mirror of
https://github.com/musistudio/claude-code-router.git
synced 2026-01-29 22:02:05 +00:00
fix docs
This commit is contained in:
musistudio
@@ -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
|
||||
|
||||
@@ -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.
|
||||
|
||||
|
||||
@@ -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
|
||||
|
||||
@@ -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`:
|
||||
|
||||
@@ -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
|
||||
|
||||
@@ -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
|
||||
|
||||
@@ -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: {
|
||||
|
||||
@@ -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'"
|
||||
}
|
||||
}
|
||||
|
||||
@@ -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：
|
||||
|
||||
@@ -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
|
||||
|
||||
@@ -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
|
||||
|
||||
@@ -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": {
|
||||
|
||||
@@ -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',
|
||||
],
|
||||
},
|
||||
],
|
||||
},
|
||||
],
|
||||
};
|
||||
|
||||
|
||||
@@ -44,7 +44,7 @@
|
||||
}
|
||||
},
|
||||
"engines": {
|
||||
"node": ">=18.0.0",
|
||||
"node": ">=20.0.0",
|
||||
"pnpm": ">=8.0.0"
|
||||
}
|
||||
}
|
||||
|
||||
@@ -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 ${scenarioType}`);
|
||||
return null;
|
||||
}
|
||||
|
||||
/**
|
||||
|
||||
@@ -39,6 +39,7 @@ declare module "fastify" {
|
||||
interface FastifyRequest {
|
||||
provider?: string;
|
||||
model?: string;
|
||||
scenarioType?: string;
|
||||
}
|
||||
interface FastifyInstance {
|
||||
_server?: Server;
|
||||
@@ -266,6 +267,7 @@ export { sessionUsageCache };
|
||||
export { router };
|
||||
export { calculateTokenCount };
|
||||
export { searchProjectBySession };
|
||||
export type { RouterScenarioType, RouterFallbackConfig } from "./utils/router";
|
||||
export { ConfigService } from "./services/config";
|
||||
export { ProviderService } from "./services/provider";
|
||||
export { TransformerService } from "./services/transformer";
|
||||
|
||||
@@ -126,7 +126,7 @@ const getUseModel = async (
|
||||
tokenCount: number,
|
||||
configService: ConfigService,
|
||||
lastUsage?: Usage | undefined
|
||||
) => {
|
||||
): Promise<{ model: string; scenarioType: RouterScenarioType }> => {
|
||||
const projectSpecificRouter = await getProjectSpecificRouter(req, configService);
|
||||
const providers = configService.get<any[]>("providers") || [];
|
||||
const Router = projectSpecificRouter || configService.get("Router");
|
||||
@@ -140,9 +140,9 @@ const getUseModel = async (
|
||||
(m: any) => m.toLowerCase() === model
|
||||
);
|
||||
if (finalProvider && finalModel) {
|
||||
return `${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;
|
||||
};
|
||||
|
||||
Reference in New Issue
Block a user