feat: mofin_health 加 LLM 端点错误日志监控（规范5.5）

deploy/profile-scripts/mofin_health.py

@@ -996,6 +996,64 @@ def build_report():
     except Exception:
         self_check['recent_repairs'] = []
 
+    # ── LLM端点健康检查（2026-08-10新增：监控LLM调用失败，防静默故障）──
+    # 背景：opencode.ai deepseek-v4-flash 端点故障时，gateway进程活着但LLM全挂，
+    #       监控只查端口/进程发现不了。扫描hermes错误日志弥补。
+    llm_health = {"ok": True, "errors_recent": [], "scan_window_min": 60}
+    _LLM_ERR_PATTERNS = [
+        "RemoteProtocolError", "Stream stale", "empty stream",
+        "peer closed connection", "finish_reason", "API call failed",
+    ]
+    try:
+        _err_logs = [
+            Path('/home/hmo/.hermes/profiles/position-analyst/logs/errors.log'),
+            Path('/home/hmo/.hermes/profiles/position-analyst/logs/agent.log'),
+        ]
+        _recent_errs = []
+        _scan_cutoff = (now.timestamp() - 60 * llm_health["scan_window_min"])
+        for _lf in _err_logs:
+            if not _lf.exists():
+                continue
+            try:
+                _lines = _lf.read_text(encoding='utf-8', errors='ignore').splitlines()[-5000:]
+                for _line in _lines:
+                    # 提取时间戳（行首 YYYY-MM-DD HH:MM:SS）
+                    if len(_line) < 19:
+                        continue
+                    try:
+                        _ts = datetime.strptime(_line[:19], "%Y-%m-%d %H:%M:%S")
+                    except Exception:
+                        continue
+                    if _ts.timestamp() < _scan_cutoff:
+                        continue
+                    if any(_p in _line for _p in _LLM_ERR_PATTERNS):
+                        _recent_errs.append({
+                            "log": _lf.name, "time": _ts.strftime("%H:%M"),
+                            "msg": _line[20:180],
+                        })
+            except Exception:
+                pass
+        # 去重+限量
+        _seen = set()
+        _uniq = []
+        for _e in _recent_errs:
+            _k = _e["time"] + _e["msg"][:50]
+            if _k not in _seen:
+                _seen.add(_k)
+                _uniq.append(_e)
+        llm_health["errors_recent"] = _uniq[:10]
+        llm_health["error_count"] = len(_uniq)
+        llm_health["ok"] = len(_uniq) == 0
+        self_check["llm_health"] = llm_health
+        # 若LLM故障，功能树根节点标 fail
+        if not llm_health["ok"]:
+            feature_tree["status"] = "fail"
+            feature_tree.setdefault("warn_detail", {})["llm"] = (
+                f"LLM端点最近{llm_health['scan_window_min']}分钟{llm_health['error_count']}次错误"
+                f"（{llm_health['errors_recent'][0]['log'] if llm_health['errors_recent'] else '?'}）")
+    except Exception as _e:
+        self_check["llm_health"] = {"ok": True, "error": str(_e)[:100]}
+
     # ── 写JSON ──
     report = {
         "generated_at": now.strftime("%Y-%m-%d %H:%M:%S"),