artifacts/api-server/src/lib/models.ts

import { config } from "./config";

export type AdapterId =
  | "anthropic"
  | "openai"
  | "gemini"
  | "deepseek"
  | "kimi"
  | "zhipu"
  | "minimax"
  | "qwen"
  | "codex-cli";

/**
 * Provider source bucket — every model declares which call paths can
 * serve it. Availability is now a per-source check rather than a global
 * profile gate, so e.g. MiniMax/Kimi work whether or not the hosted
 * Replit integrations are configured. `DOATLAS_PROFILE` is downgraded
 * to a *preference filter* in `listModels` (it controls ordering /
 * default model picking, not whether a configured key works).
 */
export type ModelSource =
  | "replit-integration"
  | "byok-direct"
  | "china-direct";

export interface ModelInfo {
  id: string;
  name: string;
  provider: string;
  description: string;
  context_length: number;
  capabilities: Array<"text" | "vision" | "tools" | "streaming">;
  pricing: { input_per_1k: number; output_per_1k: number };
  default: boolean;
  available: boolean;
  // Internal routing hints (not exposed to client).
  tier: "S" | "A";
  adapter: AdapterId;
  upstreamModel: string;
  /**
   * Source buckets this model can be served from. The first matching
   * bucket with credentials wins; we never try to fall through to a
   * different source mid-call.
   */
  sources: ModelSource[];
}

const ALL_MODELS: ModelInfo[] = [
  // -- S tier (旗舰) --
  {
    id: "mdl_claude-sonnet-4-6",
    name: "DoAtlas Pro",
    provider: "DoAtlas",
    description:
      "旗舰研究模型，深度推理与代码能力最强，适合复杂文献综述、机制推理与多步分析。",
    context_length: 200_000,
    capabilities: ["text", "vision", "tools", "streaming"],
    pricing: { input_per_1k: 0, output_per_1k: 0 },
    default: false,
    available: false,
    tier: "S",
    adapter: "anthropic",
    upstreamModel: "claude-sonnet-4-6",
    sources: ["replit-integration", "byok-direct"],
  },
  {
    id: "mdl_gpt-5_2",
    name: "DoAtlas Pro X",
    provider: "DoAtlas",
    description:
      "多模态旗舰模型，视觉与文本兼顾，适合图表解读、影像分析与通用研究问答。",
    context_length: 128_000,
    capabilities: ["text", "vision", "tools", "streaming"],
    pricing: { input_per_1k: 0, output_per_1k: 0 },
    default: false,
    available: false,
    tier: "S",
    adapter: "openai",
    upstreamModel: "gpt-5.2",
    sources: ["replit-integration", "byok-direct"],
  },
  {
    id: "mdl_gemini-2_5-pro",
    name: "DoAtlas Vision",
    provider: "DoAtlas",
    description:
      "Gemini 2.5 Pro 多模态旗舰，超长上下文与图表理解，适合长文档解析与多模态研究问答。",
    context_length: 1_000_000,
    capabilities: ["text", "vision", "tools", "streaming"],
    pricing: { input_per_1k: 0, output_per_1k: 0 },
    default: false,
    available: false,
    tier: "S",
    adapter: "gemini",
    upstreamModel: "gemini-2.5-pro",
    sources: ["replit-integration", "byok-direct"],
  },
  {
    id: "mdl_deepseek-r1",
    name: "DoAtlas Reason Pro",
    provider: "DoAtlas",
    description:
      "DeepSeek R1 推理模型，专注链式推理与数学/科学论证，适合机制假设与复杂证据综合。",
    context_length: 128_000,
    capabilities: ["text", "tools", "streaming"],
    pricing: { input_per_1k: 0, output_per_1k: 0 },
    default: false,
    available: false,
    tier: "S",
    adapter: "deepseek",
    upstreamModel: "deepseek-reasoner",
    sources: ["china-direct"],
  },
  {
    id: "mdl_codex-cli",
    name: "DoAtlas Code",
    provider: "DoAtlas Local",
    description:
      "本地代码助手，运行在你自己的机器上，可读写本地文件与执行命令；需在本地部署中配置。",
    context_length: 200_000,
    capabilities: ["text", "tools", "streaming"],
    pricing: { input_per_1k: 0, output_per_1k: 0 },
    default: false,
    available: false,
    tier: "S",
    adapter: "codex-cli",
    upstreamModel: "codex-cli",
    sources: ["byok-direct"],
  },

  // -- A tier (通用) --
  {
    id: "mdl_glm-4_6",
    name: "DoAtlas Air",
    provider: "DoAtlas",
    description:
      "智谱 GLM-4 系列，长上下文 128k，响应快速，适合日常问答与长文本处理。",
    context_length: 131_072,
    capabilities: ["text", "tools", "streaming"],
    pricing: { input_per_1k: 0, output_per_1k: 0 },
    default: false,
    available: false,
    tier: "A",
    adapter: "zhipu",
    upstreamModel: "glm-4-plus",
    sources: ["china-direct"],
  },
  {
    id: "mdl_minimax-m2",
    name: "DoAtlas Search",
    provider: "DoAtlas",
    description:
      "MiniMax M2 智能体模型，配合统一 web_search 工具实现实时联网检索，适合需要最新资料的研究问答。",
    context_length: 200_000,
    capabilities: ["text", "tools", "streaming"],
    pricing: { input_per_1k: 0, output_per_1k: 0 },
    default: false,
    available: false,
    tier: "A",
    adapter: "minimax",
    upstreamModel: "MiniMax-M2.7-highspeed",
    sources: ["china-direct"],
  },
  {
    id: "mdl_deepseek-v3",
    name: "DoAtlas Chat",
    provider: "DoAtlas",
    description:
      "DeepSeek V3 通用对话模型，性价比高，适合日常问答与代码补全。",
    context_length: 128_000,
    capabilities: ["text", "tools", "streaming"],
    pricing: { input_per_1k: 0, output_per_1k: 0 },
    default: false,
    available: false,
    tier: "A",
    adapter: "deepseek",
    upstreamModel: "deepseek-chat",
    sources: ["china-direct"],
  },
  {
    id: "mdl_kimi-k2",
    name: "DoAtlas Long",
    provider: "DoAtlas",
    description:
      "Kimi 长文本模型，超长上下文与文档检索能力强，适合大文档/多 PDF 阅读。",
    context_length: 256_000,
    capabilities: ["text", "tools", "streaming"],
    pricing: { input_per_1k: 0, output_per_1k: 0 },
    default: false,
    available: false,
    tier: "A",
    adapter: "kimi",
    upstreamModel: "kimi-k2-0905-preview",
    sources: ["china-direct"],
  },
  {
    id: "mdl_qwen-max",
    name: "DoAtlas Bilingual",
    provider: "DoAtlas",
    description:
      "通义千问 Max 双语通用模型，中英推理与工具调用稳定，适合中文研究场景。",
    context_length: 131_072,
    capabilities: ["text", "tools", "streaming"],
    pricing: { input_per_1k: 0, output_per_1k: 0 },
    default: false,
    available: false,
    tier: "A",
    adapter: "qwen",
    upstreamModel: "qwen-max",
    sources: ["china-direct"],
  },
  {
    id: "mdl_gemini-2_5-flash",
    name: "DoAtlas Vision Lite",
    provider: "DoAtlas",
    description:
      "Gemini 2.5 Flash 轻量多模态模型，速度快、成本低，适合实时问答与轻量视觉任务。",
    context_length: 1_000_000,
    capabilities: ["text", "vision", "tools", "streaming"],
    pricing: { input_per_1k: 0, output_per_1k: 0 },
    default: false,
    available: false,
    tier: "A",
    adapter: "gemini",
    upstreamModel: "gemini-2.5-flash",
    sources: ["replit-integration", "byok-direct"],
  },
];

/**
 * Adapter-by-source availability table. A model is available when at
 * least one of its declared sources has credentials configured.
 */
function sourceAvailable(adapter: AdapterId, source: ModelSource): boolean {
  switch (adapter) {
    case "anthropic":
      if (source === "replit-integration") return config.hasAnthropicCloud;
      if (source === "byok-direct") return config.hasAnthropicDirect;
      return false;
    case "openai":
      if (source === "replit-integration") return config.hasOpenAICloud;
      if (source === "byok-direct") return config.hasOpenAIDirect;
      return false;
    case "gemini":
      if (source === "replit-integration") return config.hasGeminiCloud;
      if (source === "byok-direct") return config.hasGeminiDirect;
      return false;
    case "deepseek":
      return source === "china-direct" && config.hasDeepseek;
    case "kimi":
      return source === "china-direct" && config.hasKimi;
    case "zhipu":
      return source === "china-direct" && config.hasZhipu;
    case "minimax":
      return source === "china-direct" && config.hasMinimax;
    case "qwen":
      return source === "china-direct" && config.hasQwen;
    case "codex-cli":
      return source === "byok-direct" && config.hasCodexCli;
    default:
      return false;
  }
}

function isAvailable(m: ModelInfo): boolean {
  return m.sources.some((s) => sourceAvailable(m.adapter, s));
}

/**
 * Return the source actually used to call this model. The first
 * configured source from the model's `sources` list wins.
 */
export function resolveSource(m: ModelInfo): ModelSource | null {
  for (const s of m.sources) {
    if (sourceAvailable(m.adapter, s)) return s;
  }
  return null;
}

export function listModels(): ModelInfo[] {
  const profile = config.profile;
  const out = ALL_MODELS.map((m) => ({
    ...m,
    available: isAvailable(m),
  }));
  // Profile is now a preference filter for default selection only.
  // "cloud" prefers replit-integration/byok-direct models; "local"
  // accepts everything that's available.
  const preferCloud = profile === "cloud";
  const candidates = preferCloud
    ? out.filter(
        (m) =>
          m.available &&
          m.sources.some((s) => s === "replit-integration" || s === "byok-direct"),
      )
    : out.filter((m) => m.available);
  const first = candidates[0] ?? out.find((m) => m.available);
  if (first) {
    const target = out.find((m) => m.id === first.id);
    if (target) target.default = true;
  }
  return out;
}

export function findModel(id: string | null | undefined): ModelInfo | null {
  if (!id) return null;
  return listModels().find((m) => m.id === id) ?? null;
}

export function pickDefaultModel(): ModelInfo | null {
  return listModels().find((m) => m.default) ?? null;
}

export function publicModel(m: ModelInfo) {
  // Strip internal routing fields.
  const { tier: _t, adapter: _a, upstreamModel: _u, sources: _s, ...pub } = m;
  return pub;
}