SRCmail/lib/ai/__tests__/local-discovery.test.ts

import { describe, expect, it, vi, beforeEach, afterEach } from 'vitest';
import {
  discoverLocalOllama,
  recommendDefaultModel,
  largestModel,
  isLocalDiscoveryDismissed,
  dismissLocalDiscovery,
  type DiscoveredLocalModel,
} from '../local-discovery';

/**
 * Real model list from this machine's Ollama (`curl 127.0.0.1:11434/api/tags`,
 * 2026-08-06) — used as the test fixture rather than invented data, per the
 * explicit instruction to use the real local runtime as the test case for
 * "which queries are required and how to add most of the modules
 * automatically". Sizes/params/capabilities are copied verbatim.
 */
const REAL_MACHINE_MODELS: DiscoveredLocalModel[] = [
  { name: 'nomic-embed-text:latest', capabilities: ['embedding'], parameterSize: '137M', sizeBytes: 274_302_450 },
  { name: 'qwen2.5:32b', capabilities: ['completion', 'tools'], parameterSize: '32.8B', sizeBytes: 19_851_349_669 },
  { name: 'gemma4:12b-mlx', capabilities: ['completion', 'tools', 'thinking'], parameterSize: '', sizeBytes: 9_977_519_169 },
  { name: 'gemma4:latest', capabilities: ['completion', 'tools', 'thinking'], parameterSize: '8.0B', sizeBytes: 9_608_350_718 },
  { name: 'deepseek-r1:32b', capabilities: ['completion', 'thinking'], parameterSize: '32.8B', sizeBytes: 19_851_337_809 },
  { name: 'deepseek-r1:14b', capabilities: ['completion', 'thinking'], parameterSize: '14.8B', sizeBytes: 8_988_112_209 },
  { name: 'llama3.2:latest', capabilities: ['completion', 'tools'], parameterSize: '3.2B', sizeBytes: 2_019_393_189 },
  { name: 'hermes3:8b', capabilities: ['completion', 'tools'], parameterSize: '8B', sizeBytes: 4_661_227_000 },
  { name: 'qwen3:latest', capabilities: ['completion', 'tools', 'thinking'], parameterSize: '8.2B', sizeBytes: 5_200_000_000 },
  { name: 'gemma4:e4b', capabilities: ['completion', 'tools', 'thinking'], parameterSize: '8.0B', sizeBytes: 9_600_000_000 },
  { name: 'qwen3.5:latest', capabilities: ['vision', 'completion', 'tools', 'thinking'], parameterSize: '9.7B', sizeBytes: 6_600_000_000 },
];

describe('recommendDefaultModel', () => {
  it('picks the smallest non-"thinking" chat model from a real mixed fleet', () => {
    // llama3.2 (2.0GB) is the smallest completion-capable, non-reasoning
    // model on this real machine — everything smaller is embedding-only.
    expect(recommendDefaultModel(REAL_MACHINE_MODELS)).toBe('llama3.2:latest');
  });

  it('never recommends an embedding-only model', () => {
    const onlyEmbedding = [REAL_MACHINE_MODELS[0]]; // nomic-embed-text
    expect(recommendDefaultModel(onlyEmbedding)).toBeNull();
  });

  it('falls back to the smallest "thinking" model when nothing else qualifies', () => {
    const onlyReasoning = REAL_MACHINE_MODELS.filter((m) => m.capabilities.includes('thinking') && !m.capabilities.includes('vision'));
    // Smallest of the thinking-only pool here is qwen3 (5.2GB) before gemma4 variants.
    expect(recommendDefaultModel(onlyReasoning)).toBe('qwen3:latest');
  });

  it('returns null when no models are chat-capable at all', () => {
    expect(recommendDefaultModel([])).toBeNull();
  });
});

describe('largestModel', () => {
  it('picks the biggest chat-capable model — qwen2.5:32b, by 11,860 bytes over deepseek-r1:32b', () => {
    // Both are ~19.85GB on this real machine (same base size class), but
    // qwen2.5:32b's actual manifest is very slightly larger — not a tie.
    expect(largestModel(REAL_MACHINE_MODELS)).toBe('qwen2.5:32b');
  });

  it('excludes embedding-only models even though they can be tiny or huge', () => {
    expect(largestModel([REAL_MACHINE_MODELS[0]])).toBeNull();
  });
});

describe('discoverLocalOllama', () => {
  const originalFetch = global.fetch;

  afterEach(() => {
    global.fetch = originalFetch;
    vi.restoreAllMocks();
  });

  it('parses a real-shaped /api/tags response into DiscoveredLocalModel[]', async () => {
    global.fetch = vi.fn().mockResolvedValue({
      ok: true,
      json: async () => ({
        models: [
          { name: 'llama3.2:latest', capabilities: ['completion', 'tools'], size: 2_019_393_189, details: { parameter_size: '3.2B' } },
          { name: 'nomic-embed-text:latest', capabilities: ['embedding'], size: 274_302_450, details: { parameter_size: '137M' } },
        ],
      }),
    }) as unknown as typeof fetch;

    const result = await discoverLocalOllama(['http://127.0.0.1:11434']);
    expect(result).not.toBeNull();
    expect(result?.baseUrl).toBe('http://127.0.0.1:11434');
    expect(result?.models).toHaveLength(2);
    expect(result?.models[0]).toEqual({
      name: 'llama3.2:latest',
      capabilities: ['completion', 'tools'],
      parameterSize: '3.2B',
      sizeBytes: 2_019_393_189,
    });
  });

  it('requires exactly one query — a single /api/tags call, no follow-up /api/show requests', async () => {
    const fetchMock = vi.fn().mockResolvedValue({
      ok: true,
      json: async () => ({ models: [{ name: 'llama3.2:latest', capabilities: ['completion'], size: 1, details: {} }] }),
    });
    global.fetch = fetchMock as unknown as typeof fetch;

    await discoverLocalOllama(['http://127.0.0.1:11434']);
    expect(fetchMock).toHaveBeenCalledTimes(1);
    expect(fetchMock).toHaveBeenCalledWith('http://127.0.0.1:11434/api/tags', expect.anything());
  });

  it('falls through to the next candidate base URL when the first is unreachable', async () => {
    const fetchMock = vi.fn()
      .mockRejectedValueOnce(new Error('connection refused'))
      .mockResolvedValueOnce({
        ok: true,
        json: async () => ({ models: [{ name: 'llama3.2:latest', capabilities: ['completion'], size: 1, details: {} }] }),
      });
    global.fetch = fetchMock as unknown as typeof fetch;

    const result = await discoverLocalOllama(['http://127.0.0.1:11434', 'http://localhost:11434']);
    expect(result?.baseUrl).toBe('http://localhost:11434');
    expect(fetchMock).toHaveBeenCalledTimes(2);
  });

  it('returns null when nothing answers on any candidate', async () => {
    global.fetch = vi.fn().mockRejectedValue(new Error('connection refused')) as unknown as typeof fetch;
    const result = await discoverLocalOllama(['http://127.0.0.1:11434', 'http://localhost:11434']);
    expect(result).toBeNull();
  });

  it('returns null (not an empty-models result) when Ollama answers with zero models installed', async () => {
    global.fetch = vi.fn().mockResolvedValue({ ok: true, json: async () => ({ models: [] }) }) as unknown as typeof fetch;
    const result = await discoverLocalOllama(['http://127.0.0.1:11434']);
    expect(result).toBeNull();
  });
});

describe('dismissal persistence', () => {
  beforeEach(() => {
    window.localStorage.clear();
  });

  it('is not dismissed by default', () => {
    expect(isLocalDiscoveryDismissed()).toBe(false);
  });

  it('persists a dismissal across calls', () => {
    dismissLocalDiscovery();
    expect(isLocalDiscoveryDismissed()).toBe(true);
  });
});