V10.420 guard 111 Ollama LAN access

2026-05-24 15:27:11 +08:00
parent 0d81839447
commit 2f1ba21352
4 changed files with 123 additions and 2 deletions
--- 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 = "V10.419"
+SYSTEM_VERSION = "V10.420"
 LOG_FILE_PATH = os.path.join(BASE_DIR, 'logs/system.log')
 public_url = PUBLIC_URL  # 用於模板顯示

--- a/docs/AI_INTELLIGENCE_MODULE_SOT.md
+++ b/docs/AI_INTELLIGENCE_MODULE_SOT.md
@@ -2,7 +2,7 @@

 > **最後更新**: 2026-05-24 (台北時間)
 > **狀態**: 🟢 四 AI Agent 自動化閉環已落地；LLM 路由紅線升級為 Ollama-first 三主機級聯，Gemini 備援預設關閉
-> **適用版本**: V10.419
+> **適用版本**: V10.420

 ---

@@ -31,6 +31,7 @@
 - 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。OpenClaw 報告型路徑的業務 keep-alive 預設 `5m`；Code Review 以 `CODE_REVIEW_ALLOW_111_FALLBACK=false`、Hermes 以 `HERMES_ALLOW_111_FALLBACK=false` 預設跳過 111，避免 16GB RAM 主機與 GCP-B 被長駐 runner、長輸出與 24h keep-alive 壓到高 load。
+- 111 的 LAN 入口必須經 `scripts/ops/ollama111_allow_proxy.py` allowlist proxy：真實 Ollama 綁 `127.0.0.1:11434`，proxy 綁 `192.168.0.111:11434`，預設只允許 111 本機與 188 生產宿主；110 / 121 / 其他 LAN client 不能直接打 111，避免跨專案 CI 或 VM 繞過 momo-pro router 載入 7B+ runner。
 - 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
--- a/docs/memory/history_logs.md
+++ b/docs/memory/history_logs.md
@@ -13,6 +13,7 @@
 ## 📅 詳細更新日誌 (考古存檔)

 ### 2026-05-24：PChome 近門檻身份回收第二輪
+- **V10.420 111 Ollama LAN allowlist proxy**: 追查 111 高負載時確認來源不是 momo-pro，而是 110 上 `awoooi-cd` 臨時測試與 121 VMware VM 直接打 `192.168.0.111:11434`，繞過 `ai_calls` 與 momo-pro router 載入 7B runner。新增 `scripts/ops/ollama111_allow_proxy.py`，將真實 Ollama 收斂到 `127.0.0.1:11434`，由 user-space proxy 綁 `192.168.0.111:11434` 並預設只允許 111 本機與 188 生產宿主；110 / 121 會被 reset，111 fallback 保留給 momo production。
 - **V10.419 Dr.Hsieh LabSmart 精華品線防錯配**: marketplace matcher 追加 `dr_hsieh_labsmart_line_conflict`，只針對 Dr.Hsieh/達特醫的 `LabSmart Hi-Tech` / `LabSmart Classic` 精華被拿去對 `神經醯胺多重修復保濕精華液` 的近門檻錯配做 hard veto；同品牌同容量但不同產品線不再因規格相同停在 `true_low_confidence` 或被誤推進比價。
 - **V10.419 production pilot**: 正式回刷 SKU `10413050` / `10413051`，兩筆 Dr.Hsieh LabSmart 精華 vs 神經醯胺多重修復精華皆由 `true_low_confidence` 轉為 `identity_veto`，`diagnostic_codes=["dr_hsieh_labsmart_line_conflict"]`；`matched` 維持 1619、`true_low_confidence` 753→751、`identity_veto` 4011→4013，無正式 `competitor_prices` 覆寫。
 - **V10.418 bge-m3 一致性檢查不打 111**: `verify_embedding_consistency()` 預設只比對 GCP-A / GCP-B，不再每週把 111 Mac 納入 bge-m3 背景驗證；新增 `EMBED_CONSISTENCY_INCLUDE_111=false` 預設，只有救急需要驗證 fallback 模型時才 opt-in。這補上 V10.417 後仍會由監測任務載入 111/GCP-B embedding runner 的缺口。
--- a/scripts/ops/ollama111_allow_proxy.py
+++ b/scripts/ops/ollama111_allow_proxy.py
@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+"""
+User-space allowlist TCP proxy for the 111 Ollama fallback.
+
+Purpose:
+  - Keep the real Ollama server bound to 127.0.0.1:11434.
+  - Expose 192.168.0.111:11434 only to approved momo-pro clients.
+  - Reject noisy LAN / VM clients without requiring sudo/pfctl.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import ipaddress
+import logging
+import os
+import signal
+from typing import Iterable
+
+
+LISTEN_HOST = os.getenv("OLLAMA111_PROXY_LISTEN_HOST", "192.168.0.111")
+LISTEN_PORT = int(os.getenv("OLLAMA111_PROXY_LISTEN_PORT", "11434"))
+TARGET_HOST = os.getenv("OLLAMA111_PROXY_TARGET_HOST", "127.0.0.1")
+TARGET_PORT = int(os.getenv("OLLAMA111_PROXY_TARGET_PORT", "11434"))
+ALLOWED_CIDRS = tuple(
+    item.strip()
+    for item in os.getenv(
+        "OLLAMA111_PROXY_ALLOWED_CIDRS",
+        "127.0.0.1/32,192.168.0.80/32,192.168.0.111/32,192.168.0.188/32",
+    ).split(",")
+    if item.strip()
+)
+
+
+def _allowed_networks() -> tuple[ipaddress._BaseNetwork, ...]:
+    return tuple(ipaddress.ip_network(item, strict=False) for item in ALLOWED_CIDRS)
+
+
+ALLOWED_NETWORKS = _allowed_networks()
+
+
+def _is_allowed(peer_ip: str) -> bool:
+    try:
+        ip = ipaddress.ip_address(peer_ip)
+    except ValueError:
+        return False
+    return any(ip in network for network in ALLOWED_NETWORKS)
+
+
+async def _pipe(reader: asyncio.StreamReader, writer: asyncio.StreamWriter) -> None:
+    try:
+        while True:
+            data = await reader.read(65536)
+            if not data:
+                break
+            writer.write(data)
+            await writer.drain()
+    except (ConnectionError, asyncio.CancelledError):
+        pass
+    finally:
+        try:
+            writer.close()
+            await writer.wait_closed()
+        except Exception:
+            pass
+
+
+async def _handle_client(reader: asyncio.StreamReader, writer: asyncio.StreamWriter) -> None:
+    peer = writer.get_extra_info("peername")
+    peer_ip = peer[0] if peer else "unknown"
+    if not _is_allowed(peer_ip):
+        logging.warning("reject peer=%s allowed=%s", peer_ip, ",".join(ALLOWED_CIDRS))
+        writer.close()
+        await writer.wait_closed()
+        return
+
+    try:
+        target_reader, target_writer = await asyncio.open_connection(TARGET_HOST, TARGET_PORT)
+    except Exception as exc:
+        logging.error("target connect failed peer=%s target=%s:%s err=%r", peer_ip, TARGET_HOST, TARGET_PORT, exc)
+        writer.close()
+        await writer.wait_closed()
+        return
+
+    logging.info("proxy peer=%s -> %s:%s", peer_ip, TARGET_HOST, TARGET_PORT)
+    await asyncio.gather(
+        _pipe(reader, target_writer),
+        _pipe(target_reader, writer),
+    )
+
+
+async def _main() -> None:
+    logging.basicConfig(
+        level=os.getenv("OLLAMA111_PROXY_LOG_LEVEL", "INFO"),
+        format="%(asctime)s %(levelname)s %(message)s",
+    )
+    server = await asyncio.start_server(_handle_client, LISTEN_HOST, LISTEN_PORT)
+    sockets = ", ".join(str(sock.getsockname()) for sock in (server.sockets or []))
+    logging.info(
+        "ollama111 allow proxy listening=%s target=%s:%s allowed=%s",
+        sockets,
+        TARGET_HOST,
+        TARGET_PORT,
+        ",".join(ALLOWED_CIDRS),
+    )
+
+    stop = asyncio.Event()
+    loop = asyncio.get_running_loop()
+    for sig in (signal.SIGINT, signal.SIGTERM):
+        loop.add_signal_handler(sig, stop.set)
+
+    async with server:
+        await stop.wait()
+        server.close()
+        await server.wait_closed()
+
+
+if __name__ == "__main__":
+    asyncio.run(_main())