@ebowwa/coder 0.7.64 → 0.7.65

package/packages/src/core/providers/index.ts

@@ -0,0 +1,419 @@
+/**
+ * Provider Registry - Central registry for all LLM providers
+ *
+ * Exports:
+ * - PROVIDERS: All provider configurations
+ * - getProvider(): Get provider by name
+ * - getProviderForModel(): Detect provider from model name
+ * - resolveProvider(): Resolve provider with API key
+ * - RollingKeysManager integration for multi-key rotation
+ */
+
+import type {
+  ProviderName,
+  ProviderConfig,
+  ProviderHealth,
+  ResolvedProvider,
+} from "./types.js";
+import { DEFAULT_ROUTING_CONFIG, BACKOFF_CONFIG } from "./types.js";
+import { RollingKeyManager } from "@ebowwa/rolling-keys";
+
+// ============================================================
+// Provider Configurations
+// ============================================================
+
+/**
+ * All available provider configurations
+ */
+export const PROVIDERS: Record<ProviderName, ProviderConfig> = {
+  // ============================================================
+  // Z.AI / Zhipu (GLM Models)
+  // ============================================================
+  zhipu: {
+    name: "zhipu",
+    displayName: "Z.AI (GLM)",
+    endpoint: process.env.ZHIPU_BASE_URL || "https://api.z.ai/api/coding/paas/v4",
+    authHeader: "Authorization",
+    apiKeyEnv: ["Z_AI_API_KEY", "ZAI_API_KEY", "GLM_API_KEY", "ZHIPU_API_KEY"],
+    format: "openai",
+    defaultModel: "GLM-4.7",
+    models: [
+      // GLM-5 (quota: 3x peak, 2x off-peak)
+      "GLM-5", "glm-5",
+      // GLM-4.x series (shared quota)
+      "GLM-4.7", "glm-4.7",
+      "GLM-4.6", "glm-4.6",
+      "GLM-4.5V", "glm-4.5v",  // Vision variant
+      "GLM-4.5", "glm-4.5",
+      "GLM-4.5-Air", "glm-4.5-air",
+    ],
+    supportsStreaming: true,
+    supportsToolCalling: true,
+    supportsVision: true,
+    supportsThinking: true,
+  },
+
+  // ============================================================
+  // MiniMax (M2.5 Model)
+  // ============================================================
+  minimax: {
+    name: "minimax",
+    displayName: "MiniMax",
+    endpoint: process.env.MINIMAX_BASE_URL || "https://api.minimax.io/anthropic",
+    authHeader: "ANTHROPIC_AUTH_TOKEN", // Anthropic-compatible
+    apiKeyEnv: ["MINIMAX_API_KEY"],
+    format: "anthropic",
+    defaultModel: "MiniMax-M2.5",
+    models: ["MiniMax-M2.5", "minimax-m2.5"],
+    supportsStreaming: true,
+    supportsToolCalling: true,
+    supportsVision: true,
+    supportsThinking: false,
+  },
+
+  // ============================================================
+  // OpenAI (Future)
+  // ============================================================
+  openai: {
+    name: "openai",
+    displayName: "OpenAI",
+    endpoint: process.env.OPENAI_BASE_URL || "https://api.openai.com/v1",
+    authHeader: "Authorization",
+    apiKeyEnv: ["OPENAI_API_KEY"],
+    format: "openai",
+    defaultModel: "gpt-4-turbo",
+    models: ["gpt-4-turbo", "gpt-4", "gpt-3.5-turbo"],
+    supportsStreaming: true,
+    supportsToolCalling: true,
+    supportsVision: true,
+    supportsThinking: false,
+  },
+
+  // ============================================================
+  // Anthropic (Stub - Not Implemented)
+  // ============================================================
+  anthropic: {
+    name: "anthropic",
+    displayName: "Anthropic (Not Implemented)",
+    endpoint: process.env.ANTHROPIC_BASE_URL || "https://api.anthropic.com",
+    authHeader: "x-api-key",
+    apiKeyEnv: ["ANTHROPIC_API_KEY"],
+    format: "anthropic",
+    defaultModel: "claude-sonnet-4-6",
+    models: ["claude-opus-4-6", "claude-sonnet-4-6", "claude-haiku-4-5"],
+    supportsStreaming: true,
+    supportsToolCalling: true,
+    supportsVision: true,
+    supportsThinking: true,
+  },
+};
+
+// ============================================================
+// Provider Health Tracking
+// ============================================================
+
+/**
+ * Health status for each provider
+ */
+const providerHealth: Record<ProviderName, ProviderHealth> = {
+  zhipu: { name: "zhipu", healthy: true, failureCount: 0, avgLatencyMs: 0, totalRequests: 0 },
+  minimax: { name: "minimax", healthy: true, failureCount: 0, avgLatencyMs: 0, totalRequests: 0 },
+  openai: { name: "openai", healthy: true, failureCount: 0, avgLatencyMs: 0, totalRequests: 0 },
+  anthropic: { name: "anthropic", healthy: false, failureCount: 999, avgLatencyMs: 0, totalRequests: 0 },
+};
+
+// ============================================================
+// Rolling Keys Support (using @ebowwa/rolling-keys)
+// ============================================================
+
+/**
+ * Rolling key managers for each provider
+ */
+const keyManagers: Partial<Record<ProviderName, RollingKeyManager>> = {};
+
+/**
+ * Get or create rolling key manager for a provider
+ */
+function getKeyManager(provider: ProviderName): RollingKeyManager | null {
+  if (keyManagers[provider]) {
+    return keyManagers[provider]!;
+  }
+
+  const config = PROVIDERS[provider];
+  if (!config || config.apiKeyEnv.length === 0) {
+    return null;
+  }
+
+  try {
+    // Create manager with first env var as primary, optional plural for array
+    const manager = new RollingKeyManager({
+      keysEnvVar: config.apiKeyEnv[0] + "S", // e.g., Z_AI_API_KEYS
+      singleKeyEnvVar: config.apiKeyEnv[0],  // e.g., Z_AI_API_KEY
+    });
+
+    // Only cache if it has keys
+    if (manager.getKeyCount() > 0) {
+      keyManagers[provider] = manager;
+      return manager;
+    }
+  } catch {
+    // No keys configured for this provider
+    return null;
+  }
+
+  return null;
+}
+
+/**
+ * Get next API key using round-robin rotation with health tracking
+ */
+function getNextKey(provider: ProviderName): string | null {
+  const manager = getKeyManager(provider);
+  if (!manager) {
+    return null;
+  }
+
+  const result = manager.getNextKey();
+  return result?.key ?? null;
+}
+
+/**
+ * Record key success (for health tracking)
+ * Note: RollingKeyManager tracks success internally via key rotation
+ */
+function recordKeySuccess(_provider: ProviderName): void {
+  // Success is tracked implicitly through continued key usage
+  // RollingKeyManager handles health recovery automatically
+}
+
+/**
+ * Record key failure (triggers backoff)
+ */
+function recordKeyFailure(provider: ProviderName, error?: unknown): void {
+  const manager = getKeyManager(provider);
+  if (manager && error) {
+    manager.handleError(error);
+  }
+}
+
+// ============================================================
+// Provider Resolution
+// ============================================================
+
+/**
+ * Model name to provider mapping
+ */
+const MODEL_TO_PROVIDER: Record<string, ProviderName> = {
+  // Zhipu / Z.AI models (coding plan - shared quota)
+  "glm-5": "zhipu",        // 3x peak, 2x off-peak
+  "glm-4.7": "zhipu",
+  "glm-4.6": "zhipu",
+  "glm-4.5v": "zhipu",     // Vision variant
+  "glm-4.5-air": "zhipu",
+  "glm-4.5": "zhipu",
+  "glm-4-plus": "zhipu",
+
+  // MiniMax models
+  "minimax-m2.5": "minimax",
+  "minimax-m2": "minimax",
+  "abab6.5-chat": "minimax",
+
+  // OpenAI models
+  "gpt-4": "openai",
+  "gpt-4-turbo": "openai",
+  "gpt-3.5-turbo": "openai",
+  "gpt-4o": "openai",
+
+  // Anthropic models (stub)
+  "claude-opus-4-6": "anthropic",
+  "claude-sonnet-4-6": "anthropic",
+  "claude-haiku-4-5": "anthropic",
+};
+
+/**
+ * Get provider configuration by name
+ */
+export function getProvider(name: ProviderName): ProviderConfig | undefined {
+  return PROVIDERS[name];
+}
+
+/**
+ * Detect provider from model name
+ */
+export function getProviderForModel(model: string): ProviderName {
+  // Normalize model name
+  const normalizedModel = model.toLowerCase();
+
+  // Direct lookup
+  if (MODEL_TO_PROVIDER[normalizedModel]) {
+    return MODEL_TO_PROVIDER[normalizedModel];
+  }
+
+  // Partial match
+  for (const [modelPrefix, provider] of Object.entries(MODEL_TO_PROVIDER)) {
+    if (normalizedModel.includes(modelPrefix) || modelPrefix.includes(normalizedModel)) {
+      return provider;
+    }
+  }
+
+  // Check provider model lists
+  for (const [providerName, config] of Object.entries(PROVIDERS)) {
+    if (config.models.some((m) => m.toLowerCase() === normalizedModel)) {
+      return providerName as ProviderName;
+    }
+  }
+
+  // Default to zhipu
+  return "zhipu";
+}
+
+/**
+ * Resolve provider with API key and endpoint
+ */
+export function resolveProvider(
+  model: string,
+  preferredProvider?: ProviderName
+): ResolvedProvider | null {
+  // Determine provider
+  const providerName = preferredProvider || getProviderForModel(model);
+  const config = PROVIDERS[providerName];
+
+  if (!config) {
+    console.error(`Unknown provider: ${providerName}`);
+    return null;
+  }
+
+  // Get API key
+  const apiKey = getNextKey(providerName);
+  if (!apiKey) {
+    console.error(`No API key found for provider: ${providerName}`);
+    return null;
+  }
+
+  // Build endpoint URL
+  let endpoint = config.endpoint;
+  if (config.format === "openai") {
+    endpoint = `${endpoint}/chat/completions`;
+  } else {
+    endpoint = `${endpoint}/v1/messages`;
+  }
+
+  // Resolve model name
+  const resolvedModel = config.models.includes(model)
+    ? model
+    : config.defaultModel;
+
+  return {
+    config,
+    apiKey,
+    endpoint,
+    model: resolvedModel,
+  };
+}
+
+/**
+ * Check if provider is healthy
+ */
+export function isProviderHealthy(provider: ProviderName): boolean {
+  const health = providerHealth[provider];
+  if (!health) return false;
+
+  // Check if in backoff period
+  if (health.backoffUntil && health.backoffUntil > Date.now()) {
+    return false;
+  }
+
+  return health.healthy;
+}
+
+/**
+ * Record provider success
+ */
+export function recordProviderSuccess(provider: ProviderName, latencyMs: number): void {
+  const health = providerHealth[provider];
+  if (!health) return;
+
+  health.healthy = true;
+  health.lastSuccess = Date.now();
+  health.failureCount = 0;
+  health.backoffUntil = undefined;
+  health.totalRequests++;
+
+  // Update average latency
+  health.avgLatencyMs = Math.round(
+    (health.avgLatencyMs * (health.totalRequests - 1) + latencyMs) / health.totalRequests
+  );
+}
+
+/**
+ * Record provider failure
+ */
+export function recordProviderFailure(provider: ProviderName): void {
+  const health = providerHealth[provider];
+  if (!health) return;
+
+  health.failureCount++;
+  health.lastFailure = Date.now();
+  health.totalRequests++;
+
+  // Apply backoff after threshold
+  if (health.failureCount >= BACKOFF_CONFIG.failureThreshold) {
+    health.healthy = false;
+    const backoffMs = Math.min(
+      BACKOFF_CONFIG.baseMs * Math.pow(2, health.failureCount - BACKOFF_CONFIG.failureThreshold),
+      BACKOFF_CONFIG.maxMs
+    );
+    health.backoffUntil = Date.now() + backoffMs;
+    console.warn(
+      `[Provider] ${provider} unhealthy after ${health.failureCount} failures, ` +
+        `backoff for ${Math.round(backoffMs / 1000)}s`
+    );
+  }
+}
+
+/**
+ * Get health status for all providers
+ */
+export function getProviderHealth(): Record<ProviderName, ProviderHealth> {
+  return { ...providerHealth };
+}
+
+/**
+ * Get list of healthy providers
+ */
+export function getHealthyProviders(): ProviderName[] {
+  return (Object.keys(PROVIDERS) as ProviderName[]).filter(isProviderHealthy);
+}
+
+/**
+ * Get next healthy provider from fallback chain
+ */
+export function getNextHealthyProvider(
+  excludeProvider?: ProviderName
+): ProviderName | null {
+  const chain = DEFAULT_ROUTING_CONFIG.fallbackChain;
+
+  for (const provider of chain) {
+    if (excludeProvider && provider === excludeProvider) continue;
+    if (isProviderHealthy(provider)) {
+      return provider;
+    }
+  }
+
+  return null;
+}
+
+// ============================================================
+// Initialization
+// ============================================================
+
+// Key pools are loaded lazily via getKeyManager()
+
+// Re-export types
+export type {
+  ProviderName,
+  ProviderConfig,
+  ProviderHealth,
+  ResolvedProvider,
+  ProviderRoutingConfig,
+} from "./types.js";

package/packages/src/core/providers/types.ts

@@ -0,0 +1,132 @@
+/**
+ * Provider Types - Generic LLM Provider Abstraction
+ *
+ * Supports multiple providers:
+ * - zhipu (Z.AI / GLM models)
+ * - minimax (MiniMax M2.5)
+ * - openai (future)
+ * - anthropic (stub/commented only)
+ */
+
+// ============================================================
+// Provider Types
+// ============================================================
+
+/**
+ * Supported provider names
+ */
+export type ProviderName = "zhipu" | "minimax" | "openai" | "anthropic";
+
+/**
+ * API format type - determines request/response format
+ */
+export type APIFormat = "anthropic" | "openai";
+
+/**
+ * Provider configuration
+ */
+export interface ProviderConfig {
+  /** Provider identifier */
+  name: ProviderName;
+  /** Human-readable display name */
+  displayName: string;
+  /** API endpoint base URL */
+  endpoint: string;
+  /** Authentication header name */
+  authHeader: string;
+  /** Environment variable names for API keys (checked in order) */
+  apiKeyEnv: string[];
+  /** API format (anthropic or openai compatible) */
+  format: APIFormat;
+  /** Default model for this provider */
+  defaultModel: string;
+  /** List of supported model IDs */
+  models: string[];
+  /** Whether provider supports streaming */
+  supportsStreaming: boolean;
+  /** Whether provider supports tool calling */
+  supportsToolCalling: boolean;
+  /** Whether provider supports vision/images */
+  supportsVision: boolean;
+  /** Whether provider supports extended thinking */
+  supportsThinking: boolean;
+}
+
+/**
+ * Provider health status
+ */
+export interface ProviderHealth {
+  /** Provider name */
+  name: ProviderName;
+  /** Whether provider is healthy */
+  healthy: boolean;
+  /** Last successful request timestamp */
+  lastSuccess?: number;
+  /** Last failure timestamp */
+  lastFailure?: number;
+  /** Consecutive failure count */
+  failureCount: number;
+  /** Backoff until timestamp (if in cooldown) */
+  backoffUntil?: number;
+  /** Average latency in ms */
+  avgLatencyMs: number;
+  /** Total requests made */
+  totalRequests: number;
+}
+
+/**
+ * Provider routing configuration
+ */
+export interface ProviderRoutingConfig {
+  /** Enable fallback between providers */
+  fallbackEnabled: boolean;
+  /** Fallback chain (order of providers to try) */
+  fallbackChain: ProviderName[];
+  /** Latency threshold in ms before trying next provider */
+  latencyThresholdMs: number;
+  /** Enable health tracking */
+  healthTracking: boolean;
+  /** Enable load balancing across healthy providers */
+  loadBalancing: boolean;
+}
+
+/**
+ * Resolved provider info for a request
+ */
+export interface ResolvedProvider {
+  /** Provider configuration */
+  config: ProviderConfig;
+  /** API key to use */
+  apiKey: string;
+  /** Full endpoint URL */
+  endpoint: string;
+  /** Model to use (may be mapped from requested model) */
+  model: string;
+}
+
+// ============================================================
+// Default Configuration
+// ============================================================
+
+/**
+ * Default provider routing configuration
+ */
+export const DEFAULT_ROUTING_CONFIG: ProviderRoutingConfig = {
+  fallbackEnabled: true,
+  fallbackChain: ["zhipu", "minimax"],
+  latencyThresholdMs: 30000, // 30 seconds
+  healthTracking: true,
+  loadBalancing: false, // Disabled by default - use primary provider
+};
+
+/**
+ * Backoff configuration for health tracking
+ */
+export const BACKOFF_CONFIG = {
+  /** Base backoff in ms */
+  baseMs: 60000, // 1 minute
+  /** Maximum backoff in ms */
+  maxMs: 3600000, // 1 hour
+  /** Number of failures before backoff */
+  failureThreshold: 3,
+} as const;

package/packages/src/core/retry.ts

@@ -69,6 +69,16 @@ function isRetryableError(
     return true;
   }
 
+  // Check for transient API errors (empty/malformed responses)
+  if (error.message.includes("No message received") ||
+      error.message.includes("empty response") ||
+      error.message.includes("invalid response") ||
+      error.message.includes("unexpected EOF") ||
+      error.message.includes("connection reset") ||
+      error.message.includes("aborted")) {
+    return true;
+  }
+
   // Check for specific status codes in error message
   for (const code of retryableStatusCodes) {
     if (error.message.includes(String(code))) {