metona-ai-desktop/electron/harness/adapters/__tests__/deepseek-vision.test.ts

/**
 * DeepSeekAdapter 多模态（vision 模型）请求格式测试（v0.5.4）
 *
 * 背景：DeepSeek 新增 vision 实验模型 deepseek-v4-flash-vision-exp
 * （OpenAI image_url content parts 格式）。适配器行为：
 * - vision 模型：带 images 的消息 content 转换为 [{type:'text'},{type:'image_url'}] parts
 * - 非 vision 模型：images 静默丢弃（共享层行为，防 API 400）
 *
 * 测试策略（契约级）：mock fetch 记录真实请求体断言。
 */

import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';

vi.mock('electron-log', () => ({
  default: { info: vi.fn(), warn: vi.fn(), error: vi.fn(), debug: vi.fn() },
}));

import { DeepSeekAdapter } from '../deepseek.adapter';
import type { MetonaRequest } from '../../types';

const mockFetch = vi.fn();
vi.stubGlobal('fetch', mockFetch);

beforeEach(() => {
  mockFetch.mockReset();
});
afterEach(() => {
  mockFetch.mockReset();
});

function makeAdapter(model: string): DeepSeekAdapter {
  return new DeepSeekAdapter({
    provider: 'deepseek',
    baseURL: 'https://api.deepseek.com',
    apiKey: 'sk-test',
    defaultModel: model,
  });
}

function makeRequest(images?: Array<{ url: string }>): MetonaRequest {
  return {
    meta: {
      sessionId: 's',
      iteration: 1,
      requestId: 'r',
      timestamp: Date.now(),
      agentVersion: '1',
    },
    systemPrompt: { roleDefinition: 'sys', outputConstraints: '', safetyGuidelines: '' },
    messages: [{ role: 'user', content: '这张图片里有什么？', images, timestamp: Date.now() }],
    params: { temperature: 0, stream: false, thinkingEnabled: false },
  };
}

function okResponse(): Response {
  return {
    ok: true,
    status: 200,
    json: async () => ({ choices: [{ message: { content: 'ok' } }], usage: {} }),
  } as unknown as Response;
}

function requestBody(): { model: string; messages: Array<Record<string, unknown>> } {
  const [, init] = mockFetch.mock.calls[0] as [string, RequestInit];
  return JSON.parse(init.body as string);
}

describe('DeepSeek vision 模型多模态请求格式（v0.5.4）', () => {
  it('vision 模型：带图片的消息转换为 image_url content parts', async () => {
    mockFetch.mockResolvedValue(okResponse());
    const adapter = makeAdapter('deepseek-v4-flash-vision-exp');

    await adapter.send(makeRequest([{ url: 'data:image/jpeg;base64,TESTPIC' }]));

    const body = requestBody();
    expect(body.model).toBe('deepseek-v4-flash-vision-exp');
    // system 消息 + user 消息（含 content parts）
    expect(body.messages).toHaveLength(2);
    const userMsg = body.messages[1];
    expect(userMsg.role).toBe('user');
    expect(Array.isArray(userMsg.content)).toBe(true);
    const parts = userMsg.content as Array<Record<string, unknown>>;
    expect(parts[0]).toEqual({ type: 'text', text: '这张图片里有什么？' });
    expect(parts[1]).toEqual({
      type: 'image_url',
      image_url: { url: 'data:image/jpeg;base64,TESTPIC' },
    });
  });

  it('非 vision 模型：images 被静默丢弃（content 保持纯文本）', async () => {
    mockFetch.mockResolvedValue(okResponse());
    const adapter = makeAdapter('deepseek-v4-pro');

    await adapter.send(makeRequest([{ url: 'data:image/jpeg;base64,TESTPIC' }]));

    const body = requestBody();
    const userMsg = body.messages[1];
    // 非 vision 模型 content 保持字符串（不转 parts，不发图片 → 不会 400）
    expect(userMsg.content).toBe('这张图片里有什么？');
  });

  it('vision 模型 max_tokens 钳制到 8192（MODEL_INFO 上限）', async () => {
    mockFetch.mockResolvedValue(okResponse());
    const adapter = makeAdapter('deepseek-v4-flash-vision-exp');

    await adapter.send({
      ...makeRequest(),
      params: { maxTokens: 63_488, temperature: 0, stream: false },
    } as MetonaRequest);

    const body = JSON.parse((mockFetch.mock.calls[0] as [string, RequestInit])[1].body as string);
    expect(body.max_tokens).toBe(8_192);
  });

  it('vision 模型无图片时不转换（content 保持纯文本）', async () => {
    mockFetch.mockResolvedValue(okResponse());
    const adapter = makeAdapter('deepseek-v4-flash-vision-exp');

    await adapter.send(makeRequest(undefined));

    const body = requestBody();
    const userMsg = body.messages[1];
    expect(userMsg.content).toBe('这张图片里有什么？');
  });
});