From c39b2e9f3ee338057ce574eb868dfd651252120f Mon Sep 17 00:00:00 2001
From: zizi <zizi2000mail@gmail.com>
Date: Wed, 15 Jul 2026 14:44:20 +0800
Subject: [PATCH] =?UTF-8?q?=E6=A1=86=E6=9E=B6:=20chat()=E6=92=9E=E6=A8=A1?=
 =?UTF-8?q?=E5=9E=8Bmax=5Ftokens=E4=B8=8A=E9=99=90=E8=87=AA=E9=80=82?=
 =?UTF-8?q?=E5=BA=94=E9=99=8D=E6=A1=A3=E9=87=8D=E5=8F=91(M2.7=E4=B8=8A?=
 =?UTF-8?q?=E9=99=90196608=E5=AE=9E=E6=B5=8B,=E4=BF=AE=E9=99=8D=E7=BA=A7?=
 =?UTF-8?q?=E9=93=BE400=E6=96=AD=E9=93=BE)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .claude/skills/llm/scripts/llm.py | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/.claude/skills/llm/scripts/llm.py b/.claude/skills/llm/scripts/llm.py
index 7aa03bc..287c302 100644
--- a/.claude/skills/llm/scripts/llm.py
+++ b/.claude/skills/llm/scripts/llm.py
@@ -59,6 +59,17 @@ def chat(prompt, model=DEFAULT_MODEL, max_tokens=512000, temperature=0.2,
             r = s.post(f"{BASE}/v1/chat/completions",
                        headers={"Authorization": f"Bearer {TOKEN}"},
                        json=payload, timeout=timeout)
+            # 各模型 max_tokens 上限不一（实测 M3 收 512k、M2.7 上限 196608 报 400）：
+            # 撞上限自适应降档重发，保证降级链换模型时不被参数掀翻。
+            # 必须在敏感检查之前处理——重发后的响应仍要完整走敏感/429/5xx 判定链
+            if r.status_code == 400 and "max tokens" in r.text.lower():
+                cur = payload.get("max_tokens", 0)
+                payload["max_tokens"] = 196608 if cur > 196608 else 12000
+                print(f"[llm] {model} max_tokens={cur} 超模型上限，降为 {payload['max_tokens']} 重发",
+                      file=sys.stderr)
+                r = s.post(f"{BASE}/v1/chat/completions",
+                           headers={"Authorization": f"Bearer {TOKEN}"},
+                           json=payload, timeout=timeout)
             # 内容安全拦截：同模型退避重试必再敏感，立即抛 SensitiveError 交上层
             # 降级换模型，不在此浪费退避（否则一敏感章空烧 3 次，实测占放量请求 23%）
             if r.status_code >= 500 and "sensitive" in r.text.lower():