fix(config): 为火山渠道白名单 reasoning_effort，修复思考强度被静默丢弃

LiteLLM 的 VolcEngineChatConfig.get_supported_openai_params() 白名单不含
reasoning_effort，客户端一传就抛 UnsupportedParamsError；叠加 drop_params
后更糟 —— Router._deployment_params_with_request_reasoning_override 会先剥掉
部署侧 extra_body/thinking，drop_params 再把客户端参数丢掉，两者同归于尽，
上游 body 皆无该字段，方舟回落默认 high（静默失效）。

改动：17 个 volcengine 条目全部加 allowed_openai_params: [reasoning_effort]。
- 透传组（huoshan/deepseek-v4-*、doubao-seed-2-1-*、glm-5.2）：客户端显式指定生效
- 档位别名组（doubao-seed-nothink/-minimal/-low/-medium/-high、fanyi）：
  extra_body 钉死值为默认档位，客户端显式传参可覆盖；不传则走钉死档位

语义为「默认档位 + 客户端可覆盖」，非硬锁：LiteLLM 无纯配置手段实现强制覆盖。
实测依据见 config.yaml 注释（本地 echo 抓上游请求体）。

config.yaml

@@ -36,40 +36,6 @@ model_list:
 
 
 
-### 
-
-
-  #日常使用ds v3.2  ，强制think  
-#  - model_name: deepseek-v3.2
-#    litellm_params:
-#      <<: *token_pricing
-#      model: volcengine/deepseek-v3-2-251201
-#      api_base: os.environ/HUOSHAN_BASE_URL
-#      api_key: os.environ/HUOSHAN_API_KEY 
-#       说明中建议设置30分钟(1800秒)以上，避免深度思考过程中断连接
-#      timeout: 1800       
-#       关键配置2：启用思考模式
-#       extra_body 中的参数会被直接合并到请求的 JSON body 中
-#      extra_headers:
-#        Connection: "close"
-#      extra_body:
-#        thinking:
-#          type: "enabled"
-#    model_info:
-#      max_input_tokens: 128000   # 允许的最大输入长度（单位：token）
-#      max_context_length: 128000 # 上下文窗口上限
-          
-#  - model_name: deepseek-v3.2-nothink
-#    litellm_params:
-#      <<: *token_pricing
-#      model: volcengine/deepseek-v3-2-251201
-#      api_base: os.environ/HUOSHAN_BASE_URL
-#      api_key: os.environ/HUOSHAN_API_KEY
-#      extra_headers:
-#        Connection: "close" 
-#      extra_body:
-#        thinking:
-#          type: "disabled" 
  
   
   - model_name: huoshan/deepseek-v4-flash
@@ -78,6 +44,7 @@ model_list:
       model: volcengine/deepseek-v4-flash-ga-260731
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
 
   - model_name: huoshan/deepseek-v4-pro
@@ -86,101 +53,11 @@ model_list:
       model: volcengine/deepseek-v4-pro-ga-260813
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
   
         
-  - model_name: doubao-seed-1-8
+     
-    litellm_params:
-      <<: *token_pricing
-      model: volcengine/doubao-seed-1-8-251228
-      api_base: os.environ/HUOSHAN_BASE_URL
-      api_key: os.environ/HUOSHAN_API_KEY
-      drop_params: true
-      extra_headers:
-        Connection: "close" 
-
-  - model_name: doubao-seed-1-8-nothink
-    litellm_params:
-      <<: *token_pricing
-      model: volcengine/doubao-seed-1-8-251228
-      api_base: os.environ/HUOSHAN_BASE_URL
-      api_key: os.environ/HUOSHAN_API_KEY
-      extra_headers:
-        Connection: "close"
-      drop_params: true 
-      extra_body:
-        reasoning_effort: "minimal"        
-
-
-#  - model_name: doubao-seed-2-0-pro
-#    litellm_params:
-#      <<: *token_pricing
-#      model: volcengine/doubao-seed-2-0-pro-260215
-#      api_base: os.environ/HUOSHAN_BASE_URL
-#      api_key: os.environ/HUOSHAN_API_KEY
-#      timeout: 1800 
-#      drop_params: true 
-#      extra_headers:
-#        Connection: "close"
-
-        
-#  - model_name: doubao-seed-2-0-lite
-#    litellm_params:
-#      <<: *token_pricing
-#      model: volcengine/doubao-seed-2-0-lite-260215
-#      api_base: os.environ/HUOSHAN_BASE_URL
-#      api_key: os.environ/HUOSHAN_API_KEY
-#      drop_params: true
-#      extra_headers:
-#        Connection: "close"
-
-  - model_name: doubao-seed-2-0-code
-    litellm_params:
-      <<: *token_pricing
-      model: volcengine/doubao-seed-2-0-code-preview-260215
-      api_base: os.environ/HUOSHAN_BASE_URL
-      api_key: os.environ/HUOSHAN_API_KEY
-      drop_params: true
-      extra_headers:
-        Connection: "close"
-    model_info:
-      max_input_tokens: 224000   # 允许的最大输入长度（单位：token）
-      max_context_length: 256000 # 上下文窗口上限
-
-#禁用思考，用于标题等
-  - model_name: doubao-seed-2-0-code-nothink
-    litellm_params:
-      <<: *token_pricing
-      model: volcengine/doubao-seed-2-0-code-preview-260215
-      api_base: os.environ/HUOSHAN_BASE_URL
-      api_key: os.environ/HUOSHAN_API_KEY
-      drop_params: true
-      extra_headers:
-        Connection: "close"
-      extra_body:
-#        thinking:
-#          type: "disable"
-        reasoning_effort: "minimal"
-    model_info:
-      max_input_tokens: 224000   # 允许的最大输入长度（单位：token）
-      max_context_length: 256000 # 上下文窗口上限
-
-  # doubao-seed-2-0-code 强制中等思考，用作回退
-  - model_name: doubao-seed-2-0-code-medium
-    litellm_params:
-      <<: *token_pricing
-      model: volcengine/doubao-seed-2-0-code-preview-260215
-      api_base: os.environ/HUOSHAN_BASE_URL
-      api_key: os.environ/HUOSHAN_API_KEY
-      timeout: 1800
-      drop_params: true  
-      extra_headers:
-        Connection: "close"
-      extra_body:
-        reasoning_effort: "medium"
-    model_info:
-      max_input_tokens: 224000   # 允许的最大输入长度（单位：token）
-      max_context_length: 256000 # 上下文窗口上限        
 
   - model_name: doubao-seed-2-1-turbo
     litellm_params:
@@ -188,6 +65,7 @@ model_list:
       model: volcengine/doubao-seed-2-1-turbo-260628
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
       extra_headers:
         Connection: "close"
@@ -201,6 +79,7 @@ model_list:
       model: volcengine/doubao-seed-2-1-pro-260628
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
       extra_headers:
         Connection: "close"
@@ -215,6 +94,17 @@ model_list:
       model: volcengine/glm-5-2-260617
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY 
+      # 方舟 glm-5-2-260617 支持 reasoning_effort（none/minimal/high/max，默认 high；
+      # 官方参数页 + 实测 curl 均确认），但 LiteLLM 的 VolcEngineChatConfig
+      # get_supported_openai_params() 白名单里没有它 —— 它是 openai-like 的固定
+      # 参数表，不含 reasoning_effort —— 客户端一传就抛 UnsupportedParamsError。
+      # allowed_openai_params 是官方逃生口：get_optional_params 用它在校验前
+      # extend supported_params，随后 _apply_openai_param_overrides 把该参数从
+      # non_default_params 原样写回 optional_params，最终透传到请求体。
+      # 不用 drop_params：那是把该模型所有不支持参数静默丢弃，会掩盖其它真错。
+      # 注意 thinking.type 无需此项，VolcEngineChatConfig 已原生白名单。
+      allowed_openai_params:
+        - reasoning_effort
       extra_headers:
         Connection: "close"
     model_info:
@@ -222,18 +112,22 @@ model_list:
       max_context_length: 1024000 # 上下文窗口上限
 
   #=======================豆包汇总nothink
+  # 【思考档位别名组】doubao-seed-nothink / -minimal / -low / -medium / -high / fanyi
+  # 语义：别名后缀即【默认】档位，非硬锁。
+  #   - 客户端不传 reasoning_effort：按 extra_body 钉的档位发给方舟。
+  #   - 客户端显式传 reasoning_effort：以客户端为准（覆盖默认档位）。
+  # 为什么钉不成硬锁（实测，本地 echo 抓上游 body）：
+  #   请求带 reasoning_effort 时，Router._deployment_params_with_request_reasoning_override
+  #   会【无条件】剥掉部署侧 extra_body.reasoning_effort / thinking；
+  #   此时若没有 allowed_openai_params 放行，drop_params 又把客户端这个参数丢掉，
+  #   两者同归于尽 ⇒ 上游 body 两者都没有 ⇒ 方舟回落默认 high（静默失效）。
+  #   故必须加 allowed_openai_params：让客户端值能通过校验，使结果至少是
+  #   「客户端说了算」而不是「悄悄变回 high」。
+  #   LiteLLM 当前无纯配置手段做到「不论客户端传什么一律覆盖」；要硬锁只能加
+  #   pre-call hook（本仓库当前不采用）。
+  # 注：方舟 doubao 上 minimal 等价于关闭思考（实测 reasoning_tokens=0），
+  #     故 nothink/minimal/fanyi 三个别名默认档位统一用 minimal。
 
-  - model_name: doubao-seed-nothink
-    litellm_params:
-      <<: *token_pricing
-      model: volcengine/doubao-seed-2-0-code-preview-260215
-      api_base: os.environ/HUOSHAN_BASE_URL
-      api_key: os.environ/HUOSHAN_API_KEY
-      drop_params: true
-      extra_headers:
-        Connection: "close"
-      extra_body:
-        reasoning_effort: "minimal"
 
   - model_name: doubao-seed-nothink
     litellm_params:
@@ -241,6 +135,7 @@ model_list:
       model: volcengine/doubao-seed-2-1-pro-260628
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
       extra_headers:
         Connection: "close"
@@ -253,6 +148,7 @@ model_list:
       model: volcengine/doubao-seed-2-1-turbo-260628
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
       extra_headers:
         Connection: "close"
@@ -263,17 +159,6 @@ model_list:
 
   #=======================豆包汇总思考等级minimal
 
-  - model_name: doubao-seed-minimal
-    litellm_params:
-      <<: *token_pricing
-      model: volcengine/doubao-seed-2-0-code-preview-260215
-      api_base: os.environ/HUOSHAN_BASE_URL
-      api_key: os.environ/HUOSHAN_API_KEY
-      drop_params: true
-      extra_headers:
-        Connection: "close"
-      extra_body:
-        reasoning_effort: "minimal"
 
   - model_name: doubao-seed-minimal
     litellm_params:
@@ -281,6 +166,7 @@ model_list:
       model: volcengine/doubao-seed-2-1-pro-260628
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
       extra_headers:
         Connection: "close"
@@ -293,6 +179,7 @@ model_list:
       model: volcengine/doubao-seed-2-1-turbo-260628
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
       extra_headers:
         Connection: "close"
@@ -302,29 +189,7 @@ model_list:
 
   #=======================豆包汇总思考等级low
 
-  - model_name: doubao-seed-1-8
-    litellm_params:
-      <<: *token_pricing
-      model: volcengine/doubao-seed-1-8-251228
-      api_base: os.environ/HUOSHAN_BASE_URL
-      api_key: os.environ/HUOSHAN_API_KEY
-      drop_params: true
-      extra_headers:
-        Connection: "close" 
-      extra_body:
-        reasoning_effort: "high"
 
-  - model_name: doubao-seed-low
-    litellm_params:
-      <<: *token_pricing
-      model: volcengine/doubao-seed-2-0-code-preview-260215
-      api_base: os.environ/HUOSHAN_BASE_URL
-      api_key: os.environ/HUOSHAN_API_KEY
-      drop_params: true
-      extra_headers:
-        Connection: "close"
-      extra_body:
-        reasoning_effort: "low"
 
   - model_name: doubao-seed-low
     litellm_params:
@@ -332,6 +197,7 @@ model_list:
       model: volcengine/doubao-seed-2-1-pro-260628
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
       extra_headers:
         Connection: "close"
@@ -344,6 +210,7 @@ model_list:
       model: volcengine/doubao-seed-2-1-turbo-260628
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
       extra_headers:
         Connection: "close"
@@ -353,17 +220,6 @@ model_list:
 
   #=======================豆包汇总思考等级medium
 
-  - model_name: doubao-seed-medium
-    litellm_params:
-      <<: *token_pricing
-      model: volcengine/doubao-seed-2-0-code-preview-260215
-      api_base: os.environ/HUOSHAN_BASE_URL
-      api_key: os.environ/HUOSHAN_API_KEY
-      drop_params: true
-      extra_headers:
-        Connection: "close"
-      extra_body:
-        reasoning_effort: "medium"
 
   - model_name: doubao-seed-medium
     litellm_params:
@@ -371,6 +227,7 @@ model_list:
       model: volcengine/doubao-seed-2-1-pro-260628
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
       extra_headers:
         Connection: "close"
@@ -383,6 +240,7 @@ model_list:
       model: volcengine/doubao-seed-2-1-turbo-260628
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
       extra_headers:
         Connection: "close"
@@ -392,17 +250,7 @@ model_list:
 
   #=======================豆包汇总思考等级high
 
-  - model_name: doubao-seed-high
+
-    litellm_params:
-      <<: *token_pricing
-      model: volcengine/doubao-seed-2-0-code-preview-260215
-      api_base: os.environ/HUOSHAN_BASE_URL
-      api_key: os.environ/HUOSHAN_API_KEY
-      drop_params: true
-      extra_headers:
-        Connection: "close"
-      extra_body:
-        reasoning_effort: "high"
 
   - model_name: doubao-seed-high
     litellm_params:
@@ -410,6 +258,7 @@ model_list:
       model: volcengine/doubao-seed-2-1-pro-260628
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
       extra_headers:
         Connection: "close"
@@ -422,6 +271,7 @@ model_list:
       model: volcengine/doubao-seed-2-1-turbo-260628
       api_base: os.environ/HUOSHAN_BASE_URL
       api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
       drop_params: true
       extra_headers:
         Connection: "close"
@@ -434,30 +284,29 @@ model_list:
   - model_name: fanyi
     litellm_params:
       <<: *token_pricing
-      model: volcengine/doubao-seed-1-8-251228
+      model: volcengine/doubao-seed-2-1-pro-260628
       api_base: os.environ/HUOSHAN_BASE_URL
-      api_key: os.environ/HUOSHAN_API_KEY 
+      api_key: os.environ/HUOSHAN_API_KEY
-      timeout: 120 
+      allowed_openai_params: ["reasoning_effort"]
-      stream: false 
       drop_params: true
       extra_headers:
-        Connection: "close"               
+        Connection: "close"
       extra_body:
-        reasoning_effort: "minimal"    
+        reasoning_effort: "minimal"
+
+  - model_name: fanyi
+    litellm_params:
+      <<: *token_pricing
+      model: volcengine/doubao-seed-2-1-turbo-260628
+      api_base: os.environ/HUOSHAN_BASE_URL
+      api_key: os.environ/HUOSHAN_API_KEY
+      allowed_openai_params: ["reasoning_effort"]
+      drop_params: true
+      extra_headers:
+        Connection: "close"
+      extra_body:
+        reasoning_effort: "minimal"   
 
-#  - model_name: fanyi
-#    litellm_params:
-#      <<: *token_pricing
-#      model: volcengine/doubao-seed-2-0-lite-260215
-#      api_base: os.environ/HUOSHAN_BASE_URL
-#      api_key: os.environ/HUOSHAN_API_KEY
-#      timeout: 120 
-#      stream: false 
-#      drop_params: true
-#      extra_headers:
-#        Connection: "close"               
-#      extra_body:
-#        reasoning_effort: "minimal"
 
           
 
@@ -469,7 +318,7 @@ litellm_settings:
   #drop_params: true  #启用drop_params，这样LiteLLM会忽略不支持的参数，但参数仍然会传递给后端API
   num_retries: 3 # retry call 3 times on each model_name (e.g. zephyr-beta)
   fallbacks: [
-    {"doubao-seed-high": ["deepseek-v4-pro"]} 
+    #{"doubao-seed-high": ["deepseek-v4-pro"]} 
   ]
   #default_fallbacks: ["deepseek-v4-flash"]
   request_timeout: 120 # raise Timeout error if call takes longer than 10s. Sets litellm.request_timeout 

docker-compose.yml

@@ -21,8 +21,28 @@ services:
       TZ: Asia/Shanghai
       # ⚠️ 必须设置这个环境变量，否则 Python 找不到你的脚本
       PYTHONPATH: "/app"
-      DATABASE_URL: "postgresql://litellm_app:shriek-sixfold-ethanol@192.168.1.34:5432/litellm"
+      # ⚠️ 必须保留 client_encoding=UTF8。
+      # 本库 litellm 是 SQL_ASCII 编码（server/client_encoding 均为 SQL_ASCII，
+      # datcollate=C）。psycopg 对 SQL_ASCII 连接的 TextLoader 直接返回 bytes
+      # 而非 str（psycopg/types/string.py:110-120），于是 litellm_proxy_extras
+      # 里 _MIGRATION_TS_RE.match(migration_name)、
+      # _TRANSIENT_INDEX_SUFFIX_RE.search(index.name) 等对 bytes 用字符串正则，
+      # 抛 TypeError: cannot use a string pattern on a bytes-like object。
+      # 该参数经 libpq 作为启动参数下发（fe-protocol3.c:2503），使连接按 UTF8
+      # 协商；psycopg 读服务端确认值决定解码（pq_ctypes.py:229-232），
+      # 从而正常返回 str。Prisma 侧无影响：quaint 丢弃未知参数
+      # （prisma-engines quaint/src/connector/postgres/url.rs:459），
+      # 且其驱动 tokio-postgres 本就固定发送 client_encoding=UTF8。
+      # 已全库校验：450 个文本列 / 43548 行 / 14.5MB 全部是合法 UTF-8，
+      # 不存在解码失败风险。
+      DATABASE_URL: "postgresql://litellm_app:shriek-sixfold-ethanol@192.168.1.34:5432/litellm?client_encoding=UTF8"
       STORE_MODEL_IN_DB: "True"
+      # v2 migration resolver 的额外兜底（可去掉以启用 v2）：
+      # main-latest 自 2026-09-20 起默认启用 v2（上游 commit fac518dbf7），
+      # v2 会在启动时调用 _warn_if_db_ahead_of_head() —— 即上面那个 TypeError
+      # 的触发点。上面的 client_encoding 修复已从根因上解除该崩溃，
+      # 但 v2 还引入迁移协调锁等新机制；保守起见先沿用更新前的 v1 路径。
+      USE_V2_MIGRATION_RESOLVER: "false"
       # --- Modified: 添加 Redis 配置 ---
       REDIS_HOST: "redis"       # 对应下方 redis 服务的名字
       REDIS_PORT: "6379"

volc_limit_sync.py

@@ -138,6 +138,24 @@ redis.call('EXPIRE', KEYS[2], tonumber(ARGV[3]))
 return 1
 """
 
+# Lua 原子释放：结算时配对扣减 inflight，floor 0 防止迟到结算的幽灵释放
+# 误删别的请求的预留（TTL 过期后迟到 DECRBY 导致负数 → DEL 全部）。
+# SET 后补 EXPIRE，保留剩余预留的 TTL（不因 SET 重置为永久）。
+# KEYS[1]=inflight_key, ARGV[1]=reserve, ARGV[2]=inflight TTL
+# 返回: 扣减后剩余 inflight 值（>=0）
+RELEASE_SCRIPT = """
+local cur = tonumber(redis.call('GET', KEYS[1]) or '0')
+local reserve = tonumber(ARGV[1])
+local new_val = cur - reserve
+if new_val <= 0 then
+    redis.call('DEL', KEYS[1])
+    return 0
+end
+redis.call('SET', KEYS[1], new_val)
+redis.call('EXPIRE', KEYS[1], tonumber(ARGV[2]))
+return new_val
+"""
+
 # stable_limit 缺失时（首次同步前/新增模型/TTL 过期/Redis 数据丢失）的保守兜底限额。
 # 免费池常态 ~2M（首开赠送/昨日采多少返多少），若 fallback 裸奔到采集上限 2M，
 # 余额不足或已耗尽的模型会被烧穿扣费。900k 对 0 余额模型仍是 9 倍优惠，
@@ -927,20 +945,18 @@ def _atomic_reserve(clean_model, api_base, effective_limit, reserve):
 def _release_inflight(clean_model, api_base, reserve):
     """结算配对释放：请求完成后把 pre-hook 预扣的 reserve 从 inflight 账户扣除。
 
-    reserve<=0 或 inflight 已过期（TTL 兜底释放）时无操作。失败/丢失的预留
+    用 Lua 原子脚本做 DECRBY + floor 0，防止 TTL 过期后迟到结算的幽灵释放
-    由 TTL 自动归零，这里只做正常结算路径的精确回收。
+    把 DECRBY 打到负数 → 误删别的请求的预留。reserve<=0 或脚本失败时无操作，
+    inflight 残留由 TTL 兜底归零。
     """
     if not redis_client or not reserve:
         return
     try:
         inflight_key = get_inflight_key(clean_model, api_base)
-        cur = float(redis_client.get(inflight_key) or 0)
+        redis_client.eval(
-        if cur <= 0:
+            RELEASE_SCRIPT, 1, inflight_key, int(reserve),
-            return
+            int(PREBLOCK_INFLIGHT_TTL),
-        # DECRBY 到负数说明 inflight 已部分过期，直接 DEL 归零即可
+        )
-        new_val = redis_client.decrby(inflight_key, int(reserve))
-        if int(new_val) <= 0:
-            redis_client.delete(inflight_key)
     except Exception as e:
         log(f"❌ _release_inflight failed: {e}", level="ERROR")
 
@@ -1142,6 +1158,22 @@ class DailyLimitHandler(CustomLogger):
         # 输出侧：思考模型 usage 含 reasoning tokens 且不受 max_tokens 限制，
         # 故取 max(请求声明 max_tokens, THINKING_RESERVE_FLOOR) 兜底思考开销，
         # 防短文本 + 小 max_tokens 请求 reserve 低估（实测 34 vs 实际 150+）。
+        # 幂等防重：litellm 对同一请求可能调用两次 async_pre_call_hook
+        # （proxy 级 + deployment 级），第二次预扣会重复占 inflight 且
+        # 结算只释放一次 → 残留累积。已有 reserve 标记时跳过预扣（含
+        # token_counter 精确计数等昂贵路径）直接放行。
+        # 假设：deployment-level hook 收到的是同一 data dict 引用（litellm
+        # pre_call_hook 循环体 data=result 赋值，就地修改会带过去）。若
+        # litellm 升级为深拷贝 data 传给第二次 hook，此检查失效，需改用
+        # 请求级唯一 ID（如 litellm_call_id）做去重。部署后留意日志确认
+        # 第二次 pre-hook 是否被跳过（Pre-reserve 只应打印一次）。
+        try:
+            _existing_meta = data.get("metadata") or {}
+            if isinstance(_existing_meta, dict) and PREBLOCK_RESERVE_META_KEY in _existing_meta:
+                return data
+        except Exception:
+            pass
+
         try:
             declared = (
                 data.get("max_tokens")