feat(p13): OllamaResponse token 補欄 + COST_TABLE 補新模型 + retry 鏈 unit test

Operation Ollama-First v5.0 / Phase 13 補強

(A) services/ollama_service.py — OllamaResponse 加 input_tokens/output_tokens
- A4 Phase 1 已知 limitation 修補：openclaw_bot_main token=0 假數據誤導日報
- generate() 解 prompt_eval_count + eval_count 寫 OllamaResponse
- 影響：ai_call_logger 收到正確 token 數，token 日報 Ollama 占比準確

(B) services/ai_call_logger.py — COST_TABLE 補 GCP 已拉/候選模型
- qwen2.5:7b-instruct (Phase 3 A7 OpenClaw Q&A 預設)
- qwen3:14b (Phase 3 A9 Nemotron + A7 升級候選)
- qwen2.5-coder:32b (Phase 8 候選)
- qwen2-vl:7b (Phase 13+ PPT vision 候選)
- deepseek-r1:14b / gemma3:4b (推理增強 / 輕量)
- 全部 cost=0（Ollama 自架）
- 解 logger.warning「unknown model cost」誤報

(J) tests/test_ollama_retry_chain.py (10 unit tests) — 驗 hotfix e862a90/6572d52
- T1 self.host @property lazy resolve
- T2 explicit host 凍結不 retry
- T3 generate 第一台 timeout → 第二台成功（核心 retry 鏈）
- T4 三主機都失敗 → success=False
- T5 cache 卡同主機 → break 不無限迴圈
- T6 Phase 13 token 解析驗證
- T7-T9 generate_embedding 同類驗證
- T10 mark_unhealthy 清 resolve cache

regression: 全戰役 14 test 檔仍 zero regression

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

services/ai_call_logger.py

@@ -55,11 +55,17 @@ COST_TABLE: Dict[str, Dict[str, float]] = {
     'claude-opus-4-7':    {'in': 15.0, 'out': 75.0},     # 程式碼 #1，Arena Elo 1548
     'claude-sonnet-4-6':  {'in': 3.0,  'out': 15.0},     # agentic 平衡
     'claude-haiku-4-5':   {'in': 0.8,  'out': 4.0},      # 輕量快速
-    # Ollama 自架 (全 0)
-    'hermes3:latest':     {'in': 0.0, 'out': 0.0},
-    'qwen2.5-coder:7b':   {'in': 0.0, 'out': 0.0},
-    'llama3.1:8b':        {'in': 0.0, 'out': 0.0},
-    'bge-m3:latest':      {'in': 0.0, 'out': 0.0},
+    # Ollama 自架 (全 0；Phase 8/13 GCP 拉模型陸續加入)
+    'hermes3:latest':       {'in': 0.0, 'out': 0.0},
+    'qwen2.5-coder:7b':     {'in': 0.0, 'out': 0.0},
+    'qwen2.5-coder:32b':    {'in': 0.0, 'out': 0.0},  # Phase 8 GCP 候選（待拉）
+    'qwen2.5:7b-instruct':  {'in': 0.0, 'out': 0.0},  # Phase 3 A7 OpenClaw Q&A 預設（GCP 已拉）
+    'qwen3:14b':            {'in': 0.0, 'out': 0.0},  # Phase 3 A9 Nemotron + A7 升級（GCP 已拉）
+    'qwen2-vl:7b':          {'in': 0.0, 'out': 0.0},  # Phase 13 PPT vision 候選
+    'deepseek-r1:14b':      {'in': 0.0, 'out': 0.0},  # 推理增強候選
+    'gemma3:4b':            {'in': 0.0, 'out': 0.0},  # 輕量 sales_copy 候選
+    'llama3.1:8b':          {'in': 0.0, 'out': 0.0},
+    'bge-m3:latest':        {'in': 0.0, 'out': 0.0},
 }
 
 

services/ollama_service.py

@@ -150,13 +150,21 @@ def get_host_label(host: str) -> str:
 
 @dataclass
 class OllamaResponse:
-    """Ollama 回應結構"""
+    """Ollama 回應結構
+
+    Phase 13 補強（A4 已知 limitation 修補）：
+      新增 input_tokens / output_tokens 欄位，由 /api/generate 回應的
+      prompt_eval_count + eval_count 解析。解 ai_call_logger 的
+      openclaw_bot_main token=0 假數據問題（誤導 token 日報統計）。
+    """
     success: bool
     content: str
     model: str
     error: Optional[str] = None
     total_duration: Optional[float] = None
     host: Optional[str] = None
+    input_tokens: int = 0     # prompt_eval_count
+    output_tokens: int = 0    # eval_count
 
 
 class OllamaService:
@@ -264,6 +272,9 @@ class OllamaService:
                         model=model,
                         total_duration=data.get('total_duration', 0) / 1e9,
                         host=current_host,
+                        # Phase 13 補強：解 token=0 假數據（A4 已知 limitation）
+                        input_tokens=int(data.get('prompt_eval_count', 0) or 0),
+                        output_tokens=int(data.get('eval_count', 0) or 0),
                     )
                 # HTTP 非 200：標 unhealthy + 嘗試下一台
                 last_error = f"HTTP {response.status_code}: {response.text[:200]}"

tests/test_ollama_retry_chain.py

@@ -0,0 +1,251 @@
+"""
+tests/test_ollama_retry_chain.py
+─────────────────────────────────────────────────────────────────
+Operation Ollama-First v5.0 / Phase 13 補強 — 三主機 retry 鏈驗證
+
+驗證 hotfix e862a90 (generate retry) + 6572d52 (embed retry) 的邏輯：
+  1. self.host @property lazy resolve
+  2. generate / generate_embedding 失敗自動 retry 下一台主機（最多 3 次）
+  3. mark_unhealthy 後下次 self.host 取新主機
+  4. caller 顯式 host=... 凍結不 retry
+  5. 三主機都失敗 → return failure（不無限迴圈）
+  6. Phase 13 補強：OllamaResponse.input_tokens/output_tokens 解析
+
+紀律：不打真實 Ollama，全 mock requests.post。
+"""
+
+from unittest.mock import patch, MagicMock
+
+import pytest
+
+
+# ═══════════════════════════════════════════════════════════════════════════
+# Fixtures
+# ═══════════════════════════════════════════════════════════════════════════
+
+@pytest.fixture(autouse=True)
+def _reset_state():
+    """每 test 清 unhealthy marks + resolve cache"""
+    import services.ollama_service as oss
+    oss._unhealthy_marks.clear()
+    oss._resolved_host_cache['host'] = None
+    oss._resolved_host_cache['ts'] = 0
+    yield
+    oss._unhealthy_marks.clear()
+    oss._resolved_host_cache['host'] = None
+    oss._resolved_host_cache['ts'] = 0
+
+
+# ═══════════════════════════════════════════════════════════════════════════
+# T1: self.host 改 @property（不再凍結 instance）
+# ═══════════════════════════════════════════════════════════════════════════
+
+def test_self_host_is_lazy_property_not_frozen():
+    """OllamaService.host 應為 property，每次存取走 resolve_ollama_host()"""
+    from services.ollama_service import OllamaService
+
+    svc = OllamaService()
+    # property 不在 __init__ 凍結
+    assert svc._explicit_host is None
+
+    # 存取 .host 走 resolve
+    with patch('services.ollama_service.resolve_ollama_host', return_value='http://test:11434') as mock_resolve:
+        h1 = svc.host
+        h2 = svc.host
+    assert h1 == h2 == 'http://test:11434'
+    assert mock_resolve.call_count >= 2  # 每次 access 都呼叫
+
+
+def test_explicit_host_freezes():
+    """caller 顯式傳 host → _explicit_host 凍結，不走 lazy resolve"""
+    from services.ollama_service import OllamaService
+
+    svc = OllamaService(host='http://explicit:11434')
+    assert svc._explicit_host == 'http://explicit:11434'
+
+    with patch('services.ollama_service.resolve_ollama_host') as mock_resolve:
+        h = svc.host
+    assert h == 'http://explicit:11434'
+    mock_resolve.assert_not_called()
+
+
+# ═══════════════════════════════════════════════════════════════════════════
+# T2: generate 三主機 retry 鏈（核心 hotfix e862a90）
+# ═══════════════════════════════════════════════════════════════════════════
+
+def test_generate_retries_on_first_host_timeout():
+    """第一台 timeout → mark_unhealthy → 第二台成功"""
+    import requests
+    from services.ollama_service import OllamaService
+
+    svc = OllamaService()
+
+    # 序列：第一次 timeout，第二次成功
+    fake_ok = MagicMock(status_code=200)
+    fake_ok.json.return_value = {
+        'response': 'OK from secondary',
+        'prompt_eval_count': 100, 'eval_count': 50,
+    }
+    call_results = [requests.Timeout('first host'), fake_ok]
+
+    # mock host property 回傳兩台不同主機
+    hosts = ['http://host1:11434', 'http://host2:11434']
+
+    with patch('services.ollama_service.resolve_ollama_host', side_effect=hosts), \
+         patch('services.ollama_service.requests.post', side_effect=call_results) as mock_post:
+        resp = svc.generate('test prompt')
+
+    assert resp.success is True
+    assert resp.content == 'OK from secondary'
+    assert resp.input_tokens == 100
+    assert resp.output_tokens == 50
+    assert mock_post.call_count == 2  # 第一次 timeout, 第二次成功
+
+
+def test_generate_all_three_hosts_fail_returns_failure():
+    """三主機都失敗 → success=False，不無限迴圈"""
+    import requests
+    from services.ollama_service import OllamaService
+
+    svc = OllamaService()
+    hosts = ['http://h1:11434', 'http://h2:11434', 'http://h3:11434']
+
+    with patch('services.ollama_service.resolve_ollama_host', side_effect=hosts), \
+         patch('services.ollama_service.requests.post',
+               side_effect=requests.Timeout('all timeout')) as mock_post:
+        resp = svc.generate('test')
+
+    assert resp.success is False
+    assert mock_post.call_count == 3  # 試了 3 次
+    assert 'all 3 hosts failed' in (resp.error or '')
+
+
+def test_generate_same_host_breaks_loop():
+    """resolve 連 3 次回同主機（cache 沒過期）→ 不無限迴圈"""
+    import requests
+    from services.ollama_service import OllamaService
+
+    svc = OllamaService()
+    same_host = 'http://stuck:11434'
+
+    with patch('services.ollama_service.resolve_ollama_host', return_value=same_host), \
+         patch('services.ollama_service.requests.post',
+               side_effect=requests.Timeout('stuck')) as mock_post:
+        resp = svc.generate('test')
+
+    assert resp.success is False
+    # 第一次試 stuck → 失敗 → 第二次 self.host 還是 stuck（cache 還在）→ break
+    assert mock_post.call_count == 1
+
+
+def test_generate_token_parsing_phase13():
+    """Phase 13 補強：OllamaResponse 解 prompt_eval_count + eval_count"""
+    from services.ollama_service import OllamaService
+
+    svc = OllamaService()
+    fake_resp = MagicMock(status_code=200)
+    fake_resp.json.return_value = {
+        'response': 'hello',
+        'prompt_eval_count': 250,
+        'eval_count': 80,
+        'total_duration': 1500000000,  # 1.5s in nanoseconds
+    }
+
+    with patch('services.ollama_service.resolve_ollama_host', return_value='http://x:11434'), \
+         patch('services.ollama_service.requests.post', return_value=fake_resp):
+        resp = svc.generate('test')
+
+    assert resp.success is True
+    assert resp.input_tokens == 250
+    assert resp.output_tokens == 80
+    assert resp.total_duration == 1.5  # 轉換為秒
+
+
+# ═══════════════════════════════════════════════════════════════════════════
+# T3: generate_embedding 三主機 retry（hotfix 6572d52）
+# ═══════════════════════════════════════════════════════════════════════════
+
+def test_embedding_retries_on_first_host_timeout():
+    """embed 第一台 timeout → 第二台成功"""
+    import requests
+    from services.ollama_service import OllamaService
+
+    svc = OllamaService()
+    fake_ok = MagicMock(status_code=200)
+    fake_ok.json.return_value = {'embeddings': [[0.1, 0.2, 0.3]]}
+
+    hosts = ['http://h1:11434', 'http://h2:11434']
+    call_seq = [requests.Timeout('h1'), fake_ok]
+
+    with patch('services.ollama_service.resolve_ollama_host', side_effect=hosts), \
+         patch.dict('os.environ', {}, clear=False), \
+         patch('services.ollama_service.requests.post', side_effect=call_seq) as mock_post:
+        # 確保不走 EMBEDDING_HOST env
+        import os
+        os.environ.pop('EMBEDDING_HOST', None)
+        vec = svc.generate_embedding('test text')
+
+    assert vec == [0.1, 0.2, 0.3]
+    assert mock_post.call_count == 2  # 第二台成功
+
+
+def test_embedding_explicit_host_no_retry():
+    """caller 顯式 host=... → 不 retry，失敗即回 []"""
+    import requests
+    from services.ollama_service import OllamaService
+
+    svc = OllamaService()
+
+    with patch('services.ollama_service.requests.post',
+               side_effect=requests.Timeout('explicit host fail')) as mock_post:
+        vec = svc.generate_embedding('test', host='http://explicit:11434')
+
+    assert vec == []
+    # 顯式 host 凍結不 retry → 嘗試一次（/api/embed），可能再試 /api/embeddings legacy
+    # 但 retry 鏈不啟動（沒有 for-loop 迭代不同主機）
+    # 我們驗：post 呼叫次數 ≤ 2（最多主路徑+legacy 各一次，不會 retry 不同主機）
+    assert mock_post.call_count <= 2
+
+
+def test_embedding_all_three_hosts_fail_returns_empty():
+    """embed 三主機都失敗 → 回 []"""
+    import requests
+    from services.ollama_service import OllamaService
+
+    svc = OllamaService()
+    hosts = ['http://h1:11434', 'http://h2:11434', 'http://h3:11434']
+
+    with patch('services.ollama_service.resolve_ollama_host', side_effect=hosts), \
+         patch.dict('os.environ', {}, clear=False), \
+         patch('services.ollama_service.requests.post',
+               side_effect=requests.Timeout('all fail')) as mock_post:
+        import os
+        os.environ.pop('EMBEDDING_HOST', None)
+        vec = svc.generate_embedding('test')
+
+    assert vec == []
+    # 三主機 retry，每次主路徑+legacy = 6 次（但 legacy 也是 timeout）
+    # 實際看 _embed_one 邏輯：timeout 的 except 直接 mark_unhealthy 不試 legacy
+    # 所以是 3 次（主路徑 timeout × 3 主機）
+    assert mock_post.call_count == 3
+
+
+# ═══════════════════════════════════════════════════════════════════════════
+# T4: mark_unhealthy 觸發 cache 失效（驗 self.host 取新主機）
+# ═══════════════════════════════════════════════════════════════════════════
+
+def test_mark_unhealthy_invalidates_cache_for_next_host():
+    """generate 失敗 mark_unhealthy → 下次 self.host 走 resolve 取新主機"""
+    import services.ollama_service as oss
+    from services.ollama_service import OllamaService
+
+    svc = OllamaService()
+
+    # 先 cache 一個主機
+    oss._resolved_host_cache['host'] = 'http://primary:11434'
+    oss._resolved_host_cache['ts'] = __import__('time').time()
+
+    # mark_unhealthy 應清空 cache
+    oss.mark_unhealthy('http://primary:11434')
+
+    assert oss._resolved_host_cache['host'] is None