From ff39dbcfa9ff94ecfee6d60623794aae96fe3ef0 Mon Sep 17 00:00:00 2001
From: xxm <xxm@mofin.local>
Date: Mon, 17 Aug 2026 11:47:07 +0800
Subject: [PATCH] =?UTF-8?q?fix:=20llm=5Fclient=E6=8C=89OCG-Router=E6=96=87?=
 =?UTF-8?q?=E6=A1=A3=E2=80=94=E2=80=94max=5Ftokens=E9=BB=98=E8=AE=A4None?=
 =?UTF-8?q?=E4=B8=8D=E5=86=99=E5=85=A5payload(=E6=8E=A8=E7=90=86=E6=A8=A1?=
 =?UTF-8?q?=E5=9E=8B=E7=83=A7=E5=85=89=E6=A0=B9=E5=9B=A0),timeout150?=
 =?UTF-8?q?=E2=86=92180?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 deploy/profile-scripts/llm_client.py | 21 +++++++++++++--------
 1 file changed, 13 insertions(+), 8 deletions(-)

diff --git a/deploy/profile-scripts/llm_client.py b/deploy/profile-scripts/llm_client.py
index 686bcd24..ca468f87 100644
--- a/deploy/profile-scripts/llm_client.py
+++ b/deploy/profile-scripts/llm_client.py
@@ -215,8 +215,13 @@ def gateway_alive(timeout=5):
         return False
 
 
-def call_llm(prompt, model=None, max_tokens=4096, timeout=150,
+def call_llm(prompt, model=None, max_tokens=None, timeout=180,
              retries=1, backoff=20, system=None, concurrent=False):
+    # 2026-08-17 OCG-Router文档对齐：
+    #   max_tokens 默认 None=不写入 payload（deepseek-v4-flash 是推理模型，指定 max_tokens
+    #   会被 reasoning tokens 烧光 → finish=length 空输出。不指定用模型默认 1M 上下文）。
+    #   timeout 150→180（文档普通请求建议）。
+    #   retries/backoff 保持 1/20：6 通道链下多重试=疯狂重试（文档禁止项），20s 退避防限流。
     """调用 LLM：OCG 直连优先，hermes gateway 兜底。带重试和结构化日志。
 
     Args:
@@ -235,18 +240,18 @@ def call_llm(prompt, model=None, max_tokens=4096, timeout=150,
         永远不抛异常到调用方。
     """
     model_name = model or REASSESS_MODEL
-    # OCG/deepseek 对过小 max_tokens 会短路返回空（2026-07-22 实测 max_tokens=64 必空）
-    if max_tokens < 512:
+    # 仅显式传入时才钳制下限（防过小短路，2026-07-22 实测 max_tokens=64 必空）；
+    # None = 不指定（遵循文档：推理模型不设上限）
+    if max_tokens is not None and max_tokens < 512:
         max_tokens = 512
     messages = []
     if system:
         messages.append({"role": "system", "content": system})
     messages.append({"role": "user", "content": prompt})
-    payload = json.dumps({
-        "model": model_name,
-        "messages": messages,
-        "max_tokens": max_tokens,
-    }).encode()
+    _req = {"model": model_name, "messages": messages}
+    if max_tokens is not None:
+        _req["max_tokens"] = max_tokens
+    payload = json.dumps(_req).encode()
 
     # ── 构建通道列表：SenseNova(快,同款model) → OCG路由代理(本地,自动选key) → key池直连 → gateway 兜底 ──
     # 2026-08-13 老莫修复 ocg_router bug 后，路由代理作为 OCG 主通道（自动选最空闲key+故障冷却）。