From 65d38d46325a28bd611552a53e9b93c51cf865a9 Mon Sep 17 00:00:00 2001
From: OoO <ooo@MacBook-Pro.local>
Date: Sun, 24 May 2026 13:32:26 +0800
Subject: [PATCH] V10.407 shorten ai runner residency

---
 .env.example                              |  3 +++
 config.py                                 |  2 +-
 docs/AI_INTELLIGENCE_MODULE_SOT.md        |  4 ++--
 docs/memory/history_logs.md               |  2 ++
 services/hermes_analyst_service.py        |  3 ++-
 services/marketplace_product_matcher.py   |  8 ++++++++
 services/openclaw_strategist_service.py   |  3 ++-
 tests/test_hermes_ollama_cascade.py       |  4 ++++
 tests/test_marketplace_product_matcher.py | 23 +++++++++++++++++++++++
 tests/test_qwen3_runtime_usage.py         |  3 +++
 10 files changed, 50 insertions(+), 5 deletions(-)

diff --git a/.env.example b/.env.example
index 9463535..3ee9f96 100644
--- a/.env.example
+++ b/.env.example
@@ -133,6 +133,8 @@ HERMES_URL=
 
 # [預設 120] Hermes 推理 timeout（秒）；批量 300 筆預估 ~90s
 HERMES_TIMEOUT=120
+# [預設 5m] Hermes runner 熱駐留；禁止回到 24h，避免 GCP-B/111 長駐高負載
+HERMES_KEEP_ALIVE=5m
 
 # [選填] Embedding 服務主機；留空時自動走同一條 Ollama 三主機級聯
 EMBEDDING_HOST=
@@ -385,6 +387,7 @@ NEMOTRON_OLLAMA_TIMEOUT=180
 OPENCLAW_STRATEGY_OLLAMA_MODEL=qwen3:14b
 OPENCLAW_STRATEGY_OLLAMA_TIMEOUT=90
 OPENCLAW_STRATEGY_OLLAMA_NUM_PREDICT=2048
+OPENCLAW_STRATEGY_OLLAMA_KEEP_ALIVE=5m
 
 # [預設 OFF] MCP Router；需先部署 docker-compose.mcp.yml 並完成健康檢查再開
 MCP_ROUTER_ENABLED=false
diff --git a/config.py b/config.py
index e55bc4e..8993933 100644
--- a/config.py
+++ b/config.py
@@ -325,7 +325,7 @@ YOUTUBE_API_KEY = os.getenv('YOUTUBE_API_KEY', '')
 # ==========================================
 # 系統版本與路徑
 # ==========================================
-SYSTEM_VERSION = "V10.406"
+SYSTEM_VERSION = "V10.408"
 LOG_FILE_PATH = os.path.join(BASE_DIR, 'logs/system.log')
 public_url = PUBLIC_URL  # 用於模板顯示
 
diff --git a/docs/AI_INTELLIGENCE_MODULE_SOT.md b/docs/AI_INTELLIGENCE_MODULE_SOT.md
index 109c2fa..06fe3dd 100644
--- a/docs/AI_INTELLIGENCE_MODULE_SOT.md
+++ b/docs/AI_INTELLIGENCE_MODULE_SOT.md
@@ -22,13 +22,13 @@
 - Code Review OpenClaw assessment 保持主機順序 GCP-A → GCP-B → 111，但可使用主機適配本地模型：GCP-A `qwen2.5-coder:7b`、GCP-B `gemma3:4b`；落到 111 時由 `OllamaService` 降級到 `llama3.2:latest`。Code Review 的 Ollama `keep_alive` 預設為 `5m`，不得再用 `24h` 長駐 runner 壓住 GCP-B/111。三段本地 Ollama 全失敗後才允許雲端備援。
 - OpenClaw Telegram Q&A 主路徑也不得綁單一 host：`_call_qwen3_qa()` 必須透過 `OllamaService` 跑 GCP-A → GCP-B → 111，並把實際落點寫入 `ai_calls.provider`。
 - OpenClaw Telegram 圖片商品辨識也必須 Ollama-first：`_identify_product_name_with_ollama_vision()` 透過 `OllamaService` 嘗試 GCP-A → GCP-B → 111；Gemini 只允許以 `openclaw_bot_image_gemini` caller 作為失敗後備援。
-- OpenClaw 週報、月報、Meta analysis、日報洞察、Telegram PPT 分析與 MCP fallback 也必須 Ollama-first；Gemini caller 只能帶 `_gemini_fallback` 或明確 fallback caller 語意，且不得先於 Ollama/NIM 被呼叫。
+- OpenClaw 週報、月報、Meta analysis、日報洞察、Telegram PPT 分析與 MCP fallback 也必須 Ollama-first；Gemini caller 只能帶 `_gemini_fallback` 或明確 fallback caller 語意，且不得先於 Ollama/NIM 被呼叫。OpenClaw strategy 的 Ollama `keep_alive` 預設為 `5m`，避免報告型任務把 GCP-B/111 runner 長駐 24h。
 - OpenClaw 週報、月報、Meta analysis、日報洞察與每日報告的 Gemini/NIM 備援 caller 必須登錄在 caller registry、AI 觀測台 agent group 與 Telegram 狀態統計，避免 fallback 用量被歸類為未知或漏算。
 - Gemini API 出站有第二道 kill switch：`GEMINI_FALLBACK_ENABLED` 預設為 `false`。即使 `GEMINI_API_KEY` 存在，通用 AI fallback、OpenClaw 報告/QA/PPT/圖片、MCP Grounding 與 Code Review L3 都不得呼叫 Gemini；只有操作員明確設為 `true` 時，Gemini 才能作緊急備援。
 - `docker-compose.yml` 的 `momo-app`、`scheduler`、`telegram-bot` 必須明確設定 `GEMINI_API_HARD_DISABLED=${GEMINI_API_HARD_DISABLED:-true}` 與 `GEMINI_FALLBACK_ENABLED=${GEMINI_FALLBACK_ENABLED:-false}`；`.env` 可保留 `GEMINI_API_KEY`，但不得因 key 存在就讓核心容器產生 Gemini 付費出站。
 - Gemini 不可被任何狀態面板或 router 推薦為主提供者：`AIProviderService._get_recommended_provider()` 不得回傳 `gemini`，只能顯示為 fallback 狀態；`llm_model_router` 的 `ea_engine` 若收到 `gemini-*` default 必須改回 `hermes3:latest`，需要深推理時才升本地 `deepseek-r1:14b`。
 - ElephantAlpha prompt / agent registry 不得再把 OpenClaw 描述為 Gemini 主模型；OpenClaw 是 `qwen2.5-coder:7b` / `qwen3:14b` Ollama-first 策略師，Gemini 僅能在 guard 顯式解鎖後作 emergency fallback。
-- 111 `192.168.0.111` 只是最後一道 Mac fallback，不承接 7B+、vision、long-context 模型長駐；`OllamaService.generate()` 落到 111 時會將 `qwen3`、`deepseek-r1`、`hermes3`、`qwen2.5*`、`gemma3`、`llava`、`minicpm-v` 與 7B+ 模型依 `OLLAMA_111_MODEL_DOWNGRADE_PATTERNS` 降級到 `OLLAMA_111_MODEL_FALLBACK=llama3.2:latest`，並以 `OLLAMA_111_KEEP_ALIVE=5m`、`OLLAMA_111_MAX_TIMEOUT=20`、`OLLAMA_111_NUM_CTX=4096`、`OLLAMA_111_NUM_PREDICT=512` 封頂，避免 16GB RAM 主機被大 context runner、長輸出與 24h keep-alive 壓到 swap。
+- 111 `192.168.0.111` 只是最後一道 Mac fallback，不承接 7B+、vision、long-context 模型長駐；`OllamaService.generate()` 落到 111 時會將 `qwen3`、`deepseek-r1`、`hermes3`、`qwen2.5*`、`gemma3`、`llava`、`minicpm-v` 與 7B+ 模型依 `OLLAMA_111_MODEL_DOWNGRADE_PATTERNS` 降級到 `OLLAMA_111_MODEL_FALLBACK=llama3.2:latest`，並以 `OLLAMA_111_KEEP_ALIVE=5m`、`OLLAMA_111_MAX_TIMEOUT=20`、`OLLAMA_111_NUM_CTX=4096`、`OLLAMA_111_NUM_PREDICT=512` 封頂。Hermes / OpenClaw / Code Review 路徑的業務 keep-alive 也預設 `5m`，避免 16GB RAM 主機與 GCP-B 被長駐 runner、長輸出與 24h keep-alive 壓到高 load。
 - ElephantAlpha 的 `price_drop_alert` / `market_opportunity` Telegram HITL 告警必須把同款證據獨立呈現，至少包含 `match_type`、`price_basis`、`alert_tier` 與 `match_score`；沒有高信心同款與總價可比證據時，不得把 PChome/MOMO 價差寫成可直接跟價建議。
 
 ## 一、四 AI Agent 路由架構
diff --git a/docs/memory/history_logs.md b/docs/memory/history_logs.md
index 33a5da5..e293746 100644
--- a/docs/memory/history_logs.md
+++ b/docs/memory/history_logs.md
@@ -13,6 +13,8 @@
 ## 📅 詳細更新日誌 (考古存檔)
 
 ### 2026-05-24：PChome 近門檻身份回收第二輪
+- **V10.408 OPI 指甲油 catalog review-only 回收**: marketplace matcher 針對 OPI 類光繚指甲油加入同系列 catalog focused identity，只在「白日夢遊」或「驕傲果凍」系列名雙方一致時進 `identity_review`；12色/11色視為可選色號數差異，不當作販售件數，跨系列仍維持 suppress。下一輪回刷會搭配 KATE 怪獸級持色唇膏限量款既有規則候選，延續不放寬全域門檻的近門檻回收策略。
+- **V10.407 Hermes/OpenClaw runner 熱駐留收斂**: V10.406 後續觀測顯示 GCP-B 仍會被 Hermes/OpenClaw 的 `24h` keep-alive runner 壓高 load，導致 GCP-B generate timeout 後轉落 111。`HERMES_KEEP_ALIVE` 與 `OPENCLAW_STRATEGY_OLLAMA_KEEP_ALIVE` 改為 env 可控且預設 `5m`，並補測試禁止 OpenClaw strategy 再硬寫 `keep_alive="24h"`；GCP-A 目前從 188 仍逾時，需另行修復主機/防火牆。
 - **V10.406 Code Review Ollama keep-alive 收斂**: production audit 顯示 Gemini 24h 已為 0，但 GCP-A `34.143.170.20:11434` 從 188 逾時、GCP-B 曾因多個 Ollama runner 長駐造成 generate timeout，導致部署後 Code Review 轉落 111。`CODE_REVIEW_OLLAMA_KEEP_ALIVE` 預設由 `24h` 改為 `5m`，讓 GCP-B/111 的 code review runner 不再長時間常駐；實測已重啟 GCP-B Ollama 並確認 `gemma3:4b` 可於約 6.2s 完成短生成。
 - **V10.405 MCP manual fetch handoff gate**: 新增 `mcp_manual_fetch_handoff` read-only builder、GET/POST endpoint、UI handoff package 審核面板與 deployment readiness smoke target，讓 runtime promotion package 搭配操作員公開頁面、節流、無登入/反爬、無 DB、無 scheduler 確認後，只放行到人工 fetch gate operator review；API/UI 不保存 payload、不打 health、不開 DB、不抓外站、不掛 scheduler，也不會自動打開 manual fetch。
 - **V10.404 Hermes 競價威脅漏斗只吃 direct alert**: `HermesAnalystService.fetch_candidates()` 的 `competitor_prices` JOIN 新增硬條件，只讓 `match_type=exact`、`price_basis=total_price`、`alert_tier=price_alert_exact` 的 identity_v2 配對進入 Hermes 競價威脅分析；`identity_review`、`unit_price_review` 與 `suppress` 仍保留在 dashboard / 人工覆核資料流，但不再消耗 Hermes token 或被上游視為直接價格威脅候選。production fresh 配對分布顯示直接告警約 497 筆、覆核型約 216 筆，本版將兩者在 AI 威脅入口切開。同版 matcher 追加 KATE 怪獸級持色唇膏、植村秀武士刀眉筆筆蕊、The Forest 焦糖楓葉擴香禮盒等近門檻 review-only 回收線，仍只進人工覆核，不直接價格告警。
diff --git a/services/hermes_analyst_service.py b/services/hermes_analyst_service.py
index e3af272..7f9bcd8 100644
--- a/services/hermes_analyst_service.py
+++ b/services/hermes_analyst_service.py
@@ -15,6 +15,7 @@ Hermes 3 競價情報分析服務 (Module 2)
 import asyncio
 import json
 import logging
+import os
 import re
 import time
 import uuid
@@ -33,7 +34,7 @@ logger = logging.getLogger(__name__)
 from config import HERMES_TIMEOUT
 
 HERMES_MODEL = "hermes3:latest"
-HERMES_KEEP_ALIVE = "24h"  # ADR-012：保持模型熱駐留，避免被別模型擠下後冷啟動 30+s timeout
+HERMES_KEEP_ALIVE = os.getenv("HERMES_KEEP_ALIVE", "5m")
 TOP_N = 20  # 輸出前 N 個威脅，控制 NemoTron 每次消耗配額
 
 
diff --git a/services/marketplace_product_matcher.py b/services/marketplace_product_matcher.py
index e356654..b0480d0 100644
--- a/services/marketplace_product_matcher.py
+++ b/services/marketplace_product_matcher.py
@@ -470,6 +470,7 @@ FOCUSED_IDENTITY_REVIEW_ONLY_REASONS = {
     "im_meme_fixx_cool_setting_spray",
     "so_natural_fixx_setting_spray_catalog",
     "kate_monster_lipstick_catalog",
+    "opi_gel_polish_series_catalog",
     "shu_auto_hard_formula_refill_catalog",
     "the_forest_maple_diffuser_flower_brandless",
 }
@@ -2951,6 +2952,13 @@ def _has_focused_low_score_exact_identity_line(left: ProductIdentity, right: Pro
         and "怪獸級持色唇膏" in right_text
     ):
         return "kate_monster_lipstick_catalog"
+    if (
+        "opi" in (left.brand_tokens & right.brand_tokens)
+        and "類光繚指甲油" in left_text
+        and "類光繚指甲油" in right_text
+        and any(series in left_text and series in right_text for series in ("白日夢遊", "驕傲果凍"))
+    ):
+        return "opi_gel_polish_series_catalog"
     if (
         {"shu uemura", "植村秀"} & (left.brand_tokens & right.brand_tokens)
         and "自動武士刀眉筆" in left_text
diff --git a/services/openclaw_strategist_service.py b/services/openclaw_strategist_service.py
index 35a93c1..ea79f04 100644
--- a/services/openclaw_strategist_service.py
+++ b/services/openclaw_strategist_service.py
@@ -81,6 +81,7 @@ OPENCLAW_STRATEGY_OLLAMA_MODEL = os.getenv(
 )
 OPENCLAW_STRATEGY_OLLAMA_TIMEOUT = int(os.getenv('OPENCLAW_STRATEGY_OLLAMA_TIMEOUT', '120'))
 OPENCLAW_STRATEGY_OLLAMA_NUM_PREDICT = int(os.getenv('OPENCLAW_STRATEGY_OLLAMA_NUM_PREDICT', '2048'))
+OPENCLAW_STRATEGY_OLLAMA_KEEP_ALIVE = os.getenv('OPENCLAW_STRATEGY_OLLAMA_KEEP_ALIVE', '5m')
 
 # 繁體中文強制 system prompt（A2 黃燈警訊「Qwen 繁中短板」緩解策略）
 QWEN3_TC_SYSTEM_PROMPT = """你是 momo 電商情報分析師「OpenClaw」。
@@ -1106,7 +1107,7 @@ def _call_ollama_strategy(
                 system_prompt=system_prompt,
                 temperature=temperature,
                 timeout=timeout_s,
-                keep_alive="24h",
+                keep_alive=OPENCLAW_STRATEGY_OLLAMA_KEEP_ALIVE,
                 options={"num_predict": predict},
             )
             ctx.set_provider(get_provider_tag(resp.host or ""))
diff --git a/tests/test_hermes_ollama_cascade.py b/tests/test_hermes_ollama_cascade.py
index de3b50a..f339ba2 100644
--- a/tests/test_hermes_ollama_cascade.py
+++ b/tests/test_hermes_ollama_cascade.py
@@ -154,3 +154,7 @@ def test_hermes_candidate_sql_only_joins_direct_price_alert_matches():
     assert "match_type_exact" in sql_text
     assert "price_basis_total_price" in sql_text
     assert "alert_tier_price_alert_exact" in sql_text
+
+
+def test_hermes_keep_alive_defaults_to_short_runner_residency():
+    assert hermes_mod.HERMES_KEEP_ALIVE == "5m"
diff --git a/tests/test_marketplace_product_matcher.py b/tests/test_marketplace_product_matcher.py
index 4e064ce..2e8b9fe 100644
--- a/tests/test_marketplace_product_matcher.py
+++ b/tests/test_marketplace_product_matcher.py
@@ -887,6 +887,14 @@ def test_marketplace_matcher_keeps_catalog_variant_recoveries_in_identity_review
             "【The Forest 癒森林】焦糖楓葉香氛擴香花禮盒 含30ml品牌香氛油(居家香氛/香水精油/香氛擴香花/擴香禮物)",
             "焦糖楓葉香氛擴香花禮盒 含30ml品牌香氛油",
         ),
+        (
+            "【O.P.I】類光繚指甲油 12色任選1瓶(小銀蓋/如膠似漆白日夢遊系列指彩/官方直營)",
+            "OPI 如膠似漆白日夢遊系列 類光繚指甲油12色任選",
+        ),
+        (
+            "【O.P.I】類光繚指甲油 12色任選1瓶(小銀蓋/如膠似漆驕傲果凍系列指彩/官方直營)",
+            "OPI 如膠似漆驕傲果凍系列 類光繚指甲油11色任選",
+        ),
     ]
 
     for momo_name, competitor_name in cases:
@@ -1199,6 +1207,16 @@ def test_marketplace_matcher_promotes_focused_low_score_exact_identity_lines():
             "焦糖楓葉香氛擴香花禮盒 含30ml品牌香氛油",
             "focused_exact_identity_the_forest_maple_diffuser_flower_brandless",
         ),
+        (
+            "【O.P.I】類光繚指甲油 12色任選1瓶(小銀蓋/如膠似漆白日夢遊系列指彩/官方直營)",
+            "OPI 如膠似漆白日夢遊系列 類光繚指甲油12色任選",
+            "focused_exact_identity_opi_gel_polish_series_catalog",
+        ),
+        (
+            "【O.P.I】類光繚指甲油 12色任選1瓶(小銀蓋/如膠似漆驕傲果凍系列指彩/官方直營)",
+            "OPI 如膠似漆驕傲果凍系列 類光繚指甲油11色任選",
+            "focused_exact_identity_opi_gel_polish_series_catalog",
+        ),
         (
             "【GATSBY】爆水擦澡濕巾24張入(涼感乾洗澡)",
             "GATSBY 爆水擦澡濕巾24張入(240g)",
@@ -1252,6 +1270,10 @@ def test_marketplace_matcher_keeps_high_variant_low_score_lines_outside_focused_
         "【YSL】官方直營 恆久完美持久柔霧蜜粉餅(任選1款/皮革蜜粉)",
         "YSL 恆久完美持久柔霧蜜粉餅(7.5g) [百貨公司專櫃貨]",
     )
+    opi_series_gap = score_marketplace_match(
+        "【O.P.I】類光繚指甲油 12色任選1瓶(小銀蓋/如膠似漆白日夢遊系列指彩/官方直營)",
+        "OPI 如膠似漆驕傲果凍系列 類光繚指甲油11色任選",
+    )
 
     for diagnostics in (
         lush,
@@ -1263,6 +1285,7 @@ def test_marketplace_matcher_keeps_high_variant_low_score_lines_outside_focused_
         lancome_line_gap,
         mac_finish_gap,
         ysl_powder_variant_guard,
+        opi_series_gap,
     ):
         assert diagnostics.score < 0.76
         assert not any(reason.startswith("focused_exact_identity_") for reason in diagnostics.reasons)
diff --git a/tests/test_qwen3_runtime_usage.py b/tests/test_qwen3_runtime_usage.py
index 3cecbae..77c9e5e 100644
--- a/tests/test_qwen3_runtime_usage.py
+++ b/tests/test_qwen3_runtime_usage.py
@@ -13,6 +13,9 @@ def test_qwen3_is_active_runtime_model_not_unused_ollama_weight():
     assert "def _call_qwen3_qa(" in openclaw_source
     assert "OllamaService(model=OPENCLAW_QA_OLLAMA_MODEL)" in openclaw_source
     assert "OPENCLAW_QA_OLLAMA_HOST" not in openclaw_source
+    assert "OPENCLAW_STRATEGY_OLLAMA_KEEP_ALIVE = os.getenv('OPENCLAW_STRATEGY_OLLAMA_KEEP_ALIVE', '5m')" in openclaw_source
+    assert "keep_alive=OPENCLAW_STRATEGY_OLLAMA_KEEP_ALIVE" in openclaw_source
+    assert 'keep_alive="24h"' not in openclaw_source
     assert 'NEMOTRON_OLLAMA_MODEL = os.getenv("NEMOTRON_OLLAMA_MODEL", "qwen3:14b")' in nemotron_source
     assert "def _call_qwen3_dispatch(" in nemotron_source
     assert "for _attempt in range(3):" in nemotron_source