V10.556 補 GCP-B 模型 fallback 防 111 過載

tests/test_ollama_retry_chain.py

@@ -167,6 +167,50 @@ def test_generate_forces_final_fallback_when_unhealthy_ttl_expires_mid_request()
     assert 'all 3 hosts failed' in (resp.error or '')
 
 
+def test_generate_uses_secondary_model_fallback_before_111():
+    """GCP-B 缺 coder 模型時，先改用 secondary fallback，不直接推到 111。"""
+    from services import ollama_service as oss
+    from services.ollama_service import OllamaService
+
+    svc = OllamaService()
+    fake_ok = MagicMock(status_code=200)
+    fake_ok.json.return_value = {
+        'response': 'OK from secondary fallback',
+        'prompt_eval_count': 10,
+        'eval_count': 5,
+    }
+
+    with patch('services.ollama_service.resolve_ollama_host', return_value=oss.OLLAMA_HOST_SECONDARY), \
+         patch('services.ollama_service.requests.post', return_value=fake_ok) as mock_post:
+        resp = svc.generate('test prompt', model='qwen2.5-coder:7b')
+
+    assert resp.success is True
+    assert resp.host == oss.OLLAMA_HOST_SECONDARY
+    assert resp.model == 'gemma3:4b'
+    payload = mock_post.call_args.kwargs['json']
+    assert payload['model'] == 'gemma3:4b'
+
+
+def test_generate_model_404_does_not_mark_host_unhealthy():
+    """模型不存在是 model availability，不應把整台 GCP-B 標成 unhealthy。"""
+    from services import ollama_service as oss
+    from services.ollama_service import OllamaService
+
+    svc = OllamaService()
+    fake_404 = MagicMock(status_code=404, text='{"error":"model not found"}')
+    fake_ok = MagicMock(status_code=200)
+    fake_ok.json.return_value = {'response': 'OK', 'prompt_eval_count': 3, 'eval_count': 2}
+
+    with patch('services.ollama_service.resolve_ollama_host', side_effect=[
+        oss.OLLAMA_HOST_SECONDARY,
+        oss.OLLAMA_HOST_FALLBACK,
+    ]), patch('services.ollama_service.requests.post', side_effect=[fake_404, fake_ok]):
+        resp = svc.generate('test prompt', model='tiny-missing-model')
+
+    assert resp.success is True
+    assert oss.OLLAMA_HOST_SECONDARY not in oss._unhealthy_marks
+
+
 def test_generate_can_disable_111_fallback_for_batch_llm_work():
     """批量 LLM 任務可選擇只跑 GCP-A/GCP-B，避免 111 承接長分析。"""
     import requests