diff --git a/volc_limit_sync.py b/volc_limit_sync.py
index 991ce0b..c485672 100644
--- a/volc_limit_sync.py
+++ b/volc_limit_sync.py
@@ -97,6 +97,11 @@ BUFFER = 300000  # User requested 30w buffer
 CHECK_INTERVAL = 10
 THRESHOLD = 300000
 
+# 前置闸门预留量：请求开始判定 usage + reserve >= limit 即 429。
+# reserve 优先取本次请求自带 max_tokens（真实上限，零采集损失），
+# 未设置时 fallback 本常量。堵住"并发请求穿过闸门后集体撞线"的窗口。
+PREBLOCK_RESERVE_TOKENS = int(os.getenv("PREBLOCK_RESERVE_TOKENS", "50000"))
+
 # stable_limit 缺失时（首次同步前/新增模型/TTL 过期）的保守兜底限额。
 # 免费模型余额常态 < 2M（昨日采多少返多少），若 fallback 裸奔到采集上限 2M，
 # 余额不足的模型会被烧穿扣费。宁可短时收紧，不可裸奔。
@@ -597,6 +602,65 @@ def restore_blocked_models():
         log(f"❌ Restore check failed: {e}", level="ERROR")
 
 
+def restore_model_if_healthy(clean_model, allowed_hashes=None):
+    """
+    若指定模型已被软踢（_daily_limit_backup）且当前限额已恢复（usage < effective limit），
+    立即恢复其路由条目。供同步守护在写健康限额后调用，不等 restore_blocked_models 的
+    30min/4h 闸门；否则额度到账后模型会带额度空转被禁（有额度但 no healthy deployments）。
+    allowed_hashes: 限定匹配的 api_base hash 集合（None = 不限，按 clean 名匹配）。
+    返回恢复的条目数；任何异常 fail-safe 返回 0。
+    """
+    if not LITELLM_AVAILABLE or not redis_client:
+        return 0
+    try:
+        from litellm.proxy.proxy_server import llm_router
+
+        if not llm_router or not llm_router.model_list:
+            return 0
+
+        restored = 0
+        for i, model_entry in enumerate(llm_router.model_list):
+            backup = model_entry.get("_daily_limit_backup")
+            if not backup:
+                continue
+
+            params = model_entry.get("litellm_params", {})
+            m_model = params.get("model", "")
+            if clean_model_name(m_model) != clean_model:
+                continue
+
+            m_base = params.get("api_base", "")
+            if allowed_hashes is not None and generate_api_base_hash(
+                m_base
+            ) not in allowed_hashes:
+                continue
+
+            effective_limit = get_effective_limit(clean_model, m_base)
+            should_restore = False
+            if effective_limit is None:
+                should_restore = True
+            else:
+                usage_key = get_usage_key(clean_model, m_base)
+                current_val = float(redis_client.get(usage_key) or 0)
+                if current_val < float(effective_limit):
+                    should_restore = True
+
+            if should_restore:
+                dprint(f"♻️ [RESTORE] Restoring {backup['model_name']} (Quota OK)")
+                llm_router.model_list[i]["model_name"] = backup["model_name"]
+                llm_router.model_list[i]["litellm_params"]["api_base"] = m_base
+                del llm_router.model_list[i]["_daily_limit_backup"]
+                restored += 1
+
+        if restored > 0:
+            log(f"✅ Restored {restored} entry/entries of {clean_model}.", level="INFO")
+        return restored
+
+    except Exception as e:
+        log(f"❌ Restore failed for {clean_model}: {e}", level="ERROR")
+        return 0
+
+
 def startup_audit():
     """
     🚀 Startup Audit:
@@ -661,10 +725,12 @@ def startup_audit():
         log(f"❌ Startup audit failed: {e}", level="ERROR")
 
 
-def _check_request_block(model_name):
+def _check_request_block(model_name, reserve=0.0):
     """
     请求前闸门：对外模型名 → 判定是否超限。
     只拦截 MONITORED_MODEL_PAIRS 内的火山监控条目；newapi 等非监控渠道天然放行。
+    reserve: 本次请求的预估消耗（max_tokens 或常量兜底），usage + reserve >= limit 即拦截，
+    避免并发请求在闸门检查后集体撞线（限额是低频快照，用量是实时累计）。
     返回 (clean_model, usage, limit) 表示超限，None 表示放行。任何异常 fail-open。
     """
     if not redis_client or not LITELLM_AVAILABLE:
@@ -705,7 +771,7 @@ def _check_request_block(model_name):
             val = redis_client.get(usage_key)
             usage = float(val or 0)
 
-            if usage >= float(effective_limit):
+            if usage + float(reserve) >= float(effective_limit):
                 return (clean, usage, effective_limit)
 
         return None
@@ -744,13 +810,24 @@ class DailyLimitHandler(CustomLogger):
         if not data or not data.get("model"):
             return None
 
-        blocked = _check_request_block(data["model"])
+        # 预留本次请求的预估消耗：优先真实 max_tokens（零采集损失），
+        # 未设置/非法时 fallback 常量，避免并发请求集体撞线超限。
+        try:
+            reserve = float(
+                data.get("max_tokens")
+                or data.get("max_completion_tokens")
+                or PREBLOCK_RESERVE_TOKENS
+            )
+        except (TypeError, ValueError):
+            reserve = float(PREBLOCK_RESERVE_TOKENS)
+
+        blocked = _check_request_block(data["model"], reserve=reserve)
         if not blocked:
             return None
 
         clean, usage, limit = blocked
         log(
-            f"🚫 [PRE-BLOCK] {data['model']} daily limit reached ({int(usage):,}/{int(limit):,})",
+            f"🚫 [PRE-BLOCK] {data['model']} daily limit reached ({int(usage):,}+{int(reserve):,}/{int(limit):,})",
             level="WARNING",
         )
         raise RateLimitError(
@@ -1435,6 +1512,11 @@ def active_monitor_loop():
                                 flush=True,
                             )
 
+                        # 即时恢复：同步发现限额已恢复（usage < effective limit）时立刻解除软踢，
+                        # 不等 restore_blocked_models 的 30min/4h 闸门 —— 否则日返还额度到账后
+                        # 模型仍被禁（有额度但 no healthy deployments），直到下一次恢复检查。
+                        restore_model_if_healthy(target, target_hashes)
+
                         # Track min remaining for frequency control
                         configured_models_found = True