audio_codebuddy/llm_patch.py

#!/usr/bin/env python3
import sys
p = "/data/ai/audio/server/src/services/llm/index.ts"
src = open(p).read()

old = """async function invokeWithRetry(
  fn: (modelId: string) => Promise<any>,
  modelId: string,
  callType: string = 'llm_chat',
): Promise<any> {
  const tried = new Map<string, any>();
  let cur = modelId;
  let lastErr: any = null;

  for (let round = 0; round < HA_RETRY_CONFIG.maxFallbackRounds; round++) {
    // 关键: 选下一个未 tried 的 node (而不是用 resolveProviderKey 取第一个匹配)
    const node = resolveNodeForModel(cur, tried);
    if (!node) {
      console.warn(`[LLM-HA] 已尝试所有 node，停止 fallback`);
      break;
    }
    const fp = node.provider.vendor + ':' + ((node.provider as any)._apiKey || '').slice(-12);
    tried.set(fp, { model: cur });
    let providerError: any = null;

    for (let attempt = 0; attempt < HA_RETRY_CONFIG.maxRetriesPerProvider; attempt++) {
      try {
        const result = await _runWithTriedNodes(new Set(tried.keys()), () => fn(cur));
        return result;
      } catch (error: any) {
        const em = error?.message || String(error);
        lastErr = error;
        providerError = error;
        if (isBusinessError(em) && !isNetworkError(em)) {
          console.error(`[LLM-HA] ${fp}/${cur} 业务错误，立即抛出: ${em.substring(0, 100)}`);
          throw error;
        }
        if (isNetworkError(em)) {
          console.warn(`[LLM-HA] ${fp}/${cur} 网络错误，立即切换: ${em.substring(0, 100)}`);
          break;
        }
        if (attempt < HA_RETRY_CONFIG.maxRetriesPerProvider - 1) {
          const delay = backoffWithJitter(attempt);
          console.log(`[LLM-HA] ${fp}/${cur} 第${attempt + 1}次失败: ${em.substring(0, 100)}, ${delay}ms后重试`);
          await new Promise(r => setTimeout(r, delay));
        }
      }
    }"""

new = """// 单次 LLM 调用超时：防止某个 provider 永久 hang 而 invokeWithRetry
// 永远得不到 error 来触发切换。
// 设为 3 分钟，比 ChatOpenAI 内部 timeout 略大，确保正常请求有足够时间响应。
const LLM_INVOKE_TIMEOUT_MS = 180_000;

function invokeWithTimeout<T>(fn: () => Promise<T>, ms: number): Promise<T> {
  let timer: ReturnType<typeof setTimeout> | null = null;
  const timeoutPromise = new Promise<never>((_, reject) => {
    timer = setTimeout(
      () => reject(new Error(`Request timeout after ${ms}ms (forced by invokeWithTimeout)`)),
      ms,
    );
  });
  return Promise.race([fn(), timeoutPromise]).finally(() => {
    if (timer) clearTimeout(timer);
  }) as Promise<T>;
}

async function invokeWithRetry(
  fn: (modelId: string) => Promise<any>,
  modelId: string,
  callType: string = 'llm_chat',
): Promise<any> {
  const tried = new Map<string, any>();
  let cur = modelId;
  let lastErr: any = null;

  for (let round = 0; round < HA_RETRY_CONFIG.maxFallbackRounds; round++) {
    // 关键: 选下一个未 tried 的 node (而不是用 resolveProviderKey 取第一个匹配)
    const node = resolveNodeForModel(cur, tried);
    if (!node) {
      console.warn(`[LLM-HA] 已尝试所有 node，停止 fallback`);
      break;
    }
    const fp = node.provider.vendor + ':' + ((node.provider as any)._apiKey || '').slice(-12);
    tried.set(fp, { model: cur });
    let providerError: any = null;

    for (let attempt = 0; attempt < HA_RETRY_CONFIG.maxRetriesPerProvider; attempt++) {
      try {
        const result = await invokeWithTimeout(
          () => _runWithTriedNodes(new Set(tried.keys()), () => fn(cur)),
          LLM_INVOKE_TIMEOUT_MS,
        );
        return result;
      } catch (error: any) {
        const em = error?.message || String(error);
        lastErr = error;
        providerError = error;
        if (isBusinessError(em) && !isNetworkError(em)) {
          console.error(`[LLM-HA] ${fp}/${cur} 业务错误，立即抛出: ${em.substring(0, 100)}`);
          throw error;
        }
        if (isNetworkError(em)) {
          console.warn(`[LLM-HA] ${fp}/${cur} 网络错误，立即切换: ${em.substring(0, 100)}`);
          break;
        }
        if (attempt < HA_RETRY_CONFIG.maxRetriesPerProvider - 1) {
          const delay = backoffWithJitter(attempt);
          console.log(`[LLM-HA] ${fp}/${cur} 第${attempt + 1}次失败: ${em.substring(0, 100)}, ${delay}ms后重试`);
          await new Promise(r => setTimeout(r, delay));
        }
      }
    }"""

if old not in src:
    print("ERROR: old block not found")
    sys.exit(1)

src = src.replace(old, new, 1)
open(p, 'w').write(src)
print("patched llm/index.ts OK")