feat: add dynamic x-api-provider and reasoning_effort support

- Add per-model provider configuration in config.json - Implement getModelProvider() to fetch provider from model config - Update all header generators to accept dynamic provider parameter - Add reasoning_effort field handling for common endpoint type - Support auto/low/medium/high/off reasoning levels for common models This enables flexible multi-provider support and reasoning control across different endpoint types (anthropic, openai, common).
2025-11-19 01:25:01 +08:00
parent c31b680d95
commit 62a384f34b
6 changed files with 70 additions and 22 deletions
--- a/config.js
+++ b/config.js
@@ -64,6 +64,11 @@ export function getModelReasoning(modelId) {
  return null;
 }

+export function getModelProvider(modelId) {
+  const model = getModelById(modelId);
+  return model?.provider || null;
+}
+
 export function getUserAgent() {
  return getCurrentUserAgent();
 }
--- a/config.json
+++ b/config.json
@@ -25,48 +25,64 @@
      "name": "Opus 4.1",
      "id": "claude-opus-4-1-20250805",
      "type": "anthropic",
-      "reasoning": "auto"
+      "reasoning": "auto",
+      "provider": "anthropic"
    },
    {
      "name": "Haiku 4.5",
      "id": "claude-haiku-4-5-20251001",
      "type": "anthropic",
-      "reasoning": "auto"
+      "reasoning": "auto",
+      "provider": "anthropic"
    },
    {
      "name": "Sonnet 4.5",
      "id": "claude-sonnet-4-5-20250929",
      "type": "anthropic",
-      "reasoning": "auto"
+      "reasoning": "auto",
+      "provider": "anthropic"
    },
    {
      "name": "GPT-5",
      "id": "gpt-5-2025-08-07",
      "type": "openai",
-      "reasoning": "auto"
+      "reasoning": "auto",
+      "provider": "openai"
    },
    {
      "name": "GPT-5-Codex",
      "id": "gpt-5-codex",
      "type": "openai",
-      "reasoning": "off"
+      "reasoning": "off",
+      "provider": "openai"
    },
    {
      "name": "GPT-5.1",
      "id": "gpt-5.1",
      "type": "openai",
-      "reasoning": "auto"
+      "reasoning": "auto",
+      "provider": "openai"
    },
    {
      "name": "GPT-5.1-Codex",
      "id": "gpt-5.1-codex",
      "type": "openai",
-      "reasoning": "off"
+      "reasoning": "off",
+      "provider": "openai"
    },
    {
      "name": "GLM-4.6",
      "id": "glm-4.6",
-      "type": "common"
+      "type": "common",
+      "reasoning": "off",
+      "provider": "fireworks"
+    },
+    {
+      "name": "Gemini-3-Pro",
+      "id": "gemini-3-pro-preview",
+      "type": "common",
+      "reasoning": "auto",
+      "provider": "google"
    }
  ],
  "dev_mode": false,
--- a/routes.js
+++ b/routes.js
@@ -1,6 +1,6 @@
 import express from 'express';
 import fetch from 'node-fetch';
-import { getConfig, getModelById, getEndpointByType, getSystemPrompt, getModelReasoning, getRedirectedModelId } from './config.js';
+import { getConfig, getModelById, getEndpointByType, getSystemPrompt, getModelReasoning, getRedirectedModelId, getModelProvider } from './config.js';
 import { logInfo, logDebug, logError, logRequest, logResponse } from './logger.js';
 import { transformToAnthropic, getAnthropicHeaders } from './transformers/request-anthropic.js';
 import { transformToOpenAI, getOpenAIHeaders } from './transformers/request-openai.js';
@@ -129,16 +129,19 @@ async function handleChatCompletions(req, res) {
    // Update request body with redirected model ID before transformation
    const requestWithRedirectedModel = { ...openaiRequest, model: modelId };

+    // Get provider from model config
+    const provider = getModelProvider(modelId);
+
    if (model.type === 'anthropic') {
      transformedRequest = transformToAnthropic(requestWithRedirectedModel);
      const isStreaming = openaiRequest.stream === true;
-      headers = getAnthropicHeaders(authHeader, clientHeaders, isStreaming, modelId);
+      headers = getAnthropicHeaders(authHeader, clientHeaders, isStreaming, modelId, provider);
    } else if (model.type === 'openai') {
      transformedRequest = transformToOpenAI(requestWithRedirectedModel);
-      headers = getOpenAIHeaders(authHeader, clientHeaders);
+      headers = getOpenAIHeaders(authHeader, clientHeaders, provider);
    } else if (model.type === 'common') {
      transformedRequest = transformToCommon(requestWithRedirectedModel);
-      headers = getCommonHeaders(authHeader, clientHeaders);
+      headers = getCommonHeaders(authHeader, clientHeaders, provider);
    } else {
      return res.status(500).json({ error: `Unknown endpoint type: ${model.type}` });
    }
@@ -285,8 +288,11 @@ async function handleDirectResponses(req, res) {

    const clientHeaders = req.headers;
    
+    // Get provider from model config
+    const provider = getModelProvider(modelId);
+    
    // 获取 headers
-    const headers = getOpenAIHeaders(authHeader, clientHeaders);
+    const headers = getOpenAIHeaders(authHeader, clientHeaders, provider);

    // 注入系统提示到 instructions 字段，并更新重定向后的模型ID
    const systemPrompt = getSystemPrompt();
@@ -427,9 +433,12 @@ async function handleDirectMessages(req, res) {

    const clientHeaders = req.headers;
    
+    // Get provider from model config
+    const provider = getModelProvider(modelId);
+    
    // 获取 headers
    const isStreaming = anthropicRequest.stream === true;
-    const headers = getAnthropicHeaders(authHeader, clientHeaders, isStreaming, modelId);
+    const headers = getAnthropicHeaders(authHeader, clientHeaders, isStreaming, modelId, provider);

    // 注入系统提示到 system 字段，并更新重定向后的模型ID
    const systemPrompt = getSystemPrompt();
@@ -576,7 +585,11 @@ async function handleCountTokens(req, res) {
    }

    const clientHeaders = req.headers;
-    const headers = getAnthropicHeaders(authHeader, clientHeaders, false, modelId);
+    
+    // Get provider from model config
+    const provider = getModelProvider(modelId);
+    
+    const headers = getAnthropicHeaders(authHeader, clientHeaders, false, modelId, provider);

    // 构建 count_tokens 端点 URL
    const countTokensUrl = endpoint.base_url.replace('/v1/messages', '/v1/messages/count_tokens');
--- a/transformers/request-anthropic.js
+++ b/transformers/request-anthropic.js
@@ -154,7 +154,7 @@ export function transformToAnthropic(openaiRequest) {
  return anthropicRequest;
 }

-export function getAnthropicHeaders(authHeader, clientHeaders = {}, isStreaming = true, modelId = null) {
+export function getAnthropicHeaders(authHeader, clientHeaders = {}, isStreaming = true, modelId = null, provider = 'anthropic') {
  // Generate unique IDs if not provided
  const sessionId = clientHeaders['x-session-id'] || generateUUID();
  const messageId = clientHeaders['x-assistant-message-id'] || generateUUID();
@@ -165,7 +165,7 @@ export function getAnthropicHeaders(authHeader, clientHeaders = {}, isStreaming
    'anthropic-version': clientHeaders['anthropic-version'] || '2023-06-01',
    'authorization': authHeader || '',
    'x-api-key': 'placeholder',
-    'x-api-provider': 'anthropic',
+    'x-api-provider': provider,
    'x-factory-client': 'cli',
    'x-session-id': sessionId,
    'x-assistant-message-id': messageId,
--- a/transformers/request-common.js
+++ b/transformers/request-common.js
@@ -1,5 +1,5 @@
 import { logDebug } from '../logger.js';
-import { getSystemPrompt, getUserAgent } from '../config.js';
+import { getSystemPrompt, getUserAgent, getModelReasoning } from '../config.js';

 export function transformToCommon(openaiRequest) {
  logDebug('Transforming OpenAI request to Common format');
@@ -39,11 +39,25 @@ export function transformToCommon(openaiRequest) {
    }
  }

+  // Handle reasoning_effort field based on model configuration
+  const reasoningLevel = getModelReasoning(openaiRequest.model);
+  if (reasoningLevel === 'auto') {
+    // Auto mode: preserve original request's reasoning_effort field exactly as-is
+    // If original request has reasoning_effort field, keep it; otherwise don't add one
+  } else if (reasoningLevel && ['low', 'medium', 'high'].includes(reasoningLevel)) {
+    // Specific level: override with model configuration
+    commonRequest.reasoning_effort = reasoningLevel;
+  } else {
+    // Off or invalid: explicitly remove reasoning_effort field
+    // This ensures any reasoning_effort field from the original request is deleted
+    delete commonRequest.reasoning_effort;
+  }
+
  logDebug('Transformed Common request', commonRequest);
  return commonRequest;
 }

-export function getCommonHeaders(authHeader, clientHeaders = {}) {
+export function getCommonHeaders(authHeader, clientHeaders = {}, provider = 'baseten') {
  // Generate unique IDs if not provided
  const sessionId = clientHeaders['x-session-id'] || generateUUID();
  const messageId = clientHeaders['x-assistant-message-id'] || generateUUID();
@@ -52,7 +66,7 @@ export function getCommonHeaders(authHeader, clientHeaders = {}) {
    'accept': 'application/json',
    'content-type': 'application/json',
    'authorization': authHeader || '',
-    'x-api-provider': 'baseten',
+    'x-api-provider': provider,
    'x-factory-client': 'cli',
    'x-session-id': sessionId,
    'x-assistant-message-id': messageId,
--- a/transformers/request-openai.js
+++ b/transformers/request-openai.js
@@ -133,7 +133,7 @@ export function transformToOpenAI(openaiRequest) {
  return targetRequest;
 }

-export function getOpenAIHeaders(authHeader, clientHeaders = {}) {
+export function getOpenAIHeaders(authHeader, clientHeaders = {}, provider = 'openai') {
  // Generate unique IDs if not provided
  const sessionId = clientHeaders['x-session-id'] || generateUUID();
  const messageId = clientHeaders['x-assistant-message-id'] || generateUUID();
@@ -141,7 +141,7 @@ export function getOpenAIHeaders(authHeader, clientHeaders = {}) {
  const headers = {
    'content-type': 'application/json',
    'authorization': authHeader || '',
-    'x-api-provider': 'openai',
+    'x-api-provider': provider,
    'x-factory-client': 'cli',
    'x-session-id': sessionId,
    'x-assistant-message-id': messageId,