                const chunk = JSON.parse(data);
+
+                // 添加自定义元数据
+                if (chunk.choices && chunk.choices[0]) {
+                  chunk.choices[0].metadata = this.metadata;
+                }
+
+                // 记录日志以便调试
+                this.logger?.debug({
+                  chunk,
+                  context: context.req.id
+                }, '转换流式数据块');
+
+                const modifiedLine = `data: ${JSON.stringify(chunk)}\n\n`;
+                controller.enqueue(encoder.encode(modifiedLine));
+              } catch (parseError) {
+                // 如果解析失败，透传原始行
+                controller.enqueue(encoder.encode(line + '\n'));
+              }
+            }
+          }
+        } catch (error) {
+          this.logger?.error({ error }, '流式转换错误');
+          controller.error(error);
+        } finally {
+          controller.close();
+          reader.releaseLock();
+        }
+      }
+    });
+
+    return new Response(transformedStream, {
+      status: response.status,
+      statusText: response.statusText,
+      headers: {
+        'Content-Type': 'text/event-stream',
+        'Cache-Control': 'no-cache',
+        'Connection': 'keep-alive'
+      }
+    });
+  }
+};
+```
+
+### 真实示例：推理内容转换器
+
+这是基于代码库中实际的 `reasoning.transformer.ts`。
+
+```typescript
+// reasoning-transformer.ts
+import { Transformer, TransformerOptions } from "@musistudio/llms";
+
+export class ReasoningTransformer implements Transformer {
+  static TransformerName = "reasoning";
+  enable: boolean;
+
+  constructor(private readonly options?: TransformerOptions) {
+    this.enable = this.options?.enable ?? true;
+  }
+
+  // 转换请求以添加推理参数
+  async transformRequestIn(request: UnifiedChatRequest): Promise<UnifiedChatRequest> {
+    if (!this.enable) {
+      request.thinking = {
+        type: "disabled",
+        budget_tokens: -1,
+      };
+      request.enable_thinking = false;
+      return request;
+    }
+
+    if (request.reasoning) {
+      request.thinking = {
+        type: "enabled",
+        budget_tokens: request.reasoning.max_tokens,
+      };
+      request.enable_thinking = true;
+    }
+    return request;
+  }
+
+  // 转换响应以将 reasoning_content 转换为 thinking 格式
+  async transformResponseOut(response: Response): Promise<Response> {
+    if (!this.enable) return response;
+
+    const contentType = response.headers.get("Content-Type");
+
+    // 处理非流式响应
+    if (contentType?.includes("application/json")) {
+      const jsonResponse = await response.json();
+      if (jsonResponse.choices[0]?.message.reasoning_content) {
+        jsonResponse.thinking = {
+          content: jsonResponse.choices[0].message.reasoning_content
+        };
+      }
+      return new Response(JSON.stringify(jsonResponse), {
+        status: response.status,
+        statusText: response.statusText,
+        headers: response.headers,
+      });
+    }
+
+    // 处理流式响应
+    if (contentType?.includes("stream")) {
+      // [流式转换代码在这里]
+      // 参见代码库中的完整实现
+    }
+
+    return response;
   }
 }
 ```
 
-### tooluse
+## 转换器注册
 
-通过 `tool_choice` 参数优化某些模型的工具使用：
+### 方法 1：静态名称（基于类）
 
-```json
-{
-  "transformer": {
-    "use": ["tooluse"]
+在 TypeScript/ES6 中创建转换器时使用：
+
+```typescript
+export class MyTransformer implements Transformer {
+  static TransformerName = "my-transformer";
+
+  async transformRequestIn(request: UnifiedChatRequest): Promise<any> {
+    // 转换逻辑
+    return request;
   }
 }
 ```
 
-### reasoning
+### 方法 2：实例名称（基于实例）
 
-用于处理 `reasoning_content` 字段：
+用于 JavaScript 转换器：
 
-```json
-{
-  "transformer": {
-    "use": ["reasoning"]
+```javascript
+module.exports = class MyTransformer {
+  constructor(options) {
+    this.name = 'my-transformer';
+    this.options = options;
   }
-}
-```
 
-### sampling
-
-用于处理采样信息字段，如 `temperature`、`top_p`、`top_k` 和 `repetition_penalty`：
-
-```json
-{
-  "transformer": {
-    "use": ["sampling"]
+  async transformRequestIn(request, provider, context) {
+    // 转换逻辑
+    return request;
   }
-}
-```
-
-### enhancetool
-
-对 LLM 返回的工具调用参数增加一层容错处理（注意：这会导致不再流式返回工具调用信息）：
-
-```json
-{
-  "transformer": {
-    "use": ["enhancetool"]
-  }
-}
-```
-
-### cleancache
-
-清除请求中的 `cache_control` 字段：
-
-```json
-{
-  "transformer": {
-    "use": ["cleancache"]
-  }
-}
-```
-
-### vertex-gemini
-
-处理使用 Vertex 鉴权的 Gemini API：
-
-```json
-{
-  "transformer": {
-    "use": ["vertex-gemini"]
-  }
-}
+};
 ```
 
 ## 应用转换器
 
-### 全局应用
+### 全局应用（提供商级别）
 
-应用于提供商的所有请求：
+为提供商的所有请求应用：
 
 ```json
 {
   "Providers": [
     {
-      "name": "deepseek",
-      "api_base_url": "https://api.deepseek.com/chat/completions",
-      "api_key": "your-api-key",
-      "transformer": {
-        "use": ["deepseek"]
-      }
+      "NAME": "deepseek",
+      "HOST": "https://api.deepseek.com",
+      "APIKEY": "your-api-key",
+      "transformers": ["anthropic"]
     }
   ]
 }
@@ -199,84 +603,189 @@ OpenRouter 转换器还支持 `provider` 路由参数，以指定 OpenRouter 应
 
 ### 模型特定应用
 
-应用于特定模型：
-
-```json
-{
-  "name": "deepseek",
-  "transformer": {
-    "use": ["deepseek"],
-    "deepseek-chat": {
-      "use": ["tooluse"]
-    }
-  }
-}
-```
-
-### 传递选项
-
-某些转换器接受选项：
-
-```json
-{
-  "transformer": {
-    "use": [
-      ["maxtoken", { "max_tokens": 8192 }]
-    ]
-  }
-}
-```
-
-## 自定义转换器
-
-创建自定义转换器插件：
-
-1. 创建转换器文件：
-
-```javascript
-module.exports = {
-  name: 'my-transformer',
-  transformRequest: async (req, config) => {
-    // 修改请求
-    return req;
-  },
-  transformResponse: async (res, config) => {
-    // 修改响应
-    return res;
-  }
-};
-```
-
-2. 在配置中加载：
+仅应用于特定模型：
 
 ```json
 {
   "transformers": [
     {
-      "path": "/path/to/transformer.js",
+      "name": "maxtoken",
       "options": {
-        "key": "value"
+        "max_tokens": 8192
+      },
+      "models": ["deepseek,deepseek-chat"]
+    }
+  ]
+}
+```
+
+注意：模型格式为 `provider,model`（例如 `deepseek,deepseek-chat`）。
+
+### 全局转换器（所有提供商）
+
+将转换器应用于所有提供商：
+
+```json
+{
+  "transformers": [
+    {
+      "name": "custom-logger",
+      "path": "/path/to/custom-logger.js"
+    }
+  ]
+}
+```
+
+### 传递选项
+
+某些转换器接受配置选项：
+
+```json
+{
+  "transformers": [
+    {
+      "name": "maxtoken",
+      "options": {
+        "max_tokens": 8192
+      }
+    },
+    {
+      "name": "customparams",
+      "options": {
+        "custom_param_1": "value1",
+        "custom_param_2": 42
       }
     }
   ]
 }
 ```
 
-## 实验性转换器
+## 最佳实践
 
-### gemini-cli（实验性）
+### 1. 不可变性
 
-通过 Gemini CLI 对 Gemini 的非官方支持。
+始终创建新对象而不是修改现有对象：
 
-### qwen-cli（实验性）
+```javascript
+// 不好的做法
+async transformRequestIn(request) {
+  request.max_tokens = 4096;
+  return request;
+}
 
-通过 Qwen CLI 对 qwen3-coder-plus 的非官方支持。
+// 好的做法
+async transformRequestIn(request) {
+  return {
+    ...request,
+    max_tokens: request.max_tokens || 4096
+  };
+}
+```
 
-### rovo-cli（实验性）
+### 2. 错误处理
 
-通过 Atlassian Rovo Dev CLI 对 GPT-5 的非官方支持。
+始终优雅地处理错误：
+
+```javascript
+async transformResponseIn(response) {
+  try {
+    const data = await response.json();
+    // 处理数据
+    return new Response(JSON.stringify(processedData), {
+      status: response.status,
+      headers: response.headers
+    });
+  } catch (error) {
+    this.logger?.error({ error }, '转换失败');
+    // 如果转换失败，返回原始响应
+    return response;
+  }
+}
+```
+
+### 3. 日志记录
+
+使用注入的 logger 进行调试：
+
+```javascript
+async transformRequestIn(request, provider, context) {
+  this.logger?.debug({
+    model: request.model,
+    provider: provider.name
+  }, '转换请求');
+
+  // 转换逻辑
+
+  return modifiedRequest;
+}
+```
+
+### 4. 流处理
+
+处理流式响应时，始终：
+- 使用缓冲区处理不完整的数据块
+- 正确释放 reader 锁
+- 处理流中的错误
+- 完成时关闭 controller
+
+```javascript
+const transformedStream = new ReadableStream({
+  start: async (controller) => {
+    const reader = response.body.getReader();
+    let buffer = '';
+
+    try {
+      while (true) {
+        const { done, value } = await reader.read();
+        if (done) break;
+
+        // 处理流...
+      }
+    } catch (error) {
+      controller.error(error);
+    } finally {
+      controller.close();
+      reader.releaseLock();
+    }
+  }
+});
+```
+
+### 5. 上下文使用
+
+`context` 参数包含有用信息：
+
+```javascript
+async transformRequestIn(request, provider, context) {
+  // 访问请求 ID
+  const requestId = context.req.id;
+
+  // 访问原始请求
+  const originalRequest = context.req.original;
+
+  // 转换逻辑
+}
+```
+
+## 测试转换器
+
+### 手动测试
+
+1. 将转换器添加到配置
+2. 启动服务器：`ccr restart`
+3. 检查日志：`tail -f ~/.claude-code-router/logs/ccr-*.log`
+4. 发出测试请求
+5. 验证输出
+
+### 调试技巧
+
+- 添加日志记录以跟踪转换步骤
+- 使用流式和非流式请求进行测试
+- 使用无效输入验证错误处理
+- 检查错误时是否返回原始响应
 
 ## 下一步
 
-- [高级主题](/zh/docs/advanced/custom-router) - 高级路由自定义
-- [Agent](/zh/docs/advanced/agents) - 使用 Agent 扩展功能
+- [高级主题](/docs/server/advanced/custom-router) - 高级路由自定义
+- [Agents](/docs/server/advanced/agents) - 使用 agents 扩展
+- [核心包](/docs/server/intro) - 了解 @musistudio/llms
diff --git a/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/server/intro.md b/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/server/intro.md
index ddb3ff3..438c4fa 100644
--- a/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/server/intro.md
+++ b/docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/server/intro.md
@@ -11,17 +11,103 @@ Claude Code Router Server 是一个核心服务组件，负责将 Claude Code 
 ## 架构概述
 
 ```
-┌─────────────┐     ┌──────────────────┐     ┌──────────────┐
-│ Claude Code │────▶│ CCR Server       │────▶│ LLM Provider │
-│   Client    │     │  (Router +       │     │  (OpenAI/    │
-└─────────────┘     │   Transformer)   │     │   Gemini/etc)│
-                    └──────────────────┘     └──────────────┘
+┌─────────────┐     ┌─────────────────────────────┐     ┌──────────────┐
+│ Claude Code │────▶│ CCR Server                  │────▶│ LLM Provider │
+│   Client    │     │  ┌─────────────────────┐    │     │  (OpenAI/    │
+└─────────────┘     │  │ @musistudio/llms    │    │     │   Gemini/etc)│
+                    │  │ (核心包)             │    │     └──────────────┘
+                    │  │ - 请求转换           │    │
+                    │  │ - 响应转换           │    │
+                    │  │ - 认证处理           │    │
+                    │  └─────────────────────┘    │
+                    │                             │
+                    │  - 路由逻辑                 │
+                    │  - Agent 系统               │
+                    │  - 配置管理                 │
+                    └─────────────────────────────┘
                            │
                            ├─ Web UI
                            ├─ Config API
                            └─ Logs API
 ```
 
+## 核心包：@musistudio/llms
+
+服务器构建于 **@musistudio/llms** 之上，这是一个通用的 LLM API 转换库，提供了核心的请求/响应转换能力。
+
+### 什么是 @musistudio/llms？
+
+`@musistudio/llms` 是一个独立的 npm 包（`@musistudio/llms`），负责处理：
+
+- **API 格式转换**：在不同的 LLM 提供商 API 之间转换（Anthropic、OpenAI、Gemini 等）
+- **请求/响应转换**：将请求和响应转换为统一格式
+- **认证处理**：管理不同提供商的认证方法
+- **流式响应支持**：处理来自不同提供商的流式响应
+- **转换器系统**：提供可扩展的架构来添加新的提供商
+
+### 核心概念
+
+#### 1. 统一请求/响应格式
+
+核心包定义了统一格式（`UnifiedChatRequest`、`UnifiedChatResponse`），抽象了提供商特定的差异：
+
+```typescript
+interface UnifiedChatRequest {
+  messages: UnifiedMessage[];
+  model: string;
+  max_tokens?: number;
+  temperature?: number;
+  stream?: boolean;
+  tools?: UnifiedTool[];
+  tool_choice?: any;
+  reasoning?: {
+    effort?: ThinkLevel;
+    max_tokens?: number;
+    enabled?: boolean;
+  };
+}
+```
+
+#### 2. 转换器接口
+
+所有转换器都实现一个通用接口：
+
+```typescript
+interface Transformer {
+  transformRequestIn?: (request: UnifiedChatRequest, provider: LLMProvider, context: TransformerContext) => Promise<any>;
+  transformRequestOut?: (request: any, context: TransformerContext) => Promise<UnifiedChatRequest>;
+  transformResponseIn?: (response: Response, context?: TransformerContext) => Promise<Response>;
+  transformResponseOut?: (response: Response, context: TransformerContext) => Promise<Response>;
+  endPoint?: string;
+  name?: string;
+  auth?: (request: any, provider: LLMProvider, context: TransformerContext) => Promise<any>;
+}
+```
+
+#### 3. 内置转换器
+
+核心包包含以下转换器：
+- **anthropic**：Anthropic API 格式
+- **openai**：OpenAI API 格式
+- **gemini**：Google Gemini API 格式
+- **deepseek**：DeepSeek API 格式
+- **groq**：Groq API 格式
+- **openrouter**：OpenRouter API 格式
+- 等等...
+
+### 与 CCR Server 的集成
+
+CCR server 通过以下方式集成 `@musistudio/llms`：
+
+1. **转换器服务**（`packages/core/src/services/transformer.ts`）：管理转换器的注册和实例化
+2. **提供商配置**：将提供商配置映射到核心包的 LLMProvider 接口
+3. **请求管道**：在请求处理过程中按顺序应用转换器
+4. **自定义转换器**：支持加载外部转换器插件
+
+### 版本和更新
+
+`@musistudio/llms` 的当前版本是 `1.0.51`。它作为独立的 npm 包发布，可以独立使用或作为 CCR Server 的一部分使用。
+
 ## 核心功能
 
 ### 1. 请求路由
diff --git a/docs/i18n/zh-CN/docusaurus-theme-classic/navbar.json b/docs/i18n/zh-CN/docusaurus-theme-classic/navbar.json
index 547f110..1519c01 100644
--- a/docs/i18n/zh-CN/docusaurus-theme-classic/navbar.json
+++ b/docs/i18n/zh-CN/docusaurus-theme-classic/navbar.json
@@ -8,11 +8,11 @@
     "description": "The alt text of navbar logo"
   },
   "item.label.Documentation": {
-    "message": "Documentation",
+    "message": "文档",
     "description": "Navbar item with label Documentation"
   },
   "item.label.Blog": {
-    "message": "Blog",
+    "message": "博客",
     "description": "Navbar item with label Blog"
   },
   "item.label.GitHub": {
diff --git a/docs/sidebars.ts b/docs/sidebars.ts
index 9fa17f9..18d4b4e 100644
--- a/docs/sidebars.ts
+++ b/docs/sidebars.ts
@@ -2,6 +2,54 @@ import type { SidebarsConfig } from '@docusaurus/plugin-content-docs';
 
 const sidebars: SidebarsConfig = {
   tutorialSidebar: [
+    {
+      type: 'category',
+      label: 'CLI',
+      link: {
+        type: 'generated-index',
+        title: 'Claude Code Router CLI',
+        description: 'Command-line tool usage guide',
+        slug: 'category/cli',
+      },
+      items: [
+        'cli/intro',
+        'cli/installation',
+        'cli/quick-start',
+        {
+          type: 'category',
+          label: 'Commands',
+          link: {
+            type: 'generated-index',
+            title: 'CLI Commands',
+            description: 'Complete command reference',
+            slug: 'category/cli-commands',
+          },
+          items: [
+            'cli/commands/start',
+            'cli/commands/model',
+            'cli/commands/status',
+            'cli/commands/statusline',
+            'cli/commands/preset',
+            'cli/commands/other',
+          ],
+        },
+        {
+          type: 'category',
+          label: 'Configuration',
+          key: 'cli-configuration-category',
+          link: {
+            type: 'generated-index',
+            title: 'CLI Configuration',
+            description: 'CLI configuration guide',
+            slug: 'category/cli-config',
+          },
+          items: [
+            'cli/config/basic',
+            'cli/config/project-level',
+          ],
+        },
+      ],
+    },
     {
       type: 'category',
       label: 'Server',
@@ -63,54 +111,6 @@ const sidebars: SidebarsConfig = {
         },
       ],
     },
-    {
-      type: 'category',
-      label: 'CLI',
-      link: {
-        type: 'generated-index',
-        title: 'Claude Code Router CLI',
-        description: 'Command-line tool usage guide',
-        slug: 'category/cli',
-      },
-      items: [
-        'cli/intro',
-        'cli/installation',
-        'cli/quick-start',
-        {
-          type: 'category',
-          label: 'Commands',
-          link: {
-            type: 'generated-index',
-            title: 'CLI Commands',
-            description: 'Complete command reference',
-            slug: 'category/cli-commands',
-          },
-          items: [
-            'cli/commands/start',
-            'cli/commands/model',
-            'cli/commands/status',
-            'cli/commands/statusline',
-            'cli/commands/preset',
-            'cli/commands/other',
-          ],
-        },
-        {
-          type: 'category',
-          label: 'Configuration',
-          key: 'cli-configuration-category',
-          link: {
-            type: 'generated-index',
-            title: 'CLI Configuration',
-            description: 'CLI configuration guide',
-            slug: 'category/cli-config',
-          },
-          items: [
-            'cli/config/basic',
-            'cli/config/project-level',
-          ],
-        },
-      ],
-    },
   ],
 };
 
diff --git a/package.json b/package.json
index 2e3f8b8..b7d3b00 100644
--- a/package.json
+++ b/package.json
@@ -44,7 +44,7 @@
     }
   },
   "engines": {
-    "node": ">=18.0.0",
+    "node": ">=20.0.0",
     "pnpm": ">=8.0.0"
   }
 }
diff --git a/packages/core/src/api/routes.ts b/packages/core/src/api/routes.ts
index 5ef76a9..bfe24aa 100644
--- a/packages/core/src/api/routes.ts
+++ b/packages/core/src/api/routes.ts
@@ -50,44 +50,147 @@ async function handleTransformerEndpoint(
     );
   }
 
-  // Process request transformer chain
-  const { requestBody, config, bypass } = await processRequestTransformers(
-    body,
-    provider,
-    transformer,
-    req.headers,
-    {
-      req,
-    }
-  );
+  try {
+    // Process request transformer chain
+    const { requestBody, config, bypass } = await processRequestTransformers(
+      body,
+      provider,
+      transformer,
+      req.headers,
+      {
+        req,
+      }
+    );
 
-  // Send request to LLM provider
-  const response = await sendRequestToProvider(
-    requestBody,
-    config,
-    provider,
-    fastify,
-    bypass,
-    transformer,
-    {
-      req,
-    }
-  );
+    // Send request to LLM provider
+    const response = await sendRequestToProvider(
+      requestBody,
+      config,
+      provider,
+      fastify,
+      bypass,
+      transformer,
+      {
+        req,
+      }
+    );
 
-  // Process response transformer chain
-  const finalResponse = await processResponseTransformers(
-    requestBody,
-    response,
-    provider,
-    transformer,
-    bypass,
-    {
-      req,
-    }
-  );
+    // Process response transformer chain
+    const finalResponse = await processResponseTransformers(
+      requestBody,
+      response,
+      provider,
+      transformer,
+      bypass,
+      {
+        req,
+      }
+    );
 
-  // Format and return response
-  return formatResponse(finalResponse, reply, body);
+    // Format and return response
+    return formatResponse(finalResponse, reply, body);
+  } catch (error: any) {
+    // Handle fallback if error occurs
+    if (error.code === 'provider_response_error') {
+      const fallbackResult = await handleFallback(req, reply, fastify, transformer, error);
+      if (fallbackResult) {
+        return fallbackResult;
+      }
+    }
+    throw error;
+  }
+}
+
+/**
+ * Handle fallback logic when request fails
+ * Tries each fallback model in sequence until one succeeds
+ */
+async function handleFallback(
+  req: FastifyRequest,
+  reply: FastifyReply,
+  fastify: FastifyInstance,
+  transformer: any,
+  error: any
+): Promise<any> {
+  const scenarioType = (req as any).scenarioType || 'default';
+  const fallbackConfig = fastify.configService.get<any>('fallback');
+
+  if (!fallbackConfig || !fallbackConfig[scenarioType]) {
+    return null;
+  }
+
+  const fallbackList = fallbackConfig[scenarioType] as string[];
+  if (!Array.isArray(fallbackList) || fallbackList.length === 0) {
+    return null;
+  }
+
+  req.log.warn(`Request failed for ${(req as any).scenarioType}, trying ${fallbackList.length} fallback models`);
+
+  // Try each fallback model in sequence
+  for (const fallbackModel of fallbackList) {
+    try {
+      req.log.info(`Trying fallback model: ${fallbackModel}`);
+
+      // Update request with fallback model
+      const newBody = { ...(req.body as any) };
+      const [fallbackProvider, ...fallbackModelName] = fallbackModel.split(',');
+      newBody.model = fallbackModelName.join(',');
+
+      // Create new request object with updated provider and body
+      const newReq = {
+        ...req,
+        provider: fallbackProvider,
+        body: newBody,
+      };
+
+      const provider = fastify.providerService.getProvider(fallbackProvider);
+      if (!provider) {
+        req.log.warn(`Fallback provider '${fallbackProvider}' not found, skipping`);
+        continue;
+      }
+
+      // Process request transformer chain
+      const { requestBody, config, bypass } = await processRequestTransformers(
+        newBody,
+        provider,
+        transformer,
+        req.headers,
+        { req: newReq }
+      );
+
+      // Send request to LLM provider
+      const response = await sendRequestToProvider(
+        requestBody,
+        config,
+        provider,
+        fastify,
+        bypass,
+        transformer,
+        { req: newReq }
+      );
+
+      // Process response transformer chain
+      const finalResponse = await processResponseTransformers(
+        requestBody,
+        response,
+        provider,
+        transformer,
+        bypass,
+        { req: newReq }
+      );
+
+      req.log.info(`Fallback model ${fallbackModel} succeeded`);
+
+      // Format and return response
+      return formatResponse(finalResponse, reply, newBody);
+    } catch (fallbackError: any) {
+      req.log.warn(`Fallback model ${fallbackModel} failed: ${fallbackError.message}`);
+      continue;
+    }
+  }
+
+  req.log.error(`All fallback models failed for yichu ${scenarioType}`);
+  return null;
 }
 
 /**
diff --git a/packages/core/src/server.ts b/packages/core/src/server.ts
index 9dc34bc..c988c2e 100644
--- a/packages/core/src/server.ts
+++ b/packages/core/src/server.ts
@@ -39,6 +39,7 @@ declare module "fastify" {
   interface FastifyRequest {
     provider?: string;
     model?: string;
+    scenarioType?: string;
   }
   interface FastifyInstance {
     _server?: Server;
@@ -266,6 +267,7 @@ export { sessionUsageCache };
 export { router };
 export { calculateTokenCount };
 export { searchProjectBySession };
+export type { RouterScenarioType, RouterFallbackConfig } from "./utils/router";
 export { ConfigService } from "./services/config";
 export { ProviderService } from "./services/provider";
 export { TransformerService } from "./services/transformer";
diff --git a/packages/core/src/utils/router.ts b/packages/core/src/utils/router.ts
index 1789e8a..f3d4fa0 100644
--- a/packages/core/src/utils/router.ts
+++ b/packages/core/src/utils/router.ts
@@ -126,7 +126,7 @@ const getUseModel = async (
   tokenCount: number,
   configService: ConfigService,
   lastUsage?: Usage | undefined
-) => {
+): Promise<{ model: string; scenarioType: RouterScenarioType }> => {
   const projectSpecificRouter = await getProjectSpecificRouter(req, configService);
   const providers = configService.get<any[]>("providers") || [];
   const Router = projectSpecificRouter || configService.get("Router");
@@ -140,9 +140,9 @@ const getUseModel = async (
       (m: any) => m.toLowerCase() === model
     );
     if (finalProvider && finalModel) {
-      return `${finalProvider.name},${finalModel}`;
+      return { model: `${finalProvider.name},${finalModel}`, scenarioType: 'default' };
     }
-    return req.body.model;
+    return { model: req.body.model, scenarioType: 'default' };
   }
 
   // if tokenCount is greater than the configured threshold, use the long context model
@@ -156,7 +156,7 @@ const getUseModel = async (
     req.log.info(
       `Using long context model due to token count: ${tokenCount}, threshold: ${longContextThreshold}`
     );
-    return Router.longContext;
+    return { model: Router.longContext, scenarioType: 'longContext' };
   }
   if (
     req.body?.system?.length > 1 &&
@@ -170,7 +170,7 @@ const getUseModel = async (
         `<CCR-SUBAGENT-MODEL>${model[1]}</CCR-SUBAGENT-MODEL>`,
         ""
       );
-      return model[1];
+      return { model: model[1], scenarioType: 'default' };
     }
   }
   // Use the background model for any Claude Haiku variant
@@ -181,7 +181,7 @@ const getUseModel = async (
     globalRouter?.background
   ) {
     req.log.info(`Using background model for ${req.body.model}`);
-    return globalRouter.background;
+    return { model: globalRouter.background, scenarioType: 'background' };
   }
   // The priority of websearch must be higher than thinking.
   if (
@@ -189,14 +189,14 @@ const getUseModel = async (
     req.body.tools.some((tool: any) => tool.type?.startsWith("web_search")) &&
     Router?.webSearch
   ) {
-    return Router.webSearch;
+    return { model: Router.webSearch, scenarioType: 'webSearch' };
   }
   // if exits thinking, use the think model
   if (req.body.thinking && Router?.think) {
     req.log.info(`Using think model for ${req.body.thinking}`);
-    return Router.think;
+    return { model: Router.think, scenarioType: 'think' };
   }
-  return Router?.default;
+  return { model: Router?.default, scenarioType: 'default' };
 };
 
 export interface RouterContext {
@@ -205,6 +205,16 @@ export interface RouterContext {
   event?: any;
 }
 
+export type RouterScenarioType = 'default' | 'background' | 'think' | 'longContext' | 'webSearch';
+
+export interface RouterFallbackConfig {
+  default?: string[];
+  background?: string[];
+  think?: string[];
+  longContext?: string[];
+  webSearch?: string[];
+}
+
 export const router = async (req: any, _res: any, context: RouterContext) => {
   const { configService, event } = context;
   // Parse sessionId from metadata.user_id
@@ -247,9 +257,6 @@ export const router = async (req: any, _res: any, context: RouterContext) => {
         tokenizerConfig
       );
       tokenCount = result.tokenCount;
-      req.log.debug(
-        `Token count: ${tokenCount} (tokenizer: ${result.tokenizerUsed}, cached: ${result.cached})`
-      );
     } else {
       // Legacy fallback
       tokenCount = calculateTokenCount(
@@ -273,13 +280,19 @@ export const router = async (req: any, _res: any, context: RouterContext) => {
       }
     }
     if (!model) {
-      model = await getUseModel(req, tokenCount, configService, lastMessageUsage);
+      const result = await getUseModel(req, tokenCount, configService, lastMessageUsage);
+      model = result.model;
+      req.scenarioType = result.scenarioType;
+    } else {
+      // Custom router doesn't provide scenario type, default to 'default'
+      req.scenarioType = 'default';
     }
     req.body.model = model;
   } catch (error: any) {
     req.log.error(`Error in router middleware: ${error.message}`);
     const Router = configService.get("Router");
     req.body.model = Router?.default;
+    req.scenarioType = 'default';
   }
   return;
 };