breakpilot-compliance/admin-compliance/lib/sdk/drafting-engine/__tests__/llm-cascade.test.ts

/**
 * Tests fuer die Drafting-Engine LLM-Kaskade (OVH -> Ollama) + Stream-Parser.
 */

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

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

describe('llm-cascade parser', () => {
  it('parseOllamaLine extrahiert message.content', async () => {
    const { parseOllamaLine } = await import('../llm-cascade')
    expect(parseOllamaLine('{"message":{"content":"X"}}')).toBe('X')
    expect(parseOllamaLine('')).toBeNull()
    expect(parseOllamaLine('kaputt')).toBeNull()
  })

  it('parseSSELine extrahiert delta.content', async () => {
    const { parseSSELine } = await import('../llm-cascade')
    expect(parseSSELine('data: {"choices":[{"delta":{"content":"Y"}}]}')).toBe('Y')
    expect(parseSSELine('data: [DONE]')).toBeNull()
    expect(parseSSELine('event: ping')).toBeNull()
  })
})

describe('cascadeComplete', () => {
  beforeEach(() => {
    vi.resetModules()
    mockFetch.mockReset()
  })
  afterEach(() => {
    vi.unstubAllEnvs()
  })

  it('nutzt OVH zuerst wenn konfiguriert (json + response_format)', async () => {
    vi.stubEnv('OVH_LLM_URL', 'https://ovh.test')
    vi.stubEnv('OVH_LLM_MODEL', 'gpt-oss-120b')
    vi.stubEnv('OVH_LLM_KEY', 'k')
    const { cascadeComplete } = await import('../llm-cascade')
    mockFetch.mockResolvedValueOnce({
      ok: true,
      json: async () => ({ choices: [{ message: { content: '{"ok":1}' } }], usage: { completion_tokens: 42 } }),
    })
    const r = await cascadeComplete([{ role: 'user', content: 'hi' }], { json: true, maxTokens: 1000 })
    expect(r).toEqual({ content: '{"ok":1}', tokensUsed: 42, provider: 'ovh' })
    const [url, opts] = mockFetch.mock.calls[0]
    expect(url).toContain('/v1/chat/completions')
    const body = JSON.parse(opts.body)
    expect(body.response_format).toEqual({ type: 'json_object' })
    expect(body.stream).toBe(false)
  })

  it('faellt auf Ollama zurueck wenn OVH nicht konfiguriert ist', async () => {
    const { cascadeComplete } = await import('../llm-cascade')
    mockFetch.mockResolvedValueOnce({
      ok: true,
      json: async () => ({ message: { content: 'hallo' }, eval_count: 7 }),
    })
    const r = await cascadeComplete([{ role: 'user', content: 'hi' }])
    expect(r).toEqual({ content: 'hallo', tokensUsed: 7, provider: 'ollama' })
    expect(mockFetch.mock.calls[0][0]).toContain('/api/chat')
  })

  it('faellt auf Ollama zurueck wenn OVH einen Fehler liefert', async () => {
    vi.stubEnv('OVH_LLM_URL', 'https://ovh.test')
    vi.stubEnv('OVH_LLM_MODEL', 'gpt-oss-120b')
    const { cascadeComplete } = await import('../llm-cascade')
    mockFetch
      .mockResolvedValueOnce({ ok: false, status: 502 })
      .mockResolvedValueOnce({ ok: true, json: async () => ({ message: { content: 'fallback' }, eval_count: 3 }) })
    const r = await cascadeComplete([{ role: 'user', content: 'hi' }])
    expect(r?.provider).toBe('ollama')
    expect(r?.content).toBe('fallback')
  })

  it('liefert null wenn weder OVH noch Ollama antworten', async () => {
    const { cascadeComplete } = await import('../llm-cascade')
    mockFetch.mockResolvedValue({ ok: false, status: 500 })
    expect(await cascadeComplete([{ role: 'user', content: 'hi' }])).toBeNull()
  })
})