message-convert.js

/**
 * message-convert.js - 消息格式转换
 *
 * 将 OpenAI / Anthropic 的 messages 数组转换为
 * chataibot.pro 接受的单一 text 字符串
 *
 * chataibot.pro 限制单条消息 2500 字符。
 * 超长文本处理策略:
 *   1. 优先压缩: 截断 tool results → 删除中间轮次 → 压缩 system
 *   2. 仍超限 → 拆分为多条消息 (context prefill + final)
 */

import config from './config.js';
import { serializeTools, serializeToolsAnthropic } from './tool-prompt.js';

/**
 * chataibot.pro 单条消息字符上限
 * 实际限制 2500，留 100 余量
 */
const MAX_TEXT_LEN = 2400;

/**
 * 从 content 字段提取纯文本
 * OpenAI/Anthropic 的 content 可能是字符串或多模态数组
 */
function extractText(content) {
  if (typeof content === 'string') return content;
  if (Array.isArray(content)) {
    return content
      .filter(part => part.type === 'text')
      .map(part => part.text)
      .join('\n');
  }
  return String(content || '');
}

/**
 * 截断单个 [Tool Result: ...] 块的内容到指定字符数
 * 保留开头 + 末尾各一半，中间用省略标记
 */
function truncateToolResults(text, maxLen = 600) {
  return text.replace(/\[Tool Result: [^\]]*\]\n([\s\S]*?)(?=\n\[(?:User|Assistant|Tool Result|System)\]|\n```tool_calls|$)/g,
    (match, content) => {
      if (match.length <= maxLen) return match;
      const header = match.substring(0, match.indexOf('\n') + 1);
      const keep = maxLen - header.length - 30;
      if (keep <= 0) return header + '(content truncated)';
      const half = Math.floor(keep / 2);
      return header + content.substring(0, half) + '\n...(truncated)...\n' + content.substring(content.length - half);
    });
}

/**
 * 解析文本为 [System]/[User]/[Assistant] 段落
 */
function parseSegments(text) {
  const segments = [];
  const segRegex = /\[(System|User|Assistant)\] /g;
  let m;
  while ((m = segRegex.exec(text)) !== null) {
    if (segments.length > 0) {
      segments[segments.length - 1].text = text.substring(segments[segments.length - 1].start, m.index).trimEnd();
    }
    segments.push({ role: m[1], start: m.index, text: '' });
  }
  if (segments.length > 0) {
    segments[segments.length - 1].text = text.substring(segments[segments.length - 1].start).trimEnd();
  }
  return segments;
}

/**
 * 智能压缩到单条 ≤ MAX_TEXT_LEN
 *
 * 策略 (逐级加压):
 *   1. 截断 tool results 到 600/300/100/40 字符
 *   2. tool results 替换为一句话摘要
 *   3. 删除中间对话轮次 (仅保留 system + 最后 user/assistant)
 *   4. 截断 system prompt
 *   5. 硬截断
 *
 * @returns {string | null} 压缩后的文本，null 表示无法压到单条
 */
function compressToSingle(text) {
  if (text.length <= MAX_TEXT_LEN) return text;

  const originalLen = text.length;

  // Step 1: 逐步截断 tool result 内容
  let result = text;
  for (const maxToolLen of [600, 300, 100, 40]) {
    result = truncateToolResults(result, maxToolLen);
    if (result.length <= MAX_TEXT_LEN) {
      console.log(`[Convert] 压缩 ${originalLen} → ${result.length} (截断 tool results)`);
      return result;
    }
  }

  // Step 2: 所有 tool result 替换为一行标记
  result = result.replace(/\[Tool Result: [^\]]*\]\n[\s\S]*?(?=\n\[(?:User|Assistant|System)\]|\n```tool_calls|$)/g,
    (match) => {
      const nl = match.indexOf('\n');
      return match.substring(0, nl > 0 ? nl : match.length) + '\n(omitted)';
    });
  if (result.length <= MAX_TEXT_LEN) {
    console.log(`[Convert] 压缩 ${originalLen} → ${result.length} (清除 tool results)`);
    return result;
  }

  // Step 3: 拆分段落，保留 system + 最后 user (+ 最后 assistant 如果有)
  const segments = parseSegments(result);
  if (segments.length <= 2) {
    // 段落太少，跳到 step 4
  } else {
    const systemSeg = segments[0]?.role === 'System' ? segments[0] : null;
    let lastUserIdx = -1;
    for (let i = segments.length - 1; i >= 0; i--) {
      if (segments[i].role === 'User') { lastUserIdx = i; break; }
    }

    // 保留 system + 最后 user + 最后一条(如果不同)
    const mustKeep = new Set();
    if (systemSeg) mustKeep.add(0);
    if (lastUserIdx >= 0) mustKeep.add(lastUserIdx);
    mustKeep.add(segments.length - 1);

    const kept = segments.filter((_, i) => mustKeep.has(i));
    const rebuilt = kept.map(s => s.text).join('\n\n');
    if (rebuilt.length <= MAX_TEXT_LEN) {
      console.log(`[Convert] 压缩 ${originalLen} → ${rebuilt.length} (删除中间轮次)`);
      return rebuilt;
    }

    // Step 4: 截断 system prompt，给 user 消息留空间
    if (systemSeg) {
      const lastMsg = lastUserIdx >= 0 ? segments[lastUserIdx].text : segments[segments.length - 1].text;
      const budgetForSystem = MAX_TEXT_LEN - lastMsg.length - 20;
      if (budgetForSystem > 200) {
        const truncated = systemSeg.text.substring(0, budgetForSystem) + '\n\n' + lastMsg;
        console.log(`[Convert] 压缩 ${originalLen} → ${truncated.length} (截断 system)`);
        return truncated;
      }
      // 仅保留 user 消息
      if (lastMsg.length <= MAX_TEXT_LEN) {
        console.log(`[Convert] 压缩 ${originalLen} → ${lastMsg.length} (仅 user 消息)`);
        return lastMsg;
      }
    }
  }

  // Step 5: 硬截断 (保留末尾，因为最后的内容最重要)
  const truncated = text.substring(text.length - MAX_TEXT_LEN);
  console.log(`[Convert] 压缩 ${originalLen} → ${MAX_TEXT_LEN} (硬截断末尾)`);
  return truncated;
}

/**
 * 主入口: 将文本处理为 chataibot 可发送的消息数组
 *
 * 优先压缩到单条 (零额外开销);
 * 极端情况下才拆分多条 (消耗额外额度)。
 *
 * @param {string} text - 完整转换后的文本
 * @returns {string[]} 消息数组，长度 ≥ 1，每条 ≤ MAX_TEXT_LEN
 */
export function splitToChunks(text) {
  if (!text) return [''];

  // 尝试压缩到单条 (绝大多数场景走这里)
  const single = compressToSingle(text);
  if (single && single.length <= MAX_TEXT_LEN) return [single];

  // 极端情况: 压缩后仍超限 → 多消息拆分
  const originalLen = text.length;

  // 先做 tool result 截断到 300 字符 (减少总量但保留上下文)
  const compressed = truncateToolResults(text, 300);

  const segments = parseSegments(compressed);
  if (segments.length === 0) return hardSplit(compressed);

  // 贪心打包段落到 chunks
  const chunks = [];
  let currentChunk = '';

  for (const seg of segments) {
    if (seg.text.length > MAX_TEXT_LEN) {
      if (currentChunk) { chunks.push(currentChunk.trimEnd()); currentChunk = ''; }
      chunks.push(...hardSplit(seg.text));
      continue;
    }
    const candidate = currentChunk ? currentChunk + '\n\n' + seg.text : seg.text;
    if (candidate.length <= MAX_TEXT_LEN) {
      currentChunk = candidate;
    } else {
      if (currentChunk) chunks.push(currentChunk.trimEnd());
      currentChunk = seg.text;
    }
  }
  if (currentChunk) chunks.push(currentChunk.trimEnd());

  console.log(`[Convert] 多消息拆分 ${originalLen} → ${chunks.length} 条 [${chunks.map(c => c.length).join(', ')}]`);
  return chunks;
}

/**
 * 按字符硬拆，尽量在换行处断开
 */
function hardSplit(text) {
  const chunks = [];
  let pos = 0;
  while (pos < text.length) {
    if (text.length - pos <= MAX_TEXT_LEN) {
      chunks.push(text.substring(pos));
      break;
    }
    let cut = pos + MAX_TEXT_LEN;
    const lastNL = text.lastIndexOf('\n', cut);
    if (lastNL > pos + 500) cut = lastNL;
    chunks.push(text.substring(pos, cut));
    pos = cut;
    while (pos < text.length && text[pos] === '\n') pos++;
  }
  return chunks;
}

/**
 * OpenAI messages → chataibot text
 * @param {Array} messages - [{ role: 'system'|'user'|'assistant'|'tool', content }]
 * @param {Array} [tools] - OpenAI tools 数组
 * @param {*} [toolChoice] - tool_choice 参数
 */
export function openaiToText(messages, tools, toolChoice) {
  if (!messages || !messages.length) return '';

  const system = messages.filter(m => m.role === 'system').map(m => extractText(m.content)).join('\n');
  const conversation = messages.filter(m => m.role !== 'system');

  // 工具定义注入到 system prompt 末尾
  const toolPrompt = serializeTools(tools, toolChoice);

  // 处理 tool 角色消息 (工具返回结果) — 转为文本格式
  const processedConversation = [];
  for (const msg of conversation) {
    if (msg.role === 'tool') {
      // 工具返回结果，格式化为文本
      const toolName = msg.name || msg.tool_call_id || 'unknown';
      processedConversation.push({
        role: 'user',
        content: `[Tool Result: ${toolName}]\n${extractText(msg.content)}`,
      });
    } else if (msg.role === 'assistant' && msg.tool_calls) {
      // assistant 发起的工具调用 — 转为文本表示
      let callText = extractText(msg.content) || '';
      for (const tc of msg.tool_calls) {
        const fn = tc.function || {};
        callText += `\n\`\`\`tool_calls\n[{"name": "${fn.name}", "arguments": ${fn.arguments || '{}'}}]\n\`\`\``;
      }
      processedConversation.push({ role: 'assistant', content: callText.trim() });
    } else {
      processedConversation.push(msg);
    }
  }

  const fullSystem = system + toolPrompt;

  // 单轮: 只有一条 user 消息
  if (processedConversation.length === 1 && processedConversation[0].role === 'user') {
    const userText = extractText(processedConversation[0].content);
    return fullSystem ? `${fullSystem}\n\n${userText}` : userText;
  }

  // 多轮: 格式化为带角色标签的文本
  let text = '';
  if (fullSystem) text += `[System] ${fullSystem}\n\n`;

  for (const msg of processedConversation) {
    const role = msg.role === 'assistant' ? 'Assistant' : 'User';
    text += `[${role}] ${extractText(msg.content)}\n\n`;
  }

  return text.trim();
}

/**
 * Anthropic messages → chataibot text
 * @param {string|undefined} system - system prompt (Anthropic 单独字段)
 * @param {Array} messages - [{ role: 'user'|'assistant', content }]
 * @param {Array} [tools] - Anthropic tools 数组
 * @param {*} [toolChoice] - tool_choice 参数
 */
export function anthropicToText(system, messages, tools, toolChoice) {
  if (!messages || !messages.length) return extractText(system) || '';

  // 工具定义注入到 system prompt 末尾
  const toolPrompt = serializeToolsAnthropic(tools, toolChoice);
  const systemText = extractText(system) || '';
  const fullSystem = systemText + toolPrompt;

  // 处理 Anthropic content 数组中的 tool_use 和 tool_result
  const processedMessages = [];
  for (const msg of messages) {
    if (Array.isArray(msg.content)) {
      // Anthropic content 可能包含 tool_use / tool_result blocks
      const parts = [];
      for (const block of msg.content) {
        if (block.type === 'text') {
          parts.push(block.text);
        } else if (block.type === 'tool_use') {
          parts.push(`\`\`\`tool_calls\n[{"name": "${block.name}", "arguments": ${JSON.stringify(block.input || {})}}]\n\`\`\``);
        } else if (block.type === 'tool_result') {
          const resultContent = typeof block.content === 'string'
            ? block.content
            : Array.isArray(block.content)
              ? block.content.filter(c => c.type === 'text').map(c => c.text).join('\n')
              : JSON.stringify(block.content);
          parts.push(`[Tool Result: ${block.tool_use_id || 'unknown'}]\n${resultContent}`);
        }
      }
      processedMessages.push({ role: msg.role, content: parts.join('\n') });
    } else {
      processedMessages.push(msg);
    }
  }

  // 单轮
  if (processedMessages.length === 1 && processedMessages[0].role === 'user') {
    const userText = typeof processedMessages[0].content === 'string'
      ? processedMessages[0].content
      : extractText(processedMessages[0].content);
    return fullSystem ? `${fullSystem}\n\n${userText}` : userText;
  }

  // 多轮
  let text = '';
  if (fullSystem) text += `[System] ${fullSystem}\n\n`;

  for (const msg of processedMessages) {
    const role = msg.role === 'assistant' ? 'Assistant' : 'User';
    const content = typeof msg.content === 'string' ? msg.content : extractText(msg.content);
    text += `[${role}] ${content}\n\n`;
  }

  return text.trim();
}

/**
 * 解析并映射模型名
 */
export function resolveModel(requestModel) {
  if (!requestModel) return 'gpt-4o';
  // 精确匹配
  if (config.modelMapping[requestModel]) return config.modelMapping[requestModel];
  // 尝试去掉版本后缀匹配 (如 claude-3-sonnet-20240229 → claude-3-sonnet)
  const base = requestModel.replace(/-\d{8}$/, '');
  if (config.modelMapping[base]) return config.modelMapping[base];
  // 原样返回让 chataibot 自己判断
  return requestModel;
}

/**
 * 根据模型名判断所属厂商 (与官网分组一致)
 */
function getOwner(model) {
  if (model.startsWith('gpt-') || model.startsWith('o1') || model.startsWith('o3') || model.startsWith('o4')) return 'OpenAI';
  if (model.startsWith('claude-')) return 'Anthropic';
  if (model.startsWith('gemini-')) return 'Google';
  return '其他';
}

/**
 * 获取可用模型列表 (去重，只列 chataibot 原生模型名)
 */
export function getModelList() {
  const seen = new Set();
  const list = [];
  for (const target of Object.values(config.modelMapping)) {
    if (seen.has(target)) continue;
    seen.add(target);
    list.push({
      id: target,
      object: 'model',
      created: 1700000000,
      owned_by: getOwner(target),
    });
  }
  return list;
}