audio_codebuddy/fix_v3.py

PATH = '/data/ai/audio/server/dist/services/llm/index.js'
with open(PATH, 'r', encoding='utf-8') as f:
    src = f.read()

# 替换整个 invokeWithRetry 函数（旧的 + trySwitchModel 调用 + tried.has(resolveProviderKey(next)) 检查）
# 找到从 "async function invokeWithRetry" 到下一个 "async function callLLM" 的整段
OLD = """async function invokeWithRetry(fn, modelId, callType = 'llm_chat') {
    const tried = new Map();
    let cur = modelId, lastErr = null;
    for (let round = 0; round < exports.HA_RETRY_CONFIG.maxFallbackRounds; round++) {
        const provider = resolveProviderKey(cur);
        // tried check is now done in resolveNodeForModel below; keep this as a guard
        if (tried.has(provider))
            continue;
        let providerError = null;
        for (let attempt = 0; attempt < exports.HA_RETRY_CONFIG.maxRetriesPerProvider; attempt++) {
            // 用 AsyncLocalStorage 传递 tried set, 让 findVendorForModel 跳过已试 node
            const node = resolveNodeForModel(cur, tried);
            if (!node) break;
            const fp = node.provider.vendor + ':' + (node.provider._apiKey || '').slice(-12);
            tried.set(fp, { model: cur });
            try {
                const result = await _runWithTriedNodes(new Set(tried.keys()), () => fn(cur));
                return result;
            }
            catch (error) {
                const em = error?.message || String(error);
                lastErr = error;
                providerError = error;
                if (isBusinessError(em) && !isNetworkError(em)) {
                    console.error(`[LLM-HA] ${provider}/${cur} 业务错误，立即抛出: ${em.substring(0, 100)}`);
                    throw error;
                }
                if (isNetworkError(em)) {
                    console.warn(`[LLM-HA] ${provider}/${cur} 网络错误，立即切换: ${em.substring(0, 100)}`);
                    break;
                }
                if (attempt < exports.HA_RETRY_CONFIG.maxRetriesPerProvider - 1) {
                    const delay = backoffWithJitter(attempt);
                    console.log(`[LLM-HA] ${provider}/${cur} 第${attempt + 1}次失败: ${em.substring(0, 100)}, ${delay}ms后重试`);
                    await new Promise(r => setTimeout(r, delay));
                }
            }
        }
        const em = providerError?.message || 'unknown';
        const et = isNetworkError(em) ? 'network' : (isBusinessError(em) ? 'business' : 'unknown');
        tried.set(provider, { model: cur, error: em, attempts: exports.HA_RETRY_CONFIG.maxRetriesPerProvider, errorType: et });
        console.log(`[LLM-HA] ${provider}/${cur} 失败 (${et})，尝试切换供应商`);
        const next = trySwitchModel(cur, providerError);
        if (!next)
            break;
        if (tried.has(resolveProviderKey(next))) {
            console.warn(`[LLM-HA] 下一供应商已尝试过，停止 fallback`);
            break;
        }
        cur = next;
        console.log(`[LLM-HA] 切换到 ${resolveProviderKey(cur)}/${cur}`);
    }
    throw new AllProvidersFailedError(Array.from(tried.entries()).map(([provider, info]) => ({ provider, ...info })));
}"""

NEW = """async function invokeWithRetry(fn, modelId, callType = 'llm_chat') {
    const tried = new Map();
    let cur = modelId, lastErr = null;
    for (let round = 0; round < exports.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._apiKey || '').slice(-12);
        tried.set(fp, { model: cur });
        let providerError = null;
        for (let attempt = 0; attempt < exports.HA_RETRY_CONFIG.maxRetriesPerProvider; attempt++) {
            try {
                const result = await _runWithTriedNodes(new Set(tried.keys()), () => fn(cur));
                return result;
            }
            catch (error) {
                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 < exports.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));
                }
            }
        }
        // 当前 node 失败, 标记其 info
        const em = providerError?.message || 'unknown';
        const et = isNetworkError(em) ? 'network' : (isBusinessError(em) ? 'business' : 'unknown');
        tried.set(fp, { model: cur, error: em, attempts: exports.HA_RETRY_CONFIG.maxRetriesPerProvider, errorType: et });
        console.log(`[LLM-HA] ${fp}/${cur} 失败 (${et})，尝试切换供应商`);
        // 关键: 自己挑下一个未 tried 的 node, 复用 switchToNextVendorModel 的逻辑但跳过 tried
        const next = pickNextModelAfterFailure(cur, tried, node);
        if (!next) {
            console.warn(`[LLM-HA] 找不到下一个未尝试的供应商`);
            break;
        }
        cur = next;
        console.log(`[LLM-HA] 切换到 ${next}`);
    }
    throw new AllProvidersFailedError(Array.from(tried.entries()).map(([provider, info]) => ({ provider, ...info })));
}

/** 从当前 node 找下一个未 tried 的 node, 返回 next modelId */
function pickNextModelAfterFailure(currentModelId, tried, currentNode) {
    const registry = (0, provider_registry_1.getLlmRegistry)();
    const available = registry.listAvailable();
    if (available.length === 0) return null;
    // 关键: 完全按 node 顺序, 跳过 tried 节点
    const startIndex = currentNode ? (available.indexOf(currentNode) + 1) % available.length : 0;
    for (let i = 0; i < available.length; i++) {
        const idx = (startIndex + i) % available.length;
        const node = available[idx];
        if (node === currentNode) continue;
        const fp = node.provider.vendor + ':' + (node.provider._apiKey || '').slice(-12);
        if (tried.has(fp)) continue;
        // 1. 同名 model
        if (node.provider.hasModel(currentModelId)) {
            console.log(`[LLM] 同模型切换到供应商 ${node.provider.displayName}，模型 ${currentModelId}`);
            return currentModelId;
        }
        // 2. 别名 model
        for (const modelId of node.provider.textModels) {
            const cfg = node.provider.getModelConfig(modelId);
            if (cfg?.canonicalModel === currentModelId) {
                console.log(`[LLM] 别名模型切换到供应商 ${node.provider.displayName}，模型 ${modelId}`);
                return modelId;
            }
        }
        // 3. 第一个 text model
        if (node.provider.textModels.length > 0) {
            const nextId = node.provider.textModels[0];
            console.log(`[LLM] 切换到供应商 ${node.provider.displayName}，模型 ${nextId}`);
            return nextId;
        }
    }
    return null;
}"""

if OLD not in src:
    print("ERROR: OLD invokeWithRetry not found exactly")
    # 看一下当前 invokeWithRetry 实际
    import re
    m = re.search(r'async function invokeWithRetry.*?throw new AllProvidersFailedError.*?;', src, re.DOTALL)
    if m:
        print("CURRENT (first 500 chars):")
        print(m.group(0)[:500])
    raise SystemExit(1)
src = src.replace(OLD, NEW)

with open(PATH, 'w', encoding='utf-8') as f:
    f.write(src)
print("OK: invokeWithRetry v3 patched (skip tried nodes via pickNextModelAfterFailure)")