server.js

import express from 'express';
import { fal } from '@fal-ai/client';

// --- Key Management ---
const FAL_KEY_STRING = process.env.FAL_KEY;
const API_KEY = process.env.API_KEY; // 自定义 API Key 环境变量保持不变

if (!FAL_KEY_STRING) {
    console.error("Error: FAL_KEY environment variable is not set.");
    process.exit(1);
}

// 解析 FAL_KEY 字符串为数组，去除空白并过滤空值
const falKeys = FAL_KEY_STRING.split(',')
                             .map(key => key.trim())
                             .filter(key => key.length > 0);

if (falKeys.length === 0) {
    console.error("Error: FAL_KEY environment variable is set, but no valid keys were found after parsing.");
    process.exit(1);
}

console.log(`Loaded ${falKeys.length} Fal AI Keys.`);

let currentFalKeyIndex = 0; // Index of the *next* key to try for a *new* request

// --- End Key Management ---


if (!API_KEY) {
    console.error("Error: API_KEY environment variable is not set.");
    process.exit(1);
}

const app = express();
app.use(express.json({ limit: '50mb' }));
app.use(express.urlencoded({ extended: true, limit: '50mb' }));

const PORT = process.env.PORT || 3000;

// API Key 鉴权中间件 (保持不变)
const apiKeyAuth = (req, res, next) => {
    const authHeader = req.headers['authorization'];

    if (!authHeader) {
        console.warn('Unauthorized: No Authorization header provided');
        return res.status(401).json({ error: 'Unauthorized: No API Key provided' });
    }

    const authParts = authHeader.split(' ');
    if (authParts.length !== 2 || authParts[0].toLowerCase() !== 'bearer') {
        console.warn('Unauthorized: Invalid Authorization header format');
        return res.status(401).json({ error: 'Unauthorized: Invalid Authorization header format' });
    }

    const providedKey = authParts[1];
    if (providedKey !== API_KEY) {
        console.warn('Unauthorized: Invalid API Key');
        return res.status(401).json({ error: 'Unauthorized: Invalid API Key' });
    }

    next();
};

// 应用 API Key 鉴权中间件到所有 API 路由 (保持不变)
app.use(['/v1/models', '/v1/chat/completions'], apiKeyAuth);

// === 全局定义限制 === (保持不变)
const PROMPT_LIMIT = 4800;
const SYSTEM_PROMPT_LIMIT = 4800;
// === 限制定义结束 ===

// 定义 fal-ai/any-llm 支持的模型列表 (保持不变)
const FAL_SUPPORTED_MODELS = [
    "anthropic/claude-3.7-sonnet",
    "anthropic/claude-3.5-sonnet",
    "anthropic/claude-3-5-haiku",
    "anthropic/claude-3-haiku",
    "google/gemini-pro-1.5",
    "google/gemini-flash-1.5",
    "google/gemini-flash-1.5-8b",
    "google/gemini-2.0-flash-001",
    "meta-llama/llama-3.2-1b-instruct",
    "meta-llama/llama-3.2-3b-instruct",
    "meta-llama/llama-3.1-8b-instruct",
    "meta-llama/llama-3.1-70b-instruct",
    "openai/gpt-4o-mini",
    "openai/gpt-4o",
    "deepseek/deepseek-r1",
    "meta-llama/llama-4-maverick",
    "meta-llama/llama-4-scout"
];

// Helper function to get owner from model ID (保持不变)
const getOwner = (modelId) => {
    if (modelId && modelId.includes('/')) {
        return modelId.split('/')[0];
    }
    return 'fal-ai';
}

// GET /v1/models endpoint (保持不变)
app.get('/v1/models', (req, res) => {
    console.log("Received request for GET /v1/models");
    try {
        const modelsData = FAL_SUPPORTED_MODELS.map(modelId => ({
            id: modelId, object: "model", created: 1700000000, owned_by: getOwner(modelId)
        }));
        res.json({ object: "list", data: modelsData });
        console.log("Successfully returned model list.");
    } catch (error) {
        console.error("Error processing GET /v1/models:", error);
        res.status(500).json({ error: "Failed to retrieve model list." });
    }
});

// === convertMessagesToFalPrompt 函数 (保持不变) ===
function convertMessagesToFalPrompt(messages) {
    let fixed_system_prompt_content = "";
    const conversation_message_blocks = [];
    console.log(`Original messages count: ${messages.length}`);

    // 1. 分离 System 消息，格式化 User/Assistant 消息
    for (const message of messages) {
        let content = (message.content === null || message.content === undefined) ? "" : String(message.content);
        switch (message.role) {
            case 'system':
                fixed_system_prompt_content += `System: ${content}\n\n`;
                break;
            case 'user':
                conversation_message_blocks.push(`Human: ${content}\n\n`);
                break;
            case 'assistant':
                conversation_message_blocks.push(`Assistant: ${content}\n\n`);
                break;
            default:
                console.warn(`Unsupported role: ${message.role}`);
                continue;
        }
    }

    // 2. 截断合并后的 system 消息（如果超长）
    if (fixed_system_prompt_content.length > SYSTEM_PROMPT_LIMIT) {
        const originalLength = fixed_system_prompt_content.length;
        fixed_system_prompt_content = fixed_system_prompt_content.substring(0, SYSTEM_PROMPT_LIMIT);
        console.warn(`Combined system messages truncated from ${originalLength} to ${SYSTEM_PROMPT_LIMIT}`);
    }
    // 清理末尾可能多余的空白，以便后续判断和拼接
    fixed_system_prompt_content = fixed_system_prompt_content.trim();


    // 3. 计算 system_prompt 中留给对话历史的剩余空间
    let space_occupied_by_fixed_system = 0;
    if (fixed_system_prompt_content.length > 0) {
         space_occupied_by_fixed_system = fixed_system_prompt_content.length + 4; // 预留 \n\n...\n\n 的长度
    }
     const remaining_system_limit = Math.max(0, SYSTEM_PROMPT_LIMIT - space_occupied_by_fixed_system);
    console.log(`Trimmed fixed system prompt length: ${fixed_system_prompt_content.length}. Approx remaining system history limit: ${remaining_system_limit}`);


    // 4. 反向填充 User/Assistant 对话历史
    const prompt_history_blocks = [];
    const system_prompt_history_blocks = [];
    let current_prompt_length = 0;
    let current_system_history_length = 0;
    let promptFull = false;
    let systemHistoryFull = (remaining_system_limit <= 0);

    console.log(`Processing ${conversation_message_blocks.length} user/assistant messages for recency filling.`);
    for (let i = conversation_message_blocks.length - 1; i >= 0; i--) {
        const message_block = conversation_message_blocks[i];
        const block_length = message_block.length;

        if (promptFull && systemHistoryFull) {
            console.log(`Both prompt and system history slots full. Omitting older messages from index ${i}.`);
            break;
        }

        // 优先尝试放入 prompt
        if (!promptFull) {
            if (current_prompt_length + block_length <= PROMPT_LIMIT) {
                prompt_history_blocks.unshift(message_block);
                current_prompt_length += block_length;
                continue;
            } else {
                promptFull = true;
                console.log(`Prompt limit (${PROMPT_LIMIT}) reached. Trying system history slot.`);
            }
        }

        // 如果 prompt 满了，尝试放入 system_prompt 的剩余空间
        if (!systemHistoryFull) {
            if (current_system_history_length + block_length <= remaining_system_limit) {
                 system_prompt_history_blocks.unshift(message_block);
                 current_system_history_length += block_length;
                 continue;
            } else {
                 systemHistoryFull = true;
                 console.log(`System history limit (${remaining_system_limit}) reached.`);
            }
        }
    }

    // 5. *** 组合最终的 prompt 和 system_prompt (包含分隔符逻辑) ***
    const system_prompt_history_content = system_prompt_history_blocks.join('').trim();
    const final_prompt = prompt_history_blocks.join('').trim();

    // 定义分隔符
    const SEPARATOR = "\n\n-------下面是比较早之前的对话内容-----\n\n";

    let final_system_prompt = "";

    // 检查各部分是否有内容 (使用 trim 后的固定部分)
    const hasFixedSystem = fixed_system_prompt_content.length > 0;
    const hasSystemHistory = system_prompt_history_content.length > 0;

    if (hasFixedSystem && hasSystemHistory) {
        // 两部分都有，用分隔符连接
        final_system_prompt = fixed_system_prompt_content + SEPARATOR + system_prompt_history_content;
        console.log("Combining fixed system prompt and history with separator.");
    } else if (hasFixedSystem) {
        // 只有固定部分
        final_system_prompt = fixed_system_prompt_content;
        console.log("Using only fixed system prompt.");
    } else if (hasSystemHistory) {
        // 只有历史部分 (固定部分为空)
        final_system_prompt = system_prompt_history_content;
        console.log("Using only history in system prompt slot.");
    }
    // 如果两部分都为空，final_system_prompt 保持空字符串 ""

    // 6. 返回结果
    const result = {
        system_prompt: final_system_prompt, // 最终结果不需要再 trim
        prompt: final_prompt              // final_prompt 在组合前已 trim
    };

    console.log(`Final system_prompt length (Sys+Separator+Hist): ${result.system_prompt.length}`);
    console.log(`Final prompt length (Hist): ${result.prompt.length}`);

    return result;
}
// === convertMessagesToFalPrompt 函数结束 ===


// POST /v1/chat/completions endpoint (带 Key 重试逻辑)
app.post('/v1/chat/completions', async (req, res) => {
    const { model, messages, stream = false, reasoning = false, ...restOpenAIParams } = req.body;
    const requestId = `req-${Date.now()}`; // Unique ID for this incoming request

    console.log(`[${requestId}] Received chat completion request for model: ${model}, stream: ${stream}`);

    if (!FAL_SUPPORTED_MODELS.includes(model)) {
        console.warn(`[${requestId}] Warning: Requested model '${model}' is not in the explicitly supported list.`);
    }
    if (!model || !messages || !Array.isArray(messages) || messages.length === 0) {
        console.error(`[${requestId}] Invalid request parameters:`, { model, messages: Array.isArray(messages) ? messages.length : typeof messages });
        return res.status(400).json({ error: 'Missing or invalid parameters: model and messages array are required.' });
    }

    let lastError = null; // Store the last error encountered during key rotation
    let success = false; // Flag to indicate if any key succeeded

    // *** 重试循环：尝试最多 falKeys.length 次 ***
    for (let attempt = 0; attempt < falKeys.length; attempt++) {
        const keyIndexToTry = (currentFalKeyIndex + attempt) % falKeys.length;
        const selectedFalKey = falKeys[keyIndexToTry];
        console.log(`[${requestId}] Attempt ${attempt + 1}/${falKeys.length}: Trying Fal Key at index ${keyIndexToTry}`);

        try {
            // 配置 fal 客户端 (每次尝试都重新配置)
            fal.config({
                credentials: selectedFalKey,
            });

            // 准备 Fal Input (只需要准备一次)
            // 注意：如果 convertMessagesToFalPrompt 很耗时，可以移到循环外
            const { prompt, system_prompt } = convertMessagesToFalPrompt(messages);
            const falInput = {
                model: model,
                prompt: prompt,
                ...(system_prompt && { system_prompt: system_prompt }),
                reasoning: !!reasoning,
            };

            // 打印一次 Fal Input 和 Prompt 信息
            if (attempt === 0) {
                console.log(`[${requestId}] Fal Input:`, JSON.stringify(falInput, null, 2));
                console.log(`[${requestId}] Forwarding request to fal-ai with system-priority + separator + recency input:`);
                console.log(`[${requestId}] System Prompt Length:`, system_prompt?.length || 0);
                console.log(`[${requestId}] Prompt Length:`, prompt?.length || 0);
                // 为了简洁，默认注释掉详细内容的打印
                // console.log(`[${requestId}] --- System Prompt Start ---`);
                // console.log(system_prompt);
                // console.log(`[${requestId}] --- System Prompt End ---`);
                // console.log(`[${requestId}] --- Prompt Start ---`);
                // console.log(prompt);
                // console.log(`[${requestId}] --- Prompt End ---`);
            }

            // --- 执行 Fal AI 调用 ---
            if (stream) {
                // --- 流式处理 ---
                res.setHeader('Content-Type', 'text/event-stream; charset=utf-8');
                res.setHeader('Cache-Control', 'no-cache');
                res.setHeader('Connection', 'keep-alive');
                res.setHeader('Access-Control-Allow-Origin', '*');
                // 注意：Headers 只能发送一次

                let previousOutput = '';
                const falStream = await fal.stream("fal-ai/any-llm", { input: falInput });

                // 标记成功，设置下一次请求的起始 Key 索引
                success = true;
                currentFalKeyIndex = (keyIndexToTry + 1) % falKeys.length;
                console.log(`[${requestId}] Key at index ${keyIndexToTry} successful. Next request starts at index ${currentFalKeyIndex}.`);

                // 处理流
                try {
                     if (!res.headersSent) { // 确保 header 只发送一次
                         res.flushHeaders();
                         console.log(`[${requestId}] Stream headers flushed.`);
                     }
                    for await (const event of falStream) {
                        // ... (流处理逻辑基本不变，添加 requestId 用于日志) ...
                        const currentOutput = (event && typeof event.output === 'string') ? event.output : '';
                        const isPartial = (event && typeof event.partial === 'boolean') ? event.partial : true;
                        const errorInfo = (event && event.error) ? event.error : null;

                        if (errorInfo) {
                            console.error(`[${requestId}] Error received in fal stream event:`, errorInfo);
                            const errorChunk = { id: `chatcmpl-${Date.now()}-error`, object: "chat.completion.chunk", created: Math.floor(Date.now() / 1000), model: model, choices: [{ index: 0, delta: {}, finish_reason: "error", message: { role: 'assistant', content: `Fal Stream Error: ${JSON.stringify(errorInfo)}` } }] };
                            if (!res.writableEnded) {
                                res.write(`data: ${JSON.stringify(errorChunk)}\n\n`);
                            }
                            break; // Stop processing on error in stream event
                        }

                        let deltaContent = '';
                        if (currentOutput.startsWith(previousOutput)) {
                            deltaContent = currentOutput.substring(previousOutput.length);
                        } else if (currentOutput.length > 0) {
                            console.warn(`[${requestId}] Fal stream output mismatch detected. Sending full current output as delta.`, { previousLength: previousOutput.length, currentLength: currentOutput.length });
                            deltaContent = currentOutput;
                            previousOutput = ''; // Reset previous if mismatch
                        }
                        previousOutput = currentOutput;

                        // Send chunk if there's content or if it's the final chunk (isPartial is false)
                        if (deltaContent || !isPartial) {
                            const openAIChunk = { id: `chatcmpl-${Date.now()}`, object: "chat.completion.chunk", created: Math.floor(Date.now() / 1000), model: model, choices: [{ index: 0, delta: { content: deltaContent }, finish_reason: isPartial === false ? "stop" : null }] };
                             if (!res.writableEnded) {
                                res.write(`data: ${JSON.stringify(openAIChunk)}\n\n`);
                             }
                        }
                     }
                     // After the loop, ensure the [DONE] signal is sent if the stream finished normally
                     if (!res.writableEnded) {
                        res.write(`data: [DONE]\n\n`);
                        res.end();
                        console.log(`[${requestId}] Stream finished and [DONE] sent.`);
                     }
                } catch (streamError) {
                    console.error(`[${requestId}] Error during fal stream processing loop:`, streamError);
                    lastError = streamError; // Store error from stream processing
                    try {
                        // Don't mark success=false here, the key worked but the stream itself failed.
                        // The outer loop should break because the response has likely been ended.
                        if (!res.writableEnded) { // Check if we can still write to the response
                            const errorDetails = (streamError instanceof Error) ? streamError.message : JSON.stringify(streamError);
                            res.write(`data: ${JSON.stringify({ error: { message: "Stream processing error", type: "proxy_error", details: errorDetails } })}\n\n`);
                            res.write(`data: [DONE]\n\n`); // Send DONE even after error
                            res.end();
                        } else {
                             console.error(`[${requestId}] Stream already ended, cannot send error message.`);
                        }
                    } catch (finalError) {
                         console.error(`[${requestId}] Error sending stream error message to client:`, finalError);
                        if (!res.writableEnded) { res.end(); }
                    }
                     // Break the outer key retry loop as the stream failed mid-way
                     break;
                }
                // 如果流成功处理完，直接跳出重试循环
                break; // Exit the key retry loop

            } else {
                // --- 非流式处理 ---
                console.log(`[${requestId}] Executing non-stream request with key index ${keyIndexToTry}...`);
                const result = await fal.subscribe("fal-ai/any-llm", { input: falInput, logs: true });

                // 检查 Fal AI 返回的业务错误 (例如输入无效)，这种错误不应该因为换 Key 而解决
                if (result && result.error) {
                     console.error(`[${requestId}] Fal-ai returned a business error with key index ${keyIndexToTry}:`, result.error);
                     // 将此视为最终错误，不重试其他 key
                     lastError = new Error(`Fal-ai error: ${JSON.stringify(result.error)}`);
                     lastError.status = 500; // Or map from Fal error if possible, default 500
                     lastError.type = "fal_ai_error";
                     break; // Exit retry loop, no point trying other keys for bad input
                }

                console.log(`[${requestId}] Received non-stream result from fal-ai with key index ${keyIndexToTry}`);
                // console.log("Full non-stream result:", JSON.stringify(result, null, 2)); // Uncomment for detailed logs

                // 标记成功，设置下一次请求的起始 Key 索引
                success = true;
                currentFalKeyIndex = (keyIndexToTry + 1) % falKeys.length;
                console.log(`[${requestId}] Key at index ${keyIndexToTry} successful. Next request starts at index ${currentFalKeyIndex}.`);


                const openAIResponse = {
                    id: `chatcmpl-${result.requestId || Date.now()}`, object: "chat.completion", created: Math.floor(Date.now() / 1000), model: model,
                    choices: [{ index: 0, message: { role: "assistant", content: result.output || "" }, finish_reason: "stop" }],
                    usage: { prompt_tokens: null, completion_tokens: null, total_tokens: null }, // Fal doesn't provide token usage
                    system_fingerprint: null, // Fal doesn't provide system fingerprint
                    ...(result.reasoning && { fal_reasoning: result.reasoning }), // Include reasoning if present
                };
                res.json(openAIResponse);
                console.log(`[${requestId}] Returned non-stream response.`);
                break; // 成功，跳出重试循环
            }

        } catch (error) {
            lastError = error; // Store the error from this attempt
            const status = error?.status; // Fal client errors should have status
            const errorMessage = error?.body?.detail || error?.message || 'Unknown error'; // Get detailed message if possible

            console.warn(`[${requestId}] Attempt ${attempt + 1} with key index ${keyIndexToTry} failed. Status: ${status || 'N/A'}, Message: ${errorMessage}`);

            // 检查是否是与 Key 相关的错误 (401 Unauthorized, 403 Forbidden, 429 Rate Limit)
            if (status === 401 || status === 403 || status === 429) {
                console.log(`[${requestId}] Key-related error (${status}). Trying next key...`);
                // 继续循环尝试下一个 Key
            } else {
                 // 如果是其他类型的错误 (如网络问题、Fal内部服务器错误5xx、请求参数错误400等)
                 // 通常重试其他 Key 没有意义，直接中断重试
                 console.error(`[${requestId}] Unrecoverable error encountered. Status: ${status || 'N/A'}. Stopping key rotation for this request.`);
                 console.error("Error details:", error); // Log the full error object for debugging
                 break; // 跳出重试循环
            }
        }
    } // --- 结束重试循环 ---

    // 如果循环结束了还没有成功 (所有 Key 都失败了或遇到不可恢复错误)
    if (!success) {
        console.error(`[${requestId}] All Fal Key attempts failed or an unrecoverable error occurred.`);
        if (!res.headersSent) {
            const statusCode = lastError?.status || 503; // Use status from last error if available, default to 503 Service Unavailable
            const errorMessage = (lastError instanceof Error) ? lastError.message : JSON.stringify(lastError);
            // Try to extract a more specific message if available
            const detailMessage = lastError?.body?.detail || errorMessage;
            const errorType = lastError?.type || (statusCode === 401 || statusCode === 403 || statusCode === 429 ? "key_error" : "proxy_error");

             console.error(`[${requestId}] Sending final error response. Status: ${statusCode}, Type: ${errorType}, Message: ${detailMessage}`);

             // 返回一个标准的 OpenAI 错误格式
             res.status(statusCode).json({
                 object: "error",
                 message: `All Fal Key attempts failed or an unrecoverable error occurred. Last error: ${detailMessage}`,
                 type: errorType,
                 param: null,
                 code: statusCode === 429 ? "rate_limit_exceeded" : (statusCode === 401 || statusCode === 403 ? "invalid_api_key" : null)
             });
        } else if (!res.writableEnded) {
            console.error(`[${requestId}] Headers already sent, but request failed after stream started or during processing. Ending response with error chunk.`);
            // 尝试在流式响应中发送错误（如果可能）
            try {
                const errorDetails = (lastError instanceof Error) ? lastError.message : JSON.stringify(lastError);
                const detailMessage = lastError?.body?.detail || errorDetails;
                const errorChunk = { id: `chatcmpl-${Date.now()}-final-error`, object: "chat.completion.chunk", created: Math.floor(Date.now() / 1000), model: model, choices: [{ index: 0, delta: {}, finish_reason: "error", message: { role: 'assistant', content: `Proxy Error: All key attempts failed or stream processing error. Last error: ${detailMessage}` } }] };
                res.write(`data: ${JSON.stringify(errorChunk)}\n\n`);
                res.write(`data: [DONE]\n\n`);
                res.end();
            } catch (e) {
                console.error(`[${requestId}] Failed to write final error to stream:`, e);
                if (!res.writableEnded) res.end(); // Force end if possible
            }
        } else {
            console.error(`[${requestId}] Request failed, but response stream was already fully ended. Cannot send error.`);
        }
    }

});

// 启动服务器 (更新启动信息)
app.listen(PORT, () => {
    console.log(`===================================================`);
    console.log(` Fal OpenAI Proxy Server (Key Rotation with Retry + System Top + Separator + Recency)`); // 更新描述
    console.log(` Listening on port: ${PORT}`);
    console.log(` Loaded ${falKeys.length} Fal AI Keys for rotation.`);
    console.log(` Using Limits: System Prompt=${SYSTEM_PROMPT_LIMIT}, Prompt=${PROMPT_LIMIT}`);
    console.log(` API Key Auth Enabled: ${API_KEY ? 'Yes' : 'No'}`);
    console.log(` Chat Completions Endpoint: POST http://localhost:${PORT}/v1/chat/completions`);
    console.log(` Models Endpoint: GET http://localhost:${PORT}/v1/models`);
    console.log(`===================================================`);
});

// 根路径响应 (更新信息)
app.get('/', (req, res) => {
    res.send('Fal OpenAI Proxy (Key Rotation with Retry + System Top + Separator + Recency Strategy) is running.'); // 更新描述
});