feat: ocg_router key降级改指数退避——按key独立统计失败次数,失败冷却base2s×2^count指数增加(上限600s/10分钟),成功后重置计数; 空输出/报错立即换key+该key进入指数退避

gateway/scripts/ocg_router.py

@@ -67,6 +67,7 @@ _keys: list[dict] = []  # [{key_id, label, api_key, workspace_id}]
 _key_usage: dict[str, dict] = {}  # key_id → {rolling_pct, weekly_pct, monthly_pct}
 _key_failures: dict[str, float] = {}  # key_id → fail_until_timestamp
 _key_failure_reason: dict[str, str] = {}  # key_id → 最近失败原因（冷却跳过时用于报错）
+_key_fail_count: dict[str, int] = {}  # key_id → 连续失败次数（指数退避基数，2026-08-14 老莫）
 _route_stats: dict[str, int] = {}  # key_id → hit_count
 _total_hits = 0
 _current_key_id = ""
@@ -233,24 +234,44 @@ def pick_key_round_robin():
         return _keys[0]
 
 
-def mark_key_failed(kid, reason="", cooldown=FAIL_COOLDOWN_SEC):
-    """标记 key 故障，进入冷却期。记录失败原因供冷却期报错。"""
-    until = time.time() + cooldown
+def mark_key_failed(kid, reason="", cooldown=None):
+    """标记 key 故障，进入指数退避冷却期（按 key 独立统计失败次数）。
+    2026-08-14 老莫设计：失败冷却时长指数增加（base × 2^count），上限 600s（10分钟）。
+    空输出/报错 → 立即换下一个 key，同时该 key 进入指数退避冷却。"""
+    now = time.time()
     with _state_lock:
+        # 失败计数 +1（独立统计）
+        _key_fail_count[kid] = _key_fail_count.get(kid, 0) + 1
+        count = _key_fail_count[kid]
+        # 指数退避：base=2s, 每次失败 ×2，上限 600s（10分钟）
+        base = 2
+        cool = min(base * (2 ** (count - 1)), 600)
+        until = now + cool
         _key_failures[kid] = until
         if reason:
             _key_failure_reason[kid] = reason
-    log.warning("key %s marked failed until %s (%ds): %s", kid,
-                datetime.fromtimestamp(until).strftime("%H:%M:%S"), cooldown, reason)
+    log.warning("key %s marked failed (count=%d, cooldown=%ds): %s", kid, count, cool, reason)
+
+
+def reset_key_fail(kid):
+    """key 成功 → 重置失败计数（清除指数退避）"""
+    with _state_lock:
+        if kid in _key_fail_count:
+            del _key_fail_count[kid]
+        if kid in _key_failures:
+            del _key_failures[kid]
+        if kid in _key_failure_reason:
+            del _key_failure_reason[kid]
 
 
 def record_route(kid):
-    """记录一次成功路由。"""
+    """记录一次成功路由 + 重置失败计数（指数退避清除）。"""
     global _total_hits, _current_key_id
     with _state_lock:
         _route_stats[kid] = _route_stats.get(kid, 0) + 1
         _total_hits += 1
         _current_key_id = kid
+    reset_key_fail(kid)  # 成功重置失败计数
 
 
 # ═══════════════════════════════════════════════════════════════