feat(translate): service.py 接 zhipu / config 加 zhipu_* / .env.example 加 ZHIPU 配置

配 zhipu.py(zhipu 模块)一起使用:
- service.py: 新增 _zhipu_translator(),引擎选择链路 spark → zhipu → tencent → maas → agnes → local
- config.py: 新增 zhipu_api_key / zhipu_base_url / zhipu_model / zhipu_interval_sec
- .env.example: 补 ZHIPU_* 字段说明

留空 zhipu_api_key = spark 不可用时直接降级 tencent(向后兼容)

.env.example

@@ -45,6 +45,15 @@ SPARK_MODEL=lite
 # 单次调用间隔(秒),避免被限速
 SPARK_INTERVAL_SEC=1.0
 
+# ===== 智谱 GLM(第二序位翻译;glm-4-flash 免费)=====
+# 留空 = 不启用智谱(spark 不可用时直接走 tencent)
+# 控制台 https://open.bigmodel.cn/ → API Keys → 新建
+ZHIPU_API_KEY=your_zhipu_api_key
+ZHIPU_BASE_URL=https://open.bigmodel.cn/api/paas/v4
+# 模型:glm-4-flash(默认,免费) / glm-4-air / glm-4-airx / glm-4 / glm-4-plus / glm-4.5 / glm-4.6
+ZHIPU_MODEL=glm-4-flash
+ZHIPU_INTERVAL_SEC=1.0
+
 # ===== 本地翻译(降级) =====
 # 不启用就留空:不会用本地模<E59CB0>?LOCAL_TRANSLATE_ENABLED=false
 LOCAL_TRANSLATE_MODEL=nllb-200-distilled-600M

backend/app/config.py

@@ -77,6 +77,13 @@ class Settings(BaseSettings):
     spark_model: str = "lite"
     spark_interval_sec: float = 1.0
 
+    # ===== 智谱 GLM(第二序位翻译;GLM-4-Flash 免费)=====
+    # 留空 = 不启用智谱(spark 不可用时直接降级到 tencent)
+    zhipu_api_key: str = ""
+    zhipu_base_url: str = "https://open.bigmodel.cn/api/paas/v4"
+    zhipu_model: str = "glm-4-flash"
+    zhipu_interval_sec: float = 1.0
+
     @field_validator("tencent_tmt_quota_buffer")
     @classmethod
     def _check_buffer(cls, v: float) -> float:

backend/app/services/translation/service.py

@@ -2,16 +2,17 @@
 
 引擎链路(优先级降序):
 1. spark(主,Lite 免费;spark_api_password 配了才用)
-2. tencent TMT(第二级,按月配额;快满时主动切走)
-3. tencent_maas(备用,OpenAI 兼容,无配额;主失败/TMT 配额耗尽时启用)
-4. agnes(第三级,通用 LLM 做翻译;MaaS 不可用时启用 — 质量次之但够用)
-5. local(最后兜底,需 settings.local_translate_enabled=true)
+2. zhipu(第二序位,GLM-4-Flash 免费;zhipu_api_key 配了才用)
+3. tencent TMT(第三级,按月配额;快满时主动切走)
+4. tencent_maas(备用,OpenAI 兼容,无配额;主失败/TMT 配额耗尽时启用)
+5. agnes(第四级,通用 LLM 做翻译;MaaS 不可用时启用 — 质量次之但够用)
+6. local(最后兜底,需 settings.local_translate_enabled=true)
 
 注:
 - TMT 是按月计费的(腾讯云后台可能计费口径是请求字节,我们 redis 累加的是字符数,
   差异约 2-3x);用户从腾讯云后台看"已用 2M"时,我们 redis 显示约 80 万字符
 - 用户决策:以腾讯云后台数字为准,快满时降级
-- spark 是 Lite 免费,默认走它;spark 不可用时降级到 tencent(继续吃配额)。
+- spark / zhipu 都是免费模型,默认优先;不可用时降级到 tencent(继续吃配额)。
   想要完全绕开 tencent,把 TENCENTCLOUD_SECRET_ID 留空即可。
 """
 from __future__ import annotations
@@ -30,6 +31,7 @@ from app.services.translation.local import LocalTranslator
 from app.services.translation.spark import SparkTranslator
 from app.services.translation.tencent import TencentTranslator
 from app.services.translation.tencent_maas import TencentMaaSTranslator
+from app.services.translation.zhipu import ZhipuTranslator
 
 logger = logging.getLogger("news.translate.service")
 
@@ -48,6 +50,7 @@ def _month_key() -> str:
 class TranslationService:
     def __init__(self):
         self._spark: BaseTranslator | None = None
+        self._zhipu: BaseTranslator | None = None
         self._tencent: BaseTranslator | None = None
         self._tencent_maas: BaseTranslator | None = None
         self._agnes: BaseTranslator | None = None
@@ -65,8 +68,18 @@ class TranslationService:
                 self._spark = None
         return self._spark
 
+    def _zhipu_translator(self) -> BaseTranslator | None:
+        """第二序位引擎:智谱 GLM(免费)。配了 zhipu_api_key 才启用。"""
+        if self._zhipu is None and settings.zhipu_api_key:
+            try:
+                self._zhipu = ZhipuTranslator()
+            except Exception as e:
+                logger.warning("zhipu init failed: %s", e)
+                self._zhipu = None
+        return self._zhipu
+
     def _primary(self) -> BaseTranslator | None:
-        """第二级:腾讯 TMT(初始化失败返回 None 表示不可用)。"""
+        """第三级:腾讯 TMT(初始化失败返回 None 表示不可用)。"""
         if self._tencent is None:
             try:
                 self._tencent = TencentTranslator()
@@ -155,10 +168,12 @@ class TranslationService:
             return TranslationResult(text=cached, engine="cache", chars=chars, cached=True)
 
         # 2) 选引擎
-        # 优先级:spark → tencent(配额)→ maas → agnes → local
+        # 优先级:spark → zhipu → tencent(配额)→ maas → agnes → local
         engine: BaseTranslator | None = None
         if self._spark_translator() is not None:
             engine = self._spark_translator()
+        elif self._zhipu_translator() is not None:
+            engine = self._zhipu_translator()
         elif await self.can_use_tencent(chars):
             engine = self._primary()
             if engine is None:
@@ -168,7 +183,7 @@ class TranslationService:
             engine = None
 
         if engine is None:
-            # spark 不可用 + 配额耗尽 / TMT 不可用:走备用链(maas → agnes → local)
+            # spark / zhipu 不可用 + 配额耗尽 / TMT 不可用:走备用链(maas → agnes → local)
             engine = self._fallback()
             if engine is None:
                 # 全无可用:返回原文 + 标记
@@ -189,10 +204,17 @@ class TranslationService:
                 res = await engine.translate(text, source=source, target=target)
             except Exception as e:
                 logger.exception("translate failed with %s: %s", engine.name, e)
-                # 失败时按 tencent_maas → local 顺序找一个不同的 fallback
-                # spark 失败时也要走 tencent(继续吃配额,因优先级只是降低不是禁用)
+                # 失败时按 zhipu → tencent → maas → local 顺序找一个不同的 fallback
+                # spark / zhipu 失败时也要走 tencent(继续吃配额,因优先级只是降低不是禁用)
                 fb: BaseTranslator | None = None
                 if engine.name == "spark":
+                    if self._zhipu_translator() is not None:
+                        fb = self._zhipu_translator()
+                    if fb is None and await self.can_use_tencent(chars):
+                        fb = self._primary()
+                    if fb is None:
+                        fb = self._maas() if engine.name != "tencent_maas" else None
+                elif engine.name == "zhipu":
                     if await self.can_use_tencent(chars):
                         fb = self._primary()
                     if fb is None:
@@ -212,7 +234,7 @@ class TranslationService:
                 raise RuntimeError(f"translation failed for {chars} chars (engine={engine.name})")
 
         # 4) 写缓存 — 只缓存真实翻译结果;失败/降级文本不缓存(避免污染 30 天)
-        if res.engine in ("spark", "tencent", "tencent_maas", "agnes", "nllb") and not res.cached:
+        if res.engine in ("spark", "zhipu", "tencent", "tencent_maas", "agnes", "nllb") and not res.cached:
             if "[翻译失败" not in res.text and "[本条未翻译" not in res.text:
                 try:
                     await r.set(ck, res.text, ex=60 * 60 * 24 * 30)  # 30 天

backend/app/services/translation/zhipu.py

@@ -0,0 +1,108 @@
+"""智谱 GLM 翻译(OpenAI 兼容协议)。
+
+- 端点:https://open.bigmodel.cn/api/paas/v4/chat/completions
+- 模型:glm-4-flash(免费,默认) / glm-4.5 / glm-4.6 / glm-4-air 等
+- 鉴权:Bearer <api_key>
+
+设计上独立于 LlmClient,专门走 zhipu_* 配置,避免和 LLM 智能增强共用 client 的节流。
+"""
+from __future__ import annotations
+
+import asyncio
+import logging
+import random
+
+import httpx
+
+from app.config import settings
+from app.services.translation.base import BaseTranslator, TranslationResult
+
+logger = logging.getLogger("news.translate.zhipu")
+
+
+# 经过实测,这套 prompt 在 GLM-4-Flash 上输出稳定(不夹带 reasoning / 注释)。
+_SYSTEM_PROMPT = """你是一个即时翻译助手。对于用户输入的英文或日文文章,请直接输出对应的简体中文译文。严格遵守:
+
+不要进行任何分步思考、不要输出分析、不要添加注释或说明。
+
+只输出中文译文本身,不包含任何额外文字(如"翻译结果:"、"以下是中文:"等)。
+
+如果输入内容既非英文也非日文,仅回复:"仅支持英文或日文翻译为中文。" """
+
+
+class ZhipuTranslator(BaseTranslator):
+    """智谱 GLM 翻译(OpenAI 兼容协议)。"""
+
+    name = "zhipu"
+
+    def __init__(self):
+        if not settings.zhipu_api_key:
+            raise RuntimeError("Zhipu API key missing")
+        self.api_key = settings.zhipu_api_key
+        self.base_url = settings.zhipu_base_url.rstrip("/")
+        self.model = settings.zhipu_model
+        self.interval_sec = settings.zhipu_interval_sec
+
+    def is_configured(self) -> bool:
+        return bool(self.api_key)
+
+    async def translate(
+        self, text: str, source: str = "auto", target: str = "zh"
+    ) -> TranslationResult:
+        if not text.strip():
+            return TranslationResult(text=text, engine=self.name, chars=0)
+        if not self.is_configured():
+            raise RuntimeError("Zhipu API key missing")
+
+        url = f"{self.base_url}/chat/completions"
+        payload = {
+            "model": self.model,
+            "messages": [
+                {"role": "system", "content": _SYSTEM_PROMPT},
+                {"role": "user", "content": text},
+            ],
+            "temperature": 0.0,
+            "max_tokens": max(256, len(text) * 3),
+            "stream": False,
+        }
+        headers = {
+            "Authorization": f"Bearer {self.api_key}",
+            "Content-Type": "application/json",
+        }
+
+        # 简单串行 + 重试 1 次
+        last_exc: Exception | None = None
+        for attempt in range(2):
+            try:
+                async with httpx.AsyncClient(timeout=60.0) as client:
+                    r = await client.post(url, json=payload, headers=headers)
+                if r.status_code >= 500:
+                    raise RuntimeError(f"Zhipu 5xx: {r.status_code} {r.text[:200]}")
+                if r.status_code == 429:
+                    raise RuntimeError(f"Zhipu 限流: {r.text[:200]}")
+                if r.status_code != 200:
+                    raise RuntimeError(f"Zhipu {r.status_code}: {r.text[:300]}")
+                data = r.json()
+                if "error" in data:
+                    raise RuntimeError(f"Zhipu error: {data['error']}")
+                content = (
+                    data.get("choices", [{}])[0]
+                    .get("message", {})
+                    .get("content", "")
+                    .strip()
+                )
+                if not content:
+                    raise RuntimeError(f"Zhipu empty content: {r.text[:300]}")
+                await asyncio.sleep(self.interval_sec)
+                return TranslationResult(
+                    text=content, engine=self.name, chars=len(text), cached=False
+                )
+            except Exception as e:
+                last_exc = e
+                logger.warning("zhipu attempt %s failed: %s", attempt, e)
+                if attempt == 0:
+                    await asyncio.sleep(0.5 + random.random())
+                else:
+                    raise
+        assert last_exc is not None
+        raise last_exc