skill-mv/lib/api.mjs

// lib/api.mjs — 网关 HTTP 封装（chat / images / 异步 task 轮询 / 文件下载）
// 零依赖，仅用 Node 内置 fetch（Node >= 18）。

import { loadConfig, requireToken, mask } from './config.mjs';

/** 通道未开通/上游不可用时抛出，携带可操作指引 */
export class ProviderUnavailableError extends Error {
  constructor(message, { endpoint, status, hint, cause } = {}) {
    super(message);
    this.name = 'ProviderUnavailableError';
    this.endpoint = endpoint;
    this.status = status;
    this.hint = hint;
    if (cause) this.cause = cause;
  }
}

/** 网关返回了业务错误（参数错/额度等） */
export class GatewayError extends Error {
  constructor(message, { endpoint, status, body } = {}) {
    super(message);
    this.name = 'GatewayError';
    this.endpoint = endpoint;
    this.status = status;
    this.body = body;
  }
}

const DEFAULT_TIMEOUT = 120_000;

function joinUrl(base, path) {
  return `${String(base).replace(/\/+$/, '')}/${String(path).replace(/^\/+/, '')}`;
}

/**
 * 底层请求。自动带 token、超时、把网关错误翻译成可读异常。
 * @returns {Promise<{status:number, headers:Headers, body:any, text:string}>}
 */
export async function request(path, opts = {}) {
  const {
    method = 'GET',
    body,
    timeout = DEFAULT_TIMEOUT,
    base, // 覆盖 base（默认 config.baseUrl）
    headers = {},
    expectJson = true,
    raw = false,
    retries = 0,
    retryDelay = 1500,
  } = opts;

  const cfg = loadConfig();
  const token = requireToken();
  const url = /^https?:\/\//.test(path) ? path : joinUrl(base || cfg.baseUrl, path);

  let lastErr;
  for (let attempt = 0; attempt <= retries; attempt++) {
    if (attempt > 0) {
      await sleep(retryDelay * 2 ** (attempt - 1));
    }
    try {
      const res = await fetch(url, {
        method,
        headers: {
          Authorization: `Bearer ${token}`,
          Accept: 'application/json',
          ...(body !== undefined ? { 'Content-Type': 'application/json' } : {}),
          ...headers,
        },
        body: body !== undefined ? (typeof body === 'string' ? body : JSON.stringify(body)) : undefined,
        signal: AbortSignal.timeout(timeout),
      });

      const text = raw ? '' : await res.text();
      let parsed = null;
      if (expectJson && text) {
        try {
          parsed = JSON.parse(text);
        } catch {
          parsed = null;
        }
      }

      if (!res.ok) {
        const err = classifyHttpError(res.status, text, url, parsed);
        // 5xx / 429 可重试
        if ((res.status >= 500 || res.status === 429) && attempt < retries) {
          lastErr = err;
          continue;
        }
        throw err;
      }
      return { status: res.status, headers: res.headers, body: parsed, text };
    } catch (e) {
      if (e instanceof ProviderUnavailableError || e instanceof GatewayError) throw e;
      lastErr = new GatewayError(`${e.name === 'TimeoutError' ? '请求超时' : '网络错误'}: ${e.message}`, {
        endpoint: url,
      });
      if (attempt >= retries) throw lastErr;
    }
  }
  throw lastErr;
}

/** 把 HTTP 状态码 / 网关错误体翻译成「能照着做」的错误 */
function classifyHttpError(status, text, url, parsed) {
  const raw = (parsed && (parsed.error?.message || parsed.message)) || text.slice(0, 300);
  const lower = String(raw).toLowerCase();

  // 渠道未开通的典型特征
  const unavailable =
    lower.includes('no available channel') ||
    lower.includes('model_not_found') ||
    lower.includes('channel') && lower.includes('not') ||
    lower.includes('task request failed') ||
    lower.includes('upstream') && (status === 404 || status === 502 || status === 503) ||
    lower.includes('not implemented');

  const hint =
    status === 401 || status === 403
      ? '凭据无效或权限不足：检查 FEME_API_TOKEN / /opt/data/fmode-identity.json 的 newapi_token'
      : unavailable
        ? '该模型/通道在当前网关未开通。可用 `fmode-mv providers` 查看替代通道，或在 provider 配置里改换模型名。'
        : status === 429
          ? '触发限流，稍后重试或降低并发'
          : undefined;

  const message = `网关 ${status} @ ${url}\n  ${raw}` + (hint ? `\n  → ${hint}` : '');

  if (unavailable) {
    return new ProviderUnavailableError(message, { endpoint: url, status, hint });
  }
  return new GatewayError(message, { endpoint: url, status, body: raw });
}

export function sleep(ms) {
  return new Promise((r) => setTimeout(r, ms));
}

/** 下载远程文件为 Buffer */
export async function download(url, { timeout = 180_000 } = {}) {
  const res = await fetch(url, { signal: AbortSignal.timeout(timeout) });
  if (!res.ok) throw new GatewayError(`下载失败 ${res.status}: ${url}`);
  return Buffer.from(await res.arrayBuffer());
}

// ==================== Chat ====================

/**
 * 文本对话（作曲填词 / 分镜 / 口播稿都走这里）。
 * @param {{model?:string, messages:Array, system?:string, temperature?:number, max_tokens?:number, json?:boolean, timeout?:number}} o
 */
export async function chat(o = {}) {
  const {
    model = process.env.FEME_CHAT_MODEL || 'deepseek-v4-flash',
    messages = [],
    system,
    temperature = 0.8,
    max_tokens = 4096,
    json = false,
    timeout = 180_000,
    retries = 1,
  } = o;

  const msgs = system ? [{ role: 'system', content: system }, ...messages] : messages;

  const body = {
    model,
    messages: msgs,
    temperature,
    max_tokens,
    stream: false,
  };
  if (json) body.response_format = { type: 'json_object' };

  const res = await request('/chat/completions', {
    method: 'POST',
    body,
    timeout,
    retries,
  });

  const choice = res.body?.choices?.[0];
  // 部分推理模型（如 deepseek-v4）会把内容放在 reasoning_content，content 可能为空
  let content = choice?.message?.content;
  if ((content == null || content === '') && choice?.message?.reasoning_content) {
    if (json) {
      // 尝试从 reasoning 里抠出 JSON
      const m = String(choice.message.reasoning_content).match(/\{[\s\S]*\}/);
      content = m ? m[0] : choice.message.reasoning_content;
    } else {
      content = choice.message.reasoning_content;
    }
  }
  if (content == null) {
    throw new GatewayError(`chat 返回为空（model=${model}）`, { body: res.body });
  }
  return {
    content: String(content),
    model: res.body?.model || model,
    usage: res.body?.usage,
    finishReason: choice?.finish_reason || null,
    truncated: choice?.finish_reason === 'length',
  };
}

/**
 * 让模型产出 JSON 并解析。
 *
 * 现实里有两类坑，这里都自动处理：
 *   1) 推理型模型（如 deepseek-v4-flash）会把大量 token 花在推理上，
 *      导致 max_tokens 被吃满、JSON 被截断 —— finish_reason === 'length'。
 *      → 自动把 max_tokens 翻倍重试，最多升到上限。
 *   2) 输出带 markdown 围栏 / 前后废话 / 轻微格式错误。
 *      → extractJSON 已能剥离围栏；仍失败则带着错误提示让模型重出一次。
 */
export async function chatJSON(o = {}) {
  const {
    model = process.env.FEME_CHAT_MODEL || 'deepseek-v4-flash',
    messages = [],
    system,
    temperature = 0.7,
    max_tokens = 4096,
    maxTokensCap = 32768,
    escalationSteps = 2,
    repairAttempts = 1,
    timeout = 300_000,
  } = o;

  const msgs = system ? [{ role: 'system', content: system }, ...messages] : [...messages];
  let budget = max_tokens;
  let lastText = '';
  let lastErr = null;

  for (let step = 0; step <= escalationSteps; step++) {
    const r = await chat({ model, messages: msgs, temperature, max_tokens: budget, json: true, timeout });
    lastText = r.content;
    const parsed = extractJSON(r.content);
    if (parsed !== null) return { ...r, json: parsed, maxTokensUsed: budget };

    // 截断 → 加大预算重来
    if (r.truncated && budget < maxTokensCap) {
      budget = Math.min(maxTokensCap, budget * 2);
      continue;
    }
    lastErr = r;
    break;
  }

  // 修复式重试：把失败输出回喂，要求只吐合法 JSON
  for (let i = 0; i < repairAttempts; i++) {
    const tail = lastText.slice(-1200);
    const r = await chat({
      model,
      messages: [
        ...msgs,
        { role: 'assistant', content: tail },
        {
          role: 'user',
          content:
            '上面的 JSON 不完整或格式非法，无法解析。请重新输出**完整且合法**的 JSON：' +
            '不要省略任何数组元素、不要写 markdown 围栏、不要在 JSON 前后添加任何解释文字。' +
            '如果内容太长，请压缩每个字段的文字长度，但必须保证 JSON 结构完整闭合。',
        },
      ],
      temperature: 0.3,
      max_tokens: maxTokensCap,
      json: true,
      timeout,
    });
    lastText = r.content;
    const parsed = extractJSON(r.content);
    if (parsed !== null) return { ...r, json: parsed, maxTokensUsed: maxTokensCap, repaired: true };
  }

  const hint =
    lastErr?.truncated || lastText.length > 3000
      ? '\n  → 疑似输出被 max_tokens 截断。可提高 --max-tokens，或减少镜头数（--shots），或换用非推理型模型（FEME_CHAT_MODEL=doubao-seed-2-1-pro）。'
      : '';
  throw new GatewayError(
    `模型未返回可解析的 JSON（model=${model}，已尝试升配与修复）。输出尾部片段：\n${lastText.slice(-400)}${hint}`
  );
}

/** 从可能带 ```json 围栏或前后缀说明的文本中抠出 JSON */
export function extractJSON(text) {
  if (!text) return null;
  const s = String(text).trim();
  const tryParse = (t) => {
    try {
      return JSON.parse(t);
    } catch {
      return null;
    }
  };
  let v = tryParse(s);
  if (v !== null) return v;

  const fence = s.match(/```(?:json)?\s*([\s\S]*?)```/i);
  if (fence) {
    v = tryParse(fence[1].trim());
    if (v !== null) return v;
  }
  const obj = s.match(/[{[][\s\S]*[}\]]/);
  if (obj) {
    v = tryParse(obj[0]);
    if (v !== null) return v;
  }
  return null;
}

// ==================== Images ====================

/**
 * 文生图 / 图生图（关键帧）。
 * @param {{prompt:string, model?:string, size?:string, n?:number, refImage?:string|Buffer, timeout?:number}} o
 * @returns {Promise<{buffers:Buffer[], model:string, size:string}>}
 */
export async function generateImage(o = {}) {
  const {
    prompt,
    model = process.env.FEME_IMAGE_MODEL || 'gpt-image-2.5-sunburst',
    size = '1024x1024',
    n = 1,
    refImage, // 参考图（URL 或本地路径/base64）—— 用于人物一致性
    timeout = 300_000,
    retries = 1,
  } = o;
  if (!prompt) throw new GatewayError('generateImage 需要 prompt');

  const payload = { model, prompt, n, size };
  if (refImage) payload.image = await toImageInput(refImage); // 兼容 OpenAI 风格图生图

  const res = await request('/images/generations', {
    method: 'POST',
    body: payload,
    timeout,
    retries,
  });

  const items = res.body?.data || [];
  const buffers = [];
  for (const it of items) {
    if (it.b64_json) buffers.push(Buffer.from(it.b64_json, 'base64'));
    else if (it.url) buffers.push(await download(it.url));
  }
  if (!buffers.length) {
    throw new GatewayError(`图片通道未返回图像数据（model=${model}）`, { body: JSON.stringify(res.body).slice(0, 300) });
  }
  return { buffers, model: res.body?.model || model, size };
}

async function toImageInput(ref) {
  if (Buffer.isBuffer(ref)) return `data:image/png;base64,${ref.toString('base64')}`;
  const s = String(ref);
  if (/^https?:\/\//.test(s) || s.startsWith('data:')) return s;
  const { readFileSync } = await import('node:fs');
  const buf = readFileSync(s);
  const ext = s.toLowerCase().endsWith('.jpg') || s.toLowerCase().endsWith('.jpeg') ? 'jpeg' : 'png';
  return `data:image/${ext};base64,${buf.toString('base64')}`;
}

// ==================== 异步任务（音乐 / 视频 / 唇形） ====================

/**
 * 提交一个异步任务并轮询至完成。
 * 不同通道的字段名差异用 pick 函数吸收。
 *
 * @param {object} o
 * @param {string} o.submitPath  提交路径
 * @param {string} o.fetchPath   查询路径（返回单任务）
 * @param {string} o.listPath    查询路径（返回任务列表，多数通道用这个）
 * @param {object} o.payload     提交体
 * @param {number} o.interval    轮询间隔 ms
 * @param {number} o.maxWait     最长等待 ms
 * @param {(t:any)=>({status:string, progress?:number, result?:any, error?:string})} o.pick 状态解析
 */
export async function runAsyncTask(o) {
  const {
    submitPath,
    fetchPath,
    listPath,
    payload,
    interval = 5000,
    maxWait = 600_000,
    pick,
    submitBase,
    fetchBase,
    taskLabel = 'task',
    onTick,
  } = o;

  const cfg = loadConfig();
  const submitRes = await request(submitPath, {
    method: 'POST',
    body: payload,
    base: submitBase || cfg.rootUrl,
    timeout: 60_000,
    retries: 0,
  });

  const taskId =
    submitRes.body?.data ??
    submitRes.body?.id ??
    submitRes.body?.task_id ??
    submitRes.body?.taskId ??
    submitRes.body?.data?.task_id ??
    submitRes.body?.data?.id;

  if (!taskId) {
    throw new GatewayError(`提交${taskLabel}未返回任务 ID`, {
      body: JSON.stringify(submitRes.body).slice(0, 300),
    });
  }

  const started = Date.now();
  let last = null;
  while (Date.now() - started < maxWait) {
    await sleep(interval);
    const id = typeof taskId === 'object' ? taskId.id : taskId;
    const res = listPath
      ? await request(listPath, {
          method: 'POST',
          body: { ids: [id] },
          base: fetchBase || cfg.rootUrl,
          timeout: 60_000,
        })
      : await request(`${fetchPath}?ids=${encodeURIComponent(id)}`, {
          base: fetchBase || cfg.rootUrl,
          timeout: 60_000,
        });

    const raw = listPath ? res.body?.data?.[0] ?? res.body?.data ?? res.body : res.body?.data ?? res.body;
    const st = pick ? pick(raw) : { status: normalStatus(raw?.status) };
    last = st;
    if (onTick) onTick({ ...st, taskId: id, elapsed: Date.now() - started });

    if (st.status === 'succeeded' || st.status === 'complete' || st.status === 'completed') {
      return { taskId: id, result: st.result ?? raw, elapsed: Date.now() - started };
    }
    if (st.status === 'failed' || st.status === 'error') {
      throw new GatewayError(`${taskLabel}执行失败：${st.error || '未知原因'}`, {
        body: JSON.stringify(raw).slice(0, 500),
      });
    }
  }
  throw new GatewayError(
    `${taskLabel}超时（${Math.round(maxWait / 1000)}s 未完成，最后状态 ${last?.status || 'unknown'}，taskId=${last?.taskId || '-'}）`
  );
}

function normalStatus(s) {
  const v = String(s || '').toLowerCase();
  if (['success', 'succeeded', 'succeed', 'complete', 'completed', 'done'].includes(v)) return 'succeeded';
  if (['fail', 'failed', 'error', 'cancelled', 'canceled'].includes(v)) return 'failed';
  if (!v) return 'unknown';
  return 'running';
}