zsjz-ai/ai-electron/electron/service/codex/providerService.ts

import { mkdir, readFile, writeFile } from 'node:fs/promises';
import { join } from 'node:path';
import { getCodexDataDir } from './codexHome';
import { DEFAULT_API_KEY_ENV, type CodexProviderSpec, type CodexRuntime, type JsonValue } from './codexRuntime';
import type { AppliedProviderInfo, ApplyProviderInput } from './types';

/**
 * 把后端「模型管理」里的一条记录翻译成 Codex 的 model provider。
 * 全程不做任何 Codex/OpenAI 账号登录：靠 wire_api="responses" + requires_openai_auth=false + env_key。
 * apiKey 只存在于内存与子进程环境变量，落盘的 AppliedProvider 不含任何密钥。
 */

export const PROVIDER_ID = 'zsjz';

export type AppliedProvider = AppliedProviderInfo;

/** 端点协议判定：responses 直通；chat-only 走内置桥接；unsupported 双协议都没有；unreachable 网络不可达 */
export type EndpointProtocol = 'responses' | 'chat-only' | 'unsupported' | 'unreachable';

export interface EndpointProbeResult {
  protocol: EndpointProtocol;
  /** /responses 的 HTTP 状态；未探或网络失败为 null */
  responsesStatus: number | null;
  /** /chat/completions 的 HTTP 状态；未探为 null */
  chatStatus: number | null;
  /** /api/version 探到的 Ollama 版本；非 Ollama 或未探为 null */
  ollamaVersion: string | null;
  /** Ollama 低于 0.13.3（没有非状态化 /v1/responses）：提示升级，但允许走桥接 */
  ollamaNeedsUpgrade: boolean;
  detail: string;
}

const EXTRA_ALLOWLIST: readonly string[] = [
  'request_max_retries',
  'stream_max_retries',
  'stream_idle_timeout_ms',
  'query_params',
];

function providerFile(): string {
  return join(getCodexDataDir(), 'provider.json');
}

/** Codex 要求 base_url 是 API 根（形如 https://api.openai.com/v1），它自己再拼 /responses */
export function normalizeBaseUrl(raw: string | null | undefined): string | null {
  const value = (raw ?? '').trim().replace(/\/+$/u, '');
  return value || null;
}

/** 给页面的非阻断提示；返回 null 表示没问题 */
export function baseUrlHint(baseUrl: string | null): string | null {
  if (!baseUrl) return '未配置 base_url';
  if (!/^https?:\/\//u.test(baseUrl)) return 'base_url 必须以 http:// 或 https:// 开头';
  if (!/\/v\d+$/u.test(baseUrl)) {
    return 'base_url 通常应写到版本段（如 .../v1），否则 Codex 拼出的 /responses 可能 404';
  }
  return null;
}

function parseJsonRecord(raw: unknown): Record<string, unknown> | null {
  if (!raw) return null;
  if (typeof raw === 'object' && !Array.isArray(raw)) return raw as Record<string, unknown>;
  if (typeof raw !== 'string') return null;
  try {
    const parsed = JSON.parse(raw) as unknown;
    return parsed && typeof parsed === 'object' && !Array.isArray(parsed)
      ? (parsed as Record<string, unknown>)
      : null;
  } catch {
    return null;
  }
}

function readStringMap(raw: unknown): Record<string, string> | null {
  const record = parseJsonRecord(raw);
  if (!record) return null;
  const result: Record<string, string> = {};
  for (const [key, value] of Object.entries(record)) {
    if (typeof value === 'string' && value) result[key] = value;
  }
  return Object.keys(result).length ? result : null;
}

/** Codex 内置 ollama/lmstudio 走 OpenAI 兼容根（.../v1），而库里常只存到 host:port */
function normalizeOssBaseUrl(raw: string | null | undefined): string | null {
  const base = normalizeBaseUrl(raw);
  if (!base) return null;
  return /\/v\d+$/u.test(base) ? base : `${base}/v1`;
}

/** OLLAMA 用 Codex 内置 provider（Responses API，默认端口 11434），不需要 api key */
function resolveBuiltinProvider(input: ApplyProviderInput): string | null {
  if (input.builtinProvider) return input.builtinProvider;
  const type = `${input.providerType ?? ''}`.trim().toUpperCase();
  if (type === 'OLLAMA') return 'ollama';
  if (type === 'LMSTUDIO') return 'lmstudio';
  return null;
}

export function toProviderSpec(input: ApplyProviderInput): CodexProviderSpec {
  const model = `${input.modelId ?? ''}`.trim();
  if (!model) throw new Error('缺少 modelId');

  const builtinProvider = resolveBuiltinProvider(input);
  if (builtinProvider) {
    return {
      id: null,
      builtinProvider,
      model,
      name: input.name ?? `${builtinProvider} 本地模型`,
      // 内置 provider 的默认地址是 localhost，必须把库里的真实地址带上，否则连不到远端 Ollama
      baseUrl: normalizeOssBaseUrl(input.baseUrl),
    };
  }

  const baseUrl = normalizeBaseUrl(input.baseUrl);
  if (!baseUrl) throw new Error('缺少 baseUrl');

  const config = parseJsonRecord(input.config);
  const extra: Record<string, JsonValue> = {};
  for (const [key, value] of Object.entries(config ?? {})) {
    if (EXTRA_ALLOWLIST.includes(key)) extra[key] = value as JsonValue;
  }

  return {
    id: PROVIDER_ID,
    name: input.name ?? model,
    baseUrl,
    model,
    apiKey: input.apiKey ?? null,
    envKey: DEFAULT_API_KEY_ENV,
    httpHeaders: readStringMap(input.headersJson),
    extra: Object.keys(extra).length ? extra : null,
  };
}

/** config 里被丢弃的非白名单键，用于页面提示 */
export function describeDroppedKeys(input: ApplyProviderInput): string[] {
  const config = parseJsonRecord(input.config);
  return Object.keys(config ?? {}).filter((key) => !EXTRA_ALLOWLIST.includes(key));
}

/** Ollama 自 0.13.3 起提供非状态化 /v1/responses（流式 / 工具调用 / reasoning summaries） */
const OLLAMA_RESPONSES_MIN_VERSION: readonly number[] = [0, 13, 3];

interface ProbeAttempt {
  status: number | null;
  error: string | null;
}

async function postProbe(url: string, body: Record<string, unknown>, apiKey?: string | null): Promise<ProbeAttempt> {
  try {
    const response = await fetch(url, {
      method: 'POST',
      headers: {
        'content-type': 'application/json',
        ...(apiKey ? { authorization: `Bearer ${apiKey}` } : {}),
      },
      body: JSON.stringify(body),
      signal: AbortSignal.timeout(8_000),
    });
    return { status: response.status, error: null };
  } catch (error) {
    return { status: null, error: error instanceof Error ? error.message : String(error) };
  }
}

/** Ollama 的 /api/version 挂在 API 根之外（去掉 /v1 尾段）；非 Ollama 端点返回 null */
async function probeOllamaVersion(baseUrl: string): Promise<string | null> {
  const origin = baseUrl.replace(/\/v\d+$/u, '');
  try {
    const response = await fetch(`${origin}/api/version`, { signal: AbortSignal.timeout(5_000) });
    if (response.status !== 200) return null;
    const parsed = (await response.json()) as { version?: unknown };
    return typeof parsed.version === 'string' && parsed.version ? parsed.version : null;
  } catch {
    return null;
  }
}

function parseVersion(raw: string): number[] | null {
  const match = raw.trim().match(/^(\d+)\.(\d+)(?:\.(\d+))?/u);
  if (!match) return null;
  return [Number(match[1]), Number(match[2]), Number(match[3] ?? 0)];
}

function versionLt(a: number[], b: readonly number[]): boolean {
  for (let index = 0; index < 3; index += 1) {
    if (a[index] !== b[index]) return a[index] < b[index];
  }
  return false;
}

function probeResult(partial: Partial<EndpointProbeResult> & { protocol: EndpointProtocol; detail: string }): EndpointProbeResult {
  return {
    responsesStatus: null,
    chatStatus: null,
    ollamaVersion: null,
    ollamaNeedsUpgrade: false,
    ...partial,
  };
}

/**
 * 探测端点协议能力（四态判定）。
 *
 * Codex 0.155+ 只会发 Responses API：responses → 直通；chat-only → 内置桥接转换；
 * 其余两种拒绝。判定顺序：
 *  1. POST /responses（用**真实模型名**——Ollama 对未拉取的模型返回 404，不能直接判端点不支持）；
 *  2. 404/405 且是 Ollama → 查 /api/version 消歧：≥0.13.3 说明 404 只是模型没拉（判 responses），
 *     低于 0.13.3 直接判 chat-only（/chat/completions 必有）；
 *  3. 否则再 POST /chat/completions：路由存在判 chat-only，也不存在判 unsupported。
 */
export async function probeEndpoint(params: {
  baseUrl: string;
  modelId?: string | null;
  apiKey?: string | null;
  providerType?: string | null;
  builtinProvider?: string | null;
}): Promise<EndpointProbeResult> {
  const baseUrl = params.baseUrl.replace(/\/+$/u, '');
  const model = params.modelId?.trim() || 'probe';
  const isOllama =
    params.builtinProvider === 'ollama' ||
    `${params.providerType ?? ''}`.trim().toUpperCase() === 'OLLAMA';

  const responses = await postProbe(
    `${baseUrl}/responses`,
    { model, input: 'ping', max_output_tokens: 16, stream: false },
    params.apiKey,
  );
  if (responses.status === null) {
    return probeResult({ protocol: 'unreachable', detail: `端点不可达：${responses.error ?? '网络错误'}` });
  }
  if (responses.status !== 404 && responses.status !== 405) {
    return probeResult({
      protocol: 'responses',
      responsesStatus: responses.status,
      detail: `端点已实现 Responses API（HTTP ${responses.status}）`,
    });
  }

  if (isOllama) {
    const version = await probeOllamaVersion(baseUrl);
    if (version) {
      const parsed = parseVersion(version);
      if (parsed && !versionLt(parsed, OLLAMA_RESPONSES_MIN_VERSION)) {
        return probeResult({
          protocol: 'responses',
          responsesStatus: responses.status,
          ollamaVersion: version,
          detail: `Ollama ${version} 支持 Responses API；/responses 返回 404 通常只是模型「${model}」未拉取，请先 ollama pull ${model}`,
        });
      }
      return probeResult({
        protocol: 'chat-only',
        responsesStatus: responses.status,
        ollamaVersion: version,
        ollamaNeedsUpgrade: true,
        detail: `Ollama ${version} 低于 0.13.3，只提供 /chat/completions（建议升级 Ollama 以直连）`,
      });
    }
  }

  const chat = await postProbe(
    `${baseUrl}/chat/completions`,
    { model, messages: [{ role: 'user', content: 'ping' }], max_tokens: 1, stream: false },
    params.apiKey,
  );
  if (chat.status !== null && chat.status !== 404 && chat.status !== 405) {
    return probeResult({
      protocol: 'chat-only',
      responsesStatus: responses.status,
      chatStatus: chat.status,
      detail: `端点只提供 /chat/completions（/responses 返回 HTTP ${responses.status}）`,
    });
  }
  return probeResult({
    protocol: 'unsupported',
    responsesStatus: responses.status,
    chatStatus: chat.status,
    detail: `端点 Responses 与 Chat Completions 均不存在（HTTP ${responses.status}/${chat.status ?? '网络错误'}），无法对接`,
  });
}

export async function readAppliedProvider(): Promise<AppliedProvider | null> {
  try {
    const raw = JSON.parse(await readFile(providerFile(), 'utf8')) as Partial<AppliedProvider>;
    return raw.model ? (raw as AppliedProvider) : null;
  } catch {
    return null;
  }
}

async function writeAppliedProvider(value: AppliedProvider | null): Promise<void> {
  await mkdir(getCodexDataDir(), { recursive: true });
  await writeFile(providerFile(), JSON.stringify(value ?? {}, null, 2), 'utf8');
}

/** 应用一条模型配置：先探端点，再重启子进程（api_key 走 env，换 key 必须重启） */
export async function applyProvider(
  runtime: CodexRuntime,
  input: ApplyProviderInput & { modelRecordId?: string | number | null },
): Promise<AppliedProvider> {
  const spec = toProviderSpec(input);
  // 只要显式配了地址就探（含远端 Ollama）：应用成功后才发现连不上，比这里直接报错更难排查。
  // 内置 provider 没配地址时不探（默认 localhost，可能压根没起本地服务）。
  if (spec.baseUrl) {
    const probe = await probeEndpoint({
      baseUrl: spec.baseUrl,
      modelId: spec.model,
      apiKey: spec.apiKey,
      providerType: input.providerType,
      builtinProvider: spec.builtinProvider,
    });
    if (probe.protocol === 'chat-only') {
      throw new Error(`${probe.detail}。内置桥接转换尚未启用，当前请先升级端点（Ollama 需 ≥ 0.13.3）`);
    }
    if (probe.protocol !== 'responses') throw new Error(probe.detail);
  }
  await runtime.applyProvider(spec);
  const applied: AppliedProvider = {
    model: spec.model,
    modelRecordId: input.modelRecordId === undefined || input.modelRecordId === null
      ? null
      : String(input.modelRecordId),
    baseUrl: spec.baseUrl ?? null,
    providerId: spec.id,
    builtinProvider: spec.builtinProvider ?? null,
    name: spec.name ?? null,
    appliedAt: new Date().toISOString(),
  };
  await writeAppliedProvider(applied);
  return applied;
}

export async function clearProvider(runtime: CodexRuntime): Promise<void> {
  await runtime.applyProvider(null);
  await writeAppliedProvider(null);
}